1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /** 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include <sys/param.h> 42 #include <sys/arb.h> 43 #include <sys/module.h> 44 #include <sys/kernel.h> 45 #include <sys/libkern.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> 52 #include <sys/socket.h> 53 #include <sys/socketvar.h> 54 #include <sys/sysctl.h> 55 #include <sys/systm.h> 56 #ifdef STATS 57 #include <sys/qmath.h> 58 #include <sys/tree.h> 59 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 60 #endif 61 #include <sys/refcount.h> 62 #include <sys/queue.h> 63 #include <sys/eventhandler.h> 64 #include <sys/smp.h> 65 #include <sys/kthread.h> 66 #include <sys/lock.h> 67 #include <sys/mutex.h> 68 #include <sys/tim_filter.h> 69 #include <sys/time.h> 70 #include <sys/protosw.h> 71 #include <vm/uma.h> 72 #include <sys/kern_prefetch.h> 73 74 #include <net/route.h> 75 #include <net/route/nhop.h> 76 #include <net/vnet.h> 77 78 #define TCPSTATES /* for logging */ 79 80 #include <netinet/in.h> 81 #include <netinet/in_kdtrace.h> 82 #include <netinet/in_pcb.h> 83 #include <netinet/ip.h> 84 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 85 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 86 #include <netinet/ip_var.h> 87 #include <netinet/ip6.h> 88 #include <netinet6/in6_pcb.h> 89 #include <netinet6/ip6_var.h> 90 #define TCPOUTFLAGS 91 #include <netinet/tcp.h> 92 #include <netinet/tcp_fsm.h> 93 #include <netinet/tcp_seq.h> 94 #include <netinet/tcp_timer.h> 95 #include <netinet/tcp_var.h> 96 #include <netinet/tcpip.h> 97 #include <netinet/tcp_hpts.h> 98 #include <netinet/cc/cc.h> 99 #include <netinet/tcp_log_buf.h> 100 #include <netinet/tcp_ratelimit.h> 101 #include <netinet/tcp_lro.h> 102 #ifdef TCPDEBUG 103 #include <netinet/tcp_debug.h> 104 #endif /* TCPDEBUG */ 105 #ifdef TCP_OFFLOAD 106 #include <netinet/tcp_offload.h> 107 #endif 108 #ifdef INET6 109 #include <netinet6/tcp6_var.h> 110 #endif 111 #include <netinet/tcp_fastopen.h> 112 113 #include <netipsec/ipsec_support.h> 114 #include <net/if.h> 115 #include <net/if_var.h> 116 #include <net/ethernet.h> 117 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 119 #include <netipsec/ipsec.h> 120 #include <netipsec/ipsec6.h> 121 #endif /* IPSEC */ 122 123 #include <netinet/udp.h> 124 #include <netinet/udp_var.h> 125 #include <machine/in_cksum.h> 126 127 #ifdef MAC 128 #include <security/mac/mac_framework.h> 129 #endif 130 131 #include "sack_filter.h" 132 #include "tcp_bbr.h" 133 #include "rack_bbr_common.h" 134 uma_zone_t bbr_zone; 135 uma_zone_t bbr_pcb_zone; 136 137 struct sysctl_ctx_list bbr_sysctl_ctx; 138 struct sysctl_oid *bbr_sysctl_root; 139 140 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 141 (tv) = (value); \ 142 if ((u_long)(tv) < (u_long)(tvmin)) \ 143 (tv) = (tvmin); \ 144 if ((u_long)(tv) > (u_long)(tvmax)) \ 145 (tv) = (tvmax); \ 146 } while(0) 147 148 /*#define BBR_INVARIANT 1*/ 149 150 /* 151 * initial window 152 */ 153 static uint32_t bbr_def_init_win = 10; 154 static int32_t bbr_persist_min = 250000; /* 250ms */ 155 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 156 static int32_t bbr_cwnd_may_shrink = 0; 157 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 158 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 159 static int32_t bbr_hardware_pacing_limit = 8000; 160 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 161 static int32_t bbr_no_retran = 0; 162 163 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 164 static int32_t bbr_max_net_error_cnt = 10; 165 /* Should the following be dynamic too -- loss wise */ 166 static int32_t bbr_rtt_gain_thresh = 0; 167 /* Measurement controls */ 168 static int32_t bbr_use_google_algo = 1; 169 static int32_t bbr_ts_limiting = 1; 170 static int32_t bbr_ts_can_raise = 0; 171 static int32_t bbr_do_red = 600; 172 static int32_t bbr_red_scale = 20000; 173 static int32_t bbr_red_mul = 1; 174 static int32_t bbr_red_div = 2; 175 static int32_t bbr_red_growth_restrict = 1; 176 static int32_t bbr_target_is_bbunit = 0; 177 static int32_t bbr_drop_limit = 0; 178 /* 179 * How much gain do we need to see to 180 * stay in startup? 181 */ 182 static int32_t bbr_marks_rxt_sack_passed = 0; 183 static int32_t bbr_start_exit = 25; 184 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 185 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 186 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 187 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 188 * if we go back ever to where the pacer 189 * has priority over timers. 190 */ 191 static int32_t bbr_policer_call_from_rack_to = 0; 192 static int32_t bbr_policer_detection_enabled = 1; 193 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 194 * measurements before we are 195 * "good" note that 2 == 1. 196 * This is because we use a > 197 * comparison. This means if 198 * min_measure was 0, it takes 199 * num-measures > min(0) and 200 * you get 1 measurement and 201 * you are good. Set to 1, you 202 * have to have two 203 * measurements (this is done 204 * to prevent it from being ok 205 * to have no measurements). */ 206 static int32_t bbr_no_pacing_until = 4; 207 208 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 209 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 210 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 211 212 static int32_t bbr_target_cwnd_mult_limit = 8; 213 /* 214 * bbr_cwnd_min_val is the number of 215 * segments we hold to in the RTT probe 216 * state typically 4. 217 */ 218 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 219 220 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 221 222 static int32_t bbr_gain_to_target = 1; 223 static int32_t bbr_gain_gets_extra_too = 1; 224 /* 225 * bbr_high_gain is the 2/ln(2) value we need 226 * to double the sending rate in startup. This 227 * is used for both cwnd and hptsi gain's. 228 */ 229 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 230 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 231 static int32_t bbr_use_lower_gain_in_startup = 1; 232 233 /* thresholds for reduction on drain in sub-states/drain */ 234 static int32_t bbr_drain_rtt = BBR_SRTT; 235 static int32_t bbr_drain_floor = 88; 236 static int32_t google_allow_early_out = 1; 237 static int32_t google_consider_lost = 1; 238 static int32_t bbr_drain_drop_mul = 4; 239 static int32_t bbr_drain_drop_div = 5; 240 static int32_t bbr_rand_ot = 50; 241 static int32_t bbr_can_force_probertt = 0; 242 static int32_t bbr_can_adjust_probertt = 1; 243 static int32_t bbr_probertt_sets_rtt = 0; 244 static int32_t bbr_can_use_ts_for_rtt = 1; 245 static int32_t bbr_is_ratio = 0; 246 static int32_t bbr_sub_drain_app_limit = 1; 247 static int32_t bbr_prtt_slam_cwnd = 1; 248 static int32_t bbr_sub_drain_slam_cwnd = 1; 249 static int32_t bbr_slam_cwnd_in_main_drain = 1; 250 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 251 * hold */ 252 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 253 /* 254 * bbr_drain_gain is the reverse of the high_gain 255 * designed to drain back out the standing queue 256 * that is formed in startup by causing a larger 257 * hptsi gain and thus drainging the packets 258 * in flight. 259 */ 260 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 261 static int32_t bbr_rttprobe_gain = 192; 262 263 /* 264 * The cwnd_gain is the default cwnd gain applied when 265 * calculating a target cwnd. Note that the cwnd is 266 * a secondary factor in the way BBR works (see the 267 * paper and think about it, it will take some time). 268 * Basically the hptsi_gain spreads the packets out 269 * so you never get more than BDP to the peer even 270 * if the cwnd is high. In our implemenation that 271 * means in non-recovery/retransmission scenarios 272 * cwnd will never be reached by the flight-size. 273 */ 274 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 275 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 276 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 277 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 278 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 279 static int32_t bbr_ignore_data_after_close = 1; 280 static int16_t bbr_hptsi_gain[] = { 281 (BBR_UNIT *5 / 4), 282 (BBR_UNIT * 3 / 4), 283 BBR_UNIT, 284 BBR_UNIT, 285 BBR_UNIT, 286 BBR_UNIT, 287 BBR_UNIT, 288 BBR_UNIT 289 }; 290 int32_t bbr_use_rack_resend_cheat = 1; 291 int32_t bbr_sends_full_iwnd = 1; 292 293 #define BBR_HPTSI_GAIN_MAX 8 294 /* 295 * The BBR module incorporates a number of 296 * TCP ideas that have been put out into the IETF 297 * over the last few years: 298 * - Yuchung Cheng's RACK TCP (for which its named) that 299 * will stop us using the number of dup acks and instead 300 * use time as the gage of when we retransmit. 301 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 302 * of Dukkipati et.al. 303 * - Van Jacobson's et.al BBR. 304 * 305 * RACK depends on SACK, so if an endpoint arrives that 306 * cannot do SACK the state machine below will shuttle the 307 * connection back to using the "default" TCP stack that is 308 * in FreeBSD. 309 * 310 * To implement BBR and RACK the original TCP stack was first decomposed 311 * into a functional state machine with individual states 312 * for each of the possible TCP connection states. The do_segment 313 * functions role in life is to mandate the connection supports SACK 314 * initially and then assure that the RACK state matches the conenction 315 * state before calling the states do_segment function. Data processing 316 * of inbound segments also now happens in the hpts_do_segment in general 317 * with only one exception. This is so we can keep the connection on 318 * a single CPU. 319 * 320 * Each state is simplified due to the fact that the original do_segment 321 * has been decomposed and we *know* what state we are in (no 322 * switches on the state) and all tests for SACK are gone. This 323 * greatly simplifies what each state does. 324 * 325 * TCP output is also over-written with a new version since it 326 * must maintain the new rack scoreboard and has had hptsi 327 * integrated as a requirment. Still todo is to eliminate the 328 * use of the callout_() system and use the hpts for all 329 * timers as well. 330 */ 331 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 332 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 333 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 334 * free list */ 335 static int32_t bbr_tlp_thresh = 1; 336 static int32_t bbr_reorder_thresh = 2; 337 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 338 * 60,000,000 - 60 seconds */ 339 static int32_t bbr_pkt_delay = 1000; 340 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 341 static int32_t bbr_incr_timers = 1; 342 343 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 344 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 345 static int32_t bbr_exit_startup_at_loss = 1; 346 347 /* 348 * bbr_lt_bw_ratio is 1/8th 349 * bbr_lt_bw_diff is < 4 Kbit/sec 350 */ 351 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 352 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 353 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 354 * the lt_bw for */ 355 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 356 * lt_bw */ 357 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 358 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 359 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 360 361 static int32_t bbr_verbose_logging = 0; 362 /* 363 * Currently regular tcp has a rto_min of 30ms 364 * the backoff goes 12 times so that ends up 365 * being a total of 122.850 seconds before a 366 * connection is killed. 367 */ 368 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 369 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 370 371 /****************************************************/ 372 /* DEFAULT TSO SIZING (cpu performance impacting) */ 373 /****************************************************/ 374 /* What amount is our formula using to get TSO size */ 375 static int32_t bbr_hptsi_per_second = 1000; 376 377 /* 378 * For hptsi under bbr_cross_over connections what is delay 379 * target 7ms (in usec) combined with a seg_max of 2 380 * gets us close to identical google behavior in 381 * TSO size selection (possibly more 1MSS sends). 382 */ 383 static int32_t bbr_hptsi_segments_delay_tar = 7000; 384 385 /* Does pacing delay include overhead's in its time calculations? */ 386 static int32_t bbr_include_enet_oh = 0; 387 static int32_t bbr_include_ip_oh = 1; 388 static int32_t bbr_include_tcp_oh = 1; 389 static int32_t bbr_google_discount = 10; 390 391 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 392 static int32_t bbr_state_is_pkt_epoch = 0; 393 static int32_t bbr_state_drain_2_tar = 1; 394 /* What is the max the 0 - bbr_cross_over MBPS TSO target 395 * can reach using our delay target. Note that this 396 * value becomes the floor for the cross over 397 * algorithm. 398 */ 399 static int32_t bbr_hptsi_segments_max = 2; 400 static int32_t bbr_hptsi_segments_floor = 1; 401 static int32_t bbr_hptsi_utter_max = 0; 402 403 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 404 static int32_t bbr_hptsi_bytes_min = 1460; 405 static int32_t bbr_all_get_min = 0; 406 407 /* Cross over point from algo-a to algo-b */ 408 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 409 410 /* Do we deal with our restart state? */ 411 static int32_t bbr_uses_idle_restart = 0; 412 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 413 414 /* Do we allow hardware pacing? */ 415 static int32_t bbr_allow_hdwr_pacing = 0; 416 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 417 static int32_t bbr_hdwr_pace_floor = 1; 418 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 419 420 /****************************************************/ 421 static int32_t bbr_resends_use_tso = 0; 422 static int32_t bbr_tlp_max_resend = 2; 423 static int32_t bbr_sack_block_limit = 128; 424 425 #define BBR_MAX_STAT 19 426 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 427 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 428 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 429 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 430 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 431 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 432 counter_u64_t bbr_flows_whdwr_pacing; 433 counter_u64_t bbr_flows_nohdwr_pacing; 434 435 counter_u64_t bbr_nohdwr_pacing_enobuf; 436 counter_u64_t bbr_hdwr_pacing_enobuf; 437 438 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 439 440 /* 441 * Static defintions we need for forward declarations. 442 */ 443 static uint32_t 444 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 445 uint32_t useconds_time, uint64_t bw); 446 static uint32_t 447 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 448 static void 449 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 450 static void 451 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 452 static void 453 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 454 int dolog); 455 static uint32_t 456 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 457 static void 458 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 459 int32_t pkt_epoch, uint32_t losses); 460 static uint32_t 461 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, 462 struct bbr_sendmap *rsm); 463 static uint32_t 464 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 465 static uint32_t 466 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 467 struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); 468 static void 469 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 470 int32_t line); 471 static void 472 bbr_set_state_target(struct tcp_bbr *bbr, int line); 473 static void 474 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 475 static void 476 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, 477 int event, int line); 478 static void 479 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 480 static void 481 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 482 static void 483 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 484 uint32_t rtt, uint32_t line, uint8_t is_start, 485 uint16_t set); 486 static struct bbr_sendmap * 487 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 488 static __inline uint32_t 489 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 490 static void 491 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, 492 uint8_t which); 493 static void 494 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, 495 uint32_t time_since_sent, uint32_t srtt, 496 uint32_t thresh, uint32_t to); 497 static void 498 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 499 static void 500 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 501 uint32_t del_by, uint32_t cts, uint32_t sloton, 502 uint32_t prev_delay); 503 static void 504 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 505 int32_t line); 506 static void 507 bbr_stop_all_timers(struct tcpcb *tp); 508 static void 509 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 510 static void 511 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 512 static void 513 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 514 static void 515 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 516 uint32_t cts, uint32_t usecs, uint64_t bw, 517 uint32_t override, int mod); 518 static int 519 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt); 520 521 static inline uint8_t 522 bbr_state_val(struct tcp_bbr *bbr) 523 { 524 return(bbr->rc_bbr_substate); 525 } 526 527 static inline uint32_t 528 get_min_cwnd(struct tcp_bbr *bbr) 529 { 530 int mss; 531 532 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 533 bbr->r_ctl.rc_pace_max_segs); 534 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 535 return (bbr_cwnd_min_val_hs * mss); 536 else 537 return (bbr_cwnd_min_val * mss); 538 } 539 540 static uint32_t 541 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 542 { 543 uint64_t srtt, var; 544 uint64_t ret_val; 545 546 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 547 if (tp->t_srtt == 0) { 548 srtt = (uint64_t)BBR_INITIAL_RTO; 549 var = 0; 550 } else { 551 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 552 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 553 } 554 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 555 bbr_persist_min, bbr_persist_max); 556 return ((uint32_t)ret_val); 557 } 558 559 static uint32_t 560 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 561 { 562 /* 563 * Start the FR timer, we do this based on getting the first one in 564 * the rc_tmap. Note that if its NULL we must stop the timer. in all 565 * events we need to stop the running timer (if its running) before 566 * starting the new one. 567 */ 568 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 569 int32_t idx; 570 int32_t is_tlp_timer = 0; 571 struct bbr_sendmap *rsm; 572 573 if (bbr->rc_all_timers_stopped) { 574 /* All timers have been stopped none are to run */ 575 return (0); 576 } 577 if (bbr->rc_in_persist) { 578 /* We can't start any timer in persists */ 579 return (bbr_get_persists_timer_val(tp, bbr)); 580 } 581 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 582 if ((rsm == NULL) || 583 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 584 (tp->t_state < TCPS_ESTABLISHED)) { 585 /* Nothing on the send map */ 586 activate_rxt: 587 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 588 uint64_t tov; 589 590 time_since_sent = 0; 591 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 592 if (rsm) { 593 idx = rsm->r_rtr_cnt - 1; 594 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 595 tstmp_touse = rsm->r_tim_lastsent[idx]; 596 else 597 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 598 if (TSTMP_GT(tstmp_touse, cts)) 599 time_since_sent = cts - tstmp_touse; 600 } 601 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 602 if (tp->t_srtt == 0) 603 tov = BBR_INITIAL_RTO; 604 else 605 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 606 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 607 if (tp->t_rxtshift) 608 tov *= tcp_backoff[tp->t_rxtshift]; 609 if (tov > time_since_sent) 610 tov -= time_since_sent; 611 else 612 tov = bbr->r_ctl.rc_min_to; 613 TCPT_RANGESET_NOSLOP(to, tov, 614 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 615 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 616 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 617 return (to); 618 } 619 return (0); 620 } 621 if (rsm->r_flags & BBR_ACKED) { 622 rsm = bbr_find_lowest_rsm(bbr); 623 if (rsm == NULL) { 624 /* No lowest? */ 625 goto activate_rxt; 626 } 627 } 628 /* Convert from ms to usecs */ 629 if (rsm->r_flags & BBR_SACK_PASSED) { 630 if ((tp->t_flags & TF_SENTFIN) && 631 ((tp->snd_max - tp->snd_una) == 1) && 632 (rsm->r_flags & BBR_HAS_FIN)) { 633 /* 634 * We don't start a bbr rack timer if all we have is 635 * a FIN outstanding. 636 */ 637 goto activate_rxt; 638 } 639 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 640 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 641 idx = rsm->r_rtr_cnt - 1; 642 exp = rsm->r_tim_lastsent[idx] + thresh; 643 if (SEQ_GEQ(exp, cts)) { 644 to = exp - cts; 645 if (to < bbr->r_ctl.rc_min_to) { 646 to = bbr->r_ctl.rc_min_to; 647 } 648 } else { 649 to = bbr->r_ctl.rc_min_to; 650 } 651 } else { 652 /* Ok we need to do a TLP not RACK */ 653 if (bbr->rc_tlp_in_progress != 0) { 654 /* 655 * The previous send was a TLP. 656 */ 657 goto activate_rxt; 658 } 659 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 660 if (rsm == NULL) { 661 /* We found no rsm to TLP with. */ 662 goto activate_rxt; 663 } 664 if (rsm->r_flags & BBR_HAS_FIN) { 665 /* If its a FIN we don't do TLP */ 666 rsm = NULL; 667 goto activate_rxt; 668 } 669 time_since_sent = 0; 670 idx = rsm->r_rtr_cnt - 1; 671 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 672 tstmp_touse = rsm->r_tim_lastsent[idx]; 673 else 674 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 675 if (TSTMP_GT(tstmp_touse, cts)) 676 time_since_sent = cts - tstmp_touse; 677 is_tlp_timer = 1; 678 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 679 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 680 if (thresh > time_since_sent) 681 to = thresh - time_since_sent; 682 else 683 to = bbr->r_ctl.rc_min_to; 684 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 685 /* 686 * If the TLP time works out to larger than the max 687 * RTO lets not do TLP.. just RTO. 688 */ 689 goto activate_rxt; 690 } 691 if ((bbr->rc_tlp_rtx_out == 1) && 692 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 693 /* 694 * Second retransmit of the same TLP 695 * lets not. 696 */ 697 bbr->rc_tlp_rtx_out = 0; 698 goto activate_rxt; 699 } 700 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 701 /* 702 * The tail is no longer the last one I did a probe 703 * on 704 */ 705 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 706 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 707 } 708 } 709 if (is_tlp_timer == 0) { 710 BBR_STAT_INC(bbr_to_arm_rack); 711 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 712 } else { 713 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 714 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 715 /* 716 * We have exceeded how many times we can retran the 717 * current TLP timer, switch to the RTO timer. 718 */ 719 goto activate_rxt; 720 } else { 721 BBR_STAT_INC(bbr_to_arm_tlp); 722 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 723 } 724 } 725 return (to); 726 } 727 728 static inline int32_t 729 bbr_minseg(struct tcp_bbr *bbr) 730 { 731 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 732 } 733 734 static void 735 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 736 { 737 struct inpcb *inp; 738 struct hpts_diag diag; 739 uint32_t delayed_ack = 0; 740 uint32_t left = 0; 741 uint32_t hpts_timeout; 742 uint8_t stopped; 743 int32_t delay_calc = 0; 744 uint32_t prev_delay = 0; 745 746 inp = tp->t_inpcb; 747 if (tcp_in_hpts(inp)) { 748 /* A previous call is already set up */ 749 return; 750 } 751 if ((tp->t_state == TCPS_CLOSED) || 752 (tp->t_state == TCPS_LISTEN)) { 753 return; 754 } 755 stopped = bbr->rc_tmr_stopped; 756 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 757 left = bbr->r_ctl.rc_timer_exp - cts; 758 } 759 bbr->r_ctl.rc_hpts_flags = 0; 760 bbr->r_ctl.rc_timer_exp = 0; 761 prev_delay = bbr->r_ctl.rc_last_delay_val; 762 if (bbr->r_ctl.rc_last_delay_val && 763 (slot == 0)) { 764 /* 765 * If a previous pacer delay was in place we 766 * are not coming from the output side (where 767 * we calculate a delay, more likely a timer). 768 */ 769 slot = bbr->r_ctl.rc_last_delay_val; 770 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 771 /* Compensate for time passed */ 772 delay_calc = cts - bbr->rc_pacer_started; 773 if (delay_calc <= slot) 774 slot -= delay_calc; 775 } 776 } 777 /* Do we have early to make up for by pushing out the pacing time? */ 778 if (bbr->r_agg_early_set) { 779 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 780 slot += bbr->r_ctl.rc_agg_early; 781 bbr->r_ctl.rc_agg_early = 0; 782 bbr->r_agg_early_set = 0; 783 } 784 /* Are we running a total debt that needs to be compensated for? */ 785 if (bbr->r_ctl.rc_hptsi_agg_delay) { 786 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 787 /* We nuke the delay */ 788 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 789 bbr->r_ctl.rc_hptsi_agg_delay = 0; 790 } else { 791 /* We nuke some of the delay, put in a minimal 100usecs */ 792 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 793 bbr->r_ctl.rc_last_delay_val = slot = 100; 794 } 795 } 796 bbr->r_ctl.rc_last_delay_val = slot; 797 hpts_timeout = bbr_timer_start(tp, bbr, cts); 798 if (tp->t_flags & TF_DELACK) { 799 if (bbr->rc_in_persist == 0) { 800 delayed_ack = bbr_delack_time; 801 } else { 802 /* 803 * We are in persists and have 804 * gotten a new data element. 805 */ 806 if (hpts_timeout > bbr_delack_time) { 807 /* 808 * Lets make the persists timer (which acks) 809 * be the smaller of hpts_timeout and bbr_delack_time. 810 */ 811 hpts_timeout = bbr_delack_time; 812 } 813 } 814 } 815 if (delayed_ack && 816 ((hpts_timeout == 0) || 817 (delayed_ack < hpts_timeout))) { 818 /* We need a Delayed ack timer */ 819 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 820 hpts_timeout = delayed_ack; 821 } 822 if (slot) { 823 /* Mark that we have a pacing timer up */ 824 BBR_STAT_INC(bbr_paced_segments); 825 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 826 } 827 /* 828 * If no timers are going to run and we will fall off thfe hptsi 829 * wheel, we resort to a keep-alive timer if its configured. 830 */ 831 if ((hpts_timeout == 0) && 832 (slot == 0)) { 833 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 834 (tp->t_state <= TCPS_CLOSING)) { 835 /* 836 * Ok we have no timer (persists, rack, tlp, rxt or 837 * del-ack), we don't have segments being paced. So 838 * all that is left is the keepalive timer. 839 */ 840 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 841 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 842 } else { 843 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 844 } 845 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 846 } 847 } 848 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 849 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 850 /* 851 * RACK, TLP, persists and RXT timers all are restartable 852 * based on actions input .. i.e we received a packet (ack 853 * or sack) and that changes things (rw, or snd_una etc). 854 * Thus we can restart them with a new value. For 855 * keep-alive, delayed_ack we keep track of what was left 856 * and restart the timer with a smaller value. 857 */ 858 if (left < hpts_timeout) 859 hpts_timeout = left; 860 } 861 if (bbr->r_ctl.rc_incr_tmrs && slot && 862 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 863 /* 864 * If configured to do so, and the timer is either 865 * the TLP or RXT timer, we need to increase the timeout 866 * by the pacing time. Consider the bottleneck at my 867 * machine as an example, we are sending something 868 * to start a TLP on. The last packet won't be emitted 869 * fully until the pacing time (the bottleneck will hold 870 * the data in place). Once the packet is emitted that 871 * is when we want to start waiting for the TLP. This 872 * is most evident with hardware pacing (where the nic 873 * is holding the packet(s) before emitting). But it 874 * can also show up in the network so we do it for all 875 * cases. Technically we would take off one packet from 876 * this extra delay but this is easier and being more 877 * conservative is probably better. 878 */ 879 hpts_timeout += slot; 880 } 881 if (hpts_timeout) { 882 /* 883 * Hack alert for now we can't time-out over 2147 seconds (a 884 * bit more than 35min) 885 */ 886 if (hpts_timeout > 0x7ffffffe) 887 hpts_timeout = 0x7ffffffe; 888 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 889 } else 890 bbr->r_ctl.rc_timer_exp = 0; 891 if ((slot) && 892 (bbr->rc_use_google || 893 bbr->output_error_seen || 894 (slot <= hpts_timeout)) ) { 895 /* 896 * Tell LRO that it can queue packets while 897 * we pace. 898 */ 899 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 900 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 901 (bbr->rc_cwnd_limited == 0)) { 902 /* 903 * If we are not cwnd limited and we 904 * are running a rack timer we put on 905 * the do not disturbe even for sack. 906 */ 907 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 908 } else 909 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 910 bbr->rc_pacer_started = cts; 911 912 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), 913 __LINE__, &diag); 914 bbr->rc_timer_first = 0; 915 bbr->bbr_timer_src = frm; 916 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 917 bbr_log_hpts_diag(bbr, cts, &diag); 918 } else if (hpts_timeout) { 919 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), 920 __LINE__, &diag); 921 /* 922 * We add the flag here as well if the slot is set, 923 * since hpts will call in to clear the queue first before 924 * calling the output routine (which does our timers). 925 * We don't want to set the flag if its just a timer 926 * else the arrival of data might (that causes us 927 * to send more) might get delayed. Imagine being 928 * on a keep-alive timer and a request comes in for 929 * more data. 930 */ 931 if (slot) 932 bbr->rc_pacer_started = cts; 933 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 934 (bbr->rc_cwnd_limited == 0)) { 935 /* 936 * For a rack timer, don't wake us even 937 * if a sack arrives as long as we are 938 * not cwnd limited. 939 */ 940 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 941 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 942 } else { 943 /* All other timers wake us up */ 944 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 945 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 946 } 947 bbr->bbr_timer_src = frm; 948 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 949 bbr_log_hpts_diag(bbr, cts, &diag); 950 bbr->rc_timer_first = 1; 951 } 952 bbr->rc_tmr_stopped = 0; 953 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 954 } 955 956 static void 957 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 958 { 959 /* 960 * We received an ack, and then did not call send or were bounced 961 * out due to the hpts was running. Now a timer is up as well, is it 962 * the right timer? 963 */ 964 struct inpcb *inp; 965 struct bbr_sendmap *rsm; 966 uint32_t hpts_timeout; 967 int tmr_up; 968 969 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 970 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 971 return; 972 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 973 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 974 (tmr_up == PACE_TMR_RXT)) { 975 /* Should be an RXT */ 976 return; 977 } 978 inp = bbr->rc_inp; 979 if (rsm == NULL) { 980 /* Nothing outstanding? */ 981 if (tp->t_flags & TF_DELACK) { 982 if (tmr_up == PACE_TMR_DELACK) 983 /* 984 * We are supposed to have delayed ack up 985 * and we do 986 */ 987 return; 988 } else if (sbavail(&inp->inp_socket->so_snd) && 989 (tmr_up == PACE_TMR_RXT)) { 990 /* 991 * if we hit enobufs then we would expect the 992 * possibility of nothing outstanding and the RXT up 993 * (and the hptsi timer). 994 */ 995 return; 996 } else if (((V_tcp_always_keepalive || 997 inp->inp_socket->so_options & SO_KEEPALIVE) && 998 (tp->t_state <= TCPS_CLOSING)) && 999 (tmr_up == PACE_TMR_KEEP) && 1000 (tp->snd_max == tp->snd_una)) { 1001 /* We should have keep alive up and we do */ 1002 return; 1003 } 1004 } 1005 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 1006 if ((tp->t_flags & TF_SENTFIN) && 1007 ((tp->snd_max - tp->snd_una) == 1) && 1008 (rsm->r_flags & BBR_HAS_FIN)) { 1009 /* needs to be a RXT */ 1010 if (tmr_up == PACE_TMR_RXT) 1011 return; 1012 else 1013 goto wrong_timer; 1014 } else if (tmr_up == PACE_TMR_RACK) 1015 return; 1016 else 1017 goto wrong_timer; 1018 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1019 /* Rack timer has priority if we have data out */ 1020 return; 1021 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1022 ((tmr_up == PACE_TMR_TLP) || 1023 (tmr_up == PACE_TMR_RXT))) { 1024 /* 1025 * Either a TLP or RXT is fine if no sack-passed is in place 1026 * and data is outstanding. 1027 */ 1028 return; 1029 } else if (tmr_up == PACE_TMR_DELACK) { 1030 /* 1031 * If the delayed ack was going to go off before the 1032 * rtx/tlp/rack timer were going to expire, then that would 1033 * be the timer in control. Note we don't check the time 1034 * here trusting the code is correct. 1035 */ 1036 return; 1037 } 1038 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1039 ((tmr_up == PACE_TMR_RXT) || 1040 (tmr_up == PACE_TMR_TLP) || 1041 (tmr_up == PACE_TMR_RACK))) { 1042 /* 1043 * We have outstanding data and 1044 * we *do* have a RACK, TLP or RXT 1045 * timer running. We won't restart 1046 * anything here since thats probably ok we 1047 * will get called with some timer here shortly. 1048 */ 1049 return; 1050 } 1051 /* 1052 * Ok the timer originally started is not what we want now. We will 1053 * force the hpts to be stopped if any, and restart with the slot 1054 * set to what was in the saved slot. 1055 */ 1056 wrong_timer: 1057 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1058 if (tcp_in_hpts(inp)) 1059 tcp_hpts_remove(inp); 1060 bbr_timer_cancel(bbr, __LINE__, cts); 1061 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1062 0); 1063 } else { 1064 /* 1065 * Output is hptsi so we just need to switch the type of 1066 * timer. We don't bother with keep-alive, since when we 1067 * jump through the output, it will start the keep-alive if 1068 * nothing is sent. 1069 * 1070 * We only need a delayed-ack added and or the hpts_timeout. 1071 */ 1072 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1073 if (tp->t_flags & TF_DELACK) { 1074 if (hpts_timeout == 0) { 1075 hpts_timeout = bbr_delack_time; 1076 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1077 } 1078 else if (hpts_timeout > bbr_delack_time) { 1079 hpts_timeout = bbr_delack_time; 1080 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1081 } 1082 } 1083 if (hpts_timeout) { 1084 if (hpts_timeout > 0x7ffffffe) 1085 hpts_timeout = 0x7ffffffe; 1086 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1087 } 1088 } 1089 } 1090 1091 int32_t bbr_clear_lost = 0; 1092 1093 /* 1094 * Considers the two time values now (cts) and earlier. 1095 * If cts is smaller than earlier, we could have 1096 * had a sequence wrap (our counter wraps every 1097 * 70 min or so) or it could be just clock skew 1098 * getting us two different time values. Clock skew 1099 * will show up within 10ms or so. So in such 1100 * a case (where cts is behind earlier time by 1101 * less than 10ms) we return 0. Otherwise we 1102 * return the true difference between them. 1103 */ 1104 static inline uint32_t 1105 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1106 /* 1107 * Given two timestamps, the current time stamp cts, and some other 1108 * time-stamp taken in theory earlier return the difference. The 1109 * trick is here sometimes locking will get the other timestamp 1110 * after the cts. If this occurs we need to return 0. 1111 */ 1112 if (TSTMP_GEQ(cts, earlier_time)) 1113 return (cts - earlier_time); 1114 /* 1115 * cts is behind earlier_time if its less than 10ms consider it 0. 1116 * If its more than 10ms difference then we had a time wrap. Else 1117 * its just the normal locking foo. I wonder if we should not go to 1118 * 64bit TS and get rid of this issue. 1119 */ 1120 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1121 return (0); 1122 /* 1123 * Ok the time must have wrapped. So we need to answer a large 1124 * amount of time, which the normal subtraction should do. 1125 */ 1126 return (cts - earlier_time); 1127 } 1128 1129 static int 1130 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1131 { 1132 uint32_t stat; 1133 int32_t error; 1134 1135 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1136 if (error || req->newptr == NULL) 1137 return error; 1138 1139 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1140 if (error) 1141 return (error); 1142 if (stat == 1) { 1143 #ifdef BBR_INVARIANTS 1144 printf("Clearing BBR lost counters\n"); 1145 #endif 1146 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1147 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1148 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1149 } else if (stat == 2) { 1150 #ifdef BBR_INVARIANTS 1151 printf("Clearing BBR option counters\n"); 1152 #endif 1153 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1154 } else if (stat == 3) { 1155 #ifdef BBR_INVARIANTS 1156 printf("Clearing BBR stats counters\n"); 1157 #endif 1158 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1159 } else if (stat == 4) { 1160 #ifdef BBR_INVARIANTS 1161 printf("Clearing BBR out-size counters\n"); 1162 #endif 1163 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1164 } 1165 bbr_clear_lost = 0; 1166 return (0); 1167 } 1168 1169 static void 1170 bbr_init_sysctls(void) 1171 { 1172 struct sysctl_oid *bbr_probertt; 1173 struct sysctl_oid *bbr_hptsi; 1174 struct sysctl_oid *bbr_measure; 1175 struct sysctl_oid *bbr_cwnd; 1176 struct sysctl_oid *bbr_timeout; 1177 struct sysctl_oid *bbr_states; 1178 struct sysctl_oid *bbr_startup; 1179 struct sysctl_oid *bbr_policer; 1180 1181 /* Probe rtt controls */ 1182 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1183 SYSCTL_CHILDREN(bbr_sysctl_root), 1184 OID_AUTO, 1185 "probertt", 1186 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1187 ""); 1188 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1189 SYSCTL_CHILDREN(bbr_probertt), 1190 OID_AUTO, "gain", CTLFLAG_RW, 1191 &bbr_rttprobe_gain, 192, 1192 "What is the filter gain drop in probe_rtt (0=disable)?"); 1193 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1194 SYSCTL_CHILDREN(bbr_probertt), 1195 OID_AUTO, "cwnd", CTLFLAG_RW, 1196 &bbr_rtt_probe_cwndtarg, 4, 1197 "How many mss's are outstanding during probe-rtt"); 1198 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1199 SYSCTL_CHILDREN(bbr_probertt), 1200 OID_AUTO, "int", CTLFLAG_RW, 1201 &bbr_rtt_probe_limit, 4000000, 1202 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1203 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1204 SYSCTL_CHILDREN(bbr_probertt), 1205 OID_AUTO, "mintime", CTLFLAG_RW, 1206 &bbr_rtt_probe_time, 200000, 1207 "How many microseconds in probe-rtt"); 1208 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1209 SYSCTL_CHILDREN(bbr_probertt), 1210 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1211 &bbr_filter_len_sec, 6, 1212 "How long in seconds does the rttProp filter run?"); 1213 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1214 SYSCTL_CHILDREN(bbr_probertt), 1215 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1216 &bbr_drain_rtt, BBR_SRTT, 1217 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1218 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1219 SYSCTL_CHILDREN(bbr_probertt), 1220 OID_AUTO, "can_force", CTLFLAG_RW, 1221 &bbr_can_force_probertt, 0, 1222 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1223 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1224 SYSCTL_CHILDREN(bbr_probertt), 1225 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1226 &bbr_probertt_sets_rtt, 0, 1227 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1228 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1229 SYSCTL_CHILDREN(bbr_probertt), 1230 OID_AUTO, "can_adjust", CTLFLAG_RW, 1231 &bbr_can_adjust_probertt, 1, 1232 "Can we dynamically adjust the probe-rtt limits and times?"); 1233 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1234 SYSCTL_CHILDREN(bbr_probertt), 1235 OID_AUTO, "is_ratio", CTLFLAG_RW, 1236 &bbr_is_ratio, 0, 1237 "is the limit to filter a ratio?"); 1238 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1239 SYSCTL_CHILDREN(bbr_probertt), 1240 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1241 &bbr_prtt_slam_cwnd, 0, 1242 "Should we set/recover cwnd?"); 1243 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1244 SYSCTL_CHILDREN(bbr_probertt), 1245 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1246 &bbr_can_use_ts_for_rtt, 1, 1247 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1248 1249 /* Pacing controls */ 1250 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1251 SYSCTL_CHILDREN(bbr_sysctl_root), 1252 OID_AUTO, 1253 "pacing", 1254 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1255 ""); 1256 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1257 SYSCTL_CHILDREN(bbr_hptsi), 1258 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1259 &bbr_allow_hdwr_pacing, 1, 1260 "Do we allow hardware pacing?"); 1261 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1262 SYSCTL_CHILDREN(bbr_hptsi), 1263 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1264 &bbr_hardware_pacing_limit, 4000, 1265 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1266 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1267 SYSCTL_CHILDREN(bbr_hptsi), 1268 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1269 &bbr_hdwr_pace_adjust, 2, 1270 "Multiplier to calculated tso size?"); 1271 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1272 SYSCTL_CHILDREN(bbr_hptsi), 1273 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1274 &bbr_hdwr_pace_floor, 1, 1275 "Do we invoke the hardware pacing floor?"); 1276 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1277 SYSCTL_CHILDREN(bbr_hptsi), 1278 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1279 &bbr_hdwr_pacing_delay_cnt, 10, 1280 "How many packets must be sent after hdwr pacing is enabled"); 1281 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1282 SYSCTL_CHILDREN(bbr_hptsi), 1283 OID_AUTO, "bw_cross", CTLFLAG_RW, 1284 &bbr_cross_over, 3000000, 1285 "What is the point where we cross over to linux like TSO size set"); 1286 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1287 SYSCTL_CHILDREN(bbr_hptsi), 1288 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1289 &bbr_hptsi_segments_delay_tar, 7000, 1290 "What is the worse case delay target for hptsi < 48Mbp connections"); 1291 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1292 SYSCTL_CHILDREN(bbr_hptsi), 1293 OID_AUTO, "enet_oh", CTLFLAG_RW, 1294 &bbr_include_enet_oh, 0, 1295 "Do we include the ethernet overhead in calculating pacing delay?"); 1296 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1297 SYSCTL_CHILDREN(bbr_hptsi), 1298 OID_AUTO, "ip_oh", CTLFLAG_RW, 1299 &bbr_include_ip_oh, 1, 1300 "Do we include the IP overhead in calculating pacing delay?"); 1301 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1302 SYSCTL_CHILDREN(bbr_hptsi), 1303 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1304 &bbr_include_tcp_oh, 0, 1305 "Do we include the TCP overhead in calculating pacing delay?"); 1306 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1307 SYSCTL_CHILDREN(bbr_hptsi), 1308 OID_AUTO, "google_discount", CTLFLAG_RW, 1309 &bbr_google_discount, 10, 1310 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1311 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1312 SYSCTL_CHILDREN(bbr_hptsi), 1313 OID_AUTO, "all_get_min", CTLFLAG_RW, 1314 &bbr_all_get_min, 0, 1315 "If you are less than a MSS do you just get the min?"); 1316 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1317 SYSCTL_CHILDREN(bbr_hptsi), 1318 OID_AUTO, "tso_min", CTLFLAG_RW, 1319 &bbr_hptsi_bytes_min, 1460, 1320 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1321 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1322 SYSCTL_CHILDREN(bbr_hptsi), 1323 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1324 &bbr_hptsi_segments_max, 6, 1325 "For 0 -> 24Mbps what is top number of segments for TSO"); 1326 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1327 SYSCTL_CHILDREN(bbr_hptsi), 1328 OID_AUTO, "seg_floor", CTLFLAG_RW, 1329 &bbr_hptsi_segments_floor, 1, 1330 "Minimum TSO size we will fall too in segments"); 1331 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1332 SYSCTL_CHILDREN(bbr_hptsi), 1333 OID_AUTO, "utter_max", CTLFLAG_RW, 1334 &bbr_hptsi_utter_max, 0, 1335 "The absolute maximum that any pacing (outside of hardware) can be"); 1336 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1337 SYSCTL_CHILDREN(bbr_hptsi), 1338 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1339 &bbr_hptsi_per_second, 100, 1340 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1341 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1342 SYSCTL_CHILDREN(bbr_hptsi), 1343 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1344 &bbr_hptsi_max_mul, 1, 1345 "The multiplier for pace len max"); 1346 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1347 SYSCTL_CHILDREN(bbr_hptsi), 1348 OID_AUTO, "srtt_div", CTLFLAG_RW, 1349 &bbr_hptsi_max_div, 2, 1350 "The divisor for pace len max"); 1351 /* Measurement controls */ 1352 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1353 SYSCTL_CHILDREN(bbr_sysctl_root), 1354 OID_AUTO, 1355 "measure", 1356 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1357 "Measurement controls"); 1358 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1359 SYSCTL_CHILDREN(bbr_measure), 1360 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1361 &bbr_initial_bw_bps, 62500, 1362 "Minimum initial b/w in bytes per second"); 1363 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1364 SYSCTL_CHILDREN(bbr_measure), 1365 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1366 &bbr_sack_not_required, 0, 1367 "Do we allow bbr to run on connections not supporting SACK?"); 1368 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1369 SYSCTL_CHILDREN(bbr_measure), 1370 OID_AUTO, "use_google", CTLFLAG_RW, 1371 &bbr_use_google_algo, 0, 1372 "Use has close to google V1.0 has possible?"); 1373 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1374 SYSCTL_CHILDREN(bbr_measure), 1375 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1376 &bbr_ts_limiting, 1, 1377 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1378 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1379 SYSCTL_CHILDREN(bbr_measure), 1380 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1381 &bbr_ts_can_raise, 0, 1382 "Can we raise the b/w via timestamp b/w calculation?"); 1383 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1384 SYSCTL_CHILDREN(bbr_measure), 1385 OID_AUTO, "ts_delta", CTLFLAG_RW, 1386 &bbr_min_usec_delta, 20000, 1387 "How long in usec between ts of our sends in ts validation code?"); 1388 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1389 SYSCTL_CHILDREN(bbr_measure), 1390 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1391 &bbr_min_peer_delta, 20, 1392 "What min numerical value should be between the peer deltas?"); 1393 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1394 SYSCTL_CHILDREN(bbr_measure), 1395 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1396 &bbr_delta_percent, 150, 1397 "What percentage (150 = 15.0) do we allow variance for?"); 1398 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1399 SYSCTL_CHILDREN(bbr_measure), 1400 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1401 &bbr_min_measurements_req, 1, 1402 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1403 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1404 SYSCTL_CHILDREN(bbr_measure), 1405 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1406 &bbr_no_pacing_until, 4, 1407 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1408 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1409 SYSCTL_CHILDREN(bbr_measure), 1410 OID_AUTO, "quanta", CTLFLAG_RW, 1411 &bbr_quanta, 2, 1412 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1413 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1414 SYSCTL_CHILDREN(bbr_measure), 1415 OID_AUTO, "noretran", CTLFLAG_RW, 1416 &bbr_no_retran, 0, 1417 "Should google mode not use retransmission measurements for the b/w estimation?"); 1418 /* State controls */ 1419 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1420 SYSCTL_CHILDREN(bbr_sysctl_root), 1421 OID_AUTO, 1422 "states", 1423 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1424 "State controls"); 1425 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1426 SYSCTL_CHILDREN(bbr_states), 1427 OID_AUTO, "idle_restart", CTLFLAG_RW, 1428 &bbr_uses_idle_restart, 0, 1429 "Do we use a new special idle_restart state to ramp back up quickly?"); 1430 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1431 SYSCTL_CHILDREN(bbr_states), 1432 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1433 &bbr_idle_restart_threshold, 100000, 1434 "How long must we be idle before we restart??"); 1435 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1436 SYSCTL_CHILDREN(bbr_states), 1437 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1438 &bbr_state_is_pkt_epoch, 0, 1439 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1440 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1441 SYSCTL_CHILDREN(bbr_states), 1442 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1443 &bbr_rtt_gain_thresh, 0, 1444 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1445 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1446 SYSCTL_CHILDREN(bbr_states), 1447 OID_AUTO, "drain_floor", CTLFLAG_RW, 1448 &bbr_drain_floor, 88, 1449 "What is the lowest we can drain (pg) too?"); 1450 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1451 SYSCTL_CHILDREN(bbr_states), 1452 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1453 &bbr_state_drain_2_tar, 1, 1454 "Do we drain to target in drain substate?"); 1455 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1456 SYSCTL_CHILDREN(bbr_states), 1457 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1458 &bbr_gain_to_target, 1, 1459 "Does probe bw gain to target??"); 1460 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1461 SYSCTL_CHILDREN(bbr_states), 1462 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1463 &bbr_gain_gets_extra_too, 1, 1464 "Does probe bw gain get the extra time too?"); 1465 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1466 SYSCTL_CHILDREN(bbr_states), 1467 OID_AUTO, "ld_div", CTLFLAG_RW, 1468 &bbr_drain_drop_div, 5, 1469 "Long drain drop divider?"); 1470 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1471 SYSCTL_CHILDREN(bbr_states), 1472 OID_AUTO, "ld_mul", CTLFLAG_RW, 1473 &bbr_drain_drop_mul, 4, 1474 "Long drain drop multiplier?"); 1475 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1476 SYSCTL_CHILDREN(bbr_states), 1477 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1478 &bbr_rand_ot, 50, 1479 "Random discount of the ot?"); 1480 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1481 SYSCTL_CHILDREN(bbr_states), 1482 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1483 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1484 "How many packet-epochs does the b/w delivery rate last?"); 1485 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1486 SYSCTL_CHILDREN(bbr_states), 1487 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1488 &bbr_sub_drain_app_limit, 0, 1489 "Does our sub-state drain invoke app limited if its long?"); 1490 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1491 SYSCTL_CHILDREN(bbr_states), 1492 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1493 &bbr_sub_drain_slam_cwnd, 0, 1494 "Should we set/recover cwnd for sub-state drain?"); 1495 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1496 SYSCTL_CHILDREN(bbr_states), 1497 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1498 &bbr_slam_cwnd_in_main_drain, 0, 1499 "Should we set/recover cwnd for main-state drain?"); 1500 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1501 SYSCTL_CHILDREN(bbr_states), 1502 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1503 &google_allow_early_out, 1, 1504 "Should we allow google probe-bw/drain to exit early at flight target?"); 1505 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1506 SYSCTL_CHILDREN(bbr_states), 1507 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1508 &google_consider_lost, 1, 1509 "Should we have losses exit gain of probebw in google mode??"); 1510 /* Startup controls */ 1511 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1512 SYSCTL_CHILDREN(bbr_sysctl_root), 1513 OID_AUTO, 1514 "startup", 1515 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1516 "Startup controls"); 1517 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1518 SYSCTL_CHILDREN(bbr_startup), 1519 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1520 &bbr_sends_full_iwnd, 1, 1521 "Do we not pace but burst out initial windows has our TSO size?"); 1522 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1523 SYSCTL_CHILDREN(bbr_startup), 1524 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1525 &bbr_startup_loss_thresh, 2000, 1526 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1527 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1528 SYSCTL_CHILDREN(bbr_startup), 1529 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1530 &bbr_use_lower_gain_in_startup, 1, 1531 "Should we use a lower hptsi gain if we see loss in startup?"); 1532 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1533 SYSCTL_CHILDREN(bbr_startup), 1534 OID_AUTO, "gain", CTLFLAG_RW, 1535 &bbr_start_exit, 25, 1536 "What gain percent do we need to see to stay in startup??"); 1537 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1538 SYSCTL_CHILDREN(bbr_startup), 1539 OID_AUTO, "low_gain", CTLFLAG_RW, 1540 &bbr_low_start_exit, 15, 1541 "What gain percent do we need to see to stay in the lower gain startup??"); 1542 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1543 SYSCTL_CHILDREN(bbr_startup), 1544 OID_AUTO, "loss_exit", CTLFLAG_RW, 1545 &bbr_exit_startup_at_loss, 1, 1546 "Should we exit startup at loss in an epoch if we are not gaining?"); 1547 /* CWND controls */ 1548 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1549 SYSCTL_CHILDREN(bbr_sysctl_root), 1550 OID_AUTO, 1551 "cwnd", 1552 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1553 "Cwnd controls"); 1554 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1555 SYSCTL_CHILDREN(bbr_cwnd), 1556 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1557 &bbr_cwndtarget_rtt_touse, 0, 1558 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1559 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1560 SYSCTL_CHILDREN(bbr_cwnd), 1561 OID_AUTO, "may_shrink", CTLFLAG_RW, 1562 &bbr_cwnd_may_shrink, 0, 1563 "Can the cwnd shrink if it would grow to more than the target?"); 1564 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1565 SYSCTL_CHILDREN(bbr_cwnd), 1566 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1567 &bbr_target_cwnd_mult_limit, 8, 1568 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1569 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1570 SYSCTL_CHILDREN(bbr_cwnd), 1571 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1572 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1573 "What is the high-speed min cwnd (rttProp under 1ms)"); 1574 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1575 SYSCTL_CHILDREN(bbr_cwnd), 1576 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1577 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1578 "What is the min cwnd (rttProp > 1ms)"); 1579 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1580 SYSCTL_CHILDREN(bbr_cwnd), 1581 OID_AUTO, "initwin", CTLFLAG_RW, 1582 &bbr_def_init_win, 10, 1583 "What is the BBR initial window, if 0 use tcp version"); 1584 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1585 SYSCTL_CHILDREN(bbr_cwnd), 1586 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1587 &bbr_do_red, 600, 1588 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1589 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1590 SYSCTL_CHILDREN(bbr_cwnd), 1591 OID_AUTO, "red_scale", CTLFLAG_RW, 1592 &bbr_red_scale, 20000, 1593 "What RTT do we scale with?"); 1594 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1595 SYSCTL_CHILDREN(bbr_cwnd), 1596 OID_AUTO, "red_growslow", CTLFLAG_RW, 1597 &bbr_red_growth_restrict, 1, 1598 "Do we restrict cwnd growth for whats in flight?"); 1599 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1600 SYSCTL_CHILDREN(bbr_cwnd), 1601 OID_AUTO, "red_div", CTLFLAG_RW, 1602 &bbr_red_div, 2, 1603 "If we reduce whats the divisor?"); 1604 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1605 SYSCTL_CHILDREN(bbr_cwnd), 1606 OID_AUTO, "red_mul", CTLFLAG_RW, 1607 &bbr_red_mul, 1, 1608 "If we reduce whats the mulitiplier?"); 1609 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1610 SYSCTL_CHILDREN(bbr_cwnd), 1611 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1612 &bbr_target_is_bbunit, 0, 1613 "Is the state target the pacing_gain or BBR_UNIT?"); 1614 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1615 SYSCTL_CHILDREN(bbr_cwnd), 1616 OID_AUTO, "drop_limit", CTLFLAG_RW, 1617 &bbr_drop_limit, 0, 1618 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1619 1620 /* Timeout controls */ 1621 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1622 SYSCTL_CHILDREN(bbr_sysctl_root), 1623 OID_AUTO, 1624 "timeout", 1625 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1626 "Time out controls"); 1627 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1628 SYSCTL_CHILDREN(bbr_timeout), 1629 OID_AUTO, "delack", CTLFLAG_RW, 1630 &bbr_delack_time, 100000, 1631 "BBR's delayed ack time"); 1632 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1633 SYSCTL_CHILDREN(bbr_timeout), 1634 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1635 &bbr_tlp_type_to_use, 3, 1636 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1637 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1638 SYSCTL_CHILDREN(bbr_timeout), 1639 OID_AUTO, "persmin", CTLFLAG_RW, 1640 &bbr_persist_min, 250000, 1641 "What is the minimum time in microseconds between persists"); 1642 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1643 SYSCTL_CHILDREN(bbr_timeout), 1644 OID_AUTO, "persmax", CTLFLAG_RW, 1645 &bbr_persist_max, 1000000, 1646 "What is the largest delay in microseconds between persists"); 1647 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1648 SYSCTL_CHILDREN(bbr_timeout), 1649 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1650 &bbr_tlp_min, 10000, 1651 "TLP Min timeout in usecs"); 1652 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1653 SYSCTL_CHILDREN(bbr_timeout), 1654 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1655 &bbr_delayed_ack_time, 200000, 1656 "TLP delayed ack compensation value"); 1657 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1658 SYSCTL_CHILDREN(bbr_sysctl_root), 1659 OID_AUTO, "minrto", CTLFLAG_RW, 1660 &bbr_rto_min_ms, 30, 1661 "Minimum RTO in ms"); 1662 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1663 SYSCTL_CHILDREN(bbr_timeout), 1664 OID_AUTO, "maxrto", CTLFLAG_RW, 1665 &bbr_rto_max_sec, 4, 1666 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1667 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1668 SYSCTL_CHILDREN(bbr_timeout), 1669 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1670 &bbr_tlp_max_resend, 2, 1671 "How many times does TLP retry a single segment or multiple with no ACK"); 1672 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1673 SYSCTL_CHILDREN(bbr_timeout), 1674 OID_AUTO, "minto", CTLFLAG_RW, 1675 &bbr_min_to, 1000, 1676 "Minimum rack timeout in useconds"); 1677 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1678 SYSCTL_CHILDREN(bbr_timeout), 1679 OID_AUTO, "pktdelay", CTLFLAG_RW, 1680 &bbr_pkt_delay, 1000, 1681 "Extra RACK time (in useconds) besides reordering thresh"); 1682 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1683 SYSCTL_CHILDREN(bbr_timeout), 1684 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1685 &bbr_incr_timers, 1, 1686 "Increase the RXT/TLP timer by the pacing time used?"); 1687 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1688 SYSCTL_CHILDREN(bbr_timeout), 1689 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1690 &bbr_marks_rxt_sack_passed, 0, 1691 "Mark sack passed on all those not ack'd when a RXT hits?"); 1692 /* Policer controls */ 1693 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1694 SYSCTL_CHILDREN(bbr_sysctl_root), 1695 OID_AUTO, 1696 "policer", 1697 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1698 "Policer controls"); 1699 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1700 SYSCTL_CHILDREN(bbr_policer), 1701 OID_AUTO, "detect_enable", CTLFLAG_RW, 1702 &bbr_policer_detection_enabled, 1, 1703 "Is policer detection enabled??"); 1704 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1705 SYSCTL_CHILDREN(bbr_policer), 1706 OID_AUTO, "min_pes", CTLFLAG_RW, 1707 &bbr_lt_intvl_min_rtts, 4, 1708 "Minimum number of PE's?"); 1709 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1710 SYSCTL_CHILDREN(bbr_policer), 1711 