1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /** 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_ratelimit.h" 40 #include <sys/param.h> 41 #include <sys/arb.h> 42 #include <sys/module.h> 43 #include <sys/kernel.h> 44 #include <sys/libkern.h> 45 #ifdef TCP_HHOOK 46 #include <sys/hhook.h> 47 #endif 48 #include <sys/malloc.h> 49 #include <sys/mbuf.h> 50 #include <sys/proc.h> 51 #include <sys/socket.h> 52 #include <sys/socketvar.h> 53 #include <sys/sysctl.h> 54 #include <sys/systm.h> 55 #ifdef STATS 56 #include <sys/qmath.h> 57 #include <sys/tree.h> 58 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 59 #endif 60 #include <sys/refcount.h> 61 #include <sys/queue.h> 62 #include <sys/eventhandler.h> 63 #include <sys/smp.h> 64 #include <sys/kthread.h> 65 #include <sys/lock.h> 66 #include <sys/mutex.h> 67 #include <sys/tim_filter.h> 68 #include <sys/time.h> 69 #include <sys/protosw.h> 70 #include <vm/uma.h> 71 #include <sys/kern_prefetch.h> 72 73 #include <net/route.h> 74 #include <net/route/nhop.h> 75 #include <net/vnet.h> 76 77 #define TCPSTATES /* for logging */ 78 79 #include <netinet/in.h> 80 #include <netinet/in_kdtrace.h> 81 #include <netinet/in_pcb.h> 82 #include <netinet/ip.h> 83 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 84 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 85 #include <netinet/ip_var.h> 86 #include <netinet/ip6.h> 87 #include <netinet6/in6_pcb.h> 88 #include <netinet6/ip6_var.h> 89 #define TCPOUTFLAGS 90 #include <netinet/tcp.h> 91 #include <netinet/tcp_fsm.h> 92 #include <netinet/tcp_seq.h> 93 #include <netinet/tcp_timer.h> 94 #include <netinet/tcp_var.h> 95 #include <netinet/tcpip.h> 96 #include <netinet/tcp_hpts.h> 97 #include <netinet/cc/cc.h> 98 #include <netinet/tcp_log_buf.h> 99 #include <netinet/tcp_ratelimit.h> 100 #include <netinet/tcp_lro.h> 101 #ifdef TCP_OFFLOAD 102 #include <netinet/tcp_offload.h> 103 #endif 104 #ifdef INET6 105 #include <netinet6/tcp6_var.h> 106 #endif 107 #include <netinet/tcp_fastopen.h> 108 109 #include <netipsec/ipsec_support.h> 110 #include <net/if.h> 111 #include <net/if_var.h> 112 #include <net/ethernet.h> 113 114 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 115 #include <netipsec/ipsec.h> 116 #include <netipsec/ipsec6.h> 117 #endif /* IPSEC */ 118 119 #include <netinet/udp.h> 120 #include <netinet/udp_var.h> 121 #include <machine/in_cksum.h> 122 123 #ifdef MAC 124 #include <security/mac/mac_framework.h> 125 #endif 126 127 #include "sack_filter.h" 128 #include "tcp_bbr.h" 129 #include "rack_bbr_common.h" 130 uma_zone_t bbr_zone; 131 uma_zone_t bbr_pcb_zone; 132 133 struct sysctl_ctx_list bbr_sysctl_ctx; 134 struct sysctl_oid *bbr_sysctl_root; 135 136 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 137 (tv) = (value); \ 138 if ((u_long)(tv) < (u_long)(tvmin)) \ 139 (tv) = (tvmin); \ 140 if ((u_long)(tv) > (u_long)(tvmax)) \ 141 (tv) = (tvmax); \ 142 } while(0) 143 144 /*#define BBR_INVARIANT 1*/ 145 146 /* 147 * initial window 148 */ 149 static uint32_t bbr_def_init_win = 10; 150 static int32_t bbr_persist_min = 250000; /* 250ms */ 151 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 152 static int32_t bbr_cwnd_may_shrink = 0; 153 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 154 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 155 static int32_t bbr_hardware_pacing_limit = 8000; 156 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 157 static int32_t bbr_no_retran = 0; 158 159 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 160 static int32_t bbr_max_net_error_cnt = 10; 161 /* Should the following be dynamic too -- loss wise */ 162 static int32_t bbr_rtt_gain_thresh = 0; 163 /* Measurement controls */ 164 static int32_t bbr_use_google_algo = 1; 165 static int32_t bbr_ts_limiting = 1; 166 static int32_t bbr_ts_can_raise = 0; 167 static int32_t bbr_do_red = 600; 168 static int32_t bbr_red_scale = 20000; 169 static int32_t bbr_red_mul = 1; 170 static int32_t bbr_red_div = 2; 171 static int32_t bbr_red_growth_restrict = 1; 172 static int32_t bbr_target_is_bbunit = 0; 173 static int32_t bbr_drop_limit = 0; 174 /* 175 * How much gain do we need to see to 176 * stay in startup? 177 */ 178 static int32_t bbr_marks_rxt_sack_passed = 0; 179 static int32_t bbr_start_exit = 25; 180 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 181 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 182 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 183 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 184 * if we go back ever to where the pacer 185 * has priority over timers. 186 */ 187 static int32_t bbr_policer_call_from_rack_to = 0; 188 static int32_t bbr_policer_detection_enabled = 1; 189 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 190 * measurements before we are 191 * "good" note that 2 == 1. 192 * This is because we use a > 193 * comparison. This means if 194 * min_measure was 0, it takes 195 * num-measures > min(0) and 196 * you get 1 measurement and 197 * you are good. Set to 1, you 198 * have to have two 199 * measurements (this is done 200 * to prevent it from being ok 201 * to have no measurements). */ 202 static int32_t bbr_no_pacing_until = 4; 203 204 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 205 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 206 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 207 208 static int32_t bbr_target_cwnd_mult_limit = 8; 209 /* 210 * bbr_cwnd_min_val is the number of 211 * segments we hold to in the RTT probe 212 * state typically 4. 213 */ 214 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 215 216 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 217 218 static int32_t bbr_gain_to_target = 1; 219 static int32_t bbr_gain_gets_extra_too = 1; 220 /* 221 * bbr_high_gain is the 2/ln(2) value we need 222 * to double the sending rate in startup. This 223 * is used for both cwnd and hptsi gain's. 224 */ 225 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 226 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 227 static int32_t bbr_use_lower_gain_in_startup = 1; 228 229 /* thresholds for reduction on drain in sub-states/drain */ 230 static int32_t bbr_drain_rtt = BBR_SRTT; 231 static int32_t bbr_drain_floor = 88; 232 static int32_t google_allow_early_out = 1; 233 static int32_t google_consider_lost = 1; 234 static int32_t bbr_drain_drop_mul = 4; 235 static int32_t bbr_drain_drop_div = 5; 236 static int32_t bbr_rand_ot = 50; 237 static int32_t bbr_can_force_probertt = 0; 238 static int32_t bbr_can_adjust_probertt = 1; 239 static int32_t bbr_probertt_sets_rtt = 0; 240 static int32_t bbr_can_use_ts_for_rtt = 1; 241 static int32_t bbr_is_ratio = 0; 242 static int32_t bbr_sub_drain_app_limit = 1; 243 static int32_t bbr_prtt_slam_cwnd = 1; 244 static int32_t bbr_sub_drain_slam_cwnd = 1; 245 static int32_t bbr_slam_cwnd_in_main_drain = 1; 246 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 247 * hold */ 248 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 249 /* 250 * bbr_drain_gain is the reverse of the high_gain 251 * designed to drain back out the standing queue 252 * that is formed in startup by causing a larger 253 * hptsi gain and thus drainging the packets 254 * in flight. 255 */ 256 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 257 static int32_t bbr_rttprobe_gain = 192; 258 259 /* 260 * The cwnd_gain is the default cwnd gain applied when 261 * calculating a target cwnd. Note that the cwnd is 262 * a secondary factor in the way BBR works (see the 263 * paper and think about it, it will take some time). 264 * Basically the hptsi_gain spreads the packets out 265 * so you never get more than BDP to the peer even 266 * if the cwnd is high. In our implemenation that 267 * means in non-recovery/retransmission scenarios 268 * cwnd will never be reached by the flight-size. 269 */ 270 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 271 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 272 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 273 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 274 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 275 static int32_t bbr_ignore_data_after_close = 1; 276 static int16_t bbr_hptsi_gain[] = { 277 (BBR_UNIT *5 / 4), 278 (BBR_UNIT * 3 / 4), 279 BBR_UNIT, 280 BBR_UNIT, 281 BBR_UNIT, 282 BBR_UNIT, 283 BBR_UNIT, 284 BBR_UNIT 285 }; 286 int32_t bbr_use_rack_resend_cheat = 1; 287 int32_t bbr_sends_full_iwnd = 1; 288 289 #define BBR_HPTSI_GAIN_MAX 8 290 /* 291 * The BBR module incorporates a number of 292 * TCP ideas that have been put out into the IETF 293 * over the last few years: 294 * - Yuchung Cheng's RACK TCP (for which its named) that 295 * will stop us using the number of dup acks and instead 296 * use time as the gage of when we retransmit. 297 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 298 * of Dukkipati et.al. 299 * - Van Jacobson's et.al BBR. 300 * 301 * RACK depends on SACK, so if an endpoint arrives that 302 * cannot do SACK the state machine below will shuttle the 303 * connection back to using the "default" TCP stack that is 304 * in FreeBSD. 305 * 306 * To implement BBR and RACK the original TCP stack was first decomposed 307 * into a functional state machine with individual states 308 * for each of the possible TCP connection states. The do_segment 309 * functions role in life is to mandate the connection supports SACK 310 * initially and then assure that the RACK state matches the conenction 311 * state before calling the states do_segment function. Data processing 312 * of inbound segments also now happens in the hpts_do_segment in general 313 * with only one exception. This is so we can keep the connection on 314 * a single CPU. 315 * 316 * Each state is simplified due to the fact that the original do_segment 317 * has been decomposed and we *know* what state we are in (no 318 * switches on the state) and all tests for SACK are gone. This 319 * greatly simplifies what each state does. 320 * 321 * TCP output is also over-written with a new version since it 322 * must maintain the new rack scoreboard and has had hptsi 323 * integrated as a requirment. Still todo is to eliminate the 324 * use of the callout_() system and use the hpts for all 325 * timers as well. 326 */ 327 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 328 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 329 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 330 * free list */ 331 static int32_t bbr_tlp_thresh = 1; 332 static int32_t bbr_reorder_thresh = 2; 333 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 334 * 60,000,000 - 60 seconds */ 335 static int32_t bbr_pkt_delay = 1000; 336 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 337 static int32_t bbr_incr_timers = 1; 338 339 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 340 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 341 static int32_t bbr_exit_startup_at_loss = 1; 342 343 /* 344 * bbr_lt_bw_ratio is 1/8th 345 * bbr_lt_bw_diff is < 4 Kbit/sec 346 */ 347 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 348 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 349 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 350 * the lt_bw for */ 351 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 352 * lt_bw */ 353 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 354 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 355 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 356 357 static int32_t bbr_verbose_logging = 0; 358 /* 359 * Currently regular tcp has a rto_min of 30ms 360 * the backoff goes 12 times so that ends up 361 * being a total of 122.850 seconds before a 362 * connection is killed. 363 */ 364 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 365 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 366 367 /****************************************************/ 368 /* DEFAULT TSO SIZING (cpu performance impacting) */ 369 /****************************************************/ 370 /* What amount is our formula using to get TSO size */ 371 static int32_t bbr_hptsi_per_second = 1000; 372 373 /* 374 * For hptsi under bbr_cross_over connections what is delay 375 * target 7ms (in usec) combined with a seg_max of 2 376 * gets us close to identical google behavior in 377 * TSO size selection (possibly more 1MSS sends). 378 */ 379 static int32_t bbr_hptsi_segments_delay_tar = 7000; 380 381 /* Does pacing delay include overhead's in its time calculations? */ 382 static int32_t bbr_include_enet_oh = 0; 383 static int32_t bbr_include_ip_oh = 1; 384 static int32_t bbr_include_tcp_oh = 1; 385 static int32_t bbr_google_discount = 10; 386 387 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 388 static int32_t bbr_state_is_pkt_epoch = 0; 389 static int32_t bbr_state_drain_2_tar = 1; 390 /* What is the max the 0 - bbr_cross_over MBPS TSO target 391 * can reach using our delay target. Note that this 392 * value becomes the floor for the cross over 393 * algorithm. 394 */ 395 static int32_t bbr_hptsi_segments_max = 2; 396 static int32_t bbr_hptsi_segments_floor = 1; 397 static int32_t bbr_hptsi_utter_max = 0; 398 399 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 400 static int32_t bbr_hptsi_bytes_min = 1460; 401 static int32_t bbr_all_get_min = 0; 402 403 /* Cross over point from algo-a to algo-b */ 404 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 405 406 /* Do we deal with our restart state? */ 407 static int32_t bbr_uses_idle_restart = 0; 408 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 409 410 /* Do we allow hardware pacing? */ 411 static int32_t bbr_allow_hdwr_pacing = 0; 412 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 413 static int32_t bbr_hdwr_pace_floor = 1; 414 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 415 416 /****************************************************/ 417 static int32_t bbr_resends_use_tso = 0; 418 static int32_t bbr_tlp_max_resend = 2; 419 static int32_t bbr_sack_block_limit = 128; 420 421 #define BBR_MAX_STAT 19 422 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 423 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 424 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 425 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 426 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 427 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 428 counter_u64_t bbr_flows_whdwr_pacing; 429 counter_u64_t bbr_flows_nohdwr_pacing; 430 431 counter_u64_t bbr_nohdwr_pacing_enobuf; 432 counter_u64_t bbr_hdwr_pacing_enobuf; 433 434 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 435 436 /* 437 * Static defintions we need for forward declarations. 438 */ 439 static uint32_t 440 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 441 uint32_t useconds_time, uint64_t bw); 442 static uint32_t 443 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 444 static void 445 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 446 static void 447 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 448 static void 449 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 450 int dolog); 451 static uint32_t 452 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 453 static void 454 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 455 int32_t pkt_epoch, uint32_t losses); 456 static uint32_t 457 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, 458 struct bbr_sendmap *rsm); 459 static uint32_t 460 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 461 static uint32_t 462 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 463 struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); 464 static void 465 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 466 int32_t line); 467 static void 468 bbr_set_state_target(struct tcp_bbr *bbr, int line); 469 static void 470 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 471 static void 472 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, 473 int event, int line); 474 static void 475 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 476 static void 477 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 478 static void 479 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 480 uint32_t rtt, uint32_t line, uint8_t is_start, 481 uint16_t set); 482 static struct bbr_sendmap * 483 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 484 static __inline uint32_t 485 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 486 static void 487 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, 488 uint8_t which); 489 static void 490 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, 491 uint32_t time_since_sent, uint32_t srtt, 492 uint32_t thresh, uint32_t to); 493 static void 494 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 495 static void 496 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 497 uint32_t del_by, uint32_t cts, uint32_t sloton, 498 uint32_t prev_delay); 499 static void 500 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 501 int32_t line); 502 static void 503 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr); 504 static void 505 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 506 static void 507 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 508 static void 509 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 510 static void 511 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 512 uint32_t cts, uint32_t usecs, uint64_t bw, 513 uint32_t override, int mod); 514 static int bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt); 515 516 static inline uint8_t 517 bbr_state_val(struct tcp_bbr *bbr) 518 { 519 return(bbr->rc_bbr_substate); 520 } 521 522 static inline uint32_t 523 get_min_cwnd(struct tcp_bbr *bbr) 524 { 525 int mss; 526 527 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 528 bbr->r_ctl.rc_pace_max_segs); 529 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 530 return (bbr_cwnd_min_val_hs * mss); 531 else 532 return (bbr_cwnd_min_val * mss); 533 } 534 535 static uint32_t 536 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 537 { 538 uint64_t srtt, var; 539 uint64_t ret_val; 540 541 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 542 if (tp->t_srtt == 0) { 543 srtt = (uint64_t)BBR_INITIAL_RTO; 544 var = 0; 545 } else { 546 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 547 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 548 } 549 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 550 bbr_persist_min, bbr_persist_max); 551 return ((uint32_t)ret_val); 552 } 553 554 static uint32_t 555 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 556 { 557 /* 558 * Start the FR timer, we do this based on getting the first one in 559 * the rc_tmap. Note that if its NULL we must stop the timer. in all 560 * events we need to stop the running timer (if its running) before 561 * starting the new one. 562 */ 563 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 564 int32_t idx; 565 int32_t is_tlp_timer = 0; 566 struct bbr_sendmap *rsm; 567 568 if (bbr->rc_all_timers_stopped) { 569 /* All timers have been stopped none are to run */ 570 return (0); 571 } 572 if (bbr->rc_in_persist) { 573 /* We can't start any timer in persists */ 574 return (bbr_get_persists_timer_val(tp, bbr)); 575 } 576 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 577 if ((rsm == NULL) || 578 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 579 (tp->t_state < TCPS_ESTABLISHED)) { 580 /* Nothing on the send map */ 581 activate_rxt: 582 if (SEQ_LT(tp->snd_una, tp->snd_max) || 583 sbavail(&tptosocket(tp)->so_snd)) { 584 uint64_t tov; 585 586 time_since_sent = 0; 587 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 588 if (rsm) { 589 idx = rsm->r_rtr_cnt - 1; 590 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 591 tstmp_touse = rsm->r_tim_lastsent[idx]; 592 else 593 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 594 if (TSTMP_GT(tstmp_touse, cts)) 595 time_since_sent = cts - tstmp_touse; 596 } 597 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 598 if (tp->t_srtt == 0) 599 tov = BBR_INITIAL_RTO; 600 else 601 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 602 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 603 if (tp->t_rxtshift) 604 tov *= tcp_backoff[tp->t_rxtshift]; 605 if (tov > time_since_sent) 606 tov -= time_since_sent; 607 else 608 tov = bbr->r_ctl.rc_min_to; 609 TCPT_RANGESET_NOSLOP(to, tov, 610 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 611 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 612 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 613 return (to); 614 } 615 return (0); 616 } 617 if (rsm->r_flags & BBR_ACKED) { 618 rsm = bbr_find_lowest_rsm(bbr); 619 if (rsm == NULL) { 620 /* No lowest? */ 621 goto activate_rxt; 622 } 623 } 624 /* Convert from ms to usecs */ 625 if (rsm->r_flags & BBR_SACK_PASSED) { 626 if ((tp->t_flags & TF_SENTFIN) && 627 ((tp->snd_max - tp->snd_una) == 1) && 628 (rsm->r_flags & BBR_HAS_FIN)) { 629 /* 630 * We don't start a bbr rack timer if all we have is 631 * a FIN outstanding. 632 */ 633 goto activate_rxt; 634 } 635 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 636 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 637 idx = rsm->r_rtr_cnt - 1; 638 exp = rsm->r_tim_lastsent[idx] + thresh; 639 if (SEQ_GEQ(exp, cts)) { 640 to = exp - cts; 641 if (to < bbr->r_ctl.rc_min_to) { 642 to = bbr->r_ctl.rc_min_to; 643 } 644 } else { 645 to = bbr->r_ctl.rc_min_to; 646 } 647 } else { 648 /* Ok we need to do a TLP not RACK */ 649 if (bbr->rc_tlp_in_progress != 0) { 650 /* 651 * The previous send was a TLP. 652 */ 653 goto activate_rxt; 654 } 655 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 656 if (rsm == NULL) { 657 /* We found no rsm to TLP with. */ 658 goto activate_rxt; 659 } 660 if (rsm->r_flags & BBR_HAS_FIN) { 661 /* If its a FIN we don't do TLP */ 662 rsm = NULL; 663 goto activate_rxt; 664 } 665 time_since_sent = 0; 666 idx = rsm->r_rtr_cnt - 1; 667 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 668 tstmp_touse = rsm->r_tim_lastsent[idx]; 669 else 670 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 671 if (TSTMP_GT(tstmp_touse, cts)) 672 time_since_sent = cts - tstmp_touse; 673 is_tlp_timer = 1; 674 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 675 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 676 if (thresh > time_since_sent) 677 to = thresh - time_since_sent; 678 else 679 to = bbr->r_ctl.rc_min_to; 680 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 681 /* 682 * If the TLP time works out to larger than the max 683 * RTO lets not do TLP.. just RTO. 684 */ 685 goto activate_rxt; 686 } 687 if ((bbr->rc_tlp_rtx_out == 1) && 688 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 689 /* 690 * Second retransmit of the same TLP 691 * lets not. 692 */ 693 bbr->rc_tlp_rtx_out = 0; 694 goto activate_rxt; 695 } 696 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 697 /* 698 * The tail is no longer the last one I did a probe 699 * on 700 */ 701 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 702 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 703 } 704 } 705 if (is_tlp_timer == 0) { 706 BBR_STAT_INC(bbr_to_arm_rack); 707 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 708 } else { 709 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 710 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 711 /* 712 * We have exceeded how many times we can retran the 713 * current TLP timer, switch to the RTO timer. 714 */ 715 goto activate_rxt; 716 } else { 717 BBR_STAT_INC(bbr_to_arm_tlp); 718 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 719 } 720 } 721 return (to); 722 } 723 724 static inline int32_t 725 bbr_minseg(struct tcp_bbr *bbr) 726 { 727 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 728 } 729 730 static void 731 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 732 { 733 struct inpcb *inp = tptoinpcb(tp); 734 struct hpts_diag diag; 735 uint32_t delayed_ack = 0; 736 uint32_t left = 0; 737 uint32_t hpts_timeout; 738 uint8_t stopped; 739 int32_t delay_calc = 0; 740 uint32_t prev_delay = 0; 741 742 if (tcp_in_hpts(tp)) { 743 /* A previous call is already set up */ 744 return; 745 } 746 if ((tp->t_state == TCPS_CLOSED) || 747 (tp->t_state == TCPS_LISTEN)) { 748 return; 749 } 750 stopped = bbr->rc_tmr_stopped; 751 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 752 left = bbr->r_ctl.rc_timer_exp - cts; 753 } 754 bbr->r_ctl.rc_hpts_flags = 0; 755 bbr->r_ctl.rc_timer_exp = 0; 756 prev_delay = bbr->r_ctl.rc_last_delay_val; 757 if (bbr->r_ctl.rc_last_delay_val && 758 (slot == 0)) { 759 /* 760 * If a previous pacer delay was in place we 761 * are not coming from the output side (where 762 * we calculate a delay, more likely a timer). 763 */ 764 slot = bbr->r_ctl.rc_last_delay_val; 765 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 766 /* Compensate for time passed */ 767 delay_calc = cts - bbr->rc_pacer_started; 768 if (delay_calc <= slot) 769 slot -= delay_calc; 770 } 771 } 772 /* Do we have early to make up for by pushing out the pacing time? */ 773 if (bbr->r_agg_early_set) { 774 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 775 slot += bbr->r_ctl.rc_agg_early; 776 bbr->r_ctl.rc_agg_early = 0; 777 bbr->r_agg_early_set = 0; 778 } 779 /* Are we running a total debt that needs to be compensated for? */ 780 if (bbr->r_ctl.rc_hptsi_agg_delay) { 781 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 782 /* We nuke the delay */ 783 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 784 bbr->r_ctl.rc_hptsi_agg_delay = 0; 785 } else { 786 /* We nuke some of the delay, put in a minimal 100usecs */ 787 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 788 bbr->r_ctl.rc_last_delay_val = slot = 100; 789 } 790 } 791 bbr->r_ctl.rc_last_delay_val = slot; 792 hpts_timeout = bbr_timer_start(tp, bbr, cts); 793 if (tp->t_flags & TF_DELACK) { 794 if (bbr->rc_in_persist == 0) { 795 delayed_ack = bbr_delack_time; 796 } else { 797 /* 798 * We are in persists and have 799 * gotten a new data element. 800 */ 801 if (hpts_timeout > bbr_delack_time) { 802 /* 803 * Lets make the persists timer (which acks) 804 * be the smaller of hpts_timeout and bbr_delack_time. 805 */ 806 hpts_timeout = bbr_delack_time; 807 } 808 } 809 } 810 if (delayed_ack && 811 ((hpts_timeout == 0) || 812 (delayed_ack < hpts_timeout))) { 813 /* We need a Delayed ack timer */ 814 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 815 hpts_timeout = delayed_ack; 816 } 817 if (slot) { 818 /* Mark that we have a pacing timer up */ 819 BBR_STAT_INC(bbr_paced_segments); 820 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 821 } 822 /* 823 * If no timers are going to run and we will fall off thfe hptsi 824 * wheel, we resort to a keep-alive timer if its configured. 825 */ 826 if ((hpts_timeout == 0) && 827 (slot == 0)) { 828 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 829 (tp->t_state <= TCPS_CLOSING)) { 830 /* 831 * Ok we have no timer (persists, rack, tlp, rxt or 832 * del-ack), we don't have segments being paced. So 833 * all that is left is the keepalive timer. 834 */ 835 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 836 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 837 } else { 838 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 839 } 840 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 841 } 842 } 843 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 844 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 845 /* 846 * RACK, TLP, persists and RXT timers all are restartable 847 * based on actions input .. i.e we received a packet (ack 848 * or sack) and that changes things (rw, or snd_una etc). 849 * Thus we can restart them with a new value. For 850 * keep-alive, delayed_ack we keep track of what was left 851 * and restart the timer with a smaller value. 852 */ 853 if (left < hpts_timeout) 854 hpts_timeout = left; 855 } 856 if (bbr->r_ctl.rc_incr_tmrs && slot && 857 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 858 /* 859 * If configured to do so, and the timer is either 860 * the TLP or RXT timer, we need to increase the timeout 861 * by the pacing time. Consider the bottleneck at my 862 * machine as an example, we are sending something 863 * to start a TLP on. The last packet won't be emitted 864 * fully until the pacing time (the bottleneck will hold 865 * the data in place). Once the packet is emitted that 866 * is when we want to start waiting for the TLP. This 867 * is most evident with hardware pacing (where the nic 868 * is holding the packet(s) before emitting). But it 869 * can also show up in the network so we do it for all 870 * cases. Technically we would take off one packet from 871 * this extra delay but this is easier and being more 872 * conservative is probably better. 873 */ 874 hpts_timeout += slot; 875 } 876 if (hpts_timeout) { 877 /* 878 * Hack alert for now we can't time-out over 2147 seconds (a 879 * bit more than 35min) 880 */ 881 if (hpts_timeout > 0x7ffffffe) 882 hpts_timeout = 0x7ffffffe; 883 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 884 } else 885 bbr->r_ctl.rc_timer_exp = 0; 886 if ((slot) && 887 (bbr->rc_use_google || 888 bbr->output_error_seen || 889 (slot <= hpts_timeout)) ) { 890 /* 891 * Tell LRO that it can queue packets while 892 * we pace. 893 */ 894 bbr->rc_tp->t_flags2 |= TF2_MBUF_QUEUE_READY; 895 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 896 (bbr->rc_cwnd_limited == 0)) { 897 /* 898 * If we are not cwnd limited and we 899 * are running a rack timer we put on 900 * the do not disturbe even for sack. 901 */ 902 tp->t_flags2 |= TF2_DONT_SACK_QUEUE; 903 } else 904 tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE; 905 bbr->rc_pacer_started = cts; 906 907 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot), 908 __LINE__, &diag); 909 bbr->rc_timer_first = 0; 910 bbr->bbr_timer_src = frm; 911 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 912 bbr_log_hpts_diag(bbr, cts, &diag); 913 } else if (hpts_timeout) { 914 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout), 915 __LINE__, &diag); 916 /* 917 * We add the flag here as well if the slot is set, 918 * since hpts will call in to clear the queue first before 919 * calling the output routine (which does our timers). 920 * We don't want to set the flag if its just a timer 921 * else the arrival of data might (that causes us 922 * to send more) might get delayed. Imagine being 923 * on a keep-alive timer and a request comes in for 924 * more data. 925 */ 926 if (slot) 927 bbr->rc_pacer_started = cts; 928 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 929 (bbr->rc_cwnd_limited == 0)) { 930 /* 931 * For a rack timer, don't wake us even 932 * if a sack arrives as long as we are 933 * not cwnd limited. 934 */ 935 tp->t_flags2 |= (TF2_MBUF_QUEUE_READY | 936 TF2_DONT_SACK_QUEUE); 937 } else { 938 /* All other timers wake us up */ 939 tp->t_flags2 &= ~(TF2_MBUF_QUEUE_READY | 940 TF2_DONT_SACK_QUEUE); 941 } 942 bbr->bbr_timer_src = frm; 943 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 944 bbr_log_hpts_diag(bbr, cts, &diag); 945 bbr->rc_timer_first = 1; 946 } 947 bbr->rc_tmr_stopped = 0; 948 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 949 } 950 951 static void 952 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 953 { 954 /* 955 * We received an ack, and then did not call send or were bounced 956 * out due to the hpts was running. Now a timer is up as well, is it 957 * the right timer? 958 */ 959 struct inpcb *inp; 960 struct bbr_sendmap *rsm; 961 uint32_t hpts_timeout; 962 int tmr_up; 963 964 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 965 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 966 return; 967 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 968 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 969 (tmr_up == PACE_TMR_RXT)) { 970 /* Should be an RXT */ 971 return; 972 } 973 inp = bbr->rc_inp; 974 if (rsm == NULL) { 975 /* Nothing outstanding? */ 976 if (tp->t_flags & TF_DELACK) { 977 if (tmr_up == PACE_TMR_DELACK) 978 /* 979 * We are supposed to have delayed ack up 980 * and we do 981 */ 982 return; 983 } else if (sbavail(&inp->inp_socket->so_snd) && 984 (tmr_up == PACE_TMR_RXT)) { 985 /* 986 * if we hit enobufs then we would expect the 987 * possibility of nothing outstanding and the RXT up 988 * (and the hptsi timer). 989 */ 990 return; 991 } else if (((V_tcp_always_keepalive || 992 inp->inp_socket->so_options & SO_KEEPALIVE) && 993 (tp->t_state <= TCPS_CLOSING)) && 994 (tmr_up == PACE_TMR_KEEP) && 995 (tp->snd_max == tp->snd_una)) { 996 /* We should have keep alive up and we do */ 997 return; 998 } 999 } 1000 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 1001 if ((tp->t_flags & TF_SENTFIN) && 1002 ((tp->snd_max - tp->snd_una) == 1) && 1003 (rsm->r_flags & BBR_HAS_FIN)) { 1004 /* needs to be a RXT */ 1005 if (tmr_up == PACE_TMR_RXT) 1006 return; 1007 else 1008 goto wrong_timer; 1009 } else if (tmr_up == PACE_TMR_RACK) 1010 return; 1011 else 1012 goto wrong_timer; 1013 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1014 /* Rack timer has priority if we have data out */ 1015 return; 1016 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1017 ((tmr_up == PACE_TMR_TLP) || 1018 (tmr_up == PACE_TMR_RXT))) { 1019 /* 1020 * Either a TLP or RXT is fine if no sack-passed is in place 1021 * and data is outstanding. 1022 */ 1023 return; 1024 } else if (tmr_up == PACE_TMR_DELACK) { 1025 /* 1026 * If the delayed ack was going to go off before the 1027 * rtx/tlp/rack timer were going to expire, then that would 1028 * be the timer in control. Note we don't check the time 1029 * here trusting the code is correct. 1030 */ 1031 return; 1032 } 1033 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1034 ((tmr_up == PACE_TMR_RXT) || 1035 (tmr_up == PACE_TMR_TLP) || 1036 (tmr_up == PACE_TMR_RACK))) { 1037 /* 1038 * We have outstanding data and 1039 * we *do* have a RACK, TLP or RXT 1040 * timer running. We won't restart 1041 * anything here since thats probably ok we 1042 * will get called with some timer here shortly. 1043 */ 1044 return; 1045 } 1046 /* 1047 * Ok the timer originally started is not what we want now. We will 1048 * force the hpts to be stopped if any, and restart with the slot 1049 * set to what was in the saved slot. 1050 */ 1051 wrong_timer: 1052 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1053 if (tcp_in_hpts(tp)) 1054 tcp_hpts_remove(tp); 1055 bbr_timer_cancel(bbr, __LINE__, cts); 1056 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1057 0); 1058 } else { 1059 /* 1060 * Output is hptsi so we just need to switch the type of 1061 * timer. We don't bother with keep-alive, since when we 1062 * jump through the output, it will start the keep-alive if 1063 * nothing is sent. 1064 * 1065 * We only need a delayed-ack added and or the hpts_timeout. 1066 */ 1067 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1068 if (tp->t_flags & TF_DELACK) { 1069 if (hpts_timeout == 0) { 1070 hpts_timeout = bbr_delack_time; 1071 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1072 } 1073 else if (hpts_timeout > bbr_delack_time) { 1074 hpts_timeout = bbr_delack_time; 1075 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1076 } 1077 } 1078 if (hpts_timeout) { 1079 if (hpts_timeout > 0x7ffffffe) 1080 hpts_timeout = 0x7ffffffe; 1081 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1082 } 1083 } 1084 } 1085 1086 int32_t bbr_clear_lost = 0; 1087 1088 /* 1089 * Considers the two time values now (cts) and earlier. 1090 * If cts is smaller than earlier, we could have 1091 * had a sequence wrap (our counter wraps every 1092 * 70 min or so) or it could be just clock skew 1093 * getting us two different time values. Clock skew 1094 * will show up within 10ms or so. So in such 1095 * a case (where cts is behind earlier time by 1096 * less than 10ms) we return 0. Otherwise we 1097 * return the true difference between them. 1098 */ 1099 static inline uint32_t 1100 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1101 /* 1102 * Given two timestamps, the current time stamp cts, and some other 1103 * time-stamp taken in theory earlier return the difference. The 1104 * trick is here sometimes locking will get the other timestamp 1105 * after the cts. If this occurs we need to return 0. 1106 */ 1107 if (TSTMP_GEQ(cts, earlier_time)) 1108 return (cts - earlier_time); 1109 /* 1110 * cts is behind earlier_time if its less than 10ms consider it 0. 1111 * If its more than 10ms difference then we had a time wrap. Else 1112 * its just the normal locking foo. I wonder if we should not go to 1113 * 64bit TS and get rid of this issue. 1114 */ 1115 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1116 return (0); 1117 /* 1118 * Ok the time must have wrapped. So we need to answer a large 1119 * amount of time, which the normal subtraction should do. 1120 */ 1121 return (cts - earlier_time); 1122 } 1123 1124 static int 1125 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1126 { 1127 uint32_t stat; 1128 int32_t error; 1129 1130 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1131 if (error || req->newptr == NULL) 1132 return error; 1133 1134 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1135 if (error) 1136 return (error); 1137 if (stat == 1) { 1138 #ifdef BBR_INVARIANTS 1139 printf("Clearing BBR lost counters\n"); 1140 #endif 1141 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1142 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1143 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1144 } else if (stat == 2) { 1145 #ifdef BBR_INVARIANTS 1146 printf("Clearing BBR option counters\n"); 1147 #endif 1148 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1149 } else if (stat == 3) { 1150 #ifdef BBR_INVARIANTS 1151 printf("Clearing BBR stats counters\n"); 1152 #endif 1153 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1154 } else if (stat == 4) { 1155 #ifdef BBR_INVARIANTS 1156 printf("Clearing BBR out-size counters\n"); 1157 #endif 1158 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1159 } 1160 bbr_clear_lost = 0; 1161 return (0); 1162 } 1163 1164 static void 1165 bbr_init_sysctls(void) 1166 { 1167 struct sysctl_oid *bbr_probertt; 1168 struct sysctl_oid *bbr_hptsi; 1169 struct sysctl_oid *bbr_measure; 1170 struct sysctl_oid *bbr_cwnd; 1171 struct sysctl_oid *bbr_timeout; 1172 struct sysctl_oid *bbr_states; 1173 struct sysctl_oid *bbr_startup; 1174 struct sysctl_oid *bbr_policer; 1175 1176 /* Probe rtt controls */ 1177 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1178 SYSCTL_CHILDREN(bbr_sysctl_root), 1179 OID_AUTO, 1180 "probertt", 1181 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1182 ""); 1183 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1184 SYSCTL_CHILDREN(bbr_probertt), 1185 OID_AUTO, "gain", CTLFLAG_RW, 1186 &bbr_rttprobe_gain, 192, 1187 "What is the filter gain drop in probe_rtt (0=disable)?"); 1188 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1189 SYSCTL_CHILDREN(bbr_probertt), 1190 OID_AUTO, "cwnd", CTLFLAG_RW, 1191 &bbr_rtt_probe_cwndtarg, 4, 1192 "How many mss's are outstanding during probe-rtt"); 1193 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1194 SYSCTL_CHILDREN(bbr_probertt), 1195 OID_AUTO, "int", CTLFLAG_RW, 1196 &bbr_rtt_probe_limit, 4000000, 1197 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1198 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1199 SYSCTL_CHILDREN(bbr_probertt), 1200 OID_AUTO, "mintime", CTLFLAG_RW, 1201 &bbr_rtt_probe_time, 200000, 1202 "How many microseconds in probe-rtt"); 1203 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1204 SYSCTL_CHILDREN(bbr_probertt), 1205 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1206 &bbr_filter_len_sec, 6, 1207 "How long in seconds does the rttProp filter run?"); 1208 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1209 SYSCTL_CHILDREN(bbr_probertt), 1210 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1211 &bbr_drain_rtt, BBR_SRTT, 1212 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1213 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1214 SYSCTL_CHILDREN(bbr_probertt), 1215 OID_AUTO, "can_force", CTLFLAG_RW, 1216 &bbr_can_force_probertt, 0, 1217 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1218 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1219 SYSCTL_CHILDREN(bbr_probertt), 1220 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1221 &bbr_probertt_sets_rtt, 0, 1222 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1223 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1224 SYSCTL_CHILDREN(bbr_probertt), 1225 OID_AUTO, "can_adjust", CTLFLAG_RW, 1226 &bbr_can_adjust_probertt, 1, 1227 "Can we dynamically adjust the probe-rtt limits and times?"); 1228 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1229 SYSCTL_CHILDREN(bbr_probertt), 1230 OID_AUTO, "is_ratio", CTLFLAG_RW, 1231 &bbr_is_ratio, 0, 1232 "is the limit to filter a ratio?"); 1233 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1234 SYSCTL_CHILDREN(bbr_probertt), 1235 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1236 &bbr_prtt_slam_cwnd, 0, 1237 "Should we set/recover cwnd?"); 1238 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1239 SYSCTL_CHILDREN(bbr_probertt), 1240 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1241 &bbr_can_use_ts_for_rtt, 1, 1242 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1243 1244 /* Pacing controls */ 1245 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1246 SYSCTL_CHILDREN(bbr_sysctl_root), 1247 OID_AUTO, 1248 "pacing", 1249 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1250 ""); 1251 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1252 SYSCTL_CHILDREN(bbr_hptsi), 1253 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1254 &bbr_allow_hdwr_pacing, 1, 1255 "Do we allow hardware pacing?"); 1256 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1257 SYSCTL_CHILDREN(bbr_hptsi), 1258 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1259 &bbr_hardware_pacing_limit, 4000, 1260 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1261 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1262 SYSCTL_CHILDREN(bbr_hptsi), 1263 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1264 &bbr_hdwr_pace_adjust, 2, 1265 "Multiplier to calculated tso size?"); 1266 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1267 SYSCTL_CHILDREN(bbr_hptsi), 1268 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1269 &bbr_hdwr_pace_floor, 1, 1270 "Do we invoke the hardware pacing floor?"); 1271 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1272 SYSCTL_CHILDREN(bbr_hptsi), 1273 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1274 &bbr_hdwr_pacing_delay_cnt, 10, 1275 "How many packets must be sent after hdwr pacing is enabled"); 1276 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1277 SYSCTL_CHILDREN(bbr_hptsi), 1278 OID_AUTO, "bw_cross", CTLFLAG_RW, 1279 &bbr_cross_over, 3000000, 1280 "What is the point where we cross over to linux like TSO size set"); 1281 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1282 SYSCTL_CHILDREN(bbr_hptsi), 1283 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1284 &bbr_hptsi_segments_delay_tar, 7000, 1285 "What is the worse case delay target for hptsi < 48Mbp connections"); 1286 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1287 SYSCTL_CHILDREN(bbr_hptsi), 1288 OID_AUTO, "enet_oh", CTLFLAG_RW, 1289 &bbr_include_enet_oh, 0, 1290 "Do we include the ethernet overhead in calculating pacing delay?"); 1291 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1292 SYSCTL_CHILDREN(bbr_hptsi), 1293 OID_AUTO, "ip_oh", CTLFLAG_RW, 1294 &bbr_include_ip_oh, 1, 1295 "Do we include the IP overhead in calculating pacing delay?"); 1296 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1297 SYSCTL_CHILDREN(bbr_hptsi), 1298 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1299 &bbr_include_tcp_oh, 0, 1300 "Do we include the TCP overhead in calculating pacing delay?"); 1301 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1302 SYSCTL_CHILDREN(bbr_hptsi), 1303 OID_AUTO, "google_discount", CTLFLAG_RW, 1304 &bbr_google_discount, 10, 1305 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1306 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1307 SYSCTL_CHILDREN(bbr_hptsi), 1308 OID_AUTO, "all_get_min", CTLFLAG_RW, 1309 &bbr_all_get_min, 0, 1310 "If you are less than a MSS do you just get the min?"); 1311 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1312 SYSCTL_CHILDREN(bbr_hptsi), 1313 OID_AUTO, "tso_min", CTLFLAG_RW, 1314 &bbr_hptsi_bytes_min, 1460, 1315 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1316 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1317 SYSCTL_CHILDREN(bbr_hptsi), 1318 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1319 &bbr_hptsi_segments_max, 6, 1320 "For 0 -> 24Mbps what is top number of segments for TSO"); 1321 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1322 SYSCTL_CHILDREN(bbr_hptsi), 1323 OID_AUTO, "seg_floor", CTLFLAG_RW, 1324 &bbr_hptsi_segments_floor, 1, 1325 "Minimum TSO size we will fall too in segments"); 1326 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1327 SYSCTL_CHILDREN(bbr_hptsi), 1328 OID_AUTO, "utter_max", CTLFLAG_RW, 1329 &bbr_hptsi_utter_max, 0, 1330 "The absolute maximum that any pacing (outside of hardware) can be"); 1331 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1332 SYSCTL_CHILDREN(bbr_hptsi), 1333 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1334 &bbr_hptsi_per_second, 100, 1335 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1336 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1337 SYSCTL_CHILDREN(bbr_hptsi), 1338 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1339 &bbr_hptsi_max_mul, 1, 1340 "The multiplier for pace len max"); 1341 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1342 SYSCTL_CHILDREN(bbr_hptsi), 1343 OID_AUTO, "srtt_div", CTLFLAG_RW, 1344 &bbr_hptsi_max_div, 2, 1345 "The divisor for pace len max"); 1346 /* Measurement controls */ 1347 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1348 SYSCTL_CHILDREN(bbr_sysctl_root), 1349 OID_AUTO, 1350 "measure", 1351 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1352 "Measurement controls"); 1353 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1354 SYSCTL_CHILDREN(bbr_measure), 1355 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1356 &bbr_initial_bw_bps, 62500, 1357 "Minimum initial b/w in bytes per second"); 1358 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1359 SYSCTL_CHILDREN(bbr_measure), 1360 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1361 &bbr_sack_not_required, 0, 1362 "Do we allow bbr to run on connections not supporting SACK?"); 1363 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1364 SYSCTL_CHILDREN(bbr_measure), 1365 OID_AUTO, "use_google", CTLFLAG_RW, 1366 &bbr_use_google_algo, 0, 1367 "Use has close to google V1.0 has possible?"); 1368 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1369 SYSCTL_CHILDREN(bbr_measure), 1370 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1371 &bbr_ts_limiting, 1, 1372 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1373 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1374 SYSCTL_CHILDREN(bbr_measure), 1375 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1376 &bbr_ts_can_raise, 0, 1377 "Can we raise the b/w via timestamp b/w calculation?"); 1378 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1379 SYSCTL_CHILDREN(bbr_measure), 1380 OID_AUTO, "ts_delta", CTLFLAG_RW, 1381 &bbr_min_usec_delta, 20000, 1382 "How long in usec between ts of our sends in ts validation code?"); 1383 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1384 SYSCTL_CHILDREN(bbr_measure), 1385 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1386 &bbr_min_peer_delta, 20, 1387 "What min numerical value should be between the peer deltas?"); 1388 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1389 SYSCTL_CHILDREN(bbr_measure), 1390 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1391 &bbr_delta_percent, 150, 1392 "What percentage (150 = 15.0) do we allow variance for?"); 1393 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1394 SYSCTL_CHILDREN(bbr_measure), 1395 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1396 &bbr_min_measurements_req, 1, 1397 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1398 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1399 SYSCTL_CHILDREN(bbr_measure), 1400 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1401 &bbr_no_pacing_until, 4, 1402 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1403 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1404 SYSCTL_CHILDREN(bbr_measure), 1405 OID_AUTO, "quanta", CTLFLAG_RW, 1406 &bbr_quanta, 2, 1407 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1408 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1409 SYSCTL_CHILDREN(bbr_measure), 1410 OID_AUTO, "noretran", CTLFLAG_RW, 1411 &bbr_no_retran, 0, 1412 "Should google mode not use retransmission measurements for the b/w estimation?"); 1413 /* State controls */ 1414 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1415 SYSCTL_CHILDREN(bbr_sysctl_root), 1416 OID_AUTO, 1417 "states", 1418 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1419 "State controls"); 1420 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1421 SYSCTL_CHILDREN(bbr_states), 1422 OID_AUTO, "idle_restart", CTLFLAG_RW, 1423 &bbr_uses_idle_restart, 0, 1424 "Do we use a new special idle_restart state to ramp back up quickly?"); 1425 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1426 SYSCTL_CHILDREN(bbr_states), 1427 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1428 &bbr_idle_restart_threshold, 100000, 1429 "How long must we be idle before we restart??"); 1430 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1431 SYSCTL_CHILDREN(bbr_states), 1432 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1433 &bbr_state_is_pkt_epoch, 0, 1434 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1435 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1436 SYSCTL_CHILDREN(bbr_states), 1437 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1438 &bbr_rtt_gain_thresh, 0, 1439 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1440 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1441 SYSCTL_CHILDREN(bbr_states), 1442 OID_AUTO, "drain_floor", CTLFLAG_RW, 1443 &bbr_drain_floor, 88, 1444 "What is the lowest we can drain (pg) too?"); 1445 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1446 SYSCTL_CHILDREN(bbr_states), 1447 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1448 &bbr_state_drain_2_tar, 1, 1449 "Do we drain to target in drain substate?"); 1450 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1451 SYSCTL_CHILDREN(bbr_states), 1452 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1453 &bbr_gain_to_target, 1, 1454 "Does probe bw gain to target??"); 1455 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1456 SYSCTL_CHILDREN(bbr_states), 1457 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1458 &bbr_gain_gets_extra_too, 1, 1459 "Does probe bw gain get the extra time too?"); 1460 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1461 SYSCTL_CHILDREN(bbr_states), 1462 OID_AUTO, "ld_div", CTLFLAG_RW, 1463 &bbr_drain_drop_div, 5, 1464 "Long drain drop divider?"); 1465 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1466 SYSCTL_CHILDREN(bbr_states), 1467 OID_AUTO, "ld_mul", CTLFLAG_RW, 1468 &bbr_drain_drop_mul, 4, 1469 "Long drain drop multiplier?"); 1470 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1471 SYSCTL_CHILDREN(bbr_states), 1472 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1473 &bbr_rand_ot, 50, 1474 "Random discount of the ot?"); 1475 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1476 SYSCTL_CHILDREN(bbr_states), 1477 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1478 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1479 "How many packet-epochs does the b/w delivery rate last?"); 1480 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1481 SYSCTL_CHILDREN(bbr_states), 1482 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1483 &bbr_sub_drain_app_limit, 0, 1484 "Does our sub-state drain invoke app limited if its long?"); 1485 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1486 SYSCTL_CHILDREN(bbr_states), 1487 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1488 &bbr_sub_drain_slam_cwnd, 0, 1489 "Should we set/recover cwnd for sub-state drain?"); 1490 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1491 SYSCTL_CHILDREN(bbr_states), 1492 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1493 &bbr_slam_cwnd_in_main_drain, 0, 1494 "Should we set/recover cwnd for main-state