OID_AUTO, "bwdiff", CTLFLAG_RW, 1712 &bbr_lt_bw_diff, (4000/8), 1713 "Minimal bw diff?"); 1714 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1715 SYSCTL_CHILDREN(bbr_policer), 1716 OID_AUTO, "bwratio", CTLFLAG_RW, 1717 &bbr_lt_bw_ratio, 8, 1718 "Minimal bw diff?"); 1719 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1720 SYSCTL_CHILDREN(bbr_policer), 1721 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1722 &bbr_policer_call_from_rack_to, 0, 1723 "Do we call the policer detection code from a rack-timeout?"); 1724 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1725 SYSCTL_CHILDREN(bbr_policer), 1726 OID_AUTO, "false_postive", CTLFLAG_RW, 1727 &bbr_lt_intvl_fp, 0, 1728 "What packet epoch do we do false-positive detection at (0=no)?"); 1729 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1730 SYSCTL_CHILDREN(bbr_policer), 1731 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1732 &bbr_lt_loss_thresh, 196, 1733 "Loss threshold 196 = 19.6%?"); 1734 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1735 SYSCTL_CHILDREN(bbr_policer), 1736 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1737 &bbr_lt_fd_thresh, 100, 1738 "What percentage is the false detection threshold (150=15.0)?"); 1739 /* All the rest */ 1740 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1741 SYSCTL_CHILDREN(bbr_sysctl_root), 1742 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1743 &bbr_use_rack_resend_cheat, 0, 1744 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1745 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1746 SYSCTL_CHILDREN(bbr_sysctl_root), 1747 OID_AUTO, "error_paceout", CTLFLAG_RW, 1748 &bbr_error_base_paceout, 10000, 1749 "When we hit an error what is the min to pace out in usec's?"); 1750 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1751 SYSCTL_CHILDREN(bbr_sysctl_root), 1752 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1753 &bbr_max_net_error_cnt, 10, 1754 "When we hit this many errors in a row, kill the session?"); 1755 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1756 SYSCTL_CHILDREN(bbr_sysctl_root), 1757 OID_AUTO, "data_after_close", CTLFLAG_RW, 1758 &bbr_ignore_data_after_close, 1, 1759 "Do we hold off sending a RST until all pending data is ack'd"); 1760 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1761 SYSCTL_CHILDREN(bbr_sysctl_root), 1762 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1763 &bbr_resends_use_tso, 0, 1764 "Can resends use TSO?"); 1765 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1766 SYSCTL_CHILDREN(bbr_sysctl_root), 1767 OID_AUTO, "sblklimit", CTLFLAG_RW, 1768 &bbr_sack_block_limit, 128, 1769 "When do we start ignoring small sack blocks"); 1770 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1771 SYSCTL_CHILDREN(bbr_sysctl_root), 1772 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1773 &bbr_verbose_logging, 0, 1774 "Should BBR black box logging be verbose"); 1775 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1776 SYSCTL_CHILDREN(bbr_sysctl_root), 1777 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1778 &bbr_reorder_thresh, 2, 1779 "What factor for rack will be added when seeing reordering (shift right)"); 1780 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1781 SYSCTL_CHILDREN(bbr_sysctl_root), 1782 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1783 &bbr_reorder_fade, 0, 1784 "Does reorder detection fade, if so how many ms (0 means never)"); 1785 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1786 SYSCTL_CHILDREN(bbr_sysctl_root), 1787 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1788 &bbr_tlp_thresh, 1, 1789 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1790 /* Stats and counters */ 1791 /* The pacing counters for hdwr/software can't be in the array */ 1792 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1793 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1794 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1795 SYSCTL_CHILDREN(bbr_sysctl_root), 1796 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1797 &bbr_hdwr_pacing_enobuf, 1798 "Total number of enobufs for hardware paced flows"); 1799 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1800 SYSCTL_CHILDREN(bbr_sysctl_root), 1801 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1802 &bbr_nohdwr_pacing_enobuf, 1803 "Total number of enobufs for non-hardware paced flows"); 1804 1805 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1806 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1807 SYSCTL_CHILDREN(bbr_sysctl_root), 1808 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1809 &bbr_flows_whdwr_pacing, 1810 "Total number of hardware paced flows"); 1811 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1812 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1813 SYSCTL_CHILDREN(bbr_sysctl_root), 1814 OID_AUTO, "software_pacing", CTLFLAG_RD, 1815 &bbr_flows_nohdwr_pacing, 1816 "Total number of software paced flows"); 1817 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1818 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1819 OID_AUTO, "stats", CTLFLAG_RD, 1820 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1821 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1822 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1823 OID_AUTO, "opts", CTLFLAG_RD, 1824 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1825 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1826 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1827 OID_AUTO, "lost", CTLFLAG_RD, 1828 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1829 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1830 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1831 OID_AUTO, "stateresend", CTLFLAG_RD, 1832 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1833 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1834 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1835 OID_AUTO, "statetime", CTLFLAG_RD, 1836 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1837 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1838 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1839 OID_AUTO, "outsize", CTLFLAG_RD, 1840 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1841 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1842 SYSCTL_CHILDREN(bbr_sysctl_root), 1843 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1844 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1845 } 1846 1847 static void 1848 bbr_counter_destroy(void) 1849 { 1850 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1851 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1852 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1853 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1854 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1855 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1856 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1857 counter_u64_free(bbr_hdwr_pacing_enobuf); 1858 counter_u64_free(bbr_flows_whdwr_pacing); 1859 counter_u64_free(bbr_flows_nohdwr_pacing); 1860 1861 } 1862 1863 static __inline void 1864 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1865 { 1866 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1867 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1868 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1869 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1870 l->bw_inuse = bbr_get_bw(bbr); 1871 l->inflight = ctf_flight_size(bbr->rc_tp, 1872 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1873 l->applimited = bbr->r_ctl.r_app_limited_until; 1874 l->delivered = bbr->r_ctl.rc_delivered; 1875 l->timeStamp = cts; 1876 l->lost = bbr->r_ctl.rc_lost; 1877 l->bbr_state = bbr->rc_bbr_state; 1878 l->bbr_substate = bbr_state_val(bbr); 1879 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1880 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1881 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1882 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1883 l->inhpts = tcp_in_hpts(bbr->rc_inp); 1884 l->use_lt_bw = bbr->rc_lt_use_bw; 1885 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1886 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1887 } 1888 1889 static void 1890 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1891 { 1892 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1893 union tcp_log_stackspecific log; 1894 1895 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1896 log.u_bbr.flex1 = 0; 1897 log.u_bbr.flex2 = 0; 1898 log.u_bbr.flex5 = 0; 1899 log.u_bbr.flex3 = 0; 1900 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1901 log.u_bbr.flex7 = reason; 1902 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1903 log.u_bbr.flex8 = 0; 1904 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1905 &bbr->rc_inp->inp_socket->so_rcv, 1906 &bbr->rc_inp->inp_socket->so_snd, 1907 BBR_LOG_BW_RED_EV, 0, 1908 0, &log, false, &bbr->rc_tv); 1909 } 1910 } 1911 1912 static void 1913 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1914 { 1915 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1916 union tcp_log_stackspecific log; 1917 1918 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1919 log.u_bbr.flex1 = seq; 1920 log.u_bbr.flex2 = count; 1921 log.u_bbr.flex8 = mode; 1922 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1923 &bbr->rc_inp->inp_socket->so_rcv, 1924 &bbr->rc_inp->inp_socket->so_snd, 1925 BBR_LOG_LOWGAIN, 0, 1926 0, &log, false, &bbr->rc_tv); 1927 } 1928 } 1929 1930 static void 1931 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1932 uint8_t reason, uint32_t p_maxseg, int len) 1933 { 1934 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1935 union tcp_log_stackspecific log; 1936 1937 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1938 log.u_bbr.flex1 = p_maxseg; 1939 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1940 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1941 log.u_bbr.flex4 = reason; 1942 log.u_bbr.flex5 = bbr->rc_in_persist; 1943 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1944 log.u_bbr.flex7 = p_maxseg; 1945 log.u_bbr.flex8 = bbr->rc_in_persist; 1946 log.u_bbr.pkts_out = 0; 1947 log.u_bbr.applimited = len; 1948 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1949 &bbr->rc_inp->inp_socket->so_rcv, 1950 &bbr->rc_inp->inp_socket->so_snd, 1951 BBR_LOG_JUSTRET, 0, 1952 tlen, &log, false, &bbr->rc_tv); 1953 } 1954 } 1955 1956 static void 1957 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1958 { 1959 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1960 union tcp_log_stackspecific log; 1961 1962 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1963 log.u_bbr.flex1 = seq; 1964 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1965 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1966 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1967 &bbr->rc_inp->inp_socket->so_rcv, 1968 &bbr->rc_inp->inp_socket->so_snd, 1969 BBR_LOG_ENTREC, 0, 1970 0, &log, false, &bbr->rc_tv); 1971 } 1972 } 1973 1974 static void 1975 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) 1976 { 1977 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 1978 union tcp_log_stackspecific log; 1979 1980 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1981 log.u_bbr.flex1 = tso; 1982 log.u_bbr.flex2 = maxseg; 1983 log.u_bbr.flex3 = mtu; 1984 log.u_bbr.flex4 = csum_flags; 1985 TCP_LOG_EVENTP(tp, NULL, 1986 &bbr->rc_inp->inp_socket->so_rcv, 1987 &bbr->rc_inp->inp_socket->so_snd, 1988 BBR_LOG_MSGSIZE, 0, 1989 0, &log, false, &bbr->rc_tv); 1990 } 1991 } 1992 1993 static void 1994 bbr_log_flowend(struct tcp_bbr *bbr) 1995 { 1996 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1997 union tcp_log_stackspecific log; 1998 struct sockbuf *r, *s; 1999 struct timeval tv; 2000 2001 if (bbr->rc_inp->inp_socket) { 2002 r = &bbr->rc_inp->inp_socket->so_rcv; 2003 s = &bbr->rc_inp->inp_socket->so_snd; 2004 } else { 2005 r = s = NULL; 2006 } 2007 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2008 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2009 r, s, 2010 TCP_LOG_FLOWEND, 0, 2011 0, &log, false, &tv); 2012 } 2013 } 2014 2015 static void 2016 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2017 uint32_t lost, uint32_t del) 2018 { 2019 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2020 union tcp_log_stackspecific log; 2021 2022 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2023 log.u_bbr.flex1 = lost; 2024 log.u_bbr.flex2 = del; 2025 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2026 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2027 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2028 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2029 log.u_bbr.flex7 = line; 2030 log.u_bbr.flex8 = 0; 2031 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2032 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2033 &bbr->rc_inp->inp_socket->so_rcv, 2034 &bbr->rc_inp->inp_socket->so_snd, 2035 BBR_LOG_PKT_EPOCH, 0, 2036 0, &log, false, &bbr->rc_tv); 2037 } 2038 } 2039 2040 static void 2041 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2042 { 2043 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2044 union tcp_log_stackspecific log; 2045 2046 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2047 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2048 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2049 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2050 log.u_bbr.flex7 = line; 2051 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2052 &bbr->rc_inp->inp_socket->so_rcv, 2053 &bbr->rc_inp->inp_socket->so_snd, 2054 BBR_LOG_TIME_EPOCH, 0, 2055 0, &log, false, &bbr->rc_tv); 2056 } 2057 } 2058 2059 static void 2060 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2061 { 2062 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2063 union tcp_log_stackspecific log; 2064 2065 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2066 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2067 log.u_bbr.flex2 = new_tar; 2068 log.u_bbr.flex3 = line; 2069 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2070 log.u_bbr.flex5 = bbr_quanta; 2071 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2072 log.u_bbr.flex7 = bbr->rc_last_options; 2073 log.u_bbr.flex8 = meth; 2074 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2075 &bbr->rc_inp->inp_socket->so_rcv, 2076 &bbr->rc_inp->inp_socket->so_snd, 2077 BBR_LOG_STATE_TARGET, 0, 2078 0, &log, false, &bbr->rc_tv); 2079 } 2080 2081 } 2082 2083 static void 2084 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2085 { 2086 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2087 union tcp_log_stackspecific log; 2088 2089 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2090 log.u_bbr.flex1 = line; 2091 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2092 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2093 if (bbr_state_is_pkt_epoch) 2094 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2095 else 2096 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2097 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2098 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2099 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2100 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2101 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2102 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2103 &bbr->rc_inp->inp_socket->so_rcv, 2104 &bbr->rc_inp->inp_socket->so_snd, 2105 BBR_LOG_STATE, 0, 2106 0, &log, false, &bbr->rc_tv); 2107 } 2108 } 2109 2110 static void 2111 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2112 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2113 { 2114 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2115 union tcp_log_stackspecific log; 2116 2117 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2118 log.u_bbr.flex1 = line; 2119 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2120 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2121 log.u_bbr.flex4 = applied; 2122 log.u_bbr.flex5 = rtt; 2123 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2124 log.u_bbr.flex7 = cond; 2125 log.u_bbr.flex8 = reas; 2126 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2127 &bbr->rc_inp->inp_socket->so_rcv, 2128 &bbr->rc_inp->inp_socket->so_snd, 2129 BBR_LOG_RTT_SHRINKS, 0, 2130 0, &log, false, &bbr->rc_tv); 2131 } 2132 } 2133 2134 static void 2135 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2136 { 2137 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2138 union tcp_log_stackspecific log; 2139 2140 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2141 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2142 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2143 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2144 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2145 &bbr->rc_inp->inp_socket->so_rcv, 2146 &bbr->rc_inp->inp_socket->so_snd, 2147 BBR_LOG_EXITREC, 0, 2148 0, &log, false, &bbr->rc_tv); 2149 } 2150 } 2151 2152 static void 2153 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2154 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2155 { 2156 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2157 union tcp_log_stackspecific log; 2158 2159 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2160 log.u_bbr.flex1 = line; 2161 log.u_bbr.flex2 = prev_acked; 2162 log.u_bbr.flex3 = bytes_this_ack; 2163 log.u_bbr.flex4 = chg; 2164 log.u_bbr.flex5 = th_ack; 2165 log.u_bbr.flex6 = target; 2166 log.u_bbr.flex8 = meth; 2167 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2168 &bbr->rc_inp->inp_socket->so_rcv, 2169 &bbr->rc_inp->inp_socket->so_snd, 2170 BBR_LOG_CWND, 0, 2171 0, &log, false, &bbr->rc_tv); 2172 } 2173 } 2174 2175 static void 2176 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2177 { 2178 /* 2179 * Log the rtt sample we are applying to the srtt algorithm in 2180 * useconds. 2181 */ 2182 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2183 union tcp_log_stackspecific log; 2184 2185 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2186 log.u_bbr.flex1 = rtt; 2187 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2188 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2189 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2190 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2191 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2192 log.u_bbr.flex6 = tsin; 2193 log.u_bbr.flex7 = 0; 2194 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2195 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2196 &bbr->rc_inp->inp_socket->so_rcv, 2197 &bbr->rc_inp->inp_socket->so_snd, 2198 TCP_LOG_RTT, 0, 2199 0, &log, false, &bbr->rc_tv); 2200 } 2201 } 2202 2203 static void 2204 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2205 { 2206 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2207 union tcp_log_stackspecific log; 2208 2209 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2210 log.u_bbr.flex1 = time_in; 2211 log.u_bbr.flex2 = line; 2212 log.u_bbr.flex8 = enter_exit; 2213 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2214 &bbr->rc_inp->inp_socket->so_rcv, 2215 &bbr->rc_inp->inp_socket->so_snd, 2216 BBR_LOG_PERSIST, 0, 2217 0, &log, false, &bbr->rc_tv); 2218 } 2219 } 2220 static void 2221 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2222 { 2223 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2224 union tcp_log_stackspecific log; 2225 2226 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2227 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2228 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2229 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2230 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2231 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2232 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2233 &bbr->rc_inp->inp_socket->so_rcv, 2234 &bbr->rc_inp->inp_socket->so_snd, 2235 BBR_LOG_ACKCLEAR, 0, 2236 0, &log, false, &bbr->rc_tv); 2237 } 2238 } 2239 2240 static void 2241 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2242 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2243 { 2244 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2245 union tcp_log_stackspecific log; 2246 struct timeval tv; 2247 2248 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2249 log.u_bbr.flex1 = nsegs; 2250 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2251 if (m) { 2252 struct timespec ts; 2253 2254 log.u_bbr.flex3 = m->m_flags; 2255 if (m->m_flags & M_TSTMP) { 2256 mbuf_tstmp2timespec(m, &ts); 2257 tv.tv_sec = ts.tv_sec; 2258 tv.tv_usec = ts.tv_nsec / 1000; 2259 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2260 } else { 2261 log.u_bbr.lt_epoch = 0; 2262 } 2263 if (m->m_flags & M_TSTMP_LRO) { 2264 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; 2265 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000; 2266 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2267 } else { 2268 /* No arrival timestamp */ 2269 log.u_bbr.flex5 = 0; 2270 } 2271 2272 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2273 } else { 2274 log.u_bbr.flex3 = 0; 2275 log.u_bbr.flex5 = 0; 2276 log.u_bbr.flex6 = 0; 2277 log.u_bbr.pkts_out = 0; 2278 } 2279 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2280 log.u_bbr.flex7 = bbr->r_wanted_output; 2281 log.u_bbr.flex8 = bbr->rc_in_persist; 2282 TCP_LOG_EVENTP(bbr->rc_tp, th, 2283 &bbr->rc_inp->inp_socket->so_rcv, 2284 &bbr->rc_inp->inp_socket->so_snd, 2285 TCP_LOG_IN, 0, 2286 tlen, &log, true, &bbr->rc_tv); 2287 } 2288 } 2289 2290 static void 2291 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2292 { 2293 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2294 union tcp_log_stackspecific log; 2295 2296 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2297 log.u_bbr.flex1 = did_out; 2298 log.u_bbr.flex2 = nxt_pkt; 2299 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2300 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2301 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2302 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2303 log.u_bbr.flex7 = bbr->r_wanted_output; 2304 log.u_bbr.flex8 = bbr->rc_in_persist; 2305 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2306 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2307 &bbr->rc_inp->inp_socket->so_rcv, 2308 &bbr->rc_inp->inp_socket->so_snd, 2309 BBR_LOG_DOSEG_DONE, 0, 2310 0, &log, true, &bbr->rc_tv); 2311 } 2312 } 2313 2314 static void 2315 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2316 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2317 { 2318 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2319 union tcp_log_stackspecific log; 2320 2321 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2322 log.u_bbr.flex1 = line; 2323 log.u_bbr.flex2 = o_len; 2324 log.u_bbr.flex3 = segcnt; 2325 log.u_bbr.flex4 = segsiz; 2326 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2327 &bbr->rc_inp->inp_socket->so_rcv, 2328 &bbr->rc_inp->inp_socket->so_snd, 2329 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2330 len, &log, true, &bbr->rc_tv); 2331 } 2332 } 2333 2334 static void 2335 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2336 { 2337 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2338 union tcp_log_stackspecific log; 2339 2340 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2341 log.u_bbr.flex1 = timers; 2342 log.u_bbr.flex2 = ret; 2343 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2344 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2345 log.u_bbr.flex5 = cts; 2346 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2347 log.u_bbr.flex8 = hpts_calling; 2348 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2349 &bbr->rc_inp->inp_socket->so_rcv, 2350 &bbr->rc_inp->inp_socket->so_snd, 2351 BBR_LOG_TO_PROCESS, 0, 2352 0, &log, false, &bbr->rc_tv); 2353 } 2354 } 2355 2356 static void 2357 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2358 { 2359 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2360 union tcp_log_stackspecific log; 2361 uint64_t ar; 2362 2363 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2364 log.u_bbr.flex1 = bbr->bbr_timer_src; 2365 log.u_bbr.flex2 = 0; 2366 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2367 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2368 ar >>= 32; 2369 ar &= 0x00000000ffffffff; 2370 log.u_bbr.flex4 = (uint32_t)ar; 2371 ar = (uint64_t)bbr->r_ctl.rc_resend; 2372 ar &= 0x00000000ffffffff; 2373 log.u_bbr.flex5 = (uint32_t)ar; 2374 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2375 log.u_bbr.flex8 = to_num; 2376 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2377 &bbr->rc_inp->inp_socket->so_rcv, 2378 &bbr->rc_inp->inp_socket->so_snd, 2379 BBR_LOG_RTO, 0, 2380 0, &log, false, &bbr->rc_tv); 2381 } 2382 } 2383 2384 static void 2385 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2386 { 2387 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2388 union tcp_log_stackspecific log; 2389 2390 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2391 log.u_bbr.flex1 = flex1; 2392 log.u_bbr.flex2 = flex2; 2393 log.u_bbr.flex3 = flex3; 2394 log.u_bbr.flex4 = 0; 2395 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2396 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2397 log.u_bbr.flex8 = reason; 2398 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2399 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2400 &bbr->rc_inp->inp_socket->so_rcv, 2401 &bbr->rc_inp->inp_socket->so_snd, 2402 BBR_LOG_REDUCE, 0, 2403 0, &log, false, &bbr->rc_tv); 2404 } 2405 } 2406 2407 static void 2408 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2409 { 2410 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2411 union tcp_log_stackspecific log; 2412 2413 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2414 log.u_bbr.flex1 = diag->p_nxt_slot; 2415 log.u_bbr.flex2 = diag->p_cur_slot; 2416 log.u_bbr.flex3 = diag->slot_req; 2417 log.u_bbr.flex4 = diag->inp_hptsslot; 2418 log.u_bbr.flex5 = diag->slot_remaining; 2419 log.u_bbr.flex6 = diag->need_new_to; 2420 log.u_bbr.flex7 = diag->p_hpts_active; 2421 log.u_bbr.flex8 = diag->p_on_min_sleep; 2422 /* Hijack other fields as needed */ 2423 log.u_bbr.epoch = diag->have_slept; 2424 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2425 log.u_bbr.pkts_out = diag->co_ret; 2426 log.u_bbr.applimited = diag->hpts_sleep_time; 2427 log.u_bbr.delivered = diag->p_prev_slot; 2428 log.u_bbr.inflight = diag->p_runningslot; 2429 log.u_bbr.bw_inuse = diag->wheel_slot; 2430 log.u_bbr.rttProp = diag->wheel_cts; 2431 log.u_bbr.delRate = diag->maxslots; 2432 log.u_bbr.cur_del_rate = diag->p_curtick; 2433 log.u_bbr.cur_del_rate <<= 32; 2434 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2435 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2436 &bbr->rc_inp->inp_socket->so_rcv, 2437 &bbr->rc_inp->inp_socket->so_snd, 2438 BBR_LOG_HPTSDIAG, 0, 2439 0, &log, false, &bbr->rc_tv); 2440 } 2441 } 2442 2443 static void 2444 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2445 uint32_t thresh, uint32_t to) 2446 { 2447 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2448 union tcp_log_stackspecific log; 2449 2450 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2451 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2452 log.u_bbr.flex2 = time_since_sent; 2453 log.u_bbr.flex3 = srtt; 2454 log.u_bbr.flex4 = thresh; 2455 log.u_bbr.flex5 = to; 2456 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2457 log.u_bbr.flex8 = mode; 2458 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2459 &bbr->rc_inp->inp_socket->so_rcv, 2460 &bbr->rc_inp->inp_socket->so_snd, 2461 BBR_LOG_TIMERPREP, 0, 2462 0, &log, false, &bbr->rc_tv); 2463 } 2464 } 2465 2466 static void 2467 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2468 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2469 { 2470 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2471 union tcp_log_stackspecific log; 2472 2473 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2474 log.u_bbr.flex1 = usecs; 2475 log.u_bbr.flex2 = len; 2476 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2477 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2478 if (override) 2479 log.u_bbr.flex5 = (1 << 2); 2480 else 2481 log.u_bbr.flex5 = 0; 2482 log.u_bbr.flex6 = override; 2483 log.u_bbr.flex7 = gain; 2484 log.u_bbr.flex8 = mod; 2485 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2486 &bbr->rc_inp->inp_socket->so_rcv, 2487 &bbr->rc_inp->inp_socket->so_snd, 2488 BBR_LOG_HPTSI_CALC, 0, 2489 len, &log, false, &bbr->rc_tv); 2490 } 2491 } 2492 2493 static void 2494 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2495 { 2496 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2497 union tcp_log_stackspecific log; 2498 2499 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2500 2501 log.u_bbr.flex1 = bbr->bbr_timer_src; 2502 log.u_bbr.flex2 = to; 2503 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2504 log.u_bbr.flex4 = slot; 2505 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; 2506 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2507 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; 2508 log.u_bbr.flex8 = which; 2509 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2510 &bbr->rc_inp->inp_socket->so_rcv, 2511 &bbr->rc_inp->inp_socket->so_snd, 2512 BBR_LOG_TIMERSTAR, 0, 2513 0, &log, false, &bbr->rc_tv); 2514 } 2515 } 2516 2517 static void 2518 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) 2519 { 2520 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2521 union tcp_log_stackspecific log; 2522 2523 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2524 log.u_bbr.flex1 = thresh; 2525 log.u_bbr.flex2 = lro; 2526 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2527 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2528 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2529 log.u_bbr.flex6 = srtt; 2530 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2531 log.u_bbr.flex8 = frm; 2532 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2533 &bbr->rc_inp->inp_socket->so_rcv, 2534 &bbr->rc_inp->inp_socket->so_snd, 2535 BBR_LOG_THRESH_CALC, 0, 2536 0, &log, false, &bbr->rc_tv); 2537 } 2538 } 2539 2540 static void 2541 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2542 { 2543 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2544 union tcp_log_stackspecific log; 2545 2546 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2547 log.u_bbr.flex1 = line; 2548 log.u_bbr.flex2 = bbr->bbr_timer_src; 2549 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2550 log.u_bbr.flex4 = bbr->rc_in_persist; 2551 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2552 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2553 log.u_bbr.flex8 = hpts_removed; 2554 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2555 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2556 &bbr->rc_inp->inp_socket->so_rcv, 2557 &bbr->rc_inp->inp_socket->so_snd, 2558 BBR_LOG_TIMERCANC, 0, 2559 0, &log, false, &bbr->rc_tv); 2560 } 2561 } 2562 2563 static void 2564 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2565 { 2566 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2567 union tcp_log_stackspecific log; 2568 2569 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2570 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2571 log.u_bbr.flex2 = (peer_delta >> 32); 2572 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2573 log.u_bbr.flex4 = (delta >> 32); 2574 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2575 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2576 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2577 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2578 &bbr->rc_inp->inp_socket->so_rcv, 2579 &bbr->rc_inp->inp_socket->so_snd, 2580 BBR_LOG_TSTMP_VAL, 0, 2581 0, &log, false, &bbr->rc_tv); 2582 } 2583 } 2584 2585 static void 2586 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) 2587 { 2588 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2589 union tcp_log_stackspecific log; 2590 2591 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2592 log.u_bbr.flex1 = tsosz; 2593 log.u_bbr.flex2 = tls; 2594 log.u_bbr.flex3 = tcp_min_hptsi_time; 2595 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2596 log.u_bbr.flex5 = old_val; 2597 log.u_bbr.flex6 = maxseg; 2598 log.u_bbr.flex7 = bbr->rc_no_pacing; 2599 log.u_bbr.flex7 <<= 1; 2600 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2601 if (hdwr) 2602 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2603 else 2604 log.u_bbr.flex8 = bbr->rc_use_google; 2605 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2606 &bbr->rc_inp->inp_socket->so_rcv, 2607 &bbr->rc_inp->inp_socket->so_snd, 2608 BBR_LOG_BBRTSO, 0, 2609 0, &log, false, &bbr->rc_tv); 2610 } 2611 } 2612 2613 static void 2614 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2615 uint32_t flags, uint32_t line) 2616 { 2617 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2618 union tcp_log_stackspecific log; 2619 2620 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2621 log.u_bbr.flex1 = line; 2622 log.u_bbr.flex2 = rsm->r_start; 2623 log.u_bbr.flex3 = rsm->r_end; 2624 log.u_bbr.flex4 = rsm->r_delivered; 2625 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2626 log.u_bbr.flex6 = rsm->r_dupack; 2627 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2628 log.u_bbr.flex8 = rsm->r_flags; 2629 /* Hijack the pkts_out fids */ 2630 log.u_bbr.applimited = flags; 2631 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2632 &bbr->rc_inp->inp_socket->so_rcv, 2633 &bbr->rc_inp->inp_socket->so_snd, 2634 BBR_RSM_CLEARED, 0, 2635 0, &log, false, &bbr->rc_tv); 2636 } 2637 } 2638 2639 static void 2640 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2641 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2642 uint32_t flex6, uint32_t pkts_out, int flex7, 2643 uint32_t flex4, uint32_t flex1) 2644 { 2645 2646 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2647 union tcp_log_stackspecific log; 2648 2649 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2650 log.u_bbr.flex1 = flex1; 2651 log.u_bbr.flex2 = flex2; 2652 log.u_bbr.flex3 = flex3; 2653 log.u_bbr.flex4 = flex4; 2654 log.u_bbr.flex5 = flex5; 2655 log.u_bbr.flex6 = flex6; 2656 log.u_bbr.flex7 = flex7; 2657 /* Hijack the pkts_out fids */ 2658 log.u_bbr.pkts_out = pkts_out; 2659 log.u_bbr.flex8 = flex8; 2660 if (bbr->rc_ack_was_delayed) 2661 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2662 else 2663 log.u_bbr.epoch = 0; 2664 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2665 &bbr->rc_inp->inp_socket->so_rcv, 2666 &bbr->rc_inp->inp_socket->so_snd, 2667 BBR_LOG_BBRUPD, 0, 2668 flex2, &log, false, &bbr->rc_tv); 2669 } 2670 } 2671 2672 static void 2673 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2674 uint32_t newbw, uint32_t obw, uint32_t diff, 2675 uint32_t tim) 2676 { 2677 if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2678 union tcp_log_stackspecific log; 2679 2680 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2681 log.u_bbr.flex1 = reason; 2682 log.u_bbr.flex2 = newbw; 2683 log.u_bbr.flex3 = obw; 2684 log.u_bbr.flex4 = diff; 2685 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2686 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2687 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2688 log.u_bbr.pkts_out = tim; 2689 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2690 if (bbr->rc_lt_use_bw == 0) 2691 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2692 else 2693 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2694 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2695 &bbr->rc_inp->inp_socket->so_rcv, 2696 &bbr->rc_inp->inp_socket->so_snd, 2697 BBR_LOG_BWSAMP, 0, 2698 0, &log, false, &bbr->rc_tv); 2699 } 2700 } 2701 2702 static inline void 2703 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2704 { 2705 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2706 union tcp_log_stackspecific log; 2707 2708 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2709 log.u_bbr.flex1 = line; 2710 log.u_bbr.flex2 = tick; 2711 log.u_bbr.flex3 = tp->t_maxunacktime; 2712 log.u_bbr.flex4 = tp->t_acktime; 2713 log.u_bbr.flex8 = event; 2714 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2715 &bbr->rc_inp->inp_socket->so_rcv, 2716 &bbr->rc_inp->inp_socket->so_snd, 2717 BBR_LOG_PROGRESS, 0, 2718 0, &log, false, &bbr->rc_tv); 2719 } 2720 } 2721 2722 static void 2723 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2724 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2725 int error) 2726 { 2727 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2728 union tcp_log_stackspecific log; 2729 2730 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2731 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2732 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2733 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2734 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2735 log.u_bbr.bw_inuse = rate; 2736 log.u_bbr.flex5 = line; 2737 log.u_bbr.flex6 = error; 2738 log.u_bbr.flex8 = bbr->skip_gain; 2739 log.u_bbr.flex8 <<= 1; 2740 log.u_bbr.flex8 |= bbr->gain_is_limited; 2741 log.u_bbr.flex8 <<= 1; 2742 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2743 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2744 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2745 &bbr->rc_inp->inp_socket->so_rcv, 2746 &bbr->rc_inp->inp_socket->so_snd, 2747 BBR_LOG_HDWR_PACE, 0, 2748 0, &log, false, &bbr->rc_tv); 2749 } 2750 } 2751 2752 static void 2753 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) 2754 { 2755 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2756 union tcp_log_stackspecific log; 2757 2758 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2759 log.u_bbr.flex1 = slot; 2760 log.u_bbr.flex2 = del_by; 2761 log.u_bbr.flex3 = prev_delay; 2762 log.u_bbr.flex4 = line; 2763 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2764 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2765 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2766 log.u_bbr.flex8 = bbr->rc_in_persist; 2767 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2768 &bbr->rc_inp->inp_socket->so_rcv, 2769 &bbr->rc_inp->inp_socket->so_snd, 2770 BBR_LOG_BBRSND, 0, 2771 len, &log, false, &bbr->rc_tv); 2772 } 2773 } 2774 2775 static void 2776 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) 2777 { 2778 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2779 union tcp_log_stackspecific log; 2780 2781 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2782 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2783 log.u_bbr.flex2 = 0; 2784 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2785 log.u_bbr.flex4 = end; 2786 log.u_bbr.flex5 = seq; 2787 log.u_bbr.flex6 = t; 2788 log.u_bbr.flex7 = match; 2789 log.u_bbr.flex8 = flags; 2790 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2791 &bbr->rc_inp->inp_socket->so_rcv, 2792 &bbr->rc_inp->inp_socket->so_snd, 2793 BBR_LOG_BBRRTT, 0, 2794 0, &log, false, &bbr->rc_tv); 2795 } 2796 } 2797 2798 static void 2799 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2800 { 2801 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2802 union tcp_log_stackspecific log; 2803 2804 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2805 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2806 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2807 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2808 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2809 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2810 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2811 log.u_bbr.flex7 = 0; 2812 log.u_bbr.flex8 = entry_method; 2813 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2814 &bbr->rc_inp->inp_socket->so_rcv, 2815 &bbr->rc_inp->inp_socket->so_snd, 2816 BBR_LOG_EXIT_GAIN, 0, 2817 0, &log, false, &bbr->rc_tv); 2818 } 2819 } 2820 2821 static void 2822 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2823 { 2824 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2825 union tcp_log_stackspecific log; 2826 2827 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2828 /* R-HU */ 2829 log.u_bbr.flex1 = 0; 2830 log.u_bbr.flex2 = 0; 2831 log.u_bbr.flex3 = 0; 2832 log.u_bbr.flex4 = 0; 2833 log.u_bbr.flex7 = 0; 2834 log.u_bbr.flex8 = settings_desired; 2835 2836 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2837 &bbr->rc_inp->inp_socket->so_rcv, 2838 &bbr->rc_inp->inp_socket->so_snd, 2839 BBR_LOG_SETTINGS_CHG, 0, 2840 0, &log, false, &bbr->rc_tv); 2841 } 2842 } 2843 2844 /* 2845 * Returns the bw from the our filter. 2846 */ 2847 static inline uint64_t 2848 bbr_get_full_bw(struct tcp_bbr *bbr) 2849 { 2850 uint64_t bw; 2851 2852 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2853 2854 return (bw); 2855 } 2856 2857 static inline void 2858 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2859 { 2860 uint64_t calclr; 2861 uint32_t lost, del; 2862 2863 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2864 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2865 else 2866 lost = 0; 2867 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2868 if (lost == 0) { 2869 calclr = 0; 2870 } else if (del) { 2871 calclr = lost; 2872 calclr *= (uint64_t)1000; 2873 calclr /= (uint64_t)del; 2874 } else { 2875 /* Nothing delivered? 100.0% loss */ 2876 calclr = 1000; 2877 } 2878 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2879 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2880 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2881 bbr->r_ctl.rc_pkt_epoch++; 2882 if (bbr->rc_no_pacing && 2883 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2884 bbr->rc_no_pacing = 0; 2885 tcp_bbr_tso_size_check(bbr, cts); 2886 } 2887 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2888 bbr->r_ctl.rc_pkt_epoch_time = cts; 2889 /* What was our loss rate */ 2890 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2891 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2892 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2893 } 2894 2895 static inline void 2896 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2897 { 2898 uint32_t epoch_time; 2899 2900 /* Tick the RTT clock */ 2901 bbr->r_ctl.rc_rtt_epoch++; 2902 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2903 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2904 bbr->r_ctl.rc_rcv_epoch_start = cts; 2905 } 2906 2907 static inline void 2908 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2909 { 2910 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2911 bbr->rc_is_pkt_epoch_now = 1; 2912 } 2913 } 2914 2915 /* 2916 * Returns the bw from either the b/w filter 2917 * or from the lt_bw (if the connection is being 2918 * policed). 2919 */ 2920 static inline uint64_t 2921 __bbr_get_bw(struct tcp_bbr *bbr) 2922 { 2923 uint64_t bw, min_bw; 2924 uint64_t rtt; 2925 int gm_measure_cnt = 1; 2926 2927 /* 2928 * For startup we make, like google, a 2929 * minimum b/w. This is generated from the 2930 * IW and the rttProp. We do fall back to srtt 2931 * if for some reason (initial handshake) we don't 2932 * have a rttProp. We, in the worst case, fall back 2933 * to the configured min_bw (rc_initial_hptsi_bw). 2934 */ 2935 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2936 /* Attempt first to use rttProp */ 2937 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2938 if (rtt && (rtt < 0xffffffff)) { 2939 measure: 2940 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2941 ((uint64_t)1000000); 2942 min_bw /= rtt; 2943 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2944 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2945 } 2946 2947 } else if (bbr->rc_tp->t_srtt != 0) { 2948 /* No rttProp, use srtt? */ 2949 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2950 goto measure; 2951 } else { 2952 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2953 } 2954 } else 2955 min_bw = 0; 2956 2957 if ((bbr->rc_past_init_win == 0) && 2958 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2959 bbr->rc_past_init_win = 1; 2960 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2961 gm_measure_cnt = 0; 2962 if (gm_measure_cnt && 2963 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2964 (bbr->rc_past_init_win == 0))) { 2965 /* For google we use our guess rate until we get 1 measurement */ 2966 2967 use_initial_window: 2968 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2969 if (rtt && (rtt < 0xffffffff)) { 2970 /* 2971 * We have an RTT measurement. Use that in 2972 * combination with our initial window to calculate 2973 * a b/w. 2974 */ 2975 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2976 ((uint64_t)1000000); 2977 bw /= rtt; 2978 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2979 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2980 } 2981 } else { 2982 /* Drop back to the 40 and punt to a default */ 2983 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2984 } 2985 if (bw < 1) 2986 /* Probably should panic */ 2987 bw = 1; 2988 if (bw > min_bw) 2989 return (bw); 2990 else 2991 return (min_bw); 2992 } 2993 if (bbr->rc_lt_use_bw) 2994 bw = bbr->r_ctl.rc_lt_bw; 2995 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2996 bw = bbr->r_ctl.red_bw; 2997 else 2998 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2999 if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { 3000 /* 3001 * Enforce user set rate limit, keep in mind that 3002 * t_peakrate_thr is in B/s already 3003 */ 3004 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); 3005 } 3006 if (bw == 0) { 3007 /* We should not be at 0, go to the initial window then */ 3008 goto use_initial_window; 3009 } 3010 if (bw < 1) 3011 /* Probably should panic */ 3012 bw = 1; 3013 if (bw < min_bw) 3014 bw = min_bw; 3015 return (bw); 3016 } 3017 3018 static inline uint64_t 3019 bbr_get_bw(struct tcp_bbr *bbr) 3020 { 3021 uint64_t bw; 3022 3023 bw = __bbr_get_bw(bbr); 3024 return (bw); 3025 } 3026 3027 static inline void 3028 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3029 { 3030 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3031 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3032 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3033 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3034 } 3035 3036 static inline void 3037 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3038 { 3039 bbr->rc_lt_is_sampling = 0; 3040 bbr->rc_lt_use_bw = 0; 3041 bbr->r_ctl.rc_lt_bw = 0; 3042 bbr_reset_lt_bw_interval(bbr, cts); 3043 } 3044 3045 static inline void 3046 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3047 { 3048 uint64_t diff; 3049 3050 /* Do we have a previous sample? */ 3051 if (bbr->r_ctl.rc_lt_bw) { 3052 /* Get the diff in bytes per second */ 3053 if (bbr->r_ctl.rc_lt_bw > bw) 3054 diff = bbr->r_ctl.rc_lt_bw - bw; 3055 else 3056 diff = bw - bbr->r_ctl.rc_lt_bw; 3057 if ((diff <= bbr_lt_bw_diff) || 3058 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3059 /* Consider us policed */ 3060 uint32_t saved_bw; 3061 3062 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3063 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3064 bbr->rc_lt_use_bw = 1; 3065 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3066 /* 3067 * Use pkt based epoch for measuring length of 3068 * policer up 3069 */ 3070 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3071 /* 3072 * reason 4 is we need to start consider being 3073 * policed 3074 */ 3075 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3076 return; 3077 } 3078 } 3079 bbr->r_ctl.rc_lt_bw = bw; 3080 bbr_reset_lt_bw_interval(bbr, cts); 3081 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3082 } 3083 3084 static void 3085 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3086 { 3087 uint32_t ran, deduct; 3088 3089 ran = arc4random_uniform(bbr_rand_ot); 3090 if (ran) { 3091 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3092 bbr->r_ctl.rc_level_state_extra -= deduct; 3093 } 3094 } 3095 /* 3096 * Return randomly the starting state 3097 * to use in probebw. 3098 */ 3099 static uint8_t 3100 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3101 { 3102 uint32_t ran; 3103 uint8_t ret_val; 3104 3105 /* Initialize the offset to 0 */ 3106 bbr->r_ctl.rc_exta_time_gd = 0; 3107 bbr->rc_hit_state_1 = 0; 3108 bbr->r_ctl.rc_level_state_extra = 0; 3109 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3110 /* 3111 * The math works funny here :) the return value is used to set the 3112 * substate and then the state change is called which increments by 3113 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3114 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3115 * we return 1 - 7, so we dont return 0 and end up starting in 3116 * state 1 (DRAIN). 3117 */ 3118 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3119 /* Set an epoch */ 3120 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3121 bbr_set_epoch(bbr, cts, __LINE__); 3122 3123 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3124 return (ret_val); 3125 } 3126 3127 static void 3128 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3129 { 3130 uint32_t diff, d_time; 3131 uint64_t del_time, bw, lost, delivered; 3132 3133 if (bbr->r_use_policer == 0) 3134 return; 3135 if (bbr->rc_lt_use_bw) { 3136 /* We are using lt bw do we stop yet? */ 3137 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3138 if (diff > bbr_lt_bw_max_rtts) { 3139 /* Reset it all */ 3140 reset_all: 3141 bbr_reset_lt_bw_sampling(bbr, cts); 3142 if (bbr->rc_filled_pipe) { 3143 bbr_set_epoch(bbr, cts, __LINE__); 3144 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3145 bbr_substate_change(bbr, cts, __LINE__, 0); 3146 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3147 bbr_log_type_statechange(bbr, cts, __LINE__); 3148 } else { 3149 /* 3150 * This should not happen really 3151 * unless we remove the startup/drain 3152 * restrictions above. 3153 */ 3154 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3155 bbr_set_epoch(bbr, cts, __LINE__); 3156 bbr->r_ctl.rc_bbr_state_time = cts; 3157 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3158 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3159 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3160 bbr_set_state_target(bbr, __LINE__); 3161 bbr_log_type_statechange(bbr, cts, __LINE__); 3162 } 3163 /* reason 0 is to stop using lt-bw */ 3164 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3165 return; 3166 } 3167 if (bbr_lt_intvl_fp == 0) { 3168 /* Not doing false-positive detection */ 3169 return; 3170 } 3171 /* False positive detection */ 3172 if (diff == bbr_lt_intvl_fp) { 3173 /* At bbr_lt_intvl_fp we record the lost */ 3174 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3175 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3176 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3177 /* Now is our loss rate still high? */ 3178 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3179 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3180 if ((delivered == 0) || 3181 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3182 /* No still below our threshold */ 3183 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3184 } else { 3185 /* Yikes its still high, it must be a false positive */ 3186 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3187 goto reset_all; 3188 } 3189 } 3190 return; 3191 } 3192 /* 3193 * Wait for the first loss before sampling, to let the policer 3194 * exhaust its tokens and estimate the steady-state rate allowed by 3195 * the policer. Starting samples earlier includes bursts that 3196 * over-estimate the bw. 3197 */ 3198 if (bbr->rc_lt_is_sampling == 0) { 3199 /* reason 1 is to begin doing the sampling */ 3200 if (loss_detected == 0) 3201 return; 3202 bbr_reset_lt_bw_interval(bbr, cts); 3203 bbr->rc_lt_is_sampling = 1; 3204 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3205 return; 3206 } 3207 /* Now how long were we delivering long term last> */ 3208 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3209 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3210 else 3211 d_time = 0; 3212 3213 /* To avoid underestimates, reset sampling if we run out of data. */ 3214 if (bbr->r_ctl.r_app_limited_until) { 3215 /* Can not measure in app-limited state */ 3216 bbr_reset_lt_bw_sampling(bbr, cts); 3217 /* reason 2 is to reset sampling due to app limits */ 3218 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3219 return; 3220 } 3221 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3222 if (diff < bbr_lt_intvl_min_rtts) { 3223 /* 3224 * need more samples (we don't 3225 * start on a round like linux so 3226 * we need 1 more). 3227 */ 3228 /* 6 is not_enough time or no-loss */ 3229 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3230 return; 3231 } 3232 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3233 /* 3234 * For now if we wait too long, reset all sampling. We need 3235 * to do some research here, its possible that we should 3236 * base this on how much loss as occurred.. something like 3237 * if its under 10% (or some thresh) reset all otherwise 3238 * don't. Thats for phase II I guess. 3239 */ 3240 bbr_reset_lt_bw_sampling(bbr, cts); 3241 /* reason 3 is to reset sampling due too long of sampling */ 3242 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3243 return; 3244 } 3245 /* 3246 * End sampling interval when a packet is lost, so we estimate the 3247 * policer tokens were exhausted. Stopping the sampling before the 3248 * tokens are exhausted under-estimates the policed rate. 3249 */ 3250 if (loss_detected == 0) { 3251 /* 6 is not_enough time or no-loss */ 3252 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3253 return; 3254 } 3255 /* Calculate packets lost and delivered in sampling interval. */ 3256 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3257 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3258 if ((delivered == 0) || 3259 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3260 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3261 return; 3262 } 3263 if (d_time < 1000) { 3264 /* Not enough time. wait */ 3265 /* 6 is not_enough time or no-loss */ 3266 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3267 return; 3268 } 3269 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3270 /* Too long */ 3271 bbr_reset_lt_bw_sampling(bbr, cts); 3272 /* reason 3 is to reset sampling due too long of sampling */ 3273 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3274 return; 3275 } 3276 del_time = d_time; 3277 bw = delivered; 3278 bw *= (uint64_t)USECS_IN_SECOND; 3279 bw /= del_time; 3280 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3281 } 3282 3283 /* 3284 * Allocate a sendmap from our zone. 3285 */ 3286 static struct bbr_sendmap * 3287 bbr_alloc(struct tcp_bbr *bbr) 3288 { 3289 struct bbr_sendmap *rsm; 3290 3291 BBR_STAT_INC(bbr_to_alloc); 3292 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3293 if (rsm) { 3294 bbr->r_ctl.rc_num_maps_alloced++; 3295 return (rsm); 3296 } 3297 if (bbr->r_ctl.rc_free_cnt) { 3298 BBR_STAT_INC(bbr_to_alloc_emerg); 3299 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3300 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3301 bbr->r_ctl.rc_free_cnt--; 3302 return (rsm); 3303 } 3304 BBR_STAT_INC(bbr_to_alloc_failed); 3305 return (NULL); 3306 } 3307 3308 static struct bbr_sendmap * 3309 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3310 { 3311 if ((V_tcp_map_entries_limit > 0) && 3312 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3313 BBR_STAT_INC(bbr_alloc_limited); 3314 if (!bbr->alloc_limit_reported) { 3315 bbr->alloc_limit_reported = 1; 3316 BBR_STAT_INC(bbr_alloc_limited_conns); 3317 } 3318 return (NULL); 3319 } 3320 return (bbr_alloc(bbr)); 3321 } 3322 3323 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3324 static struct bbr_sendmap * 3325 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3326 { 3327 struct bbr_sendmap *rsm; 3328 3329 if (limit_type) { 3330 /* currently there is only one limit type */ 3331 if (V_tcp_map_split_limit > 0 && 3332 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3333 BBR_STAT_INC(bbr_split_limited); 3334 if (!bbr->alloc_limit_reported) { 3335 bbr->alloc_limit_reported = 1; 3336 BBR_STAT_INC(bbr_alloc_limited_conns); 3337 } 3338 return (NULL); 3339 } 3340 } 3341 3342 /* allocate and mark in the limit type, if set */ 3343 rsm = bbr_alloc(bbr); 3344 if (rsm != NULL && limit_type) { 3345 rsm->r_limit_type = limit_type; 3346 bbr->r_ctl.rc_num_split_allocs++; 3347 } 3348 return (rsm); 3349 } 3350 3351 static void 3352 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3353 { 3354 if (rsm->r_limit_type) { 3355 /* currently there is only one limit type */ 3356 bbr->r_ctl.rc_num_split_allocs--; 3357 } 3358 if (rsm->r_is_smallmap) 3359 bbr->r_ctl.rc_num_small_maps_alloced--; 3360 if (bbr->r_ctl.rc_tlp_send == rsm) 3361 bbr->r_ctl.rc_tlp_send = NULL; 3362 if (bbr->r_ctl.rc_resend == rsm) { 3363 bbr->r_ctl.rc_resend = NULL; 3364 } 3365 if (bbr->r_ctl.rc_next == rsm) 3366 bbr->r_ctl.rc_next = NULL; 3367 if (bbr->r_ctl.rc_sacklast == rsm) 3368 bbr->r_ctl.rc_sacklast = NULL; 3369 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3370 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3371 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3372 rsm->r_limit_type = 0; 3373 bbr->r_ctl.rc_free_cnt++; 3374 return; 3375 } 3376 bbr->r_ctl.rc_num_maps_alloced--; 3377 uma_zfree(bbr_zone, rsm); 3378 } 3379 3380 /* 3381 * Returns the BDP. 3382 */ 3383 static uint64_t 3384 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3385 /* 3386 * Calculate the bytes in flight needed given the bw (in bytes per 3387 * second) and the specifyed rtt in useconds. We need to put out the 3388 * returned value per RTT to match that rate. Gain will normally 3389 * raise it up from there. 3390 * 3391 * This should not overflow as long as the bandwidth is below 1 3392 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3393 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3394 */ 3395 uint64_t usec_per_sec; 3396 3397 usec_per_sec = USECS_IN_SECOND; 3398 return ((rtt * bw) / usec_per_sec); 3399 } 3400 3401 /* 3402 * Return the initial cwnd. 3403 */ 3404 static uint32_t 3405 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3406 { 3407 uint32_t i_cwnd; 3408 3409 if (bbr->rc_init_win) { 3410 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3411 } else if (V_tcp_initcwnd_segments) 3412 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3413 max(2 * tp->t_maxseg, 14600)); 3414 else if (V_tcp_do_rfc3390) 3415 i_cwnd = min(4 * tp->t_maxseg, 3416 max(2 * tp->t_maxseg, 4380)); 3417 else { 3418 /* Per RFC5681 Section 3.1 */ 3419 if (tp->t_maxseg > 2190) 3420 i_cwnd = 2 * tp->t_maxseg; 3421 else if (tp->t_maxseg > 1095) 3422 i_cwnd = 3 * tp->t_maxseg; 3423 else 3424 i_cwnd = 4 * tp->t_maxseg; 3425 } 3426 return (i_cwnd); 3427 } 3428 3429 /* 3430 * Given a specified gain, return the target 3431 * cwnd based on that gain. 3432 */ 3433 static uint32_t 3434 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3435 { 3436 uint64_t bdp, rtt; 3437 uint32_t cwnd; 3438 3439 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3440 (bbr_get_full_bw(bbr) == 0)) { 3441 /* No measurements yet */ 3442 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3443 } 3444 /* 3445 * Get bytes per RTT needed (rttProp is normally in 3446 * bbr_cwndtarget_rtt_touse) 3447 */ 3448 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3449 /* Get the bdp from the two values */ 3450 bdp = bbr_get_bw_delay_prod(rtt, bw); 3451 /* Now apply the gain */ 3452 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3453 3454 return (cwnd); 3455 } 3456 3457 static uint32_t 3458 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3459 { 3460 uint32_t cwnd, mss; 3461 3462 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3463 /* Get the base cwnd with gain rounded to a mss */ 3464 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3465 /* 3466 * Add in N (2 default since we do not have a 3467 * fq layer to trap packets in) quanta's per the I-D 3468 * section 4.2.3.2 quanta adjust. 3469 */ 3470 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3471 if (bbr->rc_use_google) { 3472 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3473 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3474 /* 3475 * The linux implementation adds 3476 * an extra 2 x mss in gain cycle which 3477 * is documented no-where except in the code. 3478 * so we add more for Neal undocumented feature 3479 */ 3480 cwnd += 2 * mss; 3481 } 3482 if ((cwnd / mss) & 0x1) { 3483 /* Round up for odd num mss */ 3484 cwnd += mss; 3485 } 3486 } 3487 /* Are we below the min cwnd? */ 3488 if (cwnd < get_min_cwnd(bbr)) 3489 return (get_min_cwnd(bbr)); 3490 return (cwnd); 3491 } 3492 3493 static uint16_t 3494 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3495 { 3496 if (gain < 1) 3497 gain = 1; 3498 return (gain); 3499 } 3500 3501 static uint32_t 3502 bbr_get_header_oh(struct tcp_bbr *bbr) 3503 { 3504 int seg_oh; 3505 3506 seg_oh = 0; 3507 if (bbr->r_ctl.rc_inc_tcp_oh) { 3508 /* Do we include TCP overhead? */ 3509 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3510 } 3511 if (bbr->r_ctl.rc_inc_ip_oh) { 3512 /* Do we include IP overhead? */ 3513 #ifdef INET6 3514 if (bbr->r_is_v6) { 3515 seg_oh += sizeof(struct ip6_hdr); 3516 } else 3517 #endif 3518 { 3519 3520 #ifdef INET 3521 seg_oh += sizeof(struct ip); 3522 #endif 3523 } 3524 } 3525 if (bbr->r_ctl.rc_inc_enet_oh) { 3526 /* Do we include the ethernet overhead? */ 3527 seg_oh += sizeof(struct ether_header); 3528 } 3529 return(seg_oh); 3530 } 3531 3532 static uint32_t 3533 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3534 { 3535 uint64_t divor, res, tim; 3536 3537 if (useconds_time == 0) 3538 return (0); 3539 gain = bbr_gain_adjust(bbr, gain); 3540 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3541 tim = useconds_time; 3542 res = (tim * bw * gain) / divor; 3543 if (res == 0) 3544 res = 1; 3545 return ((uint32_t)res); 3546 } 3547 3548 /* 3549 * Given a gain and a length return the delay in useconds that 3550 * should be used to evenly space out packets 3551 * on the connection (based on the gain factor). 