drain?"); 1495 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1496 SYSCTL_CHILDREN(bbr_states), 1497 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1498 &google_allow_early_out, 1, 1499 "Should we allow google probe-bw/drain to exit early at flight target?"); 1500 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1501 SYSCTL_CHILDREN(bbr_states), 1502 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1503 &google_consider_lost, 1, 1504 "Should we have losses exit gain of probebw in google mode??"); 1505 /* Startup controls */ 1506 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1507 SYSCTL_CHILDREN(bbr_sysctl_root), 1508 OID_AUTO, 1509 "startup", 1510 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1511 "Startup controls"); 1512 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1513 SYSCTL_CHILDREN(bbr_startup), 1514 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1515 &bbr_sends_full_iwnd, 1, 1516 "Do we not pace but burst out initial windows has our TSO size?"); 1517 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1518 SYSCTL_CHILDREN(bbr_startup), 1519 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1520 &bbr_startup_loss_thresh, 2000, 1521 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1522 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1523 SYSCTL_CHILDREN(bbr_startup), 1524 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1525 &bbr_use_lower_gain_in_startup, 1, 1526 "Should we use a lower hptsi gain if we see loss in startup?"); 1527 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1528 SYSCTL_CHILDREN(bbr_startup), 1529 OID_AUTO, "gain", CTLFLAG_RW, 1530 &bbr_start_exit, 25, 1531 "What gain percent do we need to see to stay in startup??"); 1532 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1533 SYSCTL_CHILDREN(bbr_startup), 1534 OID_AUTO, "low_gain", CTLFLAG_RW, 1535 &bbr_low_start_exit, 15, 1536 "What gain percent do we need to see to stay in the lower gain startup??"); 1537 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1538 SYSCTL_CHILDREN(bbr_startup), 1539 OID_AUTO, "loss_exit", CTLFLAG_RW, 1540 &bbr_exit_startup_at_loss, 1, 1541 "Should we exit startup at loss in an epoch if we are not gaining?"); 1542 /* CWND controls */ 1543 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1544 SYSCTL_CHILDREN(bbr_sysctl_root), 1545 OID_AUTO, 1546 "cwnd", 1547 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1548 "Cwnd controls"); 1549 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1550 SYSCTL_CHILDREN(bbr_cwnd), 1551 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1552 &bbr_cwndtarget_rtt_touse, 0, 1553 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1554 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1555 SYSCTL_CHILDREN(bbr_cwnd), 1556 OID_AUTO, "may_shrink", CTLFLAG_RW, 1557 &bbr_cwnd_may_shrink, 0, 1558 "Can the cwnd shrink if it would grow to more than the target?"); 1559 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1560 SYSCTL_CHILDREN(bbr_cwnd), 1561 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1562 &bbr_target_cwnd_mult_limit, 8, 1563 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1564 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1565 SYSCTL_CHILDREN(bbr_cwnd), 1566 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1567 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1568 "What is the high-speed min cwnd (rttProp under 1ms)"); 1569 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1570 SYSCTL_CHILDREN(bbr_cwnd), 1571 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1572 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1573 "What is the min cwnd (rttProp > 1ms)"); 1574 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1575 SYSCTL_CHILDREN(bbr_cwnd), 1576 OID_AUTO, "initwin", CTLFLAG_RW, 1577 &bbr_def_init_win, 10, 1578 "What is the BBR initial window, if 0 use tcp version"); 1579 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1580 SYSCTL_CHILDREN(bbr_cwnd), 1581 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1582 &bbr_do_red, 600, 1583 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1584 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1585 SYSCTL_CHILDREN(bbr_cwnd), 1586 OID_AUTO, "red_scale", CTLFLAG_RW, 1587 &bbr_red_scale, 20000, 1588 "What RTT do we scale with?"); 1589 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1590 SYSCTL_CHILDREN(bbr_cwnd), 1591 OID_AUTO, "red_growslow", CTLFLAG_RW, 1592 &bbr_red_growth_restrict, 1, 1593 "Do we restrict cwnd growth for whats in flight?"); 1594 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1595 SYSCTL_CHILDREN(bbr_cwnd), 1596 OID_AUTO, "red_div", CTLFLAG_RW, 1597 &bbr_red_div, 2, 1598 "If we reduce whats the divisor?"); 1599 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1600 SYSCTL_CHILDREN(bbr_cwnd), 1601 OID_AUTO, "red_mul", CTLFLAG_RW, 1602 &bbr_red_mul, 1, 1603 "If we reduce whats the mulitiplier?"); 1604 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1605 SYSCTL_CHILDREN(bbr_cwnd), 1606 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1607 &bbr_target_is_bbunit, 0, 1608 "Is the state target the pacing_gain or BBR_UNIT?"); 1609 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1610 SYSCTL_CHILDREN(bbr_cwnd), 1611 OID_AUTO, "drop_limit", CTLFLAG_RW, 1612 &bbr_drop_limit, 0, 1613 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1614 1615 /* Timeout controls */ 1616 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1617 SYSCTL_CHILDREN(bbr_sysctl_root), 1618 OID_AUTO, 1619 "timeout", 1620 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1621 "Time out controls"); 1622 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1623 SYSCTL_CHILDREN(bbr_timeout), 1624 OID_AUTO, "delack", CTLFLAG_RW, 1625 &bbr_delack_time, 100000, 1626 "BBR's delayed ack time"); 1627 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1628 SYSCTL_CHILDREN(bbr_timeout), 1629 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1630 &bbr_tlp_type_to_use, 3, 1631 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1632 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1633 SYSCTL_CHILDREN(bbr_timeout), 1634 OID_AUTO, "persmin", CTLFLAG_RW, 1635 &bbr_persist_min, 250000, 1636 "What is the minimum time in microseconds between persists"); 1637 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1638 SYSCTL_CHILDREN(bbr_timeout), 1639 OID_AUTO, "persmax", CTLFLAG_RW, 1640 &bbr_persist_max, 1000000, 1641 "What is the largest delay in microseconds between persists"); 1642 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1643 SYSCTL_CHILDREN(bbr_timeout), 1644 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1645 &bbr_tlp_min, 10000, 1646 "TLP Min timeout in usecs"); 1647 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1648 SYSCTL_CHILDREN(bbr_timeout), 1649 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1650 &bbr_delayed_ack_time, 200000, 1651 "TLP delayed ack compensation value"); 1652 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1653 SYSCTL_CHILDREN(bbr_sysctl_root), 1654 OID_AUTO, "minrto", CTLFLAG_RW, 1655 &bbr_rto_min_ms, 30, 1656 "Minimum RTO in ms"); 1657 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1658 SYSCTL_CHILDREN(bbr_timeout), 1659 OID_AUTO, "maxrto", CTLFLAG_RW, 1660 &bbr_rto_max_sec, 4, 1661 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1662 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1663 SYSCTL_CHILDREN(bbr_timeout), 1664 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1665 &bbr_tlp_max_resend, 2, 1666 "How many times does TLP retry a single segment or multiple with no ACK"); 1667 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1668 SYSCTL_CHILDREN(bbr_timeout), 1669 OID_AUTO, "minto", CTLFLAG_RW, 1670 &bbr_min_to, 1000, 1671 "Minimum rack timeout in useconds"); 1672 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1673 SYSCTL_CHILDREN(bbr_timeout), 1674 OID_AUTO, "pktdelay", CTLFLAG_RW, 1675 &bbr_pkt_delay, 1000, 1676 "Extra RACK time (in useconds) besides reordering thresh"); 1677 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1678 SYSCTL_CHILDREN(bbr_timeout), 1679 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1680 &bbr_incr_timers, 1, 1681 "Increase the RXT/TLP timer by the pacing time used?"); 1682 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1683 SYSCTL_CHILDREN(bbr_timeout), 1684 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1685 &bbr_marks_rxt_sack_passed, 0, 1686 "Mark sack passed on all those not ack'd when a RXT hits?"); 1687 /* Policer controls */ 1688 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1689 SYSCTL_CHILDREN(bbr_sysctl_root), 1690 OID_AUTO, 1691 "policer", 1692 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1693 "Policer controls"); 1694 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1695 SYSCTL_CHILDREN(bbr_policer), 1696 OID_AUTO, "detect_enable", CTLFLAG_RW, 1697 &bbr_policer_detection_enabled, 1, 1698 "Is policer detection enabled??"); 1699 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1700 SYSCTL_CHILDREN(bbr_policer), 1701 OID_AUTO, "min_pes", CTLFLAG_RW, 1702 &bbr_lt_intvl_min_rtts, 4, 1703 "Minimum number of PE's?"); 1704 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1705 SYSCTL_CHILDREN(bbr_policer), 1706 OID_AUTO, "bwdiff", CTLFLAG_RW, 1707 &bbr_lt_bw_diff, (4000/8), 1708 "Minimal bw diff?"); 1709 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1710 SYSCTL_CHILDREN(bbr_policer), 1711 OID_AUTO, "bwratio", CTLFLAG_RW, 1712 &bbr_lt_bw_ratio, 8, 1713 "Minimal bw diff?"); 1714 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1715 SYSCTL_CHILDREN(bbr_policer), 1716 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1717 &bbr_policer_call_from_rack_to, 0, 1718 "Do we call the policer detection code from a rack-timeout?"); 1719 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1720 SYSCTL_CHILDREN(bbr_policer), 1721 OID_AUTO, "false_postive", CTLFLAG_RW, 1722 &bbr_lt_intvl_fp, 0, 1723 "What packet epoch do we do false-positive detection at (0=no)?"); 1724 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1725 SYSCTL_CHILDREN(bbr_policer), 1726 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1727 &bbr_lt_loss_thresh, 196, 1728 "Loss threshold 196 = 19.6%?"); 1729 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1730 SYSCTL_CHILDREN(bbr_policer), 1731 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1732 &bbr_lt_fd_thresh, 100, 1733 "What percentage is the false detection threshold (150=15.0)?"); 1734 /* All the rest */ 1735 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1736 SYSCTL_CHILDREN(bbr_sysctl_root), 1737 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1738 &bbr_use_rack_resend_cheat, 0, 1739 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1740 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1741 SYSCTL_CHILDREN(bbr_sysctl_root), 1742 OID_AUTO, "error_paceout", CTLFLAG_RW, 1743 &bbr_error_base_paceout, 10000, 1744 "When we hit an error what is the min to pace out in usec's?"); 1745 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1746 SYSCTL_CHILDREN(bbr_sysctl_root), 1747 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1748 &bbr_max_net_error_cnt, 10, 1749 "When we hit this many errors in a row, kill the session?"); 1750 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1751 SYSCTL_CHILDREN(bbr_sysctl_root), 1752 OID_AUTO, "data_after_close", CTLFLAG_RW, 1753 &bbr_ignore_data_after_close, 1, 1754 "Do we hold off sending a RST until all pending data is ack'd"); 1755 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1756 SYSCTL_CHILDREN(bbr_sysctl_root), 1757 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1758 &bbr_resends_use_tso, 0, 1759 "Can resends use TSO?"); 1760 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1761 SYSCTL_CHILDREN(bbr_sysctl_root), 1762 OID_AUTO, "sblklimit", CTLFLAG_RW, 1763 &bbr_sack_block_limit, 128, 1764 "When do we start ignoring small sack blocks"); 1765 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1766 SYSCTL_CHILDREN(bbr_sysctl_root), 1767 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1768 &bbr_verbose_logging, 0, 1769 "Should BBR black box logging be verbose"); 1770 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1771 SYSCTL_CHILDREN(bbr_sysctl_root), 1772 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1773 &bbr_reorder_thresh, 2, 1774 "What factor for rack will be added when seeing reordering (shift right)"); 1775 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1776 SYSCTL_CHILDREN(bbr_sysctl_root), 1777 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1778 &bbr_reorder_fade, 0, 1779 "Does reorder detection fade, if so how many ms (0 means never)"); 1780 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1781 SYSCTL_CHILDREN(bbr_sysctl_root), 1782 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1783 &bbr_tlp_thresh, 1, 1784 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1785 /* Stats and counters */ 1786 /* The pacing counters for hdwr/software can't be in the array */ 1787 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1788 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1789 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1790 SYSCTL_CHILDREN(bbr_sysctl_root), 1791 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1792 &bbr_hdwr_pacing_enobuf, 1793 "Total number of enobufs for hardware paced flows"); 1794 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1795 SYSCTL_CHILDREN(bbr_sysctl_root), 1796 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1797 &bbr_nohdwr_pacing_enobuf, 1798 "Total number of enobufs for non-hardware paced flows"); 1799 1800 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1801 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1802 SYSCTL_CHILDREN(bbr_sysctl_root), 1803 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1804 &bbr_flows_whdwr_pacing, 1805 "Total number of hardware paced flows"); 1806 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1807 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1808 SYSCTL_CHILDREN(bbr_sysctl_root), 1809 OID_AUTO, "software_pacing", CTLFLAG_RD, 1810 &bbr_flows_nohdwr_pacing, 1811 "Total number of software paced flows"); 1812 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1813 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1814 OID_AUTO, "stats", CTLFLAG_RD, 1815 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1816 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1817 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1818 OID_AUTO, "opts", CTLFLAG_RD, 1819 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1820 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1821 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1822 OID_AUTO, "lost", CTLFLAG_RD, 1823 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1824 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1825 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1826 OID_AUTO, "stateresend", CTLFLAG_RD, 1827 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1828 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1829 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1830 OID_AUTO, "statetime", CTLFLAG_RD, 1831 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1832 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1833 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1834 OID_AUTO, "outsize", CTLFLAG_RD, 1835 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1836 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1837 SYSCTL_CHILDREN(bbr_sysctl_root), 1838 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1839 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1840 } 1841 1842 static void 1843 bbr_counter_destroy(void) 1844 { 1845 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1846 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1847 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1848 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1849 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1850 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1851 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1852 counter_u64_free(bbr_hdwr_pacing_enobuf); 1853 counter_u64_free(bbr_flows_whdwr_pacing); 1854 counter_u64_free(bbr_flows_nohdwr_pacing); 1855 1856 } 1857 1858 static __inline void 1859 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1860 { 1861 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1862 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1863 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1864 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1865 l->bw_inuse = bbr_get_bw(bbr); 1866 l->inflight = ctf_flight_size(bbr->rc_tp, 1867 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1868 l->applimited = bbr->r_ctl.r_app_limited_until; 1869 l->delivered = bbr->r_ctl.rc_delivered; 1870 l->timeStamp = cts; 1871 l->lost = bbr->r_ctl.rc_lost; 1872 l->bbr_state = bbr->rc_bbr_state; 1873 l->bbr_substate = bbr_state_val(bbr); 1874 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1875 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1876 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1877 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1878 l->inhpts = tcp_in_hpts(bbr->rc_tp); 1879 l->use_lt_bw = bbr->rc_lt_use_bw; 1880 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1881 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1882 } 1883 1884 static void 1885 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1886 { 1887 if (tcp_bblogging_on(bbr->rc_tp)) { 1888 union tcp_log_stackspecific log; 1889 1890 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1891 log.u_bbr.flex1 = 0; 1892 log.u_bbr.flex2 = 0; 1893 log.u_bbr.flex5 = 0; 1894 log.u_bbr.flex3 = 0; 1895 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1896 log.u_bbr.flex7 = reason; 1897 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1898 log.u_bbr.flex8 = 0; 1899 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1900 &bbr->rc_inp->inp_socket->so_rcv, 1901 &bbr->rc_inp->inp_socket->so_snd, 1902 BBR_LOG_BW_RED_EV, 0, 1903 0, &log, false, &bbr->rc_tv); 1904 } 1905 } 1906 1907 static void 1908 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1909 { 1910 if (tcp_bblogging_on(bbr->rc_tp)) { 1911 union tcp_log_stackspecific log; 1912 1913 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1914 log.u_bbr.flex1 = seq; 1915 log.u_bbr.flex2 = count; 1916 log.u_bbr.flex8 = mode; 1917 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1918 &bbr->rc_inp->inp_socket->so_rcv, 1919 &bbr->rc_inp->inp_socket->so_snd, 1920 BBR_LOG_LOWGAIN, 0, 1921 0, &log, false, &bbr->rc_tv); 1922 } 1923 } 1924 1925 static void 1926 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1927 uint8_t reason, uint32_t p_maxseg, int len) 1928 { 1929 if (tcp_bblogging_on(bbr->rc_tp)) { 1930 union tcp_log_stackspecific log; 1931 1932 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1933 log.u_bbr.flex1 = p_maxseg; 1934 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1935 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1936 log.u_bbr.flex4 = reason; 1937 log.u_bbr.flex5 = bbr->rc_in_persist; 1938 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1939 log.u_bbr.flex7 = p_maxseg; 1940 log.u_bbr.flex8 = bbr->rc_in_persist; 1941 log.u_bbr.pkts_out = 0; 1942 log.u_bbr.applimited = len; 1943 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1944 &bbr->rc_inp->inp_socket->so_rcv, 1945 &bbr->rc_inp->inp_socket->so_snd, 1946 BBR_LOG_JUSTRET, 0, 1947 tlen, &log, false, &bbr->rc_tv); 1948 } 1949 } 1950 1951 static void 1952 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1953 { 1954 if (tcp_bblogging_on(bbr->rc_tp)) { 1955 union tcp_log_stackspecific log; 1956 1957 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1958 log.u_bbr.flex1 = seq; 1959 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1960 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1961 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1962 &bbr->rc_inp->inp_socket->so_rcv, 1963 &bbr->rc_inp->inp_socket->so_snd, 1964 BBR_LOG_ENTREC, 0, 1965 0, &log, false, &bbr->rc_tv); 1966 } 1967 } 1968 1969 static void 1970 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) 1971 { 1972 if (tcp_bblogging_on(tp)) { 1973 union tcp_log_stackspecific log; 1974 1975 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1976 log.u_bbr.flex1 = tso; 1977 log.u_bbr.flex2 = maxseg; 1978 log.u_bbr.flex3 = mtu; 1979 log.u_bbr.flex4 = csum_flags; 1980 TCP_LOG_EVENTP(tp, NULL, 1981 &bbr->rc_inp->inp_socket->so_rcv, 1982 &bbr->rc_inp->inp_socket->so_snd, 1983 BBR_LOG_MSGSIZE, 0, 1984 0, &log, false, &bbr->rc_tv); 1985 } 1986 } 1987 1988 static void 1989 bbr_log_flowend(struct tcp_bbr *bbr) 1990 { 1991 if (tcp_bblogging_on(bbr->rc_tp)) { 1992 union tcp_log_stackspecific log; 1993 struct sockbuf *r, *s; 1994 struct timeval tv; 1995 1996 if (bbr->rc_inp->inp_socket) { 1997 r = &bbr->rc_inp->inp_socket->so_rcv; 1998 s = &bbr->rc_inp->inp_socket->so_snd; 1999 } else { 2000 r = s = NULL; 2001 } 2002 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2003 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2004 r, s, 2005 TCP_LOG_FLOWEND, 0, 2006 0, &log, false, &tv); 2007 } 2008 } 2009 2010 static void 2011 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2012 uint32_t lost, uint32_t del) 2013 { 2014 if (tcp_bblogging_on(bbr->rc_tp)) { 2015 union tcp_log_stackspecific log; 2016 2017 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2018 log.u_bbr.flex1 = lost; 2019 log.u_bbr.flex2 = del; 2020 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2021 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2022 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2023 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2024 log.u_bbr.flex7 = line; 2025 log.u_bbr.flex8 = 0; 2026 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2027 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2028 &bbr->rc_inp->inp_socket->so_rcv, 2029 &bbr->rc_inp->inp_socket->so_snd, 2030 BBR_LOG_PKT_EPOCH, 0, 2031 0, &log, false, &bbr->rc_tv); 2032 } 2033 } 2034 2035 static void 2036 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2037 { 2038 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2039 union tcp_log_stackspecific log; 2040 2041 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2042 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2043 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2044 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2045 log.u_bbr.flex7 = line; 2046 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2047 &bbr->rc_inp->inp_socket->so_rcv, 2048 &bbr->rc_inp->inp_socket->so_snd, 2049 BBR_LOG_TIME_EPOCH, 0, 2050 0, &log, false, &bbr->rc_tv); 2051 } 2052 } 2053 2054 static void 2055 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2056 { 2057 if (tcp_bblogging_on(bbr->rc_tp)) { 2058 union tcp_log_stackspecific log; 2059 2060 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2061 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2062 log.u_bbr.flex2 = new_tar; 2063 log.u_bbr.flex3 = line; 2064 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2065 log.u_bbr.flex5 = bbr_quanta; 2066 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2067 log.u_bbr.flex7 = bbr->rc_last_options; 2068 log.u_bbr.flex8 = meth; 2069 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2070 &bbr->rc_inp->inp_socket->so_rcv, 2071 &bbr->rc_inp->inp_socket->so_snd, 2072 BBR_LOG_STATE_TARGET, 0, 2073 0, &log, false, &bbr->rc_tv); 2074 } 2075 2076 } 2077 2078 static void 2079 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2080 { 2081 if (tcp_bblogging_on(bbr->rc_tp)) { 2082 union tcp_log_stackspecific log; 2083 2084 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2085 log.u_bbr.flex1 = line; 2086 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2087 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2088 if (bbr_state_is_pkt_epoch) 2089 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2090 else 2091 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2092 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2093 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2094 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2095 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2096 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2097 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2098 &bbr->rc_inp->inp_socket->so_rcv, 2099 &bbr->rc_inp->inp_socket->so_snd, 2100 BBR_LOG_STATE, 0, 2101 0, &log, false, &bbr->rc_tv); 2102 } 2103 } 2104 2105 static void 2106 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2107 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2108 { 2109 if (tcp_bblogging_on(bbr->rc_tp)) { 2110 union tcp_log_stackspecific log; 2111 2112 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2113 log.u_bbr.flex1 = line; 2114 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2115 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2116 log.u_bbr.flex4 = applied; 2117 log.u_bbr.flex5 = rtt; 2118 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2119 log.u_bbr.flex7 = cond; 2120 log.u_bbr.flex8 = reas; 2121 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2122 &bbr->rc_inp->inp_socket->so_rcv, 2123 &bbr->rc_inp->inp_socket->so_snd, 2124 BBR_LOG_RTT_SHRINKS, 0, 2125 0, &log, false, &bbr->rc_tv); 2126 } 2127 } 2128 2129 static void 2130 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2131 { 2132 if (tcp_bblogging_on(bbr->rc_tp)) { 2133 union tcp_log_stackspecific log; 2134 2135 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2136 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2137 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2138 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2139 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2140 &bbr->rc_inp->inp_socket->so_rcv, 2141 &bbr->rc_inp->inp_socket->so_snd, 2142 BBR_LOG_EXITREC, 0, 2143 0, &log, false, &bbr->rc_tv); 2144 } 2145 } 2146 2147 static void 2148 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2149 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2150 { 2151 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2152 union tcp_log_stackspecific log; 2153 2154 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2155 log.u_bbr.flex1 = line; 2156 log.u_bbr.flex2 = prev_acked; 2157 log.u_bbr.flex3 = bytes_this_ack; 2158 log.u_bbr.flex4 = chg; 2159 log.u_bbr.flex5 = th_ack; 2160 log.u_bbr.flex6 = target; 2161 log.u_bbr.flex8 = meth; 2162 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2163 &bbr->rc_inp->inp_socket->so_rcv, 2164 &bbr->rc_inp->inp_socket->so_snd, 2165 BBR_LOG_CWND, 0, 2166 0, &log, false, &bbr->rc_tv); 2167 } 2168 } 2169 2170 static void 2171 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2172 { 2173 /* 2174 * Log the rtt sample we are applying to the srtt algorithm in 2175 * useconds. 2176 */ 2177 if (tcp_bblogging_on(bbr->rc_tp)) { 2178 union tcp_log_stackspecific log; 2179 2180 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2181 log.u_bbr.flex1 = rtt; 2182 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2183 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2184 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2185 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2186 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2187 log.u_bbr.flex6 = tsin; 2188 log.u_bbr.flex7 = 0; 2189 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2190 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2191 &bbr->rc_inp->inp_socket->so_rcv, 2192 &bbr->rc_inp->inp_socket->so_snd, 2193 TCP_LOG_RTT, 0, 2194 0, &log, false, &bbr->rc_tv); 2195 } 2196 } 2197 2198 static void 2199 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2200 { 2201 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2202 union tcp_log_stackspecific log; 2203 2204 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2205 log.u_bbr.flex1 = time_in; 2206 log.u_bbr.flex2 = line; 2207 log.u_bbr.flex8 = enter_exit; 2208 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2209 &bbr->rc_inp->inp_socket->so_rcv, 2210 &bbr->rc_inp->inp_socket->so_snd, 2211 BBR_LOG_PERSIST, 0, 2212 0, &log, false, &bbr->rc_tv); 2213 } 2214 } 2215 static void 2216 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2217 { 2218 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2219 union tcp_log_stackspecific log; 2220 2221 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2222 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2223 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2224 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2225 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2226 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2227 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2228 &bbr->rc_inp->inp_socket->so_rcv, 2229 &bbr->rc_inp->inp_socket->so_snd, 2230 BBR_LOG_ACKCLEAR, 0, 2231 0, &log, false, &bbr->rc_tv); 2232 } 2233 } 2234 2235 static void 2236 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2237 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2238 { 2239 if (tcp_bblogging_on(bbr->rc_tp)) { 2240 union tcp_log_stackspecific log; 2241 struct timeval tv; 2242 2243 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2244 log.u_bbr.flex1 = nsegs; 2245 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2246 if (m) { 2247 struct timespec ts; 2248 2249 log.u_bbr.flex3 = m->m_flags; 2250 if (m->m_flags & M_TSTMP) { 2251 mbuf_tstmp2timespec(m, &ts); 2252 tv.tv_sec = ts.tv_sec; 2253 tv.tv_usec = ts.tv_nsec / 1000; 2254 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2255 } else { 2256 log.u_bbr.lt_epoch = 0; 2257 } 2258 if (m->m_flags & M_TSTMP_LRO) { 2259 mbuf_tstmp2timeval(m, &tv); 2260 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2261 } else { 2262 /* No arrival timestamp */ 2263 log.u_bbr.flex5 = 0; 2264 } 2265 2266 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2267 } else { 2268 log.u_bbr.flex3 = 0; 2269 log.u_bbr.flex5 = 0; 2270 log.u_bbr.flex6 = 0; 2271 log.u_bbr.pkts_out = 0; 2272 } 2273 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2274 log.u_bbr.flex7 = bbr->r_wanted_output; 2275 log.u_bbr.flex8 = bbr->rc_in_persist; 2276 TCP_LOG_EVENTP(bbr->rc_tp, th, 2277 &bbr->rc_inp->inp_socket->so_rcv, 2278 &bbr->rc_inp->inp_socket->so_snd, 2279 TCP_LOG_IN, 0, 2280 tlen, &log, true, &bbr->rc_tv); 2281 } 2282 } 2283 2284 static void 2285 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2286 { 2287 if (tcp_bblogging_on(bbr->rc_tp)) { 2288 union tcp_log_stackspecific log; 2289 2290 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2291 log.u_bbr.flex1 = did_out; 2292 log.u_bbr.flex2 = nxt_pkt; 2293 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2294 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2295 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2296 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2297 log.u_bbr.flex7 = bbr->r_wanted_output; 2298 log.u_bbr.flex8 = bbr->rc_in_persist; 2299 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2300 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2301 &bbr->rc_inp->inp_socket->so_rcv, 2302 &bbr->rc_inp->inp_socket->so_snd, 2303 BBR_LOG_DOSEG_DONE, 0, 2304 0, &log, true, &bbr->rc_tv); 2305 } 2306 } 2307 2308 static void 2309 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2310 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2311 { 2312 if (tcp_bblogging_on(bbr->rc_tp)) { 2313 union tcp_log_stackspecific log; 2314 2315 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2316 log.u_bbr.flex1 = line; 2317 log.u_bbr.flex2 = o_len; 2318 log.u_bbr.flex3 = segcnt; 2319 log.u_bbr.flex4 = segsiz; 2320 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2321 &bbr->rc_inp->inp_socket->so_rcv, 2322 &bbr->rc_inp->inp_socket->so_snd, 2323 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2324 len, &log, true, &bbr->rc_tv); 2325 } 2326 } 2327 2328 static void 2329 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2330 { 2331 if (tcp_bblogging_on(bbr->rc_tp)) { 2332 union tcp_log_stackspecific log; 2333 2334 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2335 log.u_bbr.flex1 = timers; 2336 log.u_bbr.flex2 = ret; 2337 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2338 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2339 log.u_bbr.flex5 = cts; 2340 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2341 log.u_bbr.flex8 = hpts_calling; 2342 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2343 &bbr->rc_inp->inp_socket->so_rcv, 2344 &bbr->rc_inp->inp_socket->so_snd, 2345 BBR_LOG_TO_PROCESS, 0, 2346 0, &log, false, &bbr->rc_tv); 2347 } 2348 } 2349 2350 static void 2351 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2352 { 2353 if (tcp_bblogging_on(bbr->rc_tp)) { 2354 union tcp_log_stackspecific log; 2355 uint64_t ar; 2356 2357 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2358 log.u_bbr.flex1 = bbr->bbr_timer_src; 2359 log.u_bbr.flex2 = 0; 2360 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2361 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2362 ar >>= 32; 2363 ar &= 0x00000000ffffffff; 2364 log.u_bbr.flex4 = (uint32_t)ar; 2365 ar = (uint64_t)bbr->r_ctl.rc_resend; 2366 ar &= 0x00000000ffffffff; 2367 log.u_bbr.flex5 = (uint32_t)ar; 2368 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2369 log.u_bbr.flex8 = to_num; 2370 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2371 &bbr->rc_inp->inp_socket->so_rcv, 2372 &bbr->rc_inp->inp_socket->so_snd, 2373 BBR_LOG_RTO, 0, 2374 0, &log, false, &bbr->rc_tv); 2375 } 2376 } 2377 2378 static void 2379 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2380 { 2381 if (tcp_bblogging_on(bbr->rc_tp)) { 2382 union tcp_log_stackspecific log; 2383 2384 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2385 log.u_bbr.flex1 = flex1; 2386 log.u_bbr.flex2 = flex2; 2387 log.u_bbr.flex3 = flex3; 2388 log.u_bbr.flex4 = 0; 2389 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2390 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2391 log.u_bbr.flex8 = reason; 2392 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2393 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2394 &bbr->rc_inp->inp_socket->so_rcv, 2395 &bbr->rc_inp->inp_socket->so_snd, 2396 BBR_LOG_REDUCE, 0, 2397 0, &log, false, &bbr->rc_tv); 2398 } 2399 } 2400 2401 static void 2402 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2403 { 2404 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2405 union tcp_log_stackspecific log; 2406 2407 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2408 log.u_bbr.flex1 = diag->p_nxt_slot; 2409 log.u_bbr.flex2 = diag->p_cur_slot; 2410 log.u_bbr.flex3 = diag->slot_req; 2411 log.u_bbr.flex4 = diag->inp_hptsslot; 2412 log.u_bbr.flex5 = diag->slot_remaining; 2413 log.u_bbr.flex6 = diag->need_new_to; 2414 log.u_bbr.flex7 = diag->p_hpts_active; 2415 log.u_bbr.flex8 = diag->p_on_min_sleep; 2416 /* Hijack other fields as needed */ 2417 log.u_bbr.epoch = diag->have_slept; 2418 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2419 log.u_bbr.pkts_out = diag->co_ret; 2420 log.u_bbr.applimited = diag->hpts_sleep_time; 2421 log.u_bbr.delivered = diag->p_prev_slot; 2422 log.u_bbr.inflight = diag->p_runningslot; 2423 log.u_bbr.bw_inuse = diag->wheel_slot; 2424 log.u_bbr.rttProp = diag->wheel_cts; 2425 log.u_bbr.delRate = diag->maxslots; 2426 log.u_bbr.cur_del_rate = diag->p_curtick; 2427 log.u_bbr.cur_del_rate <<= 32; 2428 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2429 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2430 &bbr->rc_inp->inp_socket->so_rcv, 2431 &bbr->rc_inp->inp_socket->so_snd, 2432 BBR_LOG_HPTSDIAG, 0, 2433 0, &log, false, &bbr->rc_tv); 2434 } 2435 } 2436 2437 static void 2438 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2439 uint32_t thresh, uint32_t to) 2440 { 2441 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2442 union tcp_log_stackspecific log; 2443 2444 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2445 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2446 log.u_bbr.flex2 = time_since_sent; 2447 log.u_bbr.flex3 = srtt; 2448 log.u_bbr.flex4 = thresh; 2449 log.u_bbr.flex5 = to; 2450 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2451 log.u_bbr.flex8 = mode; 2452 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2453 &bbr->rc_inp->inp_socket->so_rcv, 2454 &bbr->rc_inp->inp_socket->so_snd, 2455 BBR_LOG_TIMERPREP, 0, 2456 0, &log, false, &bbr->rc_tv); 2457 } 2458 } 2459 2460 static void 2461 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2462 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2463 { 2464 if (tcp_bblogging_on(bbr->rc_tp)) { 2465 union tcp_log_stackspecific log; 2466 2467 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2468 log.u_bbr.flex1 = usecs; 2469 log.u_bbr.flex2 = len; 2470 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2471 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2472 if (override) 2473 log.u_bbr.flex5 = (1 << 2); 2474 else 2475 log.u_bbr.flex5 = 0; 2476 log.u_bbr.flex6 = override; 2477 log.u_bbr.flex7 = gain; 2478 log.u_bbr.flex8 = mod; 2479 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2480 &bbr->rc_inp->inp_socket->so_rcv, 2481 &bbr->rc_inp->inp_socket->so_snd, 2482 BBR_LOG_HPTSI_CALC, 0, 2483 len, &log, false, &bbr->rc_tv); 2484 } 2485 } 2486 2487 static void 2488 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2489 { 2490 if (tcp_bblogging_on(bbr->rc_tp)) { 2491 union tcp_log_stackspecific log; 2492 2493 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2494 2495 log.u_bbr.flex1 = bbr->bbr_timer_src; 2496 log.u_bbr.flex2 = to; 2497 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2498 log.u_bbr.flex4 = slot; 2499 log.u_bbr.flex5 = bbr->rc_tp->t_hpts_slot; 2500 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2501 log.u_bbr.pkts_out = bbr->rc_tp->t_flags2; 2502 log.u_bbr.flex8 = which; 2503 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2504 &bbr->rc_inp->inp_socket->so_rcv, 2505 &bbr->rc_inp->inp_socket->so_snd, 2506 BBR_LOG_TIMERSTAR, 0, 2507 0, &log, false, &bbr->rc_tv); 2508 } 2509 } 2510 2511 static void 2512 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) 2513 { 2514 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2515 union tcp_log_stackspecific log; 2516 2517 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2518 log.u_bbr.flex1 = thresh; 2519 log.u_bbr.flex2 = lro; 2520 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2521 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2522 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2523 log.u_bbr.flex6 = srtt; 2524 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2525 log.u_bbr.flex8 = frm; 2526 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2527 &bbr->rc_inp->inp_socket->so_rcv, 2528 &bbr->rc_inp->inp_socket->so_snd, 2529 BBR_LOG_THRESH_CALC, 0, 2530 0, &log, false, &bbr->rc_tv); 2531 } 2532 } 2533 2534 static void 2535 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2536 { 2537 if (tcp_bblogging_on(bbr->rc_tp)) { 2538 union tcp_log_stackspecific log; 2539 2540 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2541 log.u_bbr.flex1 = line; 2542 log.u_bbr.flex2 = bbr->bbr_timer_src; 2543 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2544 log.u_bbr.flex4 = bbr->rc_in_persist; 2545 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2546 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2547 log.u_bbr.flex8 = hpts_removed; 2548 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2549 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2550 &bbr->rc_inp->inp_socket->so_rcv, 2551 &bbr->rc_inp->inp_socket->so_snd, 2552 BBR_LOG_TIMERCANC, 0, 2553 0, &log, false, &bbr->rc_tv); 2554 } 2555 } 2556 2557 static void 2558 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2559 { 2560 if (tcp_bblogging_on(bbr->rc_tp)) { 2561 union tcp_log_stackspecific log; 2562 2563 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2564 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2565 log.u_bbr.flex2 = (peer_delta >> 32); 2566 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2567 log.u_bbr.flex4 = (delta >> 32); 2568 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2569 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2570 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2571 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2572 &bbr->rc_inp->inp_socket->so_rcv, 2573 &bbr->rc_inp->inp_socket->so_snd, 2574 BBR_LOG_TSTMP_VAL, 0, 2575 0, &log, false, &bbr->rc_tv); 2576 } 2577 } 2578 2579 static void 2580 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) 2581 { 2582 if (tcp_bblogging_on(bbr->rc_tp)) { 2583 union tcp_log_stackspecific log; 2584 2585 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2586 log.u_bbr.flex1 = tsosz; 2587 log.u_bbr.flex2 = tls; 2588 log.u_bbr.flex3 = tcp_min_hptsi_time; 2589 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2590 log.u_bbr.flex5 = old_val; 2591 log.u_bbr.flex6 = maxseg; 2592 log.u_bbr.flex7 = bbr->rc_no_pacing; 2593 log.u_bbr.flex7 <<= 1; 2594 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2595 if (hdwr) 2596 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2597 else 2598 log.u_bbr.flex8 = bbr->rc_use_google; 2599 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2600 &bbr->rc_inp->inp_socket->so_rcv, 2601 &bbr->rc_inp->inp_socket->so_snd, 2602 BBR_LOG_BBRTSO, 0, 2603 0, &log, false, &bbr->rc_tv); 2604 } 2605 } 2606 2607 static void 2608 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2609 uint32_t flags, uint32_t line) 2610 { 2611 if (tcp_bblogging_on(bbr->rc_tp)) { 2612 union tcp_log_stackspecific log; 2613 2614 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2615 log.u_bbr.flex1 = line; 2616 log.u_bbr.flex2 = rsm->r_start; 2617 log.u_bbr.flex3 = rsm->r_end; 2618 log.u_bbr.flex4 = rsm->r_delivered; 2619 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2620 log.u_bbr.flex6 = rsm->r_dupack; 2621 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2622 log.u_bbr.flex8 = rsm->r_flags; 2623 /* Hijack the pkts_out fids */ 2624 log.u_bbr.applimited = flags; 2625 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2626 &bbr->rc_inp->inp_socket->so_rcv, 2627 &bbr->rc_inp->inp_socket->so_snd, 2628 BBR_RSM_CLEARED, 0, 2629 0, &log, false, &bbr->rc_tv); 2630 } 2631 } 2632 2633 static void 2634 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2635 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2636 uint32_t flex6, uint32_t pkts_out, int flex7, 2637 uint32_t flex4, uint32_t flex1) 2638 { 2639 2640 if (tcp_bblogging_on(bbr->rc_tp)) { 2641 union tcp_log_stackspecific log; 2642 2643 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2644 log.u_bbr.flex1 = flex1; 2645 log.u_bbr.flex2 = flex2; 2646 log.u_bbr.flex3 = flex3; 2647 log.u_bbr.flex4 = flex4; 2648 log.u_bbr.flex5 = flex5; 2649 log.u_bbr.flex6 = flex6; 2650 log.u_bbr.flex7 = flex7; 2651 /* Hijack the pkts_out fids */ 2652 log.u_bbr.pkts_out = pkts_out; 2653 log.u_bbr.flex8 = flex8; 2654 if (bbr->rc_ack_was_delayed) 2655 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2656 else 2657 log.u_bbr.epoch = 0; 2658 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2659 &bbr->rc_inp->inp_socket->so_rcv, 2660 &bbr->rc_inp->inp_socket->so_snd, 2661 BBR_LOG_BBRUPD, 0, 2662 flex2, &log, false, &bbr->rc_tv); 2663 } 2664 } 2665 2666 static void 2667 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2668 uint32_t newbw, uint32_t obw, uint32_t diff, 2669 uint32_t tim) 2670 { 2671 if (/*bbr_verbose_logging && */tcp_bblogging_on(bbr->rc_tp)) { 2672 union tcp_log_stackspecific log; 2673 2674 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2675 log.u_bbr.flex1 = reason; 2676 log.u_bbr.flex2 = newbw; 2677 log.u_bbr.flex3 = obw; 2678 log.u_bbr.flex4 = diff; 2679 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2680 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2681 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2682 log.u_bbr.pkts_out = tim; 2683 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2684 if (bbr->rc_lt_use_bw == 0) 2685 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2686 else 2687 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2688 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2689 &bbr->rc_inp->inp_socket->so_rcv, 2690 &bbr->rc_inp->inp_socket->so_snd, 2691 BBR_LOG_BWSAMP, 0, 2692 0, &log, false, &bbr->rc_tv); 2693 } 2694 } 2695 2696 static inline void 2697 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2698 { 2699 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2700 union tcp_log_stackspecific log; 2701 2702 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2703 log.u_bbr.flex1 = line; 2704 log.u_bbr.flex2 = tick; 2705 log.u_bbr.flex3 = tp->t_maxunacktime; 2706 log.u_bbr.flex4 = tp->t_acktime; 2707 log.u_bbr.flex8 = event; 2708 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2709 &bbr->rc_inp->inp_socket->so_rcv, 2710 &bbr->rc_inp->inp_socket->so_snd, 2711 BBR_LOG_PROGRESS, 0, 2712 0, &log, false, &bbr->rc_tv); 2713 } 2714 } 2715 2716 static void 2717 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2718 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2719 int error) 2720 { 2721 if (tcp_bblogging_on(bbr->rc_tp)) { 2722 union tcp_log_stackspecific log; 2723 2724 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2725 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2726 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2727 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2728 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2729 log.u_bbr.bw_inuse = rate; 2730 log.u_bbr.flex5 = line; 2731 log.u_bbr.flex6 = error; 2732 log.u_bbr.flex8 = bbr->skip_gain; 2733 log.u_bbr.flex8 <<= 1; 2734 log.u_bbr.flex8 |= bbr->gain_is_limited; 2735 log.u_bbr.flex8 <<= 1; 2736 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2737 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2738 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2739 &bbr->rc_inp->inp_socket->so_rcv, 2740 &bbr->rc_inp->inp_socket->so_snd, 2741 BBR_LOG_HDWR_PACE, 0, 2742 0, &log, false, &bbr->rc_tv); 2743 } 2744 } 2745 2746 static void 2747 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) 2748 { 2749 if (tcp_bblogging_on(bbr->rc_tp)) { 2750 union tcp_log_stackspecific log; 2751 2752 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2753 log.u_bbr.flex1 = slot; 2754 log.u_bbr.flex2 = del_by; 2755 log.u_bbr.flex3 = prev_delay; 2756 log.u_bbr.flex4 = line; 2757 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2758 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2759 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2760 log.u_bbr.flex8 = bbr->rc_in_persist; 2761 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2762 &bbr->rc_inp->inp_socket->so_rcv, 2763 &bbr->rc_inp->inp_socket->so_snd, 2764 BBR_LOG_BBRSND, 0, 2765 len, &log, false, &bbr->rc_tv); 2766 } 2767 } 2768 2769 static void 2770 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) 2771 { 2772 if (tcp_bblogging_on(bbr->rc_tp)) { 2773 union tcp_log_stackspecific log; 2774 2775 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2776 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2777 log.u_bbr.flex2 = 0; 2778 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2779 log.u_bbr.flex4 = end; 2780 log.u_bbr.flex5 = seq; 2781 log.u_bbr.flex6 = t; 2782 log.u_bbr.flex7 = match; 2783 log.u_bbr.flex8 = flags; 2784 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2785 &bbr->rc_inp->inp_socket->so_rcv, 2786 &bbr->rc_inp->inp_socket->so_snd, 2787 BBR_LOG_BBRRTT, 0, 2788 0, &log, false, &bbr->rc_tv); 2789 } 2790 } 2791 2792 static void 2793 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2794 { 2795 if (tcp_bblogging_on(bbr->rc_tp)) { 2796 union tcp_log_stackspecific log; 2797 2798 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2799 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2800 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2801 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2802 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2803 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2804 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2805 log.u_bbr.flex7 = 0; 2806 log.u_bbr.flex8 = entry_method; 2807 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2808 &bbr->rc_inp->inp_socket->so_rcv, 2809 &bbr->rc_inp->inp_socket->so_snd, 2810 BBR_LOG_EXIT_GAIN, 0, 2811 0, &log, false, &bbr->rc_tv); 2812 } 2813 } 2814 2815 static void 2816 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2817 { 2818 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2819 union tcp_log_stackspecific log; 2820 2821 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2822 /* R-HU */ 2823 log.u_bbr.flex1 = 0; 2824 log.u_bbr.flex2 = 0; 2825 log.u_bbr.flex3 = 0; 2826 log.u_bbr.flex4 = 0; 2827 log.u_bbr.flex7 = 0; 2828 log.u_bbr.flex8 = settings_desired; 2829 2830 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2831 &bbr->rc_inp->inp_socket->so_rcv, 2832 &bbr->rc_inp->inp_socket->so_snd, 2833 BBR_LOG_SETTINGS_CHG, 0, 2834 0, &log, false, &bbr->rc_tv); 2835 } 2836 } 2837 2838 /* 2839 * Returns the bw from the our filter. 2840 */ 2841 static inline uint64_t 2842 bbr_get_full_bw(struct tcp_bbr *bbr) 2843 { 2844 uint64_t bw; 2845 2846 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2847 2848 return (bw); 2849 } 2850 2851 static inline void 2852 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2853 { 2854 uint64_t calclr; 2855 uint32_t lost, del; 2856 2857 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2858 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2859 else 2860 lost = 0; 2861 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2862 if (lost == 0) { 2863 calclr = 0; 2864 } else if (del) { 2865 calclr = lost; 2866 calclr *= (uint64_t)1000; 2867 calclr /= (uint64_t)del; 2868 } else { 2869 /* Nothing delivered? 100.0% loss */ 2870 calclr = 1000; 2871 } 2872 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2873 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2874 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2875 bbr->r_ctl.rc_pkt_epoch++; 2876 if (bbr->rc_no_pacing && 2877 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2878 bbr->rc_no_pacing = 0; 2879 tcp_bbr_tso_size_check(bbr, cts); 2880 } 2881 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2882 bbr->r_ctl.rc_pkt_epoch_time = cts; 2883 /* What was our loss rate */ 2884 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2885 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2886 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2887 } 2888 2889 static inline void 2890 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2891 { 2892 uint32_t epoch_time; 2893 2894 /* Tick the RTT clock */ 2895 bbr->r_ctl.rc_rtt_epoch++; 2896 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2897 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2898 bbr->r_ctl.rc_rcv_epoch_start = cts; 2899 } 2900 2901 static inline void 2902 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2903 { 2904 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2905 bbr->rc_is_pkt_epoch_now = 1; 2906 } 2907 } 2908 2909 /* 2910 * Returns the bw from either the b/w filter 2911 * or from the lt_bw (if the connection is being 2912 * policed). 2913 */ 2914 static inline uint64_t 2915 __bbr_get_bw(struct tcp_bbr *bbr) 2916 { 2917 uint64_t bw, min_bw; 2918 uint64_t rtt; 2919 int gm_measure_cnt = 1; 2920 2921 /* 2922 * For startup we make, like google, a 2923 * minimum b/w. This is generated from the 2924 * IW and the rttProp. We do fall back to srtt 2925 * if for some reason (initial handshake) we don't 2926 * have a rttProp. We, in the worst case, fall back 2927 * to the configured min_bw (rc_initial_hptsi_bw). 2928 */ 2929 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2930 /* Attempt first to use rttProp */ 2931 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2932 if (rtt && (rtt < 0xffffffff)) { 2933 measure: 2934 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2935 ((uint64_t)1000000); 2936 min_bw /= rtt; 2937 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2938 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2939 } 2940 2941 } else if (bbr->rc_tp->t_srtt != 0) { 2942 /* No rttProp, use srtt? */ 2943 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2944 goto measure; 2945 } else { 2946 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2947 } 2948 } else 2949 min_bw = 0; 2950 2951 if ((bbr->rc_past_init_win == 0) && 2952 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2953 bbr->rc_past_init_win = 1; 2954 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2955 gm_measure_cnt = 0; 2956 if (gm_measure_cnt && 2957 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2958 (bbr->rc_past_init_win == 0))) { 2959 /* For google we use our guess rate until we get 1 measurement */ 2960 2961 use_initial_window: 2962 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2963 if (rtt && (rtt < 0xffffffff)) { 2964 /* 2965 * We have an RTT measurement. Use that in 2966 * combination with our initial window to calculate 2967 * a b/w. 2968 */ 2969 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2970 ((uint64_t)1000000); 2971 bw /= rtt; 2972 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2973 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2974 } 2975 } else { 2976 /* Drop back to the 40 and punt to a default */ 2977 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2978 } 2979 if (bw < 1) 2980 /* Probably should panic */ 2981 bw = 1; 2982 if (bw > min_bw) 2983 return (bw); 2984 else 2985 return (min_bw); 2986 } 2987 if (bbr->rc_lt_use_bw) 2988 bw = bbr->r_ctl.rc_lt_bw; 2989 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2990 bw = bbr->r_ctl.red_bw; 2991 else 2992 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2993 if (bw == 0) { 2994 /* We should not be at 0, go to the initial window then */ 2995 goto use_initial_window; 2996 } 2997 if (bw < 1) 2998 /* Probably should panic */ 2999 bw = 1; 3000 if (bw < min_bw) 3001 bw = min_bw; 3002 return (bw); 3003 } 3004 3005 static inline uint64_t 3006 bbr_get_bw(struct tcp_bbr *bbr) 3007 { 3008 uint64_t bw; 3009 3010 bw = __bbr_get_bw(bbr); 3011 return (bw); 3012 } 3013 3014 static inline void 3015 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3016 { 3017 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3018 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3019 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3020 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3021 } 3022 3023 static inline void 3024 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3025 { 3026 bbr->rc_lt_is_sampling = 0; 3027 bbr->rc_lt_use_bw = 0; 3028 bbr->r_ctl.rc_lt_bw = 0; 3029 bbr_reset_lt_bw_interval(bbr, cts); 3030 } 3031 3032 static inline void 3033 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3034 { 3035 uint64_t diff; 3036 3037 /* Do we have a previous sample? */ 3038 if (bbr->r_ctl.rc_lt_bw) { 3039 /* Get the diff in bytes per second */ 3040 if (bbr->r_ctl.rc_lt_bw > bw) 3041 diff = bbr->r_ctl.rc_lt_bw - bw; 3042 else 3043 diff = bw - bbr->r_ctl.rc_lt_bw; 3044 if ((diff <= bbr_lt_bw_diff) || 3045 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3046 /* Consider us policed */ 3047 uint32_t saved_bw; 3048 3049 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3050 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3051 bbr->rc_lt_use_bw = 1; 3052 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3053 /* 3054 * Use pkt based epoch for measuring length of 3055 * policer up 3056 */ 3057 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3058 /* 3059 * reason 4 is we need to start consider being 3060 * policed 3061 */ 3062 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3063 return; 3064 } 3065 } 3066 bbr->r_ctl.rc_lt_bw = bw; 3067 bbr_reset_lt_bw_interval(bbr, cts); 3068 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3069 } 3070 3071 static void 3072 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3073 { 3074 uint32_t ran, deduct; 3075 3076 ran = arc4random_uniform(bbr_rand_ot); 3077 if (ran) { 3078 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3079 bbr->r_ctl.rc_level_state_extra -= deduct; 3080 } 3081 } 3082 /* 3083 * Return randomly the starting state 3084 * to use in probebw. 3085 */ 3086 static uint8_t 3087 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3088 { 3089 uint32_t ran; 3090 uint8_t ret_val; 3091 3092 /* Initialize the offset to 0 */ 3093 bbr->r_ctl.rc_exta_time_gd = 0; 3094 bbr->rc_hit_state_1 = 0; 3095 bbr->r_ctl.rc_level_state_extra = 0; 3096 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3097 /* 3098 * The math works funny here :) the return value is used to set the 3099 * substate and then the state change is called which increments by 3100 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3101 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3102 * we return 1 - 7, so we dont return 0 and end up starting in 3103 * state 1 (DRAIN). 3104 */ 3105 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3106 /* Set an epoch */ 3107 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3108 bbr_set_epoch(bbr, cts, __LINE__); 3109 3110 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3111 return (ret_val); 3112 } 3113 3114 static void 3115 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3116 { 3117 uint32_t diff, d_time; 3118 uint64_t del_time, bw, lost, delivered; 3119 3120 if (bbr->r_use_policer == 0) 3121 return; 3122 if (bbr->rc_lt_use_bw) { 3123 /* We are using lt bw do we stop yet? */ 3124 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3125 if (diff > bbr_lt_bw_max_rtts) { 3126 /* Reset it all */ 3127 reset_all: 3128 bbr_reset_lt_bw_sampling(bbr, cts); 3129 if (bbr->rc_filled_pipe) { 3130 bbr_set_epoch(bbr, cts, __LINE__); 3131 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3132 bbr_substate_change(bbr, cts, __LINE__, 0); 3133 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3134 bbr_log_type_statechange(bbr, cts, __LINE__); 3135 } else { 3136 /* 3137 * This should not happen really 3138 * unless we remove the startup/drain 3139 * restrictions above. 3140 */ 3141 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3142 bbr_set_epoch(bbr, cts, __LINE__); 3143 bbr->r_ctl.rc_bbr_state_time = cts; 3144 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3145 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3146 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3147 bbr_set_state_target(bbr, __LINE__); 3148 bbr_log_type_statechange(bbr, cts, __LINE__); 3149 } 3150 /* reason 0 is to stop using lt-bw */ 3151 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3152 return; 3153 } 3154 if (bbr_lt_intvl_fp == 0) { 3155 /* Not doing false-positive detection */ 3156 return; 3157 } 3158 /* False positive detection */ 3159 if (diff == bbr_lt_intvl_fp) { 3160 /* At bbr_lt_intvl_fp we record the lost */ 3161 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3162 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3163 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3164 /* Now is our loss rate still high? */ 3165 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3166 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3167 if ((delivered == 0) || 3168 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3169 /* No still below our threshold */ 3170 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3171 } else { 3172 /* Yikes its still high, it must be a false positive */ 3173 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3174 goto reset_all; 3175 } 3176 } 3177 return; 3178 } 3179 /* 3180 * Wait for the first loss before sampling, to let the policer 3181 * exhaust its tokens and estimate the steady-state rate allowed by 3182 * the policer. Starting samples earlier includes bursts that 3183 * over-estimate the bw. 3184 */ 3185 if (bbr->rc_lt_is_sampling == 0) { 3186 /* reason 1 is to begin doing the sampling */ 3187 if (loss_detected == 0) 3188 return; 3189 bbr_reset_lt_bw_interval(bbr, cts); 3190 bbr->rc_lt_is_sampling = 1; 3191 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3192 return; 3193 } 3194 /* Now how long were we delivering long term last> */ 3195 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3196 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3197 else 3198 d_time = 0; 3199 3200 /* To avoid underestimates, reset sampling if we run out of data. */ 3201 if (bbr->r_ctl.r_app_limited_until) { 3202 /* Can not measure in app-limited state */ 3203 bbr_reset_lt_bw_sampling(bbr, cts); 3204 /* reason 2 is to reset sampling due to app limits */ 3205 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3206 return; 3207 } 3208 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3209 if (diff < bbr_lt_intvl_min_rtts) { 3210 /* 3211 * need more samples (we don't 3212 * start on a round like linux so 3213 * we need 1 more). 3214 */ 3215 /* 6 is not_enough time or no-loss */ 3216 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3217 return; 3218 } 3219 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3220 /* 3221 * For now if we wait too long, reset all sampling. We need 3222 * to do some research here, its possible that we should 3223 * base this on how much loss as occurred.. something like 3224 * if its under 10% (or some thresh) reset all otherwise 3225 * don't. Thats for phase II I guess. 3226 */ 3227 bbr_reset_lt_bw_sampling(bbr, cts); 3228 /* reason 3 is to reset sampling due too long of sampling */ 3229 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3230 return; 3231 } 3232 /* 3233 * End sampling interval when a packet is lost, so we estimate the 3234 * policer tokens were exhausted. Stopping the sampling before the 3235 * tokens are exhausted under-estimates the policed rate. 3236 */ 3237 if (loss_detected == 0) { 3238 /* 6 is not_enough time or no-loss */ 3239 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3240 return; 3241 } 3242 /* Calculate packets lost and delivered in sampling interval. */ 3243 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3244 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3245 if ((delivered == 0) || 3246 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3247 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3248 return; 3249 } 3250 if (d_time < 1000) { 3251 /* Not enough time. wait */ 3252 /* 6 is not_enough time or no-loss */ 3253 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3254 return; 3255 } 3256 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3257 /* Too long */ 3258 bbr_reset_lt_bw_sampling(bbr, cts); 3259 /* reason 3 is to reset sampling due too long of sampling */ 3260 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3261 return; 3262 } 3263 del_time = d_time; 3264 bw = delivered; 3265 bw *= (uint64_t)USECS_IN_SECOND; 3266 bw /= del_time; 3267 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3268 } 3269 3270 /* 3271 * Allocate a sendmap from our zone. 3272 */ 3273 static struct bbr_sendmap * 3274 bbr_alloc(struct tcp_bbr *bbr) 3275 { 3276 struct bbr_sendmap *rsm; 3277 3278 BBR_STAT_INC(bbr_to_alloc); 3279 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3280 if (rsm) { 3281 bbr->r_ctl.rc_num_maps_alloced++; 3282 return (rsm); 3283 } 3284 if (bbr->r_ctl.rc_free_cnt) { 3285 BBR_STAT_INC(bbr_to_alloc_emerg); 3286 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3287 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3288 bbr->r_ctl.rc_free_cnt--; 3289 return (rsm); 3290 } 3291 BBR_STAT_INC(bbr_to_alloc_failed); 3292 return (NULL); 3293 } 3294 3295 static struct bbr_sendmap * 3296 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3297 { 3298 if ((V_tcp_map_entries_limit > 0) && 3299 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3300 BBR_STAT_INC(bbr_alloc_limited); 3301 if (!bbr->alloc_limit_reported) { 3302 bbr->alloc_limit_reported = 1; 3303 BBR_STAT_INC(bbr_alloc_limited_conns); 3304 } 3305 return (NULL); 3306 } 3307 return (bbr_alloc(bbr)); 3308 } 3309 3310 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3311 static struct bbr_sendmap * 3312 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3313 { 3314 struct bbr_sendmap *rsm; 3315 3316 if (limit_type) { 3317 /* currently there is only one limit type */ 3318 if (V_tcp_map_split_limit > 0 && 3319 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3320 BBR_STAT_INC(bbr_split_limited); 3321 if (!bbr->alloc_limit_reported) { 3322 bbr->alloc_limit_reported = 1; 3323 BBR_STAT_INC(bbr_alloc_limited_conns); 3324 } 3325 return (NULL); 3326 } 3327 } 3328 3329 /* allocate and mark in the limit type, if set */ 3330 rsm = bbr_alloc(bbr); 3331 if (rsm != NULL && limit_type) { 3332 rsm->r_limit_type = limit_type; 3333 bbr->r_ctl.rc_num_split_allocs++; 3334 } 3335 return (rsm); 3336 } 3337 3338 static void 3339 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3340 { 3341 if (rsm->r_limit_type) { 3342 /* currently there is only one limit type */ 3343 bbr->r_ctl.rc_num_split_allocs--; 3344 } 3345 if (rsm->r_is_smallmap) 3346 bbr->r_ctl.rc_num_small_maps_alloced--; 3347 if (bbr->r_ctl.rc_tlp_send == rsm) 3348 bbr->r_ctl.rc_tlp_send = NULL; 3349 if (bbr->r_ctl.rc_resend == rsm) { 3350 bbr->r_ctl.rc_resend = NULL; 3351 } 3352 if (bbr->r_ctl.rc_next == rsm) 3353 bbr->r_ctl.rc_next = NULL; 3354 if (bbr->r_ctl.rc_sacklast == rsm) 3355 bbr->r_ctl.rc_sacklast = NULL; 3356 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3357 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3358 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3359 rsm->r_limit_type = 0; 3360 bbr->r_ctl.rc_free_cnt++; 3361 return; 3362 } 3363 bbr->r_ctl.rc_num_maps_alloced--; 3364 uma_zfree(bbr_zone, rsm); 3365 } 3366 3367 /* 3368 * Returns the BDP. 3369 */ 3370 static uint64_t 3371 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3372 /* 3373 * Calculate the bytes in flight needed given the bw (in bytes per 3374 * second) and the specifyed rtt in useconds. We need to put out the 3375 * returned value per RTT to match that rate. Gain will normally 3376 * raise it up from there. 3377 * 3378 * This should not overflow as long as the bandwidth is below 1 3379 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3380 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3381 */ 3382 uint64_t usec_per_sec; 3383 3384 usec_per_sec = USECS_IN_SECOND; 3385 return ((rtt * bw) / usec_per_sec); 3386 } 3387 3388 /* 3389 * Return the initial cwnd. 3390 */ 3391 static uint32_t 3392 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3393 { 3394 uint32_t i_cwnd; 3395 3396 if (bbr->rc_init_win) { 3397 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3398 } else if (V_tcp_initcwnd_segments) 3399 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3400 max(2 * tp->t_maxseg, 14600)); 3401 else if (V_tcp_do_rfc3390) 3402 i_cwnd = min(4 * tp->t_maxseg, 3403 max(2 * tp->t_maxseg, 4380)); 3404 else { 3405 /* Per RFC5681 Section 3.1 */ 3406 if (tp->t_maxseg > 2190) 3407 i_cwnd = 2 * tp->t_maxseg; 3408 else if (tp->t_maxseg > 1095) 3409 i_cwnd = 3 * tp->t_maxseg; 3410 else 3411 i_cwnd = 4 * tp->t_maxseg; 3412 } 3413 return (i_cwnd); 3414 } 3415 3416 /* 3417 * Given a specified gain, return the target 3418 * cwnd based on that gain. 3419 */ 3420 static uint32_t 3421 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3422 { 3423 uint64_t bdp, rtt; 3424 uint32_t cwnd; 3425 3426 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3427 (bbr_get_full_bw(bbr) == 0)) { 3428 /* No measurements yet */ 3429 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3430 } 3431 /* 3432 * Get bytes per RTT needed (rttProp is normally in 3433 * bbr_cwndtarget_rtt_touse) 3434 */ 3435 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3436 /* Get the bdp from the two values */ 3437 bdp = bbr_get_bw_delay_prod(rtt, bw); 3438 /* Now apply the gain */ 3439 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3440 3441 return (cwnd); 3442 } 3443 3444 static uint32_t 3445 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3446 { 3447 uint32_t cwnd, mss; 3448 3449 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3450 /* Get the base cwnd with gain rounded to a mss */ 3451 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3452 /* 3453 * Add in N (2 default since we do not have a 3454 * fq layer to trap packets in) quanta's per the I-D 3455 * section 4.2.3.2 quanta adjust. 3456 */ 3457 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3458 if (bbr->rc_use_google) { 3459 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3460 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3461 /* 3462 * The linux implementation adds 3463 * an extra 2 x mss in gain cycle which 3464 * is documented no-where except in the code. 3465 * so we add more for Neal undocumented feature 3466 */ 3467 cwnd += 2 * mss; 3468 } 3469 if ((cwnd / mss) & 0x1) { 3470 /* Round up for odd num mss */ 3471 cwnd += mss; 3472 } 3473 } 3474 /* Are we below the min cwnd? */ 3475 if (cwnd < get_min_cwnd(bbr)) 3476 return (get_min_cwnd(bbr)); 3477 return (cwnd); 3478 } 3479 3480 static uint16_t 3481 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3482 { 3483 if (gain < 1) 3484 gain = 1; 3485 return (gain); 3486 } 3487 3488 static uint32_t 3489 bbr_get_header_oh(struct tcp_bbr *bbr) 3490 { 3491 int seg_oh; 3492 3493 seg_oh = 0; 3494 if (bbr->r_ctl.rc_inc_tcp_oh) { 3495 /* Do we include TCP overhead? */ 3496 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3497 } 3498 if (bbr->r_ctl.rc_inc_ip_oh) { 3499 /* Do we include IP overhead? */ 3500 #ifdef INET6 3501 if (bbr->r_is_v6) { 3502 seg_oh += sizeof(struct ip6_hdr); 3503 } else 3504 #endif 3505 { 3506 3507 #ifdef INET 3508 seg_oh += sizeof(struct ip); 3509 #endif 3510 } 3511 } 3512 if (bbr->r_ctl.rc_inc_enet_oh) { 3513 /* Do we include the ethernet overhead? */ 3514 seg_oh += sizeof(struct ether_header); 3515 } 3516 return(seg_oh); 3517 } 3518 3519 static uint32_t 3520 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3521 { 3522 uint64_t divor, res, tim; 3523 3524 if (useconds_time == 0) 3525 return (0); 3526 gain = bbr_gain_adjust(bbr, gain); 3527 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3528 tim = useconds_time; 3529 res = (tim * bw * gain) / divor; 3530 if (res == 0) 3531 res = 1; 3532 return ((uint32_t)res); 3533 } 3534 3535 /* 3536 * Given a gain and a length return the delay in useconds that 3537 * should be used to evenly space out packets 3538 * on the connection (based on the gain factor). 3539 */ 3540 static uint32_t 3541 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3542 { 3543 uint64_t bw, lentim, res; 3544 uint32_t usecs, srtt, over = 0; 3545 uint32_t seg_oh, num_segs, maxseg; 3546 3547 if (len == 0) 3548 return (0); 3549 3550 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3551 num_segs = (len + maxseg - 1) / maxseg; 3552 if (bbr->rc_use_google == 0) { 3553 seg_oh = bbr_get_header_oh(bbr); 3554 len += (num_segs * seg_oh); 3555 } 3556 gain = bbr_gain_adjust(bbr, gain); 3557 bw = bbr_get_bw(bbr); 3558 if (bbr->rc_use_google) { 3559 uint64_t cbw; 3560 3561 /* 3562 * Reduce the b/w by the google discount 3563 * factor 10 = 1%. 3564 */ 3565 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3566 cbw /= (uint64_t)1000; 3567 /* We don't apply a discount if it results in 0 */ 3568 if (cbw > 0) 3569 bw = cbw; 3570 } 3571 lentim = ((uint64_t)len * 3572 (uint64_t)USECS_IN_SECOND * 3573 (uint64_t)BBR_UNIT); 3574 res = lentim / ((uint64_t)gain * bw); 3575 if (res == 0) 3576 res = 1; 3577 usecs = (uint32_t)res; 3578 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3579 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3580 (bbr->rc_use_google == 0) && 3581 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3582 /* 3583 * We cannot let the delay be more than 1/2 the srtt time. 3584 * Otherwise we cannot pace out or send properly. 3585 */ 3586 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3587 BBR_STAT_INC(bbr_hpts_min_time); 3588 } 3589 if (!nolog) 3590 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3591 return (usecs); 3592 } 3593 3594 static void 3595 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3596 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3597 { 3598 uint64_t bw; 3599 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3600 int32_t meth; 3601 3602 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3603 3604 #ifdef STATS 3605 if ((tp->t_flags & TF_GPUTINPROG) && 3606 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3607 /* 3608 * Strech acks and compressed acks will cause this to 3609 * oscillate but we are doing it the same way as the main 3610 * stack so it will be compariable (though possibly not 3611 * ideal). 3612 */ 3613 int32_t cgput; 3614 int64_t gput, time_stamp; 3615 3616 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3617 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3618 cgput = gput / time_stamp; 3619 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3620 cgput); 3621 if (tp->t_stats_gput_prev > 0) 3622 stats_voi_update_abs_s32(tp->t_stats, 3623 VOI_TCP_GPUT_ND, 3624 ((gput - tp->t_stats_gput_prev) * 100) / 3625 tp->t_stats_gput_prev); 3626 tp->t_flags &= ~TF_GPUTINPROG; 3627 tp->t_stats_gput_prev = cgput; 3628 } 3629 #endif 3630 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3631 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3632 /* We don't change anything in probe-rtt */ 3633 return; 3634 } 3635 maxseg = tp->t_maxseg - bbr->rc_last_options; 3636 saved_bytes = bytes_this_ack; 3637 bytes_this_ack += sack_changed; 3638 if (bytes_this_ack > prev_acked) { 3639 bytes_this_ack -= prev_acked; 3640 /* 3641 * A byte ack'd gives us a full mss 3642 * to be like linux i.e. they count packets. 3643 */ 3644 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3645 bytes_this_ack = maxseg; 3646 } else { 3647 /* Unlikely */ 3648 bytes_this_ack = 0; 3649 } 3650 cwnd = tp->snd_cwnd; 3651 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3652 if (bw) 3653 target_cwnd = bbr_get_target_cwnd(bbr, 3654 bw, 3655 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3656 else 3657 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3658 if (IN_RECOVERY(tp->t_flags) && 3659 (bbr->bbr_prev_in_rec == 0)) { 3660 /* 3661 * We are entering recovery and 3662 * thus packet conservation. 3663 */ 3664 bbr->pkt_conservation = 1; 3665 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3666 cwnd = ctf_flight_size(tp, 3667 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3668 bytes_this_ack; 3669 } 3670 if (IN_RECOVERY(tp->t_flags)) { 3671 uint32_t flight; 3672 3673 bbr->bbr_prev_in_rec = 1; 3674 if (cwnd > losses) { 3675 cwnd -= losses; 3676 if (cwnd < maxseg) 3677 cwnd = maxseg; 3678 } else 3679 cwnd = maxseg; 3680 flight = ctf_flight_size(tp, 3681 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3682 bbr_log_type_cwndupd(bbr, flight, 0, 3683 losses, 10, 0, 0, line); 3684 if (bbr->pkt_conservation) { 3685 uint32_t time_in; 3686 3687 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3688 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3689 else 3690 time_in = 0; 3691 3692 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3693 /* Clear packet conservation after an rttProp */ 3694 bbr->pkt_conservation = 0; 3695 } else { 3696 if ((flight + bytes_this_ack) > cwnd) 3697 cwnd = flight + bytes_this_ack; 3698 if (cwnd < get_min_cwnd(bbr)) 3699 cwnd = get_min_cwnd(bbr); 3700 tp->snd_cwnd = cwnd; 3701 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3702 prev_acked, 1, target_cwnd, th->th_ack, line); 3703 return; 3704 } 3705 } 3706 } else 3707 bbr->bbr_prev_in_rec = 0; 3708 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3709 bbr->r_ctl.restrict_growth--; 3710 if (bytes_this_ack > maxseg) 3711 bytes_this_ack = maxseg; 3712 } 3713 if (bbr->rc_filled_pipe) { 3714 /* 3715 * Here we have exited startup and filled the pipe. We will 3716 * thus allow the cwnd to shrink to the target. We hit here 3717 * mostly. 3718 */ 3719 uint32_t s_cwnd; 3720 3721 meth = 2; 3722 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3723 if (s_cwnd > cwnd) 3724 cwnd = s_cwnd; 3725 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3726 cwnd = s_cwnd; 3727 } else { 3728 /* 3729 * Here we are still in startup, we increase cwnd by what 3730 * has been acked. 3731 */ 3732 if ((cwnd < target_cwnd) || 3733 (bbr->rc_past_init_win == 0)) { 3734 meth = 3; 3735 cwnd += bytes_this_ack; 3736 } else { 3737 /* 3738 * Method 4 means we are at target so no gain in 3739 * startup and past the initial window. 3740 */ 3741 meth = 4; 3742 } 3743 } 3744 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3745 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3746 } 3747 3748 static void 3749 tcp_bbr_partialack(struct tcpcb *tp) 3750 { 3751 struct tcp_bbr *bbr; 3752 3753 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3754 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3755 if (ctf_flight_size(tp, 3756 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3757 tp->snd_cwnd) { 3758 bbr->r_wanted_output = 1; 3759 } 3760 } 3761 3762 static void 3763 bbr_post_recovery(struct tcpcb *tp) 3764 { 3765 struct tcp_bbr *bbr; 3766 uint32_t flight; 3767 3768 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3769 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3770 /* 3771 * Here we just exit recovery. 3772 */ 3773 EXIT_RECOVERY(tp->t_flags); 3774 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3775 bbr->r_recovery_bw = 0; 3776 tp->snd_recover = tp->snd_una; 3777 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3778 bbr->pkt_conservation = 0; 3779 if (bbr->rc_use_google == 0) { 3780 /* 3781 * For non-google mode lets 3782 * go ahead and make sure we clear 3783 * the recovery state so if we 3784 * bounce back in to recovery we 3785 * will do PC. 3786 */ 3787 bbr->bbr_prev_in_rec = 0; 3788 } 3789 bbr_log_type_exit_rec(bbr); 3790 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3791 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3792 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3793 } else { 3794 /* For probe-rtt case lets fix up its saved_cwnd */ 3795 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3796 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3797 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3798 } 3799 } 3800 flight = ctf_flight_size(tp, 3801 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3802 if ((bbr->rc_use_google == 0) && 3803 bbr_do_red) { 3804 uint64_t val, lr2use; 3805 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3806 uint32_t *cwnd_p; 3807 3808 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3809 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3810 val /= bbr_get_rtt(bbr, BBR_SRTT); 3811 ratio = (uint32_t)val; 3812 } else 3813 ratio = 1000; 3814 3815 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3816 bbr->r_ctl.recovery_lr, 21, 3817 ratio, 3818 bbr->r_ctl.rc_red_cwnd_pe, 3819 __LINE__); 3820 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3821 goto done; 3822 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3823 bbr_prtt_slam_cwnd) || 3824 (bbr_sub_drain_slam_cwnd && 3825 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3826 bbr->rc_hit_state_1 && 3827 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3828 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3829 bbr_slam_cwnd_in_main_drain)) { 3830 /* 3831 * Here we must poke at the saved cwnd 3832 * as well as the cwnd. 3833 */ 3834 cwnd = bbr->r_ctl.rc_saved_cwnd; 3835 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3836 } else { 3837 cwnd = tp->snd_cwnd; 3838 cwnd_p = &tp->snd_cwnd; 3839 } 3840 maxseg = tp->t_maxseg - bbr->rc_last_options; 3841 /* Add the overall lr with the recovery lr */ 3842 if (bbr->r_ctl.rc_lost == 0) 3843 lr2use = 0; 3844 else if (bbr->r_ctl.rc_delivered == 0) 3845 lr2use = 1000; 3846 else { 3847 lr2use = bbr->r_ctl.rc_lost * 1000; 3848 lr2use /= bbr->r_ctl.rc_delivered; 3849 } 3850 lr2use += bbr->r_ctl.recovery_lr; 3851 acks_inflight = (flight / (maxseg * 2)); 3852 if (bbr_red_scale) { 3853 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3854 lr2use /= bbr_red_scale; 3855 if ((bbr_red_growth_restrict) && 3856 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3857 bbr->r_ctl.restrict_growth += acks_inflight; 3858 } 3859 if (lr2use) { 3860 val = (uint64_t)cwnd * lr2use; 3861 val /= 1000; 3862 if (cwnd > val) 3863 newcwnd = roundup((cwnd - val), maxseg); 3864 else 3865 newcwnd = maxseg; 3866 } else { 3867 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3868 val /= (uint64_t)bbr_red_div; 3869 newcwnd = roundup((uint32_t)val, maxseg); 3870 } 3871 /* with standard delayed acks how many acks can I expect? */ 3872 if (bbr_drop_limit == 0) { 3873 /* 3874 * Anticpate how much we will 3875 * raise the cwnd based on the acks. 3876 */ 3877 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3878 /* We do enforce the min (with the acks) */ 3879 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3880 } 3881 } else { 3882 /* 3883 * A strict drop limit of N is inplace 3884 */ 3885 if (newcwnd < (bbr_drop_limit * maxseg)) { 3886 newcwnd = bbr_drop_limit * maxseg; 3887 } 3888 } 3889 /* For the next N acks do we restrict the growth */ 3890 *cwnd_p = newcwnd; 3891 if (tp->snd_cwnd > newcwnd) 3892 tp->snd_cwnd = newcwnd; 3893 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3894 (uint32_t)lr2use, 3895 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3896 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3897 } 3898 done: 3899 bbr->r_ctl.recovery_lr = 0; 3900 if (flight <= tp->snd_cwnd) { 3901 bbr->r_wanted_output = 1; 3902 } 3903 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3904 } 3905 3906 static void 3907 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3908 { 3909 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3910 /* Limit the drop in b/w to 1/2 our current filter. */ 3911 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3912 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3913 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3914 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3915 tcp_bbr_tso_size_check(bbr, cts); 3916 } 3917 3918 static void 3919 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3920 { 3921 struct tcp_bbr *bbr; 3922 3923 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3924 #ifdef STATS 3925 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3926 #endif 3927 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3928 switch (type) { 3929 case CC_NDUPACK: 3930 if (!IN_RECOVERY(tp->t_flags)) { 3931 tp->snd_recover = tp->snd_max; 3932 /* Start a new epoch */ 3933 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3934 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3935 /* 3936 * Move forward the lt epoch 3937 * so it won't count the truncated 3938 * epoch. 3939 */ 3940 bbr->r_ctl.rc_lt_epoch++; 3941 } 3942 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3943 /* 3944 * Just like the policer detection code 3945 * if we are in startup we must push 3946 * forward the last startup epoch 3947 * to hide the truncated PE. 3948 */ 3949 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3950 } 3951 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3952 ENTER_RECOVERY(tp->t_flags); 3953 bbr->rc_tlp_rtx_out = 0; 3954 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3955 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3956 if (tcp_in_hpts(bbr->rc_tp) && 3957 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3958 /* 3959 * When we enter recovery, we need to restart 3960 * any timers. This may mean we gain an agg 3961 * early, which will be made up for at the last 3962 * rxt out. 3963 */ 3964 bbr->rc_timer_first = 1; 3965 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3966 } 3967 /* 3968 * Calculate a new cwnd based on to the current 3969 * delivery rate with no gain. We get the bdp 3970 * without gaining it up like we normally would and 3971 * we use the last cur_del_rate. 3972 */ 3973 if ((bbr->rc_use_google == 0) && 3974 (bbr->r_ctl.bbr_rttprobe_gain_val || 3975 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3976 tp->snd_cwnd = ctf_flight_size(tp, 3977 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3978 (tp->t_maxseg - bbr->rc_last_options); 3979 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3980 /* We always gate to min cwnd */ 3981 tp->snd_cwnd = get_min_cwnd(bbr); 3982 } 3983 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3984 } 3985 bbr_log_type_enter_rec(bbr, rsm->r_start); 3986 } 3987 break; 3988 case CC_RTO_ERR: 3989 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 3990 /* RTO was unnecessary, so reset everything. */ 3991 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 3992 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3993 tp->snd_cwnd = tp->snd_cwnd_prev; 3994 tp->snd_ssthresh = tp->snd_ssthresh_prev; 3995 tp->snd_recover = tp->snd_recover_prev; 3996 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3997 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 3998 } 3999 tp->t_badrxtwin = 0; 4000 break; 4001 } 4002 } 4003 4004 /* 4005 * Indicate whether this ack should be delayed. We can delay the ack if 4006 * following conditions are met: 4007 * - There is no delayed ack timer in progress. 4008 * - Our last ack wasn't a 0-sized window. We never want to delay 4009 * the ack that opens up a 0-sized window. 4010 * - LRO wasn't used for this segment. We make sure by checking that the 4011 * segment size is not larger than the MSS. 4012 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4013 * connection. 4014 * - The data being acked is less than a full segment (a stretch ack 4015 * of more than a segment we should ack. 4016 * - nsegs is 1 (if its more than that we received more than 1 ack). 4017 */ 4018 #define DELAY_ACK(tp, bbr, nsegs) \ 4019 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4020 ((tp->t_flags & TF_DELACK) == 0) && \ 4021 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4022 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4023 4024 /* 4025 * Return the lowest RSM in the map of 4026 * packets still in flight that is not acked. 4027 * This should normally find on the first one 4028 * since we remove packets from the send 4029 * map after they are marked ACKED. 4030 */ 4031 static struct bbr_sendmap * 4032 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4033 { 4034 struct bbr_sendmap *rsm; 4035 4036 /* 4037 * Walk the time-order transmitted list looking for an rsm that is 4038 * not acked. This will be the one that was sent the longest time 4039 * ago that is still outstanding. 4040 */ 4041 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4042 if (rsm->r_flags & BBR_ACKED) { 4043 continue; 4044 } 4045 goto finish; 4046 } 4047 finish: 4048 return (rsm); 4049 } 4050 4051 static struct bbr_sendmap * 4052 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4053 { 4054 struct bbr_sendmap *prsm; 4055 4056 /* 4057 * Walk the sequence order list backward until we hit and arrive at 4058 * the highest seq not acked. In theory when this is called it 4059 * should be the last segment (which it was not). 4060 */ 4061 prsm = rsm; 4062 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4063 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4064 continue; 4065 } 4066 return (prsm); 4067 } 4068 return (NULL); 4069 } 4070 4071 /* 4072 * Returns to the caller the number of microseconds that 4073 * the packet can be outstanding before we think we 4074 * should have had an ack returned. 4075 */ 4076 static uint32_t 4077 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4078 { 4079 /* 4080 * lro is the flag we use to determine if we have seen reordering. 4081 * If it gets set we have seen reordering. The reorder logic either 4082 * works in one of two ways: 4083 * 4084 * If reorder-fade is configured, then we track the last time we saw 4085 * re-ordering occur. If we reach the point where enough time as 4086 * passed we no longer consider reordering has occuring. 4087 * 4088 * Or if reorder-face is 0, then once we see reordering we consider 4089 * the connection to alway be subject to reordering and just set lro 4090 * to 1. 4091 * 4092 * In the end if lro is non-zero we add the extra time for 4093 * reordering in. 4094 */ 4095 int32_t lro; 4096 uint32_t thresh, t_rxtcur; 4097 4098 if (srtt == 0) 4099 srtt = 1; 4100 if (bbr->r_ctl.rc_reorder_ts) { 4101 if (bbr->r_ctl.rc_reorder_fade) { 4102 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4103 lro = cts - bbr->r_ctl.rc_reorder_ts; 4104 if (lro == 0) { 4105 /* 4106 * No time as passed since the last 4107 * reorder, mark it as reordering. 4108 */ 4109 lro = 1; 4110 } 4111 } else { 4112 /* Negative time? */ 4113 lro = 0; 4114 } 4115 if (lro > bbr->r_ctl.rc_reorder_fade) { 4116 /* Turn off reordering seen too */ 4117 bbr->r_ctl.rc_reorder_ts = 0; 4118 lro = 0; 4119 } 4120 } else { 4121 /* Reodering does not fade */ 4122 lro = 1; 4123 } 4124 } else { 4125 lro = 0; 4126 } 4127 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4128 if (lro) { 4129 /* It must be set, if not you get 1/4 rtt */ 4130 if (bbr->r_ctl.rc_reorder_shift) 4131 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4132 else 4133 thresh += (srtt >> 2); 4134 } else { 4135 thresh += 1000; 4136 } 4137 /* We don't let the rack timeout be above a RTO */ 4138 if ((bbr->rc_tp)->t_srtt == 0) 4139 t_rxtcur = BBR_INITIAL_RTO; 4140 else 4141 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4142 if (thresh > t_rxtcur) { 4143 thresh = t_rxtcur; 4144 } 4145 /* And we don't want it above the RTO max either */ 4146 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4147 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4148 } 4149 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4150 return (thresh); 4151 } 4152 4153 /* 4154 * Return to the caller the amount of time in mico-seconds 4155 * that should be used for the TLP timer from the last 4156 * send time of this packet. 4157 */ 4158 static uint32_t 4159 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4160 struct bbr_sendmap *rsm, uint32_t srtt, 4161 uint32_t cts) 4162 { 4163 uint32_t thresh, len, maxseg, t_rxtcur; 4164 struct bbr_sendmap *prsm; 4165 4166 if (srtt == 0) 4167 srtt = 1; 4168 if (bbr->rc_tlp_threshold) 4169 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4170 else 4171 thresh = (srtt * 2); 4172 maxseg = tp->t_maxseg - bbr->rc_last_options; 4173 /* Get the previous sent packet, if any */ 4174 len = rsm->r_end - rsm->r_start; 4175 4176 /* 2.1 behavior */ 4177 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4178 if (prsm && (len <= maxseg)) { 4179 /* 4180 * Two packets outstanding, thresh should be (2*srtt) + 4181 * possible inter-packet delay (if any). 4182 */ 4183 uint32_t inter_gap = 0; 4184 int idx, nidx; 4185 4186 idx = rsm->r_rtr_cnt - 1; 4187 nidx = prsm->r_rtr_cnt - 1; 4188 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4189 /* Yes it was sent later (or at the same time) */ 4190 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4191 } 4192 thresh += inter_gap; 4193 } else if (len <= maxseg) { 4194 /* 4195 * Possibly compensate for delayed-ack. 4196 */ 4197 uint32_t alt_thresh; 4198 4199 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4200 if (alt_thresh > thresh) 4201 thresh = alt_thresh; 4202 } 4203 /* Not above the current RTO */ 4204 if (tp->t_srtt == 0) 4205 t_rxtcur = BBR_INITIAL_RTO; 4206 else 4207 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4208 4209 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4210 /* Not above an RTO */ 4211 if (thresh > t_rxtcur) { 4212 thresh = t_rxtcur; 4213 } 4214 /* Not above a RTO max */ 4215 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4216 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4217 } 4218 /* And now apply the user TLP min */ 4219 if (thresh < bbr_tlp_min) { 4220 thresh = bbr_tlp_min; 4221 } 4222 return (thresh); 4223 } 4224 4225 /* 4226 * Return one of three RTTs to use (in microseconds). 4227 */ 4228 static __inline uint32_t 4229 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4230 { 4231 uint32_t f_rtt; 4232 uint32_t srtt; 4233 4234 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4235 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4236 /* We have no rtt at all */ 4237 if (bbr->rc_tp->t_srtt == 0) 4238 f_rtt = BBR_INITIAL_RTO; 4239 else 4240 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4241 /* 4242 * Since we don't know how good the rtt is apply a 4243 * delayed-ack min 4244 */ 4245 if (f_rtt < bbr_delayed_ack_time) { 4246 f_rtt = bbr_delayed_ack_time; 4247 } 4248 } 4249 /* Take the filter version or last measured pkt-rtt */ 4250 if (rtt_type == BBR_RTT_PROP) { 4251 srtt = f_rtt; 4252 } else if (rtt_type == BBR_RTT_PKTRTT) { 4253 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4254 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4255 } else { 4256 /* No pkt rtt yet */ 4257 srtt = f_rtt; 4258 } 4259 } else if (rtt_type == BBR_RTT_RACK) { 4260 srtt = bbr->r_ctl.rc_last_rtt; 4261 /* We need to add in any internal delay for our timer */ 4262 if (bbr->rc_ack_was_delayed) 4263 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4264 } else if (rtt_type == BBR_SRTT) { 4265 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4266 } else { 4267 /* TSNH */ 4268 srtt = f_rtt; 4269 #ifdef BBR_INVARIANTS 4270 panic("Unknown rtt request type %d", rtt_type); 4271 #endif 4272 } 4273 return (srtt); 4274 } 4275 4276 static int 4277 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4278 { 4279 uint32_t thresh; 4280 4281 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4282 cts, rsm); 4283 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4284 /* It is lost (past time) */ 4285 return (1); 4286 } 4287 return (0); 4288 } 4289 4290 /* 4291 * Return a sendmap if we need to retransmit something. 4292 */ 4293 static struct bbr_sendmap * 4294 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4295 { 4296 /* 4297 * Check to see that we don't need to fall into recovery. We will 4298 * need to do so if our oldest transmit is past the time we should 4299 * have had an ack. 4300 */ 4301 4302 struct bbr_sendmap *rsm; 4303 int32_t idx; 4304 4305 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4306 /* Nothing outstanding that we know of */ 4307 return (NULL); 4308 } 4309 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4310 if (rsm == NULL) { 4311 /* Nothing in the transmit map */ 4312 return (NULL); 4313 } 4314 if (tp->t_flags & TF_SENTFIN) { 4315 /* Fin restricted, don't find anything once a fin is sent */ 4316 return (NULL); 4317 } 4318 if (rsm->r_flags & BBR_ACKED) { 4319 /* 4320 * Ok the first one is acked (this really should not happen 4321 * since we remove the from the tmap once they are acked) 4322 */ 4323 rsm = bbr_find_lowest_rsm(bbr); 4324 if (rsm == NULL) 4325 return (NULL); 4326 } 4327 idx = rsm->r_rtr_cnt - 1; 4328 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4329 /* Send timestamp is the same or less? can't be ready */ 4330 return (NULL); 4331 } 4332 /* Get our RTT time */ 4333 if (bbr_is_lost(bbr, rsm, cts) && 4334 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4335 (rsm->r_flags & BBR_SACK_PASSED))) { 4336 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4337 rsm->r_flags |= BBR_MARKED_LOST; 4338 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4339 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4340 } 4341 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4342 #ifdef BBR_INVARIANTS 4343 if ((rsm->r_end - rsm->r_start) == 0) 4344 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4345 #endif 4346 return (rsm); 4347 } 4348 return (NULL); 4349 } 4350 4351 /* 4352 * RACK Timer, here we simply do logging and house keeping. 4353 * the normal bbr_output_wtime() function will call the 4354 * appropriate thing to check if we need to do a RACK retransmit. 4355 * We return 1, saying don't proceed with bbr_output_wtime only 4356 * when all timers have been stopped (destroyed PCB?). 4357 */ 4358 static int 4359 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4360 { 4361 /* 4362 * This timer simply provides an internal trigger to send out data. 4363 * The check_recovery_mode call will see if there are needed 4364 * retransmissions, if so we will enter fast-recovery. The output 4365 * call may or may not do the same thing depending on sysctl 4366 * settings. 4367 */ 4368 uint32_t lost; 4369 4370 if (bbr->rc_all_timers_stopped) { 4371 return (1); 4372 } 4373 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4374 /* Its not time yet */ 4375 return (0); 4376 } 4377 BBR_STAT_INC(bbr_to_tot); 4378 lost = bbr->r_ctl.rc_lost; 4379 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4380 bbr_set_state(tp, bbr, 0); 4381 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4382 if (bbr->r_ctl.rc_resend == NULL) { 4383 /* Lets do the check here */ 4384 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4385 } 4386 if (bbr_policer_call_from_rack_to) 4387 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4388 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4389 return (0); 4390 } 4391 4392 static __inline void 4393 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4394 { 4395 int idx; 4396 4397 nrsm->r_start = start; 4398 nrsm->r_end = rsm->r_end; 4399 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4400 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4401 nrsm->r_flags = rsm->r_flags; 4402 /* We don't transfer forward the SYN flag */ 4403 nrsm->r_flags &= ~BBR_HAS_SYN; 4404 /* We move forward the FIN flag, not that this should happen */ 4405 rsm->r_flags &= ~BBR_HAS_FIN; 4406 nrsm->r_dupack = rsm->r_dupack; 4407 nrsm->r_rtr_bytes = 0; 4408 nrsm->r_is_gain = rsm->r_is_gain; 4409 nrsm->r_is_drain = rsm->r_is_drain; 4410 nrsm->r_delivered = rsm->r_delivered; 4411 nrsm->r_ts_valid = rsm->r_ts_valid; 4412 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4413 nrsm->r_del_time = rsm->r_del_time; 4414 nrsm->r_app_limited = rsm->r_app_limited; 4415 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4416 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4417 /* We split a piece the lower section looses any just_ret flag. */ 4418 nrsm->r_bbr_state = rsm->r_bbr_state; 4419 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4420 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4421 } 4422 rsm->r_end = nrsm->r_start; 4423 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4424 idx /= 8; 4425 /* Check if we got too small */ 4426 if ((rsm->r_is_smallmap == 0) && 4427 ((rsm->r_end - rsm->r_start) <= idx)) { 4428 bbr->r_ctl.rc_num_small_maps_alloced++; 4429 rsm->r_is_smallmap = 1; 4430 } 4431 /* Check the new one as well */ 4432 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4433 bbr->r_ctl.rc_num_small_maps_alloced++; 4434 nrsm->r_is_smallmap = 1; 4435 } 4436 } 4437 4438 static int 4439 bbr_sack_mergable(struct bbr_sendmap *at, 4440 uint32_t start, uint32_t end) 4441 { 4442 /* 4443 * Given a sack block defined by 4444 * start and end, and a current position 4445 * at. Return 1 if either side of at 4446 * would show that the block is mergable 4447 * to that side. A block to be mergable 4448 * must have overlap with the start/end 4449 * and be in the SACK'd state. 4450 */ 4451 struct bbr_sendmap *l_rsm; 4452 struct bbr_sendmap *r_rsm; 4453 4454 /* first get the either side blocks */ 4455 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4456 r_rsm = TAILQ_NEXT(at, r_next); 4457 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4458 /* Potentially mergeable */ 4459 if ((l_rsm->r_end == start) || 4460 (SEQ_LT(start, l_rsm->r_end) && 4461 SEQ_GT(end, l_rsm->r_end))) { 4462 /* 4463 * map blk |------| 4464 * sack blk |------| 4465 * <or> 4466 * map blk |------| 4467 * sack blk |------| 4468 */ 4469 return (1); 4470 } 4471 } 4472 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4473 /* Potentially mergeable */ 4474 if ((r_rsm->r_start == end) || 4475 (SEQ_LT(start, r_rsm->r_start) && 4476 SEQ_GT(end, r_rsm->r_start))) { 4477 /* 4478 * map blk |---------| 4479 * sack blk |----| 4480 * <or> 4481 * map blk |---------| 4482 * sack blk |-------| 4483 */ 4484 return (1); 4485 } 4486 } 4487 return (0); 4488 } 4489 4490 static struct bbr_sendmap * 4491 bbr_merge_rsm(struct tcp_bbr *bbr, 4492 struct bbr_sendmap *l_rsm, 4493 struct bbr_sendmap *r_rsm) 4494 { 4495 /* 4496 * We are merging two ack'd RSM's, 4497 * the l_rsm is on the left (lower seq 4498 * values) and the r_rsm is on the right 4499 * (higher seq value). The simplest way 4500 * to merge these is to move the right 4501 * one into the left. I don't think there 4502 * is any reason we need to try to find 4503 * the oldest (or last oldest retransmitted). 4504 */ 4505 l_rsm->r_end = r_rsm->r_end; 4506 if (l_rsm->r_dupack < r_rsm->r_dupack) 4507 l_rsm->r_dupack = r_rsm->r_dupack; 4508 if (r_rsm->r_rtr_bytes) 4509 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4510 if (r_rsm->r_in_tmap) { 4511 /* This really should not happen */ 4512 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4513 } 4514 if (r_rsm->r_app_limited) 4515 l_rsm->r_app_limited = r_rsm->r_app_limited; 4516 /* Now the flags */ 4517 if (r_rsm->r_flags & BBR_HAS_FIN) 4518 l_rsm->r_flags |= BBR_HAS_FIN; 4519 if (r_rsm->r_flags & BBR_TLP) 4520 l_rsm->r_flags |= BBR_TLP; 4521 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4522 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4523 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4524 /* This really should not happen */ 4525 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4526 } 4527 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4528 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4529 /* Transfer the split limit to the map we free */ 4530 r_rsm->r_limit_type = l_rsm->r_limit_type; 4531 l_rsm->r_limit_type = 0; 4532 } 4533 bbr_free(bbr, r_rsm); 4534 return(l_rsm); 4535 } 4536 4537 /* 4538 * TLP Timer, here we simply setup what segment we want to 4539 * have the TLP expire on, the normal bbr_output_wtime() will then 4540 * send it out. 4541 * 4542 * We return 1, saying don't proceed with bbr_output_wtime only 4543 * when all timers have been stopped (destroyed PCB?). 4544 */ 4545 static int 4546 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4547 { 4548 /* 4549 * Tail Loss Probe. 4550 */ 4551 struct bbr_sendmap *rsm = NULL; 4552 struct socket *so; 4553 uint32_t amm; 4554 uint32_t out, avail; 4555 uint32_t maxseg; 4556 int collapsed_win = 0; 4557 4558 if (bbr->rc_all_timers_stopped) { 4559 return (1); 4560 } 4561 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4562 /* Its not time yet */ 4563 return (0); 4564 } 4565 if (ctf_progress_timeout_check(tp, true)) { 4566 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4567 return (-ETIMEDOUT); /* tcp_drop() */ 4568 } 4569 /* Did we somehow get into persists? */ 4570 if (bbr->rc_in_persist) { 4571 return (0); 4572 } 4573 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4574 bbr_set_state(tp, bbr, 0); 4575 BBR_STAT_INC(bbr_tlp_tot); 4576 maxseg = tp->t_maxseg - bbr->rc_last_options; 4577 /* 4578 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4579 * need to figure out how to force a full MSS segment out. 4580 */ 4581 so = tptosocket(tp); 4582 avail = sbavail(&so->so_snd); 4583 out = ctf_outstanding(tp); 4584 if (out > tp->snd_wnd) { 4585 /* special case, we need a retransmission */ 4586 collapsed_win = 1; 4587 goto need_retran; 4588 } 4589 if (avail > out) { 4590 /* New data is available */ 4591 amm = avail - out; 4592 if (amm > maxseg) { 4593 amm = maxseg; 4594 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4595 /* not enough to fill a MTU and no-delay is off */ 4596 goto need_retran; 4597 } 4598 /* Set the send-new override */ 4599 if ((out + amm) <= tp->snd_wnd) { 4600 bbr->rc_tlp_new_data = 1; 4601 } else { 4602 goto need_retran; 4603 } 4604 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4605 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4606 bbr->r_ctl.rc_tlp_send = NULL; 4607 /* cap any slots */ 4608 BBR_STAT_INC(bbr_tlp_newdata); 4609 goto send; 4610 } 4611 need_retran: 4612 /* 4613 * Ok we need to arrange the last un-acked segment to be re-sent, or 4614 * optionally the first un-acked segment. 4615 */ 4616 if (collapsed_win == 0) { 4617 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4618 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4619 rsm = bbr_find_high_nonack(bbr, rsm); 4620 } 4621 if (rsm == NULL) { 4622 goto restore; 4623 } 4624 } else { 4625 /* 4626 * We must find the last segment 4627 * that was acceptable by the client. 4628 */ 4629 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4630 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4631 /* Found one */ 4632 break; 4633 } 4634 } 4635 if (rsm == NULL) { 4636 /* None? if so send the first */ 4637 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4638 if (rsm == NULL) 4639 goto restore; 4640 } 4641 } 4642 if ((rsm->r_end - rsm->r_start) > maxseg) { 4643 /* 4644 * We need to split this the last segment in two. 4645 */ 4646 struct bbr_sendmap *nrsm; 4647 4648 nrsm = bbr_alloc_full_limit(bbr); 4649 if (nrsm == NULL) { 4650 /* 4651 * We can't get memory to split, we can either just 4652 * not split it. Or retransmit the whole piece, lets 4653 * do the large send (BTLP :-) ). 