3552 */ 3553 static uint32_t 3554 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3555 { 3556 uint64_t bw, lentim, res; 3557 uint32_t usecs, srtt, over = 0; 3558 uint32_t seg_oh, num_segs, maxseg; 3559 3560 if (len == 0) 3561 return (0); 3562 3563 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3564 num_segs = (len + maxseg - 1) / maxseg; 3565 if (bbr->rc_use_google == 0) { 3566 seg_oh = bbr_get_header_oh(bbr); 3567 len += (num_segs * seg_oh); 3568 } 3569 gain = bbr_gain_adjust(bbr, gain); 3570 bw = bbr_get_bw(bbr); 3571 if (bbr->rc_use_google) { 3572 uint64_t cbw; 3573 3574 /* 3575 * Reduce the b/w by the google discount 3576 * factor 10 = 1%. 3577 */ 3578 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3579 cbw /= (uint64_t)1000; 3580 /* We don't apply a discount if it results in 0 */ 3581 if (cbw > 0) 3582 bw = cbw; 3583 } 3584 lentim = ((uint64_t)len * 3585 (uint64_t)USECS_IN_SECOND * 3586 (uint64_t)BBR_UNIT); 3587 res = lentim / ((uint64_t)gain * bw); 3588 if (res == 0) 3589 res = 1; 3590 usecs = (uint32_t)res; 3591 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3592 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3593 (bbr->rc_use_google == 0) && 3594 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3595 /* 3596 * We cannot let the delay be more than 1/2 the srtt time. 3597 * Otherwise we cannot pace out or send properly. 3598 */ 3599 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3600 BBR_STAT_INC(bbr_hpts_min_time); 3601 } 3602 if (!nolog) 3603 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3604 return (usecs); 3605 } 3606 3607 static void 3608 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3609 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3610 { 3611 INP_WLOCK_ASSERT(tp->t_inpcb); 3612 uint64_t bw; 3613 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3614 int32_t meth; 3615 3616 #ifdef STATS 3617 if ((tp->t_flags & TF_GPUTINPROG) && 3618 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3619 /* 3620 * Strech acks and compressed acks will cause this to 3621 * oscillate but we are doing it the same way as the main 3622 * stack so it will be compariable (though possibly not 3623 * ideal). 3624 */ 3625 int32_t cgput; 3626 int64_t gput, time_stamp; 3627 3628 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3629 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3630 cgput = gput / time_stamp; 3631 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3632 cgput); 3633 if (tp->t_stats_gput_prev > 0) 3634 stats_voi_update_abs_s32(tp->t_stats, 3635 VOI_TCP_GPUT_ND, 3636 ((gput - tp->t_stats_gput_prev) * 100) / 3637 tp->t_stats_gput_prev); 3638 tp->t_flags &= ~TF_GPUTINPROG; 3639 tp->t_stats_gput_prev = cgput; 3640 } 3641 #endif 3642 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3643 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3644 /* We don't change anything in probe-rtt */ 3645 return; 3646 } 3647 maxseg = tp->t_maxseg - bbr->rc_last_options; 3648 saved_bytes = bytes_this_ack; 3649 bytes_this_ack += sack_changed; 3650 if (bytes_this_ack > prev_acked) { 3651 bytes_this_ack -= prev_acked; 3652 /* 3653 * A byte ack'd gives us a full mss 3654 * to be like linux i.e. they count packets. 3655 */ 3656 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3657 bytes_this_ack = maxseg; 3658 } else { 3659 /* Unlikely */ 3660 bytes_this_ack = 0; 3661 } 3662 cwnd = tp->snd_cwnd; 3663 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3664 if (bw) 3665 target_cwnd = bbr_get_target_cwnd(bbr, 3666 bw, 3667 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3668 else 3669 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3670 if (IN_RECOVERY(tp->t_flags) && 3671 (bbr->bbr_prev_in_rec == 0)) { 3672 /* 3673 * We are entering recovery and 3674 * thus packet conservation. 3675 */ 3676 bbr->pkt_conservation = 1; 3677 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3678 cwnd = ctf_flight_size(tp, 3679 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3680 bytes_this_ack; 3681 } 3682 if (IN_RECOVERY(tp->t_flags)) { 3683 uint32_t flight; 3684 3685 bbr->bbr_prev_in_rec = 1; 3686 if (cwnd > losses) { 3687 cwnd -= losses; 3688 if (cwnd < maxseg) 3689 cwnd = maxseg; 3690 } else 3691 cwnd = maxseg; 3692 flight = ctf_flight_size(tp, 3693 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3694 bbr_log_type_cwndupd(bbr, flight, 0, 3695 losses, 10, 0, 0, line); 3696 if (bbr->pkt_conservation) { 3697 uint32_t time_in; 3698 3699 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3700 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3701 else 3702 time_in = 0; 3703 3704 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3705 /* Clear packet conservation after an rttProp */ 3706 bbr->pkt_conservation = 0; 3707 } else { 3708 if ((flight + bytes_this_ack) > cwnd) 3709 cwnd = flight + bytes_this_ack; 3710 if (cwnd < get_min_cwnd(bbr)) 3711 cwnd = get_min_cwnd(bbr); 3712 tp->snd_cwnd = cwnd; 3713 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3714 prev_acked, 1, target_cwnd, th->th_ack, line); 3715 return; 3716 } 3717 } 3718 } else 3719 bbr->bbr_prev_in_rec = 0; 3720 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3721 bbr->r_ctl.restrict_growth--; 3722 if (bytes_this_ack > maxseg) 3723 bytes_this_ack = maxseg; 3724 } 3725 if (bbr->rc_filled_pipe) { 3726 /* 3727 * Here we have exited startup and filled the pipe. We will 3728 * thus allow the cwnd to shrink to the target. We hit here 3729 * mostly. 3730 */ 3731 uint32_t s_cwnd; 3732 3733 meth = 2; 3734 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3735 if (s_cwnd > cwnd) 3736 cwnd = s_cwnd; 3737 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3738 cwnd = s_cwnd; 3739 } else { 3740 /* 3741 * Here we are still in startup, we increase cwnd by what 3742 * has been acked. 3743 */ 3744 if ((cwnd < target_cwnd) || 3745 (bbr->rc_past_init_win == 0)) { 3746 meth = 3; 3747 cwnd += bytes_this_ack; 3748 } else { 3749 /* 3750 * Method 4 means we are at target so no gain in 3751 * startup and past the initial window. 3752 */ 3753 meth = 4; 3754 } 3755 } 3756 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3757 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3758 } 3759 3760 static void 3761 tcp_bbr_partialack(struct tcpcb *tp) 3762 { 3763 struct tcp_bbr *bbr; 3764 3765 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3766 INP_WLOCK_ASSERT(tp->t_inpcb); 3767 if (ctf_flight_size(tp, 3768 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3769 tp->snd_cwnd) { 3770 bbr->r_wanted_output = 1; 3771 } 3772 } 3773 3774 static void 3775 bbr_post_recovery(struct tcpcb *tp) 3776 { 3777 struct tcp_bbr *bbr; 3778 uint32_t flight; 3779 3780 INP_WLOCK_ASSERT(tp->t_inpcb); 3781 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3782 /* 3783 * Here we just exit recovery. 3784 */ 3785 EXIT_RECOVERY(tp->t_flags); 3786 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3787 bbr->r_recovery_bw = 0; 3788 tp->snd_recover = tp->snd_una; 3789 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3790 bbr->pkt_conservation = 0; 3791 if (bbr->rc_use_google == 0) { 3792 /* 3793 * For non-google mode lets 3794 * go ahead and make sure we clear 3795 * the recovery state so if we 3796 * bounce back in to recovery we 3797 * will do PC. 3798 */ 3799 bbr->bbr_prev_in_rec = 0; 3800 } 3801 bbr_log_type_exit_rec(bbr); 3802 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3803 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3804 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3805 } else { 3806 /* For probe-rtt case lets fix up its saved_cwnd */ 3807 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3808 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3809 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3810 } 3811 } 3812 flight = ctf_flight_size(tp, 3813 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3814 if ((bbr->rc_use_google == 0) && 3815 bbr_do_red) { 3816 uint64_t val, lr2use; 3817 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3818 uint32_t *cwnd_p; 3819 3820 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3821 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3822 val /= bbr_get_rtt(bbr, BBR_SRTT); 3823 ratio = (uint32_t)val; 3824 } else 3825 ratio = 1000; 3826 3827 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3828 bbr->r_ctl.recovery_lr, 21, 3829 ratio, 3830 bbr->r_ctl.rc_red_cwnd_pe, 3831 __LINE__); 3832 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3833 goto done; 3834 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3835 bbr_prtt_slam_cwnd) || 3836 (bbr_sub_drain_slam_cwnd && 3837 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3838 bbr->rc_hit_state_1 && 3839 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3840 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3841 bbr_slam_cwnd_in_main_drain)) { 3842 /* 3843 * Here we must poke at the saved cwnd 3844 * as well as the cwnd. 3845 */ 3846 cwnd = bbr->r_ctl.rc_saved_cwnd; 3847 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3848 } else { 3849 cwnd = tp->snd_cwnd; 3850 cwnd_p = &tp->snd_cwnd; 3851 } 3852 maxseg = tp->t_maxseg - bbr->rc_last_options; 3853 /* Add the overall lr with the recovery lr */ 3854 if (bbr->r_ctl.rc_lost == 0) 3855 lr2use = 0; 3856 else if (bbr->r_ctl.rc_delivered == 0) 3857 lr2use = 1000; 3858 else { 3859 lr2use = bbr->r_ctl.rc_lost * 1000; 3860 lr2use /= bbr->r_ctl.rc_delivered; 3861 } 3862 lr2use += bbr->r_ctl.recovery_lr; 3863 acks_inflight = (flight / (maxseg * 2)); 3864 if (bbr_red_scale) { 3865 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3866 lr2use /= bbr_red_scale; 3867 if ((bbr_red_growth_restrict) && 3868 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3869 bbr->r_ctl.restrict_growth += acks_inflight; 3870 } 3871 if (lr2use) { 3872 val = (uint64_t)cwnd * lr2use; 3873 val /= 1000; 3874 if (cwnd > val) 3875 newcwnd = roundup((cwnd - val), maxseg); 3876 else 3877 newcwnd = maxseg; 3878 } else { 3879 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3880 val /= (uint64_t)bbr_red_div; 3881 newcwnd = roundup((uint32_t)val, maxseg); 3882 } 3883 /* with standard delayed acks how many acks can I expect? */ 3884 if (bbr_drop_limit == 0) { 3885 /* 3886 * Anticpate how much we will 3887 * raise the cwnd based on the acks. 3888 */ 3889 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3890 /* We do enforce the min (with the acks) */ 3891 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3892 } 3893 } else { 3894 /* 3895 * A strict drop limit of N is inplace 3896 */ 3897 if (newcwnd < (bbr_drop_limit * maxseg)) { 3898 newcwnd = bbr_drop_limit * maxseg; 3899 } 3900 } 3901 /* For the next N acks do we restrict the growth */ 3902 *cwnd_p = newcwnd; 3903 if (tp->snd_cwnd > newcwnd) 3904 tp->snd_cwnd = newcwnd; 3905 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3906 (uint32_t)lr2use, 3907 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3908 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3909 } 3910 done: 3911 bbr->r_ctl.recovery_lr = 0; 3912 if (flight <= tp->snd_cwnd) { 3913 bbr->r_wanted_output = 1; 3914 } 3915 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3916 } 3917 3918 static void 3919 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3920 { 3921 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3922 /* Limit the drop in b/w to 1/2 our current filter. */ 3923 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3924 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3925 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3926 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3927 tcp_bbr_tso_size_check(bbr, cts); 3928 } 3929 3930 static void 3931 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3932 { 3933 struct tcp_bbr *bbr; 3934 3935 INP_WLOCK_ASSERT(tp->t_inpcb); 3936 #ifdef STATS 3937 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3938 #endif 3939 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3940 switch (type) { 3941 case CC_NDUPACK: 3942 if (!IN_RECOVERY(tp->t_flags)) { 3943 tp->snd_recover = tp->snd_max; 3944 /* Start a new epoch */ 3945 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3946 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3947 /* 3948 * Move forward the lt epoch 3949 * so it won't count the truncated 3950 * epoch. 3951 */ 3952 bbr->r_ctl.rc_lt_epoch++; 3953 } 3954 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3955 /* 3956 * Just like the policer detection code 3957 * if we are in startup we must push 3958 * forward the last startup epoch 3959 * to hide the truncated PE. 3960 */ 3961 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3962 } 3963 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3964 ENTER_RECOVERY(tp->t_flags); 3965 bbr->rc_tlp_rtx_out = 0; 3966 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3967 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3968 if (tcp_in_hpts(bbr->rc_inp) && 3969 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3970 /* 3971 * When we enter recovery, we need to restart 3972 * any timers. This may mean we gain an agg 3973 * early, which will be made up for at the last 3974 * rxt out. 3975 */ 3976 bbr->rc_timer_first = 1; 3977 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3978 } 3979 /* 3980 * Calculate a new cwnd based on to the current 3981 * delivery rate with no gain. We get the bdp 3982 * without gaining it up like we normally would and 3983 * we use the last cur_del_rate. 3984 */ 3985 if ((bbr->rc_use_google == 0) && 3986 (bbr->r_ctl.bbr_rttprobe_gain_val || 3987 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3988 tp->snd_cwnd = ctf_flight_size(tp, 3989 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3990 (tp->t_maxseg - bbr->rc_last_options); 3991 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3992 /* We always gate to min cwnd */ 3993 tp->snd_cwnd = get_min_cwnd(bbr); 3994 } 3995 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3996 } 3997 bbr_log_type_enter_rec(bbr, rsm->r_start); 3998 } 3999 break; 4000 case CC_RTO_ERR: 4001 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 4002 /* RTO was unnecessary, so reset everything. */ 4003 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 4004 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 4005 tp->snd_cwnd = tp->snd_cwnd_prev; 4006 tp->snd_ssthresh = tp->snd_ssthresh_prev; 4007 tp->snd_recover = tp->snd_recover_prev; 4008 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 4009 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 4010 } 4011 tp->t_badrxtwin = 0; 4012 break; 4013 } 4014 } 4015 4016 /* 4017 * Indicate whether this ack should be delayed. We can delay the ack if 4018 * following conditions are met: 4019 * - There is no delayed ack timer in progress. 4020 * - Our last ack wasn't a 0-sized window. We never want to delay 4021 * the ack that opens up a 0-sized window. 4022 * - LRO wasn't used for this segment. We make sure by checking that the 4023 * segment size is not larger than the MSS. 4024 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4025 * connection. 4026 * - The data being acked is less than a full segment (a stretch ack 4027 * of more than a segment we should ack. 4028 * - nsegs is 1 (if its more than that we received more than 1 ack). 4029 */ 4030 #define DELAY_ACK(tp, bbr, nsegs) \ 4031 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4032 ((tp->t_flags & TF_DELACK) == 0) && \ 4033 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4034 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4035 4036 /* 4037 * Return the lowest RSM in the map of 4038 * packets still in flight that is not acked. 4039 * This should normally find on the first one 4040 * since we remove packets from the send 4041 * map after they are marked ACKED. 4042 */ 4043 static struct bbr_sendmap * 4044 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4045 { 4046 struct bbr_sendmap *rsm; 4047 4048 /* 4049 * Walk the time-order transmitted list looking for an rsm that is 4050 * not acked. This will be the one that was sent the longest time 4051 * ago that is still outstanding. 4052 */ 4053 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4054 if (rsm->r_flags & BBR_ACKED) { 4055 continue; 4056 } 4057 goto finish; 4058 } 4059 finish: 4060 return (rsm); 4061 } 4062 4063 static struct bbr_sendmap * 4064 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4065 { 4066 struct bbr_sendmap *prsm; 4067 4068 /* 4069 * Walk the sequence order list backward until we hit and arrive at 4070 * the highest seq not acked. In theory when this is called it 4071 * should be the last segment (which it was not). 4072 */ 4073 prsm = rsm; 4074 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4075 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4076 continue; 4077 } 4078 return (prsm); 4079 } 4080 return (NULL); 4081 } 4082 4083 /* 4084 * Returns to the caller the number of microseconds that 4085 * the packet can be outstanding before we think we 4086 * should have had an ack returned. 4087 */ 4088 static uint32_t 4089 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4090 { 4091 /* 4092 * lro is the flag we use to determine if we have seen reordering. 4093 * If it gets set we have seen reordering. The reorder logic either 4094 * works in one of two ways: 4095 * 4096 * If reorder-fade is configured, then we track the last time we saw 4097 * re-ordering occur. If we reach the point where enough time as 4098 * passed we no longer consider reordering has occuring. 4099 * 4100 * Or if reorder-face is 0, then once we see reordering we consider 4101 * the connection to alway be subject to reordering and just set lro 4102 * to 1. 4103 * 4104 * In the end if lro is non-zero we add the extra time for 4105 * reordering in. 4106 */ 4107 int32_t lro; 4108 uint32_t thresh, t_rxtcur; 4109 4110 if (srtt == 0) 4111 srtt = 1; 4112 if (bbr->r_ctl.rc_reorder_ts) { 4113 if (bbr->r_ctl.rc_reorder_fade) { 4114 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4115 lro = cts - bbr->r_ctl.rc_reorder_ts; 4116 if (lro == 0) { 4117 /* 4118 * No time as passed since the last 4119 * reorder, mark it as reordering. 4120 */ 4121 lro = 1; 4122 } 4123 } else { 4124 /* Negative time? */ 4125 lro = 0; 4126 } 4127 if (lro > bbr->r_ctl.rc_reorder_fade) { 4128 /* Turn off reordering seen too */ 4129 bbr->r_ctl.rc_reorder_ts = 0; 4130 lro = 0; 4131 } 4132 } else { 4133 /* Reodering does not fade */ 4134 lro = 1; 4135 } 4136 } else { 4137 lro = 0; 4138 } 4139 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4140 if (lro) { 4141 /* It must be set, if not you get 1/4 rtt */ 4142 if (bbr->r_ctl.rc_reorder_shift) 4143 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4144 else 4145 thresh += (srtt >> 2); 4146 } else { 4147 thresh += 1000; 4148 } 4149 /* We don't let the rack timeout be above a RTO */ 4150 if ((bbr->rc_tp)->t_srtt == 0) 4151 t_rxtcur = BBR_INITIAL_RTO; 4152 else 4153 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4154 if (thresh > t_rxtcur) { 4155 thresh = t_rxtcur; 4156 } 4157 /* And we don't want it above the RTO max either */ 4158 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4159 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4160 } 4161 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4162 return (thresh); 4163 } 4164 4165 /* 4166 * Return to the caller the amount of time in mico-seconds 4167 * that should be used for the TLP timer from the last 4168 * send time of this packet. 4169 */ 4170 static uint32_t 4171 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4172 struct bbr_sendmap *rsm, uint32_t srtt, 4173 uint32_t cts) 4174 { 4175 uint32_t thresh, len, maxseg, t_rxtcur; 4176 struct bbr_sendmap *prsm; 4177 4178 if (srtt == 0) 4179 srtt = 1; 4180 if (bbr->rc_tlp_threshold) 4181 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4182 else 4183 thresh = (srtt * 2); 4184 maxseg = tp->t_maxseg - bbr->rc_last_options; 4185 /* Get the previous sent packet, if any */ 4186 len = rsm->r_end - rsm->r_start; 4187 4188 /* 2.1 behavior */ 4189 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4190 if (prsm && (len <= maxseg)) { 4191 /* 4192 * Two packets outstanding, thresh should be (2*srtt) + 4193 * possible inter-packet delay (if any). 4194 */ 4195 uint32_t inter_gap = 0; 4196 int idx, nidx; 4197 4198 idx = rsm->r_rtr_cnt - 1; 4199 nidx = prsm->r_rtr_cnt - 1; 4200 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4201 /* Yes it was sent later (or at the same time) */ 4202 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4203 } 4204 thresh += inter_gap; 4205 } else if (len <= maxseg) { 4206 /* 4207 * Possibly compensate for delayed-ack. 4208 */ 4209 uint32_t alt_thresh; 4210 4211 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4212 if (alt_thresh > thresh) 4213 thresh = alt_thresh; 4214 } 4215 /* Not above the current RTO */ 4216 if (tp->t_srtt == 0) 4217 t_rxtcur = BBR_INITIAL_RTO; 4218 else 4219 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4220 4221 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4222 /* Not above an RTO */ 4223 if (thresh > t_rxtcur) { 4224 thresh = t_rxtcur; 4225 } 4226 /* Not above a RTO max */ 4227 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4228 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4229 } 4230 /* And now apply the user TLP min */ 4231 if (thresh < bbr_tlp_min) { 4232 thresh = bbr_tlp_min; 4233 } 4234 return (thresh); 4235 } 4236 4237 /* 4238 * Return one of three RTTs to use (in microseconds). 4239 */ 4240 static __inline uint32_t 4241 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4242 { 4243 uint32_t f_rtt; 4244 uint32_t srtt; 4245 4246 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4247 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4248 /* We have no rtt at all */ 4249 if (bbr->rc_tp->t_srtt == 0) 4250 f_rtt = BBR_INITIAL_RTO; 4251 else 4252 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4253 /* 4254 * Since we don't know how good the rtt is apply a 4255 * delayed-ack min 4256 */ 4257 if (f_rtt < bbr_delayed_ack_time) { 4258 f_rtt = bbr_delayed_ack_time; 4259 } 4260 } 4261 /* Take the filter version or last measured pkt-rtt */ 4262 if (rtt_type == BBR_RTT_PROP) { 4263 srtt = f_rtt; 4264 } else if (rtt_type == BBR_RTT_PKTRTT) { 4265 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4266 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4267 } else { 4268 /* No pkt rtt yet */ 4269 srtt = f_rtt; 4270 } 4271 } else if (rtt_type == BBR_RTT_RACK) { 4272 srtt = bbr->r_ctl.rc_last_rtt; 4273 /* We need to add in any internal delay for our timer */ 4274 if (bbr->rc_ack_was_delayed) 4275 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4276 } else if (rtt_type == BBR_SRTT) { 4277 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4278 } else { 4279 /* TSNH */ 4280 srtt = f_rtt; 4281 #ifdef BBR_INVARIANTS 4282 panic("Unknown rtt request type %d", rtt_type); 4283 #endif 4284 } 4285 return (srtt); 4286 } 4287 4288 static int 4289 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4290 { 4291 uint32_t thresh; 4292 4293 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4294 cts, rsm); 4295 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4296 /* It is lost (past time) */ 4297 return (1); 4298 } 4299 return (0); 4300 } 4301 4302 /* 4303 * Return a sendmap if we need to retransmit something. 4304 */ 4305 static struct bbr_sendmap * 4306 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4307 { 4308 /* 4309 * Check to see that we don't need to fall into recovery. We will 4310 * need to do so if our oldest transmit is past the time we should 4311 * have had an ack. 4312 */ 4313 4314 struct bbr_sendmap *rsm; 4315 int32_t idx; 4316 4317 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4318 /* Nothing outstanding that we know of */ 4319 return (NULL); 4320 } 4321 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4322 if (rsm == NULL) { 4323 /* Nothing in the transmit map */ 4324 return (NULL); 4325 } 4326 if (tp->t_flags & TF_SENTFIN) { 4327 /* Fin restricted, don't find anything once a fin is sent */ 4328 return (NULL); 4329 } 4330 if (rsm->r_flags & BBR_ACKED) { 4331 /* 4332 * Ok the first one is acked (this really should not happen 4333 * since we remove the from the tmap once they are acked) 4334 */ 4335 rsm = bbr_find_lowest_rsm(bbr); 4336 if (rsm == NULL) 4337 return (NULL); 4338 } 4339 idx = rsm->r_rtr_cnt - 1; 4340 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4341 /* Send timestamp is the same or less? can't be ready */ 4342 return (NULL); 4343 } 4344 /* Get our RTT time */ 4345 if (bbr_is_lost(bbr, rsm, cts) && 4346 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4347 (rsm->r_flags & BBR_SACK_PASSED))) { 4348 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4349 rsm->r_flags |= BBR_MARKED_LOST; 4350 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4351 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4352 } 4353 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4354 #ifdef BBR_INVARIANTS 4355 if ((rsm->r_end - rsm->r_start) == 0) 4356 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4357 #endif 4358 return (rsm); 4359 } 4360 return (NULL); 4361 } 4362 4363 /* 4364 * RACK Timer, here we simply do logging and house keeping. 4365 * the normal bbr_output_wtime() function will call the 4366 * appropriate thing to check if we need to do a RACK retransmit. 4367 * We return 1, saying don't proceed with bbr_output_wtime only 4368 * when all timers have been stopped (destroyed PCB?). 4369 */ 4370 static int 4371 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4372 { 4373 /* 4374 * This timer simply provides an internal trigger to send out data. 4375 * The check_recovery_mode call will see if there are needed 4376 * retransmissions, if so we will enter fast-recovery. The output 4377 * call may or may not do the same thing depending on sysctl 4378 * settings. 4379 */ 4380 uint32_t lost; 4381 4382 if (bbr->rc_all_timers_stopped) { 4383 return (1); 4384 } 4385 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4386 /* Its not time yet */ 4387 return (0); 4388 } 4389 BBR_STAT_INC(bbr_to_tot); 4390 lost = bbr->r_ctl.rc_lost; 4391 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4392 bbr_set_state(tp, bbr, 0); 4393 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4394 if (bbr->r_ctl.rc_resend == NULL) { 4395 /* Lets do the check here */ 4396 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4397 } 4398 if (bbr_policer_call_from_rack_to) 4399 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4400 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4401 return (0); 4402 } 4403 4404 static __inline void 4405 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4406 { 4407 int idx; 4408 4409 nrsm->r_start = start; 4410 nrsm->r_end = rsm->r_end; 4411 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4412 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4413 nrsm->r_flags = rsm->r_flags; 4414 /* We don't transfer forward the SYN flag */ 4415 nrsm->r_flags &= ~BBR_HAS_SYN; 4416 /* We move forward the FIN flag, not that this should happen */ 4417 rsm->r_flags &= ~BBR_HAS_FIN; 4418 nrsm->r_dupack = rsm->r_dupack; 4419 nrsm->r_rtr_bytes = 0; 4420 nrsm->r_is_gain = rsm->r_is_gain; 4421 nrsm->r_is_drain = rsm->r_is_drain; 4422 nrsm->r_delivered = rsm->r_delivered; 4423 nrsm->r_ts_valid = rsm->r_ts_valid; 4424 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4425 nrsm->r_del_time = rsm->r_del_time; 4426 nrsm->r_app_limited = rsm->r_app_limited; 4427 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4428 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4429 /* We split a piece the lower section looses any just_ret flag. */ 4430 nrsm->r_bbr_state = rsm->r_bbr_state; 4431 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4432 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4433 } 4434 rsm->r_end = nrsm->r_start; 4435 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4436 idx /= 8; 4437 /* Check if we got too small */ 4438 if ((rsm->r_is_smallmap == 0) && 4439 ((rsm->r_end - rsm->r_start) <= idx)) { 4440 bbr->r_ctl.rc_num_small_maps_alloced++; 4441 rsm->r_is_smallmap = 1; 4442 } 4443 /* Check the new one as well */ 4444 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4445 bbr->r_ctl.rc_num_small_maps_alloced++; 4446 nrsm->r_is_smallmap = 1; 4447 } 4448 } 4449 4450 static int 4451 bbr_sack_mergable(struct bbr_sendmap *at, 4452 uint32_t start, uint32_t end) 4453 { 4454 /* 4455 * Given a sack block defined by 4456 * start and end, and a current position 4457 * at. Return 1 if either side of at 4458 * would show that the block is mergable 4459 * to that side. A block to be mergable 4460 * must have overlap with the start/end 4461 * and be in the SACK'd state. 4462 */ 4463 struct bbr_sendmap *l_rsm; 4464 struct bbr_sendmap *r_rsm; 4465 4466 /* first get the either side blocks */ 4467 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4468 r_rsm = TAILQ_NEXT(at, r_next); 4469 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4470 /* Potentially mergeable */ 4471 if ((l_rsm->r_end == start) || 4472 (SEQ_LT(start, l_rsm->r_end) && 4473 SEQ_GT(end, l_rsm->r_end))) { 4474 /* 4475 * map blk |------| 4476 * sack blk |------| 4477 * <or> 4478 * map blk |------| 4479 * sack blk |------| 4480 */ 4481 return (1); 4482 } 4483 } 4484 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4485 /* Potentially mergeable */ 4486 if ((r_rsm->r_start == end) || 4487 (SEQ_LT(start, r_rsm->r_start) && 4488 SEQ_GT(end, r_rsm->r_start))) { 4489 /* 4490 * map blk |---------| 4491 * sack blk |----| 4492 * <or> 4493 * map blk |---------| 4494 * sack blk |-------| 4495 */ 4496 return (1); 4497 } 4498 } 4499 return (0); 4500 } 4501 4502 static struct bbr_sendmap * 4503 bbr_merge_rsm(struct tcp_bbr *bbr, 4504 struct bbr_sendmap *l_rsm, 4505 struct bbr_sendmap *r_rsm) 4506 { 4507 /* 4508 * We are merging two ack'd RSM's, 4509 * the l_rsm is on the left (lower seq 4510 * values) and the r_rsm is on the right 4511 * (higher seq value). The simplest way 4512 * to merge these is to move the right 4513 * one into the left. I don't think there 4514 * is any reason we need to try to find 4515 * the oldest (or last oldest retransmitted). 4516 */ 4517 l_rsm->r_end = r_rsm->r_end; 4518 if (l_rsm->r_dupack < r_rsm->r_dupack) 4519 l_rsm->r_dupack = r_rsm->r_dupack; 4520 if (r_rsm->r_rtr_bytes) 4521 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4522 if (r_rsm->r_in_tmap) { 4523 /* This really should not happen */ 4524 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4525 } 4526 if (r_rsm->r_app_limited) 4527 l_rsm->r_app_limited = r_rsm->r_app_limited; 4528 /* Now the flags */ 4529 if (r_rsm->r_flags & BBR_HAS_FIN) 4530 l_rsm->r_flags |= BBR_HAS_FIN; 4531 if (r_rsm->r_flags & BBR_TLP) 4532 l_rsm->r_flags |= BBR_TLP; 4533 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4534 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4535 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4536 /* This really should not happen */ 4537 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4538 } 4539 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4540 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4541 /* Transfer the split limit to the map we free */ 4542 r_rsm->r_limit_type = l_rsm->r_limit_type; 4543 l_rsm->r_limit_type = 0; 4544 } 4545 bbr_free(bbr, r_rsm); 4546 return(l_rsm); 4547 } 4548 4549 /* 4550 * TLP Timer, here we simply setup what segment we want to 4551 * have the TLP expire on, the normal bbr_output_wtime() will then 4552 * send it out. 4553 * 4554 * We return 1, saying don't proceed with bbr_output_wtime only 4555 * when all timers have been stopped (destroyed PCB?). 4556 */ 4557 static int 4558 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4559 { 4560 /* 4561 * Tail Loss Probe. 4562 */ 4563 struct bbr_sendmap *rsm = NULL; 4564 struct socket *so; 4565 uint32_t amm; 4566 uint32_t out, avail; 4567 uint32_t maxseg; 4568 int collapsed_win = 0; 4569 4570 if (bbr->rc_all_timers_stopped) { 4571 return (1); 4572 } 4573 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4574 /* Its not time yet */ 4575 return (0); 4576 } 4577 if (ctf_progress_timeout_check(tp, true)) { 4578 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4579 return (-ETIMEDOUT); /* tcp_drop() */ 4580 } 4581 /* Did we somehow get into persists? */ 4582 if (bbr->rc_in_persist) { 4583 return (0); 4584 } 4585 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4586 bbr_set_state(tp, bbr, 0); 4587 BBR_STAT_INC(bbr_tlp_tot); 4588 maxseg = tp->t_maxseg - bbr->rc_last_options; 4589 /* 4590 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4591 * need to figure out how to force a full MSS segment out. 4592 */ 4593 so = tp->t_inpcb->inp_socket; 4594 avail = sbavail(&so->so_snd); 4595 out = ctf_outstanding(tp); 4596 if (out > tp->snd_wnd) { 4597 /* special case, we need a retransmission */ 4598 collapsed_win = 1; 4599 goto need_retran; 4600 } 4601 if (avail > out) { 4602 /* New data is available */ 4603 amm = avail - out; 4604 if (amm > maxseg) { 4605 amm = maxseg; 4606 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4607 /* not enough to fill a MTU and no-delay is off */ 4608 goto need_retran; 4609 } 4610 /* Set the send-new override */ 4611 if ((out + amm) <= tp->snd_wnd) { 4612 bbr->rc_tlp_new_data = 1; 4613 } else { 4614 goto need_retran; 4615 } 4616 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4617 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4618 bbr->r_ctl.rc_tlp_send = NULL; 4619 /* cap any slots */ 4620 BBR_STAT_INC(bbr_tlp_newdata); 4621 goto send; 4622 } 4623 need_retran: 4624 /* 4625 * Ok we need to arrange the last un-acked segment to be re-sent, or 4626 * optionally the first un-acked segment. 4627 */ 4628 if (collapsed_win == 0) { 4629 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4630 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4631 rsm = bbr_find_high_nonack(bbr, rsm); 4632 } 4633 if (rsm == NULL) { 4634 goto restore; 4635 } 4636 } else { 4637 /* 4638 * We must find the last segment 4639 * that was acceptable by the client. 4640 */ 4641 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4642 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4643 /* Found one */ 4644 break; 4645 } 4646 } 4647 if (rsm == NULL) { 4648 /* None? if so send the first */ 4649 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4650 if (rsm == NULL) 4651 goto restore; 4652 } 4653 } 4654 if ((rsm->r_end - rsm->r_start) > maxseg) { 4655 /* 4656 * We need to split this the last segment in two. 4657 */ 4658 struct bbr_sendmap *nrsm; 4659 4660 nrsm = bbr_alloc_full_limit(bbr); 4661 if (nrsm == NULL) { 4662 /* 4663 * We can't get memory to split, we can either just 4664 * not split it. Or retransmit the whole piece, lets 4665 * do the large send (BTLP :-) ). 4666 */ 4667 goto go_for_it; 4668 } 4669 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4670 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4671 if (rsm->r_in_tmap) { 4672 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4673 nrsm->r_in_tmap = 1; 4674 } 4675 rsm->r_flags &= (~BBR_HAS_FIN); 4676 rsm = nrsm; 4677 } 4678 go_for_it: 4679 bbr->r_ctl.rc_tlp_send = rsm; 4680 bbr->rc_tlp_rtx_out = 1; 4681 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4682 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4683 tp->t_rxtshift++; 4684 } else { 4685 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4686 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4687 } 4688 send: 4689 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4690 /* 4691 * Can't [re]/transmit a segment we have retranmitted the 4692 * max times. We need the retransmit timer to take over. 4693 */ 4694 restore: 4695 bbr->rc_tlp_new_data = 0; 4696 bbr->r_ctl.rc_tlp_send = NULL; 4697 if (rsm) 4698 rsm->r_flags &= ~BBR_TLP; 4699 BBR_STAT_INC(bbr_tlp_retran_fail); 4700 return (0); 4701 } else if (rsm) { 4702 rsm->r_flags |= BBR_TLP; 4703 } 4704 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4705 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4706 /* 4707 * We have retransmitted to many times for TLP. Switch to 4708 * the regular RTO timer 4709 */ 4710 goto restore; 4711 } 4712 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4713 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4714 return (0); 4715 } 4716 4717 /* 4718 * Delayed ack Timer, here we simply need to setup the 4719 * ACK_NOW flag and remove the DELACK flag. From there 4720 * the output routine will send the ack out. 4721 * 4722 * We only return 1, saying don't proceed, if all timers 4723 * are stopped (destroyed PCB?). 4724 */ 4725 static int 4726 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4727 { 4728 if (bbr->rc_all_timers_stopped) { 4729 return (1); 4730 } 4731 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4732 tp->t_flags &= ~TF_DELACK; 4733 tp->t_flags |= TF_ACKNOW; 4734 KMOD_TCPSTAT_INC(tcps_delack); 4735 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4736 return (0); 4737 } 4738 4739 /* 4740 * Here we send a KEEP-ALIVE like probe to the 4741 * peer, we do not send data. 4742 * 4743 * We only return 1, saying don't proceed, if all timers 4744 * are stopped (destroyed PCB?). 4745 */ 4746 static int 4747 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4748 { 4749 struct tcptemp *t_template; 4750 int32_t retval = 1; 4751 4752 if (bbr->rc_all_timers_stopped) { 4753 return (1); 4754 } 4755 if (bbr->rc_in_persist == 0) 4756 return (0); 4757 KASSERT(tp->t_inpcb != NULL, 4758 ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 4759 /* 4760 * Persistence timer into zero window. Force a byte to be output, if 4761 * possible. 4762 */ 4763 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4764 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4765 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4766 /* 4767 * Have we exceeded the user specified progress time? 4768 */ 4769 if (ctf_progress_timeout_check(tp, true)) { 4770 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4771 return (-ETIMEDOUT); /* tcp_drop() */ 4772 } 4773 /* 4774 * Hack: if the peer is dead/unreachable, we do not time out if the 4775 * window is closed. After a full backoff, drop the connection if 4776 * the idle time (no responses to probes) reaches the maximum 4777 * backoff that we would use if retransmitting. 4778 */ 4779 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 4780 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4781 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4782 KMOD_TCPSTAT_INC(tcps_persistdrop); 4783 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4784 return (-ETIMEDOUT); /* tcp_drop() */ 4785 } 4786 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4787 tp->snd_una == tp->snd_max) { 4788 bbr_exit_persist(tp, bbr, cts, __LINE__); 4789 retval = 0; 4790 goto out; 4791 } 4792 /* 4793 * If the user has closed the socket then drop a persisting 4794 * connection after a much reduced timeout. 4795 */ 4796 if (tp->t_state > TCPS_CLOSE_WAIT && 4797 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4798 KMOD_TCPSTAT_INC(tcps_persistdrop); 4799 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4800 return (-ETIMEDOUT); /* tcp_drop() */ 4801 } 4802 t_template = tcpip_maketemplate(bbr->rc_inp); 4803 if (t_template) { 4804 tcp_respond(tp, t_template->tt_ipgen, 4805 &t_template->tt_t, (struct mbuf *)NULL, 4806 tp->rcv_nxt, tp->snd_una - 1, 0); 4807 /* This sends an ack */ 4808 if (tp->t_flags & TF_DELACK) 4809 tp->t_flags &= ~TF_DELACK; 4810 free(t_template, M_TEMP); 4811 } 4812 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 4813 tp->t_rxtshift++; 4814 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4815 out: 4816 return (retval); 4817 } 4818 4819 /* 4820 * If a keepalive goes off, we had no other timers 4821 * happening. We always return 1 here since this 4822 * routine either drops the connection or sends 4823 * out a segment with respond. 4824 */ 4825 static int 4826 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4827 { 4828 struct tcptemp *t_template; 4829 struct inpcb *inp; 4830 4831 if (bbr->rc_all_timers_stopped) { 4832 return (1); 4833 } 4834 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4835 inp = tp->t_inpcb; 4836 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4837 /* 4838 * Keep-alive timer went off; send something or drop connection if 4839 * idle for too long. 4840 */ 4841 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4842 if (tp->t_state < TCPS_ESTABLISHED) 4843 goto dropit; 4844 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4845 tp->t_state <= TCPS_CLOSING) { 4846 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4847 goto dropit; 4848 /* 4849 * Send a packet designed to force a response if the peer is 4850 * up and reachable: either an ACK if the connection is 4851 * still alive, or an RST if the peer has closed the 4852 * connection due to timeout or reboot. Using sequence 4853 * number tp->snd_una-1 causes the transmitted zero-length 4854 * segment to lie outside the receive window; by the 4855 * protocol spec, this requires the correspondent TCP to 4856 * respond. 4857 */ 4858 KMOD_TCPSTAT_INC(tcps_keepprobe); 4859 t_template = tcpip_maketemplate(inp); 4860 if (t_template) { 4861 tcp_respond(tp, t_template->tt_ipgen, 4862 &t_template->tt_t, (struct mbuf *)NULL, 4863 tp->rcv_nxt, tp->snd_una - 1, 0); 4864 free(t_template, M_TEMP); 4865 } 4866 } 4867 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4868 return (1); 4869 dropit: 4870 KMOD_TCPSTAT_INC(tcps_keepdrops); 4871 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4872 return (-ETIMEDOUT); /* tcp_drop() */ 4873 } 4874 4875 /* 4876 * Retransmit helper function, clear up all the ack 4877 * flags and take care of important book keeping. 4878 */ 4879 static void 4880 bbr_remxt_tmr(struct tcpcb *tp) 4881 { 4882 /* 4883 * The retransmit timer went off, all sack'd blocks must be 4884 * un-acked. 4885 */ 4886 struct bbr_sendmap *rsm, *trsm = NULL; 4887 struct tcp_bbr *bbr; 4888 uint32_t cts, lost; 4889 4890 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4891 cts = tcp_get_usecs(&bbr->rc_tv); 4892 lost = bbr->r_ctl.rc_lost; 4893 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4894 bbr_set_state(tp, bbr, 0); 4895 4896 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4897 if (rsm->r_flags & BBR_ACKED) { 4898 uint32_t old_flags; 4899 4900 rsm->r_dupack = 0; 4901 if (rsm->r_in_tmap == 0) { 4902 /* We must re-add it back to the tlist */ 4903 if (trsm == NULL) { 4904 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4905 } else { 4906 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4907 } 4908 rsm->r_in_tmap = 1; 4909 } 4910 old_flags = rsm->r_flags; 4911 rsm->r_flags |= BBR_RXT_CLEARED; 4912 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4913 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4914 } else { 4915 if ((tp->t_state < TCPS_ESTABLISHED) && 4916 (rsm->r_start == tp->snd_una)) { 4917 /* 4918 * Special case for TCP FO. Where 4919 * we sent more data beyond the snd_max. 4920 * We don't mark that as lost and stop here. 4921 */ 4922 break; 4923 } 4924 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4925 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4926 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4927 } 4928 if (bbr_marks_rxt_sack_passed) { 4929 /* 4930 * With this option, we will rack out 4931 * in 1ms increments the rest of the packets. 4932 */ 4933 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4934 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4935 } else { 4936 /* 4937 * With this option we only mark them lost 4938 * and remove all sack'd markings. We will run 4939 * another RXT or a TLP. This will cause 4940 * us to eventually send more based on what 4941 * ack's come in. 4942 */ 4943 rsm->r_flags |= BBR_MARKED_LOST; 4944 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4945 rsm->r_flags &= ~BBR_SACK_PASSED; 4946 } 4947 } 4948 trsm = rsm; 4949 } 4950 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4951 /* Clear the count (we just un-acked them) */ 4952 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4953 bbr->rc_tlp_new_data = 0; 4954 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4955 /* zap the behindness on a rxt */ 4956 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4957 bbr->r_agg_early_set = 0; 4958 bbr->r_ctl.rc_agg_early = 0; 4959 bbr->rc_tlp_rtx_out = 0; 4960 bbr->r_ctl.rc_sacked = 0; 4961 bbr->r_ctl.rc_sacklast = NULL; 4962 bbr->r_timer_override = 1; 4963 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4964 } 4965 4966 /* 4967 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4968 * we will setup to retransmit the lowest seq number outstanding. 4969 */ 4970 static int 4971 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4972 { 4973 int32_t rexmt; 4974 int32_t retval = 0; 4975 bool isipv6; 4976 4977 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4978 if (bbr->rc_all_timers_stopped) { 4979 return (1); 4980 } 4981 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4982 (tp->snd_una == tp->snd_max)) { 4983 /* Nothing outstanding .. nothing to do */ 4984 return (0); 4985 } 4986 /* 4987 * Retransmission timer went off. Message has not been acked within 4988 * retransmit interval. Back off to a longer retransmit interval 4989 * and retransmit one segment. 4990 */ 4991 if (ctf_progress_timeout_check(tp, true)) { 4992 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4993 return (-ETIMEDOUT); /* tcp_drop() */ 4994 } 4995 bbr_remxt_tmr(tp); 4996 if ((bbr->r_ctl.rc_resend == NULL) || 4997 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 4998 /* 4999 * If the rwnd collapsed on 5000 * the one we are retransmitting 5001 * it does not count against the 5002 * rxt count. 5003 */ 5004 tp->t_rxtshift++; 5005 } 5006 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { 5007 tp->t_rxtshift = TCP_MAXRXTSHIFT; 5008 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 5009 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 5010 /* XXXGL: previously t_softerror was casted to uint16_t */ 5011 MPASS(tp->t_softerror >= 0); 5012 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 5013 return (retval); /* tcp_drop() */ 5014 } 5015 if (tp->t_state == TCPS_SYN_SENT) { 5016 /* 5017 * If the SYN was retransmitted, indicate CWND to be limited 5018 * to 1 segment in cc_conn_init(). 5019 */ 5020 tp->snd_cwnd = 1; 5021 } else if (tp->t_rxtshift == 1) { 5022 /* 5023 * first retransmit; record ssthresh and cwnd so they can be 5024 * recovered if this turns out to be a "bad" retransmit. A 5025 * retransmit is considered "bad" if an ACK for this segment 5026 * is received within RTT/2 interval; the assumption here is 5027 * that the ACK was already in flight. See "On Estimating 5028 * End-to-End Network Path Properties" by Allman and Paxson 5029 * for more details. 5030 */ 5031 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5032 if (!IN_RECOVERY(tp->t_flags)) { 5033 tp->snd_cwnd_prev = tp->snd_cwnd; 5034 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5035 tp->snd_recover_prev = tp->snd_recover; 5036 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5037 tp->t_flags |= TF_PREVVALID; 5038 } else { 5039 tp->t_flags &= ~TF_PREVVALID; 5040 } 5041 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5042 } else { 5043 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5044 tp->t_flags &= ~TF_PREVVALID; 5045 } 5046 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5047 if ((tp->t_state == TCPS_SYN_SENT) || 5048 (tp->t_state == TCPS_SYN_RECEIVED)) 5049 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5050 else 5051 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5052 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5053 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5054 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5055 /* 5056 * We enter the path for PLMTUD if connection is established or, if 5057 * connection is FIN_WAIT_1 status, reason for the last is that if 5058 * amount of data we send is very small, we could send it in couple 5059 * of packets and process straight to FIN. In that case we won't 5060 * catch ESTABLISHED state. 5061 */ 5062 #ifdef INET6 5063 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; 5064 #else 5065 isipv6 = false; 5066 #endif 5067 if (((V_tcp_pmtud_blackhole_detect == 1) || 5068 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5069 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5070 ((tp->t_state == TCPS_ESTABLISHED) || 5071 (tp->t_state == TCPS_FIN_WAIT_1))) { 5072 /* 5073 * Idea here is that at each stage of mtu probe (usually, 5074 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5075 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5076 * should take care of that. 5077 */ 5078 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5079 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5080 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5081 tp->t_rxtshift % 2 == 0)) { 5082 /* 5083 * Enter Path MTU Black-hole Detection mechanism: - 5084 * Disable Path MTU Discovery (IP "DF" bit). - 5085 * Reduce MTU to lower value than what we negotiated 5086 * with peer. 5087 */ 5088 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5089 /* 5090 * Record that we may have found a black 5091 * hole. 5092 */ 5093 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5094 /* Keep track of previous MSS. */ 5095 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5096 } 5097 /* 5098 * Reduce the MSS to blackhole value or to the 5099 * default in an attempt to retransmit. 5100 */ 5101 #ifdef INET6 5102 isipv6 = bbr->r_is_v6; 5103 if (isipv6 && 5104 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5105 /* Use the sysctl tuneable blackhole MSS. */ 5106 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5107 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5108 } else if (isipv6) { 5109 /* Use the default MSS. */ 5110 tp->t_maxseg = V_tcp_v6mssdflt; 5111 /* 5112 * Disable Path MTU Discovery when we switch 5113 * to minmss. 5114 */ 5115 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5116 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5117 } 5118 #endif 5119 #if defined(INET6) && defined(INET) 5120 else 5121 #endif 5122 #ifdef INET 5123 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5124 /* Use the sysctl tuneable blackhole MSS. */ 5125 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5126 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5127 } else { 5128 /* Use the default MSS. */ 5129 tp->t_maxseg = V_tcp_mssdflt; 5130 /* 5131 * Disable Path MTU Discovery when we switch 5132 * to minmss. 5133 */ 5134 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5135 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5136 } 5137 #endif 5138 } else { 5139 /* 5140 * If further retransmissions are still unsuccessful 5141 * with a lowered MTU, maybe this isn't a blackhole 5142 * and we restore the previous MSS and blackhole 5143 * detection flags. The limit '6' is determined by 5144 * giving each probe stage (1448, 1188, 524) 2 5145 * chances to recover. 5146 */ 5147 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5148 (tp->t_rxtshift >= 6)) { 5149 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5150 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5151 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5152 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5153 } 5154 } 5155 } 5156 /* 5157 * Disable RFC1323 and SACK if we haven't got any response to our 5158 * third SYN to work-around some broken terminal servers (most of 5159 * which have hopefully been retired) that have bad VJ header 5160 * compression code which trashes TCP segments containing 5161 * unknown-to-them TCP options. 5162 */ 5163 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5164 (tp->t_rxtshift == 3)) 5165 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5166 /* 5167 * If we backed off this far, our srtt estimate is probably bogus. 5168 * Clobber it so we'll take the next rtt measurement as our srtt; 5169 * move the current srtt into rttvar to keep the current retransmit 5170 * times until then. 5171 */ 5172 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5173 #ifdef INET6 5174 if (bbr->r_is_v6) 5175 in6_losing(tp->t_inpcb); 5176 else 5177 #endif 5178 in_losing(tp->t_inpcb); 5179 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5180 tp->t_srtt = 0; 5181 } 5182 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5183 tp->snd_recover = tp->snd_max; 5184 tp->t_flags |= TF_ACKNOW; 5185 tp->t_rtttime = 0; 5186 5187 return (retval); 5188 } 5189 5190 static int 5191 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5192 { 5193 int32_t ret = 0; 5194 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5195 5196 if (timers == 0) { 5197 return (0); 5198 } 5199 if (tp->t_state == TCPS_LISTEN) { 5200 /* no timers on listen sockets */ 5201 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5202 return (0); 5203 return (1); 5204 } 5205 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5206 uint32_t left; 5207 5208 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5209 ret = -1; 5210 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5211 return (0); 5212 } 5213 if (hpts_calling == 0) { 5214 ret = -2; 5215 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5216 return (0); 5217 } 5218 /* 5219 * Ok our timer went off early and we are not paced false 5220 * alarm, go back to sleep. 5221 */ 5222 left = bbr->r_ctl.rc_timer_exp - cts; 5223 ret = -3; 5224 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5225 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left)); 5226 return (1); 5227 } 5228 bbr->rc_tmr_stopped = 0; 5229 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5230 if (timers & PACE_TMR_DELACK) { 5231 ret = bbr_timeout_delack(tp, bbr, cts); 5232 } else if (timers & PACE_TMR_PERSIT) { 5233 ret = bbr_timeout_persist(tp, bbr, cts); 5234 } else if (timers & PACE_TMR_RACK) { 5235 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5236 ret = bbr_timeout_rack(tp, bbr, cts); 5237 } else if (timers & PACE_TMR_TLP) { 5238 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5239 ret = bbr_timeout_tlp(tp, bbr, cts); 5240 } else if (timers & PACE_TMR_RXT) { 5241 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5242 ret = bbr_timeout_rxt(tp, bbr, cts); 5243 } else if (timers & PACE_TMR_KEEP) { 5244 ret = bbr_timeout_keepalive(tp, bbr, cts); 5245 } 5246 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5247 return (ret); 5248 } 5249 5250 static void 5251 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5252 { 5253 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5254 uint8_t hpts_removed = 0; 5255 5256 if (tcp_in_hpts(bbr->rc_inp) && 5257 (bbr->rc_timer_first == 1)) { 5258 /* 5259 * If we are canceling timer's when we have the 5260 * timer ahead of the output being paced. We also 5261 * must remove ourselves from the hpts. 5262 */ 5263 hpts_removed = 1; 5264 tcp_hpts_remove(bbr->rc_inp); 5265 if (bbr->r_ctl.rc_last_delay_val) { 5266 /* Update the last hptsi delay too */ 5267 uint32_t time_since_send; 5268 5269 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5270 time_since_send = cts - bbr->rc_pacer_started; 5271 else 5272 time_since_send = 0; 5273 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5274 /* Cut down our slot time */ 5275 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5276 } else { 5277 bbr->r_ctl.rc_last_delay_val = 0; 5278 } 5279 bbr->rc_pacer_started = cts; 5280 } 5281 } 5282 bbr->rc_timer_first = 0; 5283 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5284 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5285 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5286 } 5287 } 5288 5289 static void 5290 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type) 5291 { 5292 struct tcp_bbr *bbr; 5293 5294 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5295 bbr->rc_all_timers_stopped = 1; 5296 return; 5297 } 5298 5299 /* 5300 * stop all timers always returning 0. 5301 */ 5302 static int 5303 bbr_stopall(struct tcpcb *tp) 5304 { 5305 return (0); 5306 } 5307 5308 static void 5309 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 5310 { 5311 return; 5312 } 5313 5314 /* 5315 * return true if a bbr timer (rack or tlp) is active. 5316 */ 5317 static int 5318 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type) 5319 { 5320 return (0); 5321 } 5322 5323 static uint32_t 5324 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5325 { 5326 struct bbr_sendmap *rsm; 5327 5328 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5329 if ((rsm == NULL) || (u_rsm == rsm)) 5330 return (cts); 5331 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5332 } 5333 5334 static void 5335 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5336 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5337 { 5338 int32_t idx; 5339 5340 rsm->r_rtr_cnt++; 5341 rsm->r_dupack = 0; 5342 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5343 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5344 rsm->r_flags |= BBR_OVERMAX; 5345 } 5346 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5347 /* Take off the collapsed flag at rxt */ 5348 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5349 } 5350 if (rsm->r_flags & BBR_MARKED_LOST) { 5351 /* We have retransmitted, its no longer lost */ 5352 rsm->r_flags &= ~BBR_MARKED_LOST; 5353 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5354 } 5355 if (rsm->r_flags & BBR_RXT_CLEARED) { 5356 /* 5357 * We hit a RXT timer on it and 5358 * we cleared the "acked" flag. 5359 * We now have it going back into 5360 * flight, we can remove the cleared 5361 * flag and possibly do accounting on 5362 * this piece. 5363 */ 5364 rsm->r_flags &= ~BBR_RXT_CLEARED; 5365 } 5366 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5367 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5368 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5369 } 5370 idx = rsm->r_rtr_cnt - 1; 5371 rsm->r_tim_lastsent[idx] = cts; 5372 rsm->r_pacing_delay = pacing_time; 5373 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5374 rsm->r_ts_valid = bbr->rc_ts_valid; 5375 if (bbr->rc_ts_valid) 5376 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5377 if (bbr->r_ctl.r_app_limited_until) 5378 rsm->r_app_limited = 1; 5379 else 5380 rsm->r_app_limited = 0; 5381 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5382 rsm->r_bbr_state = bbr_state_val(bbr); 5383 else 5384 rsm->r_bbr_state = 8; 5385 if (rsm->r_flags & BBR_ACKED) { 5386 /* Problably MTU discovery messing with us */ 5387 uint32_t old_flags; 5388 5389 old_flags = rsm->r_flags; 5390 rsm->r_flags &= ~BBR_ACKED; 5391 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5392 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5393 if (bbr->r_ctl.rc_sacked == 0) 5394 bbr->r_ctl.rc_sacklast = NULL; 5395 } 5396 if (rsm->r_in_tmap) { 5397 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5398 } 5399 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5400 rsm->r_in_tmap = 1; 5401 if (rsm->r_flags & BBR_SACK_PASSED) { 5402 /* We have retransmitted due to the SACK pass */ 5403 rsm->r_flags &= ~BBR_SACK_PASSED; 5404 rsm->r_flags |= BBR_WAS_SACKPASS; 5405 } 5406 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5407 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5408 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5409 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5410 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5411 rsm->r_is_gain = 1; 5412 rsm->r_is_drain = 0; 5413 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5414 rsm->r_is_drain = 1; 5415 rsm->r_is_gain = 0; 5416 } else { 5417 rsm->r_is_drain = 0; 5418 rsm->r_is_gain = 0; 5419 } 5420 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5421 } 5422 5423 /* 5424 * Returns 0, or the sequence where we stopped 5425 * updating. We also update the lenp to be the amount 5426 * of data left. 5427 */ 5428 5429 static uint32_t 5430 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5431 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5432 { 5433 /* 5434 * We (re-)transmitted starting at rsm->r_start for some length 5435 * (possibly less than r_end. 5436 */ 5437 struct bbr_sendmap *nrsm; 5438 uint32_t c_end; 5439 int32_t len; 5440 5441 len = *lenp; 5442 c_end = rsm->r_start + len; 5443 if (SEQ_GEQ(c_end, rsm->r_end)) { 5444 /* 5445 * We retransmitted the whole piece or more than the whole 5446 * slopping into the next rsm. 5447 */ 5448 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5449 if (c_end == rsm->r_end) { 5450 *lenp = 0; 5451 return (0); 5452 } else { 5453 int32_t act_len; 5454 5455 /* Hangs over the end return whats left */ 5456 act_len = rsm->r_end - rsm->r_start; 5457 *lenp = (len - act_len); 5458 return (rsm->r_end); 5459 } 5460 /* We don't get out of this block. */ 5461 } 5462 /* 5463 * Here we retransmitted less than the whole thing which means we 5464 * have to split this into what was transmitted and what was not. 5465 */ 5466 nrsm = bbr_alloc_full_limit(bbr); 5467 if (nrsm == NULL) { 5468 *lenp = 0; 5469 return (0); 5470 } 5471 /* 5472 * So here we are going to take the original rsm and make it what we 5473 * retransmitted. nrsm will be the tail portion we did not 5474 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5475 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5476 * 1, 6 and the new piece will be 6, 11. 