4654 */ 4655 goto go_for_it; 4656 } 4657 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4658 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4659 if (rsm->r_in_tmap) { 4660 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4661 nrsm->r_in_tmap = 1; 4662 } 4663 rsm->r_flags &= (~BBR_HAS_FIN); 4664 rsm = nrsm; 4665 } 4666 go_for_it: 4667 bbr->r_ctl.rc_tlp_send = rsm; 4668 bbr->rc_tlp_rtx_out = 1; 4669 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4670 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4671 tp->t_rxtshift++; 4672 } else { 4673 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4674 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4675 } 4676 send: 4677 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4678 /* 4679 * Can't [re]/transmit a segment we have retransmitted the 4680 * max times. We need the retransmit timer to take over. 4681 */ 4682 restore: 4683 bbr->rc_tlp_new_data = 0; 4684 bbr->r_ctl.rc_tlp_send = NULL; 4685 if (rsm) 4686 rsm->r_flags &= ~BBR_TLP; 4687 BBR_STAT_INC(bbr_tlp_retran_fail); 4688 return (0); 4689 } else if (rsm) { 4690 rsm->r_flags |= BBR_TLP; 4691 } 4692 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4693 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4694 /* 4695 * We have retransmitted to many times for TLP. Switch to 4696 * the regular RTO timer 4697 */ 4698 goto restore; 4699 } 4700 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4701 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4702 return (0); 4703 } 4704 4705 /* 4706 * Delayed ack Timer, here we simply need to setup the 4707 * ACK_NOW flag and remove the DELACK flag. From there 4708 * the output routine will send the ack out. 4709 * 4710 * We only return 1, saying don't proceed, if all timers 4711 * are stopped (destroyed PCB?). 4712 */ 4713 static int 4714 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4715 { 4716 if (bbr->rc_all_timers_stopped) { 4717 return (1); 4718 } 4719 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4720 tp->t_flags &= ~TF_DELACK; 4721 tp->t_flags |= TF_ACKNOW; 4722 KMOD_TCPSTAT_INC(tcps_delack); 4723 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4724 return (0); 4725 } 4726 4727 /* 4728 * Here we send a KEEP-ALIVE like probe to the 4729 * peer, we do not send data. 4730 * 4731 * We only return 1, saying don't proceed, if all timers 4732 * are stopped (destroyed PCB?). 4733 */ 4734 static int 4735 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4736 { 4737 struct tcptemp *t_template; 4738 int32_t retval = 1; 4739 4740 if (bbr->rc_all_timers_stopped) { 4741 return (1); 4742 } 4743 if (bbr->rc_in_persist == 0) 4744 return (0); 4745 4746 /* 4747 * Persistence timer into zero window. Force a byte to be output, if 4748 * possible. 4749 */ 4750 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4751 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4752 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4753 /* 4754 * Have we exceeded the user specified progress time? 4755 */ 4756 if (ctf_progress_timeout_check(tp, true)) { 4757 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4758 return (-ETIMEDOUT); /* tcp_drop() */ 4759 } 4760 /* 4761 * Hack: if the peer is dead/unreachable, we do not time out if the 4762 * window is closed. After a full backoff, drop the connection if 4763 * the idle time (no responses to probes) reaches the maximum 4764 * backoff that we would use if retransmitting. 4765 */ 4766 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 4767 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4768 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4769 KMOD_TCPSTAT_INC(tcps_persistdrop); 4770 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4771 return (-ETIMEDOUT); /* tcp_drop() */ 4772 } 4773 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4774 tp->snd_una == tp->snd_max) { 4775 bbr_exit_persist(tp, bbr, cts, __LINE__); 4776 retval = 0; 4777 goto out; 4778 } 4779 /* 4780 * If the user has closed the socket then drop a persisting 4781 * connection after a much reduced timeout. 4782 */ 4783 if (tp->t_state > TCPS_CLOSE_WAIT && 4784 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4785 KMOD_TCPSTAT_INC(tcps_persistdrop); 4786 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4787 return (-ETIMEDOUT); /* tcp_drop() */ 4788 } 4789 t_template = tcpip_maketemplate(bbr->rc_inp); 4790 if (t_template) { 4791 tcp_respond(tp, t_template->tt_ipgen, 4792 &t_template->tt_t, (struct mbuf *)NULL, 4793 tp->rcv_nxt, tp->snd_una - 1, 0); 4794 /* This sends an ack */ 4795 if (tp->t_flags & TF_DELACK) 4796 tp->t_flags &= ~TF_DELACK; 4797 free(t_template, M_TEMP); 4798 } 4799 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 4800 tp->t_rxtshift++; 4801 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4802 out: 4803 return (retval); 4804 } 4805 4806 /* 4807 * If a keepalive goes off, we had no other timers 4808 * happening. We always return 1 here since this 4809 * routine either drops the connection or sends 4810 * out a segment with respond. 4811 */ 4812 static int 4813 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4814 { 4815 struct tcptemp *t_template; 4816 struct inpcb *inp = tptoinpcb(tp); 4817 4818 if (bbr->rc_all_timers_stopped) { 4819 return (1); 4820 } 4821 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4822 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4823 /* 4824 * Keep-alive timer went off; send something or drop connection if 4825 * idle for too long. 4826 */ 4827 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4828 if (tp->t_state < TCPS_ESTABLISHED) 4829 goto dropit; 4830 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4831 tp->t_state <= TCPS_CLOSING) { 4832 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4833 goto dropit; 4834 /* 4835 * Send a packet designed to force a response if the peer is 4836 * up and reachable: either an ACK if the connection is 4837 * still alive, or an RST if the peer has closed the 4838 * connection due to timeout or reboot. Using sequence 4839 * number tp->snd_una-1 causes the transmitted zero-length 4840 * segment to lie outside the receive window; by the 4841 * protocol spec, this requires the correspondent TCP to 4842 * respond. 4843 */ 4844 KMOD_TCPSTAT_INC(tcps_keepprobe); 4845 t_template = tcpip_maketemplate(inp); 4846 if (t_template) { 4847 tcp_respond(tp, t_template->tt_ipgen, 4848 &t_template->tt_t, (struct mbuf *)NULL, 4849 tp->rcv_nxt, tp->snd_una - 1, 0); 4850 free(t_template, M_TEMP); 4851 } 4852 } 4853 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4854 return (1); 4855 dropit: 4856 KMOD_TCPSTAT_INC(tcps_keepdrops); 4857 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4858 return (-ETIMEDOUT); /* tcp_drop() */ 4859 } 4860 4861 /* 4862 * Retransmit helper function, clear up all the ack 4863 * flags and take care of important book keeping. 4864 */ 4865 static void 4866 bbr_remxt_tmr(struct tcpcb *tp) 4867 { 4868 /* 4869 * The retransmit timer went off, all sack'd blocks must be 4870 * un-acked. 4871 */ 4872 struct bbr_sendmap *rsm, *trsm = NULL; 4873 struct tcp_bbr *bbr; 4874 uint32_t cts, lost; 4875 4876 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4877 cts = tcp_get_usecs(&bbr->rc_tv); 4878 lost = bbr->r_ctl.rc_lost; 4879 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4880 bbr_set_state(tp, bbr, 0); 4881 4882 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4883 if (rsm->r_flags & BBR_ACKED) { 4884 uint32_t old_flags; 4885 4886 rsm->r_dupack = 0; 4887 if (rsm->r_in_tmap == 0) { 4888 /* We must re-add it back to the tlist */ 4889 if (trsm == NULL) { 4890 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4891 } else { 4892 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4893 } 4894 rsm->r_in_tmap = 1; 4895 } 4896 old_flags = rsm->r_flags; 4897 rsm->r_flags |= BBR_RXT_CLEARED; 4898 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4899 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4900 } else { 4901 if ((tp->t_state < TCPS_ESTABLISHED) && 4902 (rsm->r_start == tp->snd_una)) { 4903 /* 4904 * Special case for TCP FO. Where 4905 * we sent more data beyond the snd_max. 4906 * We don't mark that as lost and stop here. 4907 */ 4908 break; 4909 } 4910 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4911 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4912 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4913 } 4914 if (bbr_marks_rxt_sack_passed) { 4915 /* 4916 * With this option, we will rack out 4917 * in 1ms increments the rest of the packets. 4918 */ 4919 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4920 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4921 } else { 4922 /* 4923 * With this option we only mark them lost 4924 * and remove all sack'd markings. We will run 4925 * another RXT or a TLP. This will cause 4926 * us to eventually send more based on what 4927 * ack's come in. 4928 */ 4929 rsm->r_flags |= BBR_MARKED_LOST; 4930 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4931 rsm->r_flags &= ~BBR_SACK_PASSED; 4932 } 4933 } 4934 trsm = rsm; 4935 } 4936 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4937 /* Clear the count (we just un-acked them) */ 4938 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4939 bbr->rc_tlp_new_data = 0; 4940 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4941 /* zap the behindness on a rxt */ 4942 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4943 bbr->r_agg_early_set = 0; 4944 bbr->r_ctl.rc_agg_early = 0; 4945 bbr->rc_tlp_rtx_out = 0; 4946 bbr->r_ctl.rc_sacked = 0; 4947 bbr->r_ctl.rc_sacklast = NULL; 4948 bbr->r_timer_override = 1; 4949 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4950 } 4951 4952 /* 4953 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4954 * we will setup to retransmit the lowest seq number outstanding. 4955 */ 4956 static int 4957 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4958 { 4959 struct inpcb *inp = tptoinpcb(tp); 4960 int32_t rexmt; 4961 int32_t retval = 0; 4962 bool isipv6; 4963 4964 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4965 if (bbr->rc_all_timers_stopped) { 4966 return (1); 4967 } 4968 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4969 (tp->snd_una == tp->snd_max)) { 4970 /* Nothing outstanding .. nothing to do */ 4971 return (0); 4972 } 4973 /* 4974 * Retransmission timer went off. Message has not been acked within 4975 * retransmit interval. Back off to a longer retransmit interval 4976 * and retransmit one segment. 4977 */ 4978 if (ctf_progress_timeout_check(tp, true)) { 4979 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4980 return (-ETIMEDOUT); /* tcp_drop() */ 4981 } 4982 bbr_remxt_tmr(tp); 4983 if ((bbr->r_ctl.rc_resend == NULL) || 4984 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 4985 /* 4986 * If the rwnd collapsed on 4987 * the one we are retransmitting 4988 * it does not count against the 4989 * rxt count. 4990 */ 4991 tp->t_rxtshift++; 4992 } 4993 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { 4994 tp->t_rxtshift = TCP_MAXRXTSHIFT; 4995 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 4996 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 4997 /* XXXGL: previously t_softerror was casted to uint16_t */ 4998 MPASS(tp->t_softerror >= 0); 4999 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 5000 return (retval); /* tcp_drop() */ 5001 } 5002 if (tp->t_state == TCPS_SYN_SENT) { 5003 /* 5004 * If the SYN was retransmitted, indicate CWND to be limited 5005 * to 1 segment in cc_conn_init(). 5006 */ 5007 tp->snd_cwnd = 1; 5008 } else if (tp->t_rxtshift == 1) { 5009 /* 5010 * first retransmit; record ssthresh and cwnd so they can be 5011 * recovered if this turns out to be a "bad" retransmit. A 5012 * retransmit is considered "bad" if an ACK for this segment 5013 * is received within RTT/2 interval; the assumption here is 5014 * that the ACK was already in flight. See "On Estimating 5015 * End-to-End Network Path Properties" by Allman and Paxson 5016 * for more details. 5017 */ 5018 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5019 if (!IN_RECOVERY(tp->t_flags)) { 5020 tp->snd_cwnd_prev = tp->snd_cwnd; 5021 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5022 tp->snd_recover_prev = tp->snd_recover; 5023 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5024 tp->t_flags |= TF_PREVVALID; 5025 } else { 5026 tp->t_flags &= ~TF_PREVVALID; 5027 } 5028 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5029 } else { 5030 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5031 tp->t_flags &= ~TF_PREVVALID; 5032 } 5033 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5034 if ((tp->t_state == TCPS_SYN_SENT) || 5035 (tp->t_state == TCPS_SYN_RECEIVED)) 5036 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5037 else 5038 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5039 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5040 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5041 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5042 /* 5043 * We enter the path for PLMTUD if connection is established or, if 5044 * connection is FIN_WAIT_1 status, reason for the last is that if 5045 * amount of data we send is very small, we could send it in couple 5046 * of packets and process straight to FIN. In that case we won't 5047 * catch ESTABLISHED state. 5048 */ 5049 #ifdef INET6 5050 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false; 5051 #else 5052 isipv6 = false; 5053 #endif 5054 if (((V_tcp_pmtud_blackhole_detect == 1) || 5055 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5056 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5057 ((tp->t_state == TCPS_ESTABLISHED) || 5058 (tp->t_state == TCPS_FIN_WAIT_1))) { 5059 /* 5060 * Idea here is that at each stage of mtu probe (usually, 5061 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5062 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5063 * should take care of that. 5064 */ 5065 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5066 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5067 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5068 tp->t_rxtshift % 2 == 0)) { 5069 /* 5070 * Enter Path MTU Black-hole Detection mechanism: - 5071 * Disable Path MTU Discovery (IP "DF" bit). - 5072 * Reduce MTU to lower value than what we negotiated 5073 * with peer. 5074 */ 5075 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5076 /* 5077 * Record that we may have found a black 5078 * hole. 5079 */ 5080 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5081 /* Keep track of previous MSS. */ 5082 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5083 } 5084 /* 5085 * Reduce the MSS to blackhole value or to the 5086 * default in an attempt to retransmit. 5087 */ 5088 #ifdef INET6 5089 isipv6 = bbr->r_is_v6; 5090 if (isipv6 && 5091 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5092 /* Use the sysctl tuneable blackhole MSS. */ 5093 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5094 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5095 } else if (isipv6) { 5096 /* Use the default MSS. */ 5097 tp->t_maxseg = V_tcp_v6mssdflt; 5098 /* 5099 * Disable Path MTU Discovery when we switch 5100 * to minmss. 5101 */ 5102 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5103 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5104 } 5105 #endif 5106 #if defined(INET6) && defined(INET) 5107 else 5108 #endif 5109 #ifdef INET 5110 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5111 /* Use the sysctl tuneable blackhole MSS. */ 5112 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5113 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5114 } else { 5115 /* Use the default MSS. */ 5116 tp->t_maxseg = V_tcp_mssdflt; 5117 /* 5118 * Disable Path MTU Discovery when we switch 5119 * to minmss. 5120 */ 5121 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5122 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5123 } 5124 #endif 5125 } else { 5126 /* 5127 * If further retransmissions are still unsuccessful 5128 * with a lowered MTU, maybe this isn't a blackhole 5129 * and we restore the previous MSS and blackhole 5130 * detection flags. The limit '6' is determined by 5131 * giving each probe stage (1448, 1188, 524) 2 5132 * chances to recover. 5133 */ 5134 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5135 (tp->t_rxtshift >= 6)) { 5136 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5137 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5138 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5139 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5140 } 5141 } 5142 } 5143 /* 5144 * Disable RFC1323 and SACK if we haven't got any response to our 5145 * third SYN to work-around some broken terminal servers (most of 5146 * which have hopefully been retired) that have bad VJ header 5147 * compression code which trashes TCP segments containing 5148 * unknown-to-them TCP options. 5149 */ 5150 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5151 (tp->t_rxtshift == 3)) 5152 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5153 /* 5154 * If we backed off this far, our srtt estimate is probably bogus. 5155 * Clobber it so we'll take the next rtt measurement as our srtt; 5156 * move the current srtt into rttvar to keep the current retransmit 5157 * times until then. 5158 */ 5159 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5160 #ifdef INET6 5161 if (bbr->r_is_v6) 5162 in6_losing(inp); 5163 else 5164 #endif 5165 in_losing(inp); 5166 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5167 tp->t_srtt = 0; 5168 } 5169 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5170 tp->snd_recover = tp->snd_max; 5171 tp->t_flags |= TF_ACKNOW; 5172 tp->t_rtttime = 0; 5173 5174 return (retval); 5175 } 5176 5177 static int 5178 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5179 { 5180 int32_t ret = 0; 5181 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5182 5183 if (timers == 0) { 5184 return (0); 5185 } 5186 if (tp->t_state == TCPS_LISTEN) { 5187 /* no timers on listen sockets */ 5188 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5189 return (0); 5190 return (1); 5191 } 5192 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5193 uint32_t left; 5194 5195 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5196 ret = -1; 5197 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5198 return (0); 5199 } 5200 if (hpts_calling == 0) { 5201 ret = -2; 5202 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5203 return (0); 5204 } 5205 /* 5206 * Ok our timer went off early and we are not paced false 5207 * alarm, go back to sleep. 5208 */ 5209 left = bbr->r_ctl.rc_timer_exp - cts; 5210 ret = -3; 5211 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5212 tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(left)); 5213 return (1); 5214 } 5215 bbr->rc_tmr_stopped = 0; 5216 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5217 if (timers & PACE_TMR_DELACK) { 5218 ret = bbr_timeout_delack(tp, bbr, cts); 5219 } else if (timers & PACE_TMR_PERSIT) { 5220 ret = bbr_timeout_persist(tp, bbr, cts); 5221 } else if (timers & PACE_TMR_RACK) { 5222 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5223 ret = bbr_timeout_rack(tp, bbr, cts); 5224 } else if (timers & PACE_TMR_TLP) { 5225 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5226 ret = bbr_timeout_tlp(tp, bbr, cts); 5227 } else if (timers & PACE_TMR_RXT) { 5228 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5229 ret = bbr_timeout_rxt(tp, bbr, cts); 5230 } else if (timers & PACE_TMR_KEEP) { 5231 ret = bbr_timeout_keepalive(tp, bbr, cts); 5232 } 5233 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5234 return (ret); 5235 } 5236 5237 static void 5238 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5239 { 5240 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5241 uint8_t hpts_removed = 0; 5242 5243 if (tcp_in_hpts(bbr->rc_tp) && 5244 (bbr->rc_timer_first == 1)) { 5245 /* 5246 * If we are canceling timer's when we have the 5247 * timer ahead of the output being paced. We also 5248 * must remove ourselves from the hpts. 5249 */ 5250 hpts_removed = 1; 5251 tcp_hpts_remove(bbr->rc_tp); 5252 if (bbr->r_ctl.rc_last_delay_val) { 5253 /* Update the last hptsi delay too */ 5254 uint32_t time_since_send; 5255 5256 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5257 time_since_send = cts - bbr->rc_pacer_started; 5258 else 5259 time_since_send = 0; 5260 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5261 /* Cut down our slot time */ 5262 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5263 } else { 5264 bbr->r_ctl.rc_last_delay_val = 0; 5265 } 5266 bbr->rc_pacer_started = cts; 5267 } 5268 } 5269 bbr->rc_timer_first = 0; 5270 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5271 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5272 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5273 } 5274 } 5275 5276 static int 5277 bbr_stopall(struct tcpcb *tp) 5278 { 5279 struct tcp_bbr *bbr; 5280 5281 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5282 bbr->rc_all_timers_stopped = 1; 5283 return (0); 5284 } 5285 5286 static uint32_t 5287 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5288 { 5289 struct bbr_sendmap *rsm; 5290 5291 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5292 if ((rsm == NULL) || (u_rsm == rsm)) 5293 return (cts); 5294 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5295 } 5296 5297 static void 5298 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5299 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5300 { 5301 int32_t idx; 5302 5303 rsm->r_rtr_cnt++; 5304 rsm->r_dupack = 0; 5305 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5306 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5307 rsm->r_flags |= BBR_OVERMAX; 5308 } 5309 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5310 /* Take off the collapsed flag at rxt */ 5311 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5312 } 5313 if (rsm->r_flags & BBR_MARKED_LOST) { 5314 /* We have retransmitted, its no longer lost */ 5315 rsm->r_flags &= ~BBR_MARKED_LOST; 5316 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5317 } 5318 if (rsm->r_flags & BBR_RXT_CLEARED) { 5319 /* 5320 * We hit a RXT timer on it and 5321 * we cleared the "acked" flag. 5322 * We now have it going back into 5323 * flight, we can remove the cleared 5324 * flag and possibly do accounting on 5325 * this piece. 5326 */ 5327 rsm->r_flags &= ~BBR_RXT_CLEARED; 5328 } 5329 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5330 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5331 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5332 } 5333 idx = rsm->r_rtr_cnt - 1; 5334 rsm->r_tim_lastsent[idx] = cts; 5335 rsm->r_pacing_delay = pacing_time; 5336 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5337 rsm->r_ts_valid = bbr->rc_ts_valid; 5338 if (bbr->rc_ts_valid) 5339 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5340 if (bbr->r_ctl.r_app_limited_until) 5341 rsm->r_app_limited = 1; 5342 else 5343 rsm->r_app_limited = 0; 5344 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5345 rsm->r_bbr_state = bbr_state_val(bbr); 5346 else 5347 rsm->r_bbr_state = 8; 5348 if (rsm->r_flags & BBR_ACKED) { 5349 /* Problably MTU discovery messing with us */ 5350 uint32_t old_flags; 5351 5352 old_flags = rsm->r_flags; 5353 rsm->r_flags &= ~BBR_ACKED; 5354 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5355 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5356 if (bbr->r_ctl.rc_sacked == 0) 5357 bbr->r_ctl.rc_sacklast = NULL; 5358 } 5359 if (rsm->r_in_tmap) { 5360 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5361 } 5362 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5363 rsm->r_in_tmap = 1; 5364 if (rsm->r_flags & BBR_SACK_PASSED) { 5365 /* We have retransmitted due to the SACK pass */ 5366 rsm->r_flags &= ~BBR_SACK_PASSED; 5367 rsm->r_flags |= BBR_WAS_SACKPASS; 5368 } 5369 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5370 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5371 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5372 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5373 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5374 rsm->r_is_gain = 1; 5375 rsm->r_is_drain = 0; 5376 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5377 rsm->r_is_drain = 1; 5378 rsm->r_is_gain = 0; 5379 } else { 5380 rsm->r_is_drain = 0; 5381 rsm->r_is_gain = 0; 5382 } 5383 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5384 } 5385 5386 /* 5387 * Returns 0, or the sequence where we stopped 5388 * updating. We also update the lenp to be the amount 5389 * of data left. 5390 */ 5391 5392 static uint32_t 5393 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5394 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5395 { 5396 /* 5397 * We (re-)transmitted starting at rsm->r_start for some length 5398 * (possibly less than r_end. 5399 */ 5400 struct bbr_sendmap *nrsm; 5401 uint32_t c_end; 5402 int32_t len; 5403 5404 len = *lenp; 5405 c_end = rsm->r_start + len; 5406 if (SEQ_GEQ(c_end, rsm->r_end)) { 5407 /* 5408 * We retransmitted the whole piece or more than the whole 5409 * slopping into the next rsm. 5410 */ 5411 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5412 if (c_end == rsm->r_end) { 5413 *lenp = 0; 5414 return (0); 5415 } else { 5416 int32_t act_len; 5417 5418 /* Hangs over the end return whats left */ 5419 act_len = rsm->r_end - rsm->r_start; 5420 *lenp = (len - act_len); 5421 return (rsm->r_end); 5422 } 5423 /* We don't get out of this block. */ 5424 } 5425 /* 5426 * Here we retransmitted less than the whole thing which means we 5427 * have to split this into what was transmitted and what was not. 5428 */ 5429 nrsm = bbr_alloc_full_limit(bbr); 5430 if (nrsm == NULL) { 5431 *lenp = 0; 5432 return (0); 5433 } 5434 /* 5435 * So here we are going to take the original rsm and make it what we 5436 * retransmitted. nrsm will be the tail portion we did not 5437 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5438 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5439 * 1, 6 and the new piece will be 6, 11. 5440 */ 5441 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5442 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5443 nrsm->r_dupack = 0; 5444 if (rsm->r_in_tmap) { 5445 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5446 nrsm->r_in_tmap = 1; 5447 } 5448 rsm->r_flags &= (~BBR_HAS_FIN); 5449 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5450 *lenp = 0; 5451 return (0); 5452 } 5453 5454 static uint64_t 5455 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5456 { 5457 uint64_t bw; 5458 5459 bw = bbr_get_bw(bbr); 5460 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5461 bw /= (uint64_t)BBR_UNIT; 5462 return(bw); 5463 } 5464 5465 static void 5466 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5467 uint64_t act_rate, uint64_t rate_wanted) 5468 { 5469 /* 5470 * We could not get a full gains worth 5471 * of rate. 5472 */ 5473 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5474 /* we can't even get the real rate */ 5475 uint64_t red; 5476 5477 bbr->skip_gain = 1; 5478 bbr->gain_is_limited = 0; 5479 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5480 if (red) 5481 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5482 } else { 5483 /* We can use a lower gain */ 5484 bbr->skip_gain = 0; 5485 bbr->gain_is_limited = 1; 5486 } 5487 } 5488 5489 static void 5490 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5491 { 5492 const struct tcp_hwrate_limit_table *nrte; 5493 int error, rate = -1; 5494 5495 if (bbr->r_ctl.crte == NULL) 5496 return; 5497 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5498 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5499 /* Lost our routes? */ 5500 /* Clear the way for a re-attempt */ 5501 bbr->bbr_attempt_hdwr_pace = 0; 5502 lost_rate: 5503 bbr->gain_is_limited = 0; 5504 bbr->skip_gain = 0; 5505 bbr->bbr_hdrw_pacing = 0; 5506 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5507 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5508 tcp_bbr_tso_size_check(bbr, cts); 5509 return; 5510 } 5511 rate = bbr_get_hardware_rate(bbr); 5512 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5513 bbr->rc_tp, 5514 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5515 rate, 5516 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5517 &error, NULL); 5518 if (nrte == NULL) { 5519 goto lost_rate; 5520 } 5521 if (nrte != bbr->r_ctl.crte) { 5522 bbr->r_ctl.crte = nrte; 5523 if (error == 0) { 5524 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5525 if (bbr->r_ctl.crte->rate < rate) { 5526 /* We have a problem */ 5527 bbr_setup_less_of_rate(bbr, cts, 5528 bbr->r_ctl.crte->rate, rate); 5529 } else { 5530 /* We are good */ 5531 bbr->gain_is_limited = 0; 5532 bbr->skip_gain = 0; 5533 } 5534 } else { 5535 /* A failure should release the tag */ 5536 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5537 bbr->gain_is_limited = 0; 5538 bbr->skip_gain = 0; 5539 bbr->bbr_hdrw_pacing = 0; 5540 } 5541 bbr_type_log_hdwr_pacing(bbr, 5542 bbr->r_ctl.crte->ptbl->rs_ifp, 5543 rate, 5544 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5545 __LINE__, 5546 cts, 5547 error); 5548 } 5549 } 5550 5551 static void 5552 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5553 { 5554 /* 5555 * If we have hardware pacing support 5556 * we need to factor that in for our 5557 * TSO size. 5558 */ 5559 const struct tcp_hwrate_limit_table *rlp; 5560 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5561 5562 if ((bbr->bbr_hdrw_pacing == 0) || 5563 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5564 (bbr->r_ctl.crte == NULL)) 5565 return; 5566 if (bbr->hw_pacing_set == 0) { 5567 /* Not yet by the hdwr pacing count delay */ 5568 return; 5569 } 5570 if (bbr_hdwr_pace_adjust == 0) { 5571 /* No adjustment */ 5572 return; 5573 } 5574 rlp = bbr->r_ctl.crte; 5575 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5576 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5577 else 5578 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5579 /* 5580 * So lets first get the 5581 * time we will take between 5582 * TSO sized sends currently without 5583 * hardware help. 5584 */ 5585 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5586 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5587 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5588 hdwr_delay *= rlp->time_between; 5589 if (cur_delay > hdwr_delay) 5590 delta = cur_delay - hdwr_delay; 5591 else 5592 delta = 0; 5593 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5594 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5595 1); 5596 if (delta && 5597 (delta < (max(rlp->time_between, 5598 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5599 /* 5600 * Now lets divide by the pacing 5601 * time between each segment the 5602 * hardware sends rounding up and 5603 * derive a bytes from that. We multiply 5604 * that by bbr_hdwr_pace_adjust to get 5605 * more bang for our buck. 5606 * 5607 * The goal is to have the software pacer 5608 * waiting no more than an additional 5609 * pacing delay if we can (without the 5610 * compensation i.e. x bbr_hdwr_pace_adjust). 5611 */ 5612 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5613 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5614 seg_sz *= bbr_hdwr_pace_adjust; 5615 if (bbr_hdwr_pace_floor && 5616 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5617 /* Currently hardware paces 5618 * out rs_min_seg segments at a time. 5619 * We need to make sure we always send at least 5620 * a full burst of bbr_hdwr_pace_floor down. 5621 */ 5622 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5623 } 5624 seg_sz *= maxseg; 5625 } else if (delta == 0) { 5626 /* 5627 * The highest pacing rate is 5628 * above our b/w gained. This means 5629 * we probably are going quite fast at 5630 * the hardware highest rate. Lets just multiply 5631 * the calculated TSO size by the 5632 * multiplier factor (its probably 5633 * 4 segments in the default config for 5634 * mlx). 5635 */ 5636 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5637 if (bbr_hdwr_pace_floor && 5638 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5639 /* Currently hardware paces 5640 * out rs_min_seg segments at a time. 5641 * We need to make sure we always send at least 5642 * a full burst of bbr_hdwr_pace_floor down. 5643 */ 5644 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5645 } 5646 } else { 5647 /* 5648 * The pacing time difference is so 5649 * big that the hardware will 5650 * pace out more rapidly then we 5651 * really want and then we 5652 * will have a long delay. Lets just keep 5653 * the same TSO size so its as if 5654 * we were not using hdwr pacing (we 5655 * just gain a bit of spacing from the 5656 * hardware if seg_sz > 1). 5657 */ 5658 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5659 } 5660 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5661 new_tso = seg_sz; 5662 else 5663 new_tso = bbr->r_ctl.rc_pace_max_segs; 5664 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5665 new_tso = PACE_MAX_IP_BYTES - maxseg; 5666 5667 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5668 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5669 bbr->r_ctl.rc_pace_max_segs = new_tso; 5670 } 5671 } 5672 5673 static void 5674 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5675 { 5676 uint64_t bw; 5677 uint32_t old_tso = 0, new_tso; 5678 uint32_t maxseg, bytes; 5679 uint32_t tls_seg=0; 5680 /* 5681 * Google/linux uses the following algorithm to determine 5682 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5683 * 5684 * bytes = bw_in_bytes_per_second / 1000 5685 * bytes = min(bytes, 64k) 5686 * tso_segs = bytes / MSS 5687 * if (bw < 1.2Mbs) 5688 * min_tso_segs = 1 5689 * else 5690 * min_tso_segs = 2 5691 * tso_segs = max(tso_segs, min_tso_segs) 5692 * 5693 * * Note apply a device specific limit (we apply this in the 5694 * tcp_m_copym). 5695 * Note that before the initial measurement is made google bursts out 5696 * a full iwnd just like new-reno/cubic. 5697 * 5698 * We do not use this algorithm. Instead we 5699 * use a two phased approach: 5700 * 5701 * if ( bw <= per-tcb-cross-over) 5702 * goal_tso = calculate how much with this bw we 5703 * can send in goal-time seconds. 5704 * if (goal_tso > mss) 5705 * seg = goal_tso / mss 5706 * tso = seg * mss 5707 * else 5708 * tso = mss 5709 * if (tso > per-tcb-max) 5710 * tso = per-tcb-max 5711 * else if ( bw > 512Mbps) 5712 * tso = max-tso (64k/mss) 5713 * else 5714 * goal_tso = bw / per-tcb-divsor 5715 * seg = (goal_tso + mss-1)/mss 5716 * tso = seg * mss 5717 * 5718 * if (tso < per-tcb-floor) 5719 * tso = per-tcb-floor 5720 * if (tso > per-tcb-utter_max) 5721 * tso = per-tcb-utter_max 5722 * 5723 * Note the default per-tcb-divisor is 1000 (same as google). 5724 * the goal cross over is 30Mbps however. To recreate googles 5725 * algorithm you need to set: 5726 * 5727 * cross-over = 23,168,000 bps 5728 * goal-time = 18000 5729 * per-tcb-max = 2 5730 * per-tcb-divisor = 1000 5731 * per-tcb-floor = 1 5732 * 5733 * This will get you "google bbr" behavior with respect to tso size. 5734 * 5735 * Note we do set anything TSO size until we are past the initial 5736 * window. Before that we gnerally use either a single MSS 5737 * or we use the full IW size (so we burst a IW at a time) 5738 */ 5739 5740 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5741 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5742 } else { 5743 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5744 } 5745 old_tso = bbr->r_ctl.rc_pace_max_segs; 5746 if (bbr->rc_past_init_win == 0) { 5747 /* 5748 * Not enough data has been acknowledged to make a 5749 * judgement. Set up the initial TSO based on if we 5750 * are sending a full IW at once or not. 5751 */ 5752 if (bbr->rc_use_google) 5753 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5754 else if (bbr->bbr_init_win_cheat) 5755 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5756 else 5757 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5758 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5759 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5760 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5761 bbr->r_ctl.rc_pace_max_segs = maxseg; 5762 } 5763 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5764 bbr_adjust_for_hw_pacing(bbr, cts); 5765 return; 5766 } 5767 /** 5768 * Now lets set the TSO goal based on our delivery rate in 5769 * bytes per second. Note we only do this if 5770 * we have acked at least the initial cwnd worth of data. 5771 */ 5772 bw = bbr_get_bw(bbr); 5773 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5774 (bbr->rc_use_google == 0)) { 5775 /* We clamp to one MSS in recovery */ 5776 new_tso = maxseg; 5777 } else if (bbr->rc_use_google) { 5778 int min_tso_segs; 5779 5780 /* Google considers the gain too */ 5781 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5782 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5783 bw /= BBR_UNIT; 5784 } 5785 bytes = bw / 1024; 5786 if (bytes > (64 * 1024)) 5787 bytes = 64 * 1024; 5788 new_tso = bytes / maxseg; 5789 if (bw < ONE_POINT_TWO_MEG) 5790 min_tso_segs = 1; 5791 else 5792 min_tso_segs = 2; 5793 if (new_tso < min_tso_segs) 5794 new_tso = min_tso_segs; 5795 new_tso *= maxseg; 5796 } else if (bbr->rc_no_pacing) { 5797 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5798 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5799 /* 5800 * Calculate the worse case b/w TSO if we are inserting no 5801 * more than a delay_target number of TSO's. 5802 */ 5803 uint32_t tso_len, min_tso; 5804 5805 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5806 if (tso_len > maxseg) { 5807 new_tso = tso_len / maxseg; 5808 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5809 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5810 new_tso *= maxseg; 5811 } else { 5812 /* 5813 * less than a full sized frame yikes.. long rtt or 5814 * low bw? 5815 */ 5816 min_tso = bbr_minseg(bbr); 5817 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5818 new_tso = rounddown(tso_len, min_tso); 5819 else 5820 new_tso = min_tso; 5821 } 5822 } else if (bw > FIVETWELVE_MBPS) { 5823 /* 5824 * This guy is so fast b/w wise that we can TSO as large as 5825 * possible of segments that the NIC will allow. 5826 */ 5827 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5828 } else { 5829 /* 5830 * This formula is based on attempting to send a segment or 5831 * more every bbr_hptsi_per_second. The default is 1000 5832 * which means you are targeting what you can send every 1ms 5833 * based on the peers bw. 5834 * 5835 * If the number drops to say 500, then you are looking more 5836 * at 2ms and you will raise how much we send in a single 5837 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5838 * trade off of course is you will send more at once and 5839 * thus tend to clump up the sends into larger "bursts" 5840 * building a queue. 5841 */ 5842 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5843 new_tso = roundup(bw, (uint64_t)maxseg); 5844 /* 5845 * Gate the floor to match what our lower than 48Mbps 5846 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5847 * becomes the floor for this calculation. 5848 */ 5849 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5850 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5851 } 5852 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5853 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5854 if (new_tso > PACE_MAX_IP_BYTES) 5855 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5856 /* Enforce an utter maximum. */ 5857 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5858 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5859 } 5860 if (old_tso != new_tso) { 5861 /* Only log changes */ 5862 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5863 bbr->r_ctl.rc_pace_max_segs = new_tso; 5864 } 5865 /* We have hardware pacing! */ 5866 bbr_adjust_for_hw_pacing(bbr, cts); 5867 } 5868 5869 static void 5870 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5871 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5872 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5873 struct sockbuf *sb) 5874 { 5875 5876 struct bbr_sendmap *rsm, *nrsm; 5877 register uint32_t snd_max, snd_una; 5878 uint32_t pacing_time; 5879 /* 5880 * Add to the RACK log of packets in flight or retransmitted. If 5881 * there is a TS option we will use the TS echoed, if not we will 5882 * grab a TS. 5883 * 5884 * Retransmissions will increment the count and move the ts to its 5885 * proper place. Note that if options do not include TS's then we 5886 * won't be able to effectively use the ACK for an RTT on a retran. 5887 * 5888 * Notes about r_start and r_end. Lets consider a send starting at 5889 * sequence 1 for 10 bytes. In such an example the r_start would be 5890 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5891 * This means that r_end is actually the first sequence for the next 5892 * slot (11). 5893 * 5894 */ 5895 INP_WLOCK_ASSERT(tptoinpcb(tp)); 5896 if (err) { 5897 /* 5898 * We don't log errors -- we could but snd_max does not 5899 * advance in this case either. 5900 */ 5901 return; 5902 } 5903 if (th_flags & TH_RST) { 5904 /* 5905 * We don't log resets and we return immediately from 5906 * sending 5907 */ 5908 *abandon = 1; 5909 return; 5910 } 5911 snd_una = tp->snd_una; 5912 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5913 /* 5914 * The call to bbr_log_output is made before bumping 5915 * snd_max. This means we can record one extra byte on a SYN 5916 * or FIN if seq_out is adding more on and a FIN is present 5917 * (and we are not resending). 5918 */ 5919 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5920 len++; 5921 if (th_flags & TH_FIN) 5922 len++; 5923 } 5924 if (SEQ_LEQ((seq_out + len), snd_una)) { 5925 /* Are sending an old segment to induce an ack (keep-alive)? */ 5926 return; 5927 } 5928 if (SEQ_LT(seq_out, snd_una)) { 5929 /* huh? should we panic? */ 5930 uint32_t end; 5931 5932 end = seq_out + len; 5933 seq_out = snd_una; 5934 len = end - seq_out; 5935 } 5936 snd_max = tp->snd_max; 5937 if (len == 0) { 5938 /* We don't log zero window probes */ 5939 return; 5940 } 5941 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5942 /* First question is it a retransmission? */ 5943 if (seq_out == snd_max) { 5944 again: 5945 rsm = bbr_alloc(bbr); 5946 if (rsm == NULL) { 5947 return; 5948 } 5949 rsm->r_flags = 0; 5950 if (th_flags & TH_SYN) 5951 rsm->r_flags |= BBR_HAS_SYN; 5952 if (th_flags & TH_FIN) 5953 rsm->r_flags |= BBR_HAS_FIN; 5954 rsm->r_tim_lastsent[0] = cts; 5955 rsm->r_rtr_cnt = 1; 5956 rsm->r_rtr_bytes = 0; 5957 rsm->r_start = seq_out; 5958 rsm->r_end = rsm->r_start + len; 5959 rsm->r_dupack = 0; 5960 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5961 rsm->r_pacing_delay = pacing_time; 5962 rsm->r_ts_valid = bbr->rc_ts_valid; 5963 if (bbr->rc_ts_valid) 5964 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5965 rsm->r_del_time = bbr->r_ctl.rc_del_time; 5966 if (bbr->r_ctl.r_app_limited_until) 5967 rsm->r_app_limited = 1; 5968 else 5969 rsm->r_app_limited = 0; 5970 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5971 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5972 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5973 /* 5974 * Here we must also add in this rsm since snd_max 5975 * is updated after we return from a new send. 5976 */ 5977 rsm->r_flight_at_send += len; 5978 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 5979 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5980 rsm->r_in_tmap = 1; 5981 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5982 rsm->r_bbr_state = bbr_state_val(bbr); 5983 else 5984 rsm->r_bbr_state = 8; 5985 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5986 rsm->r_is_gain = 1; 5987 rsm->r_is_drain = 0; 5988 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5989 rsm->r_is_drain = 1; 5990 rsm->r_is_gain = 0; 5991 } else { 5992 rsm->r_is_drain = 0; 5993 rsm->r_is_gain = 0; 5994 } 5995 return; 5996 } 5997 /* 5998 * If we reach here its a retransmission and we need to find it. 5999 */ 6000 more: 6001 if (hintrsm && (hintrsm->r_start == seq_out)) { 6002 rsm = hintrsm; 6003 hintrsm = NULL; 6004 } else if (bbr->r_ctl.rc_next) { 6005 /* We have a hint from a previous run */ 6006 rsm = bbr->r_ctl.rc_next; 6007 } else { 6008 /* No hints sorry */ 6009 rsm = NULL; 6010 } 6011 if ((rsm) && (rsm->r_start == seq_out)) { 6012 /* 6013 * We used rc_next or hintrsm to retransmit, hopefully the 6014 * likely case. 6015 */ 6016 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6017 if (len == 0) { 6018 return; 6019 } else { 6020 goto more; 6021 } 6022 } 6023 /* Ok it was not the last pointer go through it the hard way. */ 6024 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6025 if (rsm->r_start == seq_out) { 6026 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6027 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6028 if (len == 0) { 6029 return; 6030 } else { 6031 continue; 6032 } 6033 } 6034 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6035 /* Transmitted within this piece */ 6036 /* 6037 * Ok we must split off the front and then let the 6038 * update do the rest 6039 */ 6040 nrsm = bbr_alloc_full_limit(bbr); 6041 if (nrsm == NULL) { 6042 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6043 return; 6044 } 6045 /* 6046 * copy rsm to nrsm and then trim the front of rsm 6047 * to not include this part. 6048 */ 6049 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6050 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6051 if (rsm->r_in_tmap) { 6052 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6053 nrsm->r_in_tmap = 1; 6054 } 6055 rsm->r_flags &= (~BBR_HAS_FIN); 6056 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6057 if (len == 0) { 6058 return; 6059 } 6060 } 6061 } 6062 /* 6063 * Hmm not found in map did they retransmit both old and on into the 6064 * new? 6065 */ 6066 if (seq_out == tp->snd_max) { 6067 goto again; 6068 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6069 #ifdef BBR_INVARIANTS 6070 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6071 seq_out, len, tp->snd_una, tp->snd_max); 6072 printf("Starting Dump of all rack entries\n"); 6073 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6074 printf("rsm:%p start:%u end:%u\n", 6075 rsm, rsm->r_start, rsm->r_end); 6076 } 6077 printf("Dump complete\n"); 6078 panic("seq_out not found rack:%p tp:%p", 6079 bbr, tp); 6080 #endif 6081 } else { 6082 #ifdef BBR_INVARIANTS 6083 /* 6084 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6085 * flag) 6086 */ 6087 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6088 seq_out, len, tp->snd_max, tp); 6089 #endif 6090 } 6091 } 6092 6093 static void 6094 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6095 { 6096 /* 6097 * Collapse timeout back the cum-ack moved. 6098 */ 6099 tp->t_rxtshift = 0; 6100 tp->t_softerror = 0; 6101 } 6102 6103 static void 6104 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6105 { 6106 bbr->rtt_valid = 1; 6107 bbr->r_ctl.cur_rtt = rtt_usecs; 6108 bbr->r_ctl.ts_in = tsin; 6109 if (rsm_send_time) 6110 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6111 } 6112 6113 static void 6114 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6115 { 6116 /** 6117 * We have in our bbr control: 6118 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6119 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6120 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6121 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6122 * 6123 * Now we can calculate the time between the sends by doing: 6124 * 6125 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6126 * 6127 * And the peer's time between receiving them by doing: 6128 * 6129 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6130 * 6131 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6132 * We also may find that we can't use the timestamps if say we see 6133 * that the peer_delta indicates that though we may have taken 10ms to 6134 * pace out the data, it only saw 1ms between the two packets. This would 6135 * indicate that somewhere on the path is a batching entity that is giving 6136 * out time-slices of the actual b/w. This would mean we could not use 6137 * reliably the peers timestamps. 6138 * 6139 * We expect delta > peer_delta initially. Until we figure out the 6140 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6141 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6142 * then its 10ms vs our usec. If the peer is running a usec clock we would 6143 * put a 1 there. If the value is faster then ours, we will disable the 6144 * use of timestamps (though we could revist this later if we find it to be not 6145 * just an isolated one or two flows)). 6146 * 6147 * To detect the batching middle boxes we will come up with our compensation and 6148 * if with it in place, we find the peer is drastically off (by some margin) in 6149 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6150 * 6151 */ 6152 uint64_t delta, peer_delta, delta_up; 6153 6154 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6155 if (delta < bbr_min_usec_delta) { 6156 /* 6157 * Have not seen a min amount of time 6158 * between our send times so we can 6159 * make a determination of the timestamp 6160 * yet. 6161 */ 6162 return; 6163 } 6164 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6165 if (peer_delta < bbr_min_peer_delta) { 6166 /* 6167 * We may have enough in the form of 6168 * our delta but the peers number 6169 * has not changed that much. It could 6170 * be its clock ratio is such that 6171 * we need more data (10ms tick) or 6172 * there may be other compression scenarios 6173 * going on. In any event we need the 6174 * spread to be larger. 6175 */ 6176 return; 6177 } 6178 /* Ok lets first see which way our delta is going */ 6179 if (peer_delta > delta) { 6180 /* Very unlikely, the peer without 6181 * compensation shows that it saw 6182 * the two sends arrive further apart 6183 * then we saw then in micro-seconds. 6184 */ 6185 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6186 /* well it looks like the peer is a micro-second clock. */ 6187 bbr->rc_ts_clock_set = 1; 6188 bbr->r_ctl.bbr_peer_tsratio = 1; 6189 } else { 6190 bbr->rc_ts_cant_be_used = 1; 6191 bbr->rc_ts_clock_set = 1; 6192 } 6193 return; 6194 } 6195 /* Ok we know that the peer_delta is smaller than our send distance */ 6196 bbr->rc_ts_clock_set = 1; 6197 /* First question is it within the percentage that they are using usec time? */ 6198 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6199 if ((peer_delta + delta_up) >= delta) { 6200 /* Its a usec clock */ 6201 bbr->r_ctl.bbr_peer_tsratio = 1; 6202 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6203 return; 6204 } 6205 /* Ok if not usec, what about 10usec (though unlikely)? */ 6206 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6207 if (((peer_delta * 10) + delta_up) >= delta) { 6208 bbr->r_ctl.bbr_peer_tsratio = 10; 6209 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6210 return; 6211 } 6212 /* And what about 100usec (though again unlikely)? */ 6213 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6214 if (((peer_delta * 100) + delta_up) >= delta) { 6215 bbr->r_ctl.bbr_peer_tsratio = 100; 6216 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6217 return; 6218 } 6219 /* And how about 1 msec (the most likely one)? */ 6220 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6221 if (((peer_delta * 1000) + delta_up) >= delta) { 6222 bbr->r_ctl.bbr_peer_tsratio = 1000; 6223 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6224 return; 6225 } 6226 /* Ok if not msec could it be 10 msec? */ 6227 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6228 if (((peer_delta * 10000) + delta_up) >= delta) { 6229 bbr->r_ctl.bbr_peer_tsratio = 10000; 6230 return; 6231 } 6232 /* If we fall down here the clock tick so slowly we can't use it */ 6233 bbr->rc_ts_cant_be_used = 1; 6234 bbr->r_ctl.bbr_peer_tsratio = 0; 6235 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6236 } 6237 6238 /* 6239 * Collect new round-trip time estimate 6240 * and update averages and current timeout. 6241 */ 6242 static void 6243 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6244 { 6245 int32_t delta; 6246 uint32_t rtt, tsin; 6247 int32_t rtt_ticks; 6248 6249 if (bbr->rtt_valid == 0) 6250 /* No valid sample */ 6251 return; 6252 6253 rtt = bbr->r_ctl.cur_rtt; 6254 tsin = bbr->r_ctl.ts_in; 6255 if (bbr->rc_prtt_set_ts) { 6256 /* 6257 * We are to force feed the rttProp filter due 6258 * to an entry into PROBE_RTT. This assures 6259 * that the times are sync'd between when we 6260 * go into PROBE_RTT and the filter expiration. 6261 * 6262 * Google does not use a true filter, so they do 6263 * this implicitly since they only keep one value 6264 * and when they enter probe-rtt they update the 6265 * value to the newest rtt. 6266 */ 6267 uint32_t rtt_prop; 6268 6269 bbr->rc_prtt_set_ts = 0; 6270 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6271 if (rtt > rtt_prop) 6272 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6273 else 6274 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6275 } 6276 #ifdef STATS 6277 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rtt)); 6278 #endif 6279 if (bbr->rc_ack_was_delayed) 6280 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6281 6282 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6283 bbr->r_ctl.rc_lowest_rtt = rtt; 6284 bbr_log_rtt_sample(bbr, rtt, tsin); 6285 if (bbr->r_init_rtt) { 6286 /* 6287 * The initial rtt is not-trusted, nuke it and lets get 6288 * our first valid measurement in. 6289 */ 6290 bbr->r_init_rtt = 0; 6291 tp->t_srtt = 0; 6292 } 6293 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6294 /* 6295 * So we have not yet figured out 6296 * what the peers TSTMP value is 6297 * in (most likely ms). We need a 6298 * series of cum-ack's to determine 6299 * this reliably. 6300 */ 6301 if (bbr->rc_ack_is_cumack) { 6302 if (bbr->rc_ts_data_set) { 6303 /* Lets attempt to determine the timestamp granularity. */ 6304 bbr_make_timestamp_determination(bbr); 6305 } else { 6306 bbr->rc_ts_data_set = 1; 6307 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6308 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6309 } 6310 } else { 6311 /* 6312 * We have to have consecutive acks 6313 * reset any "filled" state to none. 6314 */ 6315 bbr->rc_ts_data_set = 0; 6316 } 6317 } 6318 /* Round it up */ 6319 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6320 if (rtt_ticks == 0) 6321 rtt_ticks = 1; 6322 if (tp->t_srtt != 0) { 6323 /* 6324 * srtt is stored as fixed point with 5 bits after the 6325 * binary point (i.e., scaled by 8). The following magic is 6326 * equivalent to the smoothing algorithm in rfc793 with an 6327 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6328 * Adjust rtt to origin 0. 6329 */ 6330 6331 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6332 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6333 6334 tp->t_srtt += delta; 6335 if (tp->t_srtt <= 0) 6336 tp->t_srtt = 1; 6337 6338 /* 6339 * We accumulate a smoothed rtt variance (actually, a 6340 * smoothed mean difference), then set the retransmit timer 6341 * to smoothed rtt + 4 times the smoothed variance. rttvar 6342 * is stored as fixed point with 4 bits after the binary 6343 * point (scaled by 16). The following is equivalent to 6344 * rfc793 smoothing with an alpha of .75 (rttvar = 6345 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6346 * wired-in beta. 6347 */ 6348 if (delta < 0) 6349 delta = -delta; 6350 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6351 tp->t_rttvar += delta; 6352 if (tp->t_rttvar <= 0) 6353 tp->t_rttvar = 1; 6354 } else { 6355 /* 6356 * No rtt measurement yet - use the unsmoothed rtt. Set the 6357 * variance to half the rtt (so our first retransmit happens 6358 * at 3*rtt). 6359 */ 6360 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6361 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6362 } 6363 KMOD_TCPSTAT_INC(tcps_rttupdated); 6364 if (tp->t_rttupdated < UCHAR_MAX) 6365 tp->t_rttupdated++; 6366 #ifdef STATS 6367 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6368 #endif 6369 /* 6370 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6371 * way we do the smoothing, srtt and rttvar will each average +1/2 6372 * tick of bias. When we compute the retransmit timer, we want 1/2 6373 * tick of rounding and 1 extra tick because of +-1/2 tick 6374 * uncertainty in the firing of the timer. The bias will give us 6375 * exactly the 1.5 tick we need. But, because the bias is 6376 * statistical, we have to test that we don't drop below the minimum 6377 * feasible timer (which is 2 ticks). 6378 */ 6379 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6380 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6381 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6382 6383 /* 6384 * We received an ack for a packet that wasn't retransmitted; it is 6385 * probably safe to discard any error indications we've received 6386 * recently. This isn't quite right, but close enough for now (a 6387 * route might have failed after we sent a segment, and the return 6388 * path might not be symmetrical). 6389 */ 6390 tp->t_softerror = 0; 6391 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6392 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6393 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6394 } 6395 6396 static void 6397 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6398 { 6399 bbr->r_ctl.rc_rtt_shrinks = cts; 6400 if (bbr_can_force_probertt && 6401 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6402 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6403 /* 6404 * We should enter probe-rtt its been too long 6405 * since we have been there. 6406 */ 6407 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6408 } else 6409 bbr_check_probe_rtt_limits(bbr, cts); 6410 } 6411 6412 static void 6413 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6414 { 6415 uint64_t orig_bw; 6416 6417 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6418 /* We never apply a zero measurement */ 6419 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6420 0, 0, 0, 0, 0, 0); 6421 return; 6422 } 6423 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6424 bbr->r_ctl.r_measurement_count++; 6425 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6426 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6427 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6428 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6429 0, 0, 0, 0, 0, 0); 6430 if (orig_bw && 6431 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6432 if (bbr->bbr_hdrw_pacing) { 6433 /* 6434 * Apply a new rate to the hardware 6435 * possibly. 6436 */ 6437 bbr_update_hardware_pacing_rate(bbr, cts); 6438 } 6439 bbr_set_state_target(bbr, __LINE__); 6440 tcp_bbr_tso_size_check(bbr, cts); 6441 if (bbr->r_recovery_bw) { 6442 bbr_setup_red_bw(bbr, cts); 6443 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6444 } 6445 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6446 tcp_bbr_tso_size_check(bbr, cts); 6447 } 6448 6449 static void 6450 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6451 { 6452 if (bbr->rc_in_persist == 0) { 6453 /* We log only when not in persist */ 6454 /* Translate to a Bytes Per Second */ 6455 uint64_t tim, bw, ts_diff, ts_bw; 6456 uint32_t delivered; 6457 6458 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6459 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6460 else 6461 tim = 1; 6462 /* 6463 * Now that we have processed the tim (skipping the sample 6464 * or possibly updating the time, go ahead and 6465 * calculate the cdr. 6466 */ 6467 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6468 bw = (uint64_t)delivered; 6469 bw *= (uint64_t)USECS_IN_SECOND; 6470 bw /= tim; 6471 if (bw == 0) { 6472 /* We must have a calculatable amount */ 6473 return; 6474 } 6475 /* 6476 * If we are using this b/w shove it in now so we 6477 * can see in the trace viewer if it gets over-ridden. 6478 */ 6479 if (rsm->r_ts_valid && 6480 bbr->rc_ts_valid && 6481 bbr->rc_ts_clock_set && 6482 (bbr->rc_ts_cant_be_used == 0) && 6483 bbr->rc_use_ts_limit) { 6484 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6485 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6486 if ((delivered == 0) || 6487 (rtt < 1000)) { 6488 /* Can't use the ts */ 6489 bbr_log_type_bbrupd(bbr, 61, cts, 6490 ts_diff, 6491 bbr->r_ctl.last_inbound_ts, 6492 rsm->r_del_ack_ts, 0, 6493 0, 0, 0, delivered); 6494 } else { 6495 ts_bw = (uint64_t)delivered; 6496 ts_bw *= (uint64_t)USECS_IN_SECOND; 6497 ts_bw /= ts_diff; 6498 bbr_log_type_bbrupd(bbr, 62, cts, 6499 (ts_bw >> 32), 6500 (ts_bw & 0xffffffff), 0, 0, 6501 0, 0, ts_diff, delivered); 6502 if ((bbr->ts_can_raise) && 6503 (ts_bw > bw)) { 6504 bbr_log_type_bbrupd(bbr, 8, cts, 6505 delivered, 6506 ts_diff, 6507 (bw >> 32), 6508 (bw & 0x00000000ffffffff), 6509 0, 0, 0, 0); 6510 bw = ts_bw; 6511 } else if (ts_bw && (ts_bw < bw)) { 6512 bbr_log_type_bbrupd(bbr, 7, cts, 6513 delivered, 6514 ts_diff, 6515 (bw >> 32), 6516 (bw & 0x00000000ffffffff), 6517 0, 0, 0, 0); 6518 bw = ts_bw; 6519 } 6520 } 6521 } 6522 if (rsm->r_first_sent_time && 6523 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6524 uint64_t sbw, sti; 6525 /* 6526 * We use what was in flight at the time of our 6527 * send and the size of this send to figure 6528 * out what we have been sending at (amount). 6529 * For the time we take from the time of 6530 * the send of the first send outstanding 6531 * until this send plus this sends pacing 6532 * time. This gives us a good calculation 6533 * as to the rate we have been sending at. 6534 */ 6535 6536 sbw = (uint64_t)(rsm->r_flight_at_send); 6537 sbw *= (uint64_t)USECS_IN_SECOND; 6538 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6539 sti += rsm->r_pacing_delay; 6540 sbw /= sti; 6541 if (sbw < bw) { 6542 bbr_log_type_bbrupd(bbr, 6, cts, 6543 delivered, 6544 (uint32_t)sti, 6545 (bw >> 32), 6546 (uint32_t)bw, 6547 rsm->r_first_sent_time, 0, (sbw >> 32), 6548 (uint32_t)sbw); 6549 bw = sbw; 6550 } 6551 } 6552 /* Use the google algorithm for b/w measurements */ 6553 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6554 if ((rsm->r_app_limited == 0) || 6555 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6556 tcp_bbr_commit_bw(bbr, cts); 6557 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6558 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6559 } 6560 } 6561 } 6562 6563 static void 6564 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6565 { 6566 if (bbr->rc_in_persist == 0) { 6567 /* We log only when not in persist */ 6568 /* Translate to a Bytes Per Second */ 6569 uint64_t tim, bw; 6570 uint32_t delivered; 6571 int no_apply = 0; 6572 6573 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6574 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6575 else 6576 tim = 1; 6577 /* 6578 * Now that we have processed the tim (skipping the sample 6579 * or possibly updating the time, go ahead and 6580 * calculate the cdr. 6581 */ 6582 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6583 bw = (uint64_t)delivered; 6584 bw *= (uint64_t)USECS_IN_SECOND; 6585 bw /= tim; 6586 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6587 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6588 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6589 6590 no_apply = 1; 6591 } 6592 /* 6593 * If we are using this b/w shove it in now so we 6594 * can see in the trace viewer if it gets over-ridden. 6595 */ 6596 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6597 /* Gate by the sending rate */ 6598 if (rsm->r_first_sent_time && 6599 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6600 uint64_t sbw, sti; 6601 /* 6602 * We use what was in flight at the time of our 6603 * send and the size of this send to figure 6604 * out what we have been sending at (amount). 6605 * For the time we take from the time of 6606 * the send of the first send outstanding 6607 * until this send plus this sends pacing 6608 * time. This gives us a good calculation 6609 * as to the rate we have been sending at. 6610 */ 6611 6612 sbw = (uint64_t)(rsm->r_flight_at_send); 6613 sbw *= (uint64_t)USECS_IN_SECOND; 6614 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6615 sti += rsm->r_pacing_delay; 6616 sbw /= sti; 6617 if (sbw < bw) { 6618 bbr_log_type_bbrupd(bbr, 6, cts, 6619 delivered, 6620 (uint32_t)sti, 6621 (bw >> 32), 6622 (uint32_t)bw, 6623 rsm->r_first_sent_time, 0, (sbw >> 32), 6624 (uint32_t)sbw); 6625 bw = sbw; 6626 } 6627 if ((sti > tim) && 6628 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6629 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6630 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6631 no_apply = 1; 6632 } else 6633 no_apply = 0; 6634 } 6635 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6636 if ((no_apply == 0) && 6637 ((rsm->r_app_limited == 0) || 6638 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6639 tcp_bbr_commit_bw(bbr, cts); 6640 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6641 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6642 } 6643 } 6644 } 6645 6646 static void 6647 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6648 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6649 { 6650 uint64_t old_rttprop; 6651 6652 /* Update our delivery time and amount */ 6653 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6654 bbr->r_ctl.rc_del_time = cts; 6655 if (rtt == 0) { 6656 /* 6657 * 0 means its a retransmit, for now we don't use these for 6658 * the rest of BBR. 6659 */ 6660 return; 6661 } 6662 if ((bbr->rc_use_google == 0) && 6663 (match != BBR_RTT_BY_EXACTMATCH) && 6664 (match != BBR_RTT_BY_TIMESTAMP)){ 6665 /* 6666 * We get a lot of rtt updates, lets not pay attention to 6667 * any that are not an exact match. That way we don't have 6668 * to worry about timestamps and the whole nonsense of 6669 * unsure if its a retransmission etc (if we ever had the 6670 * timestamp fixed to always have the last thing sent this 6671 * would not be a issue). 6672 */ 6673 return; 6674 } 6675 if ((bbr_no_retran && bbr->rc_use_google) && 6676 (match != BBR_RTT_BY_EXACTMATCH) && 6677 (match != BBR_RTT_BY_TIMESTAMP)){ 6678 /* 6679 * We only do measurements in google mode 6680 * with bbr_no_retran on for sure things. 6681 */ 6682 return; 6683 } 6684 /* Only update srtt if we know by exact match */ 6685 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6686 if (ack_type == BBR_CUM_ACKED) 6687 bbr->rc_ack_is_cumack = 1; 6688 else 6689 bbr->rc_ack_is_cumack = 0; 6690 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6691 /* 6692 * Note the following code differs to the original 6693 * BBR spec. It calls for <= not <. However after a 6694 * long discussion in email with Neal, he acknowledged 6695 * that it should be < than so that we will have flows 6696 * going into probe-rtt (we were seeing cases where that 6697 * did not happen and caused ugly things to occur). We 6698 * have added this agreed upon fix to our code base. 6699 */ 6700 if (rtt < old_rttprop) { 6701 /* Update when we last saw a rtt drop */ 6702 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6703 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6704 } 6705 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6706 match, rsm->r_start, rsm->r_flags); 6707 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6708 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6709 /* 6710 * The RTT-prop moved, reset the target (may be a 6711 * nop for some states). 6712 */ 6713 bbr_set_state_target(bbr, __LINE__); 6714 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6715 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6716 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6717 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6718 /* It went up */ 6719 bbr_check_probe_rtt_limits(bbr, cts); 6720 } 6721 if ((bbr->rc_use_google == 0) && 6722 (match == BBR_RTT_BY_TIMESTAMP)) { 6723 /* 6724 * We don't do b/w update with 6725 * these since they are not really 6726 * reliable. 6727 */ 6728 return; 6729 } 6730 if (bbr->r_ctl.r_app_limited_until && 6731 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6732 /* We are no longer app-limited */ 6733 bbr->r_ctl.r_app_limited_until = 0; 6734 } 6735 if (bbr->rc_use_google) { 6736 bbr_google_measurement(bbr, rsm, rtt, cts); 6737 } else { 6738 bbr_nf_measurement(bbr, rsm, rtt, cts); 6739 } 6740 } 6741 6742 /* 6743 * Convert a timestamp that the main stack 6744 * uses (milliseconds) into one that bbr uses 6745 * (microseconds). Return that converted timestamp. 6746 */ 6747 static uint32_t 6748 bbr_ts_convert(uint32_t cts) { 6749 uint32_t sec, msec; 6750 6751 sec = cts / MS_IN_USEC; 6752 msec = cts - (MS_IN_USEC * sec); 6753 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6754 } 6755 6756 /* 6757 * Return 0 if we did not update the RTT time, return 6758 * 1 if we did. 6759 */ 6760 static int 6761 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6762 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6763 { 6764 int32_t i; 6765 uint32_t t, uts = 0; 6766 6767 if ((rsm->r_flags & BBR_ACKED) || 6768 (rsm->r_flags & BBR_WAS_RENEGED) || 6769 (rsm->r_flags & BBR_RXT_CLEARED)) { 6770 /* Already done */ 6771 return (0); 6772 } 6773 if (rsm->r_rtt_not_allowed) { 6774 /* Not allowed */ 6775 return (0); 6776 } 6777 if (rsm->r_rtr_cnt == 1) { 6778 /* 6779 * Only one transmit. Hopefully the normal case. 6780 */ 6781 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6782 t = cts - rsm->r_tim_lastsent[0]; 6783 else 6784 t = 1; 6785 if ((int)t <= 0) 6786 t = 1; 6787 bbr->r_ctl.rc_last_rtt = t; 6788 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6789 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6790 return (1); 6791 } 6792 /* Convert to usecs */ 6793 if ((bbr_can_use_ts_for_rtt == 1) && 6794 (bbr->rc_use_google == 1) && 6795 (ack_type == BBR_CUM_ACKED) && 6796 (to->to_flags & TOF_TS) && 6797 (to->to_tsecr != 0)) { 6798 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6799 if (t < 1) 6800 t = 1; 6801 t *= MS_IN_USEC; 6802 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6803 BBR_RTT_BY_TIMESTAMP, 6804 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6805 ack_type, to); 6806 return (1); 6807 } 6808 uts = bbr_ts_convert(to->to_tsecr); 6809 if ((to->to_flags & TOF_TS) && 6810 (to->to_tsecr != 0) && 6811 (ack_type == BBR_CUM_ACKED) && 6812 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6813 /* 6814 * Now which timestamp does it match? In this block the ACK 6815 * may be coming from a previous transmission. 6816 */ 6817 uint32_t fudge; 6818 6819 fudge = BBR_TIMER_FUDGE; 6820 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6821 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6822 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6823 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6824 t = cts - rsm->r_tim_lastsent[i]; 6825 else 6826 t = 1; 6827 if ((int)t <= 0) 6828 t = 1; 6829 bbr->r_ctl.rc_last_rtt = t; 6830 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6831 rsm->r_tim_lastsent[i], ack_type, to); 6832 if ((i + 1) < rsm->r_rtr_cnt) { 6833 /* Likely */ 6834 return (0); 6835 } else if (rsm->r_flags & BBR_TLP) { 6836 bbr->rc_tlp_rtx_out = 0; 6837 } 6838 return (1); 6839 } 6840 } 6841 /* Fall through if we can't find a matching timestamp */ 6842 } 6843 /* 6844 * Ok its a SACK block that we retransmitted. or a windows 6845 * machine without timestamps. We can tell nothing from the 6846 * time-stamp since its not there or the time the peer last 6847 * recieved a segment that moved forward its cum-ack point. 6848 * 6849 * Lets look at the last retransmit and see what we can tell 6850 * (with BBR for space we only keep 2 note we have to keep 6851 * at least 2 so the map can not be condensed more). 6852 */ 6853 i = rsm->r_rtr_cnt - 1; 6854 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6855 t = cts - rsm->r_tim_lastsent[i]; 6856 else 6857 goto not_sure; 6858 if (t < bbr->r_ctl.rc_lowest_rtt) { 6859 /* 6860 * We retransmitted and the ack came back in less 6861 * than the smallest rtt we have observed in the 6862 * windowed rtt. We most likey did an improper 6863 * retransmit as outlined in 4.2 Step 3 point 2 in 6864 * the rack-draft. 6865 * 6866 * Use the prior transmission to update all the 6867 * information as long as there is only one prior 6868 * transmission. 6869 */ 6870 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6871 #ifdef BBR_INVARIANTS 6872 if (rsm->r_rtr_cnt == 1) 6873 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6874 #endif 6875 i = rsm->r_rtr_cnt - 2; 6876 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6877 t = cts - rsm->r_tim_lastsent[i]; 6878 else 6879 t = 1; 6880 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6881 rsm->r_tim_lastsent[i], ack_type, to); 6882 return (0); 6883 } else { 6884 /* 6885 * Too many prior transmissions, just 6886 * updated BBR delivered 6887 */ 6888 not_sure: 6889 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6890 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6891 } 6892 } else { 6893 /* 6894 * We retransmitted it and the retransmit did the 6895 * job. 6896 */ 6897 if (rsm->r_flags & BBR_TLP) 6898 bbr->rc_tlp_rtx_out = 0; 6899 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6900 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6901 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6902 else 6903 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6904 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6905 return (1); 6906 } 6907 return (0); 6908 } 6909 6910 /* 6911 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6912 */ 6913 static void 6914 bbr_log_sack_passed(struct tcpcb *tp, 6915 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6916 { 6917 struct bbr_sendmap *nrsm; 6918 6919 nrsm = rsm; 6920 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6921 bbr_head, r_tnext) { 6922 if (nrsm == rsm) { 6923 /* Skip original segment he is acked */ 6924 continue; 6925 } 6926 if (nrsm->r_flags & BBR_ACKED) { 6927 /* Skip ack'd segments */ 6928 continue; 6929 } 6930 if (nrsm->r_flags & BBR_SACK_PASSED) { 6931 /* 6932 * We found one that is already marked 6933 * passed, we have been here before and 6934 * so all others below this are marked. 6935 */ 6936 break; 6937 } 6938 BBR_STAT_INC(bbr_sack_passed); 6939 nrsm->r_flags |= BBR_SACK_PASSED; 6940 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6941 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6942 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6943 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6944 nrsm->r_flags |= BBR_MARKED_LOST; 6945 } 6946 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6947 } 6948 } 6949 6950 /* 6951 * Returns the number of bytes that were 6952 * newly ack'd by sack blocks. 6953 */ 6954 static uint32_t 6955 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6956 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6957 { 6958 int32_t times = 0; 6959 uint32_t start, end, changed = 0; 6960 struct bbr_sendmap *rsm, *nrsm; 6961 int32_t used_ref = 1; 6962 uint8_t went_back = 0, went_fwd = 0; 6963 6964 start = sack->start; 6965 end = sack->end; 6966 rsm = *prsm; 6967 if (rsm == NULL) 6968 used_ref = 0; 6969 6970 /* Do we locate the block behind where we last were? */ 6971 if (rsm && SEQ_LT(start, rsm->r_start)) { 6972 went_back = 1; 6973 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 6974 if (SEQ_GEQ(start, rsm->r_start) && 6975 SEQ_LT(start, rsm->r_end)) { 6976 goto do_rest_ofb; 6977 } 6978 } 6979 } 6980 start_at_beginning: 6981 went_fwd = 1; 6982 /* 6983 * Ok lets locate the block where this guy is fwd from rsm (if its 6984 * set) 6985 */ 6986 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 6987 if (SEQ_GEQ(start, rsm->r_start) && 6988 SEQ_LT(start, rsm->r_end)) { 6989 break; 6990 } 6991 } 6992 do_rest_ofb: 6993 if (rsm == NULL) { 6994 /* 6995 * This happens when we get duplicate sack blocks with the 6996 * same end. For example SACK 4: 100 SACK 3: 100 The sort 6997 * will not change there location so we would just start at 6998 * the end of the first one and get lost. 6999 */ 7000 if (tp->t_flags & TF_SENTFIN) { 7001 /* 7002 * Check to see if we have not logged the FIN that 7003 * went out. 7004 */ 7005 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7006 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7007 /* 7008 * Ok we did not get the FIN logged. 7009 */ 7010 nrsm->r_end++; 7011 rsm = nrsm; 7012 goto do_rest_ofb; 7013 } 7014 } 7015 if (times == 1) { 7016 #ifdef BBR_INVARIANTS 7017 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7018 tp, bbr, sack, to, prsm); 7019 #else 7020 goto out; 7021 #endif 7022 } 7023 times++; 7024 BBR_STAT_INC(bbr_sack_proc_restart); 7025 rsm = NULL; 7026 goto start_at_beginning; 7027 } 7028 /* Ok we have an ACK for some piece of rsm */ 7029 if (rsm->r_start != start) { 7030 /* 7031 * Need to split this in two pieces the before and after. 7032 */ 7033 if (bbr_sack_mergable(rsm, start, end)) 7034 nrsm = bbr_alloc_full_limit(bbr); 7035 else 7036 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7037 if (nrsm == NULL) { 7038 /* We could not allocate ignore the sack */ 7039 struct sackblk blk; 7040 7041 blk.start = start; 7042 blk.end = end; 7043 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7044 goto out; 7045 } 7046 bbr_clone_rsm(bbr, nrsm, rsm, start); 7047 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7048 if (rsm->r_in_tmap) { 7049 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7050 nrsm->r_in_tmap = 1; 7051 } 7052 rsm->r_flags &= (~BBR_HAS_FIN); 7053 rsm = nrsm; 7054 } 7055 if (SEQ_GEQ(end, rsm->r_end)) { 7056 /* 7057 * The end of this block is either beyond this guy or right 7058 * at this guy. 7059 */ 7060 if ((rsm->r_flags & BBR_ACKED) == 0) { 7061 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7062 changed += (rsm->r_end - rsm->r_start); 7063 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7064 bbr_log_sack_passed(tp, bbr, rsm); 7065 if (rsm->r_flags & BBR_MARKED_LOST) { 7066 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7067 } 7068 /* Is Reordering occuring? */ 7069 if (rsm->r_flags & BBR_SACK_PASSED) { 7070 BBR_STAT_INC(bbr_reorder_seen); 7071 bbr->r_ctl.rc_reorder_ts = cts; 7072 if (rsm->r_flags & BBR_MARKED_LOST) { 7073 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7074 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7075 /* LT sampling also needs adjustment */ 7076 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7077 } 7078 } 7079 rsm->r_flags |= BBR_ACKED; 7080 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7081 if (rsm->r_in_tmap) { 7082 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7083 rsm->r_in_tmap = 0; 7084 } 7085 } 7086 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7087 if (end == rsm->r_end) { 7088 /* This block only - done */ 7089 goto out; 7090 } 7091 /* There is more not coverend by this rsm move on */ 7092 start = rsm->r_end; 7093 nrsm = TAILQ_NEXT(rsm, r_next); 7094 rsm = nrsm; 7095 times = 0; 7096 goto do_rest_ofb; 7097 } 7098 if (rsm->r_flags & BBR_ACKED) { 7099 /* Been here done that */ 7100 goto out; 7101 } 7102 /* Ok we need to split off this one at the tail */ 7103 if (bbr_sack_mergable(rsm, start, end)) 7104 nrsm = bbr_alloc_full_limit(bbr); 7105 else 7106 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7107 if (nrsm == NULL) { 7108 /* failed XXXrrs what can we do but loose the sack info? */ 7109 struct sackblk blk; 7110 7111 blk.start = start; 7112 blk.end = end; 7113 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7114 goto out; 7115 } 7116 /* Clone it */ 7117 bbr_clone_rsm(bbr, nrsm, rsm, end); 7118 /* The sack block does not cover this guy fully */ 7119 rsm->r_flags &= (~BBR_HAS_FIN); 7120 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7121 if (rsm->r_in_tmap) { 7122 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7123 nrsm->r_in_tmap = 1; 7124 } 7125 nrsm->r_dupack = 0; 7126 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7127 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7128 changed += (rsm->r_end - rsm->r_start); 7129 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7130 bbr_log_sack_passed(tp, bbr, rsm); 7131 /* Is Reordering occuring? */ 7132 if (rsm->r_flags & BBR_MARKED_LOST) { 7133 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7134 } 7135 if (rsm->r_flags & BBR_SACK_PASSED) { 7136 BBR_STAT_INC(bbr_reorder_seen); 7137 bbr->r_ctl.rc_reorder_ts = cts; 7138 if (rsm->r_flags & BBR_MARKED_LOST) { 7139 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7140 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7141 /* LT sampling also needs adjustment */ 7142 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7143 } 7144 } 7145 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7146 rsm->r_flags |= BBR_ACKED; 7147 if (rsm->r_in_tmap) { 7148 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7149 rsm->r_in_tmap = 0; 7150 } 7151 out: 7152 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7153 /* 7154 * Now can we merge this newly acked 7155 * block with either the previous or 7156 * next block? 7157 */ 7158 nrsm = TAILQ_NEXT(rsm, r_next); 7159 if (nrsm && 7160 (nrsm->r_flags & BBR_ACKED)) { 7161 /* yep this and next can be merged */ 7162 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7163 } 7164 /* Now what about the previous? */ 7165 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7166 if (nrsm && 7167 (nrsm->r_flags & BBR_ACKED)) { 7168 /* yep the previous and this can be merged */ 7169 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7170 } 7171 } 7172 if (used_ref == 0) { 7173 BBR_STAT_INC(bbr_sack_proc_all); 7174 } else { 7175 BBR_STAT_INC(bbr_sack_proc_short); 7176 } 7177 if (went_fwd && went_back) { 7178 BBR_STAT_INC(bbr_sack_search_both); 7179 } else if (went_fwd) { 7180 BBR_STAT_INC(bbr_sack_search_fwd); 7181 } else if (went_back) { 7182 BBR_STAT_INC(bbr_sack_search_back); 7183 } 7184 /* Save off where the next seq is */ 7185 if (rsm) 7186 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7187 else 7188 bbr->r_ctl.rc_sacklast = NULL; 7189 *prsm = rsm; 7190 return (changed); 7191 } 7192 7193 static void inline 7194 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7195 { 7196 struct bbr_sendmap *tmap; 7197 7198 BBR_STAT_INC(bbr_reneges_seen); 7199 tmap = NULL; 7200 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7201 /* Its no longer sacked, mark it so */ 7202 uint32_t oflags; 7203 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7204 #ifdef BBR_INVARIANTS 7205 if (rsm->r_in_tmap) { 7206 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7207 bbr, rsm, rsm->r_flags); 7208 } 7209 #endif 7210 oflags = rsm->r_flags; 7211 if (rsm->r_flags & BBR_MARKED_LOST) { 7212 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7213 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7214 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7215 /* LT sampling also needs adjustment */ 7216 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7217 } 7218 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7219 rsm->r_flags |= BBR_WAS_RENEGED; 7220 rsm->r_flags |= BBR_RXT_CLEARED; 7221 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7222 /* Rebuild it into our tmap */ 7223 if (tmap == NULL) { 7224 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7225 tmap = rsm; 7226 } else { 7227 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7228 tmap = rsm; 7229 } 7230 tmap->r_in_tmap = 1; 7231 /* 7232 * XXXrrs Delivered? Should we do anything here? 7233 * 7234 * Of course we don't on a rxt timeout so maybe its ok that 7235 * we don't? 7236 * 7237 * For now lets not. 7238 */ 7239 rsm = TAILQ_NEXT(rsm, r_next); 7240 } 7241 /* 7242 * Now lets possibly clear the sack filter so we start recognizing 7243 * sacks that cover this area. 7244 */ 7245 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7246 } 7247 7248 static void 7249 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7250 { 7251 struct tcp_bbr *bbr; 7252 struct bbr_sendmap *rsm; 7253 uint32_t cts; 7254 7255 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7256 cts = bbr->r_ctl.rc_rcvtime; 7257 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7258 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7259 if ((rsm->r_end - rsm->r_start) <= 1) { 7260 /* Log out the SYN completely */ 7261 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7262 rsm->r_rtr_bytes = 0; 7263 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7264 if (rsm->r_in_tmap) { 7265 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7266 rsm->r_in_tmap = 0; 7267 } 7268 if (bbr->r_ctl.rc_next == rsm) { 7269 /* scoot along the marker */ 7270 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7271 } 7272 if (to != NULL) 7273 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7274 bbr_free(bbr, rsm); 7275 } else { 7276 /* There is more (Fast open)? strip out SYN. */ 7277 rsm->r_flags &= ~BBR_HAS_SYN; 7278 rsm->r_start++; 7279 } 7280 } 7281 } 7282 7283 /* 7284 * Returns the number of bytes that were 7285 * acknowledged by SACK blocks. 7286 */ 7287 7288 static uint32_t 7289 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7290 uint32_t *prev_acked) 7291 { 7292 uint32_t changed, last_seq, entered_recovery = 0; 7293 struct tcp_bbr *bbr; 7294 struct bbr_sendmap *rsm; 7295 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7296 register uint32_t th_ack; 7297 int32_t i, j, k, new_sb, num_sack_blks = 0; 7298 uint32_t cts, acked, ack_point, sack_changed = 0; 7299 uint32_t p_maxseg, maxseg, p_acked = 0; 7300 7301 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7302 if (tcp_get_flags(th) & TH_RST) { 7303 /* We don't log resets */ 7304 return (0); 7305 } 7306 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7307 cts = bbr->r_ctl.rc_rcvtime; 7308 7309 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7310 changed = 0; 7311 maxseg = tp->t_maxseg - bbr->rc_last_options; 7312 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7313 th_ack = th->th_ack; 7314 if (SEQ_GT(th_ack, tp->snd_una)) { 7315 acked = th_ack - tp->snd_una; 7316 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7317 bbr->rc_tp->t_acktime = ticks; 7318 } else 7319 acked = 0; 7320 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7321 /* Only sent here for sack processing */ 7322 goto proc_sack; 7323 } 7324 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7325 changed = th_ack - rsm->r_start; 7326 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7327 /* 7328 * For the SYN incoming case we will not have called 7329 * tcp_output for the sending of the SYN, so there will be 7330 * no map. All other cases should probably be a panic. 7331 */ 7332 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7333 /* 7334 * We have a timestamp that can be used to generate 7335 * an initial RTT. 7336 */ 7337 uint32_t ts, now, rtt; 7338 7339 ts = bbr_ts_convert(to->to_tsecr); 7340 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7341 rtt = now - ts; 7342 if (rtt < 1) 7343 rtt = 1; 7344 bbr_log_type_bbrrttprop(bbr, rtt, 7345 tp->iss, 0, cts, 7346 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7347 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7348 changed = 1; 7349 bbr->r_wanted_output = 1; 7350 goto out; 7351 } 7352 goto proc_sack; 7353 } else if (rsm == NULL) { 7354 goto out; 7355 } 7356 if (changed) { 7357 /* 7358 * The ACK point is advancing to th_ack, we must drop off 7359 * the packets in the rack log and calculate any eligble 7360 * RTT's. 7361 */ 7362 bbr->r_wanted_output = 1; 7363 more: 7364 if (rsm == NULL) { 7365 if (tp->t_flags & TF_SENTFIN) { 7366 /* if we send a FIN we will not hav a map */ 7367 goto proc_sack; 7368 } 7369 #ifdef BBR_INVARIANTS 7370 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7371 tp, 7372 th, tp->t_state, bbr, 7373 tp->snd_una, tp->snd_max, changed); 7374 #endif 7375 goto proc_sack; 7376 } 7377 } 7378 if (SEQ_LT(th_ack, rsm->r_start)) { 7379 /* Huh map is missing this */ 7380 #ifdef BBR_INVARIANTS 7381 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7382 rsm->r_start, 7383 th_ack, tp->t_state, 7384 bbr->r_state, bbr); 7385 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7386 #endif 7387 goto proc_sack; 7388 } else if (th_ack == rsm->r_start) { 7389 /* None here to ack */ 7390 goto proc_sack; 7391 } 7392 /* 7393 * Clear the dup ack counter, it will 7394 * either be freed or if there is some 7395 * remaining we need to start it at zero. 7396 */ 7397 rsm->r_dupack = 0; 7398 /* Now do we consume the whole thing? */ 7399 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7400 /* Its all consumed. */ 7401 uint32_t left; 7402 7403 if (rsm->r_flags & BBR_ACKED) { 7404 /* 7405 * It was acked on the scoreboard -- remove it from 7406 * total 7407 */ 7408 p_acked += (rsm->r_end - rsm->r_start); 7409 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7410 if (bbr->r_ctl.rc_sacked == 0) 7411 bbr->r_ctl.rc_sacklast = NULL; 7412 } else { 7413 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7414 if (rsm->r_flags & BBR_MARKED_LOST) { 7415 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7416 } 7417 if (rsm->r_flags & BBR_SACK_PASSED) { 7418 /* 7419 * There are acked segments ACKED on the 7420 * scoreboard further up. We are seeing 7421 * reordering. 7422 */ 7423 BBR_STAT_INC(bbr_reorder_seen); 7424 bbr->r_ctl.rc_reorder_ts = cts; 7425 if (rsm->r_flags & BBR_MARKED_LOST) { 7426 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7427 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7428 /* LT sampling also needs adjustment */ 7429 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7430 } 7431 } 7432 rsm->r_flags &= ~BBR_MARKED_LOST; 7433 } 7434 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7435 rsm->r_rtr_bytes = 0; 7436 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7437 if (rsm->r_in_tmap) { 7438 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7439 rsm->r_in_tmap = 0; 7440 } 7441 if (bbr->r_ctl.rc_next == rsm) { 7442 /* scoot along the marker */ 7443 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7444 } 7445 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7446 /* Adjust the packet counts */ 7447 left = th_ack - rsm->r_end; 7448 /* Free back to zone */ 7449 bbr_free(bbr, rsm); 7450 if (left) { 7451 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7452 goto more; 7453 } 7454 goto proc_sack; 7455 } 7456 if (rsm->r_flags & BBR_ACKED) { 7457 /* 7458 * It was acked on the scoreboard -- remove it from total 7459 * for the part being cum-acked. 7460 */ 7461 p_acked += (rsm->r_end - rsm->r_start); 7462 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7463 if (bbr->r_ctl.rc_sacked == 0) 7464 bbr->r_ctl.rc_sacklast = NULL; 7465 } else { 7466 /* 7467 * It was acked up to th_ack point for the first time 7468 */ 7469 struct bbr_sendmap lrsm; 7470 7471 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7472 lrsm.r_end = th_ack; 7473 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7474 } 7475 if ((rsm->r_flags & BBR_MARKED_LOST) && 7476 ((rsm->r_flags & BBR_ACKED) == 0)) { 7477 /* 7478 * It was marked lost and partly ack'd now 7479 * for the first time. We lower the rc_lost_bytes 7480 * and still leave it MARKED. 7481 */ 7482 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7483 } 7484 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7485 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7486 rsm->r_rtr_bytes = 0; 7487 /* adjust packet count */ 7488 rsm->r_start = th_ack; 7489 proc_sack: 7490 /* Check for reneging */ 7491 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7492 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7493 /* 7494 * The peer has moved snd_una up to the edge of this send, 7495 * i.e. one that it had previously acked. The only way that 7496 * can be true if the peer threw away data (space issues) 7497 * that it had previously sacked (else it would have given 7498 * us snd_una up to (rsm->r_end). We need to undo the acked 7499 * markings here. 7500 * 7501 * Note we have to look to make sure th_ack is our 7502 * rsm->r_start in case we get an old ack where th_ack is 7503 * behind snd_una. 7504 */ 7505 bbr_peer_reneges(bbr, rsm, th->th_ack); 7506 } 7507 if ((to->to_flags & TOF_SACK) == 0) { 7508 /* We are done nothing left to log */ 7509 goto out; 7510 } 7511 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7512 if (rsm) { 7513 last_seq = rsm->r_end; 7514 } else { 7515 last_seq = tp->snd_max; 7516 } 7517 /* Sack block processing */ 7518 if (SEQ_GT(th_ack, tp->snd_una)) 7519 ack_point = th_ack; 7520 else 7521 ack_point = tp->snd_una; 7522 for (i = 0; i < to->to_nsacks; i++) { 7523 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7524 &sack, sizeof(sack)); 7525 sack.start = ntohl(sack.start); 7526 sack.end = ntohl(sack.end); 7527 if (SEQ_GT(sack.end, sack.start) && 7528 SEQ_GT(sack.start, ack_point) && 7529 SEQ_LT(sack.start, tp->snd_max) && 7530 SEQ_GT(sack.end, ack_point) && 7531 SEQ_LEQ(sack.end, tp->snd_max)) { 7532 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7533 (SEQ_LT(sack.end, last_seq)) && 7534 ((sack.end - sack.start) < (p_maxseg / 8))) { 7535 /* 7536 * Not the last piece and its smaller than 7537 * 1/8th of a p_maxseg. We ignore this. 7538 */ 7539 BBR_STAT_INC(bbr_runt_sacks); 7540 continue; 7541 } 7542 sack_blocks[num_sack_blks] = sack; 7543 num_sack_blks++; 7544 } else if (SEQ_LEQ(sack.start, th_ack) && 7545 SEQ_LEQ(sack.end, th_ack)) { 7546 /* 7547 * Its a D-SACK block. 7548 */ 7549 tcp_record_dsack(tp, sack.start, sack.end, 0); 7550 } 7551 } 7552 if (num_sack_blks == 0) 7553 goto out; 7554 /* 7555 * Sort the SACK blocks so we can update the rack scoreboard with 7556 * just one pass. 7557 */ 7558 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, 7559 num_sack_blks, th->th_ack); 7560 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7561 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7562 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7563 num_sack_blks = new_sb; 7564 if (num_sack_blks < 2) { 7565 goto do_sack_work; 7566 } 7567 /* Sort the sacks */ 7568 for (i = 0; i < num_sack_blks; i++) { 7569 for (j = i + 1; j < num_sack_blks; j++) { 7570 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7571 sack = sack_blocks[i]; 7572 sack_blocks[i] = sack_blocks[j]; 7573 sack_blocks[j] = sack; 7574 } 7575 } 7576 } 7577 /* 7578 * Now are any of the sack block ends the same (yes some 7579 * implememtations send these)? 7580 */ 7581 again: 7582 if (num_sack_blks > 1) { 7583 for (i = 0; i < num_sack_blks; i++) { 7584 for (j = i + 1; j < num_sack_blks; j++) { 7585 if (sack_blocks[i].end == sack_blocks[j].end) { 7586 /* 7587 * Ok these two have the same end we 7588 * want the smallest end and then 7589 * throw away the larger and start 7590 * again. 7591 */ 7592 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7593 /* 7594 * The second block covers 7595 * more area use that 7596 */ 7597 sack_blocks[i].start = sack_blocks[j].start; 7598 } 7599 /* 7600 * Now collapse out the dup-sack and 7601 * lower the count 7602 */ 7603 for (k = (j + 1); k < num_sack_blks; k++) { 7604 sack_blocks[j].start = sack_blocks[k].start; 7605 sack_blocks[j].end = sack_blocks[k].end; 7606 j++; 7607 } 7608 num_sack_blks--; 7609 goto again; 7610 } 7611 } 7612 } 7613 } 7614 do_sack_work: 7615 rsm = bbr->r_ctl.rc_sacklast; 7616 for (i = 0; i < num_sack_blks; i++) { 7617 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7618 if (acked) { 7619 bbr->r_wanted_output = 1; 7620 changed += acked; 7621 sack_changed += acked; 7622 } 7623 } 7624 out: 7625 *prev_acked = p_acked; 7626 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7627 /* 7628 * Ok we have a high probability that we need to go in to 7629 * recovery since we have data sack'd 7630 */ 7631 struct bbr_sendmap *rsm; 7632 7633 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7634 if (rsm) { 7635 /* Enter recovery */ 7636 entered_recovery = 1; 7637 bbr->r_wanted_output = 1; 7638 /* 7639 * When we enter recovery we need to assure we send 7640 * one packet. 7641 */ 7642 if (bbr->r_ctl.rc_resend == NULL) { 7643 bbr->r_ctl.rc_resend = rsm; 7644 } 7645 } 7646 } 7647 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7648 /* 7649 * See if we need to rack-retransmit anything if so set it 7650 * up as the thing to resend assuming something else is not 7651 * already in that position. 7652 */ 7653 if (bbr->r_ctl.rc_resend == NULL) { 7654 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7655 } 7656 } 7657 /* 7658 * We return the amount that changed via sack, this is used by the 7659 * ack-received code to augment what was changed between th_ack <-> 7660 * snd_una. 7661 */ 7662 return (sack_changed); 7663 } 7664 7665 static void 7666 bbr_strike_dupack(struct tcp_bbr *bbr) 7667 { 7668 struct bbr_sendmap *rsm; 7669 7670 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7671 if (rsm && (rsm->r_dupack < 0xff)) { 7672 rsm->r_dupack++; 7673 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7674 bbr->r_wanted_output = 1; 7675 } 7676 } 7677 7678 /* 7679 * Return value of 1, we do not need to call bbr_process_data(). 7680 * return value of 0, bbr_process_data can be called. 7681 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7682 * its unlocked and probably unsafe to touch the TCB. 7683 */ 7684 static int 7685 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7686 struct tcpcb *tp, struct tcpopt *to, 7687 uint32_t tiwin, int32_t tlen, 7688 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7689 { 7690 int32_t ourfinisacked = 0; 7691 int32_t acked_amount; 7692 uint16_t nsegs; 7693 int32_t acked; 7694 uint32_t lost, sack_changed = 0; 7695 struct mbuf *mfree; 7696 struct tcp_bbr *bbr; 7697 uint32_t prev_acked = 0; 7698 7699 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7700 lost = bbr->r_ctl.rc_lost; 7701 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7702 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7703 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7704 bbr->r_wanted_output = 1; 7705 return (1); 7706 } 7707 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7708 /* Process the ack */ 7709 if (bbr->rc_in_persist) 7710 tp->t_rxtshift = 0; 7711 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7712 bbr_strike_dupack(bbr); 7713 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7714 } 7715 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7716 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7717 /* 7718 * Old ack, behind the last one rcv'd or a duplicate ack 7719 * with SACK info. 7720 */ 7721 if (th->th_ack == tp->snd_una) { 7722 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7723 if (bbr->r_state == TCPS_SYN_SENT) { 7724 /* 7725 * Special case on where we sent SYN. When 7726 * the SYN-ACK is processed in syn_sent 7727 * state it bumps the snd_una. This causes 7728 * us to hit here even though we did ack 1 7729 * byte. 7730 * 7731 * Go through the nothing left case so we 7732 * send data. 7733 */ 7734 goto nothing_left; 7735 } 7736 } 7737 return (0); 7738 } 7739 /* 7740 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7741 * something we sent. 7742 */ 7743 if (tp->t_flags & TF_NEEDSYN) { 7744 /* 7745 * T/TCP: Connection was half-synchronized, and our SYN has 7746 * been ACK'd (so connection is now fully synchronized). Go 7747 * to non-starred state, increment snd_una for ACK of SYN, 7748 * and check if we can do window scaling. 7749 */ 7750 tp->t_flags &= ~TF_NEEDSYN; 7751 tp->snd_una++; 7752 /* Do window scaling? */ 7753 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7754 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7755 tp->rcv_scale = tp->request_r_scale; 7756 /* Send window already scaled. */ 7757 } 7758 } 7759 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7760 7761 acked = BYTES_THIS_ACK(tp, th); 7762 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7763 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7764 7765 /* 7766 * If we just performed our first retransmit, and the ACK arrives 7767 * within our recovery window, then it was a mistake to do the 7768 * retransmit in the first place. Recover our original cwnd and 7769 * ssthresh, and proceed to transmit where we left off. 7770 */ 7771 if (tp->t_flags & TF_PREVVALID) { 7772 tp->t_flags &= ~TF_PREVVALID; 7773 if (tp->t_rxtshift == 1 && 7774 (int)(ticks - tp->t_badrxtwin) < 0) 7775 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7776 } 7777 SOCKBUF_LOCK(&so->so_snd); 7778 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7779 tp->snd_wnd -= acked_amount; 7780 mfree = sbcut_locked(&so->so_snd, acked_amount); 7781 /* NB: sowwakeup_locked() does an implicit unlock. */ 7782 sowwakeup_locked(so); 7783 m_freem(mfree); 7784 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7785 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7786 } 7787 tp->snd_una = th->th_ack; 7788 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7789 if (IN_RECOVERY(tp->t_flags)) { 7790 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7791 (SEQ_LT(th->th_ack, tp->snd_max))) { 7792 tcp_bbr_partialack(tp); 7793 } else { 7794 bbr_post_recovery(tp); 7795 } 7796 } 7797 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7798 tp->snd_recover = tp->snd_una; 7799 } 7800 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7801 tp->snd_nxt = tp->snd_max; 7802 } 7803 if (tp->snd_una == tp->snd_max) { 7804 /* Nothing left outstanding */ 7805 nothing_left: 7806 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7807 if (sbavail(&so->so_snd) == 0) 7808 bbr->rc_tp->t_acktime = 0; 7809 if ((sbused(&so->so_snd) == 0) && 7810 (tp->t_flags & TF_SENTFIN)) { 7811 ourfinisacked = 1; 7812 } 7813 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7814 if (bbr->rc_in_persist == 0) { 7815 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7816 } 7817 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7818 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7819 /* 7820 * We invalidate the last ack here since we 7821 * don't want to transfer forward the time 7822 * for our sum's calculations. 7823 */ 7824 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7825 (sbavail(&so->so_snd) == 0) && 7826 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7827 /* 7828 * The socket was gone and the peer sent data, time 7829 * to reset him. 7830 */ 7831 *ret_val = 1; 7832 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7833 /* tcp_close will kill the inp pre-log the Reset */ 7834 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7835 tp = tcp_close(tp); 7836 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7837 BBR_STAT_INC(bbr_dropped_af_data); 7838 return (1); 7839 } 7840 /* Set need output so persist might get set */ 7841 bbr->r_wanted_output = 1; 7842 } 7843 if (ofia) 7844 *ofia = ourfinisacked; 7845 return (0); 7846 } 7847 7848 static void 7849 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7850 { 7851 if (bbr->rc_in_persist == 0) { 7852 bbr_timer_cancel(bbr, __LINE__, cts); 7853 bbr->r_ctl.rc_last_delay_val = 0; 7854 tp->t_rxtshift = 0; 7855 bbr->rc_in_persist = 1; 7856 bbr->r_ctl.rc_went_idle_time = cts; 7857 /* We should be capped when rw went to 0 but just in case */ 7858 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7859 /* Time freezes for the state, so do the accounting now */ 7860 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7861 uint32_t time_in; 7862 7863 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7864 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7865 int32_t idx; 7866 7867 idx = bbr_state_val(bbr); 7868 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7869 } else { 7870 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7871 } 7872 } 7873 bbr->r_ctl.rc_bbr_state_time = cts; 7874 } 7875 } 7876 7877 static void 7878 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7879 { 7880 /* 7881 * Note that if idle time does not exceed our 7882 * threshold, we do nothing continuing the state 7883 * transitions we were last walking through. 7884 */ 7885 if (idle_time >= bbr_idle_restart_threshold) { 7886 if (bbr->rc_use_idle_restart) { 7887 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7888 /* 7889 * Set our target using BBR_UNIT, so 7890 * we increase at a dramatic rate but 7891 * we stop when we get the pipe 7892 * full again for our current b/w estimate. 7893 */ 7894 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7895 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7896 bbr_set_state_target(bbr, __LINE__); 7897 /* Now setup our gains to ramp up */ 7898 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7899 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7900 bbr_log_type_statechange(bbr, cts, __LINE__); 7901 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7902 bbr_substate_change(bbr, cts, __LINE__, 1); 7903 } 7904 } 7905 } 7906 7907 static void 7908 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7909 { 7910 uint32_t idle_time; 7911 7912 if (bbr->rc_in_persist == 0) 7913 return; 7914 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7915 bbr->rc_in_persist = 0; 7916 bbr->rc_hit_state_1 = 0; 7917 bbr->r_ctl.rc_del_time = cts; 7918 /* 7919 * We invalidate the last ack here since we 7920 * don't want to transfer forward the time 7921 * for our sum's calculations. 