5477 */ 5478 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5479 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5480 nrsm->r_dupack = 0; 5481 if (rsm->r_in_tmap) { 5482 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5483 nrsm->r_in_tmap = 1; 5484 } 5485 rsm->r_flags &= (~BBR_HAS_FIN); 5486 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5487 *lenp = 0; 5488 return (0); 5489 } 5490 5491 static uint64_t 5492 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5493 { 5494 uint64_t bw; 5495 5496 bw = bbr_get_bw(bbr); 5497 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5498 bw /= (uint64_t)BBR_UNIT; 5499 return(bw); 5500 } 5501 5502 static void 5503 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5504 uint64_t act_rate, uint64_t rate_wanted) 5505 { 5506 /* 5507 * We could not get a full gains worth 5508 * of rate. 5509 */ 5510 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5511 /* we can't even get the real rate */ 5512 uint64_t red; 5513 5514 bbr->skip_gain = 1; 5515 bbr->gain_is_limited = 0; 5516 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5517 if (red) 5518 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5519 } else { 5520 /* We can use a lower gain */ 5521 bbr->skip_gain = 0; 5522 bbr->gain_is_limited = 1; 5523 } 5524 } 5525 5526 static void 5527 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5528 { 5529 const struct tcp_hwrate_limit_table *nrte; 5530 int error, rate = -1; 5531 5532 if (bbr->r_ctl.crte == NULL) 5533 return; 5534 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5535 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5536 /* Lost our routes? */ 5537 /* Clear the way for a re-attempt */ 5538 bbr->bbr_attempt_hdwr_pace = 0; 5539 lost_rate: 5540 bbr->gain_is_limited = 0; 5541 bbr->skip_gain = 0; 5542 bbr->bbr_hdrw_pacing = 0; 5543 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5544 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5545 tcp_bbr_tso_size_check(bbr, cts); 5546 return; 5547 } 5548 rate = bbr_get_hardware_rate(bbr); 5549 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5550 bbr->rc_tp, 5551 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5552 rate, 5553 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5554 &error, NULL); 5555 if (nrte == NULL) { 5556 goto lost_rate; 5557 } 5558 if (nrte != bbr->r_ctl.crte) { 5559 bbr->r_ctl.crte = nrte; 5560 if (error == 0) { 5561 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5562 if (bbr->r_ctl.crte->rate < rate) { 5563 /* We have a problem */ 5564 bbr_setup_less_of_rate(bbr, cts, 5565 bbr->r_ctl.crte->rate, rate); 5566 } else { 5567 /* We are good */ 5568 bbr->gain_is_limited = 0; 5569 bbr->skip_gain = 0; 5570 } 5571 } else { 5572 /* A failure should release the tag */ 5573 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5574 bbr->gain_is_limited = 0; 5575 bbr->skip_gain = 0; 5576 bbr->bbr_hdrw_pacing = 0; 5577 } 5578 bbr_type_log_hdwr_pacing(bbr, 5579 bbr->r_ctl.crte->ptbl->rs_ifp, 5580 rate, 5581 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5582 __LINE__, 5583 cts, 5584 error); 5585 } 5586 } 5587 5588 static void 5589 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5590 { 5591 /* 5592 * If we have hardware pacing support 5593 * we need to factor that in for our 5594 * TSO size. 5595 */ 5596 const struct tcp_hwrate_limit_table *rlp; 5597 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5598 5599 if ((bbr->bbr_hdrw_pacing == 0) || 5600 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5601 (bbr->r_ctl.crte == NULL)) 5602 return; 5603 if (bbr->hw_pacing_set == 0) { 5604 /* Not yet by the hdwr pacing count delay */ 5605 return; 5606 } 5607 if (bbr_hdwr_pace_adjust == 0) { 5608 /* No adjustment */ 5609 return; 5610 } 5611 rlp = bbr->r_ctl.crte; 5612 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5613 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5614 else 5615 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5616 /* 5617 * So lets first get the 5618 * time we will take between 5619 * TSO sized sends currently without 5620 * hardware help. 5621 */ 5622 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5623 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5624 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5625 hdwr_delay *= rlp->time_between; 5626 if (cur_delay > hdwr_delay) 5627 delta = cur_delay - hdwr_delay; 5628 else 5629 delta = 0; 5630 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5631 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5632 1); 5633 if (delta && 5634 (delta < (max(rlp->time_between, 5635 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5636 /* 5637 * Now lets divide by the pacing 5638 * time between each segment the 5639 * hardware sends rounding up and 5640 * derive a bytes from that. We multiply 5641 * that by bbr_hdwr_pace_adjust to get 5642 * more bang for our buck. 5643 * 5644 * The goal is to have the software pacer 5645 * waiting no more than an additional 5646 * pacing delay if we can (without the 5647 * compensation i.e. x bbr_hdwr_pace_adjust). 5648 */ 5649 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5650 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5651 seg_sz *= bbr_hdwr_pace_adjust; 5652 if (bbr_hdwr_pace_floor && 5653 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5654 /* Currently hardware paces 5655 * out rs_min_seg segments at a time. 5656 * We need to make sure we always send at least 5657 * a full burst of bbr_hdwr_pace_floor down. 5658 */ 5659 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5660 } 5661 seg_sz *= maxseg; 5662 } else if (delta == 0) { 5663 /* 5664 * The highest pacing rate is 5665 * above our b/w gained. This means 5666 * we probably are going quite fast at 5667 * the hardware highest rate. Lets just multiply 5668 * the calculated TSO size by the 5669 * multiplier factor (its probably 5670 * 4 segments in the default config for 5671 * mlx). 5672 */ 5673 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5674 if (bbr_hdwr_pace_floor && 5675 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5676 /* Currently hardware paces 5677 * out rs_min_seg segments at a time. 5678 * We need to make sure we always send at least 5679 * a full burst of bbr_hdwr_pace_floor down. 5680 */ 5681 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5682 } 5683 } else { 5684 /* 5685 * The pacing time difference is so 5686 * big that the hardware will 5687 * pace out more rapidly then we 5688 * really want and then we 5689 * will have a long delay. Lets just keep 5690 * the same TSO size so its as if 5691 * we were not using hdwr pacing (we 5692 * just gain a bit of spacing from the 5693 * hardware if seg_sz > 1). 5694 */ 5695 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5696 } 5697 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5698 new_tso = seg_sz; 5699 else 5700 new_tso = bbr->r_ctl.rc_pace_max_segs; 5701 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5702 new_tso = PACE_MAX_IP_BYTES - maxseg; 5703 5704 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5705 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5706 bbr->r_ctl.rc_pace_max_segs = new_tso; 5707 } 5708 } 5709 5710 static void 5711 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5712 { 5713 uint64_t bw; 5714 uint32_t old_tso = 0, new_tso; 5715 uint32_t maxseg, bytes; 5716 uint32_t tls_seg=0; 5717 /* 5718 * Google/linux uses the following algorithm to determine 5719 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5720 * 5721 * bytes = bw_in_bytes_per_second / 1000 5722 * bytes = min(bytes, 64k) 5723 * tso_segs = bytes / MSS 5724 * if (bw < 1.2Mbs) 5725 * min_tso_segs = 1 5726 * else 5727 * min_tso_segs = 2 5728 * tso_segs = max(tso_segs, min_tso_segs) 5729 * 5730 * * Note apply a device specific limit (we apply this in the 5731 * tcp_m_copym). 5732 * Note that before the initial measurement is made google bursts out 5733 * a full iwnd just like new-reno/cubic. 5734 * 5735 * We do not use this algorithm. Instead we 5736 * use a two phased approach: 5737 * 5738 * if ( bw <= per-tcb-cross-over) 5739 * goal_tso = calculate how much with this bw we 5740 * can send in goal-time seconds. 5741 * if (goal_tso > mss) 5742 * seg = goal_tso / mss 5743 * tso = seg * mss 5744 * else 5745 * tso = mss 5746 * if (tso > per-tcb-max) 5747 * tso = per-tcb-max 5748 * else if ( bw > 512Mbps) 5749 * tso = max-tso (64k/mss) 5750 * else 5751 * goal_tso = bw / per-tcb-divsor 5752 * seg = (goal_tso + mss-1)/mss 5753 * tso = seg * mss 5754 * 5755 * if (tso < per-tcb-floor) 5756 * tso = per-tcb-floor 5757 * if (tso > per-tcb-utter_max) 5758 * tso = per-tcb-utter_max 5759 * 5760 * Note the default per-tcb-divisor is 1000 (same as google). 5761 * the goal cross over is 30Mbps however. To recreate googles 5762 * algorithm you need to set: 5763 * 5764 * cross-over = 23,168,000 bps 5765 * goal-time = 18000 5766 * per-tcb-max = 2 5767 * per-tcb-divisor = 1000 5768 * per-tcb-floor = 1 5769 * 5770 * This will get you "google bbr" behavior with respect to tso size. 5771 * 5772 * Note we do set anything TSO size until we are past the initial 5773 * window. Before that we gnerally use either a single MSS 5774 * or we use the full IW size (so we burst a IW at a time) 5775 */ 5776 5777 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5778 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5779 } else { 5780 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5781 } 5782 old_tso = bbr->r_ctl.rc_pace_max_segs; 5783 if (bbr->rc_past_init_win == 0) { 5784 /* 5785 * Not enough data has been acknowledged to make a 5786 * judgement. Set up the initial TSO based on if we 5787 * are sending a full IW at once or not. 5788 */ 5789 if (bbr->rc_use_google) 5790 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5791 else if (bbr->bbr_init_win_cheat) 5792 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5793 else 5794 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5795 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5796 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5797 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5798 bbr->r_ctl.rc_pace_max_segs = maxseg; 5799 } 5800 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5801 bbr_adjust_for_hw_pacing(bbr, cts); 5802 return; 5803 } 5804 /** 5805 * Now lets set the TSO goal based on our delivery rate in 5806 * bytes per second. Note we only do this if 5807 * we have acked at least the initial cwnd worth of data. 5808 */ 5809 bw = bbr_get_bw(bbr); 5810 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5811 (bbr->rc_use_google == 0)) { 5812 /* We clamp to one MSS in recovery */ 5813 new_tso = maxseg; 5814 } else if (bbr->rc_use_google) { 5815 int min_tso_segs; 5816 5817 /* Google considers the gain too */ 5818 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5819 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5820 bw /= BBR_UNIT; 5821 } 5822 bytes = bw / 1024; 5823 if (bytes > (64 * 1024)) 5824 bytes = 64 * 1024; 5825 new_tso = bytes / maxseg; 5826 if (bw < ONE_POINT_TWO_MEG) 5827 min_tso_segs = 1; 5828 else 5829 min_tso_segs = 2; 5830 if (new_tso < min_tso_segs) 5831 new_tso = min_tso_segs; 5832 new_tso *= maxseg; 5833 } else if (bbr->rc_no_pacing) { 5834 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5835 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5836 /* 5837 * Calculate the worse case b/w TSO if we are inserting no 5838 * more than a delay_target number of TSO's. 5839 */ 5840 uint32_t tso_len, min_tso; 5841 5842 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5843 if (tso_len > maxseg) { 5844 new_tso = tso_len / maxseg; 5845 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5846 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5847 new_tso *= maxseg; 5848 } else { 5849 /* 5850 * less than a full sized frame yikes.. long rtt or 5851 * low bw? 5852 */ 5853 min_tso = bbr_minseg(bbr); 5854 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5855 new_tso = rounddown(tso_len, min_tso); 5856 else 5857 new_tso = min_tso; 5858 } 5859 } else if (bw > FIVETWELVE_MBPS) { 5860 /* 5861 * This guy is so fast b/w wise that we can TSO as large as 5862 * possible of segments that the NIC will allow. 5863 */ 5864 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5865 } else { 5866 /* 5867 * This formula is based on attempting to send a segment or 5868 * more every bbr_hptsi_per_second. The default is 1000 5869 * which means you are targeting what you can send every 1ms 5870 * based on the peers bw. 5871 * 5872 * If the number drops to say 500, then you are looking more 5873 * at 2ms and you will raise how much we send in a single 5874 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5875 * trade off of course is you will send more at once and 5876 * thus tend to clump up the sends into larger "bursts" 5877 * building a queue. 5878 */ 5879 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5880 new_tso = roundup(bw, (uint64_t)maxseg); 5881 /* 5882 * Gate the floor to match what our lower than 48Mbps 5883 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5884 * becomes the floor for this calculation. 5885 */ 5886 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5887 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5888 } 5889 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5890 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5891 if (new_tso > PACE_MAX_IP_BYTES) 5892 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5893 /* Enforce an utter maximum. */ 5894 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5895 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5896 } 5897 if (old_tso != new_tso) { 5898 /* Only log changes */ 5899 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5900 bbr->r_ctl.rc_pace_max_segs = new_tso; 5901 } 5902 /* We have hardware pacing! */ 5903 bbr_adjust_for_hw_pacing(bbr, cts); 5904 } 5905 5906 static void 5907 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5908 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5909 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5910 struct sockbuf *sb) 5911 { 5912 5913 struct bbr_sendmap *rsm, *nrsm; 5914 register uint32_t snd_max, snd_una; 5915 uint32_t pacing_time; 5916 /* 5917 * Add to the RACK log of packets in flight or retransmitted. If 5918 * there is a TS option we will use the TS echoed, if not we will 5919 * grab a TS. 5920 * 5921 * Retransmissions will increment the count and move the ts to its 5922 * proper place. Note that if options do not include TS's then we 5923 * won't be able to effectively use the ACK for an RTT on a retran. 5924 * 5925 * Notes about r_start and r_end. Lets consider a send starting at 5926 * sequence 1 for 10 bytes. In such an example the r_start would be 5927 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5928 * This means that r_end is actually the first sequence for the next 5929 * slot (11). 5930 * 5931 */ 5932 INP_WLOCK_ASSERT(tp->t_inpcb); 5933 if (err) { 5934 /* 5935 * We don't log errors -- we could but snd_max does not 5936 * advance in this case either. 5937 */ 5938 return; 5939 } 5940 if (th_flags & TH_RST) { 5941 /* 5942 * We don't log resets and we return immediately from 5943 * sending 5944 */ 5945 *abandon = 1; 5946 return; 5947 } 5948 snd_una = tp->snd_una; 5949 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5950 /* 5951 * The call to bbr_log_output is made before bumping 5952 * snd_max. This means we can record one extra byte on a SYN 5953 * or FIN if seq_out is adding more on and a FIN is present 5954 * (and we are not resending). 5955 */ 5956 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5957 len++; 5958 if (th_flags & TH_FIN) 5959 len++; 5960 } 5961 if (SEQ_LEQ((seq_out + len), snd_una)) { 5962 /* Are sending an old segment to induce an ack (keep-alive)? */ 5963 return; 5964 } 5965 if (SEQ_LT(seq_out, snd_una)) { 5966 /* huh? should we panic? */ 5967 uint32_t end; 5968 5969 end = seq_out + len; 5970 seq_out = snd_una; 5971 len = end - seq_out; 5972 } 5973 snd_max = tp->snd_max; 5974 if (len == 0) { 5975 /* We don't log zero window probes */ 5976 return; 5977 } 5978 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5979 /* First question is it a retransmission? */ 5980 if (seq_out == snd_max) { 5981 again: 5982 rsm = bbr_alloc(bbr); 5983 if (rsm == NULL) { 5984 return; 5985 } 5986 rsm->r_flags = 0; 5987 if (th_flags & TH_SYN) 5988 rsm->r_flags |= BBR_HAS_SYN; 5989 if (th_flags & TH_FIN) 5990 rsm->r_flags |= BBR_HAS_FIN; 5991 rsm->r_tim_lastsent[0] = cts; 5992 rsm->r_rtr_cnt = 1; 5993 rsm->r_rtr_bytes = 0; 5994 rsm->r_start = seq_out; 5995 rsm->r_end = rsm->r_start + len; 5996 rsm->r_dupack = 0; 5997 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5998 rsm->r_pacing_delay = pacing_time; 5999 rsm->r_ts_valid = bbr->rc_ts_valid; 6000 if (bbr->rc_ts_valid) 6001 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 6002 rsm->r_del_time = bbr->r_ctl.rc_del_time; 6003 if (bbr->r_ctl.r_app_limited_until) 6004 rsm->r_app_limited = 1; 6005 else 6006 rsm->r_app_limited = 0; 6007 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 6008 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 6009 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 6010 /* 6011 * Here we must also add in this rsm since snd_max 6012 * is updated after we return from a new send. 6013 */ 6014 rsm->r_flight_at_send += len; 6015 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 6016 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 6017 rsm->r_in_tmap = 1; 6018 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 6019 rsm->r_bbr_state = bbr_state_val(bbr); 6020 else 6021 rsm->r_bbr_state = 8; 6022 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 6023 rsm->r_is_gain = 1; 6024 rsm->r_is_drain = 0; 6025 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 6026 rsm->r_is_drain = 1; 6027 rsm->r_is_gain = 0; 6028 } else { 6029 rsm->r_is_drain = 0; 6030 rsm->r_is_gain = 0; 6031 } 6032 return; 6033 } 6034 /* 6035 * If we reach here its a retransmission and we need to find it. 6036 */ 6037 more: 6038 if (hintrsm && (hintrsm->r_start == seq_out)) { 6039 rsm = hintrsm; 6040 hintrsm = NULL; 6041 } else if (bbr->r_ctl.rc_next) { 6042 /* We have a hint from a previous run */ 6043 rsm = bbr->r_ctl.rc_next; 6044 } else { 6045 /* No hints sorry */ 6046 rsm = NULL; 6047 } 6048 if ((rsm) && (rsm->r_start == seq_out)) { 6049 /* 6050 * We used rc_next or hintrsm to retransmit, hopefully the 6051 * likely case. 6052 */ 6053 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6054 if (len == 0) { 6055 return; 6056 } else { 6057 goto more; 6058 } 6059 } 6060 /* Ok it was not the last pointer go through it the hard way. */ 6061 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6062 if (rsm->r_start == seq_out) { 6063 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6064 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6065 if (len == 0) { 6066 return; 6067 } else { 6068 continue; 6069 } 6070 } 6071 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6072 /* Transmitted within this piece */ 6073 /* 6074 * Ok we must split off the front and then let the 6075 * update do the rest 6076 */ 6077 nrsm = bbr_alloc_full_limit(bbr); 6078 if (nrsm == NULL) { 6079 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6080 return; 6081 } 6082 /* 6083 * copy rsm to nrsm and then trim the front of rsm 6084 * to not include this part. 6085 */ 6086 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6087 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6088 if (rsm->r_in_tmap) { 6089 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6090 nrsm->r_in_tmap = 1; 6091 } 6092 rsm->r_flags &= (~BBR_HAS_FIN); 6093 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6094 if (len == 0) { 6095 return; 6096 } 6097 } 6098 } 6099 /* 6100 * Hmm not found in map did they retransmit both old and on into the 6101 * new? 6102 */ 6103 if (seq_out == tp->snd_max) { 6104 goto again; 6105 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6106 #ifdef BBR_INVARIANTS 6107 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6108 seq_out, len, tp->snd_una, tp->snd_max); 6109 printf("Starting Dump of all rack entries\n"); 6110 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6111 printf("rsm:%p start:%u end:%u\n", 6112 rsm, rsm->r_start, rsm->r_end); 6113 } 6114 printf("Dump complete\n"); 6115 panic("seq_out not found rack:%p tp:%p", 6116 bbr, tp); 6117 #endif 6118 } else { 6119 #ifdef BBR_INVARIANTS 6120 /* 6121 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6122 * flag) 6123 */ 6124 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6125 seq_out, len, tp->snd_max, tp); 6126 #endif 6127 } 6128 } 6129 6130 static void 6131 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6132 { 6133 /* 6134 * Collapse timeout back the cum-ack moved. 6135 */ 6136 tp->t_rxtshift = 0; 6137 tp->t_softerror = 0; 6138 } 6139 6140 static void 6141 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6142 { 6143 bbr->rtt_valid = 1; 6144 bbr->r_ctl.cur_rtt = rtt_usecs; 6145 bbr->r_ctl.ts_in = tsin; 6146 if (rsm_send_time) 6147 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6148 } 6149 6150 static void 6151 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6152 { 6153 /** 6154 * We have in our bbr control: 6155 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6156 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6157 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6158 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6159 * 6160 * Now we can calculate the time between the sends by doing: 6161 * 6162 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6163 * 6164 * And the peer's time between receiving them by doing: 6165 * 6166 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6167 * 6168 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6169 * We also may find that we can't use the timestamps if say we see 6170 * that the peer_delta indicates that though we may have taken 10ms to 6171 * pace out the data, it only saw 1ms between the two packets. This would 6172 * indicate that somewhere on the path is a batching entity that is giving 6173 * out time-slices of the actual b/w. This would mean we could not use 6174 * reliably the peers timestamps. 6175 * 6176 * We expect delta > peer_delta initially. Until we figure out the 6177 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6178 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6179 * then its 10ms vs our usec. If the peer is running a usec clock we would 6180 * put a 1 there. If the value is faster then ours, we will disable the 6181 * use of timestamps (though we could revist this later if we find it to be not 6182 * just an isolated one or two flows)). 6183 * 6184 * To detect the batching middle boxes we will come up with our compensation and 6185 * if with it in place, we find the peer is drastically off (by some margin) in 6186 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6187 * 6188 */ 6189 uint64_t delta, peer_delta, delta_up; 6190 6191 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6192 if (delta < bbr_min_usec_delta) { 6193 /* 6194 * Have not seen a min amount of time 6195 * between our send times so we can 6196 * make a determination of the timestamp 6197 * yet. 6198 */ 6199 return; 6200 } 6201 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6202 if (peer_delta < bbr_min_peer_delta) { 6203 /* 6204 * We may have enough in the form of 6205 * our delta but the peers number 6206 * has not changed that much. It could 6207 * be its clock ratio is such that 6208 * we need more data (10ms tick) or 6209 * there may be other compression scenarios 6210 * going on. In any event we need the 6211 * spread to be larger. 6212 */ 6213 return; 6214 } 6215 /* Ok lets first see which way our delta is going */ 6216 if (peer_delta > delta) { 6217 /* Very unlikely, the peer without 6218 * compensation shows that it saw 6219 * the two sends arrive further apart 6220 * then we saw then in micro-seconds. 6221 */ 6222 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6223 /* well it looks like the peer is a micro-second clock. */ 6224 bbr->rc_ts_clock_set = 1; 6225 bbr->r_ctl.bbr_peer_tsratio = 1; 6226 } else { 6227 bbr->rc_ts_cant_be_used = 1; 6228 bbr->rc_ts_clock_set = 1; 6229 } 6230 return; 6231 } 6232 /* Ok we know that the peer_delta is smaller than our send distance */ 6233 bbr->rc_ts_clock_set = 1; 6234 /* First question is it within the percentage that they are using usec time? */ 6235 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6236 if ((peer_delta + delta_up) >= delta) { 6237 /* Its a usec clock */ 6238 bbr->r_ctl.bbr_peer_tsratio = 1; 6239 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6240 return; 6241 } 6242 /* Ok if not usec, what about 10usec (though unlikely)? */ 6243 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6244 if (((peer_delta * 10) + delta_up) >= delta) { 6245 bbr->r_ctl.bbr_peer_tsratio = 10; 6246 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6247 return; 6248 } 6249 /* And what about 100usec (though again unlikely)? */ 6250 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6251 if (((peer_delta * 100) + delta_up) >= delta) { 6252 bbr->r_ctl.bbr_peer_tsratio = 100; 6253 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6254 return; 6255 } 6256 /* And how about 1 msec (the most likely one)? */ 6257 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6258 if (((peer_delta * 1000) + delta_up) >= delta) { 6259 bbr->r_ctl.bbr_peer_tsratio = 1000; 6260 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6261 return; 6262 } 6263 /* Ok if not msec could it be 10 msec? */ 6264 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6265 if (((peer_delta * 10000) + delta_up) >= delta) { 6266 bbr->r_ctl.bbr_peer_tsratio = 10000; 6267 return; 6268 } 6269 /* If we fall down here the clock tick so slowly we can't use it */ 6270 bbr->rc_ts_cant_be_used = 1; 6271 bbr->r_ctl.bbr_peer_tsratio = 0; 6272 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6273 } 6274 6275 /* 6276 * Collect new round-trip time estimate 6277 * and update averages and current timeout. 6278 */ 6279 static void 6280 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6281 { 6282 int32_t delta; 6283 uint32_t rtt, tsin; 6284 int32_t rtt_ticks; 6285 6286 if (bbr->rtt_valid == 0) 6287 /* No valid sample */ 6288 return; 6289 6290 rtt = bbr->r_ctl.cur_rtt; 6291 tsin = bbr->r_ctl.ts_in; 6292 if (bbr->rc_prtt_set_ts) { 6293 /* 6294 * We are to force feed the rttProp filter due 6295 * to an entry into PROBE_RTT. This assures 6296 * that the times are sync'd between when we 6297 * go into PROBE_RTT and the filter expiration. 6298 * 6299 * Google does not use a true filter, so they do 6300 * this implicitly since they only keep one value 6301 * and when they enter probe-rtt they update the 6302 * value to the newest rtt. 6303 */ 6304 uint32_t rtt_prop; 6305 6306 bbr->rc_prtt_set_ts = 0; 6307 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6308 if (rtt > rtt_prop) 6309 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6310 else 6311 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6312 } 6313 if (bbr->rc_ack_was_delayed) 6314 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6315 6316 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6317 bbr->r_ctl.rc_lowest_rtt = rtt; 6318 bbr_log_rtt_sample(bbr, rtt, tsin); 6319 if (bbr->r_init_rtt) { 6320 /* 6321 * The initial rtt is not-trusted, nuke it and lets get 6322 * our first valid measurement in. 6323 */ 6324 bbr->r_init_rtt = 0; 6325 tp->t_srtt = 0; 6326 } 6327 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6328 /* 6329 * So we have not yet figured out 6330 * what the peers TSTMP value is 6331 * in (most likely ms). We need a 6332 * series of cum-ack's to determine 6333 * this reliably. 6334 */ 6335 if (bbr->rc_ack_is_cumack) { 6336 if (bbr->rc_ts_data_set) { 6337 /* Lets attempt to determine the timestamp granularity. */ 6338 bbr_make_timestamp_determination(bbr); 6339 } else { 6340 bbr->rc_ts_data_set = 1; 6341 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6342 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6343 } 6344 } else { 6345 /* 6346 * We have to have consecutive acks 6347 * reset any "filled" state to none. 6348 */ 6349 bbr->rc_ts_data_set = 0; 6350 } 6351 } 6352 /* Round it up */ 6353 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6354 if (rtt_ticks == 0) 6355 rtt_ticks = 1; 6356 if (tp->t_srtt != 0) { 6357 /* 6358 * srtt is stored as fixed point with 5 bits after the 6359 * binary point (i.e., scaled by 8). The following magic is 6360 * equivalent to the smoothing algorithm in rfc793 with an 6361 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6362 * Adjust rtt to origin 0. 6363 */ 6364 6365 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6366 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6367 6368 tp->t_srtt += delta; 6369 if (tp->t_srtt <= 0) 6370 tp->t_srtt = 1; 6371 6372 /* 6373 * We accumulate a smoothed rtt variance (actually, a 6374 * smoothed mean difference), then set the retransmit timer 6375 * to smoothed rtt + 4 times the smoothed variance. rttvar 6376 * is stored as fixed point with 4 bits after the binary 6377 * point (scaled by 16). The following is equivalent to 6378 * rfc793 smoothing with an alpha of .75 (rttvar = 6379 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6380 * wired-in beta. 6381 */ 6382 if (delta < 0) 6383 delta = -delta; 6384 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6385 tp->t_rttvar += delta; 6386 if (tp->t_rttvar <= 0) 6387 tp->t_rttvar = 1; 6388 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 6389 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 6390 } else { 6391 /* 6392 * No rtt measurement yet - use the unsmoothed rtt. Set the 6393 * variance to half the rtt (so our first retransmit happens 6394 * at 3*rtt). 6395 */ 6396 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6397 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6398 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 6399 } 6400 KMOD_TCPSTAT_INC(tcps_rttupdated); 6401 tp->t_rttupdated++; 6402 #ifdef STATS 6403 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6404 #endif 6405 /* 6406 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6407 * way we do the smoothing, srtt and rttvar will each average +1/2 6408 * tick of bias. When we compute the retransmit timer, we want 1/2 6409 * tick of rounding and 1 extra tick because of +-1/2 tick 6410 * uncertainty in the firing of the timer. The bias will give us 6411 * exactly the 1.5 tick we need. But, because the bias is 6412 * statistical, we have to test that we don't drop below the minimum 6413 * feasible timer (which is 2 ticks). 6414 */ 6415 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6416 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6417 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6418 6419 /* 6420 * We received an ack for a packet that wasn't retransmitted; it is 6421 * probably safe to discard any error indications we've received 6422 * recently. This isn't quite right, but close enough for now (a 6423 * route might have failed after we sent a segment, and the return 6424 * path might not be symmetrical). 6425 */ 6426 tp->t_softerror = 0; 6427 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6428 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6429 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6430 } 6431 6432 static void 6433 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6434 { 6435 bbr->r_ctl.rc_rtt_shrinks = cts; 6436 if (bbr_can_force_probertt && 6437 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6438 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6439 /* 6440 * We should enter probe-rtt its been too long 6441 * since we have been there. 6442 */ 6443 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6444 } else 6445 bbr_check_probe_rtt_limits(bbr, cts); 6446 } 6447 6448 static void 6449 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6450 { 6451 uint64_t orig_bw; 6452 6453 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6454 /* We never apply a zero measurement */ 6455 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6456 0, 0, 0, 0, 0, 0); 6457 return; 6458 } 6459 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6460 bbr->r_ctl.r_measurement_count++; 6461 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6462 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6463 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6464 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6465 0, 0, 0, 0, 0, 0); 6466 if (orig_bw && 6467 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6468 if (bbr->bbr_hdrw_pacing) { 6469 /* 6470 * Apply a new rate to the hardware 6471 * possibly. 6472 */ 6473 bbr_update_hardware_pacing_rate(bbr, cts); 6474 } 6475 bbr_set_state_target(bbr, __LINE__); 6476 tcp_bbr_tso_size_check(bbr, cts); 6477 if (bbr->r_recovery_bw) { 6478 bbr_setup_red_bw(bbr, cts); 6479 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6480 } 6481 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6482 tcp_bbr_tso_size_check(bbr, cts); 6483 } 6484 6485 static void 6486 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6487 { 6488 if (bbr->rc_in_persist == 0) { 6489 /* We log only when not in persist */ 6490 /* Translate to a Bytes Per Second */ 6491 uint64_t tim, bw, ts_diff, ts_bw; 6492 uint32_t delivered; 6493 6494 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6495 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6496 else 6497 tim = 1; 6498 /* 6499 * Now that we have processed the tim (skipping the sample 6500 * or possibly updating the time, go ahead and 6501 * calculate the cdr. 6502 */ 6503 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6504 bw = (uint64_t)delivered; 6505 bw *= (uint64_t)USECS_IN_SECOND; 6506 bw /= tim; 6507 if (bw == 0) { 6508 /* We must have a calculatable amount */ 6509 return; 6510 } 6511 /* 6512 * If we are using this b/w shove it in now so we 6513 * can see in the trace viewer if it gets over-ridden. 6514 */ 6515 if (rsm->r_ts_valid && 6516 bbr->rc_ts_valid && 6517 bbr->rc_ts_clock_set && 6518 (bbr->rc_ts_cant_be_used == 0) && 6519 bbr->rc_use_ts_limit) { 6520 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6521 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6522 if ((delivered == 0) || 6523 (rtt < 1000)) { 6524 /* Can't use the ts */ 6525 bbr_log_type_bbrupd(bbr, 61, cts, 6526 ts_diff, 6527 bbr->r_ctl.last_inbound_ts, 6528 rsm->r_del_ack_ts, 0, 6529 0, 0, 0, delivered); 6530 } else { 6531 ts_bw = (uint64_t)delivered; 6532 ts_bw *= (uint64_t)USECS_IN_SECOND; 6533 ts_bw /= ts_diff; 6534 bbr_log_type_bbrupd(bbr, 62, cts, 6535 (ts_bw >> 32), 6536 (ts_bw & 0xffffffff), 0, 0, 6537 0, 0, ts_diff, delivered); 6538 if ((bbr->ts_can_raise) && 6539 (ts_bw > bw)) { 6540 bbr_log_type_bbrupd(bbr, 8, cts, 6541 delivered, 6542 ts_diff, 6543 (bw >> 32), 6544 (bw & 0x00000000ffffffff), 6545 0, 0, 0, 0); 6546 bw = ts_bw; 6547 } else if (ts_bw && (ts_bw < bw)) { 6548 bbr_log_type_bbrupd(bbr, 7, cts, 6549 delivered, 6550 ts_diff, 6551 (bw >> 32), 6552 (bw & 0x00000000ffffffff), 6553 0, 0, 0, 0); 6554 bw = ts_bw; 6555 } 6556 } 6557 } 6558 if (rsm->r_first_sent_time && 6559 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6560 uint64_t sbw, sti; 6561 /* 6562 * We use what was in flight at the time of our 6563 * send and the size of this send to figure 6564 * out what we have been sending at (amount). 6565 * For the time we take from the time of 6566 * the send of the first send outstanding 6567 * until this send plus this sends pacing 6568 * time. This gives us a good calculation 6569 * as to the rate we have been sending at. 6570 */ 6571 6572 sbw = (uint64_t)(rsm->r_flight_at_send); 6573 sbw *= (uint64_t)USECS_IN_SECOND; 6574 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6575 sti += rsm->r_pacing_delay; 6576 sbw /= sti; 6577 if (sbw < bw) { 6578 bbr_log_type_bbrupd(bbr, 6, cts, 6579 delivered, 6580 (uint32_t)sti, 6581 (bw >> 32), 6582 (uint32_t)bw, 6583 rsm->r_first_sent_time, 0, (sbw >> 32), 6584 (uint32_t)sbw); 6585 bw = sbw; 6586 } 6587 } 6588 /* Use the google algorithm for b/w measurements */ 6589 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6590 if ((rsm->r_app_limited == 0) || 6591 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6592 tcp_bbr_commit_bw(bbr, cts); 6593 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6594 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6595 } 6596 } 6597 } 6598 6599 static void 6600 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6601 { 6602 if (bbr->rc_in_persist == 0) { 6603 /* We log only when not in persist */ 6604 /* Translate to a Bytes Per Second */ 6605 uint64_t tim, bw; 6606 uint32_t delivered; 6607 int no_apply = 0; 6608 6609 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6610 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6611 else 6612 tim = 1; 6613 /* 6614 * Now that we have processed the tim (skipping the sample 6615 * or possibly updating the time, go ahead and 6616 * calculate the cdr. 6617 */ 6618 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6619 bw = (uint64_t)delivered; 6620 bw *= (uint64_t)USECS_IN_SECOND; 6621 bw /= tim; 6622 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6623 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6624 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6625 6626 no_apply = 1; 6627 } 6628 /* 6629 * If we are using this b/w shove it in now so we 6630 * can see in the trace viewer if it gets over-ridden. 6631 */ 6632 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6633 /* Gate by the sending rate */ 6634 if (rsm->r_first_sent_time && 6635 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6636 uint64_t sbw, sti; 6637 /* 6638 * We use what was in flight at the time of our 6639 * send and the size of this send to figure 6640 * out what we have been sending at (amount). 6641 * For the time we take from the time of 6642 * the send of the first send outstanding 6643 * until this send plus this sends pacing 6644 * time. This gives us a good calculation 6645 * as to the rate we have been sending at. 6646 */ 6647 6648 sbw = (uint64_t)(rsm->r_flight_at_send); 6649 sbw *= (uint64_t)USECS_IN_SECOND; 6650 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6651 sti += rsm->r_pacing_delay; 6652 sbw /= sti; 6653 if (sbw < bw) { 6654 bbr_log_type_bbrupd(bbr, 6, cts, 6655 delivered, 6656 (uint32_t)sti, 6657 (bw >> 32), 6658 (uint32_t)bw, 6659 rsm->r_first_sent_time, 0, (sbw >> 32), 6660 (uint32_t)sbw); 6661 bw = sbw; 6662 } 6663 if ((sti > tim) && 6664 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6665 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6666 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6667 no_apply = 1; 6668 } else 6669 no_apply = 0; 6670 } 6671 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6672 if ((no_apply == 0) && 6673 ((rsm->r_app_limited == 0) || 6674 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6675 tcp_bbr_commit_bw(bbr, cts); 6676 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6677 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6678 } 6679 } 6680 } 6681 6682 static void 6683 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6684 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6685 { 6686 uint64_t old_rttprop; 6687 6688 /* Update our delivery time and amount */ 6689 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6690 bbr->r_ctl.rc_del_time = cts; 6691 if (rtt == 0) { 6692 /* 6693 * 0 means its a retransmit, for now we don't use these for 6694 * the rest of BBR. 6695 */ 6696 return; 6697 } 6698 if ((bbr->rc_use_google == 0) && 6699 (match != BBR_RTT_BY_EXACTMATCH) && 6700 (match != BBR_RTT_BY_TIMESTAMP)){ 6701 /* 6702 * We get a lot of rtt updates, lets not pay attention to 6703 * any that are not an exact match. That way we don't have 6704 * to worry about timestamps and the whole nonsense of 6705 * unsure if its a retransmission etc (if we ever had the 6706 * timestamp fixed to always have the last thing sent this 6707 * would not be a issue). 6708 */ 6709 return; 6710 } 6711 if ((bbr_no_retran && bbr->rc_use_google) && 6712 (match != BBR_RTT_BY_EXACTMATCH) && 6713 (match != BBR_RTT_BY_TIMESTAMP)){ 6714 /* 6715 * We only do measurements in google mode 6716 * with bbr_no_retran on for sure things. 6717 */ 6718 return; 6719 } 6720 /* Only update srtt if we know by exact match */ 6721 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6722 if (ack_type == BBR_CUM_ACKED) 6723 bbr->rc_ack_is_cumack = 1; 6724 else 6725 bbr->rc_ack_is_cumack = 0; 6726 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6727 /* 6728 * Note the following code differs to the original 6729 * BBR spec. It calls for <= not <. However after a 6730 * long discussion in email with Neal, he acknowledged 6731 * that it should be < than so that we will have flows 6732 * going into probe-rtt (we were seeing cases where that 6733 * did not happen and caused ugly things to occur). We 6734 * have added this agreed upon fix to our code base. 6735 */ 6736 if (rtt < old_rttprop) { 6737 /* Update when we last saw a rtt drop */ 6738 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6739 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6740 } 6741 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6742 match, rsm->r_start, rsm->r_flags); 6743 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6744 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6745 /* 6746 * The RTT-prop moved, reset the target (may be a 6747 * nop for some states). 6748 */ 6749 bbr_set_state_target(bbr, __LINE__); 6750 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6751 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6752 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6753 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6754 /* It went up */ 6755 bbr_check_probe_rtt_limits(bbr, cts); 6756 } 6757 if ((bbr->rc_use_google == 0) && 6758 (match == BBR_RTT_BY_TIMESTAMP)) { 6759 /* 6760 * We don't do b/w update with 6761 * these since they are not really 6762 * reliable. 6763 */ 6764 return; 6765 } 6766 if (bbr->r_ctl.r_app_limited_until && 6767 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6768 /* We are no longer app-limited */ 6769 bbr->r_ctl.r_app_limited_until = 0; 6770 } 6771 if (bbr->rc_use_google) { 6772 bbr_google_measurement(bbr, rsm, rtt, cts); 6773 } else { 6774 bbr_nf_measurement(bbr, rsm, rtt, cts); 6775 } 6776 } 6777 6778 /* 6779 * Convert a timestamp that the main stack 6780 * uses (milliseconds) into one that bbr uses 6781 * (microseconds). Return that converted timestamp. 6782 */ 6783 static uint32_t 6784 bbr_ts_convert(uint32_t cts) { 6785 uint32_t sec, msec; 6786 6787 sec = cts / MS_IN_USEC; 6788 msec = cts - (MS_IN_USEC * sec); 6789 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6790 } 6791 6792 /* 6793 * Return 0 if we did not update the RTT time, return 6794 * 1 if we did. 6795 */ 6796 static int 6797 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6798 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6799 { 6800 int32_t i; 6801 uint32_t t, uts = 0; 6802 6803 if ((rsm->r_flags & BBR_ACKED) || 6804 (rsm->r_flags & BBR_WAS_RENEGED) || 6805 (rsm->r_flags & BBR_RXT_CLEARED)) { 6806 /* Already done */ 6807 return (0); 6808 } 6809 if (rsm->r_rtt_not_allowed) { 6810 /* Not allowed */ 6811 return (0); 6812 } 6813 if (rsm->r_rtr_cnt == 1) { 6814 /* 6815 * Only one transmit. Hopefully the normal case. 6816 */ 6817 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6818 t = cts - rsm->r_tim_lastsent[0]; 6819 else 6820 t = 1; 6821 if ((int)t <= 0) 6822 t = 1; 6823 bbr->r_ctl.rc_last_rtt = t; 6824 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6825 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6826 return (1); 6827 } 6828 /* Convert to usecs */ 6829 if ((bbr_can_use_ts_for_rtt == 1) && 6830 (bbr->rc_use_google == 1) && 6831 (ack_type == BBR_CUM_ACKED) && 6832 (to->to_flags & TOF_TS) && 6833 (to->to_tsecr != 0)) { 6834 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6835 if (t < 1) 6836 t = 1; 6837 t *= MS_IN_USEC; 6838 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6839 BBR_RTT_BY_TIMESTAMP, 6840 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6841 ack_type, to); 6842 return (1); 6843 } 6844 uts = bbr_ts_convert(to->to_tsecr); 6845 if ((to->to_flags & TOF_TS) && 6846 (to->to_tsecr != 0) && 6847 (ack_type == BBR_CUM_ACKED) && 6848 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6849 /* 6850 * Now which timestamp does it match? In this block the ACK 6851 * may be coming from a previous transmission. 6852 */ 6853 uint32_t fudge; 6854 6855 fudge = BBR_TIMER_FUDGE; 6856 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6857 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6858 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6859 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6860 t = cts - rsm->r_tim_lastsent[i]; 6861 else 6862 t = 1; 6863 if ((int)t <= 0) 6864 t = 1; 6865 bbr->r_ctl.rc_last_rtt = t; 6866 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6867 rsm->r_tim_lastsent[i], ack_type, to); 6868 if ((i + 1) < rsm->r_rtr_cnt) { 6869 /* Likely */ 6870 return (0); 6871 } else if (rsm->r_flags & BBR_TLP) { 6872 bbr->rc_tlp_rtx_out = 0; 6873 } 6874 return (1); 6875 } 6876 } 6877 /* Fall through if we can't find a matching timestamp */ 6878 } 6879 /* 6880 * Ok its a SACK block that we retransmitted. or a windows 6881 * machine without timestamps. We can tell nothing from the 6882 * time-stamp since its not there or the time the peer last 6883 * recieved a segment that moved forward its cum-ack point. 6884 * 6885 * Lets look at the last retransmit and see what we can tell 6886 * (with BBR for space we only keep 2 note we have to keep 6887 * at least 2 so the map can not be condensed more). 6888 */ 6889 i = rsm->r_rtr_cnt - 1; 6890 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6891 t = cts - rsm->r_tim_lastsent[i]; 6892 else 6893 goto not_sure; 6894 if (t < bbr->r_ctl.rc_lowest_rtt) { 6895 /* 6896 * We retransmitted and the ack came back in less 6897 * than the smallest rtt we have observed in the 6898 * windowed rtt. We most likey did an improper 6899 * retransmit as outlined in 4.2 Step 3 point 2 in 6900 * the rack-draft. 6901 * 6902 * Use the prior transmission to update all the 6903 * information as long as there is only one prior 6904 * transmission. 6905 */ 6906 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6907 #ifdef BBR_INVARIANTS 6908 if (rsm->r_rtr_cnt == 1) 6909 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6910 #endif 6911 i = rsm->r_rtr_cnt - 2; 6912 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6913 t = cts - rsm->r_tim_lastsent[i]; 6914 else 6915 t = 1; 6916 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6917 rsm->r_tim_lastsent[i], ack_type, to); 6918 return (0); 6919 } else { 6920 /* 6921 * Too many prior transmissions, just 6922 * updated BBR delivered 6923 */ 6924 not_sure: 6925 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6926 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6927 } 6928 } else { 6929 /* 6930 * We retransmitted it and the retransmit did the 6931 * job. 6932 */ 6933 if (rsm->r_flags & BBR_TLP) 6934 bbr->rc_tlp_rtx_out = 0; 6935 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6936 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6937 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6938 else 6939 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6940 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6941 return (1); 6942 } 6943 return (0); 6944 } 6945 6946 /* 6947 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6948 */ 6949 static void 6950 bbr_log_sack_passed(struct tcpcb *tp, 6951 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6952 { 6953 struct bbr_sendmap *nrsm; 6954 6955 nrsm = rsm; 6956 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6957 bbr_head, r_tnext) { 6958 if (nrsm == rsm) { 6959 /* Skip orginal segment he is acked */ 6960 continue; 6961 } 6962 if (nrsm->r_flags & BBR_ACKED) { 6963 /* Skip ack'd segments */ 6964 continue; 6965 } 6966 if (nrsm->r_flags & BBR_SACK_PASSED) { 6967 /* 6968 * We found one that is already marked 6969 * passed, we have been here before and 6970 * so all others below this are marked. 6971 */ 6972 break; 6973 } 6974 BBR_STAT_INC(bbr_sack_passed); 6975 nrsm->r_flags |= BBR_SACK_PASSED; 6976 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6977 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6978 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6979 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6980 nrsm->r_flags |= BBR_MARKED_LOST; 6981 } 6982 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6983 } 6984 } 6985 6986 /* 6987 * Returns the number of bytes that were 6988 * newly ack'd by sack blocks. 6989 */ 6990 static uint32_t 6991 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6992 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6993 { 6994 int32_t times = 0; 6995 uint32_t start, end, changed = 0; 6996 struct bbr_sendmap *rsm, *nrsm; 6997 int32_t used_ref = 1; 6998 uint8_t went_back = 0, went_fwd = 0; 6999 7000 start = sack->start; 7001 end = sack->end; 7002 rsm = *prsm; 7003 if (rsm == NULL) 7004 used_ref = 0; 7005 7006 /* Do we locate the block behind where we last were? */ 7007 if (rsm && SEQ_LT(start, rsm->r_start)) { 7008 went_back = 1; 7009 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 7010 if (SEQ_GEQ(start, rsm->r_start) && 7011 SEQ_LT(start, rsm->r_end)) { 7012 goto do_rest_ofb; 7013 } 7014 } 7015 } 7016 start_at_beginning: 7017 went_fwd = 1; 7018 /* 7019 * Ok lets locate the block where this guy is fwd from rsm (if its 7020 * set) 7021 */ 7022 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 7023 if (SEQ_GEQ(start, rsm->r_start) && 7024 SEQ_LT(start, rsm->r_end)) { 7025 break; 7026 } 7027 } 7028 do_rest_ofb: 7029 if (rsm == NULL) { 7030 /* 7031 * This happens when we get duplicate sack blocks with the 7032 * same end. For example SACK 4: 100 SACK 3: 100 The sort 7033 * will not change there location so we would just start at 7034 * the end of the first one and get lost. 7035 */ 7036 if (tp->t_flags & TF_SENTFIN) { 7037 /* 7038 * Check to see if we have not logged the FIN that 7039 * went out. 7040 */ 7041 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7042 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7043 /* 7044 * Ok we did not get the FIN logged. 7045 */ 7046 nrsm->r_end++; 7047 rsm = nrsm; 7048 goto do_rest_ofb; 7049 } 7050 } 7051 if (times == 1) { 7052 #ifdef BBR_INVARIANTS 7053 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7054 tp, bbr, sack, to, prsm); 7055 #else 7056 goto out; 7057 #endif 7058 } 7059 times++; 7060 BBR_STAT_INC(bbr_sack_proc_restart); 7061 rsm = NULL; 7062 goto start_at_beginning; 7063 } 7064 /* Ok we have an ACK for some piece of rsm */ 7065 if (rsm->r_start != start) { 7066 /* 7067 * Need to split this in two pieces the before and after. 7068 */ 7069 if (bbr_sack_mergable(rsm, start, end)) 7070 nrsm = bbr_alloc_full_limit(bbr); 7071 else 7072 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7073 if (nrsm == NULL) { 7074 /* We could not allocate ignore the sack */ 7075 struct sackblk blk; 7076 7077 blk.start = start; 7078 blk.end = end; 7079 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7080 goto out; 7081 } 7082 bbr_clone_rsm(bbr, nrsm, rsm, start); 7083 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7084 if (rsm->r_in_tmap) { 7085 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7086 nrsm->r_in_tmap = 1; 7087 } 7088 rsm->r_flags &= (~BBR_HAS_FIN); 7089 rsm = nrsm; 7090 } 7091 if (SEQ_GEQ(end, rsm->r_end)) { 7092 /* 7093 * The end of this block is either beyond this guy or right 7094 * at this guy. 7095 */ 7096 if ((rsm->r_flags & BBR_ACKED) == 0) { 7097 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7098 changed += (rsm->r_end - rsm->r_start); 7099 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7100 bbr_log_sack_passed(tp, bbr, rsm); 7101 if (rsm->r_flags & BBR_MARKED_LOST) { 7102 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7103 } 7104 /* Is Reordering occuring? */ 7105 if (rsm->r_flags & BBR_SACK_PASSED) { 7106 BBR_STAT_INC(bbr_reorder_seen); 7107 bbr->r_ctl.rc_reorder_ts = cts; 7108 if (rsm->r_flags & BBR_MARKED_LOST) { 7109 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7110 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7111 /* LT sampling also needs adjustment */ 7112 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7113 } 7114 } 7115 rsm->r_flags |= BBR_ACKED; 7116 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7117 if (rsm->r_in_tmap) { 7118 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7119 rsm->r_in_tmap = 0; 7120 } 7121 } 7122 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7123 if (end == rsm->r_end) { 7124 /* This block only - done */ 7125 goto out; 7126 } 7127 /* There is more not coverend by this rsm move on */ 7128 start = rsm->r_end; 7129 nrsm = TAILQ_NEXT(rsm, r_next); 7130 rsm = nrsm; 7131 times = 0; 7132 goto do_rest_ofb; 7133 } 7134 if (rsm->r_flags & BBR_ACKED) { 7135 /* Been here done that */ 7136 goto out; 7137 } 7138 /* Ok we need to split off this one at the tail */ 7139 if (bbr_sack_mergable(rsm, start, end)) 7140 nrsm = bbr_alloc_full_limit(bbr); 7141 else 7142 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7143 if (nrsm == NULL) { 7144 /* failed XXXrrs what can we do but loose the sack info? */ 7145 struct sackblk blk; 7146 7147 blk.start = start; 7148 blk.end = end; 7149 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7150 goto out; 7151 } 7152 /* Clone it */ 7153 bbr_clone_rsm(bbr, nrsm, rsm, end); 7154 /* The sack block does not cover this guy fully */ 7155 rsm->r_flags &= (~BBR_HAS_FIN); 7156 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7157 if (rsm->r_in_tmap) { 7158 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7159 nrsm->r_in_tmap = 1; 7160 } 7161 nrsm->r_dupack = 0; 7162 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7163 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7164 changed += (rsm->r_end - rsm->r_start); 7165 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7166 bbr_log_sack_passed(tp, bbr, rsm); 7167 /* Is Reordering occuring? */ 7168 if (rsm->r_flags & BBR_MARKED_LOST) { 7169 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7170 } 7171 if (rsm->r_flags & BBR_SACK_PASSED) { 7172 BBR_STAT_INC(bbr_reorder_seen); 7173 bbr->r_ctl.rc_reorder_ts = cts; 7174 if (rsm->r_flags & BBR_MARKED_LOST) { 7175 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7176 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7177 /* LT sampling also needs adjustment */ 7178 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7179 } 7180 } 7181 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7182 rsm->r_flags |= BBR_ACKED; 7183 if (rsm->r_in_tmap) { 7184 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7185 rsm->r_in_tmap = 0; 7186 } 7187 out: 7188 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7189 /* 7190 * Now can we merge this newly acked 7191 * block with either the previous or 7192 * next block? 7193 */ 7194 nrsm = TAILQ_NEXT(rsm, r_next); 7195 if (nrsm && 7196 (nrsm->r_flags & BBR_ACKED)) { 7197 /* yep this and next can be merged */ 7198 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7199 } 7200 /* Now what about the previous? */ 7201 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7202 if (nrsm && 7203 (nrsm->r_flags & BBR_ACKED)) { 7204 /* yep the previous and this can be merged */ 7205 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7206 } 7207 } 7208 if (used_ref == 0) { 7209 BBR_STAT_INC(bbr_sack_proc_all); 7210 } else { 7211 BBR_STAT_INC(bbr_sack_proc_short); 7212 } 7213 if (went_fwd && went_back) { 7214 BBR_STAT_INC(bbr_sack_search_both); 7215 } else if (went_fwd) { 7216 BBR_STAT_INC(bbr_sack_search_fwd); 7217 } else if (went_back) { 7218 BBR_STAT_INC(bbr_sack_search_back); 7219 } 7220 /* Save off where the next seq is */ 7221 if (rsm) 7222 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7223 else 7224 bbr->r_ctl.rc_sacklast = NULL; 7225 *prsm = rsm; 7226 return (changed); 7227 } 7228 7229 static void inline 7230 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7231 { 7232 struct bbr_sendmap *tmap; 7233 7234 BBR_STAT_INC(bbr_reneges_seen); 7235 tmap = NULL; 7236 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7237 /* Its no longer sacked, mark it so */ 7238 uint32_t oflags; 7239 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7240 #ifdef BBR_INVARIANTS 7241 if (rsm->r_in_tmap) { 7242 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7243 bbr, rsm, rsm->r_flags); 7244 } 7245 #endif 7246 oflags = rsm->r_flags; 7247 if (rsm->r_flags & BBR_MARKED_LOST) { 7248 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7249 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7250 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7251 /* LT sampling also needs adjustment */ 7252 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7253 } 7254 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7255 rsm->r_flags |= BBR_WAS_RENEGED; 7256 rsm->r_flags |= BBR_RXT_CLEARED; 7257 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7258 /* Rebuild it into our tmap */ 7259 if (tmap == NULL) { 7260 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7261 tmap = rsm; 7262 } else { 7263 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7264 tmap = rsm; 7265 } 7266 tmap->r_in_tmap = 1; 7267 /* 7268 * XXXrrs Delivered? Should we do anything here? 7269 * 7270 * Of course we don't on a rxt timeout so maybe its ok that 7271 * we don't? 7272 * 7273 * For now lets not. 7274 */ 7275 rsm = TAILQ_NEXT(rsm, r_next); 7276 } 7277 /* 7278 * Now lets possibly clear the sack filter so we start recognizing 7279 * sacks that cover this area. 7280 */ 7281 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7282 } 7283 7284 static void 7285 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7286 { 7287 struct tcp_bbr *bbr; 7288 struct bbr_sendmap *rsm; 7289 uint32_t cts; 7290 7291 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7292 cts = bbr->r_ctl.rc_rcvtime; 7293 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7294 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7295 if ((rsm->r_end - rsm->r_start) <= 1) { 7296 /* Log out the SYN completely */ 7297 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7298 rsm->r_rtr_bytes = 0; 7299 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7300 if (rsm->r_in_tmap) { 7301 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7302 rsm->r_in_tmap = 0; 7303 } 7304 if (bbr->r_ctl.rc_next == rsm) { 7305 /* scoot along the marker */ 7306 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7307 } 7308 if (to != NULL) 7309 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7310 bbr_free(bbr, rsm); 7311 } else { 7312 /* There is more (Fast open)? strip out SYN. */ 7313 rsm->r_flags &= ~BBR_HAS_SYN; 7314 rsm->r_start++; 7315 } 7316 } 7317 } 7318 7319 /* 7320 * Returns the number of bytes that were 7321 * acknowledged by SACK blocks. 7322 */ 7323 7324 static uint32_t 7325 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7326 uint32_t *prev_acked) 7327 { 7328 uint32_t changed, last_seq, entered_recovery = 0; 7329 struct tcp_bbr *bbr; 7330 struct bbr_sendmap *rsm; 7331 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7332 register uint32_t th_ack; 7333 int32_t i, j, k, new_sb, num_sack_blks = 0; 7334 uint32_t cts, acked, ack_point, sack_changed = 0; 7335 uint32_t p_maxseg, maxseg, p_acked = 0; 7336 7337 INP_WLOCK_ASSERT(tp->t_inpcb); 7338 if (tcp_get_flags(th) & TH_RST) { 7339 /* We don't log resets */ 7340 return (0); 7341 } 7342 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7343 cts = bbr->r_ctl.rc_rcvtime; 7344 7345 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7346 changed = 0; 7347 maxseg = tp->t_maxseg - bbr->rc_last_options; 7348 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7349 th_ack = th->th_ack; 7350 if (SEQ_GT(th_ack, tp->snd_una)) { 7351 acked = th_ack - tp->snd_una; 7352 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7353 bbr->rc_tp->t_acktime = ticks; 7354 } else 7355 acked = 0; 7356 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7357 /* Only sent here for sack processing */ 7358 goto proc_sack; 7359 } 7360 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7361 changed = th_ack - rsm->r_start; 7362 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7363 /* 7364 * For the SYN incoming case we will not have called 7365 * tcp_output for the sending of the SYN, so there will be 7366 * no map. All other cases should probably be a panic. 7367 */ 7368 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7369 /* 7370 * We have a timestamp that can be used to generate 7371 * an initial RTT. 7372 */ 7373 uint32_t ts, now, rtt; 7374 7375 ts = bbr_ts_convert(to->to_tsecr); 7376 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7377 rtt = now - ts; 7378 if (rtt < 1) 7379 rtt = 1; 7380 bbr_log_type_bbrrttprop(bbr, rtt, 7381 tp->iss, 0, cts, 7382 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7383 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7384 changed = 1; 7385 bbr->r_wanted_output = 1; 7386 goto out; 7387 } 7388 goto proc_sack; 7389 } else if (rsm == NULL) { 7390 goto out; 7391 } 7392 if (changed) { 7393 /* 7394 * The ACK point is advancing to th_ack, we must drop off 7395 * the packets in the rack log and calculate any eligble 7396 * RTT's. 7397 */ 7398 bbr->r_wanted_output = 1; 7399 more: 7400 if (rsm == NULL) { 7401 if (tp->t_flags & TF_SENTFIN) { 7402 /* if we send a FIN we will not hav a map */ 7403 goto proc_sack; 7404 } 7405 #ifdef BBR_INVARIANTS 7406 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7407 tp, 7408 th, tp->t_state, bbr, 7409 tp->snd_una, tp->snd_max, changed); 7410 #endif 7411 goto proc_sack; 7412 } 7413 } 7414 if (SEQ_LT(th_ack, rsm->r_start)) { 7415 /* Huh map is missing this */ 7416 #ifdef BBR_INVARIANTS 7417 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7418 rsm->r_start, 7419 th_ack, tp->t_state, 7420 bbr->r_state, bbr); 7421 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7422 #endif 7423 goto proc_sack; 7424 } else if (th_ack == rsm->r_start) { 7425 /* None here to ack */ 7426 goto proc_sack; 7427 } 7428 /* 7429 * Clear the dup ack counter, it will 7430 * either be freed or if there is some 7431 * remaining we need to start it at zero. 7432 */ 7433 rsm->r_dupack = 0; 7434 /* Now do we consume the whole thing? */ 7435 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7436 /* Its all consumed. */ 7437 uint32_t left; 7438 7439 if (rsm->r_flags & BBR_ACKED) { 7440 /* 7441 * It was acked on the scoreboard -- remove it from 7442 * total 7443 */ 7444 p_acked += (rsm->r_end - rsm->r_start); 7445 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7446 if (bbr->r_ctl.rc_sacked == 0) 7447 bbr->r_ctl.rc_sacklast = NULL; 7448 } else { 7449 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7450 if (rsm->r_flags & BBR_MARKED_LOST) { 7451 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7452 } 7453 if (rsm->r_flags & BBR_SACK_PASSED) { 7454 /* 7455 * There are acked segments ACKED on the 7456 * scoreboard further up. We are seeing 7457 * reordering. 