7922 */ 7923 if (tcp_in_hpts(bbr->rc_tp)) { 7924 tcp_hpts_remove(bbr->rc_tp); 7925 bbr->rc_timer_first = 0; 7926 bbr->r_ctl.rc_hpts_flags = 0; 7927 bbr->r_ctl.rc_last_delay_val = 0; 7928 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7929 bbr->r_agg_early_set = 0; 7930 bbr->r_ctl.rc_agg_early = 0; 7931 } 7932 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7933 if (idle_time >= bbr_rtt_probe_time) { 7934 /* 7935 * This qualifies as a RTT_PROBE session since we drop the 7936 * data outstanding to nothing and waited more than 7937 * bbr_rtt_probe_time. 7938 */ 7939 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7940 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7941 } 7942 tp->t_rxtshift = 0; 7943 /* 7944 * If in probeBW and we have persisted more than an RTT lets do 7945 * special handling. 7946 */ 7947 /* Force a time based epoch */ 7948 bbr_set_epoch(bbr, cts, __LINE__); 7949 /* 7950 * Setup the lost so we don't count anything against the guy 7951 * we have been stuck with during persists. 7952 */ 7953 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7954 /* Time un-freezes for the state */ 7955 bbr->r_ctl.rc_bbr_state_time = cts; 7956 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7957 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7958 /* 7959 * If we are going back to probe-bw 7960 * or probe_rtt, we may need to possibly 7961 * do a fast restart. 7962 */ 7963 bbr_restart_after_idle(bbr, cts, idle_time); 7964 } 7965 } 7966 7967 static void 7968 bbr_collapsed_window(struct tcp_bbr *bbr) 7969 { 7970 /* 7971 * Now we must walk the 7972 * send map and divide the 7973 * ones left stranded. These 7974 * guys can't cause us to abort 7975 * the connection and are really 7976 * "unsent". However if a buggy 7977 * client actually did keep some 7978 * of the data i.e. collapsed the win 7979 * and refused to ack and then opened 7980 * the win and acked that data. We would 7981 * get into an ack war, the simplier 7982 * method then of just pretending we 7983 * did not send those segments something 7984 * won't work. 7985 */ 7986 struct bbr_sendmap *rsm, *nrsm; 7987 tcp_seq max_seq; 7988 uint32_t maxseg; 7989 int can_split = 0; 7990 int fnd = 0; 7991 7992 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 7993 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 7994 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 7995 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 7996 /* Find the first seq past or at maxseq */ 7997 if (rsm->r_flags & BBR_RWND_COLLAPSED) 7998 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 7999 if (SEQ_GEQ(max_seq, rsm->r_start) && 8000 SEQ_GEQ(rsm->r_end, max_seq)) { 8001 fnd = 1; 8002 break; 8003 } 8004 } 8005 bbr->rc_has_collapsed = 0; 8006 if (!fnd) { 8007 /* Nothing to do strange */ 8008 return; 8009 } 8010 /* 8011 * Now can we split? 8012 * 8013 * We don't want to split if splitting 8014 * would generate too many small segments 8015 * less we let an attacker fragment our 8016 * send_map and leave us out of memory. 8017 */ 8018 if ((max_seq != rsm->r_start) && 8019 (max_seq != rsm->r_end)){ 8020 /* can we split? */ 8021 int res1, res2; 8022 8023 res1 = max_seq - rsm->r_start; 8024 res2 = rsm->r_end - max_seq; 8025 if ((res1 >= (maxseg/8)) && 8026 (res2 >= (maxseg/8))) { 8027 /* No small pieces here */ 8028 can_split = 1; 8029 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8030 /* We are under the limit */ 8031 can_split = 1; 8032 } 8033 } 8034 /* Ok do we need to split this rsm? */ 8035 if (max_seq == rsm->r_start) { 8036 /* It's this guy no split required */ 8037 nrsm = rsm; 8038 } else if (max_seq == rsm->r_end) { 8039 /* It's the next one no split required. */ 8040 nrsm = TAILQ_NEXT(rsm, r_next); 8041 if (nrsm == NULL) { 8042 /* Huh? */ 8043 return; 8044 } 8045 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8046 /* yep we need to split it */ 8047 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8048 if (nrsm == NULL) { 8049 /* failed XXXrrs what can we do mark the whole? */ 8050 nrsm = rsm; 8051 goto no_split; 8052 } 8053 /* Clone it */ 8054 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8055 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8056 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8057 if (rsm->r_in_tmap) { 8058 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8059 nrsm->r_in_tmap = 1; 8060 } 8061 } else { 8062 /* 8063 * Split not allowed just start here just 8064 * use this guy. 8065 */ 8066 nrsm = rsm; 8067 } 8068 no_split: 8069 BBR_STAT_INC(bbr_collapsed_win); 8070 /* reuse fnd as a count */ 8071 fnd = 0; 8072 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8073 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8074 fnd++; 8075 bbr->rc_has_collapsed = 1; 8076 } 8077 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8078 } 8079 8080 static void 8081 bbr_un_collapse_window(struct tcp_bbr *bbr) 8082 { 8083 struct bbr_sendmap *rsm; 8084 int cleared = 0; 8085 8086 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8087 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8088 /* Clear the flag */ 8089 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8090 cleared++; 8091 } else 8092 break; 8093 } 8094 bbr_log_type_rwnd_collapse(bbr, 8095 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8096 bbr->rc_has_collapsed = 0; 8097 } 8098 8099 /* 8100 * Return value of 1, the TCB is unlocked and most 8101 * likely gone, return value of 0, the TCB is still 8102 * locked. 8103 */ 8104 static int 8105 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8106 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8107 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8108 { 8109 /* 8110 * Update window information. Don't look at window if no ACK: TAC's 8111 * send garbage on first SYN. 8112 */ 8113 uint16_t nsegs; 8114 int32_t tfo_syn; 8115 struct tcp_bbr *bbr; 8116 8117 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8118 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8119 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8120 if ((thflags & TH_ACK) && 8121 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8122 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8123 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8124 /* keep track of pure window updates */ 8125 if (tlen == 0 && 8126 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8127 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8128 tp->snd_wnd = tiwin; 8129 tp->snd_wl1 = th->th_seq; 8130 tp->snd_wl2 = th->th_ack; 8131 if (tp->snd_wnd > tp->max_sndwnd) 8132 tp->max_sndwnd = tp->snd_wnd; 8133 bbr->r_wanted_output = 1; 8134 } else if (thflags & TH_ACK) { 8135 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8136 tp->snd_wnd = tiwin; 8137 tp->snd_wl1 = th->th_seq; 8138 tp->snd_wl2 = th->th_ack; 8139 } 8140 } 8141 if (tp->snd_wnd < ctf_outstanding(tp)) 8142 /* The peer collapsed its window on us */ 8143 bbr_collapsed_window(bbr); 8144 else if (bbr->rc_has_collapsed) 8145 bbr_un_collapse_window(bbr); 8146 /* Was persist timer active and now we have window space? */ 8147 if ((bbr->rc_in_persist != 0) && 8148 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8149 bbr_minseg(bbr)))) { 8150 /* 8151 * Make the rate persist at end of persist mode if idle long 8152 * enough 8153 */ 8154 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8155 8156 /* Make sure we output to start the timer */ 8157 bbr->r_wanted_output = 1; 8158 } 8159 /* Do we need to enter persist? */ 8160 if ((bbr->rc_in_persist == 0) && 8161 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8162 TCPS_HAVEESTABLISHED(tp->t_state) && 8163 (tp->snd_max == tp->snd_una) && 8164 sbavail(&so->so_snd) && 8165 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8166 /* No send window.. we must enter persist */ 8167 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8168 } 8169 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8170 m_freem(m); 8171 return (0); 8172 } 8173 /* 8174 * We don't support urgent data but 8175 * drag along the up just to make sure 8176 * if there is a stack switch no one 8177 * is surprised. 8178 */ 8179 tp->rcv_up = tp->rcv_nxt; 8180 8181 /* 8182 * Process the segment text, merging it into the TCP sequencing 8183 * queue, and arranging for acknowledgment of receipt if necessary. 8184 * This process logically involves adjusting tp->rcv_wnd as data is 8185 * presented to the user (this happens in tcp_usrreq.c, case 8186 * PRU_RCVD). If a FIN has already been received on this connection 8187 * then we just ignore the text. 8188 */ 8189 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8190 IS_FASTOPEN(tp->t_flags)); 8191 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8192 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8193 tcp_seq save_start = th->th_seq; 8194 tcp_seq save_rnxt = tp->rcv_nxt; 8195 int save_tlen = tlen; 8196 8197 m_adj(m, drop_hdrlen); /* delayed header drop */ 8198 /* 8199 * Insert segment which includes th into TCP reassembly 8200 * queue with control block tp. Set thflags to whether 8201 * reassembly now includes a segment with FIN. This handles 8202 * the common case inline (segment is the next to be 8203 * received on an established connection, and the queue is 8204 * empty), avoiding linkage into and removal from the queue 8205 * and repetition of various conversions. Set DELACK for 8206 * segments received in order, but ack immediately when 8207 * segments are out of order (so fast retransmit can work). 8208 */ 8209 if (th->th_seq == tp->rcv_nxt && 8210 SEGQ_EMPTY(tp) && 8211 (TCPS_HAVEESTABLISHED(tp->t_state) || 8212 tfo_syn)) { 8213 #ifdef NETFLIX_SB_LIMITS 8214 u_int mcnt, appended; 8215 8216 if (so->so_rcv.sb_shlim) { 8217 mcnt = m_memcnt(m); 8218 appended = 0; 8219 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8220 CFO_NOSLEEP, NULL) == false) { 8221 counter_u64_add(tcp_sb_shlim_fails, 1); 8222 m_freem(m); 8223 return (0); 8224 } 8225 } 8226 8227 #endif 8228 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8229 bbr->bbr_segs_rcvd += max(1, nsegs); 8230 tp->t_flags |= TF_DELACK; 8231 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8232 } else { 8233 bbr->r_wanted_output = 1; 8234 tp->t_flags |= TF_ACKNOW; 8235 } 8236 tp->rcv_nxt += tlen; 8237 if (tlen && 8238 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8239 (tp->t_fbyte_in == 0)) { 8240 tp->t_fbyte_in = ticks; 8241 if (tp->t_fbyte_in == 0) 8242 tp->t_fbyte_in = 1; 8243 if (tp->t_fbyte_out && tp->t_fbyte_in) 8244 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8245 } 8246 thflags = tcp_get_flags(th) & TH_FIN; 8247 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8248 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8249 SOCKBUF_LOCK(&so->so_rcv); 8250 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8251 m_freem(m); 8252 else 8253 #ifdef NETFLIX_SB_LIMITS 8254 appended = 8255 #endif 8256 sbappendstream_locked(&so->so_rcv, m, 0); 8257 /* NB: sorwakeup_locked() does an implicit unlock. */ 8258 sorwakeup_locked(so); 8259 #ifdef NETFLIX_SB_LIMITS 8260 if (so->so_rcv.sb_shlim && appended != mcnt) 8261 counter_fo_release(so->so_rcv.sb_shlim, 8262 mcnt - appended); 8263 #endif 8264 8265 } else { 8266 /* 8267 * XXX: Due to the header drop above "th" is 8268 * theoretically invalid by now. Fortunately 8269 * m_adj() doesn't actually frees any mbufs when 8270 * trimming from the head. 8271 */ 8272 tcp_seq temp = save_start; 8273 8274 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8275 tp->t_flags |= TF_ACKNOW; 8276 if (tp->t_flags & TF_WAKESOR) { 8277 tp->t_flags &= ~TF_WAKESOR; 8278 /* NB: sorwakeup_locked() does an implicit unlock. */ 8279 sorwakeup_locked(so); 8280 } 8281 } 8282 if ((tp->t_flags & TF_SACK_PERMIT) && 8283 (save_tlen > 0) && 8284 TCPS_HAVEESTABLISHED(tp->t_state)) { 8285 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8286 /* 8287 * DSACK actually handled in the fastpath 8288 * above. 8289 */ 8290 tcp_update_sack_list(tp, save_start, 8291 save_start + save_tlen); 8292 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8293 if ((tp->rcv_numsacks >= 1) && 8294 (tp->sackblks[0].end == save_start)) { 8295 /* 8296 * Partial overlap, recorded at todrop 8297 * above. 8298 */ 8299 tcp_update_sack_list(tp, 8300 tp->sackblks[0].start, 8301 tp->sackblks[0].end); 8302 } else { 8303 tcp_update_dsack_list(tp, save_start, 8304 save_start + save_tlen); 8305 } 8306 } else if (tlen >= save_tlen) { 8307 /* Update of sackblks. */ 8308 tcp_update_dsack_list(tp, save_start, 8309 save_start + save_tlen); 8310 } else if (tlen > 0) { 8311 tcp_update_dsack_list(tp, save_start, 8312 save_start + tlen); 8313 } 8314 } 8315 } else { 8316 m_freem(m); 8317 thflags &= ~TH_FIN; 8318 } 8319 8320 /* 8321 * If FIN is received ACK the FIN and let the user know that the 8322 * connection is closing. 8323 */ 8324 if (thflags & TH_FIN) { 8325 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8326 /* The socket upcall is handled by socantrcvmore. */ 8327 socantrcvmore(so); 8328 /* 8329 * If connection is half-synchronized (ie NEEDSYN 8330 * flag on) then delay ACK, so it may be piggybacked 8331 * when SYN is sent. Otherwise, since we received a 8332 * FIN then no more input can be expected, send ACK 8333 * now. 8334 */ 8335 if (tp->t_flags & TF_NEEDSYN) { 8336 tp->t_flags |= TF_DELACK; 8337 bbr_timer_cancel(bbr, 8338 __LINE__, bbr->r_ctl.rc_rcvtime); 8339 } else { 8340 tp->t_flags |= TF_ACKNOW; 8341 } 8342 tp->rcv_nxt++; 8343 } 8344 switch (tp->t_state) { 8345 /* 8346 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8347 * CLOSE_WAIT state. 8348 */ 8349 case TCPS_SYN_RECEIVED: 8350 tp->t_starttime = ticks; 8351 /* FALLTHROUGH */ 8352 case TCPS_ESTABLISHED: 8353 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8354 break; 8355 8356 /* 8357 * If still in FIN_WAIT_1 STATE FIN has not been 8358 * acked so enter the CLOSING state. 8359 */ 8360 case TCPS_FIN_WAIT_1: 8361 tcp_state_change(tp, TCPS_CLOSING); 8362 break; 8363 8364 /* 8365 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8366 * starting the time-wait timer, turning off the 8367 * other standard timers. 8368 */ 8369 case TCPS_FIN_WAIT_2: 8370 bbr->rc_timer_first = 1; 8371 bbr_timer_cancel(bbr, 8372 __LINE__, bbr->r_ctl.rc_rcvtime); 8373 tcp_twstart(tp); 8374 return (1); 8375 } 8376 } 8377 /* 8378 * Return any desired output. 8379 */ 8380 if ((tp->t_flags & TF_ACKNOW) || 8381 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8382 bbr->r_wanted_output = 1; 8383 } 8384 return (0); 8385 } 8386 8387 /* 8388 * Here nothing is really faster, its just that we 8389 * have broken out the fast-data path also just like 8390 * the fast-ack. Return 1 if we processed the packet 8391 * return 0 if you need to take the "slow-path". 8392 */ 8393 static int 8394 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8395 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8396 uint32_t tiwin, int32_t nxt_pkt) 8397 { 8398 uint16_t nsegs; 8399 int32_t newsize = 0; /* automatic sockbuf scaling */ 8400 struct tcp_bbr *bbr; 8401 #ifdef NETFLIX_SB_LIMITS 8402 u_int mcnt, appended; 8403 #endif 8404 8405 /* On the hpts and we would have called output */ 8406 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8407 8408 /* 8409 * If last ACK falls within this segment's sequence numbers, record 8410 * the timestamp. NOTE that the test is modified according to the 8411 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8412 */ 8413 if (bbr->r_ctl.rc_resend != NULL) { 8414 return (0); 8415 } 8416 if (tiwin && tiwin != tp->snd_wnd) { 8417 return (0); 8418 } 8419 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8420 return (0); 8421 } 8422 if (__predict_false((to->to_flags & TOF_TS) && 8423 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8424 return (0); 8425 } 8426 if (__predict_false((th->th_ack != tp->snd_una))) { 8427 return (0); 8428 } 8429 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8430 return (0); 8431 } 8432 if ((to->to_flags & TOF_TS) != 0 && 8433 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8434 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8435 tp->ts_recent = to->to_tsval; 8436 } 8437 /* 8438 * This is a pure, in-sequence data packet with nothing on the 8439 * reassembly queue and we have enough buffer space to take it. 8440 */ 8441 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8442 8443 #ifdef NETFLIX_SB_LIMITS 8444 if (so->so_rcv.sb_shlim) { 8445 mcnt = m_memcnt(m); 8446 appended = 0; 8447 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8448 CFO_NOSLEEP, NULL) == false) { 8449 counter_u64_add(tcp_sb_shlim_fails, 1); 8450 m_freem(m); 8451 return (1); 8452 } 8453 } 8454 #endif 8455 /* Clean receiver SACK report if present */ 8456 if (tp->rcv_numsacks) 8457 tcp_clean_sackreport(tp); 8458 KMOD_TCPSTAT_INC(tcps_preddat); 8459 tp->rcv_nxt += tlen; 8460 if (tlen && 8461 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8462 (tp->t_fbyte_in == 0)) { 8463 tp->t_fbyte_in = ticks; 8464 if (tp->t_fbyte_in == 0) 8465 tp->t_fbyte_in = 1; 8466 if (tp->t_fbyte_out && tp->t_fbyte_in) 8467 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8468 } 8469 /* 8470 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8471 */ 8472 tp->snd_wl1 = th->th_seq; 8473 /* 8474 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8475 */ 8476 tp->rcv_up = tp->rcv_nxt; 8477 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8478 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8479 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8480 8481 /* Add data to socket buffer. */ 8482 SOCKBUF_LOCK(&so->so_rcv); 8483 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8484 m_freem(m); 8485 } else { 8486 /* 8487 * Set new socket buffer size. Give up when limit is 8488 * reached. 8489 */ 8490 if (newsize) 8491 if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) 8492 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8493 m_adj(m, drop_hdrlen); /* delayed header drop */ 8494 8495 #ifdef NETFLIX_SB_LIMITS 8496 appended = 8497 #endif 8498 sbappendstream_locked(&so->so_rcv, m, 0); 8499 ctf_calc_rwin(so, tp); 8500 } 8501 /* NB: sorwakeup_locked() does an implicit unlock. */ 8502 sorwakeup_locked(so); 8503 #ifdef NETFLIX_SB_LIMITS 8504 if (so->so_rcv.sb_shlim && mcnt != appended) 8505 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8506 #endif 8507 if (DELAY_ACK(tp, bbr, nsegs)) { 8508 bbr->bbr_segs_rcvd += max(1, nsegs); 8509 tp->t_flags |= TF_DELACK; 8510 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8511 } else { 8512 bbr->r_wanted_output = 1; 8513 tp->t_flags |= TF_ACKNOW; 8514 } 8515 return (1); 8516 } 8517 8518 /* 8519 * This subfunction is used to try to highly optimize the 8520 * fast path. We again allow window updates that are 8521 * in sequence to remain in the fast-path. We also add 8522 * in the __predict's to attempt to help the compiler. 8523 * Note that if we return a 0, then we can *not* process 8524 * it and the caller should push the packet into the 8525 * slow-path. If we return 1, then all is well and 8526 * the packet is fully processed. 8527 */ 8528 static int 8529 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8530 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8531 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8532 { 8533 int32_t acked; 8534 uint16_t nsegs; 8535 uint32_t sack_changed; 8536 uint32_t prev_acked = 0; 8537 struct tcp_bbr *bbr; 8538 8539 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8540 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8541 return (0); 8542 } 8543 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8544 /* Above what we have sent? */ 8545 return (0); 8546 } 8547 if (__predict_false(tiwin == 0)) { 8548 /* zero window */ 8549 return (0); 8550 } 8551 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8552 /* We need a SYN or a FIN, unlikely.. */ 8553 return (0); 8554 } 8555 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8556 /* Timestamp is behind .. old ack with seq wrap? */ 8557 return (0); 8558 } 8559 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8560 /* Still recovering */ 8561 return (0); 8562 } 8563 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8564 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8565 /* We are retransmitting */ 8566 return (0); 8567 } 8568 if (__predict_false(bbr->rc_in_persist != 0)) { 8569 /* In persist mode */ 8570 return (0); 8571 } 8572 if (bbr->r_ctl.rc_sacked) { 8573 /* We have sack holes on our scoreboard */ 8574 return (0); 8575 } 8576 /* Ok if we reach here, we can process a fast-ack */ 8577 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8578 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8579 /* 8580 * We never detect loss in fast ack [we can't 8581 * have a sack and can't be in recovery so 8582 * we always pass 0 (nothing detected)]. 8583 */ 8584 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8585 /* Did the window get updated? */ 8586 if (tiwin != tp->snd_wnd) { 8587 tp->snd_wnd = tiwin; 8588 tp->snd_wl1 = th->th_seq; 8589 if (tp->snd_wnd > tp->max_sndwnd) 8590 tp->max_sndwnd = tp->snd_wnd; 8591 } 8592 /* Do we need to exit persists? */ 8593 if ((bbr->rc_in_persist != 0) && 8594 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8595 bbr_minseg(bbr)))) { 8596 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8597 bbr->r_wanted_output = 1; 8598 } 8599 /* Do we need to enter persists? */ 8600 if ((bbr->rc_in_persist == 0) && 8601 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8602 TCPS_HAVEESTABLISHED(tp->t_state) && 8603 (tp->snd_max == tp->snd_una) && 8604 sbavail(&so->so_snd) && 8605 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8606 /* No send window.. we must enter persist */ 8607 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8608 } 8609 /* 8610 * If last ACK falls within this segment's sequence numbers, record 8611 * the timestamp. NOTE that the test is modified according to the 8612 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8613 */ 8614 if ((to->to_flags & TOF_TS) != 0 && 8615 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8616 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8617 tp->ts_recent = to->to_tsval; 8618 } 8619 /* 8620 * This is a pure ack for outstanding data. 8621 */ 8622 KMOD_TCPSTAT_INC(tcps_predack); 8623 8624 /* 8625 * "bad retransmit" recovery. 8626 */ 8627 if (tp->t_flags & TF_PREVVALID) { 8628 tp->t_flags &= ~TF_PREVVALID; 8629 if (tp->t_rxtshift == 1 && 8630 (int)(ticks - tp->t_badrxtwin) < 0) 8631 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8632 } 8633 /* 8634 * Recalculate the transmit timer / rtt. 8635 * 8636 * Some boxes send broken timestamp replies during the SYN+ACK 8637 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8638 * and blow up the retransmit timer. 8639 */ 8640 acked = BYTES_THIS_ACK(tp, th); 8641 8642 #ifdef TCP_HHOOK 8643 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8644 hhook_run_tcp_est_in(tp, th, to); 8645 #endif 8646 8647 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8648 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8649 sbdrop(&so->so_snd, acked); 8650 8651 if (SEQ_GT(th->th_ack, tp->snd_una)) 8652 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8653 tp->snd_una = th->th_ack; 8654 if (tp->snd_wnd < ctf_outstanding(tp)) 8655 /* The peer collapsed its window on us */ 8656 bbr_collapsed_window(bbr); 8657 else if (bbr->rc_has_collapsed) 8658 bbr_un_collapse_window(bbr); 8659 8660 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8661 tp->snd_recover = tp->snd_una; 8662 } 8663 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8664 /* 8665 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8666 */ 8667 tp->snd_wl2 = th->th_ack; 8668 m_freem(m); 8669 /* 8670 * If all outstanding data are acked, stop retransmit timer, 8671 * otherwise restart timer using current (possibly backed-off) 8672 * value. If process is waiting for space, wakeup/selwakeup/signal. 8673 * If data are ready to send, let tcp_output decide between more 8674 * output or persist. 8675 * Wake up the socket if we have room to write more. 8676 */ 8677 sowwakeup(so); 8678 if (tp->snd_una == tp->snd_max) { 8679 /* Nothing left outstanding */ 8680 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8681 if (sbavail(&so->so_snd) == 0) 8682 bbr->rc_tp->t_acktime = 0; 8683 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8684 if (bbr->rc_in_persist == 0) { 8685 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8686 } 8687 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8688 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8689 /* 8690 * We invalidate the last ack here since we 8691 * don't want to transfer forward the time 8692 * for our sum's calculations. 8693 */ 8694 bbr->r_wanted_output = 1; 8695 } 8696 if (sbavail(&so->so_snd)) { 8697 bbr->r_wanted_output = 1; 8698 } 8699 return (1); 8700 } 8701 8702 /* 8703 * Return value of 1, the TCB is unlocked and most 8704 * likely gone, return value of 0, the TCB is still 8705 * locked. 8706 */ 8707 static int 8708 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8709 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8710 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8711 { 8712 int32_t todrop; 8713 int32_t ourfinisacked = 0; 8714 struct tcp_bbr *bbr; 8715 int32_t ret_val = 0; 8716 8717 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8718 8719 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8720 ctf_calc_rwin(so, tp); 8721 /* 8722 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8723 * SYN, drop the input. if seg contains a RST, then drop the 8724 * connection. if seg does not contain SYN, then drop it. Otherwise 8725 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8726 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8727 * not support ECN so we will not say we are capable. if SYN has 8728 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8729 * segment to be acked (eventually) continue processing rest of 8730 * data/controls, beginning with URG 8731 */ 8732 if ((thflags & TH_ACK) && 8733 (SEQ_LEQ(th->th_ack, tp->iss) || 8734 SEQ_GT(th->th_ack, tp->snd_max))) { 8735 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8736 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8737 return (1); 8738 } 8739 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8740 TCP_PROBE5(connect__refused, NULL, tp, 8741 mtod(m, const char *), tp, th); 8742 tp = tcp_drop(tp, ECONNREFUSED); 8743 ctf_do_drop(m, tp); 8744 return (1); 8745 } 8746 if (thflags & TH_RST) { 8747 ctf_do_drop(m, tp); 8748 return (1); 8749 } 8750 if (!(thflags & TH_SYN)) { 8751 ctf_do_drop(m, tp); 8752 return (1); 8753 } 8754 tp->irs = th->th_seq; 8755 tcp_rcvseqinit(tp); 8756 if (thflags & TH_ACK) { 8757 int tfo_partial = 0; 8758 8759 KMOD_TCPSTAT_INC(tcps_connects); 8760 soisconnected(so); 8761 #ifdef MAC 8762 mac_socketpeer_set_from_mbuf(m, so); 8763 #endif 8764 /* Do window scaling on this connection? */ 8765 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8766 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8767 tp->rcv_scale = tp->request_r_scale; 8768 } 8769 tp->rcv_adv += min(tp->rcv_wnd, 8770 TCP_MAXWIN << tp->rcv_scale); 8771 /* 8772 * If not all the data that was sent in the TFO SYN 8773 * has been acked, resend the remainder right away. 8774 */ 8775 if (IS_FASTOPEN(tp->t_flags) && 8776 (tp->snd_una != tp->snd_max)) { 8777 tp->snd_nxt = th->th_ack; 8778 tfo_partial = 1; 8779 } 8780 /* 8781 * If there's data, delay ACK; if there's also a FIN ACKNOW 8782 * will be turned on later. 8783 */ 8784 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8785 bbr->bbr_segs_rcvd += 1; 8786 tp->t_flags |= TF_DELACK; 8787 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8788 } else { 8789 bbr->r_wanted_output = 1; 8790 tp->t_flags |= TF_ACKNOW; 8791 } 8792 if (SEQ_GT(th->th_ack, tp->iss)) { 8793 /* 8794 * The SYN is acked 8795 * handle it specially. 8796 */ 8797 bbr_log_syn(tp, to); 8798 } 8799 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8800 /* 8801 * We advance snd_una for the 8802 * fast open case. If th_ack is 8803 * acknowledging data beyond 8804 * snd_una we can't just call 8805 * ack-processing since the 8806 * data stream in our send-map 8807 * will start at snd_una + 1 (one 8808 * beyond the SYN). If its just 8809 * equal we don't need to do that 8810 * and there is no send_map. 8811 */ 8812 tp->snd_una++; 8813 } 8814 /* 8815 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8816 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8817 */ 8818 tp->t_starttime = ticks; 8819 if (tp->t_flags & TF_NEEDFIN) { 8820 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8821 tp->t_flags &= ~TF_NEEDFIN; 8822 thflags &= ~TH_SYN; 8823 } else { 8824 tcp_state_change(tp, TCPS_ESTABLISHED); 8825 TCP_PROBE5(connect__established, NULL, tp, 8826 mtod(m, const char *), tp, th); 8827 cc_conn_init(tp); 8828 } 8829 } else { 8830 /* 8831 * Received initial SYN in SYN-SENT[*] state => simultaneous 8832 * open. If segment contains CC option and there is a 8833 * cached CC, apply TAO test. If it succeeds, connection is * 8834 * half-synchronized. Otherwise, do 3-way handshake: 8835 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8836 * there was no CC option, clear cached CC value. 8837 */ 8838 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); 8839 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8840 } 8841 /* 8842 * Advance th->th_seq to correspond to first data byte. If data, 8843 * trim to stay within window, dropping FIN if necessary. 8844 */ 8845 th->th_seq++; 8846 if (tlen > tp->rcv_wnd) { 8847 todrop = tlen - tp->rcv_wnd; 8848 m_adj(m, -todrop); 8849 tlen = tp->rcv_wnd; 8850 thflags &= ~TH_FIN; 8851 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8852 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8853 } 8854 tp->snd_wl1 = th->th_seq - 1; 8855 tp->rcv_up = th->th_seq; 8856 /* 8857 * Client side of transaction: already sent SYN and data. If the 8858 * remote host used T/TCP to validate the SYN, our data will be 8859 * ACK'd; if so, enter normal data segment processing in the middle 8860 * of step 5, ack processing. Otherwise, goto step 6. 8861 */ 8862 if (thflags & TH_ACK) { 8863 if ((to->to_flags & TOF_TS) != 0) { 8864 uint32_t t, rtt; 8865 8866 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8867 if (TSTMP_GEQ(t, to->to_tsecr)) { 8868 rtt = t - to->to_tsecr; 8869 if (rtt == 0) { 8870 rtt = 1; 8871 } 8872 rtt *= MS_IN_USEC; 8873 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8874 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8875 rtt, bbr->r_ctl.rc_rcvtime); 8876 } 8877 } 8878 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8879 return (ret_val); 8880 /* We may have changed to FIN_WAIT_1 above */ 8881 if (tp->t_state == TCPS_FIN_WAIT_1) { 8882 /* 8883 * In FIN_WAIT_1 STATE in addition to the processing 8884 * for the ESTABLISHED state if our FIN is now 8885 * acknowledged then enter FIN_WAIT_2. 8886 */ 8887 if (ourfinisacked) { 8888 /* 8889 * If we can't receive any more data, then 8890 * closing user can proceed. Starting the 8891 * timer is contrary to the specification, 8892 * but if we don't get a FIN we'll hang 8893 * forever. 8894 * 8895 * XXXjl: we should release the tp also, and 8896 * use a compressed state. 8897 */ 8898 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8899 soisdisconnected(so); 8900 tcp_timer_activate(tp, TT_2MSL, 8901 (tcp_fast_finwait2_recycle ? 8902 tcp_finwait2_timeout : 8903 TP_MAXIDLE(tp))); 8904 } 8905 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8906 } 8907 } 8908 } 8909 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8910 tiwin, thflags, nxt_pkt)); 8911 } 8912 8913 /* 8914 * Return value of 1, the TCB is unlocked and most 8915 * likely gone, return value of 0, the TCB is still 8916 * locked. 8917 */ 8918 static int 8919 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8920 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8921 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8922 { 8923 int32_t ourfinisacked = 0; 8924 int32_t ret_val; 8925 struct tcp_bbr *bbr; 8926 8927 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8928 8929 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8930 ctf_calc_rwin(so, tp); 8931 if ((thflags & TH_ACK) && 8932 (SEQ_LEQ(th->th_ack, tp->snd_una) || 8933 SEQ_GT(th->th_ack, tp->snd_max))) { 8934 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8935 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8936 return (1); 8937 } 8938 if (IS_FASTOPEN(tp->t_flags)) { 8939 /* 8940 * When a TFO connection is in SYN_RECEIVED, the only valid 8941 * packets are the initial SYN, a retransmit/copy of the 8942 * initial SYN (possibly with a subset of the original 8943 * data), a valid ACK, a FIN, or a RST. 8944 */ 8945 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 8946 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8947 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8948 return (1); 8949 } else if (thflags & TH_SYN) { 8950 /* non-initial SYN is ignored */ 8951 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 8952 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 8953 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 8954 ctf_do_drop(m, NULL); 8955 return (0); 8956 } 8957 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 8958 ctf_do_drop(m, NULL); 8959 return (0); 8960 } 8961 } 8962 if ((thflags & TH_RST) || 8963 (tp->t_fin_is_rst && (thflags & TH_FIN))) 8964 return (ctf_process_rst(m, th, so, tp)); 8965 /* 8966 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 8967 * it's less than ts_recent, drop it. 8968 */ 8969 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 8970 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 8971 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 8972 return (ret_val); 8973 } 8974 /* 8975 * In the SYN-RECEIVED state, validate that the packet belongs to 8976 * this connection before trimming the data to fit the receive 8977 * window. Check the sequence number versus IRS since we know the 8978 * sequence numbers haven't wrapped. This is a partial fix for the 8979 * "LAND" DoS attack. 8980 */ 8981 if (SEQ_LT(th->th_seq, tp->irs)) { 8982 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8983 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8984 return (1); 8985 } 8986 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 8987 return (ret_val); 8988 } 8989 /* 8990 * If last ACK falls within this segment's sequence numbers, record 8991 * its timestamp. NOTE: 1) That the test incorporates suggestions 8992 * from the latest proposal of the tcplw@cray.com list (Braden 8993 * 1993/04/26). 2) That updating only on newer timestamps interferes 8994 * with our earlier PAWS tests, so this check should be solely 8995 * predicated on the sequence space of this segment. 3) That we 8996 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 8997 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 8998 * SEG.Len, This modified check allows us to overcome RFC1323's 8999 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9000 * p.869. In such cases, we can still calculate the RTT correctly 9001 * when RCV.NXT == Last.ACK.Sent. 9002 */ 9003 if ((to->to_flags & TOF_TS) != 0 && 9004 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9005 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9006 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9007 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9008 tp->ts_recent = to->to_tsval; 9009 } 9010 tp->snd_wnd = tiwin; 9011 /* 9012 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9013 * is on (half-synchronized state), then queue data for later 9014 * processing; else drop segment and return. 9015 */ 9016 if ((thflags & TH_ACK) == 0) { 9017 if (IS_FASTOPEN(tp->t_flags)) { 9018 cc_conn_init(tp); 9019 } 9020 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9021 tiwin, thflags, nxt_pkt)); 9022 } 9023 KMOD_TCPSTAT_INC(tcps_connects); 9024 if (tp->t_flags & TF_SONOTCONN) { 9025 tp->t_flags &= ~TF_SONOTCONN; 9026 soisconnected(so); 9027 } 9028 /* Do window scaling? */ 9029 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9030 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9031 tp->rcv_scale = tp->request_r_scale; 9032 } 9033 /* 9034 * ok for the first time in lets see if we can use the ts to figure 9035 * out what the initial RTT was. 9036 */ 9037 if ((to->to_flags & TOF_TS) != 0) { 9038 uint32_t t, rtt; 9039 9040 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9041 if (TSTMP_GEQ(t, to->to_tsecr)) { 9042 rtt = t - to->to_tsecr; 9043 if (rtt == 0) { 9044 rtt = 1; 9045 } 9046 rtt *= MS_IN_USEC; 9047 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9048 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9049 } 9050 } 9051 /* Drop off any SYN in the send map (probably not there) */ 9052 if (thflags & TH_ACK) 9053 bbr_log_syn(tp, to); 9054 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 9055 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9056 tp->t_tfo_pending = NULL; 9057 } 9058 /* 9059 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9060 * FIN-WAIT-1 9061 */ 9062 tp->t_starttime = ticks; 9063 if (tp->t_flags & TF_NEEDFIN) { 9064 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9065 tp->t_flags &= ~TF_NEEDFIN; 9066 } else { 9067 tcp_state_change(tp, TCPS_ESTABLISHED); 9068 TCP_PROBE5(accept__established, NULL, tp, 9069 mtod(m, const char *), tp, th); 9070 /* 9071 * TFO connections call cc_conn_init() during SYN 9072 * processing. Calling it again here for such connections 9073 * is not harmless as it would undo the snd_cwnd reduction 9074 * that occurs when a TFO SYN|ACK is retransmitted. 9075 */ 9076 if (!IS_FASTOPEN(tp->t_flags)) 9077 cc_conn_init(tp); 9078 } 9079 /* 9080 * Account for the ACK of our SYN prior to 9081 * regular ACK processing below, except for 9082 * simultaneous SYN, which is handled later. 9083 */ 9084 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9085 tp->snd_una++; 9086 /* 9087 * If segment contains data or ACK, will call tcp_reass() later; if 9088 * not, do so now to pass queued data to user. 9089 */ 9090 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9091 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9092 (struct mbuf *)0); 9093 if (tp->t_flags & TF_WAKESOR) { 9094 tp->t_flags &= ~TF_WAKESOR; 9095 /* NB: sorwakeup_locked() does an implicit unlock. */ 9096 sorwakeup_locked(so); 9097 } 9098 } 9099 tp->snd_wl1 = th->th_seq - 1; 9100 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9101 return (ret_val); 9102 } 9103 if (tp->t_state == TCPS_FIN_WAIT_1) { 9104 /* We could have went to FIN_WAIT_1 (or EST) above */ 9105 /* 9106 * In FIN_WAIT_1 STATE in addition to the processing for the 9107 * ESTABLISHED state if our FIN is now acknowledged then 9108 * enter FIN_WAIT_2. 9109 */ 9110 if (ourfinisacked) { 9111 /* 9112 * If we can't receive any more data, then closing 9113 * user can proceed. Starting the timer is contrary 9114 * to the specification, but if we don't get a FIN 9115 * we'll hang forever. 9116 * 9117 * XXXjl: we should release the tp also, and use a 9118 * compressed state. 9119 */ 9120 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9121 soisdisconnected(so); 9122 tcp_timer_activate(tp, TT_2MSL, 9123 (tcp_fast_finwait2_recycle ? 9124 tcp_finwait2_timeout : 9125 TP_MAXIDLE(tp))); 9126 } 9127 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9128 } 9129 } 9130 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9131 tiwin, thflags, nxt_pkt)); 9132 } 9133 9134 /* 9135 * Return value of 1, the TCB is unlocked and most 9136 * likely gone, return value of 0, the TCB is still 9137 * locked. 9138 */ 9139 static int 9140 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9141 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9142 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9143 { 9144 struct tcp_bbr *bbr; 9145 int32_t ret_val; 9146 9147 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9148 9149 /* 9150 * Header prediction: check for the two common cases of a 9151 * uni-directional data xfer. If the packet has no control flags, 9152 * is in-sequence, the window didn't change and we're not 9153 * retransmitting, it's a candidate. If the length is zero and the 9154 * ack moved forward, we're the sender side of the xfer. Just free 9155 * the data acked & wake any higher level process that was blocked 9156 * waiting for space. If the length is non-zero and the ack didn't 9157 * move, we're the receiver side. If we're getting packets in-order 9158 * (the reassembly queue is empty), add the data toc The socket 9159 * buffer and note that we need a delayed ack. Make sure that the 9160 * hidden state-flags are also off. Since we check for 9161 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9162 */ 9163 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9164 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9165 /* 9166 * If we have delived under 4 segments increase the initial 9167 * window if raised by the peer. We use this to determine 9168 * dynamic and static rwnd's at the end of a connection. 9169 */ 9170 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9171 } 9172 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9173 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9174 __predict_true(SEGQ_EMPTY(tp)) && 9175 __predict_true(th->th_seq == tp->rcv_nxt)) { 9176 if (tlen == 0) { 9177 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9178 tiwin, nxt_pkt, iptos)) { 9179 return (0); 9180 } 9181 } else { 9182 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9183 tiwin, nxt_pkt)) { 9184 return (0); 9185 } 9186 } 9187 } 9188 ctf_calc_rwin(so, tp); 9189 9190 if ((thflags & TH_RST) || 9191 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9192 return (ctf_process_rst(m, th, so, tp)); 9193 /* 9194 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9195 * synchronized state. 9196 */ 9197 if (thflags & TH_SYN) { 9198 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9199 return (ret_val); 9200 } 9201 /* 9202 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9203 * it's less than ts_recent, drop it. 9204 */ 9205 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9206 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9207 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9208 return (ret_val); 9209 } 9210 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9211 return (ret_val); 9212 } 9213 /* 9214 * If last ACK falls within this segment's sequence numbers, record 9215 * its timestamp. NOTE: 1) That the test incorporates suggestions 9216 * from the latest proposal of the tcplw@cray.com list (Braden 9217 * 1993/04/26). 2) That updating only on newer timestamps interferes 9218 * with our earlier PAWS tests, so this check should be solely 9219 * predicated on the sequence space of this segment. 3) That we 9220 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9221 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9222 * SEG.Len, This modified check allows us to overcome RFC1323's 9223 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9224 * p.869. In such cases, we can still calculate the RTT correctly 9225 * when RCV.NXT == Last.ACK.Sent. 9226 */ 9227 if ((to->to_flags & TOF_TS) != 0 && 9228 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9229 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9230 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9231 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9232 tp->ts_recent = to->to_tsval; 9233 } 9234 /* 9235 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9236 * is on (half-synchronized state), then queue data for later 9237 * processing; else drop segment and return. 9238 */ 9239 if ((thflags & TH_ACK) == 0) { 9240 if (tp->t_flags & TF_NEEDSYN) { 9241 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9242 tiwin, thflags, nxt_pkt)); 9243 } else if (tp->t_flags & TF_ACKNOW) { 9244 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9245 bbr->r_wanted_output = 1; 9246 return (ret_val); 9247 } else { 9248 ctf_do_drop(m, NULL); 9249 return (0); 9250 } 9251 } 9252 /* 9253 * Ack processing. 9254 */ 9255 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9256 return (ret_val); 9257 } 9258 if (sbavail(&so->so_snd)) { 9259 if (ctf_progress_timeout_check(tp, true)) { 9260 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9261 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9262 return (1); 9263 } 9264 } 9265 /* State changes only happen in bbr_process_data() */ 9266 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9267 tiwin, thflags, nxt_pkt)); 9268 } 9269 9270 /* 9271 * Return value of 1, the TCB is unlocked and most 9272 * likely gone, return value of 0, the TCB is still 9273 * locked. 9274 */ 9275 static int 9276 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9277 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9278 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9279 { 9280 struct tcp_bbr *bbr; 9281 int32_t ret_val; 9282 9283 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9284 9285 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9286 ctf_calc_rwin(so, tp); 9287 if ((thflags & TH_RST) || 9288 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9289 return (ctf_process_rst(m, th, so, tp)); 9290 /* 9291 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9292 * synchronized state. 9293 */ 9294 if (thflags & TH_SYN) { 9295 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9296 return (ret_val); 9297 } 9298 /* 9299 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9300 * it's less than ts_recent, drop it. 9301 */ 9302 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9303 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9304 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9305 return (ret_val); 9306 } 9307 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9308 return (ret_val); 9309 } 9310 /* 9311 * If last ACK falls within this segment's sequence numbers, record 9312 * its timestamp. NOTE: 1) That the test incorporates suggestions 9313 * from the latest proposal of the tcplw@cray.com list (Braden 9314 * 1993/04/26). 2) That updating only on newer timestamps interferes 9315 * with our earlier PAWS tests, so this check should be solely 9316 * predicated on the sequence space of this segment. 3) That we 9317 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9318 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9319 * SEG.Len, This modified check allows us to overcome RFC1323's 9320 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9321 * p.869. In such cases, we can still calculate the RTT correctly 9322 * when RCV.NXT == Last.ACK.Sent. 9323 */ 9324 if ((to->to_flags & TOF_TS) != 0 && 9325 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9326 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9327 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9328 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9329 tp->ts_recent = to->to_tsval; 9330 } 9331 /* 9332 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9333 * is on (half-synchronized state), then queue data for later 9334 * processing; else drop segment and return. 9335 */ 9336 if ((thflags & TH_ACK) == 0) { 9337 if (tp->t_flags & TF_NEEDSYN) { 9338 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9339 tiwin, thflags, nxt_pkt)); 9340 } else if (tp->t_flags & TF_ACKNOW) { 9341 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9342 bbr->r_wanted_output = 1; 9343 return (ret_val); 9344 } else { 9345 ctf_do_drop(m, NULL); 9346 return (0); 9347 } 9348 } 9349 /* 9350 * Ack processing. 9351 */ 9352 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9353 return (ret_val); 9354 } 9355 if (sbavail(&so->so_snd)) { 9356 if (ctf_progress_timeout_check(tp, true)) { 9357 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9358 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9359 return (1); 9360 } 9361 } 9362 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9363 tiwin, thflags, nxt_pkt)); 9364 } 9365 9366 static int 9367 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9368 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9369 { 9370 9371 if (bbr->rc_allow_data_af_clo == 0) { 9372 close_now: 9373 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9374 /* tcp_close will kill the inp pre-log the Reset */ 9375 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9376 tp = tcp_close(tp); 9377 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9378 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9379 return (1); 9380 } 9381 if (sbavail(&so->so_snd) == 0) 9382 goto close_now; 9383 /* Ok we allow data that is ignored and a followup reset */ 9384 tp->rcv_nxt = th->th_seq + *tlen; 9385 tp->t_flags2 |= TF2_DROP_AF_DATA; 9386 bbr->r_wanted_output = 1; 9387 *tlen = 0; 9388 return (0); 9389 } 9390 9391 /* 9392 * Return value of 1, the TCB is unlocked and most 9393 * likely gone, return value of 0, the TCB is still 9394 * locked. 9395 */ 9396 static int 9397 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9398 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9399 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9400 { 9401 int32_t ourfinisacked = 0; 9402 int32_t ret_val; 9403 struct tcp_bbr *bbr; 9404 9405 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9406 9407 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9408 ctf_calc_rwin(so, tp); 9409 if ((thflags & TH_RST) || 9410 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9411 return (ctf_process_rst(m, th, so, tp)); 9412 /* 9413 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9414 * synchronized state. 9415 */ 9416 if (thflags & TH_SYN) { 9417 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9418 return (ret_val); 9419 } 9420 /* 9421 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9422 * it's less than ts_recent, drop it. 9423 */ 9424 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9425 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9426 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9427 return (ret_val); 9428 } 9429 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9430 return (ret_val); 9431 } 9432 /* 9433 * If new data are received on a connection after the user processes 9434 * are gone, then RST the other end. 9435 * We call a new function now so we might continue and setup 9436 * to reset at all data being ack'd. 9437 */ 9438 if ((tp->t_flags & TF_CLOSED) && tlen && 9439 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9440 return (1); 9441 /* 9442 * If last ACK falls within this segment's sequence numbers, record 9443 * its timestamp. NOTE: 1) That the test incorporates suggestions 9444 * from the latest proposal of the tcplw@cray.com list (Braden 9445 * 1993/04/26). 2) That updating only on newer timestamps interferes 9446 * with our earlier PAWS tests, so this check should be solely 9447 * predicated on the sequence space of this segment. 3) That we 9448 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9449 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9450 * SEG.Len, This modified check allows us to overcome RFC1323's 9451 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9452 * p.869. In such cases, we can still calculate the RTT correctly 9453 * when RCV.NXT == Last.ACK.Sent. 9454 */ 9455 if ((to->to_flags & TOF_TS) != 0 && 9456 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9457 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9458 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9459 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9460 tp->ts_recent = to->to_tsval; 9461 } 9462 /* 9463 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9464 * is on (half-synchronized state), then queue data for later 9465 * processing; else drop segment and return. 9466 */ 9467 if ((thflags & TH_ACK) == 0) { 9468 if (tp->t_flags & TF_NEEDSYN) { 9469 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9470 tiwin, thflags, nxt_pkt)); 9471 } else if (tp->t_flags & TF_ACKNOW) { 9472 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9473 bbr->r_wanted_output = 1; 9474 return (ret_val); 9475 } else { 9476 ctf_do_drop(m, NULL); 9477 return (0); 9478 } 9479 } 9480 /* 9481 * Ack processing. 9482 */ 9483 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9484 return (ret_val); 9485 } 9486 if (ourfinisacked) { 9487 /* 9488 * If we can't receive any more data, then closing user can 9489 * proceed. Starting the timer is contrary to the 9490 * specification, but if we don't get a FIN we'll hang 9491 * forever. 