7458 */ 7459 BBR_STAT_INC(bbr_reorder_seen); 7460 bbr->r_ctl.rc_reorder_ts = cts; 7461 if (rsm->r_flags & BBR_MARKED_LOST) { 7462 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7463 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7464 /* LT sampling also needs adjustment */ 7465 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7466 } 7467 } 7468 rsm->r_flags &= ~BBR_MARKED_LOST; 7469 } 7470 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7471 rsm->r_rtr_bytes = 0; 7472 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7473 if (rsm->r_in_tmap) { 7474 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7475 rsm->r_in_tmap = 0; 7476 } 7477 if (bbr->r_ctl.rc_next == rsm) { 7478 /* scoot along the marker */ 7479 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7480 } 7481 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7482 /* Adjust the packet counts */ 7483 left = th_ack - rsm->r_end; 7484 /* Free back to zone */ 7485 bbr_free(bbr, rsm); 7486 if (left) { 7487 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7488 goto more; 7489 } 7490 goto proc_sack; 7491 } 7492 if (rsm->r_flags & BBR_ACKED) { 7493 /* 7494 * It was acked on the scoreboard -- remove it from total 7495 * for the part being cum-acked. 7496 */ 7497 p_acked += (rsm->r_end - rsm->r_start); 7498 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7499 if (bbr->r_ctl.rc_sacked == 0) 7500 bbr->r_ctl.rc_sacklast = NULL; 7501 } else { 7502 /* 7503 * It was acked up to th_ack point for the first time 7504 */ 7505 struct bbr_sendmap lrsm; 7506 7507 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7508 lrsm.r_end = th_ack; 7509 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7510 } 7511 if ((rsm->r_flags & BBR_MARKED_LOST) && 7512 ((rsm->r_flags & BBR_ACKED) == 0)) { 7513 /* 7514 * It was marked lost and partly ack'd now 7515 * for the first time. We lower the rc_lost_bytes 7516 * and still leave it MARKED. 7517 */ 7518 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7519 } 7520 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7521 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7522 rsm->r_rtr_bytes = 0; 7523 /* adjust packet count */ 7524 rsm->r_start = th_ack; 7525 proc_sack: 7526 /* Check for reneging */ 7527 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7528 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7529 /* 7530 * The peer has moved snd_una up to the edge of this send, 7531 * i.e. one that it had previously acked. The only way that 7532 * can be true if the peer threw away data (space issues) 7533 * that it had previously sacked (else it would have given 7534 * us snd_una up to (rsm->r_end). We need to undo the acked 7535 * markings here. 7536 * 7537 * Note we have to look to make sure th_ack is our 7538 * rsm->r_start in case we get an old ack where th_ack is 7539 * behind snd_una. 7540 */ 7541 bbr_peer_reneges(bbr, rsm, th->th_ack); 7542 } 7543 if ((to->to_flags & TOF_SACK) == 0) { 7544 /* We are done nothing left to log */ 7545 goto out; 7546 } 7547 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7548 if (rsm) { 7549 last_seq = rsm->r_end; 7550 } else { 7551 last_seq = tp->snd_max; 7552 } 7553 /* Sack block processing */ 7554 if (SEQ_GT(th_ack, tp->snd_una)) 7555 ack_point = th_ack; 7556 else 7557 ack_point = tp->snd_una; 7558 for (i = 0; i < to->to_nsacks; i++) { 7559 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7560 &sack, sizeof(sack)); 7561 sack.start = ntohl(sack.start); 7562 sack.end = ntohl(sack.end); 7563 if (SEQ_GT(sack.end, sack.start) && 7564 SEQ_GT(sack.start, ack_point) && 7565 SEQ_LT(sack.start, tp->snd_max) && 7566 SEQ_GT(sack.end, ack_point) && 7567 SEQ_LEQ(sack.end, tp->snd_max)) { 7568 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7569 (SEQ_LT(sack.end, last_seq)) && 7570 ((sack.end - sack.start) < (p_maxseg / 8))) { 7571 /* 7572 * Not the last piece and its smaller than 7573 * 1/8th of a p_maxseg. We ignore this. 7574 */ 7575 BBR_STAT_INC(bbr_runt_sacks); 7576 continue; 7577 } 7578 sack_blocks[num_sack_blks] = sack; 7579 num_sack_blks++; 7580 } else if (SEQ_LEQ(sack.start, th_ack) && 7581 SEQ_LEQ(sack.end, th_ack)) { 7582 /* 7583 * Its a D-SACK block. 7584 */ 7585 tcp_record_dsack(tp, sack.start, sack.end, 0); 7586 } 7587 } 7588 if (num_sack_blks == 0) 7589 goto out; 7590 /* 7591 * Sort the SACK blocks so we can update the rack scoreboard with 7592 * just one pass. 7593 */ 7594 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, 7595 num_sack_blks, th->th_ack); 7596 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7597 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7598 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7599 num_sack_blks = new_sb; 7600 if (num_sack_blks < 2) { 7601 goto do_sack_work; 7602 } 7603 /* Sort the sacks */ 7604 for (i = 0; i < num_sack_blks; i++) { 7605 for (j = i + 1; j < num_sack_blks; j++) { 7606 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7607 sack = sack_blocks[i]; 7608 sack_blocks[i] = sack_blocks[j]; 7609 sack_blocks[j] = sack; 7610 } 7611 } 7612 } 7613 /* 7614 * Now are any of the sack block ends the same (yes some 7615 * implememtations send these)? 7616 */ 7617 again: 7618 if (num_sack_blks > 1) { 7619 for (i = 0; i < num_sack_blks; i++) { 7620 for (j = i + 1; j < num_sack_blks; j++) { 7621 if (sack_blocks[i].end == sack_blocks[j].end) { 7622 /* 7623 * Ok these two have the same end we 7624 * want the smallest end and then 7625 * throw away the larger and start 7626 * again. 7627 */ 7628 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7629 /* 7630 * The second block covers 7631 * more area use that 7632 */ 7633 sack_blocks[i].start = sack_blocks[j].start; 7634 } 7635 /* 7636 * Now collapse out the dup-sack and 7637 * lower the count 7638 */ 7639 for (k = (j + 1); k < num_sack_blks; k++) { 7640 sack_blocks[j].start = sack_blocks[k].start; 7641 sack_blocks[j].end = sack_blocks[k].end; 7642 j++; 7643 } 7644 num_sack_blks--; 7645 goto again; 7646 } 7647 } 7648 } 7649 } 7650 do_sack_work: 7651 rsm = bbr->r_ctl.rc_sacklast; 7652 for (i = 0; i < num_sack_blks; i++) { 7653 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7654 if (acked) { 7655 bbr->r_wanted_output = 1; 7656 changed += acked; 7657 sack_changed += acked; 7658 } 7659 } 7660 out: 7661 *prev_acked = p_acked; 7662 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7663 /* 7664 * Ok we have a high probability that we need to go in to 7665 * recovery since we have data sack'd 7666 */ 7667 struct bbr_sendmap *rsm; 7668 7669 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7670 if (rsm) { 7671 /* Enter recovery */ 7672 entered_recovery = 1; 7673 bbr->r_wanted_output = 1; 7674 /* 7675 * When we enter recovery we need to assure we send 7676 * one packet. 7677 */ 7678 if (bbr->r_ctl.rc_resend == NULL) { 7679 bbr->r_ctl.rc_resend = rsm; 7680 } 7681 } 7682 } 7683 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7684 /* 7685 * See if we need to rack-retransmit anything if so set it 7686 * up as the thing to resend assuming something else is not 7687 * already in that position. 7688 */ 7689 if (bbr->r_ctl.rc_resend == NULL) { 7690 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7691 } 7692 } 7693 /* 7694 * We return the amount that changed via sack, this is used by the 7695 * ack-received code to augment what was changed between th_ack <-> 7696 * snd_una. 7697 */ 7698 return (sack_changed); 7699 } 7700 7701 static void 7702 bbr_strike_dupack(struct tcp_bbr *bbr) 7703 { 7704 struct bbr_sendmap *rsm; 7705 7706 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7707 if (rsm && (rsm->r_dupack < 0xff)) { 7708 rsm->r_dupack++; 7709 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7710 bbr->r_wanted_output = 1; 7711 } 7712 } 7713 7714 /* 7715 * Return value of 1, we do not need to call bbr_process_data(). 7716 * return value of 0, bbr_process_data can be called. 7717 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7718 * its unlocked and probably unsafe to touch the TCB. 7719 */ 7720 static int 7721 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7722 struct tcpcb *tp, struct tcpopt *to, 7723 uint32_t tiwin, int32_t tlen, 7724 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7725 { 7726 int32_t ourfinisacked = 0; 7727 int32_t acked_amount; 7728 uint16_t nsegs; 7729 int32_t acked; 7730 uint32_t lost, sack_changed = 0; 7731 struct mbuf *mfree; 7732 struct tcp_bbr *bbr; 7733 uint32_t prev_acked = 0; 7734 7735 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7736 lost = bbr->r_ctl.rc_lost; 7737 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7738 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7739 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7740 bbr->r_wanted_output = 1; 7741 return (1); 7742 } 7743 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7744 /* Process the ack */ 7745 if (bbr->rc_in_persist) 7746 tp->t_rxtshift = 0; 7747 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7748 bbr_strike_dupack(bbr); 7749 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7750 } 7751 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7752 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7753 /* 7754 * Old ack, behind the last one rcv'd or a duplicate ack 7755 * with SACK info. 7756 */ 7757 if (th->th_ack == tp->snd_una) { 7758 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7759 if (bbr->r_state == TCPS_SYN_SENT) { 7760 /* 7761 * Special case on where we sent SYN. When 7762 * the SYN-ACK is processed in syn_sent 7763 * state it bumps the snd_una. This causes 7764 * us to hit here even though we did ack 1 7765 * byte. 7766 * 7767 * Go through the nothing left case so we 7768 * send data. 7769 */ 7770 goto nothing_left; 7771 } 7772 } 7773 return (0); 7774 } 7775 /* 7776 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7777 * something we sent. 7778 */ 7779 if (tp->t_flags & TF_NEEDSYN) { 7780 /* 7781 * T/TCP: Connection was half-synchronized, and our SYN has 7782 * been ACK'd (so connection is now fully synchronized). Go 7783 * to non-starred state, increment snd_una for ACK of SYN, 7784 * and check if we can do window scaling. 7785 */ 7786 tp->t_flags &= ~TF_NEEDSYN; 7787 tp->snd_una++; 7788 /* Do window scaling? */ 7789 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7790 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7791 tp->rcv_scale = tp->request_r_scale; 7792 /* Send window already scaled. */ 7793 } 7794 } 7795 INP_WLOCK_ASSERT(tp->t_inpcb); 7796 7797 acked = BYTES_THIS_ACK(tp, th); 7798 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7799 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7800 7801 /* 7802 * If we just performed our first retransmit, and the ACK arrives 7803 * within our recovery window, then it was a mistake to do the 7804 * retransmit in the first place. Recover our original cwnd and 7805 * ssthresh, and proceed to transmit where we left off. 7806 */ 7807 if (tp->t_flags & TF_PREVVALID) { 7808 tp->t_flags &= ~TF_PREVVALID; 7809 if (tp->t_rxtshift == 1 && 7810 (int)(ticks - tp->t_badrxtwin) < 0) 7811 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7812 } 7813 SOCKBUF_LOCK(&so->so_snd); 7814 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7815 tp->snd_wnd -= acked_amount; 7816 mfree = sbcut_locked(&so->so_snd, acked_amount); 7817 /* NB: sowwakeup_locked() does an implicit unlock. */ 7818 sowwakeup_locked(so); 7819 m_freem(mfree); 7820 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7821 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7822 } 7823 tp->snd_una = th->th_ack; 7824 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7825 if (IN_RECOVERY(tp->t_flags)) { 7826 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7827 (SEQ_LT(th->th_ack, tp->snd_max))) { 7828 tcp_bbr_partialack(tp); 7829 } else { 7830 bbr_post_recovery(tp); 7831 } 7832 } 7833 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7834 tp->snd_recover = tp->snd_una; 7835 } 7836 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7837 tp->snd_nxt = tp->snd_max; 7838 } 7839 if (tp->snd_una == tp->snd_max) { 7840 /* Nothing left outstanding */ 7841 nothing_left: 7842 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7843 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) 7844 bbr->rc_tp->t_acktime = 0; 7845 if ((sbused(&so->so_snd) == 0) && 7846 (tp->t_flags & TF_SENTFIN)) { 7847 ourfinisacked = 1; 7848 } 7849 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7850 if (bbr->rc_in_persist == 0) { 7851 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7852 } 7853 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7854 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7855 /* 7856 * We invalidate the last ack here since we 7857 * don't want to transfer forward the time 7858 * for our sum's calculations. 7859 */ 7860 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7861 (sbavail(&so->so_snd) == 0) && 7862 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7863 /* 7864 * The socket was gone and the peer sent data, time 7865 * to reset him. 7866 */ 7867 *ret_val = 1; 7868 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7869 /* tcp_close will kill the inp pre-log the Reset */ 7870 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7871 tp = tcp_close(tp); 7872 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7873 BBR_STAT_INC(bbr_dropped_af_data); 7874 return (1); 7875 } 7876 /* Set need output so persist might get set */ 7877 bbr->r_wanted_output = 1; 7878 } 7879 if (ofia) 7880 *ofia = ourfinisacked; 7881 return (0); 7882 } 7883 7884 static void 7885 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7886 { 7887 if (bbr->rc_in_persist == 0) { 7888 bbr_timer_cancel(bbr, __LINE__, cts); 7889 bbr->r_ctl.rc_last_delay_val = 0; 7890 tp->t_rxtshift = 0; 7891 bbr->rc_in_persist = 1; 7892 bbr->r_ctl.rc_went_idle_time = cts; 7893 /* We should be capped when rw went to 0 but just in case */ 7894 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7895 /* Time freezes for the state, so do the accounting now */ 7896 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7897 uint32_t time_in; 7898 7899 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7900 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7901 int32_t idx; 7902 7903 idx = bbr_state_val(bbr); 7904 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7905 } else { 7906 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7907 } 7908 } 7909 bbr->r_ctl.rc_bbr_state_time = cts; 7910 } 7911 } 7912 7913 static void 7914 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7915 { 7916 /* 7917 * Note that if idle time does not exceed our 7918 * threshold, we do nothing continuing the state 7919 * transitions we were last walking through. 7920 */ 7921 if (idle_time >= bbr_idle_restart_threshold) { 7922 if (bbr->rc_use_idle_restart) { 7923 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7924 /* 7925 * Set our target using BBR_UNIT, so 7926 * we increase at a dramatic rate but 7927 * we stop when we get the pipe 7928 * full again for our current b/w estimate. 7929 */ 7930 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7931 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7932 bbr_set_state_target(bbr, __LINE__); 7933 /* Now setup our gains to ramp up */ 7934 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7935 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7936 bbr_log_type_statechange(bbr, cts, __LINE__); 7937 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7938 bbr_substate_change(bbr, cts, __LINE__, 1); 7939 } 7940 } 7941 } 7942 7943 static void 7944 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7945 { 7946 uint32_t idle_time; 7947 7948 if (bbr->rc_in_persist == 0) 7949 return; 7950 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7951 bbr->rc_in_persist = 0; 7952 bbr->rc_hit_state_1 = 0; 7953 bbr->r_ctl.rc_del_time = cts; 7954 /* 7955 * We invalidate the last ack here since we 7956 * don't want to transfer forward the time 7957 * for our sum's calculations. 7958 */ 7959 if (tcp_in_hpts(bbr->rc_inp)) { 7960 tcp_hpts_remove(bbr->rc_inp); 7961 bbr->rc_timer_first = 0; 7962 bbr->r_ctl.rc_hpts_flags = 0; 7963 bbr->r_ctl.rc_last_delay_val = 0; 7964 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7965 bbr->r_agg_early_set = 0; 7966 bbr->r_ctl.rc_agg_early = 0; 7967 } 7968 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7969 if (idle_time >= bbr_rtt_probe_time) { 7970 /* 7971 * This qualifies as a RTT_PROBE session since we drop the 7972 * data outstanding to nothing and waited more than 7973 * bbr_rtt_probe_time. 7974 */ 7975 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7976 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7977 } 7978 tp->t_rxtshift = 0; 7979 /* 7980 * If in probeBW and we have persisted more than an RTT lets do 7981 * special handling. 7982 */ 7983 /* Force a time based epoch */ 7984 bbr_set_epoch(bbr, cts, __LINE__); 7985 /* 7986 * Setup the lost so we don't count anything against the guy 7987 * we have been stuck with during persists. 7988 */ 7989 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7990 /* Time un-freezes for the state */ 7991 bbr->r_ctl.rc_bbr_state_time = cts; 7992 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7993 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7994 /* 7995 * If we are going back to probe-bw 7996 * or probe_rtt, we may need to possibly 7997 * do a fast restart. 7998 */ 7999 bbr_restart_after_idle(bbr, cts, idle_time); 8000 } 8001 } 8002 8003 static void 8004 bbr_collapsed_window(struct tcp_bbr *bbr) 8005 { 8006 /* 8007 * Now we must walk the 8008 * send map and divide the 8009 * ones left stranded. These 8010 * guys can't cause us to abort 8011 * the connection and are really 8012 * "unsent". However if a buggy 8013 * client actually did keep some 8014 * of the data i.e. collapsed the win 8015 * and refused to ack and then opened 8016 * the win and acked that data. We would 8017 * get into an ack war, the simplier 8018 * method then of just pretending we 8019 * did not send those segments something 8020 * won't work. 8021 */ 8022 struct bbr_sendmap *rsm, *nrsm; 8023 tcp_seq max_seq; 8024 uint32_t maxseg; 8025 int can_split = 0; 8026 int fnd = 0; 8027 8028 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 8029 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 8030 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 8031 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 8032 /* Find the first seq past or at maxseq */ 8033 if (rsm->r_flags & BBR_RWND_COLLAPSED) 8034 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8035 if (SEQ_GEQ(max_seq, rsm->r_start) && 8036 SEQ_GEQ(rsm->r_end, max_seq)) { 8037 fnd = 1; 8038 break; 8039 } 8040 } 8041 bbr->rc_has_collapsed = 0; 8042 if (!fnd) { 8043 /* Nothing to do strange */ 8044 return; 8045 } 8046 /* 8047 * Now can we split? 8048 * 8049 * We don't want to split if splitting 8050 * would generate too many small segments 8051 * less we let an attacker fragment our 8052 * send_map and leave us out of memory. 8053 */ 8054 if ((max_seq != rsm->r_start) && 8055 (max_seq != rsm->r_end)){ 8056 /* can we split? */ 8057 int res1, res2; 8058 8059 res1 = max_seq - rsm->r_start; 8060 res2 = rsm->r_end - max_seq; 8061 if ((res1 >= (maxseg/8)) && 8062 (res2 >= (maxseg/8))) { 8063 /* No small pieces here */ 8064 can_split = 1; 8065 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8066 /* We are under the limit */ 8067 can_split = 1; 8068 } 8069 } 8070 /* Ok do we need to split this rsm? */ 8071 if (max_seq == rsm->r_start) { 8072 /* It's this guy no split required */ 8073 nrsm = rsm; 8074 } else if (max_seq == rsm->r_end) { 8075 /* It's the next one no split required. */ 8076 nrsm = TAILQ_NEXT(rsm, r_next); 8077 if (nrsm == NULL) { 8078 /* Huh? */ 8079 return; 8080 } 8081 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8082 /* yep we need to split it */ 8083 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8084 if (nrsm == NULL) { 8085 /* failed XXXrrs what can we do mark the whole? */ 8086 nrsm = rsm; 8087 goto no_split; 8088 } 8089 /* Clone it */ 8090 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8091 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8092 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8093 if (rsm->r_in_tmap) { 8094 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8095 nrsm->r_in_tmap = 1; 8096 } 8097 } else { 8098 /* 8099 * Split not allowed just start here just 8100 * use this guy. 8101 */ 8102 nrsm = rsm; 8103 } 8104 no_split: 8105 BBR_STAT_INC(bbr_collapsed_win); 8106 /* reuse fnd as a count */ 8107 fnd = 0; 8108 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8109 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8110 fnd++; 8111 bbr->rc_has_collapsed = 1; 8112 } 8113 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8114 } 8115 8116 static void 8117 bbr_un_collapse_window(struct tcp_bbr *bbr) 8118 { 8119 struct bbr_sendmap *rsm; 8120 int cleared = 0; 8121 8122 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8123 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8124 /* Clear the flag */ 8125 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8126 cleared++; 8127 } else 8128 break; 8129 } 8130 bbr_log_type_rwnd_collapse(bbr, 8131 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8132 bbr->rc_has_collapsed = 0; 8133 } 8134 8135 /* 8136 * Return value of 1, the TCB is unlocked and most 8137 * likely gone, return value of 0, the TCB is still 8138 * locked. 8139 */ 8140 static int 8141 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8142 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8143 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8144 { 8145 /* 8146 * Update window information. Don't look at window if no ACK: TAC's 8147 * send garbage on first SYN. 8148 */ 8149 uint16_t nsegs; 8150 int32_t tfo_syn; 8151 struct tcp_bbr *bbr; 8152 8153 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8154 INP_WLOCK_ASSERT(tp->t_inpcb); 8155 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8156 if ((thflags & TH_ACK) && 8157 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8158 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8159 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8160 /* keep track of pure window updates */ 8161 if (tlen == 0 && 8162 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8163 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8164 tp->snd_wnd = tiwin; 8165 tp->snd_wl1 = th->th_seq; 8166 tp->snd_wl2 = th->th_ack; 8167 if (tp->snd_wnd > tp->max_sndwnd) 8168 tp->max_sndwnd = tp->snd_wnd; 8169 bbr->r_wanted_output = 1; 8170 } else if (thflags & TH_ACK) { 8171 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8172 tp->snd_wnd = tiwin; 8173 tp->snd_wl1 = th->th_seq; 8174 tp->snd_wl2 = th->th_ack; 8175 } 8176 } 8177 if (tp->snd_wnd < ctf_outstanding(tp)) 8178 /* The peer collapsed its window on us */ 8179 bbr_collapsed_window(bbr); 8180 else if (bbr->rc_has_collapsed) 8181 bbr_un_collapse_window(bbr); 8182 /* Was persist timer active and now we have window space? */ 8183 if ((bbr->rc_in_persist != 0) && 8184 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8185 bbr_minseg(bbr)))) { 8186 /* 8187 * Make the rate persist at end of persist mode if idle long 8188 * enough 8189 */ 8190 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8191 8192 /* Make sure we output to start the timer */ 8193 bbr->r_wanted_output = 1; 8194 } 8195 /* Do we need to enter persist? */ 8196 if ((bbr->rc_in_persist == 0) && 8197 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8198 TCPS_HAVEESTABLISHED(tp->t_state) && 8199 (tp->snd_max == tp->snd_una) && 8200 sbavail(&tp->t_inpcb->inp_socket->so_snd) && 8201 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { 8202 /* No send window.. we must enter persist */ 8203 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8204 } 8205 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8206 m_freem(m); 8207 return (0); 8208 } 8209 /* 8210 * We don't support urgent data but 8211 * drag along the up just to make sure 8212 * if there is a stack switch no one 8213 * is surprised. 8214 */ 8215 tp->rcv_up = tp->rcv_nxt; 8216 INP_WLOCK_ASSERT(tp->t_inpcb); 8217 8218 /* 8219 * Process the segment text, merging it into the TCP sequencing 8220 * queue, and arranging for acknowledgment of receipt if necessary. 8221 * This process logically involves adjusting tp->rcv_wnd as data is 8222 * presented to the user (this happens in tcp_usrreq.c, case 8223 * PRU_RCVD). If a FIN has already been received on this connection 8224 * then we just ignore the text. 8225 */ 8226 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8227 IS_FASTOPEN(tp->t_flags)); 8228 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8229 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8230 tcp_seq save_start = th->th_seq; 8231 tcp_seq save_rnxt = tp->rcv_nxt; 8232 int save_tlen = tlen; 8233 8234 m_adj(m, drop_hdrlen); /* delayed header drop */ 8235 /* 8236 * Insert segment which includes th into TCP reassembly 8237 * queue with control block tp. Set thflags to whether 8238 * reassembly now includes a segment with FIN. This handles 8239 * the common case inline (segment is the next to be 8240 * received on an established connection, and the queue is 8241 * empty), avoiding linkage into and removal from the queue 8242 * and repetition of various conversions. Set DELACK for 8243 * segments received in order, but ack immediately when 8244 * segments are out of order (so fast retransmit can work). 8245 */ 8246 if (th->th_seq == tp->rcv_nxt && 8247 SEGQ_EMPTY(tp) && 8248 (TCPS_HAVEESTABLISHED(tp->t_state) || 8249 tfo_syn)) { 8250 #ifdef NETFLIX_SB_LIMITS 8251 u_int mcnt, appended; 8252 8253 if (so->so_rcv.sb_shlim) { 8254 mcnt = m_memcnt(m); 8255 appended = 0; 8256 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8257 CFO_NOSLEEP, NULL) == false) { 8258 counter_u64_add(tcp_sb_shlim_fails, 1); 8259 m_freem(m); 8260 return (0); 8261 } 8262 } 8263 8264 #endif 8265 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8266 bbr->bbr_segs_rcvd += max(1, nsegs); 8267 tp->t_flags |= TF_DELACK; 8268 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8269 } else { 8270 bbr->r_wanted_output = 1; 8271 tp->t_flags |= TF_ACKNOW; 8272 } 8273 tp->rcv_nxt += tlen; 8274 if (tlen && 8275 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8276 (tp->t_fbyte_in == 0)) { 8277 tp->t_fbyte_in = ticks; 8278 if (tp->t_fbyte_in == 0) 8279 tp->t_fbyte_in = 1; 8280 if (tp->t_fbyte_out && tp->t_fbyte_in) 8281 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8282 } 8283 thflags = tcp_get_flags(th) & TH_FIN; 8284 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8285 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8286 SOCKBUF_LOCK(&so->so_rcv); 8287 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8288 m_freem(m); 8289 else 8290 #ifdef NETFLIX_SB_LIMITS 8291 appended = 8292 #endif 8293 sbappendstream_locked(&so->so_rcv, m, 0); 8294 /* NB: sorwakeup_locked() does an implicit unlock. */ 8295 sorwakeup_locked(so); 8296 #ifdef NETFLIX_SB_LIMITS 8297 if (so->so_rcv.sb_shlim && appended != mcnt) 8298 counter_fo_release(so->so_rcv.sb_shlim, 8299 mcnt - appended); 8300 #endif 8301 8302 } else { 8303 /* 8304 * XXX: Due to the header drop above "th" is 8305 * theoretically invalid by now. Fortunately 8306 * m_adj() doesn't actually frees any mbufs when 8307 * trimming from the head. 8308 */ 8309 tcp_seq temp = save_start; 8310 8311 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8312 tp->t_flags |= TF_ACKNOW; 8313 if (tp->t_flags & TF_WAKESOR) { 8314 tp->t_flags &= ~TF_WAKESOR; 8315 /* NB: sorwakeup_locked() does an implicit unlock. */ 8316 sorwakeup_locked(so); 8317 } 8318 } 8319 if ((tp->t_flags & TF_SACK_PERMIT) && 8320 (save_tlen > 0) && 8321 TCPS_HAVEESTABLISHED(tp->t_state)) { 8322 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8323 /* 8324 * DSACK actually handled in the fastpath 8325 * above. 8326 */ 8327 tcp_update_sack_list(tp, save_start, 8328 save_start + save_tlen); 8329 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8330 if ((tp->rcv_numsacks >= 1) && 8331 (tp->sackblks[0].end == save_start)) { 8332 /* 8333 * Partial overlap, recorded at todrop 8334 * above. 8335 */ 8336 tcp_update_sack_list(tp, 8337 tp->sackblks[0].start, 8338 tp->sackblks[0].end); 8339 } else { 8340 tcp_update_dsack_list(tp, save_start, 8341 save_start + save_tlen); 8342 } 8343 } else if (tlen >= save_tlen) { 8344 /* Update of sackblks. */ 8345 tcp_update_dsack_list(tp, save_start, 8346 save_start + save_tlen); 8347 } else if (tlen > 0) { 8348 tcp_update_dsack_list(tp, save_start, 8349 save_start + tlen); 8350 } 8351 } 8352 } else { 8353 m_freem(m); 8354 thflags &= ~TH_FIN; 8355 } 8356 8357 /* 8358 * If FIN is received ACK the FIN and let the user know that the 8359 * connection is closing. 8360 */ 8361 if (thflags & TH_FIN) { 8362 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8363 /* The socket upcall is handled by socantrcvmore. */ 8364 socantrcvmore(so); 8365 /* 8366 * If connection is half-synchronized (ie NEEDSYN 8367 * flag on) then delay ACK, so it may be piggybacked 8368 * when SYN is sent. Otherwise, since we received a 8369 * FIN then no more input can be expected, send ACK 8370 * now. 8371 */ 8372 if (tp->t_flags & TF_NEEDSYN) { 8373 tp->t_flags |= TF_DELACK; 8374 bbr_timer_cancel(bbr, 8375 __LINE__, bbr->r_ctl.rc_rcvtime); 8376 } else { 8377 tp->t_flags |= TF_ACKNOW; 8378 } 8379 tp->rcv_nxt++; 8380 } 8381 switch (tp->t_state) { 8382 /* 8383 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8384 * CLOSE_WAIT state. 8385 */ 8386 case TCPS_SYN_RECEIVED: 8387 tp->t_starttime = ticks; 8388 /* FALLTHROUGH */ 8389 case TCPS_ESTABLISHED: 8390 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8391 break; 8392 8393 /* 8394 * If still in FIN_WAIT_1 STATE FIN has not been 8395 * acked so enter the CLOSING state. 8396 */ 8397 case TCPS_FIN_WAIT_1: 8398 tcp_state_change(tp, TCPS_CLOSING); 8399 break; 8400 8401 /* 8402 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8403 * starting the time-wait timer, turning off the 8404 * other standard timers. 8405 */ 8406 case TCPS_FIN_WAIT_2: 8407 bbr->rc_timer_first = 1; 8408 bbr_timer_cancel(bbr, 8409 __LINE__, bbr->r_ctl.rc_rcvtime); 8410 INP_WLOCK_ASSERT(tp->t_inpcb); 8411 tcp_twstart(tp); 8412 return (1); 8413 } 8414 } 8415 /* 8416 * Return any desired output. 8417 */ 8418 if ((tp->t_flags & TF_ACKNOW) || 8419 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8420 bbr->r_wanted_output = 1; 8421 } 8422 INP_WLOCK_ASSERT(tp->t_inpcb); 8423 return (0); 8424 } 8425 8426 /* 8427 * Here nothing is really faster, its just that we 8428 * have broken out the fast-data path also just like 8429 * the fast-ack. Return 1 if we processed the packet 8430 * return 0 if you need to take the "slow-path". 8431 */ 8432 static int 8433 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8434 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8435 uint32_t tiwin, int32_t nxt_pkt) 8436 { 8437 uint16_t nsegs; 8438 int32_t newsize = 0; /* automatic sockbuf scaling */ 8439 struct tcp_bbr *bbr; 8440 #ifdef NETFLIX_SB_LIMITS 8441 u_int mcnt, appended; 8442 #endif 8443 #ifdef TCPDEBUG 8444 /* 8445 * The size of tcp_saveipgen must be the size of the max ip header, 8446 * now IPv6. 8447 */ 8448 u_char tcp_saveipgen[IP6_HDR_LEN]; 8449 struct tcphdr tcp_savetcp; 8450 short ostate = 0; 8451 8452 #endif 8453 /* On the hpts and we would have called output */ 8454 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8455 8456 /* 8457 * If last ACK falls within this segment's sequence numbers, record 8458 * the timestamp. NOTE that the test is modified according to the 8459 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8460 */ 8461 if (bbr->r_ctl.rc_resend != NULL) { 8462 return (0); 8463 } 8464 if (tiwin && tiwin != tp->snd_wnd) { 8465 return (0); 8466 } 8467 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8468 return (0); 8469 } 8470 if (__predict_false((to->to_flags & TOF_TS) && 8471 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8472 return (0); 8473 } 8474 if (__predict_false((th->th_ack != tp->snd_una))) { 8475 return (0); 8476 } 8477 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8478 return (0); 8479 } 8480 if ((to->to_flags & TOF_TS) != 0 && 8481 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8482 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8483 tp->ts_recent = to->to_tsval; 8484 } 8485 /* 8486 * This is a pure, in-sequence data packet with nothing on the 8487 * reassembly queue and we have enough buffer space to take it. 8488 */ 8489 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8490 8491 #ifdef NETFLIX_SB_LIMITS 8492 if (so->so_rcv.sb_shlim) { 8493 mcnt = m_memcnt(m); 8494 appended = 0; 8495 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8496 CFO_NOSLEEP, NULL) == false) { 8497 counter_u64_add(tcp_sb_shlim_fails, 1); 8498 m_freem(m); 8499 return (1); 8500 } 8501 } 8502 #endif 8503 /* Clean receiver SACK report if present */ 8504 if (tp->rcv_numsacks) 8505 tcp_clean_sackreport(tp); 8506 KMOD_TCPSTAT_INC(tcps_preddat); 8507 tp->rcv_nxt += tlen; 8508 if (tlen && 8509 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8510 (tp->t_fbyte_in == 0)) { 8511 tp->t_fbyte_in = ticks; 8512 if (tp->t_fbyte_in == 0) 8513 tp->t_fbyte_in = 1; 8514 if (tp->t_fbyte_out && tp->t_fbyte_in) 8515 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8516 } 8517 /* 8518 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8519 */ 8520 tp->snd_wl1 = th->th_seq; 8521 /* 8522 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8523 */ 8524 tp->rcv_up = tp->rcv_nxt; 8525 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8526 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8527 #ifdef TCPDEBUG 8528 if (so->so_options & SO_DEBUG) 8529 tcp_trace(TA_INPUT, ostate, tp, 8530 (void *)tcp_saveipgen, &tcp_savetcp, 0); 8531 #endif 8532 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8533 8534 /* Add data to socket buffer. */ 8535 SOCKBUF_LOCK(&so->so_rcv); 8536 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8537 m_freem(m); 8538 } else { 8539 /* 8540 * Set new socket buffer size. Give up when limit is 8541 * reached. 8542 */ 8543 if (newsize) 8544 if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) 8545 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8546 m_adj(m, drop_hdrlen); /* delayed header drop */ 8547 8548 #ifdef NETFLIX_SB_LIMITS 8549 appended = 8550 #endif 8551 sbappendstream_locked(&so->so_rcv, m, 0); 8552 ctf_calc_rwin(so, tp); 8553 } 8554 /* NB: sorwakeup_locked() does an implicit unlock. */ 8555 sorwakeup_locked(so); 8556 #ifdef NETFLIX_SB_LIMITS 8557 if (so->so_rcv.sb_shlim && mcnt != appended) 8558 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8559 #endif 8560 if (DELAY_ACK(tp, bbr, nsegs)) { 8561 bbr->bbr_segs_rcvd += max(1, nsegs); 8562 tp->t_flags |= TF_DELACK; 8563 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8564 } else { 8565 bbr->r_wanted_output = 1; 8566 tp->t_flags |= TF_ACKNOW; 8567 } 8568 return (1); 8569 } 8570 8571 /* 8572 * This subfunction is used to try to highly optimize the 8573 * fast path. We again allow window updates that are 8574 * in sequence to remain in the fast-path. We also add 8575 * in the __predict's to attempt to help the compiler. 8576 * Note that if we return a 0, then we can *not* process 8577 * it and the caller should push the packet into the 8578 * slow-path. If we return 1, then all is well and 8579 * the packet is fully processed. 8580 */ 8581 static int 8582 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8583 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8584 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8585 { 8586 int32_t acked; 8587 uint16_t nsegs; 8588 uint32_t sack_changed; 8589 #ifdef TCPDEBUG 8590 /* 8591 * The size of tcp_saveipgen must be the size of the max ip header, 8592 * now IPv6. 8593 */ 8594 u_char tcp_saveipgen[IP6_HDR_LEN]; 8595 struct tcphdr tcp_savetcp; 8596 short ostate = 0; 8597 8598 #endif 8599 uint32_t prev_acked = 0; 8600 struct tcp_bbr *bbr; 8601 8602 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8603 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8604 return (0); 8605 } 8606 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8607 /* Above what we have sent? */ 8608 return (0); 8609 } 8610 if (__predict_false(tiwin == 0)) { 8611 /* zero window */ 8612 return (0); 8613 } 8614 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8615 /* We need a SYN or a FIN, unlikely.. */ 8616 return (0); 8617 } 8618 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8619 /* Timestamp is behind .. old ack with seq wrap? */ 8620 return (0); 8621 } 8622 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8623 /* Still recovering */ 8624 return (0); 8625 } 8626 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8627 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8628 /* We are retransmitting */ 8629 return (0); 8630 } 8631 if (__predict_false(bbr->rc_in_persist != 0)) { 8632 /* In persist mode */ 8633 return (0); 8634 } 8635 if (bbr->r_ctl.rc_sacked) { 8636 /* We have sack holes on our scoreboard */ 8637 return (0); 8638 } 8639 /* Ok if we reach here, we can process a fast-ack */ 8640 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8641 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8642 /* 8643 * We never detect loss in fast ack [we can't 8644 * have a sack and can't be in recovery so 8645 * we always pass 0 (nothing detected)]. 8646 */ 8647 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8648 /* Did the window get updated? */ 8649 if (tiwin != tp->snd_wnd) { 8650 tp->snd_wnd = tiwin; 8651 tp->snd_wl1 = th->th_seq; 8652 if (tp->snd_wnd > tp->max_sndwnd) 8653 tp->max_sndwnd = tp->snd_wnd; 8654 } 8655 /* Do we need to exit persists? */ 8656 if ((bbr->rc_in_persist != 0) && 8657 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8658 bbr_minseg(bbr)))) { 8659 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8660 bbr->r_wanted_output = 1; 8661 } 8662 /* Do we need to enter persists? */ 8663 if ((bbr->rc_in_persist == 0) && 8664 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8665 TCPS_HAVEESTABLISHED(tp->t_state) && 8666 (tp->snd_max == tp->snd_una) && 8667 sbavail(&tp->t_inpcb->inp_socket->so_snd) && 8668 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { 8669 /* No send window.. we must enter persist */ 8670 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8671 } 8672 /* 8673 * If last ACK falls within this segment's sequence numbers, record 8674 * the timestamp. NOTE that the test is modified according to the 8675 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8676 */ 8677 if ((to->to_flags & TOF_TS) != 0 && 8678 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8679 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8680 tp->ts_recent = to->to_tsval; 8681 } 8682 /* 8683 * This is a pure ack for outstanding data. 8684 */ 8685 KMOD_TCPSTAT_INC(tcps_predack); 8686 8687 /* 8688 * "bad retransmit" recovery. 8689 */ 8690 if (tp->t_flags & TF_PREVVALID) { 8691 tp->t_flags &= ~TF_PREVVALID; 8692 if (tp->t_rxtshift == 1 && 8693 (int)(ticks - tp->t_badrxtwin) < 0) 8694 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8695 } 8696 /* 8697 * Recalculate the transmit timer / rtt. 8698 * 8699 * Some boxes send broken timestamp replies during the SYN+ACK 8700 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8701 * and blow up the retransmit timer. 8702 */ 8703 acked = BYTES_THIS_ACK(tp, th); 8704 8705 #ifdef TCP_HHOOK 8706 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8707 hhook_run_tcp_est_in(tp, th, to); 8708 #endif 8709 8710 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8711 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8712 sbdrop(&so->so_snd, acked); 8713 8714 if (SEQ_GT(th->th_ack, tp->snd_una)) 8715 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8716 tp->snd_una = th->th_ack; 8717 if (tp->snd_wnd < ctf_outstanding(tp)) 8718 /* The peer collapsed its window on us */ 8719 bbr_collapsed_window(bbr); 8720 else if (bbr->rc_has_collapsed) 8721 bbr_un_collapse_window(bbr); 8722 8723 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8724 tp->snd_recover = tp->snd_una; 8725 } 8726 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8727 /* 8728 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8729 */ 8730 tp->snd_wl2 = th->th_ack; 8731 m_freem(m); 8732 /* 8733 * If all outstanding data are acked, stop retransmit timer, 8734 * otherwise restart timer using current (possibly backed-off) 8735 * value. If process is waiting for space, wakeup/selwakeup/signal. 8736 * If data are ready to send, let tcp_output decide between more 8737 * output or persist. 8738 */ 8739 #ifdef TCPDEBUG 8740 if (so->so_options & SO_DEBUG) 8741 tcp_trace(TA_INPUT, ostate, tp, 8742 (void *)tcp_saveipgen, 8743 &tcp_savetcp, 0); 8744 #endif 8745 /* Wake up the socket if we have room to write more */ 8746 sowwakeup(so); 8747 if (tp->snd_una == tp->snd_max) { 8748 /* Nothing left outstanding */ 8749 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8750 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) 8751 bbr->rc_tp->t_acktime = 0; 8752 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8753 if (bbr->rc_in_persist == 0) { 8754 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8755 } 8756 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8757 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8758 /* 8759 * We invalidate the last ack here since we 8760 * don't want to transfer forward the time 8761 * for our sum's calculations. 8762 */ 8763 bbr->r_wanted_output = 1; 8764 } 8765 if (sbavail(&so->so_snd)) { 8766 bbr->r_wanted_output = 1; 8767 } 8768 return (1); 8769 } 8770 8771 /* 8772 * Return value of 1, the TCB is unlocked and most 8773 * likely gone, return value of 0, the TCB is still 8774 * locked. 8775 */ 8776 static int 8777 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8778 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8779 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8780 { 8781 int32_t todrop; 8782 int32_t ourfinisacked = 0; 8783 struct tcp_bbr *bbr; 8784 int32_t ret_val = 0; 8785 8786 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8787 ctf_calc_rwin(so, tp); 8788 /* 8789 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8790 * SYN, drop the input. if seg contains a RST, then drop the 8791 * connection. if seg does not contain SYN, then drop it. Otherwise 8792 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8793 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8794 * not support ECN so we will not say we are capable. if SYN has 8795 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8796 * segment to be acked (eventually) continue processing rest of 8797 * data/controls, beginning with URG 8798 */ 8799 if ((thflags & TH_ACK) && 8800 (SEQ_LEQ(th->th_ack, tp->iss) || 8801 SEQ_GT(th->th_ack, tp->snd_max))) { 8802 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8803 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8804 return (1); 8805 } 8806 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8807 TCP_PROBE5(connect__refused, NULL, tp, 8808 mtod(m, const char *), tp, th); 8809 tp = tcp_drop(tp, ECONNREFUSED); 8810 ctf_do_drop(m, tp); 8811 return (1); 8812 } 8813 if (thflags & TH_RST) { 8814 ctf_do_drop(m, tp); 8815 return (1); 8816 } 8817 if (!(thflags & TH_SYN)) { 8818 ctf_do_drop(m, tp); 8819 return (1); 8820 } 8821 tp->irs = th->th_seq; 8822 tcp_rcvseqinit(tp); 8823 if (thflags & TH_ACK) { 8824 int tfo_partial = 0; 8825 8826 KMOD_TCPSTAT_INC(tcps_connects); 8827 soisconnected(so); 8828 #ifdef MAC 8829 mac_socketpeer_set_from_mbuf(m, so); 8830 #endif 8831 /* Do window scaling on this connection? */ 8832 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8833 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8834 tp->rcv_scale = tp->request_r_scale; 8835 } 8836 tp->rcv_adv += min(tp->rcv_wnd, 8837 TCP_MAXWIN << tp->rcv_scale); 8838 /* 8839 * If not all the data that was sent in the TFO SYN 8840 * has been acked, resend the remainder right away. 8841 */ 8842 if (IS_FASTOPEN(tp->t_flags) && 8843 (tp->snd_una != tp->snd_max)) { 8844 tp->snd_nxt = th->th_ack; 8845 tfo_partial = 1; 8846 } 8847 /* 8848 * If there's data, delay ACK; if there's also a FIN ACKNOW 8849 * will be turned on later. 8850 */ 8851 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8852 bbr->bbr_segs_rcvd += 1; 8853 tp->t_flags |= TF_DELACK; 8854 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8855 } else { 8856 bbr->r_wanted_output = 1; 8857 tp->t_flags |= TF_ACKNOW; 8858 } 8859 if (SEQ_GT(th->th_ack, tp->iss)) { 8860 /* 8861 * The SYN is acked 8862 * handle it specially. 8863 */ 8864 bbr_log_syn(tp, to); 8865 } 8866 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8867 /* 8868 * We advance snd_una for the 8869 * fast open case. If th_ack is 8870 * acknowledging data beyond 8871 * snd_una we can't just call 8872 * ack-processing since the 8873 * data stream in our send-map 8874 * will start at snd_una + 1 (one 8875 * beyond the SYN). If its just 8876 * equal we don't need to do that 8877 * and there is no send_map. 8878 */ 8879 tp->snd_una++; 8880 } 8881 /* 8882 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8883 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8884 */ 8885 tp->t_starttime = ticks; 8886 if (tp->t_flags & TF_NEEDFIN) { 8887 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8888 tp->t_flags &= ~TF_NEEDFIN; 8889 thflags &= ~TH_SYN; 8890 } else { 8891 tcp_state_change(tp, TCPS_ESTABLISHED); 8892 TCP_PROBE5(connect__established, NULL, tp, 8893 mtod(m, const char *), tp, th); 8894 cc_conn_init(tp); 8895 } 8896 } else { 8897 /* 8898 * Received initial SYN in SYN-SENT[*] state => simultaneous 8899 * open. If segment contains CC option and there is a 8900 * cached CC, apply TAO test. If it succeeds, connection is * 8901 * half-synchronized. Otherwise, do 3-way handshake: 8902 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8903 * there was no CC option, clear cached CC value. 8904 */ 8905 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 8906 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8907 } 8908 INP_WLOCK_ASSERT(tp->t_inpcb); 8909 /* 8910 * Advance th->th_seq to correspond to first data byte. If data, 8911 * trim to stay within window, dropping FIN if necessary. 8912 */ 8913 th->th_seq++; 8914 if (tlen > tp->rcv_wnd) { 8915 todrop = tlen - tp->rcv_wnd; 8916 m_adj(m, -todrop); 8917 tlen = tp->rcv_wnd; 8918 thflags &= ~TH_FIN; 8919 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8920 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8921 } 8922 tp->snd_wl1 = th->th_seq - 1; 8923 tp->rcv_up = th->th_seq; 8924 /* 8925 * Client side of transaction: already sent SYN and data. If the 8926 * remote host used T/TCP to validate the SYN, our data will be 8927 * ACK'd; if so, enter normal data segment processing in the middle 8928 * of step 5, ack processing. Otherwise, goto step 6. 8929 */ 8930 if (thflags & TH_ACK) { 8931 if ((to->to_flags & TOF_TS) != 0) { 8932 uint32_t t, rtt; 8933 8934 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8935 if (TSTMP_GEQ(t, to->to_tsecr)) { 8936 rtt = t - to->to_tsecr; 8937 if (rtt == 0) { 8938 rtt = 1; 8939 } 8940 rtt *= MS_IN_USEC; 8941 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8942 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8943 rtt, bbr->r_ctl.rc_rcvtime); 8944 } 8945 } 8946 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8947 return (ret_val); 8948 /* We may have changed to FIN_WAIT_1 above */ 8949 if (tp->t_state == TCPS_FIN_WAIT_1) { 8950 /* 8951 * In FIN_WAIT_1 STATE in addition to the processing 8952 * for the ESTABLISHED state if our FIN is now 8953 * acknowledged then enter FIN_WAIT_2. 8954 */ 8955 if (ourfinisacked) { 8956 /* 8957 * If we can't receive any more data, then 8958 * closing user can proceed. Starting the 8959 * timer is contrary to the specification, 8960 * but if we don't get a FIN we'll hang 8961 * forever. 8962 * 8963 * XXXjl: we should release the tp also, and 8964 * use a compressed state. 8965 */ 8966 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8967 soisdisconnected(so); 8968 tcp_timer_activate(tp, TT_2MSL, 8969 (tcp_fast_finwait2_recycle ? 8970 tcp_finwait2_timeout : 8971 TP_MAXIDLE(tp))); 8972 } 8973 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8974 } 8975 } 8976 } 8977 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8978 tiwin, thflags, nxt_pkt)); 8979 } 8980 8981 /* 8982 * Return value of 1, the TCB is unlocked and most 8983 * likely gone, return value of 0, the TCB is still 8984 * locked. 8985 */ 8986 static int 8987 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8988 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8989 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8990 { 8991 int32_t ourfinisacked = 0; 8992 int32_t ret_val; 8993 struct tcp_bbr *bbr; 8994 8995 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8996 ctf_calc_rwin(so, tp); 8997 if ((thflags & TH_ACK) && 8998 (SEQ_LEQ(th->th_ack, tp->snd_una) || 8999 SEQ_GT(th->th_ack, tp->snd_max))) { 9000 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9001 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9002 return (1); 9003 } 9004 if (IS_FASTOPEN(tp->t_flags)) { 9005 /* 9006 * When a TFO connection is in SYN_RECEIVED, the only valid 9007 * packets are the initial SYN, a retransmit/copy of the 9008 * initial SYN (possibly with a subset of the original 9009 * data), a valid ACK, a FIN, or a RST. 9010 */ 9011 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 9012 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9013 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9014 return (1); 9015 } else if (thflags & TH_SYN) { 9016 /* non-initial SYN is ignored */ 9017 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 9018 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 9019 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 9020 ctf_do_drop(m, NULL); 9021 return (0); 9022 } 9023 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 9024 ctf_do_drop(m, NULL); 9025 return (0); 9026 } 9027 } 9028 if ((thflags & TH_RST) || 9029 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9030 return (ctf_process_rst(m, th, so, tp)); 9031 /* 9032 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9033 * it's less than ts_recent, drop it. 9034 */ 9035 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9036 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9037 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9038 return (ret_val); 9039 } 9040 /* 9041 * In the SYN-RECEIVED state, validate that the packet belongs to 9042 * this connection before trimming the data to fit the receive 9043 * window. Check the sequence number versus IRS since we know the 9044 * sequence numbers haven't wrapped. This is a partial fix for the 9045 * "LAND" DoS attack. 9046 */ 9047 if (SEQ_LT(th->th_seq, tp->irs)) { 9048 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9049 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9050 return (1); 9051 } 9052 INP_WLOCK_ASSERT(tp->t_inpcb); 9053 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9054 return (ret_val); 9055 } 9056 /* 9057 * If last ACK falls within this segment's sequence numbers, record 9058 * its timestamp. NOTE: 1) That the test incorporates suggestions 9059 * from the latest proposal of the tcplw@cray.com list (Braden 9060 * 1993/04/26). 2) That updating only on newer timestamps interferes 9061 * with our earlier PAWS tests, so this check should be solely 9062 * predicated on the sequence space of this segment. 3) That we 9063 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9064 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9065 * SEG.Len, This modified check allows us to overcome RFC1323's 9066 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9067 * p.869. In such cases, we can still calculate the RTT correctly 9068 * when RCV.NXT == Last.ACK.Sent. 9069 */ 9070 if ((to->to_flags & TOF_TS) != 0 && 9071 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9072 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9073 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9074 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9075 tp->ts_recent = to->to_tsval; 9076 } 9077 tp->snd_wnd = tiwin; 9078 /* 9079 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9080 * is on (half-synchronized state), then queue data for later 9081 * processing; else drop segment and return. 9082 */ 9083 if ((thflags & TH_ACK) == 0) { 9084 if (IS_FASTOPEN(tp->t_flags)) { 9085 cc_conn_init(tp); 9086 } 9087 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9088 tiwin, thflags, nxt_pkt)); 9089 } 9090 KMOD_TCPSTAT_INC(tcps_connects); 9091 soisconnected(so); 9092 /* Do window scaling? */ 9093 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9094 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9095 tp->rcv_scale = tp->request_r_scale; 9096 } 9097 /* 9098 * ok for the first time in lets see if we can use the ts to figure 9099 * out what the initial RTT was. 9100 */ 9101 if ((to->to_flags & TOF_TS) != 0) { 9102 uint32_t t, rtt; 9103 9104 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9105 if (TSTMP_GEQ(t, to->to_tsecr)) { 9106 rtt = t - to->to_tsecr; 9107 if (rtt == 0) { 9108 rtt = 1; 9109 } 9110 rtt *= MS_IN_USEC; 9111 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9112 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9113 } 9114 } 9115 /* Drop off any SYN in the send map (probably not there) */ 9116 if (thflags & TH_ACK) 9117 bbr_log_syn(tp, to); 9118 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 9119 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9120 tp->t_tfo_pending = NULL; 9121 } 9122 /* 9123 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9124 * FIN-WAIT-1 9125 */ 9126 tp->t_starttime = ticks; 9127 if (tp->t_flags & TF_NEEDFIN) { 9128 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9129 tp->t_flags &= ~TF_NEEDFIN; 9130 } else { 9131 tcp_state_change(tp, TCPS_ESTABLISHED); 9132 TCP_PROBE5(accept__established, NULL, tp, 9133 mtod(m, const char *), tp, th); 9134 /* 9135 * TFO connections call cc_conn_init() during SYN 9136 * processing. Calling it again here for such connections 9137 * is not harmless as it would undo the snd_cwnd reduction 9138 * that occurs when a TFO SYN|ACK is retransmitted. 9139 */ 9140 if (!IS_FASTOPEN(tp->t_flags)) 9141 cc_conn_init(tp); 9142 } 9143 /* 9144 * Account for the ACK of our SYN prior to 9145 * regular ACK processing below, except for 9146 * simultaneous SYN, which is handled later. 9147 */ 9148 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9149 tp->snd_una++; 9150 /* 9151 * If segment contains data or ACK, will call tcp_reass() later; if 9152 * not, do so now to pass queued data to user. 9153 */ 9154 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9155 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9156 (struct mbuf *)0); 9157 if (tp->t_flags & TF_WAKESOR) { 9158 tp->t_flags &= ~TF_WAKESOR; 9159 /* NB: sorwakeup_locked() does an implicit unlock. */ 9160 sorwakeup_locked(so); 9161 } 9162 } 9163 tp->snd_wl1 = th->th_seq - 1; 9164 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9165 return (ret_val); 9166 } 9167 if (tp->t_state == TCPS_FIN_WAIT_1) { 9168 /* We could have went to FIN_WAIT_1 (or EST) above */ 9169 /* 9170 * In FIN_WAIT_1 STATE in addition to the processing for the 9171 * ESTABLISHED state if our FIN is now acknowledged then 9172 * enter FIN_WAIT_2. 9173 */ 9174 if (ourfinisacked) { 9175 /* 9176 * If we can't receive any more data, then closing 9177 * user can proceed. Starting the timer is contrary 9178 * to the specification, but if we don't get a FIN 9179 * we'll hang forever. 9180 * 9181 * XXXjl: we should release the tp also, and use a 9182 * compressed state. 9183 */ 9184 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9185 soisdisconnected(so); 9186 tcp_timer_activate(tp, TT_2MSL, 9187 (tcp_fast_finwait2_recycle ? 9188 tcp_finwait2_timeout : 9189 TP_MAXIDLE(tp))); 9190 } 9191 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9192 } 9193 } 9194 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9195 tiwin, thflags, nxt_pkt)); 9196 } 9197 9198 /* 9199 * Return value of 1, the TCB is unlocked and most 9200 * likely gone, return value of 0, the TCB is still 9201 * locked. 9202 */ 9203 static int 9204 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9205 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9206 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9207 { 9208 struct tcp_bbr *bbr; 9209 int32_t ret_val; 9210 9211 /* 9212 * Header prediction: check for the two common cases of a 9213 * uni-directional data xfer. If the packet has no control flags, 9214 * is in-sequence, the window didn't change and we're not 9215 * retransmitting, it's a candidate. If the length is zero and the 9216 * ack moved forward, we're the sender side of the xfer. Just free 9217 * the data acked & wake any higher level process that was blocked 9218 * waiting for space. If the length is non-zero and the ack didn't 9219 * move, we're the receiver side. If we're getting packets in-order 9220 * (the reassembly queue is empty), add the data toc The socket 9221 * buffer and note that we need a delayed ack. Make sure that the 9222 * hidden state-flags are also off. Since we check for 9223 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9224 */ 9225 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9226 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9227 /* 9228 * If we have delived under 4 segments increase the initial 9229 * window if raised by the peer. We use this to determine 9230 * dynamic and static rwnd's at the end of a connection. 9231 */ 9232 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9233 } 9234 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9235 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9236 __predict_true(SEGQ_EMPTY(tp)) && 9237 __predict_true(th->th_seq == tp->rcv_nxt)) { 9238 if (tlen == 0) { 9239 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9240 tiwin, nxt_pkt, iptos)) { 9241 return (0); 9242 } 9243 } else { 9244 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9245 tiwin, nxt_pkt)) { 9246 return (0); 9247 } 9248 } 9249 } 9250 ctf_calc_rwin(so, tp); 9251 9252 if ((thflags & TH_RST) || 9253 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9254 return (ctf_process_rst(m, th, so, tp)); 9255 /* 9256 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9257 * synchronized state. 9258 */ 9259 if (thflags & TH_SYN) { 9260 ctf_challenge_ack(m, th, tp, &ret_val); 9261 return (ret_val); 9262 } 9263 /* 9264 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9265 * it's less than ts_recent, drop it. 9266 */ 9267 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9268 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9269 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9270 return (ret_val); 9271 } 9272 INP_WLOCK_ASSERT(tp->t_inpcb); 9273 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9274 return (ret_val); 9275 } 9276 /* 9277 * If last ACK falls within this segment's sequence numbers, record 9278 * its timestamp. NOTE: 1) That the test incorporates suggestions 9279 * from the latest proposal of the tcplw@cray.com list (Braden 9280 * 1993/04/26). 2) That updating only on newer timestamps interferes 9281 * with our earlier PAWS tests, so this check should be solely 9282 * predicated on the sequence space of this segment. 3) That we 9283 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9284 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9285 * SEG.Len, This modified check allows us to overcome RFC1323's 9286 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9287 * p.869. In such cases, we can still calculate the RTT correctly 9288 * when RCV.NXT == Last.ACK.Sent. 9289 */ 9290 if ((to->to_flags & TOF_TS) != 0 && 9291 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9292 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9293 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9294 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9295 tp->ts_recent = to->to_tsval; 9296 } 9297 /* 9298 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9299 * is on (half-synchronized state), then queue data for later 9300 * processing; else drop segment and return. 9301 */ 9302 if ((thflags & TH_ACK) == 0) { 9303 if (tp->t_flags & TF_NEEDSYN) { 9304 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9305 tiwin, thflags, nxt_pkt)); 9306 } else if (tp->t_flags & TF_ACKNOW) { 9307 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9308 bbr->r_wanted_output = 1; 9309 return (ret_val); 9310 } else { 9311 ctf_do_drop(m, NULL); 9312 return (0); 9313 } 9314 } 9315 /* 9316 * Ack processing. 9317 */ 9318 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9319 return (ret_val); 9320 } 9321 if (sbavail(&so->so_snd)) { 9322 if (ctf_progress_timeout_check(tp, true)) { 9323 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9324 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9325 return (1); 9326 } 9327 } 9328 /* State changes only happen in bbr_process_data() */ 9329 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9330 tiwin, thflags, nxt_pkt)); 9331 } 9332 9333 /* 9334 * Return value of 1, the TCB is unlocked and most 9335 * likely gone, return value of 0, the TCB is still 9336 * locked. 9337 */ 9338 static int 9339 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9340 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9341 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9342 { 9343 struct tcp_bbr *bbr; 9344 int32_t ret_val; 9345 9346 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9347 ctf_calc_rwin(so, tp); 9348 if ((thflags & TH_RST) || 9349 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9350 return (ctf_process_rst(m, th, so, tp)); 9351 /* 9352 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9353 * synchronized state. 9354 */ 9355 if (thflags & TH_SYN) { 9356 ctf_challenge_ack(m, th, tp, &ret_val); 9357 return (ret_val); 9358 } 9359 /* 9360 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9361 * it's less than ts_recent, drop it. 9362 */ 9363 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9364 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9365 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9366 return (ret_val); 9367 } 9368 INP_WLOCK_ASSERT(tp->t_inpcb); 9369 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9370 return (ret_val); 9371 } 9372 /* 9373 * If last ACK falls within this segment's sequence numbers, record 9374 * its timestamp. NOTE: 1) That the test incorporates suggestions 9375 * from the latest proposal of the tcplw@cray.com list (Braden 9376 * 1993/04/26). 2) That updating only on newer timestamps interferes 9377 * with our earlier PAWS tests, so this check should be solely 9378 * predicated on the sequence space of this segment. 3) That we 9379 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9380 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9381 * SEG.Len, This modified check allows us to overcome RFC1323's 9382 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9383 * p.869. In such cases, we can still calculate the RTT correctly 9384 * when RCV.NXT == Last.ACK.Sent. 9385 */ 9386 if ((to->to_flags & TOF_TS) != 0 && 9387 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9388 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9389 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9390 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9391 tp->ts_recent = to->to_tsval; 9392 } 9393 /* 9394 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9395 * is on (half-synchronized state), then queue data for later 9396 * processing; else drop segment and return. 9397 */ 9398 if ((thflags & TH_ACK) == 0) { 9399 if (tp->t_flags & TF_NEEDSYN) { 9400 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9401 tiwin, thflags, nxt_pkt)); 9402 } else if (tp->t_flags & TF_ACKNOW) { 9403 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9404 bbr->r_wanted_output = 1; 9405 return (ret_val); 9406 } else { 9407 ctf_do_drop(m, NULL); 9408 return (0); 9409 } 9410 } 9411 /* 9412 * Ack processing. 9413 */ 9414 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9415 return (ret_val); 9416 } 9417 if (sbavail(&so->so_snd)) { 9418 if (ctf_progress_timeout_check(tp, true)) { 9419 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9420 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9421 return (1); 9422 } 9423 } 9424 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9425 tiwin, thflags, nxt_pkt)); 9426 } 9427 9428 static int 9429 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9430 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9431 { 9432 9433 if (bbr->rc_allow_data_af_clo == 0) { 9434 close_now: 9435 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9436 /* tcp_close will kill the inp pre-log the Reset */ 9437 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9438 tp = tcp_close(tp); 9439 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9440 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9441 return (1); 9442 } 9443 if (sbavail(&so->so_snd) == 0) 9444 goto close_now; 9445 /* Ok we allow data that is ignored and a followup reset */ 9446 tp->rcv_nxt = th->th_seq + *tlen; 9447 tp->t_flags2 |= TF2_DROP_AF_DATA; 9448 bbr->r_wanted_output = 1; 9449 *tlen = 0; 9450 return (0); 9451 } 9452 9453 /* 9454 * Return value of 1, the TCB is unlocked and most 9455 * likely gone, return value of 0, the TCB is still 9456 * locked. 9457 */ 9458 static int 9459 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9460 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9461 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9462 { 9463 int32_t ourfinisacked = 0; 9464 int32_t ret_val; 9465 struct tcp_bbr *bbr; 9466 9467 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9468 ctf_calc_rwin(so, tp); 9469 if ((thflags & TH_RST) || 9470 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9471 return (ctf_process_rst(m, th, so, tp)); 9472 /* 9473 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9474 * synchronized state. 9475 */ 9476 if (thflags & TH_SYN) { 9477 ctf_challenge_ack(m, th, tp, &ret_val); 9478 return (ret_val); 9479 } 9480 /* 9481 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9482 * it's less than ts_recent, drop it. 9483 */ 9484 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9485 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9486 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9487 return (ret_val); 9488 } 9489 INP_WLOCK_ASSERT(tp->t_inpcb); 9490 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9491 return (ret_val); 9492 } 9493 /* 9494 * If new data are received on a connection after the user processes 9495 * are gone, then RST the other end. 9496 * We call a new function now so we might continue and setup 9497 * to reset at all data being ack'd. 9498 */ 9499 if ((tp->t_flags & TF_CLOSED) && tlen && 9500 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9501 return (1); 9502 /* 9503 * If last ACK falls within this segment's sequence numbers, record 9504 * its timestamp. NOTE: 1) That the test incorporates suggestions 9505 * from the latest proposal of the tcplw@cray.com list (Braden 9506 * 1993/04/26). 2) That updating only on newer timestamps interferes 9507 * with our earlier PAWS tests, so this check should be solely 9508 * predicated on the sequence space of this segment. 3) That we 9509 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9510 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9511 * SEG.Len, This modified check allows us to overcome RFC1323's 9512 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9513 * p.869. In such cases, we can still calculate the RTT correctly 9514 * when RCV.NXT == Last.ACK.Sent. 9515 */ 9516 if ((to->to_flags & TOF_TS) != 0 && 9517 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9518 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9519 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9520 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9521 tp->ts_recent = to->to_tsval; 9522 } 9523 /* 9524 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9525 * is on (half-synchronized state), then queue data for later 9526 * processing; else drop segment and return. 9527 */ 9528 if ((thflags & TH_ACK) == 0) { 9529 if (tp->t_flags & TF_NEEDSYN) { 9530 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9531 tiwin, thflags, nxt_pkt)); 9532 } else if (tp->t_flags & TF_ACKNOW) { 9533 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9534 bbr->r_wanted_output = 1; 9535 return (ret_val); 9536 } else { 9537 ctf_do_drop(m, NULL); 9538 return (0); 9539 } 9540 } 9541 /* 9542 * Ack processing. 9543 */ 9544 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9545 return (ret_val); 9546 } 9547 if (ourfinisacked) { 9548 /* 9549 * If we can't receive any more data, then closing user can 9550 * proceed. Starting the timer is contrary to the 9551 * specification, but if we don't get a FIN we'll hang 9552 * forever. 9553 * 9554 * XXXjl: we should release the tp also, and use a 9555 * compressed state. 9556 */ 9557 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9558 soisdisconnected(so); 9559 tcp_timer_activate(tp, TT_2MSL, 9560 (tcp_fast_finwait2_recycle ? 9561 tcp_finwait2_timeout : 9562 TP_MAXIDLE(tp))); 9563 } 9564 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9565 } 9566 if (sbavail(&so->so_snd)) { 9567 if (ctf_progress_timeout_check(tp, true)) { 9568 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9569 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9570 return (1); 9571 } 9572 } 9573 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9574 tiwin, thflags, nxt_pkt)); 9575 } 9576 9577 /* 9578 * Return value of 1, the TCB is unlocked and most 9579 * likely gone, return value of 0, the TCB is still 9580 * locked. 9581 */ 9582 static int 9583 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9584 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9585 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9586 { 9587 int32_t ourfinisacked = 0; 9588 int32_t ret_val; 9589 struct tcp_bbr *bbr; 9590 9591 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9592 ctf_calc_rwin(so, tp); 9593 if ((thflags & TH_RST) || 9594 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9595 return (ctf_process_rst(m, th, so, tp)); 9596 /* 9597 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9598 * synchronized state. 9599 */ 9600 if (thflags & TH_SYN) { 9601 ctf_challenge_ack(m, th, tp, &ret_val); 9602 return (ret_val); 9603 } 9604 /* 9605 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9606 * it's less than ts_recent, drop it. 9607 */ 9608 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9609 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9610 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9611 return (ret_val); 9612 } 9613 INP_WLOCK_ASSERT(tp->t_inpcb); 9614 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9615 return (ret_val); 9616 } 9617 /* 9618 * If new data are received on a connection after the user processes 9619 * are gone, then RST the other end. 9620 * We call a new function now so we might continue and setup 9621 * to reset at all data being ack'd. 9622 */ 9623 if ((tp->t_flags & TF_CLOSED) && tlen && 9624 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9625 return (1); 9626 /* 9627 * If last ACK falls within this segment's sequence numbers, record 9628 * its timestamp. NOTE: 1) That the test incorporates suggestions 9629 * from the latest proposal of the tcplw@cray.com list (Braden 9630 * 1993/04/26). 2) That updating only on newer timestamps interferes 9631 * with our earlier PAWS tests, so this check should be solely 9632 * predicated on the sequence space of this segment. 3) That we 9633 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9634 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9635 * SEG.Len, This modified check allows us to overcome RFC1323's 9636 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9637 * p.869. In such cases, we can still calculate the RTT correctly 9638 * when RCV.NXT == Last.ACK.Sent. 9639 */ 9640 if ((to->to_flags & TOF_TS) != 0 && 9641 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9642 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9643 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9644 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9645 tp->ts_recent = to->to_tsval; 9646 } 9647 /* 9648 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9649 * is on (half-synchronized state), then queue data for later 9650 * processing; else drop segment and return. 9651 */ 9652 if ((thflags & TH_ACK) == 0) { 9653 if (tp->t_flags & TF_NEEDSYN) { 9654 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9655 tiwin, thflags, nxt_pkt)); 9656 } else if (tp->t_flags & TF_ACKNOW) { 9657 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9658 bbr->r_wanted_output = 1; 9659 return (ret_val); 9660 } else { 9661 ctf_do_drop(m, NULL); 9662 return (0); 9663 } 9664 } 9665 /* 9666 * Ack processing. 9667 */ 9668 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9669 return (ret_val); 9670 } 9671 if (ourfinisacked) { 9672 tcp_twstart(tp); 9673 m_freem(m); 9674 return (1); 9675 } 9676 if (sbavail(&so->so_snd)) { 9677 if (ctf_progress_timeout_check(tp, true)) { 9678 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9679 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9680 return (1); 9681 } 9682 } 9683 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9684 tiwin, thflags, nxt_pkt)); 9685 } 9686 9687 /* 9688 * Return value of 1, the TCB is unlocked and most 9689 * likely gone, return value of 0, the TCB is still 9690 * locked. 9691 */ 9692 static int 9693 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9694 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9695 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9696 { 9697 int32_t ourfinisacked = 0; 9698 int32_t ret_val; 9699 struct tcp_bbr *bbr; 9700 9701 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9702 ctf_calc_rwin(so, tp); 9703 if ((thflags & TH_RST) || 9704 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9705 return (ctf_process_rst(m, th, so, tp)); 9706 /* 9707 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9708 * synchronized state. 9709 */ 9710 if (thflags & TH_SYN) { 9711 ctf_challenge_ack(m, th, tp, &ret_val); 9712 return (ret_val); 9713 } 9714 /* 9715 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9716 * it's less than ts_recent, drop it. 9717 */ 9718 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9719 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9720 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9721 return (ret_val); 9722 } 9723 INP_WLOCK_ASSERT(tp->t_inpcb); 9724 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9725 return (ret_val); 9726 } 9727 /* 9728 * If new data are received on a connection after the user processes 9729 * are gone, then RST the other end. 9730 * We call a new function now so we might continue and setup 9731 * to reset at all data being ack'd. 9732 */ 9733 if ((tp->t_flags & TF_CLOSED) && tlen && 9734 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9735 return (1); 9736 /* 9737 * If last ACK falls within this segment's sequence numbers, record 9738 * its timestamp. NOTE: 1) That the test incorporates suggestions 9739 * from the latest proposal of the tcplw@cray.com list (Braden 9740 * 1993/04/26). 2) That updating only on newer timestamps interferes 9741 * with our earlier PAWS tests, so this check should be solely 9742 * predicated on the sequence space of this segment. 3) That we 9743 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9744 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9745 * SEG.Len, This modified check allows us to overcome RFC1323's 9746 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9747 * p.869. In such cases, we can still calculate the RTT correctly 9748 * when RCV.NXT == Last.ACK.Sent. 9749 */ 9750 if ((to->to_flags & TOF_TS) != 0 && 9751 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9752 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9753 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9754 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9755 tp->ts_recent = to->to_tsval; 9756 } 9757 /* 9758 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9759 * is on (half-synchronized state), then queue data for later 9760 * processing; else drop segment and return. 9761 */ 9762 if ((thflags & TH_ACK) == 0) { 9763 if (tp->t_flags & TF_NEEDSYN) { 9764 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9765 tiwin, thflags, nxt_pkt)); 9766 } else if (tp->t_flags & TF_ACKNOW) { 9767 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9768 bbr->r_wanted_output = 1; 9769 return (ret_val); 9770 } else { 9771 ctf_do_drop(m, NULL); 9772 return (0); 9773 } 9774 } 9775 /* 9776 * case TCPS_LAST_ACK: Ack processing. 9777 */ 9778 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9779 return (ret_val); 9780 } 9781 if (ourfinisacked) { 9782 tp = tcp_close(tp); 9783 ctf_do_drop(m, tp); 9784 return (1); 9785 } 9786 if (sbavail(&so->so_snd)) { 9787 if (ctf_progress_timeout_check(tp, true)) { 9788 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9789 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9790 return (1); 9791 } 9792 } 9793 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9794 tiwin, thflags, nxt_pkt)); 9795 } 9796 9797 /* 9798 * Return value of 1, the TCB is unlocked and most 9799 * likely gone, return value of 0, the TCB is still 9800 * locked. 9801 */ 9802 static int 9803 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9804 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9805 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9806 { 9807 int32_t ourfinisacked = 0; 9808 int32_t ret_val; 9809 struct tcp_bbr *bbr; 9810 9811 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9812 ctf_calc_rwin(so, tp); 9813 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9814 if ((thflags & TH_RST) || 9815 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9816 return (ctf_process_rst(m, th, so, tp)); 9817 9818 /* 9819 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9820 * synchronized state. 9821 */ 9822 if (thflags & TH_SYN) { 9823 ctf_challenge_ack(m, th, tp, &ret_val); 9824 return (ret_val); 9825 } 9826 INP_WLOCK_ASSERT(tp->t_inpcb); 9827 /* 9828 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9829 * it's less than ts_recent, drop it. 9830 */ 9831 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9832 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9833 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9834 return (ret_val); 9835 } 9836 INP_WLOCK_ASSERT(tp->t_inpcb); 9837 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9838 return (ret_val); 9839 } 9840 /* 9841 * If new data are received on a connection after the user processes 9842 * are gone, then we may RST the other end depending on the outcome 9843 * of bbr_check_data_after_close. 9844 * We call a new function now so we might continue and setup 9845 * to reset at all data being ack'd. 9846 */ 9847 if ((tp->t_flags & TF_CLOSED) && tlen && 9848 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9849 return (1); 9850 /* 9851 * If last ACK falls within this segment's sequence numbers, record 9852 * its timestamp. NOTE: 1) That the test incorporates suggestions 9853 * from the latest proposal of the tcplw@cray.com list (Braden 9854 * 1993/04/26). 2) That updating only on newer timestamps interferes 9855 * with our earlier PAWS tests, so this check should be solely 9856 * predicated on the sequence space of this segment. 3) That we 9857 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9858 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9859 * SEG.Len, This modified check allows us to overcome RFC1323's 9860 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9861 * p.869. In such cases, we can still calculate the RTT correctly 9862 * when RCV.NXT == Last.ACK.Sent. 9863 */ 9864 INP_WLOCK_ASSERT(tp->t_inpcb); 9865 if ((to->to_flags & TOF_TS) != 0 && 9866 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9867 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9868 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9869 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9870 tp->ts_recent = to->to_tsval; 9871 } 9872 /* 9873 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9874 * is on (half-synchronized state), then queue data for later 9875 * processing; else drop segment and return. 9876 */ 9877 if ((thflags & TH_ACK) == 0) { 9878 if (tp->t_flags & TF_NEEDSYN) { 9879 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9880 tiwin, thflags, nxt_pkt)); 9881 } else if (tp->t_flags & TF_ACKNOW) { 9882 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9883 bbr->r_wanted_output = 1; 9884 return (ret_val); 9885 } else { 9886 ctf_do_drop(m, NULL); 9887 return (0); 9888 } 9889 } 9890 /* 9891 * Ack processing. 9892 */ 9893 INP_WLOCK_ASSERT(tp->t_inpcb); 9894 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9895 return (ret_val); 9896 } 9897 if (sbavail(&so->so_snd)) { 9898 if (ctf_progress_timeout_check(tp, true)) { 9899 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9900 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9901 return (1); 9902 } 9903 } 9904 INP_WLOCK_ASSERT(tp->t_inpcb); 9905 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9906 tiwin, thflags, nxt_pkt)); 9907 } 9908 9909 static void 9910 bbr_stop_all_timers(struct tcpcb *tp) 9911 { 9912 struct tcp_bbr *bbr; 9913 9914 /* 9915 * Assure no timers are running. 9916 */ 9917 if (tcp_timer_active(tp, TT_PERSIST)) { 9918 /* We enter in persists, set the flag appropriately */ 9919 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9920 bbr->rc_in_persist = 1; 9921 } 9922 tcp_timer_suspend(tp, TT_PERSIST); 9923 tcp_timer_suspend(tp, TT_REXMT); 9924 tcp_timer_suspend(tp, TT_KEEP); 9925 tcp_timer_suspend(tp, TT_DELACK); 9926 } 9927 9928 static void 9929 bbr_google_mode_on(struct tcp_bbr *bbr) 9930 { 9931 bbr->rc_use_google = 1; 9932 bbr->rc_no_pacing = 0; 9933 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9934 bbr->r_use_policer = bbr_policer_detection_enabled; 9935 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9936 bbr->bbr_use_rack_cheat = 0; 9937 bbr->r_ctl.rc_incr_tmrs = 0; 9938 bbr->r_ctl.rc_inc_tcp_oh = 0; 9939 bbr->r_ctl.rc_inc_ip_oh = 0; 9940 bbr->r_ctl.rc_inc_enet_oh = 0; 9941 reset_time(&bbr->r_ctl.rc_delrate, 9942 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9943 reset_time_small(&bbr->r_ctl.rc_rttprop, 9944 (11 * USECS_IN_SECOND)); 9945 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9946 } 9947 9948 static void 9949 bbr_google_mode_off(struct tcp_bbr *bbr) 9950 { 9951 bbr->rc_use_google = 0; 9952 bbr->r_ctl.bbr_google_discount = 0; 9953 bbr->no_pacing_until = bbr_no_pacing_until; 9954 bbr->r_use_policer = 0; 9955 if (bbr->no_pacing_until) 9956 bbr->rc_no_pacing = 1; 9957 else 9958 bbr->rc_no_pacing = 0; 9959 if (bbr_use_rack_resend_cheat) 9960 bbr->bbr_use_rack_cheat = 1; 9961 else 9962 bbr->bbr_use_rack_cheat = 0; 9963 if (bbr_incr_timers) 9964 bbr->r_ctl.rc_incr_tmrs = 1; 9965 else 9966 bbr->r_ctl.rc_incr_tmrs = 0; 9967 if (bbr_include_tcp_oh) 9968 bbr->r_ctl.rc_inc_tcp_oh = 1; 9969 else 9970 bbr->r_ctl.rc_inc_tcp_oh = 0; 9971 if (bbr_include_ip_oh) 9972 bbr->r_ctl.rc_inc_ip_oh = 1; 9973 else 9974 bbr->r_ctl.rc_inc_ip_oh = 0; 9975 if (bbr_include_enet_oh) 9976 bbr->r_ctl.rc_inc_enet_oh = 1; 9977 else 9978 bbr->r_ctl.rc_inc_enet_oh = 0; 9979 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9980 reset_time(&bbr->r_ctl.rc_delrate, 9981 bbr_num_pktepo_for_del_limit); 9982 reset_time_small(&bbr->r_ctl.rc_rttprop, 9983 (bbr_filter_len_sec * USECS_IN_SECOND)); 9984 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9985 } 9986 /* 9987 * Return 0 on success, non-zero on failure 9988 * which indicates the error (usually no memory). 9989 */ 9990 static int 9991 bbr_init(struct tcpcb *tp) 9992 { 9993 struct tcp_bbr *bbr = NULL; 9994 struct inpcb *inp; 9995 uint32_t cts; 9996 9997 tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 9998 if (tp->t_fb_ptr == NULL) { 9999 /* 10000 * We need to allocate memory but cant. The INP and INP_INFO 10001 * locks and they are recursive (happens during setup. So a 10002 * scheme to drop the locks fails :( 10003 * 10004 */ 10005 return (ENOMEM); 10006 } 10007 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10008 bbr->rtt_valid = 0; 10009 inp = tp->t_inpcb; 10010 inp->inp_flags2 |= INP_CANNOT_DO_ECN; 10011 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; 10012 TAILQ_INIT(&bbr->r_ctl.rc_map); 10013 TAILQ_INIT(&bbr->r_ctl.rc_free); 10014 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 10015 bbr->rc_tp = tp; 10016 if (tp->t_inpcb) { 10017 bbr->rc_inp = tp->t_inpcb; 10018 } 10019 cts = tcp_get_usecs(&bbr->rc_tv); 10020 tp->t_acktime = 0; 10021 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 10022 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 10023 bbr->rc_tlp_threshold = bbr_tlp_thresh; 10024 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 10025 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 10026 bbr->r_ctl.rc_min_to = bbr_min_to; 10027 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10028 bbr->r_ctl.bbr_lost_at_state = 0; 10029 bbr->r_ctl.rc_lost_at_startup = 0; 10030 bbr->rc_all_timers_stopped = 0; 10031 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 10032 bbr->r_ctl.rc_pkt_epoch_del = 0; 10033 bbr->r_ctl.rc_pkt_epoch = 0; 10034 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 10035 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 10036 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 10037 bbr->r_ctl.rc_went_idle_time = cts; 10038 bbr->rc_pacer_started = cts; 10039 bbr->r_ctl.rc_pkt_epoch_time = cts; 10040 bbr->r_ctl.rc_rcvtime = cts; 10041 bbr->r_ctl.rc_bbr_state_time = cts; 10042 bbr->r_ctl.rc_del_time = cts; 10043 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 10044 bbr->r_ctl.last_in_probertt = cts; 10045 bbr->skip_gain = 0; 10046 bbr->gain_is_limited = 0; 10047 bbr->no_pacing_until = bbr_no_pacing_until; 10048 if (bbr->no_pacing_until) 10049 bbr->rc_no_pacing = 1; 10050 if (bbr_use_google_algo) { 10051 bbr->rc_no_pacing = 0; 10052 bbr->rc_use_google = 1; 10053 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 10054 bbr->r_use_policer = bbr_policer_detection_enabled; 10055 } else { 10056 bbr->rc_use_google = 0; 10057 bbr->r_ctl.bbr_google_discount = 0; 10058 bbr->r_use_policer = 0; 10059 } 10060 if (bbr_ts_limiting) 10061 bbr->rc_use_ts_limit = 1; 10062 else 10063 bbr->rc_use_ts_limit = 0; 10064 if (bbr_ts_can_raise) 10065 bbr->ts_can_raise = 1; 10066 else 10067 bbr->ts_can_raise = 0; 10068 if (V_tcp_delack_enabled == 1) 10069 tp->t_delayed_ack = 2; 10070 else if (V_tcp_delack_enabled == 0) 10071 tp->t_delayed_ack = 0; 10072 else if (V_tcp_delack_enabled < 100) 10073 tp->t_delayed_ack = V_tcp_delack_enabled; 10074 else 10075 tp->t_delayed_ack = 2; 10076 if (bbr->rc_use_google == 0) 10077 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10078 else 10079 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10080 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10081 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10082 bbr->rc_init_win = bbr_def_init_win; 10083 if (tp->t_flags & TF_REQ_TSTMP) 10084 bbr->rc_last_options = TCP_TS_OVERHEAD; 10085 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10086 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10087 bbr->r_init_rtt = 1; 10088 10089 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10090 if (bbr_allow_hdwr_pacing) 10091 bbr->bbr_hdw_pace_ena = 1; 10092 else 10093 bbr->bbr_hdw_pace_ena = 0; 10094 if (bbr_sends_full_iwnd) 10095 bbr->bbr_init_win_cheat = 1; 10096 else 10097 bbr->bbr_init_win_cheat = 0; 10098 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10099 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10100 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10101 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10102 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10103 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10104 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10105 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10106 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10107 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10108 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10109 bbr->r_ctl.rc_rtt_shrinks = cts; 10110 if (bbr->rc_use_google) { 10111 setup_time_filter(&bbr->r_ctl.rc_delrate, 10112 FILTER_TYPE_MAX, 10113 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10114 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10115 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10116 } else { 10117 setup_time_filter(&bbr->r_ctl.rc_delrate, 10118 FILTER_TYPE_MAX, 10119 bbr_num_pktepo_for_del_limit); 10120 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10121 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10122 } 10123 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10124 if (bbr_uses_idle_restart) 10125 bbr->rc_use_idle_restart = 1; 10126 else 10127 bbr->rc_use_idle_restart = 0; 10128 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10129 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10130 if (bbr_resends_use_tso) 10131 bbr->rc_resends_use_tso = 1; 10132 #ifdef NETFLIX_PEAKRATE 10133 tp->t_peakrate_thr = tp->t_maxpeakrate; 10134 #endif 10135 if (tp->snd_una != tp->snd_max) { 10136 /* Create a send map for the current outstanding data */ 10137 struct bbr_sendmap *rsm; 10138 10139 rsm = bbr_alloc(bbr); 10140 if (rsm == NULL) { 10141 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10142 tp->t_fb_ptr = NULL; 10143 return (ENOMEM); 10144 } 10145 rsm->r_rtt_not_allowed = 1; 10146 rsm->r_tim_lastsent[0] = cts; 10147 rsm->r_rtr_cnt = 1; 10148 rsm->r_rtr_bytes = 0; 10149 rsm->r_start = tp->snd_una; 10150 rsm->r_end = tp->snd_max; 10151 rsm->r_dupack = 0; 10152 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10153 rsm->r_ts_valid = 0; 10154 rsm->r_del_ack_ts = tp->ts_recent; 10155 rsm->r_del_time = cts; 10156 if (bbr->r_ctl.r_app_limited_until) 10157 rsm->r_app_limited = 1; 10158 else 10159 rsm->r_app_limited = 0; 10160 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10161 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10162 rsm->r_in_tmap = 1; 10163 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10164 rsm->r_bbr_state = bbr_state_val(bbr); 10165 else 10166 rsm->r_bbr_state = 8; 10167 } 10168 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10169 bbr->bbr_use_rack_cheat = 1; 10170 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10171 bbr->r_ctl.rc_incr_tmrs = 1; 10172 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10173 bbr->r_ctl.rc_inc_tcp_oh = 1; 10174 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10175 bbr->r_ctl.rc_inc_ip_oh = 1; 10176 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10177 bbr->r_ctl.rc_inc_enet_oh = 1; 10178 10179 bbr_log_type_statechange(bbr, cts, __LINE__); 10180 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10181 (tp->t_srtt)) { 10182 uint32_t rtt; 10183 10184 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10185 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10186 } 10187 /* announce the settings and state */ 10188 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10189 tcp_bbr_tso_size_check(bbr, cts); 10190 /* 10191 * Now call the generic function to start a timer. This will place 10192 * the TCB on the hptsi wheel if a timer is needed with appropriate 10193 * flags. 10194 */ 10195 bbr_stop_all_timers(tp); 10196 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10197 return (0); 10198 } 10199 10200 /* 10201 * Return 0 if we can accept the connection. Return 10202 * non-zero if we can't handle the connection. A EAGAIN 10203 * means you need to wait until the connection is up. 10204 * a EADDRNOTAVAIL means we can never handle the connection 10205 * (no SACK). 10206 */ 10207 static int 10208 bbr_handoff_ok(struct tcpcb *tp) 10209 { 10210 if ((tp->t_state == TCPS_CLOSED) || 10211 (tp->t_state == TCPS_LISTEN)) { 10212 /* Sure no problem though it may not stick */ 10213 return (0); 10214 } 10215 if ((tp->t_state == TCPS_SYN_SENT) || 10216 (tp->t_state == TCPS_SYN_RECEIVED)) { 10217 /* 10218 * We really don't know you have to get to ESTAB or beyond 10219 * to tell. 10220 */ 10221 return (EAGAIN); 10222 } 10223 if (tp->t_flags & TF_SENTFIN) 10224 return (EINVAL); 10225 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10226 return (0); 10227 } 10228 /* 10229 * If we reach here we don't do SACK on this connection so we can 10230 * never do rack. 10231 */ 10232 return (EINVAL); 10233 } 10234 10235 static void 10236 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10237 { 10238 if (tp->t_fb_ptr) { 10239 uint32_t calc; 10240 struct tcp_bbr *bbr; 10241 struct bbr_sendmap *rsm; 10242 10243 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10244 if (bbr->r_ctl.crte) 10245 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10246 bbr_log_flowend(bbr); 10247 bbr->rc_tp = NULL; 10248 if (tp->t_inpcb) { 10249 /* Backout any flags2 we applied */ 10250 tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN; 10251 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; 10252 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 10253 } 10254 if (bbr->bbr_hdrw_pacing) 10255 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10256 else 10257 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10258 if (bbr->r_ctl.crte != NULL) { 10259 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10260 bbr->r_ctl.crte = NULL; 10261 } 10262 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10263 while (rsm) { 10264 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10265 uma_zfree(bbr_zone, rsm); 10266 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10267 } 10268 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10269 while (rsm) { 10270 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10271 uma_zfree(bbr_zone, rsm); 10272 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10273 } 10274 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10275 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10276 BBR_STAT_INC(bbr_dynamic_rwnd); 10277 else 10278 BBR_STAT_INC(bbr_static_rwnd); 10279 bbr->r_ctl.rc_free_cnt = 0; 10280 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10281 tp->t_fb_ptr = NULL; 10282 } 10283 /* Make sure snd_nxt is correctly set */ 10284 tp->snd_nxt = tp->snd_max; 10285 } 10286 10287 static void 10288 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10289 { 10290 switch (tp->t_state) { 10291 case TCPS_SYN_SENT: 10292 bbr->r_state = TCPS_SYN_SENT; 10293 bbr->r_substate = bbr_do_syn_sent; 10294 break; 10295 case TCPS_SYN_RECEIVED: 10296 bbr->r_state = TCPS_SYN_RECEIVED; 10297 bbr->r_substate = bbr_do_syn_recv; 10298 break; 10299 case TCPS_ESTABLISHED: 10300 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10301 bbr->r_state = TCPS_ESTABLISHED; 10302 bbr->r_substate = bbr_do_established; 10303 break; 10304 case TCPS_CLOSE_WAIT: 10305 bbr->r_state = TCPS_CLOSE_WAIT; 10306 bbr->r_substate = bbr_do_close_wait; 10307 break; 10308 case TCPS_FIN_WAIT_1: 10309 bbr->r_state = TCPS_FIN_WAIT_1; 10310 bbr->r_substate = bbr_do_fin_wait_1; 10311 break; 10312 case TCPS_CLOSING: 10313 bbr->r_state = TCPS_CLOSING; 10314 bbr->r_substate = bbr_do_closing; 10315 break; 10316 case TCPS_LAST_ACK: 10317 bbr->r_state = TCPS_LAST_ACK; 10318 bbr->r_substate = bbr_do_lastack; 10319 break; 10320 case TCPS_FIN_WAIT_2: 10321 bbr->r_state = TCPS_FIN_WAIT_2; 10322 bbr->r_substate = bbr_do_fin_wait_2; 10323 break; 10324 case TCPS_LISTEN: 10325 case TCPS_CLOSED: 10326 case TCPS_TIME_WAIT: 10327 default: 10328 break; 10329 }; 10330 } 10331 10332 static void 10333 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10334 { 10335 /* 10336 * Now what state are we going into now? Is there adjustments 10337 * needed? 10338 */ 10339 int32_t old_state; 10340 10341 old_state = bbr_state_val(bbr); 10342 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10343 /* Save the lowest srtt we saw in our end of the sub-state */ 10344 bbr->rc_hit_state_1 = 0; 10345 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10346 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10347 } 10348 bbr->rc_bbr_substate++; 10349 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10350 /* Cycle back to first state-> gain */ 10351 bbr->rc_bbr_substate = 0; 10352 } 10353 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10354 /* 10355 * We enter the gain(5/4) cycle (possibly less if 10356 * shallow buffer detection is enabled) 10357 */ 10358 if (bbr->skip_gain) { 10359 /* 10360 * Hardware pacing has set our rate to 10361 * the max and limited our b/w just 10362 * do level i.e. no gain. 10363 */ 10364 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10365 } else if (bbr->gain_is_limited && 10366 bbr->bbr_hdrw_pacing && 10367 bbr->r_ctl.crte) { 10368 /* 10369 * We can't gain above the hardware pacing 10370 * rate which is less than our rate + the gain 10371 * calculate the gain needed to reach the hardware 10372 * pacing rate.. 10373 */ 10374 uint64_t bw, rate, gain_calc; 10375 10376 bw = bbr_get_bw(bbr); 10377 rate = bbr->r_ctl.crte->rate; 10378 if ((rate > bw) && 10379 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10380 gain_calc = (rate * BBR_UNIT) / bw; 10381 if (gain_calc < BBR_UNIT) 10382 gain_calc = BBR_UNIT; 10383 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10384 } else { 10385 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10386 } 10387 } else 10388 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10389 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10390 bbr->r_ctl.rc_bbr_state_atflight = cts; 10391 } else 10392 bbr->r_ctl.rc_bbr_state_atflight = 0; 10393 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10394 bbr->rc_hit_state_1 = 1; 10395 bbr->r_ctl.rc_exta_time_gd = 0; 10396 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10397 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10398 if (bbr_state_drain_2_tar) { 10399 bbr->r_ctl.rc_bbr_state_atflight = 0; 10400 } else 10401 bbr->r_ctl.rc_bbr_state_atflight = cts; 10402 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10403 } else { 10404 /* All other cycles hit here 2-7 */ 10405 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10406 if (bbr_sub_drain_slam_cwnd && 10407 (bbr->rc_use_google == 0) && 10408 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10409 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10410 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10411 } 10412 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10413 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10414 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10415 else 10416 bbr->r_ctl.rc_exta_time_gd = 0; 10417 if (bbr->r_ctl.rc_exta_time_gd) { 10418 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10419 /* Now chop up the time for each state (div by 7) */ 10420 bbr->r_ctl.rc_level_state_extra /= 7; 10421 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10422 /* Add a randomization */ 10423 bbr_randomize_extra_state_time(bbr); 10424 } 10425 } 10426 } 10427 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10428 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10429 } 10430 if (bbr->rc_use_google) { 10431 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10432 } 10433 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10434 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10435 if (dolog) 10436 bbr_log_type_statechange(bbr, cts, line); 10437 10438 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10439 uint32_t time_in; 10440 10441 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10442 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10443 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10444 } else { 10445 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10446 } 10447 } 10448 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10449 bbr_set_state_target(bbr, __LINE__); 10450 if (bbr_sub_drain_slam_cwnd && 10451 (bbr->rc_use_google == 0) && 10452 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10453 /* Slam down the cwnd */ 10454 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10455 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10456 if (bbr_sub_drain_app_limit) { 10457 /* Go app limited if we are on a long drain */ 10458 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10459 ctf_flight_size(bbr->rc_tp, 10460 (bbr->r_ctl.rc_sacked + 10461 bbr->r_ctl.rc_lost_bytes))); 10462 } 10463 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10464 } 10465 if (bbr->rc_lt_use_bw) { 10466 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10467 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10468 } 10469 /* Google changes TSO size every cycle */ 10470 if (bbr->rc_use_google) 10471 tcp_bbr_tso_size_check(bbr, cts); 10472 bbr->r_ctl.gain_epoch = cts; 10473 bbr->r_ctl.rc_bbr_state_time = cts; 10474 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10475 } 10476 10477 static void 10478 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10479 { 10480 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10481 (google_allow_early_out == 1) && 10482 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10483 /* We have reached out target flight size possibly early */ 10484 goto change_state; 10485 } 10486 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10487 return; 10488 } 10489 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10490 /* 10491 * Must be a rttProp movement forward before 10492 * we can change states. 10493 */ 10494 return; 10495 } 10496 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10497 /* 10498 * The needed time has passed but for 10499 * the gain cycle extra rules apply: 10500 * 1) If we have seen loss, we exit 10501 * 2) If we have not reached the target 10502 * we stay in GAIN (gain-to-target). 10503 */ 10504 if (google_consider_lost && losses) 10505 goto change_state; 10506 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10507 return; 10508 } 10509 } 10510 change_state: 10511 /* For gain we must reach our target, all others last 1 rttProp */ 10512 bbr_substate_change(bbr, cts, __LINE__, 1); 10513 } 10514 10515 static void 10516 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10517 { 10518 uint32_t flight, bbr_cur_cycle_time; 10519 10520 if (bbr->rc_use_google) { 10521 bbr_set_probebw_google_gains(bbr, cts, losses); 10522 return; 10523 } 10524 if (cts == 0) { 10525 /* 10526 * Never alow cts to be 0 we 10527 * do this so we can judge if 10528 * we have set a timestamp. 10529 */ 10530 cts = 1; 10531 } 10532 if (bbr_state_is_pkt_epoch) 10533 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10534 else 10535 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10536 10537 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10538 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10539 flight = ctf_flight_size(bbr->rc_tp, 10540 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10541 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10542 /* Keep it slam down */ 10543 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10544 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10545 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10546 } 10547 if (bbr_sub_drain_app_limit) { 10548 /* Go app limited if we are on a long drain */ 10549 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10550 } 10551 } 10552 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10553 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10554 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10555 /* 10556 * Still here after the same time as 10557 * the gain. We need to drain harder 10558 * for the next srtt. Reduce by a set amount 10559 * the gain drop is capped at DRAIN states 10560 * value (88). 10561 */ 10562 bbr->r_ctl.flightsize_at_drain = flight; 10563 if (bbr_drain_drop_mul && 10564 bbr_drain_drop_div && 10565 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10566 /* Use your specific drop value (def 4/5 = 20%) */ 10567 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10568 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10569 } else { 10570 /* You get drop of 20% */ 10571 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10572 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10573 } 10574 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10575 /* Reduce our gain again to the bottom */ 10576 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10577 } 10578 bbr_log_exit_gain(bbr, cts, 4); 10579 /* 10580 * Extend out so we wait another 10581 * epoch before dropping again. 10582 */ 10583 bbr->r_ctl.gain_epoch = cts; 10584 } 10585 if (flight <= bbr->r_ctl.rc_target_at_state) { 10586 if (bbr_sub_drain_slam_cwnd && 10587 (bbr->rc_use_google == 0) && 10588 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10589 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10590 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10591 } 10592 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10593 bbr_log_exit_gain(bbr, cts, 3); 10594 } 10595 } else { 10596 /* Its a gain */ 10597 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10598 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10599 goto change_state; 10600 } 10601 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10602 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10603 bbr->rc_tp->snd_wnd)) { 10604 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10605 bbr_log_exit_gain(bbr, cts, 2); 10606 } 10607 } 10608 /** 10609 * We fall through and return always one of two things has 10610 * occurred. 10611 * 1) We are still not at target 10612 * <or> 10613 * 2) We reached the target and set rc_bbr_state_atflight 10614 * which means we no longer hit this block 10615 * next time we are called. 10616 */ 10617 return; 10618 } 10619 change_state: 10620 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10621 return; 10622 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10623 /* Less than a full time-period has passed */ 10624 return; 10625 } 10626 if (bbr->r_ctl.rc_level_state_extra && 10627 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10628 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10629 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10630 /* Less than a full time-period + extra has passed */ 10631 return; 10632 } 10633 if (bbr_gain_gets_extra_too && 10634 bbr->r_ctl.rc_level_state_extra && 10635 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10636 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10637 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10638 /* Less than a full time-period + extra has passed */ 10639 return; 10640 } 10641 bbr_substate_change(bbr, cts, __LINE__, 1); 10642 } 10643 10644 static uint32_t 10645 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10646 { 10647 uint32_t mss, tar; 10648 10649 if (bbr->rc_use_google) { 10650 /* Google just uses the cwnd target */ 10651 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10652 } else { 10653 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10654 bbr->r_ctl.rc_pace_max_segs); 10655 /* Get the base cwnd with gain rounded to a mss */ 10656 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10657 gain), mss); 10658 /* Make sure it is within our min */ 10659 if (tar < get_min_cwnd(bbr)) 10660 return (get_min_cwnd(bbr)); 10661 } 10662 return (tar); 10663 } 10664 10665 static void 10666 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10667 { 10668 uint32_t tar, meth; 10669 10670 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10671 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10672 /* Special case using old probe-rtt method */ 10673 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10674 meth = 1; 10675 } else { 10676 /* Non-probe-rtt case and reduced probe-rtt */ 10677 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10678 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10679 /* For gain cycle we use the hptsi gain */ 10680 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10681 meth = 2; 10682 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10683 /* 10684 * If configured, or for google all other states 10685 * get BBR_UNIT. 10686 */ 10687 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10688 meth = 3; 10689 } else { 10690 /* 10691 * Or we set a target based on the pacing gain 10692 * for non-google mode and default (non-configured). 10693 * Note we don't set a target goal below drain (192). 10694 */ 10695 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10696 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10697 meth = 4; 10698 } else { 10699 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10700 meth = 5; 10701 } 10702 } 10703 } 10704 bbr_log_set_of_state_target(bbr, tar, line, meth); 10705 bbr->r_ctl.rc_target_at_state = tar; 10706 } 10707 10708 static void 10709 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10710 { 10711 /* Change to probe_rtt */ 10712 uint32_t time_in; 10713 10714 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10715 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10716 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10717 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10718 + bbr->r_ctl.rc_delivered); 10719 /* Setup so we force feed the filter */ 10720 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10721 bbr->rc_prtt_set_ts = 1; 10722 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10723 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10724 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10725 } 10726 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10727 bbr->r_ctl.rc_rtt_shrinks = cts; 10728 bbr->r_ctl.last_in_probertt = cts; 10729 bbr->r_ctl.rc_probertt_srttchktim = cts; 10730 bbr->r_ctl.rc_bbr_state_time = cts; 10731 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10732 /* We need to force the filter to update */ 10733 10734 if ((bbr_sub_drain_slam_cwnd) && 10735 bbr->rc_hit_state_1 && 10736 (bbr->rc_use_google == 0) && 10737 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10738 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10739 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10740 } else 10741 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10742 /* Update the lost */ 10743 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10744 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10745 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10746 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10747 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10748 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10749 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10750 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10751 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10752 } else { 10753 /* 10754 * We bring it down slowly by using a hptsi gain that is 10755 * probably 75%. This will slowly float down our outstanding 10756 * without tampering with the cwnd. 10757 */ 10758 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10759 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10760 bbr_set_state_target(bbr, __LINE__); 10761 if (bbr_prtt_slam_cwnd && 10762 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10763 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10764 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10765 } 10766 } 10767 if (ctf_flight_size(bbr->rc_tp, 10768 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10769 bbr->r_ctl.rc_target_at_state) { 10770 /* We are at target */ 10771 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10772 } else { 10773 /* We need to come down to reach target before our time begins */ 10774 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10775 } 10776 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10777 BBR_STAT_INC(bbr_enter_probertt); 10778 bbr_log_exit_gain(bbr, cts, 0); 10779 bbr_log_type_statechange(bbr, cts, line); 10780 } 10781 10782 static void 10783 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10784 { 10785 /* 10786 * Sanity check on probe-rtt intervals. 10787 * In crazy situations where we are competing 10788 * against new-reno flows with huge buffers 10789 * our rtt-prop interval could come to dominate 10790 * things if we can't get through a full set 10791 * of cycles, we need to adjust it. 10792 */ 10793 if (bbr_can_adjust_probertt && 10794 (bbr->rc_use_google == 0)) { 10795 uint16_t val = 0; 10796 uint32_t cur_rttp, fval, newval, baseval; 10797 10798 /* Are we to small and go into probe-rtt to often? */ 10799 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10800 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10801 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10802 if (bbr_is_ratio == 0) { 10803 if (fval > bbr_rtt_probe_limit) 10804 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10805 else 10806 newval = cur_rttp; 10807 } else { 10808 int mul; 10809 10810 mul = fval / bbr_rtt_probe_limit; 10811 newval = cur_rttp * mul; 10812 } 10813 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10814 bbr->r_ctl.rc_probertt_int = cur_rttp; 10815 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10816 val = 1; 10817 } else { 10818 /* 10819 * No adjustments were made 10820 * do we need to shrink it? 10821 */ 10822 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10823 if (cur_rttp <= bbr_rtt_probe_limit) { 10824 /* 10825 * Things have calmed down lets 10826 * shrink all the way to default 10827 */ 10828 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10829 reset_time_small(&bbr->r_ctl.rc_rttprop, 10830 (bbr_filter_len_sec * USECS_IN_SECOND)); 10831 cur_rttp = bbr_rtt_probe_limit; 10832 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10833 val = 2; 10834 } else { 10835 /* 10836 * Well does some adjustment make sense? 10837 */ 10838 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10839 /* We can reduce interval time some */ 10840 bbr->r_ctl.rc_probertt_int = cur_rttp; 10841 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10842 val = 3; 10843 } 10844 } 10845 } 10846 } 10847 if (val) 10848 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10849 } 10850 } 10851 10852 static void 10853 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10854 { 10855 /* Exit probe-rtt */ 10856 10857 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10858 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10859 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10860 } 10861 bbr_log_exit_gain(bbr, cts, 1); 10862 bbr->rc_hit_state_1 = 0; 10863 bbr->r_ctl.rc_rtt_shrinks = cts; 10864 bbr->r_ctl.last_in_probertt = cts; 10865 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10866 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10867 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10868 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10869 bbr->r_ctl.rc_delivered); 10870 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10871 uint32_t time_in; 10872 10873 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10874 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10875 } 10876 if (bbr->rc_filled_pipe) { 10877 /* Switch to probe_bw */ 10878 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10879 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10880 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10881 bbr_substate_change(bbr, cts, __LINE__, 0); 10882 bbr_log_type_statechange(bbr, cts, __LINE__); 10883 } else { 10884 /* Back to startup */ 10885 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10886 bbr->r_ctl.rc_bbr_state_time = cts; 10887 /* 10888 * We don't want to give a complete free 3 10889 * measurements until we exit, so we use 10890 * the number of pe's we were in probe-rtt 10891 * to add to the startup_epoch. That way 10892 * we will still retain the old state. 10893 */ 10894 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10895 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10896 /* Make sure to use the lower pg when shifting back in */ 10897 if (bbr->r_ctl.rc_lost && 10898 bbr_use_lower_gain_in_startup && 10899 (bbr->rc_use_google == 0)) 10900 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10901 else 10902 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10903 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10904 /* Probably not needed but set it anyway */ 10905 bbr_set_state_target(bbr, __LINE__); 10906 bbr_log_type_statechange(bbr, cts, __LINE__); 10907 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10908 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10909 } 10910 bbr_check_probe_rtt_limits(bbr, cts); 10911 } 10912 10913 static int32_t inline 10914 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10915 { 10916 if ((bbr->rc_past_init_win == 1) && 10917 (bbr->rc_in_persist == 0) && 10918 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10919 return (1); 10920 } 10921 if (bbr_can_force_probertt && 10922 (bbr->rc_in_persist == 0) && 10923 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10924 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10925 return (1); 10926 } 10927 return (0); 10928 } 10929 10930 static int32_t 10931 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10932 { 10933 uint64_t btlbw, gain; 10934 if (pkt_epoch == 0) { 10935 /* 10936 * Need to be on a pkt-epoch to continue. 10937 */ 10938 return (0); 10939 } 10940 btlbw = bbr_get_full_bw(bbr); 10941 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10942 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10943 if (btlbw >= gain) { 10944 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10945 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10946 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10947 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10948 } 10949 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10950 return (1); 10951 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10952 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10953 return(0); 10954 } 10955 10956 static int32_t inline 10957 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10958 { 10959 /* Have we gained 25% in the last 3 packet based epoch's? */ 10960 uint64_t btlbw, gain; 10961 int do_exit; 10962 int delta, rtt_gain; 10963 10964 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10965 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10966 /* 10967 * This qualifies as a RTT_PROBE session since we drop the 10968 * data outstanding to nothing and waited more than 10969 * bbr_rtt_probe_time. 10970 */ 10971 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10972 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10973 } 10974 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10975 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10976 return (0); 10977 } 10978 if (bbr->rc_use_google) 10979 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10980 10981 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10982 (bbr_use_lower_gain_in_startup)) { 10983 /* Drop to a lower gain 1.5 x since we saw loss */ 10984 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10985 } 10986 if (pkt_epoch == 0) { 10987 /* 10988 * Need to be on a pkt-epoch to continue. 10989 */ 10990 return (0); 10991 } 10992 if (bbr_rtt_gain_thresh) { 10993 /* 10994 * Do we allow a flow to stay 10995 * in startup with no loss and no 10996 * gain in rtt over a set threshold? 10997 */ 10998 if (bbr->r_ctl.rc_pkt_epoch_rtt && 10999 bbr->r_ctl.startup_last_srtt && 11000 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 11001 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 11002 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 11003 } else 11004 rtt_gain = 0; 11005 if ((bbr->r_ctl.startup_last_srtt == 0) || 11006 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 11007 /* First time or new lower value */ 11008 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 11009 11010 if ((bbr->r_ctl.rc_lost == 0) && 11011 (rtt_gain < bbr_rtt_gain_thresh)) { 11012 /* 11013 * No loss, and we are under 11014 * our gain threhold for 11015 * increasing RTT. 11016 */ 11017 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 11018 bbr->r_ctl.rc_bbr_last_startup_epoch++; 11019 bbr_log_startup_event(bbr, cts, rtt_gain, 11020 delta, bbr->r_ctl.startup_last_srtt, 10); 11021 return (0); 11022 } 11023 } 11024 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 11025 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 11026 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 11027 /* 11028 * We only assess if we have a new measurement when 11029 * we have no loss and are not in recovery. 11030 * Drag up by one our last_startup epoch so we will hold 11031 * the number of non-gain we have already accumulated. 11032 */ 11033 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 11034 bbr->r_ctl.rc_bbr_last_startup_epoch++; 11035 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11036 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); 11037 return (0); 11038 } 11039 /* Case where we reduced the lost (bad retransmit) */ 11040 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 11041 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11042 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 11043 btlbw = bbr_get_full_bw(bbr); 11044 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 11045 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11046 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11047 else 11048 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11049 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11050 do_exit = 0; 11051 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 11052 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 11053 if (btlbw >= gain) { 11054 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 11055 /* Update the lost so we won't exit in next set of tests */ 11056 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11057 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11058 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 11059 } 11060 if ((bbr->rc_loss_exit && 11061 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 11062 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 11063 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 11064 /* 11065 * If we had no gain, we had loss and that loss was above 11066 * our threshould, the rwnd is not constrained, and we have 11067 * had at least 3 packet epochs exit. Note that this is 11068 * switched off by sysctl. Google does not do this by the 11069 * way. 11070 */ 11071 if ((ctf_flight_size(bbr->rc_tp, 11072 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11073 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11074 do_exit = 1; 11075 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11076 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11077 } else { 11078 /* Just record an updated loss value */ 11079 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11080 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11081 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11082 } 11083 } else 11084 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11085 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11086 do_exit) { 11087 /* Return 1 to exit the startup state. */ 11088 return (1); 11089 } 11090 /* Stay in startup */ 11091 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11092 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11093 return (0); 11094 } 11095 11096 static void 11097 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11098 { 11099 /* 11100 * A tick occurred in the rtt epoch do we need to do anything? 