9492 * 9493 * XXXjl: we should release the tp also, and use a 9494 * compressed state. 9495 */ 9496 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9497 soisdisconnected(so); 9498 tcp_timer_activate(tp, TT_2MSL, 9499 (tcp_fast_finwait2_recycle ? 9500 tcp_finwait2_timeout : 9501 TP_MAXIDLE(tp))); 9502 } 9503 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9504 } 9505 if (sbavail(&so->so_snd)) { 9506 if (ctf_progress_timeout_check(tp, true)) { 9507 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9508 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9509 return (1); 9510 } 9511 } 9512 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9513 tiwin, thflags, nxt_pkt)); 9514 } 9515 9516 /* 9517 * Return value of 1, the TCB is unlocked and most 9518 * likely gone, return value of 0, the TCB is still 9519 * locked. 9520 */ 9521 static int 9522 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9523 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9524 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9525 { 9526 int32_t ourfinisacked = 0; 9527 int32_t ret_val; 9528 struct tcp_bbr *bbr; 9529 9530 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9531 9532 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9533 ctf_calc_rwin(so, tp); 9534 if ((thflags & TH_RST) || 9535 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9536 return (ctf_process_rst(m, th, so, tp)); 9537 /* 9538 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9539 * synchronized state. 9540 */ 9541 if (thflags & TH_SYN) { 9542 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9543 return (ret_val); 9544 } 9545 /* 9546 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9547 * it's less than ts_recent, drop it. 9548 */ 9549 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9550 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9551 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9552 return (ret_val); 9553 } 9554 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9555 return (ret_val); 9556 } 9557 /* 9558 * If new data are received on a connection after the user processes 9559 * are gone, then RST the other end. 9560 * We call a new function now so we might continue and setup 9561 * to reset at all data being ack'd. 9562 */ 9563 if ((tp->t_flags & TF_CLOSED) && tlen && 9564 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9565 return (1); 9566 /* 9567 * If last ACK falls within this segment's sequence numbers, record 9568 * its timestamp. NOTE: 1) That the test incorporates suggestions 9569 * from the latest proposal of the tcplw@cray.com list (Braden 9570 * 1993/04/26). 2) That updating only on newer timestamps interferes 9571 * with our earlier PAWS tests, so this check should be solely 9572 * predicated on the sequence space of this segment. 3) That we 9573 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9574 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9575 * SEG.Len, This modified check allows us to overcome RFC1323's 9576 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9577 * p.869. In such cases, we can still calculate the RTT correctly 9578 * when RCV.NXT == Last.ACK.Sent. 9579 */ 9580 if ((to->to_flags & TOF_TS) != 0 && 9581 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9582 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9583 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9584 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9585 tp->ts_recent = to->to_tsval; 9586 } 9587 /* 9588 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9589 * is on (half-synchronized state), then queue data for later 9590 * processing; else drop segment and return. 9591 */ 9592 if ((thflags & TH_ACK) == 0) { 9593 if (tp->t_flags & TF_NEEDSYN) { 9594 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9595 tiwin, thflags, nxt_pkt)); 9596 } else if (tp->t_flags & TF_ACKNOW) { 9597 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9598 bbr->r_wanted_output = 1; 9599 return (ret_val); 9600 } else { 9601 ctf_do_drop(m, NULL); 9602 return (0); 9603 } 9604 } 9605 /* 9606 * Ack processing. 9607 */ 9608 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9609 return (ret_val); 9610 } 9611 if (ourfinisacked) { 9612 tcp_twstart(tp); 9613 m_freem(m); 9614 return (1); 9615 } 9616 if (sbavail(&so->so_snd)) { 9617 if (ctf_progress_timeout_check(tp, true)) { 9618 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9619 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9620 return (1); 9621 } 9622 } 9623 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9624 tiwin, thflags, nxt_pkt)); 9625 } 9626 9627 /* 9628 * Return value of 1, the TCB is unlocked and most 9629 * likely gone, return value of 0, the TCB is still 9630 * locked. 9631 */ 9632 static int 9633 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9634 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9635 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9636 { 9637 int32_t ourfinisacked = 0; 9638 int32_t ret_val; 9639 struct tcp_bbr *bbr; 9640 9641 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9642 9643 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9644 ctf_calc_rwin(so, tp); 9645 if ((thflags & TH_RST) || 9646 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9647 return (ctf_process_rst(m, th, so, tp)); 9648 /* 9649 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9650 * synchronized state. 9651 */ 9652 if (thflags & TH_SYN) { 9653 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9654 return (ret_val); 9655 } 9656 /* 9657 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9658 * it's less than ts_recent, drop it. 9659 */ 9660 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9661 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9662 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9663 return (ret_val); 9664 } 9665 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9666 return (ret_val); 9667 } 9668 /* 9669 * If new data are received on a connection after the user processes 9670 * are gone, then RST the other end. 9671 * We call a new function now so we might continue and setup 9672 * to reset at all data being ack'd. 9673 */ 9674 if ((tp->t_flags & TF_CLOSED) && tlen && 9675 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9676 return (1); 9677 /* 9678 * If last ACK falls within this segment's sequence numbers, record 9679 * its timestamp. NOTE: 1) That the test incorporates suggestions 9680 * from the latest proposal of the tcplw@cray.com list (Braden 9681 * 1993/04/26). 2) That updating only on newer timestamps interferes 9682 * with our earlier PAWS tests, so this check should be solely 9683 * predicated on the sequence space of this segment. 3) That we 9684 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9685 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9686 * SEG.Len, This modified check allows us to overcome RFC1323's 9687 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9688 * p.869. In such cases, we can still calculate the RTT correctly 9689 * when RCV.NXT == Last.ACK.Sent. 9690 */ 9691 if ((to->to_flags & TOF_TS) != 0 && 9692 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9693 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9694 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9695 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9696 tp->ts_recent = to->to_tsval; 9697 } 9698 /* 9699 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9700 * is on (half-synchronized state), then queue data for later 9701 * processing; else drop segment and return. 9702 */ 9703 if ((thflags & TH_ACK) == 0) { 9704 if (tp->t_flags & TF_NEEDSYN) { 9705 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9706 tiwin, thflags, nxt_pkt)); 9707 } else if (tp->t_flags & TF_ACKNOW) { 9708 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9709 bbr->r_wanted_output = 1; 9710 return (ret_val); 9711 } else { 9712 ctf_do_drop(m, NULL); 9713 return (0); 9714 } 9715 } 9716 /* 9717 * case TCPS_LAST_ACK: Ack processing. 9718 */ 9719 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9720 return (ret_val); 9721 } 9722 if (ourfinisacked) { 9723 tp = tcp_close(tp); 9724 ctf_do_drop(m, tp); 9725 return (1); 9726 } 9727 if (sbavail(&so->so_snd)) { 9728 if (ctf_progress_timeout_check(tp, true)) { 9729 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9730 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9731 return (1); 9732 } 9733 } 9734 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9735 tiwin, thflags, nxt_pkt)); 9736 } 9737 9738 /* 9739 * Return value of 1, the TCB is unlocked and most 9740 * likely gone, return value of 0, the TCB is still 9741 * locked. 9742 */ 9743 static int 9744 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9745 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9746 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9747 { 9748 int32_t ourfinisacked = 0; 9749 int32_t ret_val; 9750 struct tcp_bbr *bbr; 9751 9752 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9753 9754 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9755 ctf_calc_rwin(so, tp); 9756 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9757 if ((thflags & TH_RST) || 9758 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9759 return (ctf_process_rst(m, th, so, tp)); 9760 9761 /* 9762 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9763 * synchronized state. 9764 */ 9765 if (thflags & TH_SYN) { 9766 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9767 return (ret_val); 9768 } 9769 /* 9770 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9771 * it's less than ts_recent, drop it. 9772 */ 9773 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9774 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9775 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9776 return (ret_val); 9777 } 9778 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9779 return (ret_val); 9780 } 9781 /* 9782 * If new data are received on a connection after the user processes 9783 * are gone, then we may RST the other end depending on the outcome 9784 * of bbr_check_data_after_close. 9785 * We call a new function now so we might continue and setup 9786 * to reset at all data being ack'd. 9787 */ 9788 if ((tp->t_flags & TF_CLOSED) && tlen && 9789 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9790 return (1); 9791 /* 9792 * If last ACK falls within this segment's sequence numbers, record 9793 * its timestamp. NOTE: 1) That the test incorporates suggestions 9794 * from the latest proposal of the tcplw@cray.com list (Braden 9795 * 1993/04/26). 2) That updating only on newer timestamps interferes 9796 * with our earlier PAWS tests, so this check should be solely 9797 * predicated on the sequence space of this segment. 3) That we 9798 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9799 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9800 * SEG.Len, This modified check allows us to overcome RFC1323's 9801 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9802 * p.869. In such cases, we can still calculate the RTT correctly 9803 * when RCV.NXT == Last.ACK.Sent. 9804 */ 9805 if ((to->to_flags & TOF_TS) != 0 && 9806 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9807 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9808 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9809 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9810 tp->ts_recent = to->to_tsval; 9811 } 9812 /* 9813 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9814 * is on (half-synchronized state), then queue data for later 9815 * processing; else drop segment and return. 9816 */ 9817 if ((thflags & TH_ACK) == 0) { 9818 if (tp->t_flags & TF_NEEDSYN) { 9819 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9820 tiwin, thflags, nxt_pkt)); 9821 } else if (tp->t_flags & TF_ACKNOW) { 9822 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9823 bbr->r_wanted_output = 1; 9824 return (ret_val); 9825 } else { 9826 ctf_do_drop(m, NULL); 9827 return (0); 9828 } 9829 } 9830 /* 9831 * Ack processing. 9832 */ 9833 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9834 return (ret_val); 9835 } 9836 if (sbavail(&so->so_snd)) { 9837 if (ctf_progress_timeout_check(tp, true)) { 9838 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9839 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9840 return (1); 9841 } 9842 } 9843 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9844 tiwin, thflags, nxt_pkt)); 9845 } 9846 9847 static void 9848 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr) 9849 { 9850 /* 9851 * Assure no timers are running. 9852 */ 9853 if (tcp_timer_active(tp, TT_PERSIST)) { 9854 /* We enter in persists, set the flag appropriately */ 9855 bbr->rc_in_persist = 1; 9856 } 9857 if (tcp_in_hpts(bbr->rc_tp)) { 9858 tcp_hpts_remove(bbr->rc_tp); 9859 } 9860 } 9861 9862 static void 9863 bbr_google_mode_on(struct tcp_bbr *bbr) 9864 { 9865 bbr->rc_use_google = 1; 9866 bbr->rc_no_pacing = 0; 9867 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9868 bbr->r_use_policer = bbr_policer_detection_enabled; 9869 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9870 bbr->bbr_use_rack_cheat = 0; 9871 bbr->r_ctl.rc_incr_tmrs = 0; 9872 bbr->r_ctl.rc_inc_tcp_oh = 0; 9873 bbr->r_ctl.rc_inc_ip_oh = 0; 9874 bbr->r_ctl.rc_inc_enet_oh = 0; 9875 reset_time(&bbr->r_ctl.rc_delrate, 9876 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9877 reset_time_small(&bbr->r_ctl.rc_rttprop, 9878 (11 * USECS_IN_SECOND)); 9879 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9880 } 9881 9882 static void 9883 bbr_google_mode_off(struct tcp_bbr *bbr) 9884 { 9885 bbr->rc_use_google = 0; 9886 bbr->r_ctl.bbr_google_discount = 0; 9887 bbr->no_pacing_until = bbr_no_pacing_until; 9888 bbr->r_use_policer = 0; 9889 if (bbr->no_pacing_until) 9890 bbr->rc_no_pacing = 1; 9891 else 9892 bbr->rc_no_pacing = 0; 9893 if (bbr_use_rack_resend_cheat) 9894 bbr->bbr_use_rack_cheat = 1; 9895 else 9896 bbr->bbr_use_rack_cheat = 0; 9897 if (bbr_incr_timers) 9898 bbr->r_ctl.rc_incr_tmrs = 1; 9899 else 9900 bbr->r_ctl.rc_incr_tmrs = 0; 9901 if (bbr_include_tcp_oh) 9902 bbr->r_ctl.rc_inc_tcp_oh = 1; 9903 else 9904 bbr->r_ctl.rc_inc_tcp_oh = 0; 9905 if (bbr_include_ip_oh) 9906 bbr->r_ctl.rc_inc_ip_oh = 1; 9907 else 9908 bbr->r_ctl.rc_inc_ip_oh = 0; 9909 if (bbr_include_enet_oh) 9910 bbr->r_ctl.rc_inc_enet_oh = 1; 9911 else 9912 bbr->r_ctl.rc_inc_enet_oh = 0; 9913 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9914 reset_time(&bbr->r_ctl.rc_delrate, 9915 bbr_num_pktepo_for_del_limit); 9916 reset_time_small(&bbr->r_ctl.rc_rttprop, 9917 (bbr_filter_len_sec * USECS_IN_SECOND)); 9918 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9919 } 9920 /* 9921 * Return 0 on success, non-zero on failure 9922 * which indicates the error (usually no memory). 9923 */ 9924 static int 9925 bbr_init(struct tcpcb *tp, void **ptr) 9926 { 9927 struct inpcb *inp = tptoinpcb(tp); 9928 struct tcp_bbr *bbr = NULL; 9929 uint32_t cts; 9930 9931 *ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 9932 if (*ptr == NULL) { 9933 /* 9934 * We need to allocate memory but cant. The INP and INP_INFO 9935 * locks and they are recursive (happens during setup. So a 9936 * scheme to drop the locks fails :( 9937 * 9938 */ 9939 return (ENOMEM); 9940 } 9941 bbr = (struct tcp_bbr *)*ptr; 9942 bbr->rtt_valid = 0; 9943 tp->t_flags2 |= TF2_CANNOT_DO_ECN; 9944 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ; 9945 /* Take off any undesired flags */ 9946 tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY; 9947 tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE; 9948 tp->t_flags2 &= ~TF2_MBUF_ACKCMP; 9949 tp->t_flags2 &= ~TF2_MBUF_L_ACKS; 9950 9951 TAILQ_INIT(&bbr->r_ctl.rc_map); 9952 TAILQ_INIT(&bbr->r_ctl.rc_free); 9953 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 9954 bbr->rc_tp = tp; 9955 bbr->rc_inp = inp; 9956 cts = tcp_get_usecs(&bbr->rc_tv); 9957 tp->t_acktime = 0; 9958 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 9959 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 9960 bbr->rc_tlp_threshold = bbr_tlp_thresh; 9961 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 9962 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 9963 bbr->r_ctl.rc_min_to = bbr_min_to; 9964 bbr->rc_bbr_state = BBR_STATE_STARTUP; 9965 bbr->r_ctl.bbr_lost_at_state = 0; 9966 bbr->r_ctl.rc_lost_at_startup = 0; 9967 bbr->rc_all_timers_stopped = 0; 9968 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 9969 bbr->r_ctl.rc_pkt_epoch_del = 0; 9970 bbr->r_ctl.rc_pkt_epoch = 0; 9971 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 9972 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 9973 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 9974 bbr->r_ctl.rc_went_idle_time = cts; 9975 bbr->rc_pacer_started = cts; 9976 bbr->r_ctl.rc_pkt_epoch_time = cts; 9977 bbr->r_ctl.rc_rcvtime = cts; 9978 bbr->r_ctl.rc_bbr_state_time = cts; 9979 bbr->r_ctl.rc_del_time = cts; 9980 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 9981 bbr->r_ctl.last_in_probertt = cts; 9982 bbr->skip_gain = 0; 9983 bbr->gain_is_limited = 0; 9984 bbr->no_pacing_until = bbr_no_pacing_until; 9985 if (bbr->no_pacing_until) 9986 bbr->rc_no_pacing = 1; 9987 if (bbr_use_google_algo) { 9988 bbr->rc_no_pacing = 0; 9989 bbr->rc_use_google = 1; 9990 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9991 bbr->r_use_policer = bbr_policer_detection_enabled; 9992 } else { 9993 bbr->rc_use_google = 0; 9994 bbr->r_ctl.bbr_google_discount = 0; 9995 bbr->r_use_policer = 0; 9996 } 9997 if (bbr_ts_limiting) 9998 bbr->rc_use_ts_limit = 1; 9999 else 10000 bbr->rc_use_ts_limit = 0; 10001 if (bbr_ts_can_raise) 10002 bbr->ts_can_raise = 1; 10003 else 10004 bbr->ts_can_raise = 0; 10005 if (V_tcp_delack_enabled == 1) 10006 tp->t_delayed_ack = 2; 10007 else if (V_tcp_delack_enabled == 0) 10008 tp->t_delayed_ack = 0; 10009 else if (V_tcp_delack_enabled < 100) 10010 tp->t_delayed_ack = V_tcp_delack_enabled; 10011 else 10012 tp->t_delayed_ack = 2; 10013 if (bbr->rc_use_google == 0) 10014 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10015 else 10016 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10017 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10018 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10019 bbr->rc_init_win = bbr_def_init_win; 10020 if (tp->t_flags & TF_REQ_TSTMP) 10021 bbr->rc_last_options = TCP_TS_OVERHEAD; 10022 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10023 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10024 bbr->r_init_rtt = 1; 10025 10026 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10027 if (bbr_allow_hdwr_pacing) 10028 bbr->bbr_hdw_pace_ena = 1; 10029 else 10030 bbr->bbr_hdw_pace_ena = 0; 10031 if (bbr_sends_full_iwnd) 10032 bbr->bbr_init_win_cheat = 1; 10033 else 10034 bbr->bbr_init_win_cheat = 0; 10035 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10036 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10037 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10038 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10039 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10040 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10041 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10042 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10043 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10044 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10045 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10046 bbr->r_ctl.rc_rtt_shrinks = cts; 10047 if (bbr->rc_use_google) { 10048 setup_time_filter(&bbr->r_ctl.rc_delrate, 10049 FILTER_TYPE_MAX, 10050 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10051 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10052 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10053 } else { 10054 setup_time_filter(&bbr->r_ctl.rc_delrate, 10055 FILTER_TYPE_MAX, 10056 bbr_num_pktepo_for_del_limit); 10057 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10058 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10059 } 10060 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10061 if (bbr_uses_idle_restart) 10062 bbr->rc_use_idle_restart = 1; 10063 else 10064 bbr->rc_use_idle_restart = 0; 10065 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10066 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10067 if (bbr_resends_use_tso) 10068 bbr->rc_resends_use_tso = 1; 10069 if (tp->snd_una != tp->snd_max) { 10070 /* Create a send map for the current outstanding data */ 10071 struct bbr_sendmap *rsm; 10072 10073 rsm = bbr_alloc(bbr); 10074 if (rsm == NULL) { 10075 uma_zfree(bbr_pcb_zone, *ptr); 10076 *ptr = NULL; 10077 return (ENOMEM); 10078 } 10079 rsm->r_rtt_not_allowed = 1; 10080 rsm->r_tim_lastsent[0] = cts; 10081 rsm->r_rtr_cnt = 1; 10082 rsm->r_rtr_bytes = 0; 10083 rsm->r_start = tp->snd_una; 10084 rsm->r_end = tp->snd_max; 10085 rsm->r_dupack = 0; 10086 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10087 rsm->r_ts_valid = 0; 10088 rsm->r_del_ack_ts = tp->ts_recent; 10089 rsm->r_del_time = cts; 10090 if (bbr->r_ctl.r_app_limited_until) 10091 rsm->r_app_limited = 1; 10092 else 10093 rsm->r_app_limited = 0; 10094 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10095 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10096 rsm->r_in_tmap = 1; 10097 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10098 rsm->r_bbr_state = bbr_state_val(bbr); 10099 else 10100 rsm->r_bbr_state = 8; 10101 } 10102 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10103 bbr->bbr_use_rack_cheat = 1; 10104 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10105 bbr->r_ctl.rc_incr_tmrs = 1; 10106 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10107 bbr->r_ctl.rc_inc_tcp_oh = 1; 10108 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10109 bbr->r_ctl.rc_inc_ip_oh = 1; 10110 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10111 bbr->r_ctl.rc_inc_enet_oh = 1; 10112 10113 bbr_log_type_statechange(bbr, cts, __LINE__); 10114 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10115 (tp->t_srtt)) { 10116 uint32_t rtt; 10117 10118 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10119 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10120 } 10121 /* announce the settings and state */ 10122 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10123 tcp_bbr_tso_size_check(bbr, cts); 10124 /* 10125 * Now call the generic function to start a timer. This will place 10126 * the TCB on the hptsi wheel if a timer is needed with appropriate 10127 * flags. 10128 */ 10129 bbr_stop_all_timers(tp, bbr); 10130 /* 10131 * Validate the timers are not in usec, if they are convert. 10132 * BBR should in theory move to USEC and get rid of a 10133 * lot of the TICKS_2 calls.. but for now we stay 10134 * with tick timers. 10135 */ 10136 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS); 10137 TCPT_RANGESET(tp->t_rxtcur, 10138 ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, 10139 tp->t_rttmin, TCPTV_REXMTMAX); 10140 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10141 return (0); 10142 } 10143 10144 /* 10145 * Return 0 if we can accept the connection. Return 10146 * non-zero if we can't handle the connection. A EAGAIN 10147 * means you need to wait until the connection is up. 10148 * a EADDRNOTAVAIL means we can never handle the connection 10149 * (no SACK). 10150 */ 10151 static int 10152 bbr_handoff_ok(struct tcpcb *tp) 10153 { 10154 if ((tp->t_state == TCPS_CLOSED) || 10155 (tp->t_state == TCPS_LISTEN)) { 10156 /* Sure no problem though it may not stick */ 10157 return (0); 10158 } 10159 if ((tp->t_state == TCPS_SYN_SENT) || 10160 (tp->t_state == TCPS_SYN_RECEIVED)) { 10161 /* 10162 * We really don't know you have to get to ESTAB or beyond 10163 * to tell. 10164 */ 10165 return (EAGAIN); 10166 } 10167 if (tp->t_flags & TF_SENTFIN) 10168 return (EINVAL); 10169 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10170 return (0); 10171 } 10172 /* 10173 * If we reach here we don't do SACK on this connection so we can 10174 * never do rack. 10175 */ 10176 return (EINVAL); 10177 } 10178 10179 static void 10180 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10181 { 10182 if (tp->t_fb_ptr) { 10183 uint32_t calc; 10184 struct tcp_bbr *bbr; 10185 struct bbr_sendmap *rsm; 10186 10187 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10188 if (bbr->r_ctl.crte) 10189 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10190 bbr_log_flowend(bbr); 10191 bbr->rc_tp = NULL; 10192 if (bbr->bbr_hdrw_pacing) 10193 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10194 else 10195 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10196 if (bbr->r_ctl.crte != NULL) { 10197 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10198 bbr->r_ctl.crte = NULL; 10199 } 10200 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10201 while (rsm) { 10202 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10203 uma_zfree(bbr_zone, rsm); 10204 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10205 } 10206 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10207 while (rsm) { 10208 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10209 uma_zfree(bbr_zone, rsm); 10210 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10211 } 10212 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10213 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10214 BBR_STAT_INC(bbr_dynamic_rwnd); 10215 else 10216 BBR_STAT_INC(bbr_static_rwnd); 10217 bbr->r_ctl.rc_free_cnt = 0; 10218 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10219 tp->t_fb_ptr = NULL; 10220 } 10221 /* Make sure snd_nxt is correctly set */ 10222 tp->snd_nxt = tp->snd_max; 10223 } 10224 10225 static void 10226 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10227 { 10228 switch (tp->t_state) { 10229 case TCPS_SYN_SENT: 10230 bbr->r_state = TCPS_SYN_SENT; 10231 bbr->r_substate = bbr_do_syn_sent; 10232 break; 10233 case TCPS_SYN_RECEIVED: 10234 bbr->r_state = TCPS_SYN_RECEIVED; 10235 bbr->r_substate = bbr_do_syn_recv; 10236 break; 10237 case TCPS_ESTABLISHED: 10238 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10239 bbr->r_state = TCPS_ESTABLISHED; 10240 bbr->r_substate = bbr_do_established; 10241 break; 10242 case TCPS_CLOSE_WAIT: 10243 bbr->r_state = TCPS_CLOSE_WAIT; 10244 bbr->r_substate = bbr_do_close_wait; 10245 break; 10246 case TCPS_FIN_WAIT_1: 10247 bbr->r_state = TCPS_FIN_WAIT_1; 10248 bbr->r_substate = bbr_do_fin_wait_1; 10249 break; 10250 case TCPS_CLOSING: 10251 bbr->r_state = TCPS_CLOSING; 10252 bbr->r_substate = bbr_do_closing; 10253 break; 10254 case TCPS_LAST_ACK: 10255 bbr->r_state = TCPS_LAST_ACK; 10256 bbr->r_substate = bbr_do_lastack; 10257 break; 10258 case TCPS_FIN_WAIT_2: 10259 bbr->r_state = TCPS_FIN_WAIT_2; 10260 bbr->r_substate = bbr_do_fin_wait_2; 10261 break; 10262 case TCPS_LISTEN: 10263 case TCPS_CLOSED: 10264 case TCPS_TIME_WAIT: 10265 default: 10266 break; 10267 }; 10268 } 10269 10270 static void 10271 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10272 { 10273 /* 10274 * Now what state are we going into now? Is there adjustments 10275 * needed? 10276 */ 10277 int32_t old_state; 10278 10279 old_state = bbr_state_val(bbr); 10280 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10281 /* Save the lowest srtt we saw in our end of the sub-state */ 10282 bbr->rc_hit_state_1 = 0; 10283 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10284 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10285 } 10286 bbr->rc_bbr_substate++; 10287 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10288 /* Cycle back to first state-> gain */ 10289 bbr->rc_bbr_substate = 0; 10290 } 10291 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10292 /* 10293 * We enter the gain(5/4) cycle (possibly less if 10294 * shallow buffer detection is enabled) 10295 */ 10296 if (bbr->skip_gain) { 10297 /* 10298 * Hardware pacing has set our rate to 10299 * the max and limited our b/w just 10300 * do level i.e. no gain. 10301 */ 10302 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10303 } else if (bbr->gain_is_limited && 10304 bbr->bbr_hdrw_pacing && 10305 bbr->r_ctl.crte) { 10306 /* 10307 * We can't gain above the hardware pacing 10308 * rate which is less than our rate + the gain 10309 * calculate the gain needed to reach the hardware 10310 * pacing rate.. 10311 */ 10312 uint64_t bw, rate, gain_calc; 10313 10314 bw = bbr_get_bw(bbr); 10315 rate = bbr->r_ctl.crte->rate; 10316 if ((rate > bw) && 10317 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10318 gain_calc = (rate * BBR_UNIT) / bw; 10319 if (gain_calc < BBR_UNIT) 10320 gain_calc = BBR_UNIT; 10321 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10322 } else { 10323 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10324 } 10325 } else 10326 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10327 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10328 bbr->r_ctl.rc_bbr_state_atflight = cts; 10329 } else 10330 bbr->r_ctl.rc_bbr_state_atflight = 0; 10331 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10332 bbr->rc_hit_state_1 = 1; 10333 bbr->r_ctl.rc_exta_time_gd = 0; 10334 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10335 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10336 if (bbr_state_drain_2_tar) { 10337 bbr->r_ctl.rc_bbr_state_atflight = 0; 10338 } else 10339 bbr->r_ctl.rc_bbr_state_atflight = cts; 10340 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10341 } else { 10342 /* All other cycles hit here 2-7 */ 10343 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10344 if (bbr_sub_drain_slam_cwnd && 10345 (bbr->rc_use_google == 0) && 10346 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10347 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10348 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10349 } 10350 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10351 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10352 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10353 else 10354 bbr->r_ctl.rc_exta_time_gd = 0; 10355 if (bbr->r_ctl.rc_exta_time_gd) { 10356 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10357 /* Now chop up the time for each state (div by 7) */ 10358 bbr->r_ctl.rc_level_state_extra /= 7; 10359 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10360 /* Add a randomization */ 10361 bbr_randomize_extra_state_time(bbr); 10362 } 10363 } 10364 } 10365 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10366 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10367 } 10368 if (bbr->rc_use_google) { 10369 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10370 } 10371 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10372 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10373 if (dolog) 10374 bbr_log_type_statechange(bbr, cts, line); 10375 10376 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10377 uint32_t time_in; 10378 10379 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10380 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10381 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10382 } else { 10383 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10384 } 10385 } 10386 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10387 bbr_set_state_target(bbr, __LINE__); 10388 if (bbr_sub_drain_slam_cwnd && 10389 (bbr->rc_use_google == 0) && 10390 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10391 /* Slam down the cwnd */ 10392 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10393 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10394 if (bbr_sub_drain_app_limit) { 10395 /* Go app limited if we are on a long drain */ 10396 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10397 ctf_flight_size(bbr->rc_tp, 10398 (bbr->r_ctl.rc_sacked + 10399 bbr->r_ctl.rc_lost_bytes))); 10400 } 10401 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10402 } 10403 if (bbr->rc_lt_use_bw) { 10404 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10405 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10406 } 10407 /* Google changes TSO size every cycle */ 10408 if (bbr->rc_use_google) 10409 tcp_bbr_tso_size_check(bbr, cts); 10410 bbr->r_ctl.gain_epoch = cts; 10411 bbr->r_ctl.rc_bbr_state_time = cts; 10412 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10413 } 10414 10415 static void 10416 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10417 { 10418 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10419 (google_allow_early_out == 1) && 10420 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10421 /* We have reached out target flight size possibly early */ 10422 goto change_state; 10423 } 10424 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10425 return; 10426 } 10427 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10428 /* 10429 * Must be a rttProp movement forward before 10430 * we can change states. 10431 */ 10432 return; 10433 } 10434 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10435 /* 10436 * The needed time has passed but for 10437 * the gain cycle extra rules apply: 10438 * 1) If we have seen loss, we exit 10439 * 2) If we have not reached the target 10440 * we stay in GAIN (gain-to-target). 10441 */ 10442 if (google_consider_lost && losses) 10443 goto change_state; 10444 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10445 return; 10446 } 10447 } 10448 change_state: 10449 /* For gain we must reach our target, all others last 1 rttProp */ 10450 bbr_substate_change(bbr, cts, __LINE__, 1); 10451 } 10452 10453 static void 10454 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10455 { 10456 uint32_t flight, bbr_cur_cycle_time; 10457 10458 if (bbr->rc_use_google) { 10459 bbr_set_probebw_google_gains(bbr, cts, losses); 10460 return; 10461 } 10462 if (cts == 0) { 10463 /* 10464 * Never alow cts to be 0 we 10465 * do this so we can judge if 10466 * we have set a timestamp. 10467 */ 10468 cts = 1; 10469 } 10470 if (bbr_state_is_pkt_epoch) 10471 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10472 else 10473 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10474 10475 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10476 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10477 flight = ctf_flight_size(bbr->rc_tp, 10478 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10479 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10480 /* Keep it slam down */ 10481 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10482 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10483 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10484 } 10485 if (bbr_sub_drain_app_limit) { 10486 /* Go app limited if we are on a long drain */ 10487 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10488 } 10489 } 10490 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10491 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10492 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10493 /* 10494 * Still here after the same time as 10495 * the gain. We need to drain harder 10496 * for the next srtt. Reduce by a set amount 10497 * the gain drop is capped at DRAIN states 10498 * value (88). 10499 */ 10500 bbr->r_ctl.flightsize_at_drain = flight; 10501 if (bbr_drain_drop_mul && 10502 bbr_drain_drop_div && 10503 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10504 /* Use your specific drop value (def 4/5 = 20%) */ 10505 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10506 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10507 } else { 10508 /* You get drop of 20% */ 10509 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10510 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10511 } 10512 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10513 /* Reduce our gain again to the bottom */ 10514 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10515 } 10516 bbr_log_exit_gain(bbr, cts, 4); 10517 /* 10518 * Extend out so we wait another 10519 * epoch before dropping again. 10520 */ 10521 bbr->r_ctl.gain_epoch = cts; 10522 } 10523 if (flight <= bbr->r_ctl.rc_target_at_state) { 10524 if (bbr_sub_drain_slam_cwnd && 10525 (bbr->rc_use_google == 0) && 10526 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10527 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10528 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10529 } 10530 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10531 bbr_log_exit_gain(bbr, cts, 3); 10532 } 10533 } else { 10534 /* Its a gain */ 10535 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10536 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10537 goto change_state; 10538 } 10539 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10540 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10541 bbr->rc_tp->snd_wnd)) { 10542 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10543 bbr_log_exit_gain(bbr, cts, 2); 10544 } 10545 } 10546 /** 10547 * We fall through and return always one of two things has 10548 * occurred. 10549 * 1) We are still not at target 10550 * <or> 10551 * 2) We reached the target and set rc_bbr_state_atflight 10552 * which means we no longer hit this block 10553 * next time we are called. 10554 */ 10555 return; 10556 } 10557 change_state: 10558 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10559 return; 10560 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10561 /* Less than a full time-period has passed */ 10562 return; 10563 } 10564 if (bbr->r_ctl.rc_level_state_extra && 10565 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10566 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10567 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10568 /* Less than a full time-period + extra has passed */ 10569 return; 10570 } 10571 if (bbr_gain_gets_extra_too && 10572 bbr->r_ctl.rc_level_state_extra && 10573 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10574 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10575 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10576 /* Less than a full time-period + extra has passed */ 10577 return; 10578 } 10579 bbr_substate_change(bbr, cts, __LINE__, 1); 10580 } 10581 10582 static uint32_t 10583 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10584 { 10585 uint32_t mss, tar; 10586 10587 if (bbr->rc_use_google) { 10588 /* Google just uses the cwnd target */ 10589 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10590 } else { 10591 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10592 bbr->r_ctl.rc_pace_max_segs); 10593 /* Get the base cwnd with gain rounded to a mss */ 10594 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10595 gain), mss); 10596 /* Make sure it is within our min */ 10597 if (tar < get_min_cwnd(bbr)) 10598 return (get_min_cwnd(bbr)); 10599 } 10600 return (tar); 10601 } 10602 10603 static void 10604 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10605 { 10606 uint32_t tar, meth; 10607 10608 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10609 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10610 /* Special case using old probe-rtt method */ 10611 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10612 meth = 1; 10613 } else { 10614 /* Non-probe-rtt case and reduced probe-rtt */ 10615 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10616 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10617 /* For gain cycle we use the hptsi gain */ 10618 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10619 meth = 2; 10620 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10621 /* 10622 * If configured, or for google all other states 10623 * get BBR_UNIT. 10624 */ 10625 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10626 meth = 3; 10627 } else { 10628 /* 10629 * Or we set a target based on the pacing gain 10630 * for non-google mode and default (non-configured). 10631 * Note we don't set a target goal below drain (192). 10632 */ 10633 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10634 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10635 meth = 4; 10636 } else { 10637 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10638 meth = 5; 10639 } 10640 } 10641 } 10642 bbr_log_set_of_state_target(bbr, tar, line, meth); 10643 bbr->r_ctl.rc_target_at_state = tar; 10644 } 10645 10646 static void 10647 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10648 { 10649 /* Change to probe_rtt */ 10650 uint32_t time_in; 10651 10652 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10653 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10654 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10655 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10656 + bbr->r_ctl.rc_delivered); 10657 /* Setup so we force feed the filter */ 10658 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10659 bbr->rc_prtt_set_ts = 1; 10660 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10661 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10662 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10663 } 10664 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10665 bbr->r_ctl.rc_rtt_shrinks = cts; 10666 bbr->r_ctl.last_in_probertt = cts; 10667 bbr->r_ctl.rc_probertt_srttchktim = cts; 10668 bbr->r_ctl.rc_bbr_state_time = cts; 10669 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10670 /* We need to force the filter to update */ 10671 10672 if ((bbr_sub_drain_slam_cwnd) && 10673 bbr->rc_hit_state_1 && 10674 (bbr->rc_use_google == 0) && 10675 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10676 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10677 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10678 } else 10679 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10680 /* Update the lost */ 10681 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10682 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10683 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10684 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10685 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10686 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10687 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10688 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10689 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10690 } else { 10691 /* 10692 * We bring it down slowly by using a hptsi gain that is 10693 * probably 75%. This will slowly float down our outstanding 10694 * without tampering with the cwnd. 10695 */ 10696 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10697 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10698 bbr_set_state_target(bbr, __LINE__); 10699 if (bbr_prtt_slam_cwnd && 10700 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10701 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10702 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10703 } 10704 } 10705 if (ctf_flight_size(bbr->rc_tp, 10706 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10707 bbr->r_ctl.rc_target_at_state) { 10708 /* We are at target */ 10709 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10710 } else { 10711 /* We need to come down to reach target before our time begins */ 10712 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10713 } 10714 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10715 BBR_STAT_INC(bbr_enter_probertt); 10716 bbr_log_exit_gain(bbr, cts, 0); 10717 bbr_log_type_statechange(bbr, cts, line); 10718 } 10719 10720 static void 10721 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10722 { 10723 /* 10724 * Sanity check on probe-rtt intervals. 10725 * In crazy situations where we are competing 10726 * against new-reno flows with huge buffers 10727 * our rtt-prop interval could come to dominate 10728 * things if we can't get through a full set 10729 * of cycles, we need to adjust it. 10730 */ 10731 if (bbr_can_adjust_probertt && 10732 (bbr->rc_use_google == 0)) { 10733 uint16_t val = 0; 10734 uint32_t cur_rttp, fval, newval, baseval; 10735 10736 /* Are we to small and go into probe-rtt to often? */ 10737 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10738 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10739 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10740 if (bbr_is_ratio == 0) { 10741 if (fval > bbr_rtt_probe_limit) 10742 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10743 else 10744 newval = cur_rttp; 10745 } else { 10746 int mul; 10747 10748 mul = fval / bbr_rtt_probe_limit; 10749 newval = cur_rttp * mul; 10750 } 10751 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10752 bbr->r_ctl.rc_probertt_int = cur_rttp; 10753 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10754 val = 1; 10755 } else { 10756 /* 10757 * No adjustments were made 10758 * do we need to shrink it? 10759 */ 10760 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10761 if (cur_rttp <= bbr_rtt_probe_limit) { 10762 /* 10763 * Things have calmed down lets 10764 * shrink all the way to default 10765 */ 10766 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10767 reset_time_small(&bbr->r_ctl.rc_rttprop, 10768 (bbr_filter_len_sec * USECS_IN_SECOND)); 10769 cur_rttp = bbr_rtt_probe_limit; 10770 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10771 val = 2; 10772 } else { 10773 /* 10774 * Well does some adjustment make sense? 10775 */ 10776 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10777 /* We can reduce interval time some */ 10778 bbr->r_ctl.rc_probertt_int = cur_rttp; 10779 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10780 val = 3; 10781 } 10782 } 10783 } 10784 } 10785 if (val) 10786 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10787 } 10788 } 10789 10790 static void 10791 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10792 { 10793 /* Exit probe-rtt */ 10794 10795 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10796 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10797 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10798 } 10799 bbr_log_exit_gain(bbr, cts, 1); 10800 bbr->rc_hit_state_1 = 0; 10801 bbr->r_ctl.rc_rtt_shrinks = cts; 10802 bbr->r_ctl.last_in_probertt = cts; 10803 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10804 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10805 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10806 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10807 bbr->r_ctl.rc_delivered); 10808 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10809 uint32_t time_in; 10810 10811 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10812 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10813 } 10814 if (bbr->rc_filled_pipe) { 10815 /* Switch to probe_bw */ 10816 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10817 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10818 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10819 bbr_substate_change(bbr, cts, __LINE__, 0); 10820 bbr_log_type_statechange(bbr, cts, __LINE__); 10821 } else { 10822 /* Back to startup */ 10823 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10824 bbr->r_ctl.rc_bbr_state_time = cts; 10825 /* 10826 * We don't want to give a complete free 3 10827 * measurements until we exit, so we use 10828 * the number of pe's we were in probe-rtt 10829 * to add to the startup_epoch. That way 10830 * we will still retain the old state. 10831 */ 10832 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10833 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10834 /* Make sure to use the lower pg when shifting back in */ 10835 if (bbr->r_ctl.rc_lost && 10836 bbr_use_lower_gain_in_startup && 10837 (bbr->rc_use_google == 0)) 10838 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10839 else 10840 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10841 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10842 /* Probably not needed but set it anyway */ 10843 bbr_set_state_target(bbr, __LINE__); 10844 bbr_log_type_statechange(bbr, cts, __LINE__); 10845 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10846 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10847 } 10848 bbr_check_probe_rtt_limits(bbr, cts); 10849 } 10850 10851 static int32_t inline 10852 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10853 { 10854 if ((bbr->rc_past_init_win == 1) && 10855 (bbr->rc_in_persist == 0) && 10856 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10857 return (1); 10858 } 10859 if (bbr_can_force_probertt && 10860 (bbr->rc_in_persist == 0) && 10861 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10862 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10863 return (1); 10864 } 10865 return (0); 10866 } 10867 10868 static int32_t 10869 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10870 { 10871 uint64_t btlbw, gain; 10872 if (pkt_epoch == 0) { 10873 /* 10874 * Need to be on a pkt-epoch to continue. 10875 */ 10876 return (0); 10877 } 10878 btlbw = bbr_get_full_bw(bbr); 10879 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10880 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10881 if (btlbw >= gain) { 10882 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10883 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10884 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10885 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10886 } 10887 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10888 return (1); 10889 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10890 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10891 return(0); 10892 } 10893 10894 static int32_t inline 10895 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10896 { 10897 /* Have we gained 25% in the last 3 packet based epoch's? */ 10898 uint64_t btlbw, gain; 10899 int do_exit; 10900 int delta, rtt_gain; 10901 10902 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10903 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10904 /* 10905 * This qualifies as a RTT_PROBE session since we drop the 10906 * data outstanding to nothing and waited more than 10907 * bbr_rtt_probe_time. 10908 */ 10909 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10910 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10911 } 10912 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10913 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10914 return (0); 10915 } 10916 if (bbr->rc_use_google) 10917 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10918 10919 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10920 (bbr_use_lower_gain_in_startup)) { 10921 /* Drop to a lower gain 1.5 x since we saw loss */ 10922 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10923 } 10924 if (pkt_epoch == 0) { 10925 /* 10926 * Need to be on a pkt-epoch to continue. 10927 */ 10928 return (0); 10929 } 10930 if (bbr_rtt_gain_thresh) { 10931 /* 10932 * Do we allow a flow to stay 10933 * in startup with no loss and no 10934 * gain in rtt over a set threshold? 10935 */ 10936 if (bbr->r_ctl.rc_pkt_epoch_rtt && 10937 bbr->r_ctl.startup_last_srtt && 10938 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 10939 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 10940 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 10941 } else 10942 rtt_gain = 0; 10943 if ((bbr->r_ctl.startup_last_srtt == 0) || 10944 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 10945 /* First time or new lower value */ 10946 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 10947 10948 if ((bbr->r_ctl.rc_lost == 0) && 10949 (rtt_gain < bbr_rtt_gain_thresh)) { 10950 /* 10951 * No loss, and we are under 10952 * our gain threhold for 10953 * increasing RTT. 10954 */ 10955 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10956 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10957 bbr_log_startup_event(bbr, cts, rtt_gain, 10958 delta, bbr->r_ctl.startup_last_srtt, 10); 10959 return (0); 10960 } 10961 } 10962 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 10963 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 10964 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 10965 /* 10966 * We only assess if we have a new measurement when 10967 * we have no loss and are not in recovery. 10968 * Drag up by one our last_startup epoch so we will hold 10969 * the number of non-gain we have already accumulated. 