11101 */ 11102 #ifdef BBR_INVARIANTS 11103 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11104 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11105 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11106 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11107 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11108 /* Debug code? */ 11109 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11110 } 11111 #endif 11112 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11113 /* Do we exit the startup state? */ 11114 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11115 uint32_t time_in; 11116 11117 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11118 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11119 bbr->rc_filled_pipe = 1; 11120 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11121 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11122 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11123 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11124 } else 11125 time_in = 0; 11126 if (bbr->rc_no_pacing) 11127 bbr->rc_no_pacing = 0; 11128 bbr->r_ctl.rc_bbr_state_time = cts; 11129 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11130 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11131 bbr_set_state_target(bbr, __LINE__); 11132 if ((bbr->rc_use_google == 0) && 11133 bbr_slam_cwnd_in_main_drain) { 11134 /* Here we don't have to worry about probe-rtt */ 11135 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11136 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11137 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11138 } 11139 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11140 bbr_log_type_statechange(bbr, cts, __LINE__); 11141 if (ctf_flight_size(bbr->rc_tp, 11142 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11143 bbr->r_ctl.rc_target_at_state) { 11144 /* 11145 * Switch to probe_bw if we are already 11146 * there 11147 */ 11148 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11149 bbr_substate_change(bbr, cts, __LINE__, 0); 11150 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11151 bbr_log_type_statechange(bbr, cts, __LINE__); 11152 } 11153 } 11154 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11155 uint32_t inflight; 11156 struct tcpcb *tp; 11157 11158 tp = bbr->rc_tp; 11159 inflight = ctf_flight_size(tp, 11160 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11161 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11162 /* We have reached a flight of the cwnd target */ 11163 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11164 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11165 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11166 bbr_set_state_target(bbr, __LINE__); 11167 /* 11168 * Rig it so we don't do anything crazy and 11169 * start fresh with a new randomization. 11170 */ 11171 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11172 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11173 bbr_substate_change(bbr, cts, __LINE__, 1); 11174 } 11175 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11176 /* Has in-flight reached the bdp (or less)? */ 11177 uint32_t inflight; 11178 struct tcpcb *tp; 11179 11180 tp = bbr->rc_tp; 11181 inflight = ctf_flight_size(tp, 11182 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11183 if ((bbr->rc_use_google == 0) && 11184 bbr_slam_cwnd_in_main_drain && 11185 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11186 /* 11187 * Here we don't have to worry about probe-rtt 11188 * re-slam it, but keep it slammed down. 11189 */ 11190 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11191 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11192 } 11193 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11194 /* We have drained */ 11195 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11196 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11197 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11198 uint32_t time_in; 11199 11200 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11201 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11202 } 11203 if ((bbr->rc_use_google == 0) && 11204 bbr_slam_cwnd_in_main_drain && 11205 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11206 /* Restore the cwnd */ 11207 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11208 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11209 } 11210 /* Setup probe-rtt has being done now RRS-HERE */ 11211 bbr->r_ctl.rc_rtt_shrinks = cts; 11212 bbr->r_ctl.last_in_probertt = cts; 11213 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11214 /* Randomly pick a sub-state */ 11215 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11216 bbr_substate_change(bbr, cts, __LINE__, 0); 11217 bbr_log_type_statechange(bbr, cts, __LINE__); 11218 } 11219 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11220 uint32_t flight; 11221 11222 flight = ctf_flight_size(bbr->rc_tp, 11223 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11224 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11225 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11226 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11227 /* 11228 * We must keep cwnd at the desired MSS. 11229 */ 11230 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11231 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11232 } else if ((bbr_prtt_slam_cwnd) && 11233 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11234 /* Re-slam it */ 11235 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11236 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11237 } 11238 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11239 /* Has outstanding reached our target? */ 11240 if (flight <= bbr->r_ctl.rc_target_at_state) { 11241 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11242 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11243 /* If time is exactly 0, be 1usec off */ 11244 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11245 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11246 if (bbr->rc_use_google == 0) { 11247 /* 11248 * Restore any lowering that as occurred to 11249 * reach here 11250 */ 11251 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11252 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11253 else 11254 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11255 } 11256 } 11257 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11258 (bbr->rc_use_google == 0) && 11259 bbr->r_ctl.bbr_rttprobe_gain_val && 11260 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11261 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11262 /* 11263 * We have doddled with our current hptsi 11264 * gain an srtt and have still not made it 11265 * to target, or we have increased our flight. 11266 * Lets reduce the gain by xx% 11267 * flooring the reduce at DRAIN (based on 11268 * mul/div) 11269 */ 11270 int red; 11271 11272 bbr->r_ctl.flightsize_at_drain = flight; 11273 bbr->r_ctl.rc_probertt_srttchktim = cts; 11274 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11275 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11276 /* Reduce our gain again */ 11277 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11278 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11279 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11280 /* one more chance before we give up */ 11281 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11282 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11283 } else { 11284 /* At the very bottom */ 11285 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11286 } 11287 } 11288 } 11289 if (bbr->r_ctl.rc_bbr_enters_probertt && 11290 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11291 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11292 /* Time to exit probe RTT normally */ 11293 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11294 } 11295 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11296 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11297 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11298 /* 11299 * This qualifies as a RTT_PROBE session since we 11300 * drop the data outstanding to nothing and waited 11301 * more than bbr_rtt_probe_time. 11302 */ 11303 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11304 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11305 } 11306 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11307 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11308 } else { 11309 bbr_set_probebw_gains(bbr, cts, losses); 11310 } 11311 } 11312 } 11313 11314 static void 11315 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11316 { 11317 int32_t epoch = 0; 11318 11319 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11320 bbr_set_epoch(bbr, cts, line); 11321 /* At each epoch doe lt bw sampling */ 11322 epoch = 1; 11323 } 11324 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11325 } 11326 11327 static int 11328 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, 11329 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 11330 int32_t nxt_pkt, struct timeval *tv) 11331 { 11332 int32_t thflags, retval; 11333 uint32_t cts, lcts; 11334 uint32_t tiwin; 11335 struct tcpopt to; 11336 struct tcp_bbr *bbr; 11337 struct bbr_sendmap *rsm; 11338 struct timeval ltv; 11339 int32_t did_out = 0; 11340 uint16_t nsegs; 11341 int32_t prev_state; 11342 uint32_t lost; 11343 11344 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11345 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11346 /* add in our stats */ 11347 kern_prefetch(bbr, &prev_state); 11348 prev_state = 0; 11349 thflags = tcp_get_flags(th); 11350 /* 11351 * If this is either a state-changing packet or current state isn't 11352 * established, we require a write lock on tcbinfo. Otherwise, we 11353 * allow the tcbinfo to be in either alocked or unlocked, as the 11354 * caller may have unnecessarily acquired a write lock due to a 11355 * race. 11356 */ 11357 INP_WLOCK_ASSERT(tp->t_inpcb); 11358 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11359 __func__)); 11360 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11361 __func__)); 11362 11363 tp->t_rcvtime = ticks; 11364 /* 11365 * Unscale the window into a 32-bit value. For the SYN_SENT state 11366 * the scale is zero. 11367 */ 11368 tiwin = th->th_win << tp->snd_scale; 11369 #ifdef STATS 11370 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11371 #endif 11372 11373 if (m->m_flags & M_TSTMP) { 11374 /* Prefer the hardware timestamp if present */ 11375 struct timespec ts; 11376 11377 mbuf_tstmp2timespec(m, &ts); 11378 bbr->rc_tv.tv_sec = ts.tv_sec; 11379 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11380 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11381 } else if (m->m_flags & M_TSTMP_LRO) { 11382 /* Next the arrival timestamp */ 11383 struct timespec ts; 11384 11385 mbuf_tstmp2timespec(m, &ts); 11386 bbr->rc_tv.tv_sec = ts.tv_sec; 11387 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11388 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11389 } else { 11390 /* 11391 * Ok just get the current time. 11392 */ 11393 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11394 } 11395 /* 11396 * Parse options on any incoming segment. 11397 */ 11398 tcp_dooptions(&to, (u_char *)(th + 1), 11399 (th->th_off << 2) - sizeof(struct tcphdr), 11400 (thflags & TH_SYN) ? TO_SYN : 0); 11401 11402 /* 11403 * If timestamps were negotiated during SYN/ACK and a 11404 * segment without a timestamp is received, silently drop 11405 * the segment, unless it is a RST segment or missing timestamps are 11406 * tolerated. 11407 * See section 3.2 of RFC 7323. 11408 */ 11409 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11410 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11411 retval = 0; 11412 m_freem(m); 11413 goto done_with_input; 11414 } 11415 /* 11416 * If echoed timestamp is later than the current time, fall back to 11417 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11418 * were used when this connection was established. 11419 */ 11420 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11421 to.to_tsecr -= tp->ts_offset; 11422 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11423 to.to_tsecr = 0; 11424 } 11425 /* 11426 * If its the first time in we need to take care of options and 11427 * verify we can do SACK for rack! 11428 */ 11429 if (bbr->r_state == 0) { 11430 /* 11431 * Process options only when we get SYN/ACK back. The SYN 11432 * case for incoming connections is handled in tcp_syncache. 11433 * According to RFC1323 the window field in a SYN (i.e., a 11434 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11435 * this is traditional behavior, may need to be cleaned up. 11436 */ 11437 if (bbr->rc_inp == NULL) { 11438 bbr->rc_inp = tp->t_inpcb; 11439 } 11440 /* 11441 * We need to init rc_inp here since its not init'd when 11442 * bbr_init is called 11443 */ 11444 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11445 if ((to.to_flags & TOF_SCALE) && 11446 (tp->t_flags & TF_REQ_SCALE)) { 11447 tp->t_flags |= TF_RCVD_SCALE; 11448 tp->snd_scale = to.to_wscale; 11449 } else 11450 tp->t_flags &= ~TF_REQ_SCALE; 11451 /* 11452 * Initial send window. It will be updated with the 11453 * next incoming segment to the scaled value. 11454 */ 11455 tp->snd_wnd = th->th_win; 11456 if ((to.to_flags & TOF_TS) && 11457 (tp->t_flags & TF_REQ_TSTMP)) { 11458 tp->t_flags |= TF_RCVD_TSTMP; 11459 tp->ts_recent = to.to_tsval; 11460 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11461 } else 11462 tp->t_flags &= ~TF_REQ_TSTMP; 11463 if (to.to_flags & TOF_MSS) 11464 tcp_mss(tp, to.to_mss); 11465 if ((tp->t_flags & TF_SACK_PERMIT) && 11466 (to.to_flags & TOF_SACKPERM) == 0) 11467 tp->t_flags &= ~TF_SACK_PERMIT; 11468 if (IS_FASTOPEN(tp->t_flags)) { 11469 if (to.to_flags & TOF_FASTOPEN) { 11470 uint16_t mss; 11471 11472 if (to.to_flags & TOF_MSS) 11473 mss = to.to_mss; 11474 else 11475 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 11476 mss = TCP6_MSS; 11477 else 11478 mss = TCP_MSS; 11479 tcp_fastopen_update_cache(tp, mss, 11480 to.to_tfo_len, to.to_tfo_cookie); 11481 } else 11482 tcp_fastopen_disable_path(tp); 11483 } 11484 } 11485 /* 11486 * At this point we are at the initial call. Here we decide 11487 * if we are doing RACK or not. We do this by seeing if 11488 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11489 * we switch to the default code. 11490 */ 11491 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11492 /* Bail */ 11493 tcp_switch_back_to_default(tp); 11494 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 11495 tlen, iptos); 11496 return (1); 11497 } 11498 /* Set the flag */ 11499 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 11500 tcp_set_hpts(tp->t_inpcb); 11501 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11502 } 11503 if (thflags & TH_ACK) { 11504 /* Track ack types */ 11505 if (to.to_flags & TOF_SACK) 11506 BBR_STAT_INC(bbr_acks_with_sacks); 11507 else 11508 BBR_STAT_INC(bbr_plain_acks); 11509 } 11510 /* 11511 * This is the one exception case where we set the rack state 11512 * always. All other times (timers etc) we must have a rack-state 11513 * set (so we assure we have done the checks above for SACK). 11514 */ 11515 if (thflags & TH_FIN) 11516 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11517 if (bbr->r_state != tp->t_state) 11518 bbr_set_state(tp, bbr, tiwin); 11519 11520 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11521 kern_prefetch(rsm, &prev_state); 11522 prev_state = bbr->r_state; 11523 bbr->rc_ack_was_delayed = 0; 11524 lost = bbr->r_ctl.rc_lost; 11525 bbr->rc_is_pkt_epoch_now = 0; 11526 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11527 /* Get the real time into lcts and figure the real delay */ 11528 lcts = tcp_get_usecs(<v); 11529 if (TSTMP_GT(lcts, cts)) { 11530 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11531 bbr->rc_ack_was_delayed = 1; 11532 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11533 bbr->r_ctl.highest_hdwr_delay)) 11534 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11535 } else { 11536 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11537 bbr->rc_ack_was_delayed = 0; 11538 } 11539 } else { 11540 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11541 bbr->rc_ack_was_delayed = 0; 11542 } 11543 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11544 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11545 retval = 0; 11546 m_freem(m); 11547 goto done_with_input; 11548 } 11549 /* 11550 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11551 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11552 */ 11553 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11554 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11555 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11556 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11557 return (1); 11558 } 11559 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11560 bbr->r_ctl.rc_high_rwnd = tiwin; 11561 #ifdef BBR_INVARIANTS 11562 if ((tp->t_inpcb->inp_flags & INP_DROPPED) || 11563 (tp->t_inpcb->inp_flags2 & INP_FREED)) { 11564 panic("tp:%p bbr:%p given a dropped inp:%p", 11565 tp, bbr, tp->t_inpcb); 11566 } 11567 #endif 11568 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11569 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11570 bbr->rtt_valid = 0; 11571 if (to.to_flags & TOF_TS) { 11572 bbr->rc_ts_valid = 1; 11573 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11574 } else { 11575 bbr->rc_ts_valid = 0; 11576 bbr->r_ctl.last_inbound_ts = 0; 11577 } 11578 retval = (*bbr->r_substate) (m, th, so, 11579 tp, &to, drop_hdrlen, 11580 tlen, tiwin, thflags, nxt_pkt, iptos); 11581 #ifdef BBR_INVARIANTS 11582 if ((retval == 0) && 11583 (tp->t_inpcb == NULL)) { 11584 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 11585 retval, tp, prev_state); 11586 } 11587 #endif 11588 if (nxt_pkt == 0) 11589 BBR_STAT_INC(bbr_rlock_left_ret0); 11590 else 11591 BBR_STAT_INC(bbr_rlock_left_ret1); 11592 if (retval == 0) { 11593 /* 11594 * If retval is 1 the tcb is unlocked and most likely the tp 11595 * is gone. 11596 */ 11597 INP_WLOCK_ASSERT(tp->t_inpcb); 11598 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11599 if (bbr->rc_is_pkt_epoch_now) 11600 bbr_set_pktepoch(bbr, cts, __LINE__); 11601 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11602 if (nxt_pkt == 0) { 11603 if (bbr->r_wanted_output != 0) { 11604 bbr->rc_output_starts_timer = 0; 11605 did_out = 1; 11606 if (tcp_output(tp) < 0) 11607 return (1); 11608 } else 11609 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11610 } 11611 if ((nxt_pkt == 0) && 11612 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11613 (SEQ_GT(tp->snd_max, tp->snd_una) || 11614 (tp->t_flags & TF_DELACK) || 11615 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11616 (tp->t_state <= TCPS_CLOSING)))) { 11617 /* 11618 * We could not send (probably in the hpts but 11619 * stopped the timer)? 11620 */ 11621 if ((tp->snd_max == tp->snd_una) && 11622 ((tp->t_flags & TF_DELACK) == 0) && 11623 (tcp_in_hpts(bbr->rc_inp)) && 11624 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11625 /* 11626 * keep alive not needed if we are hptsi 11627 * output yet 11628 */ 11629 ; 11630 } else { 11631 if (tcp_in_hpts(bbr->rc_inp)) { 11632 tcp_hpts_remove(bbr->rc_inp); 11633 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11634 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11635 uint32_t del; 11636 11637 del = lcts - bbr->rc_pacer_started; 11638 if (bbr->r_ctl.rc_last_delay_val > del) { 11639 BBR_STAT_INC(bbr_force_timer_start); 11640 bbr->r_ctl.rc_last_delay_val -= del; 11641 bbr->rc_pacer_started = lcts; 11642 } else { 11643 /* We are late */ 11644 bbr->r_ctl.rc_last_delay_val = 0; 11645 BBR_STAT_INC(bbr_force_output); 11646 if (tcp_output(tp) < 0) 11647 return (1); 11648 } 11649 } 11650 } 11651 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11652 0); 11653 } 11654 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11655 /* Do we have the correct timer running? */ 11656 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11657 } 11658 /* Do we have a new state */ 11659 if (bbr->r_state != tp->t_state) 11660 bbr_set_state(tp, bbr, tiwin); 11661 done_with_input: 11662 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11663 if (did_out) 11664 bbr->r_wanted_output = 0; 11665 #ifdef BBR_INVARIANTS 11666 if (tp->t_inpcb == NULL) { 11667 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 11668 did_out, 11669 retval, tp, prev_state); 11670 } 11671 #endif 11672 } 11673 return (retval); 11674 } 11675 11676 static void 11677 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 11678 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11679 { 11680 struct timeval tv; 11681 int retval; 11682 11683 /* First lets see if we have old packets */ 11684 if (tp->t_in_pkt) { 11685 if (ctf_do_queued_segments(so, tp, 1)) { 11686 m_freem(m); 11687 return; 11688 } 11689 } 11690 if (m->m_flags & M_TSTMP_LRO) { 11691 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; 11692 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; 11693 } else { 11694 /* Should not be should we kassert instead? */ 11695 tcp_get_usecs(&tv); 11696 } 11697 retval = bbr_do_segment_nounlock(m, th, so, tp, 11698 drop_hdrlen, tlen, iptos, 0, &tv); 11699 if (retval == 0) { 11700 INP_WUNLOCK(tp->t_inpcb); 11701 } 11702 } 11703 11704 /* 11705 * Return how much data can be sent without violating the 11706 * cwnd or rwnd. 11707 */ 11708 11709 static inline uint32_t 11710 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11711 uint32_t avail, int32_t sb_offset, uint32_t cts) 11712 { 11713 uint32_t len; 11714 11715 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11716 /* We never want to go over our peers rcv-window */ 11717 len = 0; 11718 } else { 11719 uint32_t flight; 11720 11721 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11722 if (flight >= sendwin) { 11723 /* 11724 * We have in flight what we are allowed by cwnd (if 11725 * it was rwnd blocking it would have hit above out 11726 * >= tp->snd_wnd). 11727 */ 11728 return (0); 11729 } 11730 len = sendwin - flight; 11731 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11732 /* We would send too much (beyond the rwnd) */ 11733 len = tp->snd_wnd - ctf_outstanding(tp); 11734 } 11735 if ((len + sb_offset) > avail) { 11736 /* 11737 * We don't have that much in the SB, how much is 11738 * there? 11739 */ 11740 len = avail - sb_offset; 11741 } 11742 } 11743 return (len); 11744 } 11745 11746 static inline void 11747 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11748 { 11749 #ifdef NETFLIX_STATS 11750 KMOD_TCPSTAT_INC(tcps_sndpack_error); 11751 KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); 11752 #endif 11753 } 11754 11755 static inline void 11756 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11757 { 11758 if (error) { 11759 bbr_do_error_accounting(tp, bbr, rsm, len, error); 11760 return; 11761 } 11762 if (rsm) { 11763 if (rsm->r_flags & BBR_TLP) { 11764 /* 11765 * TLP should not count in retran count, but in its 11766 * own bin 11767 */ 11768 #ifdef NETFLIX_STATS 11769 KMOD_TCPSTAT_INC(tcps_tlpresends); 11770 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11771 #endif 11772 } else { 11773 /* Retransmit */ 11774 tp->t_sndrexmitpack++; 11775 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11776 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11777 #ifdef STATS 11778 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11779 len); 11780 #endif 11781 } 11782 /* 11783 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11784 * sub-state 11785 */ 11786 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11787 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11788 /* Non probe_bw log in 1, 2, or 4. */ 11789 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11790 } else { 11791 /* 11792 * Log our probe state 3, and log also 5-13 to show 11793 * us the recovery sub-state for the send. This 11794 * means that 3 == (5+6+7+8+9+10+11+12+13) 11795 */ 11796 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11797 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11798 } 11799 /* Place in both 16's the totals of retransmitted */ 11800 counter_u64_add(bbr_state_lost[16], len); 11801 counter_u64_add(bbr_state_resend[16], len); 11802 /* Place in 17's the total sent */ 11803 counter_u64_add(bbr_state_resend[17], len); 11804 counter_u64_add(bbr_state_lost[17], len); 11805 11806 } else { 11807 /* New sends */ 11808 KMOD_TCPSTAT_INC(tcps_sndpack); 11809 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11810 /* Place in 17's the total sent */ 11811 counter_u64_add(bbr_state_resend[17], len); 11812 counter_u64_add(bbr_state_lost[17], len); 11813 #ifdef STATS 11814 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11815 len); 11816 #endif 11817 } 11818 } 11819 11820 static void 11821 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11822 { 11823 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11824 /* 11825 * Limit the cwnd to not be above N x the target plus whats 11826 * is outstanding. The target is based on the current b/w 11827 * estimate. 11828 */ 11829 uint32_t target; 11830 11831 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11832 target += ctf_outstanding(tp); 11833 target *= bbr_target_cwnd_mult_limit; 11834 if (tp->snd_cwnd > target) 11835 tp->snd_cwnd = target; 11836 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11837 } 11838 } 11839 11840 static int 11841 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11842 { 11843 /* 11844 * "adv" is the amount we could increase the window, taking into 11845 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11846 */ 11847 int32_t adv; 11848 int32_t oldwin; 11849 11850 adv = recwin; 11851 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11852 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11853 if (adv > oldwin) 11854 adv -= oldwin; 11855 else { 11856 /* We can't increase the window */ 11857 adv = 0; 11858 } 11859 } else 11860 oldwin = 0; 11861 11862 /* 11863 * If the new window size ends up being the same as or less 11864 * than the old size when it is scaled, then don't force 11865 * a window update. 11866 */ 11867 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11868 return (0); 11869 11870 if (adv >= (2 * maxseg) && 11871 (adv >= (so->so_rcv.sb_hiwat / 4) || 11872 recwin <= (so->so_rcv.sb_hiwat / 8) || 11873 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11874 return (1); 11875 } 11876 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11877 return (1); 11878 return (0); 11879 } 11880 11881 /* 11882 * Return 0 on success and a errno on failure to send. 11883 * Note that a 0 return may not mean we sent anything 11884 * if the TCB was on the hpts. A non-zero return 11885 * does indicate the error we got from ip[6]_output. 11886 */ 11887 static int 11888 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11889 { 11890 struct socket *so; 11891 int32_t len; 11892 uint32_t cts; 11893 uint32_t recwin, sendwin; 11894 int32_t sb_offset; 11895 int32_t flags, abandon, error = 0; 11896 struct tcp_log_buffer *lgb = NULL; 11897 struct mbuf *m; 11898 struct mbuf *mb; 11899 uint32_t if_hw_tsomaxsegcount = 0; 11900 uint32_t if_hw_tsomaxsegsize = 0; 11901 uint32_t if_hw_tsomax = 0; 11902 struct ip *ip = NULL; 11903 #ifdef TCPDEBUG 11904 struct ipovly *ipov = NULL; 11905 #endif 11906 struct tcp_bbr *bbr; 11907 struct tcphdr *th; 11908 struct udphdr *udp = NULL; 11909 u_char opt[TCP_MAXOLEN]; 11910 unsigned ipoptlen, optlen, hdrlen; 11911 unsigned ulen; 11912 uint32_t bbr_seq; 11913 uint32_t delay_calc=0; 11914 uint8_t doing_tlp = 0; 11915 uint8_t local_options; 11916 #ifdef BBR_INVARIANTS 11917 uint8_t doing_retran_from = 0; 11918 uint8_t picked_up_retran = 0; 11919 #endif 11920 uint8_t wanted_cookie = 0; 11921 uint8_t more_to_rxt=0; 11922 int32_t prefetch_so_done = 0; 11923 int32_t prefetch_rsm = 0; 11924 uint32_t tot_len = 0; 11925 uint32_t rtr_cnt = 0; 11926 uint32_t maxseg, pace_max_segs, p_maxseg; 11927 int32_t csum_flags = 0; 11928 int32_t hw_tls; 11929 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11930 unsigned ipsec_optlen = 0; 11931 11932 #endif 11933 volatile int32_t sack_rxmit; 11934 struct bbr_sendmap *rsm = NULL; 11935 int32_t tso, mtu; 11936 struct tcpopt to; 11937 int32_t slot = 0; 11938 struct inpcb *inp; 11939 struct sockbuf *sb; 11940 uint32_t hpts_calling; 11941 #ifdef INET6 11942 struct ip6_hdr *ip6 = NULL; 11943 int32_t isipv6; 11944 #endif 11945 uint8_t app_limited = BBR_JR_SENT_DATA; 11946 uint8_t filled_all = 0; 11947 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11948 /* We take a cache hit here */ 11949 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11950 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11951 inp = bbr->rc_inp; 11952 so = inp->inp_socket; 11953 sb = &so->so_snd; 11954 if (sb->sb_flags & SB_TLS_IFNET) 11955 hw_tls = 1; 11956 else 11957 hw_tls = 0; 11958 kern_prefetch(sb, &maxseg); 11959 maxseg = tp->t_maxseg - bbr->rc_last_options; 11960 if (bbr_minseg(bbr) < maxseg) { 11961 tcp_bbr_tso_size_check(bbr, cts); 11962 } 11963 /* Remove any flags that indicate we are pacing on the inp */ 11964 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11965 p_maxseg = min(maxseg, pace_max_segs); 11966 INP_WLOCK_ASSERT(inp); 11967 #ifdef TCP_OFFLOAD 11968 if (tp->t_flags & TF_TOE) 11969 return (tcp_offload_output(tp)); 11970 #endif 11971 11972 #ifdef INET6 11973 if (bbr->r_state) { 11974 /* Use the cache line loaded if possible */ 11975 isipv6 = bbr->r_is_v6; 11976 } else { 11977 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11978 } 11979 #endif 11980 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11981 tcp_in_hpts(inp)) { 11982 /* 11983 * We are on the hpts for some timer but not hptsi output. 11984 * Possibly remove from the hpts so we can send/recv etc. 11985 */ 11986 if ((tp->t_flags & TF_ACKNOW) == 0) { 11987 /* 11988 * No immediate demand right now to send an ack, but 11989 * the user may have read, making room for new data 11990 * (a window update). If so we may want to cancel 11991 * whatever timer is running (KEEP/DEL-ACK?) and 11992 * continue to send out a window update. Or we may 11993 * have gotten more data into the socket buffer to 11994 * send. 11995 */ 11996 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 11997 (long)TCP_MAXWIN << tp->rcv_scale); 11998 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 11999 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 12000 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 12001 (tp->snd_max - tp->snd_una))) { 12002 /* 12003 * Nothing new to send and no window update 12004 * is needed to send. Lets just return and 12005 * let the timer-run off. 12006 */ 12007 return (0); 12008 } 12009 } 12010 tcp_hpts_remove(inp); 12011 bbr_timer_cancel(bbr, __LINE__, cts); 12012 } 12013 if (bbr->r_ctl.rc_last_delay_val) { 12014 /* Calculate a rough delay for early escape to sending */ 12015 if (SEQ_GT(cts, bbr->rc_pacer_started)) 12016 delay_calc = cts - bbr->rc_pacer_started; 12017 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 12018 delay_calc -= bbr->r_ctl.rc_last_delay_val; 12019 else 12020 delay_calc = 0; 12021 } 12022 /* Mark that we have called bbr_output(). */ 12023 if ((bbr->r_timer_override) || 12024 (tp->t_state < TCPS_ESTABLISHED)) { 12025 /* Timeouts or early states are exempt */ 12026 if (tcp_in_hpts(inp)) 12027 tcp_hpts_remove(inp); 12028 } else if (tcp_in_hpts(inp)) { 12029 if ((bbr->r_ctl.rc_last_delay_val) && 12030 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 12031 delay_calc) { 12032 /* 12033 * We were being paced for output and the delay has 12034 * already exceeded when we were supposed to be 12035 * called, lets go ahead and pull out of the hpts 12036 * and call output. 12037 */ 12038 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 12039 bbr->r_ctl.rc_last_delay_val = 0; 12040 tcp_hpts_remove(inp); 12041 } else if (tp->t_state == TCPS_CLOSED) { 12042 bbr->r_ctl.rc_last_delay_val = 0; 12043 tcp_hpts_remove(inp); 12044 } else { 12045 /* 12046 * On the hpts, you shall not pass! even if ACKNOW 12047 * is on, we will when the hpts fires, unless of 12048 * course we are overdue. 12049 */ 12050 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 12051 return (0); 12052 } 12053 } 12054 bbr->rc_cwnd_limited = 0; 12055 if (bbr->r_ctl.rc_last_delay_val) { 12056 /* recalculate the real delay and deal with over/under */ 12057 if (SEQ_GT(cts, bbr->rc_pacer_started)) 12058 delay_calc = cts - bbr->rc_pacer_started; 12059 else 12060 delay_calc = 0; 12061 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 12062 /* Setup the delay which will be added in */ 12063 delay_calc -= bbr->r_ctl.rc_last_delay_val; 12064 else { 12065 /* 12066 * We are early setup to adjust 12067 * our slot time. 12068 */ 12069 uint64_t merged_val; 12070 12071 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 12072 bbr->r_agg_early_set = 1; 12073 if (bbr->r_ctl.rc_hptsi_agg_delay) { 12074 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 12075 /* Nope our previous late cancels out the early */ 12076 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 12077 bbr->r_agg_early_set = 0; 12078 bbr->r_ctl.rc_agg_early = 0; 12079 } else { 12080 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 12081 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12082 } 12083 } 12084 merged_val = bbr->rc_pacer_started; 12085 merged_val <<= 32; 12086 merged_val |= bbr->r_ctl.rc_last_delay_val; 12087 bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, 12088 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 12089 bbr->r_agg_early_set, 3); 12090 bbr->r_ctl.rc_last_delay_val = 0; 12091 BBR_STAT_INC(bbr_early); 12092 delay_calc = 0; 12093 } 12094 } else { 12095 /* We were not delayed due to hptsi */ 12096 if (bbr->r_agg_early_set) 12097 bbr->r_ctl.rc_agg_early = 0; 12098 bbr->r_agg_early_set = 0; 12099 delay_calc = 0; 12100 } 12101 if (delay_calc) { 12102 /* 12103 * We had a hptsi delay which means we are falling behind on 12104 * sending at the expected rate. Calculate an extra amount 12105 * of data we can send, if any, to put us back on track. 12106 */ 12107 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12108 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12109 else 12110 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12111 } 12112 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12113 if ((tp->snd_una == tp->snd_max) && 12114 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12115 (sbavail(sb))) { 12116 /* 12117 * Ok we have been idle with nothing outstanding 12118 * we possibly need to start fresh with either a new 12119 * suite of states or a fast-ramp up. 12120 */ 12121 bbr_restart_after_idle(bbr, 12122 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12123 } 12124 /* 12125 * Now was there a hptsi delay where we are behind? We only count 12126 * being behind if: a) We are not in recovery. b) There was a delay. 12127 * <and> c) We had room to send something. 12128 * 12129 */ 12130 hpts_calling = inp->inp_hpts_calls; 12131 inp->inp_hpts_calls = 0; 12132 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12133 int retval; 12134 12135 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12136 if (retval != 0) { 12137 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12138 /* 12139 * If timers want tcp_drop(), then pass error out, 12140 * otherwise suppress it. 12141 */ 12142 return (retval < 0 ? retval : 0); 12143 } 12144 } 12145 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 12146 if (hpts_calling && 12147 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12148 bbr->r_ctl.rc_last_delay_val = 0; 12149 } 12150 bbr->r_timer_override = 0; 12151 bbr->r_wanted_output = 0; 12152 /* 12153 * For TFO connections in SYN_RECEIVED, only allow the initial 12154 * SYN|ACK and those sent by the retransmit timer. 12155 */ 12156 if (IS_FASTOPEN(tp->t_flags) && 12157 ((tp->t_state == TCPS_SYN_RECEIVED) || 12158 (tp->t_state == TCPS_SYN_SENT)) && 12159 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12160 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12161 len = 0; 12162 goto just_return_nolock; 12163 } 12164 /* 12165 * Before sending anything check for a state update. For hpts 12166 * calling without input this is important. If its input calling 12167 * then this was already done. 12168 */ 12169 if (bbr->rc_use_google == 0) 12170 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12171 again: 12172 /* 12173 * If we've recently taken a timeout, snd_max will be greater than 12174 * snd_max. BBR in general does not pay much attention to snd_nxt 12175 * for historic reasons the persist timer still uses it. This means 12176 * we have to look at it. All retransmissions that are not persits 12177 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12178 * end of this routine we pull snd_nxt always up to snd_max. 12179 */ 12180 doing_tlp = 0; 12181 #ifdef BBR_INVARIANTS 12182 doing_retran_from = picked_up_retran = 0; 12183 #endif 12184 error = 0; 12185 tso = 0; 12186 slot = 0; 12187 mtu = 0; 12188 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12189 sb_offset = tp->snd_max - tp->snd_una; 12190 flags = tcp_outflags[tp->t_state]; 12191 sack_rxmit = 0; 12192 len = 0; 12193 rsm = NULL; 12194 if (flags & TH_RST) { 12195 SOCKBUF_LOCK(sb); 12196 goto send; 12197 } 12198 recheck_resend: 12199 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12200 /* We need to always have one in reserve */ 12201 rsm = bbr_alloc(bbr); 12202 if (rsm == NULL) { 12203 error = ENOMEM; 12204 /* Lie to get on the hpts */ 12205 tot_len = tp->t_maxseg; 12206 if (hpts_calling) 12207 /* Retry in a ms */ 12208 slot = 1001; 12209 goto just_return_nolock; 12210 } 12211 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12212 bbr->r_ctl.rc_free_cnt++; 12213 rsm = NULL; 12214 } 12215 /* What do we send, a resend? */ 12216 if (bbr->r_ctl.rc_resend == NULL) { 12217 /* Check for rack timeout */ 12218 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12219 if (bbr->r_ctl.rc_resend) { 12220 #ifdef BBR_INVARIANTS 12221 picked_up_retran = 1; 12222 #endif 12223 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12224 } 12225 } 12226 if (bbr->r_ctl.rc_resend) { 12227 rsm = bbr->r_ctl.rc_resend; 12228 #ifdef BBR_INVARIANTS 12229 doing_retran_from = 1; 12230 #endif 12231 /* Remove any TLP flags its a RACK or T-O */ 12232 rsm->r_flags &= ~BBR_TLP; 12233 bbr->r_ctl.rc_resend = NULL; 12234 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12235 #ifdef BBR_INVARIANTS 12236 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12237 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12238 goto recheck_resend; 12239 #else 12240 /* TSNH */ 12241 rsm = NULL; 12242 goto recheck_resend; 12243 #endif 12244 } 12245 rtr_cnt++; 12246 if (rsm->r_flags & BBR_HAS_SYN) { 12247 /* Only retransmit a SYN by itself */ 12248 len = 0; 12249 if ((flags & TH_SYN) == 0) { 12250 /* Huh something is wrong */ 12251 rsm->r_start++; 12252 if (rsm->r_start == rsm->r_end) { 12253 /* Clean it up, somehow we missed the ack? */ 12254 bbr_log_syn(tp, NULL); 12255 } else { 12256 /* TFO with data? */ 12257 rsm->r_flags &= ~BBR_HAS_SYN; 12258 len = rsm->r_end - rsm->r_start; 12259 } 12260 } else { 12261 /* Retransmitting SYN */ 12262 rsm = NULL; 12263 SOCKBUF_LOCK(sb); 12264 goto send; 12265 } 12266 } else 12267 len = rsm->r_end - rsm->r_start; 12268 if ((bbr->rc_resends_use_tso == 0) && 12269 (len > maxseg)) { 12270 len = maxseg; 12271 more_to_rxt = 1; 12272 } 12273 sb_offset = rsm->r_start - tp->snd_una; 12274 if (len > 0) { 12275 sack_rxmit = 1; 12276 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12277 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12278 min(len, maxseg)); 12279 } else { 12280 /* I dont think this can happen */ 12281 rsm = NULL; 12282 goto recheck_resend; 12283 } 12284 BBR_STAT_INC(bbr_resends_set); 12285 } else if (bbr->r_ctl.rc_tlp_send) { 12286 /* 12287 * Tail loss probe 12288 */ 12289 doing_tlp = 1; 12290 rsm = bbr->r_ctl.rc_tlp_send; 12291 bbr->r_ctl.rc_tlp_send = NULL; 12292 sack_rxmit = 1; 12293 len = rsm->r_end - rsm->r_start; 12294 rtr_cnt++; 12295 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12296 len = maxseg; 12297 12298 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12299 #ifdef BBR_INVARIANTS 12300 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12301 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12302 #else 12303 /* TSNH */ 12304 rsm = NULL; 12305 goto recheck_resend; 12306 #endif 12307 } 12308 sb_offset = rsm->r_start - tp->snd_una; 12309 BBR_STAT_INC(bbr_tlp_set); 12310 } 12311 /* 12312 * Enforce a connection sendmap count limit if set 12313 * as long as we are not retransmiting. 12314 */ 12315 if ((rsm == NULL) && 12316 (V_tcp_map_entries_limit > 0) && 12317 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12318 BBR_STAT_INC(bbr_alloc_limited); 12319 if (!bbr->alloc_limit_reported) { 12320 bbr->alloc_limit_reported = 1; 12321 BBR_STAT_INC(bbr_alloc_limited_conns); 12322 } 12323 goto just_return_nolock; 12324 } 12325 #ifdef BBR_INVARIANTS 12326 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12327 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12328 tp, bbr, rsm, sb_offset, len); 12329 } 12330 #endif 12331 /* 12332 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12333 * state flags. 12334 */ 12335 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12336 flags |= TH_FIN; 12337 if (tp->t_flags & TF_NEEDSYN) 12338 flags |= TH_SYN; 12339 12340 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12341 /* we are retransmitting the fin */ 12342 len--; 12343 if (len) { 12344 /* 12345 * When retransmitting data do *not* include the 12346 * FIN. This could happen from a TLP probe if we 12347 * allowed data with a FIN. 12348 */ 12349 flags &= ~TH_FIN; 12350 } 12351 } else if (rsm) { 12352 if (flags & TH_FIN) 12353 flags &= ~TH_FIN; 12354 } 12355 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12356 void *end_rsm; 12357 12358 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12359 if (end_rsm) 12360 kern_prefetch(end_rsm, &prefetch_rsm); 12361 prefetch_rsm = 1; 12362 } 12363 SOCKBUF_LOCK(sb); 12364 /* 12365 * If snd_nxt == snd_max and we have transmitted a FIN, the 12366 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12367 * negative length. This can also occur when TCP opens up its 12368 * congestion window while receiving additional duplicate acks after 12369 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12370 * the fast-retransmit. 12371 * 12372 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12373 * set to snd_una, the sb_offset will be 0, and the length may wind 12374 * up 0. 12375 * 12376 * If sack_rxmit is true we are retransmitting from the scoreboard 12377 * in which case len is already set. 12378 */ 12379 if (sack_rxmit == 0) { 12380 uint32_t avail; 12381 12382 avail = sbavail(sb); 12383 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12384 sb_offset = tp->snd_max - tp->snd_una; 12385 else 12386 sb_offset = 0; 12387 if (bbr->rc_tlp_new_data) { 12388 /* TLP is forcing out new data */ 12389 uint32_t tlplen; 12390 12391 doing_tlp = 1; 12392 tlplen = maxseg; 12393 12394 if (tlplen > (uint32_t)(avail - sb_offset)) { 12395 tlplen = (uint32_t)(avail - sb_offset); 12396 } 12397 if (tlplen > tp->snd_wnd) { 12398 len = tp->snd_wnd; 12399 } else { 12400 len = tlplen; 12401 } 12402 bbr->rc_tlp_new_data = 0; 12403 } else { 12404 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12405 if ((len < p_maxseg) && 12406 (bbr->rc_in_persist == 0) && 12407 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12408 ((avail - sb_offset) >= p_maxseg)) { 12409 /* 12410 * We are not completing whats in the socket 12411 * buffer (i.e. there is at least a segment 12412 * waiting to send) and we have 2 or more 12413 * segments outstanding. There is no sense 12414 * of sending a little piece. Lets defer and 12415 * and wait until we can send a whole 12416 * segment. 12417 */ 12418 len = 0; 12419 } 12420 if (bbr->rc_in_persist) { 12421 /* 12422 * We are in persists, figure out if 12423 * a retransmit is available (maybe the previous 12424 * persists we sent) or if we have to send new 12425 * data. 12426 */ 12427 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12428 if (rsm) { 12429 len = rsm->r_end - rsm->r_start; 12430 if (rsm->r_flags & BBR_HAS_FIN) 12431 len--; 12432 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12433 len = maxseg; 12434 if (len > 1) 12435 BBR_STAT_INC(bbr_persist_reneg); 12436 /* 12437 * XXXrrs we could force the len to 12438 * 1 byte here to cause the chunk to 12439 * split apart.. but that would then 12440 * mean we always retransmit it as 12441 * one byte even after the window 12442 * opens. 12443 */ 12444 sack_rxmit = 1; 12445 sb_offset = rsm->r_start - tp->snd_una; 12446 } else { 12447 /* 12448 * First time through in persists or peer 12449 * acked our one byte. Though we do have 12450 * to have something in the sb. 12451 */ 12452 len = 1; 12453 sb_offset = 0; 12454 if (avail == 0) 12455 len = 0; 12456 } 12457 } 12458 } 12459 } 12460 if (prefetch_so_done == 0) { 12461 kern_prefetch(so, &prefetch_so_done); 12462 prefetch_so_done = 1; 12463 } 12464 /* 12465 * Lop off SYN bit if it has already been sent. However, if this is 12466 * SYN-SENT state and if segment contains data and if we don't know 12467 * that foreign host supports TAO, suppress sending segment. 12468 */ 12469 if ((flags & TH_SYN) && (rsm == NULL) && 12470 SEQ_GT(tp->snd_max, tp->snd_una)) { 12471 if (tp->t_state != TCPS_SYN_RECEIVED) 12472 flags &= ~TH_SYN; 12473 /* 12474 * When sending additional segments following a TFO SYN|ACK, 12475 * do not include the SYN bit. 12476 */ 12477 if (IS_FASTOPEN(tp->t_flags) && 12478 (tp->t_state == TCPS_SYN_RECEIVED)) 12479 flags &= ~TH_SYN; 12480 sb_offset--, len++; 12481 if (sbavail(sb) == 0) 12482 len = 0; 12483 } else if ((flags & TH_SYN) && rsm) { 12484 /* 12485 * Subtract one from the len for the SYN being 12486 * retransmitted. 12487 */ 12488 len--; 12489 } 12490 /* 12491 * Be careful not to send data and/or FIN on SYN segments. This 12492 * measure is needed to prevent interoperability problems with not 12493 * fully conformant TCP implementations. 12494 */ 12495 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12496 len = 0; 12497 flags &= ~TH_FIN; 12498 } 12499 /* 12500 * On TFO sockets, ensure no data is sent in the following cases: 12501 * 12502 * - When retransmitting SYN|ACK on a passively-created socket 12503 * - When retransmitting SYN on an actively created socket 12504 * - When sending a zero-length cookie (cookie request) on an 12505 * actively created socket 12506 * - When the socket is in the CLOSED state (RST is being sent) 12507 */ 12508 if (IS_FASTOPEN(tp->t_flags) && 12509 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12510 ((tp->t_state == TCPS_SYN_SENT) && 12511 (tp->t_tfo_client_cookie_len == 0)) || 12512 (flags & TH_RST))) { 12513 len = 0; 12514 sack_rxmit = 0; 12515 rsm = NULL; 12516 } 12517 /* Without fast-open there should never be data sent on a SYN */ 12518 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 12519 len = 0; 12520 if (len <= 0) { 12521 /* 12522 * If FIN has been sent but not acked, but we haven't been 12523 * called to retransmit, len will be < 0. Otherwise, window 12524 * shrank after we sent into it. If window shrank to 0, 12525 * cancel pending retransmit, pull snd_nxt back to (closed) 12526 * window, and set the persist timer if it isn't already 12527 * going. If the window didn't close completely, just wait 12528 * for an ACK. 12529 * 12530 * We also do a general check here to ensure that we will 12531 * set the persist timer when we have data to send, but a 12532 * 0-byte window. This makes sure the persist timer is set 12533 * even if the packet hits one of the "goto send" lines 12534 * below. 12535 */ 12536 len = 0; 12537 if ((tp->snd_wnd == 0) && 12538 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12539 (tp->snd_una == tp->snd_max) && 12540 (sb_offset < (int)sbavail(sb))) { 12541 /* 12542 * Not enough room in the rwnd to send 12543 * a paced segment out. 12544 */ 12545 bbr_enter_persist(tp, bbr, cts, __LINE__); 12546 } 12547 } else if ((rsm == NULL) && 12548 (doing_tlp == 0) && 12549 (len < bbr->r_ctl.rc_pace_max_segs)) { 12550 /* 12551 * We are not sending a full segment for 12552 * some reason. Should we not send anything (think 12553 * sws or persists)? 12554 */ 12555 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12556 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12557 (len < (int)(sbavail(sb) - sb_offset))) { 12558 /* 12559 * Here the rwnd is less than 12560 * the pacing size, this is not a retransmit, 12561 * we are established and 12562 * the send is not the last in the socket buffer 12563 * lets not send, and possibly enter persists. 12564 */ 12565 len = 0; 12566 if (tp->snd_max == tp->snd_una) 12567 bbr_enter_persist(tp, bbr, cts, __LINE__); 12568 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12569 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12570 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12571 (len < (int)(sbavail(sb) - sb_offset)) && 12572 (len < bbr_minseg(bbr))) { 12573 /* 12574 * Here we are not retransmitting, and 12575 * the cwnd is not so small that we could 12576 * not send at least a min size (rxt timer 12577 * not having gone off), We have 2 segments or 12578 * more already in flight, its not the tail end 12579 * of the socket buffer and the cwnd is blocking 12580 * us from sending out minimum pacing segment size. 12581 * Lets not send anything. 12582 */ 12583 bbr->rc_cwnd_limited = 1; 12584 len = 0; 12585 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12586 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12587 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12588 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12589 (len < (int)(sbavail(sb) - sb_offset)) && 12590 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12591 /* 12592 * Here we have a send window but we have 12593 * filled it up and we can't send another pacing segment. 12594 * We also have in flight more than 2 segments 12595 * and we are not completing the sb i.e. we allow 12596 * the last bytes of the sb to go out even if 12597 * its not a full pacing segment. 12598 */ 12599 len = 0; 12600 } 12601 } 12602 /* len will be >= 0 after this point. */ 12603 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12604 tcp_sndbuf_autoscale(tp, so, sendwin); 12605 /* 12606 * 12607 */ 12608 if (bbr->rc_in_persist && 12609 len && 12610 (rsm == NULL) && 12611 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12612 /* 12613 * We are in persist, not doing a retransmit and don't have enough space 12614 * yet to send a full TSO. So is it at the end of the sb 12615 * if so we need to send else nuke to 0 and don't send. 12616 */ 12617 int sbleft; 12618 if (sbavail(sb) > sb_offset) 12619 sbleft = sbavail(sb) - sb_offset; 12620 else 12621 sbleft = 0; 12622 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12623 /* not at end of sb lets not send */ 12624 len = 0; 12625 } 12626 } 12627 /* 12628 * Decide if we can use TCP Segmentation Offloading (if supported by 12629 * hardware). 12630 * 12631 * TSO may only be used if we are in a pure bulk sending state. The 12632 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12633 * options prevent using TSO. With TSO the TCP header is the same 12634 * (except for the sequence number) for all generated packets. This 12635 * makes it impossible to transmit any options which vary per 12636 * generated segment or packet. 12637 * 12638 * IPv4 handling has a clear separation of ip options and ip header 12639 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12640 * does the right thing below to provide length of just ip options 12641 * and thus checking for ipoptlen is enough to decide if ip options 12642 * are present. 12643 */ 12644 #ifdef INET6 12645 if (isipv6) 12646 ipoptlen = ip6_optlen(inp); 12647 else 12648 #endif 12649 if (inp->inp_options) 12650 ipoptlen = inp->inp_options->m_len - 12651 offsetof(struct ipoption, ipopt_list); 12652 else 12653 ipoptlen = 0; 12654 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12655 /* 12656 * Pre-calculate here as we save another lookup into the darknesses 12657 * of IPsec that way and can actually decide if TSO is ok. 12658 */ 12659 #ifdef INET6 12660 if (isipv6 && IPSEC_ENABLED(ipv6)) 12661 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12662 #ifdef INET 12663 else 12664 #endif 12665 #endif /* INET6 */ 12666 #ifdef INET 12667 if (IPSEC_ENABLED(ipv4)) 12668 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12669 #endif /* INET */ 12670 #endif /* IPSEC */ 12671 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12672 ipoptlen += ipsec_optlen; 12673 #endif 12674 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12675 (len > maxseg) && 12676 (tp->t_port == 0) && 12677 ((tp->t_flags & TF_SIGNATURE) == 0) && 12678 tp->rcv_numsacks == 0 && 12679 ipoptlen == 0) 12680 tso = 1; 12681 12682 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12683 (long)TCP_MAXWIN << tp->rcv_scale); 12684 /* 12685 * Sender silly window avoidance. We transmit under the following 12686 * conditions when len is non-zero: 12687 * 12688 * - We have a full segment (or more with TSO) - This is the last 12689 * buffer in a write()/send() and we are either idle or running 12690 * NODELAY - we've timed out (e.g. persist timer) - we have more 12691 * then 1/2 the maximum send window's worth of data (receiver may be 12692 * limited the window size) - we need to retransmit 12693 */ 12694 if (rsm) 12695 goto send; 12696 if (len) { 12697 if (sack_rxmit) 12698 goto send; 12699 if (len >= p_maxseg) 12700 goto send; 12701 /* 12702 * NOTE! on localhost connections an 'ack' from the remote 12703 * end may occur synchronously with the output and cause us 12704 * to flush a buffer queued with moretocome. XXX 12705 * 12706 */ 12707 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12708 ((tp->t_flags & TF_NODELAY) || 12709 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12710 (tp->t_flags & TF_NOPUSH) == 0) { 12711 goto send; 12712 } 12713 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12714 goto send; 12715 } 12716 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12717 goto send; 12718 } 12719 } 12720 /* 12721 * Sending of standalone window updates. 12722 * 12723 * Window updates are important when we close our window due to a 12724 * full socket buffer and are opening it again after the application 12725 * reads data from it. Once the window has opened again and the 12726 * remote end starts to send again the ACK clock takes over and 12727 * provides the most current window information. 12728 * 12729 * We must avoid the silly window syndrome whereas every read from 12730 * the receive buffer, no matter how small, causes a window update 12731 * to be sent. We also should avoid sending a flurry of window 12732 * updates when the socket buffer had queued a lot of data and the 12733 * application is doing small reads. 12734 * 12735 * Prevent a flurry of pointless window updates by only sending an 12736 * update when we can increase the advertized window by more than 12737 * 1/4th of the socket buffer capacity. When the buffer is getting 12738 * full or is very small be more aggressive and send an update 12739 * whenever we can increase by two mss sized segments. In all other 12740 * situations the ACK's to new incoming data will carry further 12741 * window increases. 12742 * 12743 * Don't send an independent window update if a delayed ACK is 12744 * pending (it will get piggy-backed on it) or the remote side 12745 * already has done a half-close and won't send more data. Skip 12746 * this if the connection is in T/TCP half-open state. 12747 */ 12748 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12749 !(tp->t_flags & TF_DELACK) && 12750 !TCPS_HAVERCVDFIN(tp->t_state)) { 12751 /* Check to see if we should do a window update */ 12752 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12753 goto send; 12754 } 12755 /* 12756 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12757 * is also a catch-all for the retransmit timer timeout case. 12758 */ 12759 if (tp->t_flags & TF_ACKNOW) { 12760 goto send; 12761 } 12762 if (flags & TH_RST) { 12763 /* Always send a RST if one is due */ 12764 goto send; 12765 } 12766 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12767 goto send; 12768 } 12769 /* 12770 * If our state indicates that FIN should be sent and we have not 12771 * yet done so, then we need to send. 12772 */ 12773 if (flags & TH_FIN && 12774 ((tp->t_flags & TF_SENTFIN) == 0)) { 12775 goto send; 12776 } 12777 /* 12778 * No reason to send a segment, just return. 12779 */ 12780 just_return: 12781 SOCKBUF_UNLOCK(sb); 12782 just_return_nolock: 12783 if (tot_len) 12784 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12785 if (bbr->rc_no_pacing) 12786 slot = 0; 12787 if (tot_len == 0) { 12788 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12789 tp->snd_wnd) { 12790 BBR_STAT_INC(bbr_rwnd_limited); 12791 app_limited = BBR_JR_RWND_LIMITED; 12792 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12793 if ((bbr->rc_in_persist == 0) && 12794 TCPS_HAVEESTABLISHED(tp->t_state) && 12795 (tp->snd_max == tp->snd_una) && 12796 sbavail(&tp->t_inpcb->inp_socket->so_snd)) { 12797 /* No send window.. we must enter persist */ 12798 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12799 } 12800 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12801 BBR_STAT_INC(bbr_app_limited); 12802 app_limited = BBR_JR_APP_LIMITED; 12803 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12804 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12805 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12806 BBR_STAT_INC(bbr_cwnd_limited); 12807 app_limited = BBR_JR_CWND_LIMITED; 12808 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12809 bbr->r_ctl.rc_lost_bytes))); 12810 bbr->rc_cwnd_limited = 1; 12811 } else { 12812 BBR_STAT_INC(bbr_app_limited); 12813 app_limited = BBR_JR_APP_LIMITED; 12814 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12815 } 12816 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12817 bbr->r_agg_early_set = 0; 12818 bbr->r_ctl.rc_agg_early = 0; 12819 bbr->r_ctl.rc_last_delay_val = 0; 12820 } else if (bbr->rc_use_google == 0) 12821 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12822 /* Are we app limited? */ 12823 if ((app_limited == BBR_JR_APP_LIMITED) || 12824 (app_limited == BBR_JR_RWND_LIMITED)) { 12825 /** 12826 * We are application limited. 12827 */ 12828 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12829 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12830 } 12831 if (tot_len == 0) 12832 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12833 /* Dont update the time if we did not send */ 12834 bbr->r_ctl.rc_last_delay_val = 0; 12835 bbr->rc_output_starts_timer = 1; 12836 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12837 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12838 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12839 /* Make sure snd_nxt is drug up */ 12840 tp->snd_nxt = tp->snd_max; 12841 } 12842 return (error); 12843 12844 send: 12845 if (doing_tlp == 0) { 12846 /* 12847 * Data not a TLP, and its not the rxt firing. If it is the 12848 * rxt firing, we want to leave the tlp_in_progress flag on 12849 * so we don't send another TLP. It has to be a rack timer 12850 * or normal send (response to acked data) to clear the tlp 12851 * in progress flag. 12852 */ 12853 bbr->rc_tlp_in_progress = 0; 12854 bbr->rc_tlp_rtx_out = 0; 12855 } else { 12856 /* 12857 * Its a TLP. 12858 */ 12859 bbr->rc_tlp_in_progress = 1; 12860 } 12861 bbr_timer_cancel(bbr, __LINE__, cts); 12862 if (rsm == NULL) { 12863 if (sbused(sb) > 0) { 12864 /* 12865 * This is sub-optimal. We only send a stand alone 12866 * FIN on its own segment. 12867 */ 12868 if (flags & TH_FIN) { 12869 flags &= ~TH_FIN; 12870 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12871 /* Lets not send this */ 12872 slot = 0; 12873 goto just_return; 12874 } 12875 } 12876 } 12877 } else { 12878 /* 12879 * We do *not* send a FIN on a retransmit if it has data. 12880 * The if clause here where len > 1 should never come true. 12881 */ 12882 if ((len > 0) && 12883 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12884 (flags & TH_FIN))) { 12885 flags &= ~TH_FIN; 12886 len--; 12887 } 12888 } 12889 SOCKBUF_LOCK_ASSERT(sb); 12890 if (len > 0) { 12891 if ((tp->snd_una == tp->snd_max) && 12892 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12893 /* 12894 * This qualifies as a RTT_PROBE session since we 12895 * drop the data outstanding to nothing and waited 12896 * more than bbr_rtt_probe_time. 12897 */ 12898 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12899 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12900 } 12901 if (len >= maxseg) 12902 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12903 else 12904 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12905 } 12906 /* 12907 * Before ESTABLISHED, force sending of initial options unless TCP 12908 * set not to do any options. NOTE: we assume that the IP/TCP header 12909 * plus TCP options always fit in a single mbuf, leaving room for a 12910 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12911 * + optlen <= MCLBYTES 12912 */ 12913 optlen = 0; 12914 #ifdef INET6 12915 if (isipv6) 12916 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12917 else 12918 #endif 12919 hdrlen = sizeof(struct tcpiphdr); 12920 12921 /* 12922 * Compute options for segment. We only have to care about SYN and 12923 * established connection segments. Options for SYN-ACK segments 12924 * are handled in TCP syncache. 12925 */ 12926 to.to_flags = 0; 12927 local_options = 0; 12928 if ((tp->t_flags & TF_NOOPT) == 0) { 12929 /* Maximum segment size. */ 12930 if (flags & TH_SYN) { 12931 to.to_mss = tcp_mssopt(&inp->inp_inc); 12932 if (tp->t_port) 12933 to.to_mss -= V_tcp_udp_tunneling_overhead; 12934 to.to_flags |= TOF_MSS; 12935 /* 12936 * On SYN or SYN|ACK transmits on TFO connections, 12937 * only include the TFO option if it is not a 12938 * retransmit, as the presence of the TFO option may 12939 * have caused the original SYN or SYN|ACK to have 12940 * been dropped by a middlebox. 12941 */ 12942 if (IS_FASTOPEN(tp->t_flags) && 12943 (tp->t_rxtshift == 0)) { 12944 if (tp->t_state == TCPS_SYN_RECEIVED) { 12945 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12946 to.to_tfo_cookie = 12947 (u_int8_t *)&tp->t_tfo_cookie.server; 12948 to.to_flags |= TOF_FASTOPEN; 12949 wanted_cookie = 1; 12950 } else if (tp->t_state == TCPS_SYN_SENT) { 12951 to.to_tfo_len = 12952 tp->t_tfo_client_cookie_len; 12953 to.to_tfo_cookie = 12954 tp->t_tfo_cookie.client; 12955 to.to_flags |= TOF_FASTOPEN; 12956 wanted_cookie = 1; 12957 } 12958 } 12959 } 12960 /* Window scaling. */ 12961 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12962 to.to_wscale = tp->request_r_scale; 12963 to.to_flags |= TOF_SCALE; 12964 } 12965 /* Timestamps. */ 12966 if ((tp->t_flags & TF_RCVD_TSTMP) || 12967 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12968 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12969 to.to_tsecr = tp->ts_recent; 12970 to.to_flags |= TOF_TS; 12971 local_options += TCPOLEN_TIMESTAMP + 2; 12972 } 12973 /* Set receive buffer autosizing timestamp. */ 12974 if (tp->rfbuf_ts == 0 && 12975 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12976 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12977 /* Selective ACK's. */ 12978 if (flags & TH_SYN) 12979 to.to_flags |= TOF_SACKPERM; 12980 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12981 tp->rcv_numsacks > 0) { 12982 to.to_flags |= TOF_SACK; 12983 to.to_nsacks = tp->rcv_numsacks; 12984 to.to_sacks = (u_char *)tp->sackblks; 12985 } 12986 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 12987 /* TCP-MD5 (RFC2385). */ 12988 if (tp->t_flags & TF_SIGNATURE) 12989 to.to_flags |= TOF_SIGNATURE; 12990 #endif /* TCP_SIGNATURE */ 12991 12992 /* Processing the options. */ 12993 hdrlen += (optlen = tcp_addoptions(&to, opt)); 12994 /* 12995 * If we wanted a TFO option to be added, but it was unable 12996 * to fit, ensure no data is sent. 12997 */ 12998 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 12999 !(to.to_flags & TOF_FASTOPEN)) 13000 len = 0; 13001 } 13002 if (tp->t_port) { 13003 if (V_tcp_udp_tunneling_port == 0) { 13004 /* The port was removed?? */ 13005 SOCKBUF_UNLOCK(&so->so_snd); 13006 return (EHOSTUNREACH); 13007 } 13008 hdrlen += sizeof(struct udphdr); 13009 } 13010 #ifdef INET6 13011 if (isipv6) 13012 ipoptlen = ip6_optlen(tp->t_inpcb); 13013 else 13014 #endif 13015 if (tp->t_inpcb->inp_options) 13016 ipoptlen = tp->t_inpcb->inp_options->m_len - 13017 offsetof(struct ipoption, ipopt_list); 13018 else 13019 ipoptlen = 0; 13020 ipoptlen = 0; 13021 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 13022 ipoptlen += ipsec_optlen; 13023 #endif 13024 if (bbr->rc_last_options != local_options) { 13025 /* 13026 * Cache the options length this generally does not change 13027 * on a connection. We use this to calculate TSO. 13028 */ 13029 bbr->rc_last_options = local_options; 13030 } 13031 maxseg = tp->t_maxseg - (ipoptlen + optlen); 13032 p_maxseg = min(maxseg, pace_max_segs); 13033 /* 13034 * Adjust data length if insertion of options will bump the packet 13035 * length beyond the t_maxseg length. Clear the FIN bit because we 13036 * cut off the tail of the segment. 13037 */ 13038 if (len > maxseg) { 13039 if (len != 0 && (flags & TH_FIN)) { 13040 flags &= ~TH_FIN; 13041 } 13042 if (tso) { 13043 uint32_t moff; 13044 int32_t max_len; 13045 13046 /* extract TSO information */ 13047 if_hw_tsomax = tp->t_tsomax; 13048 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 13049 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 13050 KASSERT(ipoptlen == 0, 13051 ("%s: TSO can't do IP options", __func__)); 13052 13053 /* 13054 * Check if we should limit by maximum payload 13055 * length: 13056 */ 13057 if (if_hw_tsomax != 0) { 13058 /* compute maximum TSO length */ 13059 max_len = (if_hw_tsomax - hdrlen - 13060 max_linkhdr); 13061 if (max_len <= 0) { 13062 len = 0; 13063 } else if (len > max_len) { 13064 len = max_len; 13065 } 13066 } 13067 /* 13068 * Prevent the last segment from being fractional 13069 * unless the send sockbuf can be emptied: 13070 */ 13071 if ((sb_offset + len) < sbavail(sb)) { 13072 moff = len % (uint32_t)maxseg; 13073 if (moff != 0) { 13074 len -= moff; 13075 } 13076 } 13077 /* 13078 * In case there are too many small fragments don't 13079 * use TSO: 13080 */ 13081 if (len <= maxseg) { 13082 len = maxseg; 13083 tso = 0; 13084 } 13085 } else { 13086 /* Not doing TSO */ 13087 if (optlen + ipoptlen >= tp->t_maxseg) { 13088 /* 13089 * Since we don't have enough space to put 13090 * the IP header chain and the TCP header in 13091 * one packet as required by RFC 7112, don't 13092 * send it. Also ensure that at least one 13093 * byte of the payload can be put into the 13094 * TCP segment. 13095 */ 13096 SOCKBUF_UNLOCK(&so->so_snd); 13097 error = EMSGSIZE; 13098 sack_rxmit = 0; 13099 goto out; 13100 } 13101 len = maxseg; 13102 } 13103 } else { 13104 /* Not doing TSO */ 13105 if_hw_tsomaxsegcount = 0; 13106 tso = 0; 13107 } 13108 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13109 ("%s: len > IP_MAXPACKET", __func__)); 13110 #ifdef DIAGNOSTIC 13111 #ifdef INET6 13112 if (max_linkhdr + hdrlen > MCLBYTES) 13113 #else 13114 if (max_linkhdr + hdrlen > MHLEN) 13115 #endif 13116 panic("tcphdr too big"); 13117 #endif 13118 /* 13119 * This KASSERT is here to catch edge cases at a well defined place. 13120 * Before, those had triggered (random) panic conditions further 13121 * down. 13122 */ 13123 #ifdef BBR_INVARIANTS 13124 if (sack_rxmit) { 13125 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13126 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13127 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13128 } 13129 } 13130 #endif 13131 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13132 if ((len == 0) && 13133 (flags & TH_FIN) && 13134 (sbused(sb))) { 13135 /* 13136 * We have outstanding data, don't send a fin by itself!. 13137 */ 13138 slot = 0; 13139 goto just_return; 13140 } 13141 /* 13142 * Grab a header mbuf, attaching a copy of data to be transmitted, 13143 * and initialize the header from the template for sends on this 13144 * connection. 13145 */ 13146 if (len) { 13147 uint32_t moff; 13148 13149 /* 13150 * We place a limit on sending with hptsi. 13151 */ 13152 if ((rsm == NULL) && len > pace_max_segs) 13153 len = pace_max_segs; 13154 if (len <= maxseg) 13155 tso = 0; 13156 #ifdef INET6 13157 if (MHLEN < hdrlen + max_linkhdr) 13158 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13159 else 13160 #endif 13161 m = m_gethdr(M_NOWAIT, MT_DATA); 13162 13163 if (m == NULL) { 13164 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13165 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13166 SOCKBUF_UNLOCK(sb); 13167 error = ENOBUFS; 13168 sack_rxmit = 0; 13169 goto out; 13170 } 13171 m->m_data += max_linkhdr; 13172 m->m_len = hdrlen; 13173 /* 13174 * Start the m_copy functions from the closest mbuf to the 13175 * sb_offset in the socket buffer chain. 13176 */ 13177 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13178 #ifdef BBR_INVARIANTS 13179 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13180 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13181 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13182 doing_retran_from, 13183 picked_up_retran, 13184 doing_tlp); 13185 13186 #endif 13187 /* 13188 * In this messed up situation we have two choices, 13189 * a) pretend the send worked, and just start timers 13190 * and what not (not good since that may lead us 13191 * back here a lot). <or> b) Send the lowest segment 13192 * in the map. <or> c) Drop the connection. Lets do 13193 * <b> which if it continues to happen will lead to 13194 * <c> via timeouts. 13195 */ 13196 BBR_STAT_INC(bbr_offset_recovery); 13197 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13198 sb_offset = 0; 13199 if (rsm == NULL) { 13200 sack_rxmit = 0; 13201 len = sbavail(sb); 13202 } else { 13203 sack_rxmit = 1; 13204 if (rsm->r_start != tp->snd_una) { 13205 /* 13206 * Things are really messed up, <c> 13207 * is the only thing to do. 13208 */ 13209 BBR_STAT_INC(bbr_offset_drop); 13210 SOCKBUF_UNLOCK(sb); 13211 (void)m_free(m); 13212 return (-EFAULT); /* tcp_drop() */ 13213 } 13214 len = rsm->r_end - rsm->r_start; 13215 } 13216 if (len > sbavail(sb)) 13217 len = sbavail(sb); 13218 if (len > maxseg) 13219 len = maxseg; 13220 } 13221 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13222 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13223 m_copydata(mb, moff, (int)len, 13224 mtod(m, caddr_t)+hdrlen); 13225 if (rsm == NULL) 13226 sbsndptr_adv(sb, mb, len); 13227 m->m_len += len; 13228 } else { 13229 struct sockbuf *msb; 13230 13231 if (rsm) 13232 msb = NULL; 13233 else 13234 msb = sb; 13235 #ifdef BBR_INVARIANTS 13236 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13237 if (rsm) { 13238 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 ", 13239 tp, bbr, len, moff, 13240 sbavail(sb), rsm, 13241 tp->snd_una, rsm->r_flags, rsm->r_start, 13242 doing_retran_from, 13243 picked_up_retran, 13244 doing_tlp, sack_rxmit); 13245 } else { 13246 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13247 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13248 } 13249 } 13250 #endif 13251 m->m_next = tcp_m_copym( 13252 mb, moff, &len, 13253 if_hw_tsomaxsegcount, 13254 if_hw_tsomaxsegsize, msb, 13255 ((rsm == NULL) ? hw_tls : 0) 13256 #ifdef NETFLIX_COPY_ARGS 13257 , &filled_all 13258 #endif 13259 ); 13260 if (len <= maxseg) { 13261 /* 13262 * Must have ran out of mbufs for the copy 13263 * shorten it to no longer need tso. Lets 13264 * not put on sendalot since we are low on 13265 * mbufs. 13266 */ 13267 tso = 0; 13268 } 13269 if (m->m_next == NULL) { 13270 SOCKBUF_UNLOCK(sb); 13271 (void)m_free(m); 13272 error = ENOBUFS; 13273 sack_rxmit = 0; 13274 goto out; 13275 } 13276 } 13277 #ifdef BBR_INVARIANTS 13278 if (tso && len < maxseg) { 13279 panic("tp:%p tso on, but len:%d < maxseg:%d", 13280 tp, len, maxseg); 13281 } 13282 if (tso && if_hw_tsomaxsegcount) { 13283 int32_t seg_cnt = 0; 13284 struct mbuf *foo; 13285 13286 foo = m; 13287 while (foo) { 13288 seg_cnt++; 13289 foo = foo->m_next; 13290 } 13291 if (seg_cnt > if_hw_tsomaxsegcount) { 13292 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13293 } 13294 } 13295 #endif 13296 /* 13297 * If we're sending everything we've got, set PUSH. (This 13298 * will keep happy those implementations which only give 13299 * data to the user when a buffer fills or a PUSH comes in.) 13300 */ 13301 if (sb_offset + len == sbused(sb) && 13302 sbused(sb) && 13303 !(flags & TH_SYN)) { 13304 flags |= TH_PUSH; 13305 } 13306 SOCKBUF_UNLOCK(sb); 13307 } else { 13308 SOCKBUF_UNLOCK(sb); 13309 if (tp->t_flags & TF_ACKNOW) 13310 KMOD_TCPSTAT_INC(tcps_sndacks); 13311 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13312 KMOD_TCPSTAT_INC(tcps_sndctrl); 13313 else 13314 KMOD_TCPSTAT_INC(tcps_sndwinup); 13315 13316 m = m_gethdr(M_NOWAIT, MT_DATA); 13317 if (m == NULL) { 13318 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13319 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13320 error = ENOBUFS; 13321 /* Fudge the send time since we could not send */ 13322 sack_rxmit = 0; 13323 goto out; 13324 } 13325 #ifdef INET6 13326 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13327 MHLEN >= hdrlen) { 13328 M_ALIGN(m, hdrlen); 13329 } else 13330 #endif 13331 m->m_data += max_linkhdr; 13332 m->m_len = hdrlen; 13333 } 13334 SOCKBUF_UNLOCK_ASSERT(sb); 13335 m->m_pkthdr.rcvif = (struct ifnet *)0; 13336 #ifdef MAC 13337 mac_inpcb_create_mbuf(inp, m); 13338 #endif 13339 #ifdef INET6 13340 if (isipv6) { 13341 ip6 = mtod(m, struct ip6_hdr *); 13342 if (tp->t_port) { 13343 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13344 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13345 udp->uh_dport = tp->t_port; 13346 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13347 udp->uh_ulen = htons(ulen); 13348 th = (struct tcphdr *)(udp + 1); 13349 } else { 13350 th = (struct tcphdr *)(ip6 + 1); 13351 } 13352 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13353 } else 13354 #endif /* INET6 */ 13355 { 13356 ip = mtod(m, struct ip *); 13357 #ifdef TCPDEBUG 13358 ipov = (struct ipovly *)ip; 13359 #endif 13360 if (tp->t_port) { 13361 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13362 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13363 udp->uh_dport = tp->t_port; 13364 ulen = hdrlen + len - sizeof(struct ip); 13365 udp->uh_ulen = htons(ulen); 13366 th = (struct tcphdr *)(udp + 1); 13367 } else { 13368 th = (struct tcphdr *)(ip + 1); 13369 } 13370 tcpip_fillheaders(inp, tp->t_port, ip, th); 13371 } 13372 /* 13373 * If we are doing retransmissions, then snd_nxt will not reflect 13374 * the first unsent octet. For ACK only packets, we do not want the 13375 * sequence number of the retransmitted packet, we want the sequence 13376 * number of the next unsent octet. So, if there is no data (and no 13377 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13378 * ti_seq. But if we are in persist state, snd_max might reflect 13379 * one byte beyond the right edge of the window, so use snd_nxt in 13380 * that case, since we know we aren't doing a retransmission. 13381 * (retransmit and persist are mutually exclusive...) 13382 */ 13383 if (sack_rxmit == 0) { 13384 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13385 /* New data (including new persists) */ 13386 th->th_seq = htonl(tp->snd_max); 13387 bbr_seq = tp->snd_max; 13388 } else if (flags & TH_SYN) { 13389 /* Syn's always send from iss */ 13390 th->th_seq = htonl(tp->iss); 13391 bbr_seq = tp->iss; 13392 } else if (flags & TH_FIN) { 13393 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13394 /* 13395 * If we sent the fin already its 1 minus 13396 * snd_max 13397 */ 13398 th->th_seq = (htonl(tp->snd_max - 1)); 13399 bbr_seq = (tp->snd_max - 1); 13400 } else { 13401 /* First time FIN use snd_max */ 13402 th->th_seq = htonl(tp->snd_max); 13403 bbr_seq = tp->snd_max; 13404 } 13405 } else { 13406 /* 13407 * len == 0 and not persist we use snd_max, sending 13408 * an ack unless we have sent the fin then its 1 13409 * minus. 13410 */ 13411 /* 13412 * XXXRRS Question if we are in persists and we have 13413 * nothing outstanding to send and we have not sent 13414 * a FIN, we will send an ACK. In such a case it 13415 * might be better to send (tp->snd_una - 1) which 13416 * would force the peer to ack. 13417 */ 13418 if (tp->t_flags & TF_SENTFIN) { 13419 th->th_seq = htonl(tp->snd_max - 1); 13420 bbr_seq = (tp->snd_max - 1); 13421 } else { 13422 th->th_seq = htonl(tp->snd_max); 13423 bbr_seq = tp->snd_max; 13424 } 13425 } 13426 } else { 13427 /* All retransmits use the rsm to guide the send */ 13428 th->th_seq = htonl(rsm->r_start); 13429 bbr_seq = rsm->r_start; 13430 } 13431 th->th_ack = htonl(tp->rcv_nxt); 13432 if (optlen) { 13433 bcopy(opt, th + 1, optlen); 13434 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13435 } 13436 tcp_set_flags(th, flags); 13437 /* 13438 * Calculate receive window. Don't shrink window, but avoid silly 13439 * window syndrome. 13440 */ 13441 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13442 recwin < maxseg))) 13443 recwin = 0; 13444 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13445 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13446 recwin = (tp->rcv_adv - tp->rcv_nxt); 13447 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13448 recwin = TCP_MAXWIN << tp->rcv_scale; 13449 13450 /* 13451 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13452 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13453 * handled in syncache. 13454 */ 13455 if (flags & TH_SYN) 13456 th->th_win = htons((u_short) 13457 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13458 else { 13459 /* Avoid shrinking window with window scaling. */ 13460 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13461 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13462 } 13463 /* 13464 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13465 * window. This may cause the remote transmitter to stall. This 13466 * flag tells soreceive() to disable delayed acknowledgements when 13467 * draining the buffer. This can occur if the receiver is 13468 * attempting to read more data than can be buffered prior to 13469 * transmitting on the connection. 13470 */ 13471 if (th->th_win == 0) { 13472 tp->t_sndzerowin++; 13473 tp->t_flags |= TF_RXWIN0SENT; 13474 } else 13475 tp->t_flags &= ~TF_RXWIN0SENT; 13476 /* 13477 * We don't support urgent data, but drag along 13478 * the pointer in case of a stack switch. 13479 */ 13480 tp->snd_up = tp->snd_una; 13481 13482 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13483 if (to.to_flags & TOF_SIGNATURE) { 13484 /* 13485 * Calculate MD5 signature and put it into the place 13486 * determined before. NOTE: since TCP options buffer doesn't 13487 * point into mbuf's data, calculate offset and use it. 13488 */ 13489 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13490 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13491 /* 13492 * Do not send segment if the calculation of MD5 13493 * digest has failed. 13494 */ 13495 goto out; 13496 } 13497 } 13498 #endif 13499 13500 /* 13501 * Put TCP length in extended header, and then checksum extended 13502 * header and data. 13503 */ 13504 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13505 #ifdef INET6 13506 if (isipv6) { 13507 /* 13508 * ip6_plen is not need to be filled now, and will be filled 13509 * in ip6_output. 13510 */ 13511 if (tp->t_port) { 13512 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13513 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13514 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13515 th->th_sum = htons(0); 13516 UDPSTAT_INC(udps_opackets); 13517 } else { 13518 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13519 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13520 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13521 optlen + len, IPPROTO_TCP, 0); 13522 } 13523 } 13524 #endif 13525 #if defined(INET6) && defined(INET) 13526 else 13527 #endif 13528 #ifdef INET 13529 { 13530 if (tp->t_port) { 13531 m->m_pkthdr.csum_flags = CSUM_UDP; 13532 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13533 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13534 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13535 th->th_sum = htons(0); 13536 UDPSTAT_INC(udps_opackets); 13537 } else { 13538 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13539 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13540 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13541 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13542 IPPROTO_TCP + len + optlen)); 13543 } 13544 /* IP version must be set here for ipv4/ipv6 checking later */ 13545 KASSERT(ip->ip_v == IPVERSION, 13546 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13547 } 13548 #endif 13549 13550 /* 13551 * Enable TSO and specify the size of the segments. The TCP pseudo 13552 * header checksum is always provided. XXX: Fixme: This is currently 13553 * not the case for IPv6. 13554 */ 13555 if (tso) { 13556 KASSERT(len > maxseg, 13557 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13558 m->m_pkthdr.csum_flags |= CSUM_TSO; 13559 csum_flags |= CSUM_TSO; 13560 m->m_pkthdr.tso_segsz = maxseg; 13561 } 13562 KASSERT(len + hdrlen == m_length(m, NULL), 13563 ("%s: mbuf chain different than expected: %d + %u != %u", 13564 __func__, len, hdrlen, m_length(m, NULL))); 13565 13566 #ifdef TCP_HHOOK 13567 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13568 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13569 #endif 13570 #ifdef TCPDEBUG 13571 /* 13572 * Trace. 13573 */ 13574 if (so->so_options & SO_DEBUG) { 13575 u_short save = 0; 13576 13577 #ifdef INET6 13578 if (!isipv6) 13579 #endif 13580 { 13581 save = ipov->ih_len; 13582 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 13583 * (th->th_off << 2) */ ); 13584 } 13585 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 13586 #ifdef INET6 13587 if (!isipv6) 13588 #endif 13589 ipov->ih_len = save; 13590 } 13591 #endif /* TCPDEBUG */ 13592 13593 /* Log to the black box */ 13594 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 13595 union tcp_log_stackspecific log; 13596 13597 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13598 /* Record info on type of transmission */ 13599 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13600 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13601 log.u_bbr.flex3 = maxseg; 13602 log.u_bbr.flex4 = delay_calc; 13603 /* Encode filled_all into the upper flex5 bit */ 13604 log.u_bbr.flex5 = bbr->rc_past_init_win; 13605 log.u_bbr.flex5 <<= 1; 13606 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13607 log.u_bbr.flex5 <<= 29; 13608 if (filled_all) 13609 log.u_bbr.flex5 |= 0x80000000; 13610 log.u_bbr.flex5 |= tp->t_maxseg; 13611 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13612 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13613 /* lets poke in the low and the high here for debugging */ 13614 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13615 if (rsm || sack_rxmit) { 13616 if (doing_tlp) 13617 log.u_bbr.flex8 = 2; 13618 else 13619 log.u_bbr.flex8 = 1; 13620 } else { 13621 log.u_bbr.flex8 = 0; 13622 } 13623 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13624 len, &log, false, NULL, NULL, 0, tv); 13625 } else { 13626 lgb = NULL; 13627 } 13628 /* 13629 * Fill in IP length and desired time to live and send to IP level. 13630 * There should be a better way to handle ttl and tos; we could keep 13631 * them in the template, but need a way to checksum without them. 13632 */ 13633 /* 13634 * m->m_pkthdr.len should have been set before cksum calcuration, 13635 * because in6_cksum() need it. 13636 */ 13637 #ifdef INET6 13638 if (isipv6) { 13639 /* 13640 * we separately set hoplimit for every segment, since the 13641 * user might want to change the value via setsockopt. Also, 13642 * desired default hop limit might be changed via Neighbor 13643 * Discovery. 13644 */ 13645 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13646 13647 /* 13648 * Set the packet size here for the benefit of DTrace 13649 * probes. ip6_output() will set it properly; it's supposed 13650 * to include the option header lengths as well. 13651 */ 13652 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13653 13654 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13655 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13656 else 13657 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13658 13659 if (tp->t_state == TCPS_SYN_SENT) 13660 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13661 13662 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13663 /* TODO: IPv6 IP6TOS_ECT bit on */ 13664 error = ip6_output(m, inp->in6p_outputopts, 13665 &inp->inp_route6, 13666 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13667 NULL, NULL, inp); 13668 13669 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13670 mtu = inp->inp_route6.ro_nh->nh_mtu; 13671 } 13672 #endif /* INET6 */ 13673 #if defined(INET) && defined(INET6) 13674 else 13675 #endif 13676 #ifdef INET 13677 { 13678 ip->ip_len = htons(m->m_pkthdr.len); 13679 #ifdef INET6 13680 if (isipv6) 13681 ip->ip_ttl = in6_selecthlim(inp, NULL); 13682 #endif /* INET6 */ 13683 /* 13684 * If we do path MTU discovery, then we set DF on every 13685 * packet. This might not be the best thing to do according 13686 * to RFC3390 Section 2. However the tcp hostcache migitates 13687 * the problem so it affects only the first tcp connection 13688 * with a host. 13689 * 13690 * NB: Don't set DF on small MTU/MSS to have a safe 13691 * fallback. 13692 */ 13693 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13694 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13695 if (tp->t_port == 0 || len < V_tcp_minmss) { 13696 ip->ip_off |= htons(IP_DF); 13697 } 13698 } else { 13699 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13700 } 13701 13702 if (tp->t_state == TCPS_SYN_SENT) 13703 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13704 13705 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13706 13707 error = ip_output(m, inp->inp_options, &inp->inp_route, 13708 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13709 inp); 13710 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13711 mtu = inp->inp_route.ro_nh->nh_mtu; 13712 } 13713 #endif /* INET */ 13714 out: 13715 13716 if (lgb) { 13717 lgb->tlb_errno = error; 13718 lgb = NULL; 13719 } 13720 /* 13721 * In transmit state, time the transmission and arrange for the 13722 * retransmit. In persist state, just set snd_max. 13723 */ 13724 if (error == 0) { 13725 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13726 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13727 (tp->t_flags & TF_SACK_PERMIT) && 13728 tp->rcv_numsacks > 0) 13729 tcp_clean_dsack_blocks(tp); 13730 /* We sent an ack clear the bbr_segs_rcvd count */ 13731 bbr->output_error_seen = 0; 13732 bbr->oerror_cnt = 0; 13733 bbr->bbr_segs_rcvd = 0; 13734 if (len == 0) 13735 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13736 /* Do accounting for new sends */ 13737 if ((len > 0) && (rsm == NULL)) { 13738 int idx; 13739 if (tp->snd_una == tp->snd_max) { 13740 /* 13741 * Special case to match google, when 13742 * nothing is in flight the delivered 13743 * time does get updated to the current 13744 * time (see tcp_rate_bsd.c). 13745 */ 13746 bbr->r_ctl.rc_del_time = cts; 13747 } 13748 if (len >= maxseg) { 13749 idx = (len / maxseg) + 3; 13750 if (idx >= TCP_MSS_ACCT_ATIMER) 13751 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13752 else 13753 counter_u64_add(bbr_out_size[idx], 1); 13754 } else { 13755 /* smaller than a MSS */ 13756 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13757 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13758 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13759 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13760 } 13761 } 13762 } 13763 abandon = 0; 13764 /* 13765 * We must do the send accounting before we log the output, 13766 * otherwise the state of the rsm could change and we account to the 13767 * wrong bucket. 13768 */ 13769 if (len > 0) { 13770 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13771 if (error == 0) { 13772 if (tp->snd_una == tp->snd_max) 13773 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13774 } 13775 } 13776 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13777 cts, mb, &abandon, rsm, 0, sb); 13778 if (abandon) { 13779 /* 13780 * If bbr_log_output destroys the TCB or sees a TH_RST being 13781 * sent we should hit this condition. 13782 */ 13783 return (0); 13784 } 13785 if (bbr->rc_in_persist == 0) { 13786 /* 13787 * Advance snd_nxt over sequence space of this segment. 13788 */ 13789 if (error) 13790 /* We don't log or do anything with errors */ 13791 goto skip_upd; 13792 13793 if (tp->snd_una == tp->snd_max && 13794 (len || (flags & (TH_SYN | TH_FIN)))) { 13795 /* 13796 * Update the time we just added data since none was 13797 * outstanding. 13798 */ 13799 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13800 bbr->rc_tp->t_acktime = ticks; 13801 } 13802 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13803 if (flags & TH_SYN) { 13804 /* 13805 * Smack the snd_max to iss + 1 13806 * if its a FO we will add len below. 13807 */ 13808 tp->snd_max = tp->iss + 1; 13809 } 13810 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13811 tp->snd_max++; 13812 tp->t_flags |= TF_SENTFIN; 13813 } 13814 } 13815 if (sack_rxmit == 0) 13816 tp->snd_max += len; 13817 skip_upd: 13818 if ((error == 0) && len) 13819 tot_len += len; 13820 } else { 13821 /* Persists case */ 13822 int32_t xlen = len; 13823 13824 if (error) 13825 goto nomore; 13826 13827 if (flags & TH_SYN) 13828 ++xlen; 13829 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13830 ++xlen; 13831 tp->t_flags |= TF_SENTFIN; 13832 } 13833 if (xlen && (tp->snd_una == tp->snd_max)) { 13834 /* 13835 * Update the time we just added data since none was 13836 * outstanding. 13837 */ 13838 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13839 bbr->rc_tp->t_acktime = ticks; 13840 } 13841 if (sack_rxmit == 0) 13842 tp->snd_max += xlen; 13843 tot_len += (len + optlen + ipoptlen); 13844 } 13845 nomore: 13846 if (error) { 13847 /* 13848 * Failures do not advance the seq counter above. For the 13849 * case of ENOBUFS we will fall out and become ack-clocked. 13850 * capping the cwnd at the current flight. 13851 * Everything else will just have to retransmit with the timer 13852 * (no pacer). 13853 */ 13854 SOCKBUF_UNLOCK_ASSERT(sb); 13855 BBR_STAT_INC(bbr_saw_oerr); 13856 /* Clear all delay/early tracks */ 13857 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13858 bbr->r_ctl.rc_agg_early = 0; 13859 bbr->r_agg_early_set = 0; 13860 bbr->output_error_seen = 1; 13861 if (bbr->oerror_cnt < 0xf) 13862 bbr->oerror_cnt++; 13863 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13864 /* drop the session */ 13865 return (-ENETDOWN); 13866 } 13867 switch (error) { 13868 case ENOBUFS: 13869 /* 13870 * Make this guy have to get ack's to send 13871 * more but lets make sure we don't 13872 * slam him below a T-O (1MSS). 13873 */ 13874 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13875 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13876 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13877 if (tp->snd_cwnd < maxseg) 13878 tp->snd_cwnd = maxseg; 13879 } 13880 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13881 BBR_STAT_INC(bbr_saw_enobuf); 13882 if (bbr->bbr_hdrw_pacing) 13883 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13884 else 13885 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13886 /* 13887 * Here even in the enobuf's case we want to do our 13888 * state update. The reason being we may have been 13889 * called by the input function. If so we have had 13890 * things change. 13891 */ 13892 error = 0; 13893 goto enobufs; 13894 case EMSGSIZE: 13895 /* 13896 * For some reason the interface we used initially 13897 * to send segments changed to another or lowered 13898 * its MTU. If TSO was active we either got an 13899 * interface without TSO capabilits or TSO was 13900 * turned off. If we obtained mtu from ip_output() 13901 * then update it and try again. 13902 */ 13903 /* Turn on tracing (or try to) */ 13904 { 13905 int old_maxseg; 13906 13907 old_maxseg = tp->t_maxseg; 13908 BBR_STAT_INC(bbr_saw_emsgsiz); 13909 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13910 if (mtu != 0) 13911 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13912 if (old_maxseg <= tp->t_maxseg) { 13913 /* Huh it did not shrink? */ 13914 tp->t_maxseg = old_maxseg - 40; 13915 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13916 } 13917 /* 13918 * Nuke all other things that can interfere 13919 * with slot 13920 */ 13921 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13922 slot = bbr_get_pacing_delay(bbr, 13923 bbr->r_ctl.rc_bbr_hptsi_gain, 13924 (tot_len + len), cts, 0); 13925 if (slot < bbr_error_base_paceout) 13926 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13927 } else 13928 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13929 bbr->rc_output_starts_timer = 1; 13930 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13931 tot_len); 13932 return (error); 13933 } 13934 case EPERM: 13935 tp->t_softerror = error; 13936 /* Fall through */ 13937 case EHOSTDOWN: 13938 case EHOSTUNREACH: 13939 case ENETDOWN: 13940 case ENETUNREACH: 13941 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13942 tp->t_softerror = error; 13943 } 13944 /* FALLTHROUGH */ 13945 default: 13946 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13947 bbr->rc_output_starts_timer = 1; 13948 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13949 return (error); 13950 } 13951 #ifdef STATS 13952 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13953 len && 13954 (rsm == NULL) && 13955 (bbr->rc_in_persist == 0)) { 13956 tp->gput_seq = bbr_seq; 13957 tp->gput_ack = bbr_seq + 13958 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13959 tp->gput_ts = cts; 13960 tp->t_flags |= TF_GPUTINPROG; 13961 #endif 13962 } 13963 KMOD_TCPSTAT_INC(tcps_sndtotal); 13964 if ((bbr->bbr_hdw_pace_ena) && 13965 (bbr->bbr_attempt_hdwr_pace == 0) && 13966 (bbr->rc_past_init_win) && 13967 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13968 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13969 (inp->inp_route.ro_nh && 13970 inp->inp_route.ro_nh->nh_ifp)) { 13971 /* 13972 * We are past the initial window and 13973 * have at least one measurement so we 13974 * could use hardware pacing if its available. 13975 * We have an interface and we have not attempted 13976 * to setup hardware pacing, lets try to now. 13977 */ 13978 uint64_t rate_wanted; 13979 int err = 0; 13980 13981 rate_wanted = bbr_get_hardware_rate(bbr); 13982 bbr->bbr_attempt_hdwr_pace = 1; 13983 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 13984 inp->inp_route.ro_nh->nh_ifp, 13985 rate_wanted, 13986 (RS_PACING_GEQ|RS_PACING_SUB_OK), 13987 &err, NULL); 13988 if (bbr->r_ctl.crte) { 13989 bbr_type_log_hdwr_pacing(bbr, 13990 bbr->r_ctl.crte->ptbl->rs_ifp, 13991 rate_wanted, 13992 bbr->r_ctl.crte->rate, 13993 __LINE__, cts, err); 13994 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 13995 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 13996 counter_u64_add(bbr_flows_whdwr_pacing, 1); 13997 bbr->bbr_hdrw_pacing = 1; 13998 /* Now what is our gain status? */ 13999 if (bbr->r_ctl.crte->rate < rate_wanted) { 14000 /* We have a problem */ 14001 bbr_setup_less_of_rate(bbr, cts, 14002 bbr->r_ctl.crte->rate, rate_wanted); 14003 } else { 14004 /* We are good */ 14005 bbr->gain_is_limited = 0; 14006 bbr->skip_gain = 0; 14007 } 14008 tcp_bbr_tso_size_check(bbr, cts); 14009 } else { 14010 bbr_type_log_hdwr_pacing(bbr, 14011 inp->inp_route.ro_nh->nh_ifp, 14012 rate_wanted, 14013 0, 14014 __LINE__, cts, err); 14015 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 14016 } 14017 } 14018 if (bbr->bbr_hdrw_pacing) { 14019 /* 14020 * Worry about cases where the route 14021 * changes or something happened that we 14022 * lost our hardware pacing possibly during 14023 * the last ip_output call. 14024 */ 14025 if (inp->inp_snd_tag == NULL) { 14026 /* A change during ip output disabled hw pacing? */ 14027 bbr->bbr_hdrw_pacing = 0; 14028 } else if ((inp->inp_route.ro_nh == NULL) || 14029 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 14030 /* 14031 * We had an interface or route change, 14032 * detach from the current hdwr pacing 14033 * and setup to re-attempt next go 14034 * round. 14035 */ 14036 bbr->bbr_hdrw_pacing = 0; 14037 bbr->bbr_attempt_hdwr_pace = 0; 14038 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 14039 tcp_bbr_tso_size_check(bbr, cts); 14040 } 14041 } 14042 /* 14043 * Data sent (as far as we can tell). If this advertises a larger 14044 * window than any other segment, then remember the size of the 14045 * advertised window. Any pending ACK has now been sent. 14046 */ 14047 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 14048 tp->rcv_adv = tp->rcv_nxt + recwin; 14049 14050 tp->last_ack_sent = tp->rcv_nxt; 14051 if ((error == 0) && 14052 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 14053 (doing_tlp == 0) && 14054 (tso == 0) && 14055 (len > 0) && 14056 ((flags & TH_RST) == 0) && 14057 ((flags & TH_SYN) == 0) && 14058 (IN_RECOVERY(tp->t_flags) == 0) && 14059 (bbr->rc_in_persist == 0) && 14060 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 14061 /* 14062 * For non-tso we need to goto again until we have sent out 14063 * enough data to match what we are hptsi out every hptsi 14064 * interval. 14065 */ 14066 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14067 /* Make sure snd_nxt is drug up */ 14068 tp->snd_nxt = tp->snd_max; 14069 } 14070 if (rsm != NULL) { 14071 rsm = NULL; 14072 goto skip_again; 14073 } 14074 rsm = NULL; 14075 sack_rxmit = 0; 14076 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14077 goto again; 14078 } 14079 skip_again: 14080 if ((error == 0) && (flags & TH_FIN)) 14081 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 14082 if ((error == 0) && (flags & TH_RST)) 14083 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 14084 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 14085 /* 14086 * Calculate/Re-Calculate the hptsi slot in usecs based on 14087 * what we have sent so far 14088 */ 14089 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 14090 if (bbr->rc_no_pacing) 14091 slot = 0; 14092 } 14093 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14094 enobufs: 14095 if (bbr->rc_use_google == 0) 14096 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 14097 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 14098 bbr->r_ctl.rc_lost_bytes))); 14099 bbr->rc_output_starts_timer = 1; 14100 if (bbr->bbr_use_rack_cheat && 14101 (more_to_rxt || 14102 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 14103 /* Rack cheats and shotguns out all rxt's 1ms apart */ 14104 if (slot > 1000) 14105 slot = 1000; 14106 } 14107 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 14108 /* 14109 * We don't change the tso size until some number of sends 14110 * to give the hardware commands time to get down 14111 * to the interface. 14112 */ 14113 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 14114 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 14115 bbr->hw_pacing_set = 1; 14116 tcp_bbr_tso_size_check(bbr, cts); 14117 } 14118 } 14119 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 14120 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14121 /* Make sure snd_nxt is drug up */ 14122 tp->snd_nxt = tp->snd_max; 14123 } 14124 return (error); 14125 14126 } 14127 14128 /* 14129 * See bbr_output_wtime() for return values. 14130 */ 14131 static int 14132 bbr_output(struct tcpcb *tp) 14133 { 14134 int32_t ret; 14135 struct timeval tv; 14136 14137 NET_EPOCH_ASSERT(); 14138 14139 INP_WLOCK_ASSERT(tp->t_inpcb); 14140 (void)tcp_get_usecs(&tv); 14141 ret = bbr_output_wtime(tp, &tv); 14142 return (ret); 14143 } 14144 14145 static void 14146 bbr_mtu_chg(struct tcpcb *tp) 14147 { 14148 struct tcp_bbr *bbr; 14149 struct bbr_sendmap *rsm, *frsm = NULL; 14150 uint32_t maxseg; 14151 14152 /* 14153 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14154 * over the current size as SACK_PASS so a retransmit will occur. 14155 */ 14156 14157 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14158 maxseg = tp->t_maxseg - bbr->rc_last_options; 14159 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14160 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14161 /* Don't mess with ones acked (by sack?) */ 14162 if (rsm->r_flags & BBR_ACKED) 14163 continue; 14164 if ((rsm->r_end - rsm->r_start) > maxseg) { 14165 /* 14166 * We mark sack-passed on all the previous large 14167 * sends we did. This will force them to retransmit. 14168 */ 14169 rsm->r_flags |= BBR_SACK_PASSED; 14170 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14171 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14172 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14173 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14174 rsm->r_flags |= BBR_MARKED_LOST; 14175 } 14176 if (frsm == NULL) 14177 frsm = rsm; 14178 } 14179 } 14180 if (frsm) { 14181 bbr->r_ctl.rc_resend = frsm; 14182 } 14183 } 14184 14185 static int 14186 bbr_pru_options(struct tcpcb *tp, int flags) 14187 { 14188 if (flags & PRUS_OOB) 14189 return (EOPNOTSUPP); 14190 return (0); 14191 } 14192 14193 struct tcp_function_block __tcp_bbr = { 14194 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14195 .tfb_tcp_output = bbr_output, 14196 .tfb_do_queued_segments = ctf_do_queued_segments, 14197 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14198 .tfb_tcp_do_segment = bbr_do_segment, 14199 .tfb_tcp_ctloutput = bbr_ctloutput, 14200 .tfb_tcp_fb_init = bbr_init, 14201 .tfb_tcp_fb_fini = bbr_fini, 14202 .tfb_tcp_timer_stop_all = bbr_stopall, 14203 .tfb_tcp_timer_activate = bbr_timer_activate, 14204 .tfb_tcp_timer_active = bbr_timer_active, 14205 .tfb_tcp_timer_stop = bbr_timer_stop, 14206 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14207 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14208 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14209 .tfb_pru_options = bbr_pru_options, 14210 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14211 }; 14212 14213 /* 14214 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14215 * socket option arguments. When it re-acquires the lock after the copy, it 14216 * has to revalidate that the connection is still valid for the socket 14217 * option. 14218 */ 14219 static int 14220 bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt) 14221 { 14222 struct epoch_tracker et; 14223 struct tcpcb *tp; 14224 struct tcp_bbr *bbr; 14225 int32_t error = 0, optval; 14226 14227 switch (sopt->sopt_level) { 14228 case IPPROTO_IPV6: 14229 case IPPROTO_IP: 14230 return (tcp_default_ctloutput(inp, sopt)); 14231 } 14232 14233 switch (sopt->sopt_name) { 14234 case TCP_RACK_PACE_MAX_SEG: 14235 case TCP_RACK_MIN_TO: 14236 case TCP_RACK_REORD_THRESH: 14237 case TCP_RACK_REORD_FADE: 14238 case TCP_RACK_TLP_THRESH: 14239 case TCP_RACK_PKT_DELAY: 14240 case TCP_BBR_ALGORITHM: 14241 case TCP_BBR_TSLIMITS: 14242 case TCP_BBR_IWINTSO: 14243 case TCP_BBR_RECFORCE: 14244 case TCP_BBR_STARTUP_PG: 14245 case TCP_BBR_DRAIN_PG: 14246 case TCP_BBR_RWND_IS_APP: 14247 case TCP_BBR_PROBE_RTT_INT: 14248 case TCP_BBR_PROBE_RTT_GAIN: 14249 case TCP_BBR_PROBE_RTT_LEN: 14250 case TCP_BBR_STARTUP_LOSS_EXIT: 14251 case TCP_BBR_USEDEL_RATE: 14252 case TCP_BBR_MIN_RTO: 14253 case TCP_BBR_MAX_RTO: 14254 case TCP_BBR_PACE_PER_SEC: 14255 case TCP_DELACK: 14256 case TCP_BBR_PACE_DEL_TAR: 14257 case TCP_BBR_SEND_IWND_IN_TSO: 14258 case TCP_BBR_EXTRA_STATE: 14259 case TCP_BBR_UTTER_MAX_TSO: 14260 case TCP_BBR_MIN_TOPACEOUT: 14261 case TCP_BBR_FLOOR_MIN_TSO: 14262 case TCP_BBR_TSTMP_RAISES: 14263 case TCP_BBR_POLICER_DETECT: 14264 case TCP_BBR_USE_RACK_CHEAT: 14265 case TCP_DATA_AFTER_CLOSE: 14266 case TCP_BBR_HDWR_PACE: 14267 case TCP_BBR_PACE_SEG_MAX: 14268 case TCP_BBR_PACE_SEG_MIN: 14269 case TCP_BBR_PACE_CROSS: 14270 case TCP_BBR_PACE_OH: 14271 #ifdef NETFLIX_PEAKRATE 14272 case TCP_MAXPEAKRATE: 14273 #endif 14274 case TCP_BBR_TMR_PACE_OH: 14275 case TCP_BBR_RACK_RTT_USE: 14276 case TCP_BBR_RETRAN_WTSO: 14277 break; 14278 default: 14279 return (tcp_default_ctloutput(inp, sopt)); 14280 break; 14281 } 14282 INP_WUNLOCK(inp); 14283 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14284 if (error) 14285 return (error); 14286 INP_WLOCK(inp); 14287 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 14288 INP_WUNLOCK(inp); 14289 return (ECONNRESET); 14290 } 14291 tp = intotcpcb(inp); 14292 if (tp->t_fb != &__tcp_bbr) { 14293 INP_WUNLOCK(inp); 14294 return (ENOPROTOOPT); 14295 } 14296 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14297 switch (sopt->sopt_name) { 14298 case TCP_BBR_PACE_PER_SEC: 14299 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14300 bbr->r_ctl.bbr_hptsi_per_second = optval; 14301 break; 14302 case TCP_BBR_PACE_DEL_TAR: 14303 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14304 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14305 break; 14306 case TCP_BBR_PACE_SEG_MAX: 14307 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14308 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14309 break; 14310 case TCP_BBR_PACE_SEG_MIN: 14311 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14312 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14313 break; 14314 case TCP_BBR_PACE_CROSS: 14315 BBR_OPTS_INC(tcp_bbr_pace_cross); 14316 bbr->r_ctl.bbr_cross_over = optval; 14317 break; 14318 case TCP_BBR_ALGORITHM: 14319 BBR_OPTS_INC(tcp_bbr_algorithm); 14320 if (optval && (bbr->rc_use_google == 0)) { 14321 /* Turn on the google mode */ 14322 bbr_google_mode_on(bbr); 14323 if ((optval > 3) && (optval < 500)) { 14324 /* 14325 * Must be at least greater than .3% 14326 * and must be less than 50.0%. 14327 */ 14328 bbr->r_ctl.bbr_google_discount = optval; 14329 } 14330 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14331 /* Turn off the google mode */ 14332 bbr_google_mode_off(bbr); 14333 } 14334 break; 14335 case TCP_BBR_TSLIMITS: 14336 BBR_OPTS_INC(tcp_bbr_tslimits); 14337 if (optval == 1) 14338 bbr->rc_use_ts_limit = 1; 14339 else if (optval == 0) 14340 bbr->rc_use_ts_limit = 0; 14341 else 14342 error = EINVAL; 14343 break; 14344 14345 case TCP_BBR_IWINTSO: 14346 BBR_OPTS_INC(tcp_bbr_iwintso); 14347 if ((optval >= 0) && (optval < 128)) { 14348 uint32_t twin; 14349 14350 bbr->rc_init_win = optval; 14351 twin = bbr_initial_cwnd(bbr, tp); 14352 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14353 tp->snd_cwnd = twin; 14354 else 14355 error = EBUSY; 14356 } else 14357 error = EINVAL; 14358 break; 14359 case TCP_BBR_STARTUP_PG: 14360 BBR_OPTS_INC(tcp_bbr_startup_pg); 14361 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14362 bbr->r_ctl.rc_startup_pg = optval; 14363 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14364 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14365 } 14366 } else 14367 error = EINVAL; 14368 break; 14369 case TCP_BBR_DRAIN_PG: 14370 BBR_OPTS_INC(tcp_bbr_drain_pg); 14371 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14372 bbr->r_ctl.rc_drain_pg = optval; 14373 else 14374 error = EINVAL; 14375 break; 14376 case TCP_BBR_PROBE_RTT_LEN: 14377 BBR_OPTS_INC(tcp_bbr_probertt_len); 14378 if (optval <= 1) 14379 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14380 else 14381 error = EINVAL; 14382 break; 14383 case TCP_BBR_PROBE_RTT_GAIN: 14384 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14385 if (optval <= BBR_UNIT) 14386 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14387 else 14388 error = EINVAL; 14389 break; 14390 case TCP_BBR_PROBE_RTT_INT: 14391 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14392 if (optval > 1000) 14393 bbr->r_ctl.rc_probertt_int = optval; 14394 else 14395 error = EINVAL; 14396 break; 14397 case TCP_BBR_MIN_TOPACEOUT: 14398 BBR_OPTS_INC(tcp_bbr_topaceout); 14399 if (optval == 0) { 14400 bbr->no_pacing_until = 0; 14401 bbr->rc_no_pacing = 0; 14402 } else if (optval <= 0x00ff) { 14403 bbr->no_pacing_until = optval; 14404 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14405 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14406 /* Turn on no pacing */ 14407 bbr->rc_no_pacing = 1; 14408 } 14409 } else 14410 error = EINVAL; 14411 break; 14412 case TCP_BBR_STARTUP_LOSS_EXIT: 14413 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14414 bbr->rc_loss_exit = optval; 14415 break; 14416 case TCP_BBR_USEDEL_RATE: 14417 error = EINVAL; 14418 break; 14419 case TCP_BBR_MIN_RTO: 14420 BBR_OPTS_INC(tcp_bbr_min_rto); 14421 bbr->r_ctl.rc_min_rto_ms = optval; 14422 break; 14423 case TCP_BBR_MAX_RTO: 14424 BBR_OPTS_INC(tcp_bbr_max_rto); 14425 bbr->rc_max_rto_sec = optval; 14426 break; 14427 case TCP_RACK_MIN_TO: 14428 /* Minimum time between rack t-o's in ms */ 14429 BBR_OPTS_INC(tcp_rack_min_to); 14430 bbr->r_ctl.rc_min_to = optval; 14431 break; 14432 case TCP_RACK_REORD_THRESH: 14433 /* RACK reorder threshold (shift amount) */ 14434 BBR_OPTS_INC(tcp_rack_reord_thresh); 14435 if ((optval > 0) && (optval < 31)) 14436 bbr->r_ctl.rc_reorder_shift = optval; 14437 else 14438 error = EINVAL; 14439 break; 14440 case TCP_RACK_REORD_FADE: 14441 /* Does reordering fade after ms time */ 14442 BBR_OPTS_INC(tcp_rack_reord_fade); 14443 bbr->r_ctl.rc_reorder_fade = optval; 14444 break; 14445 case TCP_RACK_TLP_THRESH: 14446 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14447 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14448 if (optval) 14449 bbr->rc_tlp_threshold = optval; 14450 else 14451 error = EINVAL; 14452 break; 14453 case TCP_BBR_USE_RACK_CHEAT: 14454 BBR_OPTS_INC(tcp_use_rackcheat); 14455 if (bbr->rc_use_google) { 14456 error = EINVAL; 14457 break; 14458 } 14459 BBR_OPTS_INC(tcp_rack_cheat); 14460 if (optval) 14461 bbr->bbr_use_rack_cheat = 1; 14462 else 14463 bbr->bbr_use_rack_cheat = 0; 14464 break; 14465 case TCP_BBR_FLOOR_MIN_TSO: 14466 BBR_OPTS_INC(tcp_utter_max_tso); 14467 if ((optval >= 0) && (optval < 40)) 14468 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14469 else 14470 error = EINVAL; 14471 break; 14472 case TCP_BBR_UTTER_MAX_TSO: 14473 BBR_OPTS_INC(tcp_utter_max_tso); 14474 if ((optval >= 0) && (optval < 0xffff)) 14475 bbr->r_ctl.bbr_utter_max = optval; 14476 else 14477 error = EINVAL; 14478 break; 14479 14480 case TCP_BBR_EXTRA_STATE: 14481 BBR_OPTS_INC(tcp_extra_state); 14482 if (optval) 14483 bbr->rc_use_idle_restart = 1; 14484 else 14485 bbr->rc_use_idle_restart = 0; 14486 break; 14487 case TCP_BBR_SEND_IWND_IN_TSO: 14488 BBR_OPTS_INC(tcp_iwnd_tso); 14489 if (optval) { 14490 bbr->bbr_init_win_cheat = 1; 14491 if (bbr->rc_past_init_win == 0) { 14492 uint32_t cts; 14493 cts = tcp_get_usecs(&bbr->rc_tv); 14494 tcp_bbr_tso_size_check(bbr, cts); 14495 } 14496 } else 14497 bbr->bbr_init_win_cheat = 0; 14498 break; 14499 case TCP_BBR_HDWR_PACE: 14500 BBR_OPTS_INC(tcp_hdwr_pacing); 14501 if (optval){ 14502 bbr->bbr_hdw_pace_ena = 1; 14503 bbr->bbr_attempt_hdwr_pace = 0; 14504 } else { 14505 bbr->bbr_hdw_pace_ena = 0; 14506 #ifdef RATELIMIT 14507 if (bbr->r_ctl.crte != NULL) { 14508 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14509 bbr->r_ctl.crte = NULL; 14510 } 14511 #endif 14512 } 14513 break; 14514 14515 case TCP_DELACK: 14516 BBR_OPTS_INC(tcp_delack); 14517 if (optval < 100) { 14518 if (optval == 0) /* off */ 14519 tp->t_delayed_ack = 0; 14520 else if (optval == 1) /* on which is 2 */ 14521 tp->t_delayed_ack = 2; 14522 else /* higher than 2 and less than 100 */ 14523 tp->t_delayed_ack = optval; 14524 if (tp->t_flags & TF_DELACK) { 14525 tp->t_flags &= ~TF_DELACK; 14526 tp->t_flags |= TF_ACKNOW; 14527 NET_EPOCH_ENTER(et); 14528 bbr_output(tp); 14529 NET_EPOCH_EXIT(et); 14530 } 14531 } else 14532 error = EINVAL; 14533 break; 14534 case TCP_RACK_PKT_DELAY: 14535 /* RACK added ms i.e. rack-rtt + reord + N */ 14536 BBR_OPTS_INC(tcp_rack_pkt_delay); 14537 bbr->r_ctl.rc_pkt_delay = optval; 14538 break; 14539 #ifdef NETFLIX_PEAKRATE 14540 case TCP_MAXPEAKRATE: 14541 BBR_OPTS_INC(tcp_maxpeak); 14542 error = tcp_set_maxpeakrate(tp, optval); 14543 if (!error) 14544 tp->t_peakrate_thr = tp->t_maxpeakrate; 14545 break; 14546 #endif 14547 case TCP_BBR_RETRAN_WTSO: 14548 BBR_OPTS_INC(tcp_retran_wtso); 14549 if (optval) 14550 bbr->rc_resends_use_tso = 1; 14551 else 14552 bbr->rc_resends_use_tso = 0; 14553 break; 14554 case TCP_DATA_AFTER_CLOSE: 14555 BBR_OPTS_INC(tcp_data_ac); 14556 if (optval) 14557 bbr->rc_allow_data_af_clo = 1; 14558 else 14559 bbr->rc_allow_data_af_clo = 0; 14560 break; 14561 case TCP_BBR_POLICER_DETECT: 14562 BBR_OPTS_INC(tcp_policer_det); 14563 if (bbr->rc_use_google == 0) 14564 error = EINVAL; 14565 else if (optval) 14566 bbr->r_use_policer = 1; 14567 else 14568 bbr->r_use_policer = 0; 14569 break; 14570 14571 case TCP_BBR_TSTMP_RAISES: 14572 BBR_OPTS_INC(tcp_ts_raises); 14573 if (optval) 14574 bbr->ts_can_raise = 1; 14575 else 14576 bbr->ts_can_raise = 0; 14577 break; 14578 case TCP_BBR_TMR_PACE_OH: 14579 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14580 if (bbr->rc_use_google) { 14581 error = EINVAL; 14582 } else { 14583 if (optval) 14584 bbr->r_ctl.rc_incr_tmrs = 1; 14585 else 14586 bbr->r_ctl.rc_incr_tmrs = 0; 14587 } 14588 break; 14589 case TCP_BBR_PACE_OH: 14590 BBR_OPTS_INC(tcp_pacing_oh); 14591 if (bbr->rc_use_google) { 14592 error = EINVAL; 14593 } else { 14594 if (optval > (BBR_INCL_TCP_OH| 14595 BBR_INCL_IP_OH| 14596 BBR_INCL_ENET_OH)) { 14597 error = EINVAL; 14598 break; 14599 } 14600 if (optval & BBR_INCL_TCP_OH) 14601 bbr->r_ctl.rc_inc_tcp_oh = 1; 14602 else 14603 bbr->r_ctl.rc_inc_tcp_oh = 0; 14604 if (optval & BBR_INCL_IP_OH) 14605 bbr->r_ctl.rc_inc_ip_oh = 1; 14606 else 14607 bbr->r_ctl.rc_inc_ip_oh = 0; 14608 if (optval & BBR_INCL_ENET_OH) 14609 bbr->r_ctl.rc_inc_enet_oh = 1; 14610 else 14611 bbr->r_ctl.rc_inc_enet_oh = 0; 14612 } 14613 break; 14614 default: 14615 return (tcp_default_ctloutput(inp, sopt)); 14616 break; 14617 } 14618 #ifdef NETFLIX_STATS 14619 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14620 #endif 14621 INP_WUNLOCK(inp); 14622 return (error); 14623 } 14624 14625 /* 14626 * return 0 on success, error-num on failure 14627 */ 14628 static int 14629 bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt) 14630 { 14631 struct tcpcb *tp; 14632 struct tcp_bbr *bbr; 14633 int32_t error, optval; 14634 14635 tp = intotcpcb(inp); 14636 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14637 if (bbr == NULL) { 14638 INP_WUNLOCK(inp); 14639 return (EINVAL); 14640 } 14641 /* 14642 * Because all our options are either boolean or an int, we can just 14643 * pull everything into optval and then unlock and copy. If we ever 14644 * add a option that is not a int, then this will have quite an 14645 * impact to this routine. 14646 */ 14647 switch (sopt->sopt_name) { 14648 case TCP_BBR_PACE_PER_SEC: 14649 optval = bbr->r_ctl.bbr_hptsi_per_second; 14650 break; 14651 case TCP_BBR_PACE_DEL_TAR: 14652 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14653 break; 14654 case TCP_BBR_PACE_SEG_MAX: 14655 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14656 break; 14657 case TCP_BBR_MIN_TOPACEOUT: 14658 optval = bbr->no_pacing_until; 14659 break; 14660 case TCP_BBR_PACE_SEG_MIN: 14661 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14662 break; 14663 case TCP_BBR_PACE_CROSS: 14664 optval = bbr->r_ctl.bbr_cross_over; 14665 break; 14666 case TCP_BBR_ALGORITHM: 14667 optval = bbr->rc_use_google; 14668 break; 14669 case TCP_BBR_TSLIMITS: 14670 optval = bbr->rc_use_ts_limit; 14671 break; 14672 case TCP_BBR_IWINTSO: 14673 optval = bbr->rc_init_win; 14674 break; 14675 case TCP_BBR_STARTUP_PG: 14676 optval = bbr->r_ctl.rc_startup_pg; 14677 break; 14678 case TCP_BBR_DRAIN_PG: 14679 optval = bbr->r_ctl.rc_drain_pg; 14680 break; 14681 case TCP_BBR_PROBE_RTT_INT: 14682 optval = bbr->r_ctl.rc_probertt_int; 14683 break; 14684 case TCP_BBR_PROBE_RTT_LEN: 14685 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14686 break; 14687 case TCP_BBR_PROBE_RTT_GAIN: 14688 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14689 break; 14690 case TCP_BBR_STARTUP_LOSS_EXIT: 14691 optval = bbr->rc_loss_exit; 14692 break; 14693 case TCP_BBR_USEDEL_RATE: 14694 error = EINVAL; 14695 break; 14696 case TCP_BBR_MIN_RTO: 14697 optval = bbr->r_ctl.rc_min_rto_ms; 14698 break; 14699 case TCP_BBR_MAX_RTO: 14700 optval = bbr->rc_max_rto_sec; 14701 break; 14702 case TCP_RACK_PACE_MAX_SEG: 14703 /* Max segments in a pace */ 14704 optval = bbr->r_ctl.rc_pace_max_segs; 14705 break; 14706 case TCP_RACK_MIN_TO: 14707 /* Minimum time between rack t-o's in ms */ 14708 optval = bbr->r_ctl.rc_min_to; 14709 break; 14710 case TCP_RACK_REORD_THRESH: 14711 /* RACK reorder threshold (shift amount) */ 14712 optval = bbr->r_ctl.rc_reorder_shift; 14713 break; 14714 case TCP_RACK_REORD_FADE: 14715 /* Does reordering fade after ms time */ 14716 optval = bbr->r_ctl.rc_reorder_fade; 14717 break; 14718 case TCP_BBR_USE_RACK_CHEAT: 14719 /* Do we use the rack cheat for rxt */ 14720 optval = bbr->bbr_use_rack_cheat; 14721 break; 14722 case TCP_BBR_FLOOR_MIN_TSO: 14723 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14724 break; 14725 case TCP_BBR_UTTER_MAX_TSO: 14726 optval = bbr->r_ctl.bbr_utter_max; 14727 break; 14728 case TCP_BBR_SEND_IWND_IN_TSO: 14729 /* Do we send TSO size segments initially */ 14730 optval = bbr->bbr_init_win_cheat; 14731 break; 14732 case TCP_BBR_EXTRA_STATE: 14733 optval = bbr->rc_use_idle_restart; 14734 break; 14735 case TCP_RACK_TLP_THRESH: 14736 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14737 optval = bbr->rc_tlp_threshold; 14738 break; 14739 case TCP_RACK_PKT_DELAY: 14740 /* RACK added ms i.e. rack-rtt + reord + N */ 14741 optval = bbr->r_ctl.rc_pkt_delay; 14742 break; 14743 case TCP_BBR_RETRAN_WTSO: 14744 optval = bbr->rc_resends_use_tso; 14745 break; 14746 case TCP_DATA_AFTER_CLOSE: 14747 optval = bbr->rc_allow_data_af_clo; 14748 break; 14749 case TCP_DELACK: 14750 optval = tp->t_delayed_ack; 14751 break; 14752 case TCP_BBR_HDWR_PACE: 14753 optval = bbr->bbr_hdw_pace_ena; 14754 break; 14755 case TCP_BBR_POLICER_DETECT: 14756 optval = bbr->r_use_policer; 14757 break; 14758 case TCP_BBR_TSTMP_RAISES: 14759 optval = bbr->ts_can_raise; 14760 break; 14761 case TCP_BBR_TMR_PACE_OH: 14762 optval = bbr->r_ctl.rc_incr_tmrs; 14763 break; 14764 case TCP_BBR_PACE_OH: 14765 optval = 0; 14766 if (bbr->r_ctl.rc_inc_tcp_oh) 14767 optval |= BBR_INCL_TCP_OH; 14768 if (bbr->r_ctl.rc_inc_ip_oh) 14769 optval |= BBR_INCL_IP_OH; 14770 if (bbr->r_ctl.rc_inc_enet_oh) 14771 optval |= BBR_INCL_ENET_OH; 14772 break; 14773 default: 14774 return (tcp_default_ctloutput(inp, sopt)); 14775 break; 14776 } 14777 INP_WUNLOCK(inp); 14778 error = sooptcopyout(sopt, &optval, sizeof optval); 14779 return (error); 14780 } 14781 14782 /* 14783 * return 0 on success, error-num on failure 14784 */ 14785 static int 14786 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt) 14787 { 14788 if (sopt->sopt_dir == SOPT_SET) { 14789 return (bbr_set_sockopt(inp, sopt)); 14790 } else if (sopt->sopt_dir == SOPT_GET) { 14791 return (bbr_get_sockopt(inp, sopt)); 14792 } else { 14793 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14794 } 14795 } 14796 14797 static const char *bbr_stack_names[] = { 14798 __XSTRING(STACKNAME), 14799 #ifdef STACKALIAS 14800 __XSTRING(STACKALIAS), 14801 #endif 14802 }; 14803 14804 static bool bbr_mod_inited = false; 14805 14806 static int 14807 tcp_addbbr(module_t mod, int32_t type, void *data) 14808 { 14809 int32_t err = 0; 14810 int num_stacks; 14811 14812 switch (type) { 14813 case MOD_LOAD: 14814 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14815 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14816 sizeof(struct bbr_sendmap), 14817 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14818 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14819 sizeof(struct tcp_bbr), 14820 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14821 sysctl_ctx_init(&bbr_sysctl_ctx); 14822 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14823 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14824 OID_AUTO, 14825 #ifdef STACKALIAS 14826 __XSTRING(STACKALIAS), 14827 #else 14828 __XSTRING(STACKNAME), 14829 #endif 14830 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14831 ""); 14832 if (bbr_sysctl_root == NULL) { 14833 printf("Failed to add sysctl node\n"); 14834 err = EFAULT; 14835 goto free_uma; 14836 } 14837 bbr_init_sysctls(); 14838 num_stacks = nitems(bbr_stack_names); 14839 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14840 bbr_stack_names, &num_stacks); 14841 if (err) { 14842 printf("Failed to register %s stack name for " 14843 "%s module\n", bbr_stack_names[num_stacks], 14844 __XSTRING(MODNAME)); 14845 sysctl_ctx_free(&bbr_sysctl_ctx); 14846 free_uma: 14847 uma_zdestroy(bbr_zone); 14848 uma_zdestroy(bbr_pcb_zone); 14849 bbr_counter_destroy(); 14850 printf("Failed to register " __XSTRING(MODNAME) 14851 " module err:%d\n", err); 14852 return (err); 14853 } 14854 tcp_lro_reg_mbufq(); 14855 bbr_mod_inited = true; 14856 printf(__XSTRING(MODNAME) " is now available\n"); 14857 break; 14858 case MOD_QUIESCE: 14859 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14860 break; 14861 case MOD_UNLOAD: 14862 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14863 if (err == EBUSY) 14864 break; 14865 if (bbr_mod_inited) { 14866 uma_zdestroy(bbr_zone); 14867 uma_zdestroy(bbr_pcb_zone); 14868 sysctl_ctx_free(&bbr_sysctl_ctx); 14869 bbr_counter_destroy(); 14870 printf(__XSTRING(MODNAME) 14871 " is now no longer available\n"); 14872 bbr_mod_inited = false; 14873 } 14874 tcp_lro_dereg_mbufq(); 14875 err = 0; 14876 break; 14877 default: 14878 return (EOPNOTSUPP); 14879 } 14880 return (err); 14881 } 14882 14883 static moduledata_t tcp_bbr = { 14884 .name = __XSTRING(MODNAME), 14885 .evhand = tcp_addbbr, 14886 .priv = 0 14887 }; 14888 14889 MODULE_VERSION(MODNAME, 1); 14890 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14891 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14892