10970 */ 10971 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10972 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10973 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10974 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); 10975 return (0); 10976 } 10977 /* Case where we reduced the lost (bad retransmit) */ 10978 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 10979 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10980 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 10981 btlbw = bbr_get_full_bw(bbr); 10982 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 10983 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10984 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10985 else 10986 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10987 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10988 do_exit = 0; 10989 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 10990 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10991 if (btlbw >= gain) { 10992 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10993 /* Update the lost so we won't exit in next set of tests */ 10994 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10995 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10996 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10997 } 10998 if ((bbr->rc_loss_exit && 10999 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 11000 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 11001 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 11002 /* 11003 * If we had no gain, we had loss and that loss was above 11004 * our threshould, the rwnd is not constrained, and we have 11005 * had at least 3 packet epochs exit. Note that this is 11006 * switched off by sysctl. Google does not do this by the 11007 * way. 11008 */ 11009 if ((ctf_flight_size(bbr->rc_tp, 11010 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11011 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11012 do_exit = 1; 11013 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11014 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11015 } else { 11016 /* Just record an updated loss value */ 11017 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11018 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11019 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11020 } 11021 } else 11022 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11023 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11024 do_exit) { 11025 /* Return 1 to exit the startup state. */ 11026 return (1); 11027 } 11028 /* Stay in startup */ 11029 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11030 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11031 return (0); 11032 } 11033 11034 static void 11035 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11036 { 11037 /* 11038 * A tick occurred in the rtt epoch do we need to do anything? 11039 */ 11040 #ifdef BBR_INVARIANTS 11041 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11042 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11043 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11044 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11045 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11046 /* Debug code? */ 11047 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11048 } 11049 #endif 11050 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11051 /* Do we exit the startup state? */ 11052 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11053 uint32_t time_in; 11054 11055 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11056 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11057 bbr->rc_filled_pipe = 1; 11058 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11059 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11060 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11061 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11062 } else 11063 time_in = 0; 11064 if (bbr->rc_no_pacing) 11065 bbr->rc_no_pacing = 0; 11066 bbr->r_ctl.rc_bbr_state_time = cts; 11067 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11068 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11069 bbr_set_state_target(bbr, __LINE__); 11070 if ((bbr->rc_use_google == 0) && 11071 bbr_slam_cwnd_in_main_drain) { 11072 /* Here we don't have to worry about probe-rtt */ 11073 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11074 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11075 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11076 } 11077 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11078 bbr_log_type_statechange(bbr, cts, __LINE__); 11079 if (ctf_flight_size(bbr->rc_tp, 11080 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11081 bbr->r_ctl.rc_target_at_state) { 11082 /* 11083 * Switch to probe_bw if we are already 11084 * there 11085 */ 11086 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11087 bbr_substate_change(bbr, cts, __LINE__, 0); 11088 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11089 bbr_log_type_statechange(bbr, cts, __LINE__); 11090 } 11091 } 11092 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11093 uint32_t inflight; 11094 struct tcpcb *tp; 11095 11096 tp = bbr->rc_tp; 11097 inflight = ctf_flight_size(tp, 11098 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11099 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11100 /* We have reached a flight of the cwnd target */ 11101 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11102 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11103 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11104 bbr_set_state_target(bbr, __LINE__); 11105 /* 11106 * Rig it so we don't do anything crazy and 11107 * start fresh with a new randomization. 11108 */ 11109 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11110 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11111 bbr_substate_change(bbr, cts, __LINE__, 1); 11112 } 11113 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11114 /* Has in-flight reached the bdp (or less)? */ 11115 uint32_t inflight; 11116 struct tcpcb *tp; 11117 11118 tp = bbr->rc_tp; 11119 inflight = ctf_flight_size(tp, 11120 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11121 if ((bbr->rc_use_google == 0) && 11122 bbr_slam_cwnd_in_main_drain && 11123 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11124 /* 11125 * Here we don't have to worry about probe-rtt 11126 * re-slam it, but keep it slammed down. 11127 */ 11128 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11129 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11130 } 11131 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11132 /* We have drained */ 11133 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11134 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11135 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11136 uint32_t time_in; 11137 11138 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11139 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11140 } 11141 if ((bbr->rc_use_google == 0) && 11142 bbr_slam_cwnd_in_main_drain && 11143 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11144 /* Restore the cwnd */ 11145 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11146 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11147 } 11148 /* Setup probe-rtt has being done now RRS-HERE */ 11149 bbr->r_ctl.rc_rtt_shrinks = cts; 11150 bbr->r_ctl.last_in_probertt = cts; 11151 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11152 /* Randomly pick a sub-state */ 11153 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11154 bbr_substate_change(bbr, cts, __LINE__, 0); 11155 bbr_log_type_statechange(bbr, cts, __LINE__); 11156 } 11157 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11158 uint32_t flight; 11159 11160 flight = ctf_flight_size(bbr->rc_tp, 11161 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11162 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11163 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11164 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11165 /* 11166 * We must keep cwnd at the desired MSS. 11167 */ 11168 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11169 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11170 } else if ((bbr_prtt_slam_cwnd) && 11171 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11172 /* Re-slam it */ 11173 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11174 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11175 } 11176 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11177 /* Has outstanding reached our target? */ 11178 if (flight <= bbr->r_ctl.rc_target_at_state) { 11179 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11180 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11181 /* If time is exactly 0, be 1usec off */ 11182 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11183 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11184 if (bbr->rc_use_google == 0) { 11185 /* 11186 * Restore any lowering that as occurred to 11187 * reach here 11188 */ 11189 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11190 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11191 else 11192 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11193 } 11194 } 11195 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11196 (bbr->rc_use_google == 0) && 11197 bbr->r_ctl.bbr_rttprobe_gain_val && 11198 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11199 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11200 /* 11201 * We have doddled with our current hptsi 11202 * gain an srtt and have still not made it 11203 * to target, or we have increased our flight. 11204 * Lets reduce the gain by xx% 11205 * flooring the reduce at DRAIN (based on 11206 * mul/div) 11207 */ 11208 int red; 11209 11210 bbr->r_ctl.flightsize_at_drain = flight; 11211 bbr->r_ctl.rc_probertt_srttchktim = cts; 11212 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11213 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11214 /* Reduce our gain again */ 11215 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11216 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11217 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11218 /* one more chance before we give up */ 11219 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11220 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11221 } else { 11222 /* At the very bottom */ 11223 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11224 } 11225 } 11226 } 11227 if (bbr->r_ctl.rc_bbr_enters_probertt && 11228 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11229 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11230 /* Time to exit probe RTT normally */ 11231 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11232 } 11233 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11234 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11235 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11236 /* 11237 * This qualifies as a RTT_PROBE session since we 11238 * drop the data outstanding to nothing and waited 11239 * more than bbr_rtt_probe_time. 11240 */ 11241 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11242 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11243 } 11244 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11245 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11246 } else { 11247 bbr_set_probebw_gains(bbr, cts, losses); 11248 } 11249 } 11250 } 11251 11252 static void 11253 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11254 { 11255 int32_t epoch = 0; 11256 11257 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11258 bbr_set_epoch(bbr, cts, line); 11259 /* At each epoch doe lt bw sampling */ 11260 epoch = 1; 11261 } 11262 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11263 } 11264 11265 static int 11266 bbr_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 11267 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, 11268 struct timeval *tv) 11269 { 11270 struct inpcb *inp = tptoinpcb(tp); 11271 struct socket *so = tptosocket(tp); 11272 int32_t thflags, retval; 11273 uint32_t cts, lcts; 11274 uint32_t tiwin; 11275 struct tcpopt to; 11276 struct tcp_bbr *bbr; 11277 struct bbr_sendmap *rsm; 11278 struct timeval ltv; 11279 int32_t did_out = 0; 11280 uint16_t nsegs; 11281 int32_t prev_state; 11282 uint32_t lost; 11283 11284 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11285 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11286 /* add in our stats */ 11287 kern_prefetch(bbr, &prev_state); 11288 prev_state = 0; 11289 thflags = tcp_get_flags(th); 11290 /* 11291 * If this is either a state-changing packet or current state isn't 11292 * established, we require a write lock on tcbinfo. Otherwise, we 11293 * allow the tcbinfo to be in either alocked or unlocked, as the 11294 * caller may have unnecessarily acquired a write lock due to a 11295 * race. 11296 */ 11297 INP_WLOCK_ASSERT(tptoinpcb(tp)); 11298 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11299 __func__)); 11300 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11301 __func__)); 11302 11303 tp->t_rcvtime = ticks; 11304 /* 11305 * Unscale the window into a 32-bit value. For the SYN_SENT state 11306 * the scale is zero. 11307 */ 11308 tiwin = th->th_win << tp->snd_scale; 11309 #ifdef STATS 11310 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11311 #endif 11312 11313 if (m->m_flags & M_TSTMP) { 11314 /* Prefer the hardware timestamp if present */ 11315 struct timespec ts; 11316 11317 mbuf_tstmp2timespec(m, &ts); 11318 bbr->rc_tv.tv_sec = ts.tv_sec; 11319 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11320 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11321 } else if (m->m_flags & M_TSTMP_LRO) { 11322 /* Next the arrival timestamp */ 11323 struct timespec ts; 11324 11325 mbuf_tstmp2timespec(m, &ts); 11326 bbr->rc_tv.tv_sec = ts.tv_sec; 11327 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11328 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11329 } else { 11330 /* 11331 * Ok just get the current time. 11332 */ 11333 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11334 } 11335 /* 11336 * Parse options on any incoming segment. 11337 */ 11338 tcp_dooptions(&to, (u_char *)(th + 1), 11339 (th->th_off << 2) - sizeof(struct tcphdr), 11340 (thflags & TH_SYN) ? TO_SYN : 0); 11341 11342 /* 11343 * If timestamps were negotiated during SYN/ACK and a 11344 * segment without a timestamp is received, silently drop 11345 * the segment, unless it is a RST segment or missing timestamps are 11346 * tolerated. 11347 * See section 3.2 of RFC 7323. 11348 */ 11349 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11350 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11351 retval = 0; 11352 m_freem(m); 11353 goto done_with_input; 11354 } 11355 /* 11356 * If echoed timestamp is later than the current time, fall back to 11357 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11358 * were used when this connection was established. 11359 */ 11360 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11361 to.to_tsecr -= tp->ts_offset; 11362 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11363 to.to_tsecr = 0; 11364 } 11365 /* 11366 * If its the first time in we need to take care of options and 11367 * verify we can do SACK for rack! 11368 */ 11369 if (bbr->r_state == 0) { 11370 /* 11371 * Process options only when we get SYN/ACK back. The SYN 11372 * case for incoming connections is handled in tcp_syncache. 11373 * According to RFC1323 the window field in a SYN (i.e., a 11374 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11375 * this is traditional behavior, may need to be cleaned up. 11376 */ 11377 if (bbr->rc_inp == NULL) { 11378 bbr->rc_inp = inp; 11379 } 11380 /* 11381 * We need to init rc_inp here since its not init'd when 11382 * bbr_init is called 11383 */ 11384 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11385 if ((to.to_flags & TOF_SCALE) && 11386 (tp->t_flags & TF_REQ_SCALE)) { 11387 tp->t_flags |= TF_RCVD_SCALE; 11388 tp->snd_scale = to.to_wscale; 11389 } else 11390 tp->t_flags &= ~TF_REQ_SCALE; 11391 /* 11392 * Initial send window. It will be updated with the 11393 * next incoming segment to the scaled value. 11394 */ 11395 tp->snd_wnd = th->th_win; 11396 if ((to.to_flags & TOF_TS) && 11397 (tp->t_flags & TF_REQ_TSTMP)) { 11398 tp->t_flags |= TF_RCVD_TSTMP; 11399 tp->ts_recent = to.to_tsval; 11400 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11401 } else 11402 tp->t_flags &= ~TF_REQ_TSTMP; 11403 if (to.to_flags & TOF_MSS) 11404 tcp_mss(tp, to.to_mss); 11405 if ((tp->t_flags & TF_SACK_PERMIT) && 11406 (to.to_flags & TOF_SACKPERM) == 0) 11407 tp->t_flags &= ~TF_SACK_PERMIT; 11408 if (IS_FASTOPEN(tp->t_flags)) { 11409 if (to.to_flags & TOF_FASTOPEN) { 11410 uint16_t mss; 11411 11412 if (to.to_flags & TOF_MSS) 11413 mss = to.to_mss; 11414 else 11415 if ((inp->inp_vflag & INP_IPV6) != 0) 11416 mss = TCP6_MSS; 11417 else 11418 mss = TCP_MSS; 11419 tcp_fastopen_update_cache(tp, mss, 11420 to.to_tfo_len, to.to_tfo_cookie); 11421 } else 11422 tcp_fastopen_disable_path(tp); 11423 } 11424 } 11425 /* 11426 * At this point we are at the initial call. Here we decide 11427 * if we are doing RACK or not. We do this by seeing if 11428 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11429 * we switch to the default code. 11430 */ 11431 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11432 /* Bail */ 11433 tcp_switch_back_to_default(tp); 11434 (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen, 11435 tlen, iptos); 11436 return (1); 11437 } 11438 /* Set the flag */ 11439 bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0; 11440 tcp_set_hpts(tp); 11441 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11442 } 11443 if (thflags & TH_ACK) { 11444 /* Track ack types */ 11445 if (to.to_flags & TOF_SACK) 11446 BBR_STAT_INC(bbr_acks_with_sacks); 11447 else 11448 BBR_STAT_INC(bbr_plain_acks); 11449 } 11450 /* 11451 * This is the one exception case where we set the rack state 11452 * always. All other times (timers etc) we must have a rack-state 11453 * set (so we assure we have done the checks above for SACK). 11454 */ 11455 if (thflags & TH_FIN) 11456 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11457 if (bbr->r_state != tp->t_state) 11458 bbr_set_state(tp, bbr, tiwin); 11459 11460 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11461 kern_prefetch(rsm, &prev_state); 11462 prev_state = bbr->r_state; 11463 bbr->rc_ack_was_delayed = 0; 11464 lost = bbr->r_ctl.rc_lost; 11465 bbr->rc_is_pkt_epoch_now = 0; 11466 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11467 /* Get the real time into lcts and figure the real delay */ 11468 lcts = tcp_get_usecs(<v); 11469 if (TSTMP_GT(lcts, cts)) { 11470 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11471 bbr->rc_ack_was_delayed = 1; 11472 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11473 bbr->r_ctl.highest_hdwr_delay)) 11474 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11475 } else { 11476 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11477 bbr->rc_ack_was_delayed = 0; 11478 } 11479 } else { 11480 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11481 bbr->rc_ack_was_delayed = 0; 11482 } 11483 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11484 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11485 retval = 0; 11486 m_freem(m); 11487 goto done_with_input; 11488 } 11489 /* 11490 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11491 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11492 */ 11493 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11494 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11495 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11496 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11497 return (1); 11498 } 11499 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11500 bbr->r_ctl.rc_high_rwnd = tiwin; 11501 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11502 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11503 bbr->rtt_valid = 0; 11504 if (to.to_flags & TOF_TS) { 11505 bbr->rc_ts_valid = 1; 11506 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11507 } else { 11508 bbr->rc_ts_valid = 0; 11509 bbr->r_ctl.last_inbound_ts = 0; 11510 } 11511 retval = (*bbr->r_substate) (m, th, so, 11512 tp, &to, drop_hdrlen, 11513 tlen, tiwin, thflags, nxt_pkt, iptos); 11514 if (nxt_pkt == 0) 11515 BBR_STAT_INC(bbr_rlock_left_ret0); 11516 else 11517 BBR_STAT_INC(bbr_rlock_left_ret1); 11518 if (retval == 0) { 11519 /* 11520 * If retval is 1 the tcb is unlocked and most likely the tp 11521 * is gone. 11522 */ 11523 INP_WLOCK_ASSERT(inp); 11524 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11525 if (bbr->rc_is_pkt_epoch_now) 11526 bbr_set_pktepoch(bbr, cts, __LINE__); 11527 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11528 if (nxt_pkt == 0) { 11529 if (bbr->r_wanted_output != 0) { 11530 bbr->rc_output_starts_timer = 0; 11531 did_out = 1; 11532 if (tcp_output(tp) < 0) 11533 return (1); 11534 } else 11535 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11536 } 11537 if ((nxt_pkt == 0) && 11538 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11539 (SEQ_GT(tp->snd_max, tp->snd_una) || 11540 (tp->t_flags & TF_DELACK) || 11541 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11542 (tp->t_state <= TCPS_CLOSING)))) { 11543 /* 11544 * We could not send (probably in the hpts but 11545 * stopped the timer)? 11546 */ 11547 if ((tp->snd_max == tp->snd_una) && 11548 ((tp->t_flags & TF_DELACK) == 0) && 11549 (tcp_in_hpts(tp)) && 11550 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11551 /* 11552 * keep alive not needed if we are hptsi 11553 * output yet 11554 */ 11555 ; 11556 } else { 11557 if (tcp_in_hpts(tp)) { 11558 tcp_hpts_remove(tp); 11559 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11560 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11561 uint32_t del; 11562 11563 del = lcts - bbr->rc_pacer_started; 11564 if (bbr->r_ctl.rc_last_delay_val > del) { 11565 BBR_STAT_INC(bbr_force_timer_start); 11566 bbr->r_ctl.rc_last_delay_val -= del; 11567 bbr->rc_pacer_started = lcts; 11568 } else { 11569 /* We are late */ 11570 bbr->r_ctl.rc_last_delay_val = 0; 11571 BBR_STAT_INC(bbr_force_output); 11572 if (tcp_output(tp) < 0) 11573 return (1); 11574 } 11575 } 11576 } 11577 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11578 0); 11579 } 11580 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11581 /* Do we have the correct timer running? */ 11582 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11583 } 11584 /* Clear the flag, it may have been cleared by output but we may not have */ 11585 if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS)) 11586 tp->t_flags2 &= ~TF2_HPTS_CALLS; 11587 /* Do we have a new state */ 11588 if (bbr->r_state != tp->t_state) 11589 bbr_set_state(tp, bbr, tiwin); 11590 done_with_input: 11591 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11592 if (did_out) 11593 bbr->r_wanted_output = 0; 11594 } 11595 return (retval); 11596 } 11597 11598 static void 11599 bbr_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 11600 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11601 { 11602 struct timeval tv; 11603 int retval; 11604 11605 /* First lets see if we have old packets */ 11606 if (!STAILQ_EMPTY(&tp->t_inqueue)) { 11607 if (ctf_do_queued_segments(tp, 1)) { 11608 m_freem(m); 11609 return; 11610 } 11611 } 11612 if (m->m_flags & M_TSTMP_LRO) { 11613 mbuf_tstmp2timeval(m, &tv); 11614 } else { 11615 /* Should not be should we kassert instead? */ 11616 tcp_get_usecs(&tv); 11617 } 11618 retval = bbr_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 11619 0, &tv); 11620 if (retval == 0) { 11621 INP_WUNLOCK(tptoinpcb(tp)); 11622 } 11623 } 11624 11625 /* 11626 * Return how much data can be sent without violating the 11627 * cwnd or rwnd. 11628 */ 11629 11630 static inline uint32_t 11631 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11632 uint32_t avail, int32_t sb_offset, uint32_t cts) 11633 { 11634 uint32_t len; 11635 11636 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11637 /* We never want to go over our peers rcv-window */ 11638 len = 0; 11639 } else { 11640 uint32_t flight; 11641 11642 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11643 if (flight >= sendwin) { 11644 /* 11645 * We have in flight what we are allowed by cwnd (if 11646 * it was rwnd blocking it would have hit above out 11647 * >= tp->snd_wnd). 11648 */ 11649 return (0); 11650 } 11651 len = sendwin - flight; 11652 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11653 /* We would send too much (beyond the rwnd) */ 11654 len = tp->snd_wnd - ctf_outstanding(tp); 11655 } 11656 if ((len + sb_offset) > avail) { 11657 /* 11658 * We don't have that much in the SB, how much is 11659 * there? 11660 */ 11661 len = avail - sb_offset; 11662 } 11663 } 11664 return (len); 11665 } 11666 11667 static inline void 11668 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11669 { 11670 if (error) { 11671 return; 11672 } 11673 if (rsm) { 11674 if (rsm->r_flags & BBR_TLP) { 11675 /* 11676 * TLP should not count in retran count, but in its 11677 * own bin 11678 */ 11679 KMOD_TCPSTAT_INC(tcps_tlpresends); 11680 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11681 } else { 11682 /* Retransmit */ 11683 tp->t_sndrexmitpack++; 11684 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11685 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11686 #ifdef STATS 11687 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11688 len); 11689 #endif 11690 } 11691 /* 11692 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11693 * sub-state 11694 */ 11695 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11696 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11697 /* Non probe_bw log in 1, 2, or 4. */ 11698 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11699 } else { 11700 /* 11701 * Log our probe state 3, and log also 5-13 to show 11702 * us the recovery sub-state for the send. This 11703 * means that 3 == (5+6+7+8+9+10+11+12+13) 11704 */ 11705 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11706 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11707 } 11708 /* Place in both 16's the totals of retransmitted */ 11709 counter_u64_add(bbr_state_lost[16], len); 11710 counter_u64_add(bbr_state_resend[16], len); 11711 /* Place in 17's the total sent */ 11712 counter_u64_add(bbr_state_resend[17], len); 11713 counter_u64_add(bbr_state_lost[17], len); 11714 11715 } else { 11716 /* New sends */ 11717 KMOD_TCPSTAT_INC(tcps_sndpack); 11718 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11719 /* Place in 17's the total sent */ 11720 counter_u64_add(bbr_state_resend[17], len); 11721 counter_u64_add(bbr_state_lost[17], len); 11722 #ifdef STATS 11723 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11724 len); 11725 #endif 11726 } 11727 } 11728 11729 static void 11730 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11731 { 11732 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11733 /* 11734 * Limit the cwnd to not be above N x the target plus whats 11735 * is outstanding. The target is based on the current b/w 11736 * estimate. 11737 */ 11738 uint32_t target; 11739 11740 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11741 target += ctf_outstanding(tp); 11742 target *= bbr_target_cwnd_mult_limit; 11743 if (tp->snd_cwnd > target) 11744 tp->snd_cwnd = target; 11745 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11746 } 11747 } 11748 11749 static int 11750 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11751 { 11752 /* 11753 * "adv" is the amount we could increase the window, taking into 11754 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11755 */ 11756 int32_t adv; 11757 int32_t oldwin; 11758 11759 adv = recwin; 11760 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11761 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11762 if (adv > oldwin) 11763 adv -= oldwin; 11764 else { 11765 /* We can't increase the window */ 11766 adv = 0; 11767 } 11768 } else 11769 oldwin = 0; 11770 11771 /* 11772 * If the new window size ends up being the same as or less 11773 * than the old size when it is scaled, then don't force 11774 * a window update. 11775 */ 11776 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11777 return (0); 11778 11779 if (adv >= (2 * maxseg) && 11780 (adv >= (so->so_rcv.sb_hiwat / 4) || 11781 recwin <= (so->so_rcv.sb_hiwat / 8) || 11782 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11783 return (1); 11784 } 11785 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11786 return (1); 11787 return (0); 11788 } 11789 11790 /* 11791 * Return 0 on success and a errno on failure to send. 11792 * Note that a 0 return may not mean we sent anything 11793 * if the TCB was on the hpts. A non-zero return 11794 * does indicate the error we got from ip[6]_output. 11795 */ 11796 static int 11797 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11798 { 11799 struct socket *so; 11800 int32_t len; 11801 uint32_t cts; 11802 uint32_t recwin, sendwin; 11803 int32_t sb_offset; 11804 int32_t flags, abandon, error = 0; 11805 struct tcp_log_buffer *lgb = NULL; 11806 struct mbuf *m; 11807 struct mbuf *mb; 11808 uint32_t if_hw_tsomaxsegcount = 0; 11809 uint32_t if_hw_tsomaxsegsize = 0; 11810 uint32_t if_hw_tsomax = 0; 11811 struct ip *ip = NULL; 11812 struct tcp_bbr *bbr; 11813 struct tcphdr *th; 11814 struct udphdr *udp = NULL; 11815 u_char opt[TCP_MAXOLEN]; 11816 unsigned ipoptlen, optlen, hdrlen; 11817 unsigned ulen; 11818 uint32_t bbr_seq; 11819 uint32_t delay_calc=0; 11820 uint8_t doing_tlp = 0; 11821 uint8_t local_options; 11822 #ifdef BBR_INVARIANTS 11823 uint8_t doing_retran_from = 0; 11824 uint8_t picked_up_retran = 0; 11825 #endif 11826 uint8_t wanted_cookie = 0; 11827 uint8_t more_to_rxt=0; 11828 int32_t prefetch_so_done = 0; 11829 int32_t prefetch_rsm = 0; 11830 uint32_t tot_len = 0; 11831 uint32_t maxseg, pace_max_segs, p_maxseg; 11832 int32_t csum_flags = 0; 11833 int32_t hw_tls; 11834 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11835 unsigned ipsec_optlen = 0; 11836 11837 #endif 11838 volatile int32_t sack_rxmit; 11839 struct bbr_sendmap *rsm = NULL; 11840 int32_t tso, mtu; 11841 struct tcpopt to; 11842 int32_t slot = 0; 11843 struct inpcb *inp; 11844 struct sockbuf *sb; 11845 bool hpts_calling; 11846 #ifdef INET6 11847 struct ip6_hdr *ip6 = NULL; 11848 int32_t isipv6; 11849 #endif 11850 uint8_t app_limited = BBR_JR_SENT_DATA; 11851 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11852 /* We take a cache hit here */ 11853 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11854 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11855 inp = bbr->rc_inp; 11856 hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS); 11857 tp->t_flags2 &= ~TF2_HPTS_CALLS; 11858 so = inp->inp_socket; 11859 sb = &so->so_snd; 11860 if (tp->t_nic_ktls_xmit) 11861 hw_tls = 1; 11862 else 11863 hw_tls = 0; 11864 kern_prefetch(sb, &maxseg); 11865 maxseg = tp->t_maxseg - bbr->rc_last_options; 11866 if (bbr_minseg(bbr) < maxseg) { 11867 tcp_bbr_tso_size_check(bbr, cts); 11868 } 11869 /* Remove any flags that indicate we are pacing on the inp */ 11870 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11871 p_maxseg = min(maxseg, pace_max_segs); 11872 INP_WLOCK_ASSERT(inp); 11873 #ifdef TCP_OFFLOAD 11874 if (tp->t_flags & TF_TOE) 11875 return (tcp_offload_output(tp)); 11876 #endif 11877 11878 #ifdef INET6 11879 if (bbr->r_state) { 11880 /* Use the cache line loaded if possible */ 11881 isipv6 = bbr->r_is_v6; 11882 } else { 11883 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11884 } 11885 #endif 11886 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11887 tcp_in_hpts(tp)) { 11888 /* 11889 * We are on the hpts for some timer but not hptsi output. 11890 * Possibly remove from the hpts so we can send/recv etc. 11891 */ 11892 if ((tp->t_flags & TF_ACKNOW) == 0) { 11893 /* 11894 * No immediate demand right now to send an ack, but 11895 * the user may have read, making room for new data 11896 * (a window update). If so we may want to cancel 11897 * whatever timer is running (KEEP/DEL-ACK?) and 11898 * continue to send out a window update. Or we may 11899 * have gotten more data into the socket buffer to 11900 * send. 11901 */ 11902 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 11903 (long)TCP_MAXWIN << tp->rcv_scale); 11904 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 11905 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 11906 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 11907 (tp->snd_max - tp->snd_una))) { 11908 /* 11909 * Nothing new to send and no window update 11910 * is needed to send. Lets just return and 11911 * let the timer-run off. 11912 */ 11913 return (0); 11914 } 11915 } 11916 tcp_hpts_remove(tp); 11917 bbr_timer_cancel(bbr, __LINE__, cts); 11918 } 11919 if (bbr->r_ctl.rc_last_delay_val) { 11920 /* Calculate a rough delay for early escape to sending */ 11921 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11922 delay_calc = cts - bbr->rc_pacer_started; 11923 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11924 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11925 else 11926 delay_calc = 0; 11927 } 11928 /* Mark that we have called bbr_output(). */ 11929 if ((bbr->r_timer_override) || 11930 (tp->t_state < TCPS_ESTABLISHED)) { 11931 /* Timeouts or early states are exempt */ 11932 if (tcp_in_hpts(tp)) 11933 tcp_hpts_remove(tp); 11934 } else if (tcp_in_hpts(tp)) { 11935 if ((bbr->r_ctl.rc_last_delay_val) && 11936 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11937 delay_calc) { 11938 /* 11939 * We were being paced for output and the delay has 11940 * already exceeded when we were supposed to be 11941 * called, lets go ahead and pull out of the hpts 11942 * and call output. 11943 */ 11944 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 11945 bbr->r_ctl.rc_last_delay_val = 0; 11946 tcp_hpts_remove(tp); 11947 } else if (tp->t_state == TCPS_CLOSED) { 11948 bbr->r_ctl.rc_last_delay_val = 0; 11949 tcp_hpts_remove(tp); 11950 } else { 11951 /* 11952 * On the hpts, you shall not pass! even if ACKNOW 11953 * is on, we will when the hpts fires, unless of 11954 * course we are overdue. 11955 */ 11956 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 11957 return (0); 11958 } 11959 } 11960 bbr->rc_cwnd_limited = 0; 11961 if (bbr->r_ctl.rc_last_delay_val) { 11962 /* recalculate the real delay and deal with over/under */ 11963 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11964 delay_calc = cts - bbr->rc_pacer_started; 11965 else 11966 delay_calc = 0; 11967 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11968 /* Setup the delay which will be added in */ 11969 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11970 else { 11971 /* 11972 * We are early setup to adjust 11973 * our slot time. 11974 */ 11975 uint64_t merged_val; 11976 11977 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 11978 bbr->r_agg_early_set = 1; 11979 if (bbr->r_ctl.rc_hptsi_agg_delay) { 11980 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 11981 /* Nope our previous late cancels out the early */ 11982 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 11983 bbr->r_agg_early_set = 0; 11984 bbr->r_ctl.rc_agg_early = 0; 11985 } else { 11986 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 11987 bbr->r_ctl.rc_hptsi_agg_delay = 0; 11988 } 11989 } 11990 merged_val = bbr->rc_pacer_started; 11991 merged_val <<= 32; 11992 merged_val |= bbr->r_ctl.rc_last_delay_val; 11993 bbr_log_pacing_delay_calc(bbr, hpts_calling, 11994 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 11995 bbr->r_agg_early_set, 3); 11996 bbr->r_ctl.rc_last_delay_val = 0; 11997 BBR_STAT_INC(bbr_early); 11998 delay_calc = 0; 11999 } 12000 } else { 12001 /* We were not delayed due to hptsi */ 12002 if (bbr->r_agg_early_set) 12003 bbr->r_ctl.rc_agg_early = 0; 12004 bbr->r_agg_early_set = 0; 12005 delay_calc = 0; 12006 } 12007 if (delay_calc) { 12008 /* 12009 * We had a hptsi delay which means we are falling behind on 12010 * sending at the expected rate. Calculate an extra amount 12011 * of data we can send, if any, to put us back on track. 12012 */ 12013 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12014 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12015 else 12016 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12017 } 12018 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12019 if ((tp->snd_una == tp->snd_max) && 12020 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12021 (sbavail(sb))) { 12022 /* 12023 * Ok we have been idle with nothing outstanding 12024 * we possibly need to start fresh with either a new 12025 * suite of states or a fast-ramp up. 12026 */ 12027 bbr_restart_after_idle(bbr, 12028 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12029 } 12030 /* 12031 * Now was there a hptsi delay where we are behind? We only count 12032 * being behind if: a) We are not in recovery. b) There was a delay. 12033 * <and> c) We had room to send something. 12034 * 12035 */ 12036 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12037 int retval; 12038 12039 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12040 if (retval != 0) { 12041 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12042 /* 12043 * If timers want tcp_drop(), then pass error out, 12044 * otherwise suppress it. 12045 */ 12046 return (retval < 0 ? retval : 0); 12047 } 12048 } 12049 bbr->rc_tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY; 12050 if (hpts_calling && 12051 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12052 bbr->r_ctl.rc_last_delay_val = 0; 12053 } 12054 bbr->r_timer_override = 0; 12055 bbr->r_wanted_output = 0; 12056 /* 12057 * For TFO connections in SYN_RECEIVED, only allow the initial 12058 * SYN|ACK and those sent by the retransmit timer. 12059 */ 12060 if (IS_FASTOPEN(tp->t_flags) && 12061 ((tp->t_state == TCPS_SYN_RECEIVED) || 12062 (tp->t_state == TCPS_SYN_SENT)) && 12063 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12064 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12065 len = 0; 12066 goto just_return_nolock; 12067 } 12068 /* 12069 * Before sending anything check for a state update. For hpts 12070 * calling without input this is important. If its input calling 12071 * then this was already done. 12072 */ 12073 if (bbr->rc_use_google == 0) 12074 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12075 again: 12076 /* 12077 * If we've recently taken a timeout, snd_max will be greater than 12078 * snd_max. BBR in general does not pay much attention to snd_nxt 12079 * for historic reasons the persist timer still uses it. This means 12080 * we have to look at it. All retransmissions that are not persits 12081 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12082 * end of this routine we pull snd_nxt always up to snd_max. 12083 */ 12084 doing_tlp = 0; 12085 #ifdef BBR_INVARIANTS 12086 doing_retran_from = picked_up_retran = 0; 12087 #endif 12088 error = 0; 12089 tso = 0; 12090 slot = 0; 12091 mtu = 0; 12092 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12093 sb_offset = tp->snd_max - tp->snd_una; 12094 flags = tcp_outflags[tp->t_state]; 12095 sack_rxmit = 0; 12096 len = 0; 12097 rsm = NULL; 12098 if (flags & TH_RST) { 12099 SOCKBUF_LOCK(sb); 12100 goto send; 12101 } 12102 recheck_resend: 12103 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12104 /* We need to always have one in reserve */ 12105 rsm = bbr_alloc(bbr); 12106 if (rsm == NULL) { 12107 error = ENOMEM; 12108 /* Lie to get on the hpts */ 12109 tot_len = tp->t_maxseg; 12110 if (hpts_calling) 12111 /* Retry in a ms */ 12112 slot = 1001; 12113 goto just_return_nolock; 12114 } 12115 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12116 bbr->r_ctl.rc_free_cnt++; 12117 rsm = NULL; 12118 } 12119 /* What do we send, a resend? */ 12120 if (bbr->r_ctl.rc_resend == NULL) { 12121 /* Check for rack timeout */ 12122 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12123 if (bbr->r_ctl.rc_resend) { 12124 #ifdef BBR_INVARIANTS 12125 picked_up_retran = 1; 12126 #endif 12127 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12128 } 12129 } 12130 if (bbr->r_ctl.rc_resend) { 12131 rsm = bbr->r_ctl.rc_resend; 12132 #ifdef BBR_INVARIANTS 12133 doing_retran_from = 1; 12134 #endif 12135 /* Remove any TLP flags its a RACK or T-O */ 12136 rsm->r_flags &= ~BBR_TLP; 12137 bbr->r_ctl.rc_resend = NULL; 12138 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12139 #ifdef BBR_INVARIANTS 12140 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12141 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12142 goto recheck_resend; 12143 #else 12144 /* TSNH */ 12145 rsm = NULL; 12146 goto recheck_resend; 12147 #endif 12148 } 12149 if (rsm->r_flags & BBR_HAS_SYN) { 12150 /* Only retransmit a SYN by itself */ 12151 len = 0; 12152 if ((flags & TH_SYN) == 0) { 12153 /* Huh something is wrong */ 12154 rsm->r_start++; 12155 if (rsm->r_start == rsm->r_end) { 12156 /* Clean it up, somehow we missed the ack? */ 12157 bbr_log_syn(tp, NULL); 12158 } else { 12159 /* TFO with data? */ 12160 rsm->r_flags &= ~BBR_HAS_SYN; 12161 len = rsm->r_end - rsm->r_start; 12162 } 12163 } else { 12164 /* Retransmitting SYN */ 12165 rsm = NULL; 12166 SOCKBUF_LOCK(sb); 12167 goto send; 12168 } 12169 } else 12170 len = rsm->r_end - rsm->r_start; 12171 if ((bbr->rc_resends_use_tso == 0) && 12172 (len > maxseg)) { 12173 len = maxseg; 12174 more_to_rxt = 1; 12175 } 12176 sb_offset = rsm->r_start - tp->snd_una; 12177 if (len > 0) { 12178 sack_rxmit = 1; 12179 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12180 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12181 min(len, maxseg)); 12182 } else { 12183 /* I dont think this can happen */ 12184 rsm = NULL; 12185 goto recheck_resend; 12186 } 12187 BBR_STAT_INC(bbr_resends_set); 12188 } else if (bbr->r_ctl.rc_tlp_send) { 12189 /* 12190 * Tail loss probe 12191 */ 12192 doing_tlp = 1; 12193 rsm = bbr->r_ctl.rc_tlp_send; 12194 bbr->r_ctl.rc_tlp_send = NULL; 12195 sack_rxmit = 1; 12196 len = rsm->r_end - rsm->r_start; 12197 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12198 len = maxseg; 12199 12200 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12201 #ifdef BBR_INVARIANTS 12202 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12203 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12204 #else 12205 /* TSNH */ 12206 rsm = NULL; 12207 goto recheck_resend; 12208 #endif 12209 } 12210 sb_offset = rsm->r_start - tp->snd_una; 12211 BBR_STAT_INC(bbr_tlp_set); 12212 } 12213 /* 12214 * Enforce a connection sendmap count limit if set 12215 * as long as we are not retransmiting. 12216 */ 12217 if ((rsm == NULL) && 12218 (V_tcp_map_entries_limit > 0) && 12219 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12220 BBR_STAT_INC(bbr_alloc_limited); 12221 if (!bbr->alloc_limit_reported) { 12222 bbr->alloc_limit_reported = 1; 12223 BBR_STAT_INC(bbr_alloc_limited_conns); 12224 } 12225 goto just_return_nolock; 12226 } 12227 #ifdef BBR_INVARIANTS 12228 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12229 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12230 tp, bbr, rsm, sb_offset, len); 12231 } 12232 #endif 12233 /* 12234 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12235 * state flags. 12236 */ 12237 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12238 flags |= TH_FIN; 12239 if (tp->t_flags & TF_NEEDSYN) 12240 flags |= TH_SYN; 12241 12242 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12243 /* we are retransmitting the fin */ 12244 len--; 12245 if (len) { 12246 /* 12247 * When retransmitting data do *not* include the 12248 * FIN. This could happen from a TLP probe if we 12249 * allowed data with a FIN. 12250 */ 12251 flags &= ~TH_FIN; 12252 } 12253 } else if (rsm) { 12254 if (flags & TH_FIN) 12255 flags &= ~TH_FIN; 12256 } 12257 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12258 void *end_rsm; 12259 12260 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12261 if (end_rsm) 12262 kern_prefetch(end_rsm, &prefetch_rsm); 12263 prefetch_rsm = 1; 12264 } 12265 SOCKBUF_LOCK(sb); 12266 /* 12267 * If snd_nxt == snd_max and we have transmitted a FIN, the 12268 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12269 * negative length. This can also occur when TCP opens up its 12270 * congestion window while receiving additional duplicate acks after 12271 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12272 * the fast-retransmit. 12273 * 12274 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12275 * set to snd_una, the sb_offset will be 0, and the length may wind 12276 * up 0. 12277 * 12278 * If sack_rxmit is true we are retransmitting from the scoreboard 12279 * in which case len is already set. 12280 */ 12281 if (sack_rxmit == 0) { 12282 uint32_t avail; 12283 12284 avail = sbavail(sb); 12285 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12286 sb_offset = tp->snd_max - tp->snd_una; 12287 else 12288 sb_offset = 0; 12289 if (bbr->rc_tlp_new_data) { 12290 /* TLP is forcing out new data */ 12291 uint32_t tlplen; 12292 12293 doing_tlp = 1; 12294 tlplen = maxseg; 12295 12296 if (tlplen > (uint32_t)(avail - sb_offset)) { 12297 tlplen = (uint32_t)(avail - sb_offset); 12298 } 12299 if (tlplen > tp->snd_wnd) { 12300 len = tp->snd_wnd; 12301 } else { 12302 len = tlplen; 12303 } 12304 bbr->rc_tlp_new_data = 0; 12305 } else { 12306 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12307 if ((len < p_maxseg) && 12308 (bbr->rc_in_persist == 0) && 12309 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12310 ((avail - sb_offset) >= p_maxseg)) { 12311 /* 12312 * We are not completing whats in the socket 12313 * buffer (i.e. there is at least a segment 12314 * waiting to send) and we have 2 or more 12315 * segments outstanding. There is no sense 12316 * of sending a little piece. Lets defer and 12317 * and wait until we can send a whole 12318 * segment. 12319 */ 12320 len = 0; 12321 } 12322 if (bbr->rc_in_persist) { 12323 /* 12324 * We are in persists, figure out if 12325 * a retransmit is available (maybe the previous 12326 * persists we sent) or if we have to send new 12327 * data. 12328 */ 12329 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12330 if (rsm) { 12331 len = rsm->r_end - rsm->r_start; 12332 if (rsm->r_flags & BBR_HAS_FIN) 12333 len--; 12334 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12335 len = maxseg; 12336 if (len > 1) 12337 BBR_STAT_INC(bbr_persist_reneg); 12338 /* 12339 * XXXrrs we could force the len to 12340 * 1 byte here to cause the chunk to 12341 * split apart.. but that would then 12342 * mean we always retransmit it as 12343 * one byte even after the window 12344 * opens. 12345 */ 12346 sack_rxmit = 1; 12347 sb_offset = rsm->r_start - tp->snd_una; 12348 } else { 12349 /* 12350 * First time through in persists or peer 12351 * acked our one byte. Though we do have 12352 * to have something in the sb. 12353 */ 12354 len = 1; 12355 sb_offset = 0; 12356 if (avail == 0) 12357 len = 0; 12358 } 12359 } 12360 } 12361 } 12362 if (prefetch_so_done == 0) { 12363 kern_prefetch(so, &prefetch_so_done); 12364 prefetch_so_done = 1; 12365 } 12366 /* 12367 * Lop off SYN bit if it has already been sent. However, if this is 12368 * SYN-SENT state and if segment contains data and if we don't know 12369 * that foreign host supports TAO, suppress sending segment. 12370 */ 12371 if ((flags & TH_SYN) && (rsm == NULL) && 12372 SEQ_GT(tp->snd_max, tp->snd_una)) { 12373 if (tp->t_state != TCPS_SYN_RECEIVED) 12374 flags &= ~TH_SYN; 12375 /* 12376 * When sending additional segments following a TFO SYN|ACK, 12377 * do not include the SYN bit. 12378 */ 12379 if (IS_FASTOPEN(tp->t_flags) && 12380 (tp->t_state == TCPS_SYN_RECEIVED)) 12381 flags &= ~TH_SYN; 12382 sb_offset--, len++; 12383 if (sbavail(sb) == 0) 12384 len = 0; 12385 } else if ((flags & TH_SYN) && rsm) { 12386 /* 12387 * Subtract one from the len for the SYN being 12388 * retransmitted. 12389 */ 12390 len--; 12391 } 12392 /* 12393 * Be careful not to send data and/or FIN on SYN segments. This 12394 * measure is needed to prevent interoperability problems with not 12395 * fully conformant TCP implementations. 12396 */ 12397 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12398 len = 0; 12399 flags &= ~TH_FIN; 12400 } 12401 /* 12402 * On TFO sockets, ensure no data is sent in the following cases: 12403 * 12404 * - When retransmitting SYN|ACK on a passively-created socket 12405 * - When retransmitting SYN on an actively created socket 12406 * - When sending a zero-length cookie (cookie request) on an 12407 * actively created socket 12408 * - When the socket is in the CLOSED state (RST is being sent) 12409 */ 12410 if (IS_FASTOPEN(tp->t_flags) && 12411 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12412 ((tp->t_state == TCPS_SYN_SENT) && 12413 (tp->t_tfo_client_cookie_len == 0)) || 12414 (flags & TH_RST))) { 12415 len = 0; 12416 sack_rxmit = 0; 12417 rsm = NULL; 12418 } 12419 /* Without fast-open there should never be data sent on a SYN */ 12420 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 12421 len = 0; 12422 if (len <= 0) { 12423 /* 12424 * If FIN has been sent but not acked, but we haven't been 12425 * called to retransmit, len will be < 0. Otherwise, window 12426 * shrank after we sent into it. If window shrank to 0, 12427 * cancel pending retransmit, pull snd_nxt back to (closed) 12428 * window, and set the persist timer if it isn't already 12429 * going. If the window didn't close completely, just wait 12430 * for an ACK. 12431 * 12432 * We also do a general check here to ensure that we will 12433 * set the persist timer when we have data to send, but a 12434 * 0-byte window. This makes sure the persist timer is set 12435 * even if the packet hits one of the "goto send" lines 12436 * below. 12437 */ 12438 len = 0; 12439 if ((tp->snd_wnd == 0) && 12440 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12441 (tp->snd_una == tp->snd_max) && 12442 (sb_offset < (int)sbavail(sb))) { 12443 /* 12444 * Not enough room in the rwnd to send 12445 * a paced segment out. 12446 */ 12447 bbr_enter_persist(tp, bbr, cts, __LINE__); 12448 } 12449 } else if ((rsm == NULL) && 12450 (doing_tlp == 0) && 12451 (len < bbr->r_ctl.rc_pace_max_segs)) { 12452 /* 12453 * We are not sending a full segment for 12454 * some reason. Should we not send anything (think 12455 * sws or persists)? 12456 */ 12457 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12458 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12459 (len < (int)(sbavail(sb) - sb_offset))) { 12460 /* 12461 * Here the rwnd is less than 12462 * the pacing size, this is not a retransmit, 12463 * we are established and 12464 * the send is not the last in the socket buffer 12465 * lets not send, and possibly enter persists. 12466 */ 12467 len = 0; 12468 if (tp->snd_max == tp->snd_una) 12469 bbr_enter_persist(tp, bbr, cts, __LINE__); 12470 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12471 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12472 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12473 (len < (int)(sbavail(sb) - sb_offset)) && 12474 (len < bbr_minseg(bbr))) { 12475 /* 12476 * Here we are not retransmitting, and 12477 * the cwnd is not so small that we could 12478 * not send at least a min size (rxt timer 12479 * not having gone off), We have 2 segments or 12480 * more already in flight, its not the tail end 12481 * of the socket buffer and the cwnd is blocking 12482 * us from sending out minimum pacing segment size. 12483 * Lets not send anything. 12484 */ 12485 bbr->rc_cwnd_limited = 1; 12486 len = 0; 12487 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12488 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12489 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12490 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12491 (len < (int)(sbavail(sb) - sb_offset)) && 12492 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12493 /* 12494 * Here we have a send window but we have 12495 * filled it up and we can't send another pacing segment. 12496 * We also have in flight more than 2 segments 12497 * and we are not completing the sb i.e. we allow 12498 * the last bytes of the sb to go out even if 12499 * its not a full pacing segment. 12500 */ 12501 len = 0; 12502 } 12503 } 12504 /* len will be >= 0 after this point. */ 12505 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12506 tcp_sndbuf_autoscale(tp, so, sendwin); 12507 /* 12508 * 12509 */ 12510 if (bbr->rc_in_persist && 12511 len && 12512 (rsm == NULL) && 12513 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12514 /* 12515 * We are in persist, not doing a retransmit and don't have enough space 12516 * yet to send a full TSO. So is it at the end of the sb 12517 * if so we need to send else nuke to 0 and don't send. 12518 */ 12519 int sbleft; 12520 if (sbavail(sb) > sb_offset) 12521 sbleft = sbavail(sb) - sb_offset; 12522 else 12523 sbleft = 0; 12524 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12525 /* not at end of sb lets not send */ 12526 len = 0; 12527 } 12528 } 12529 /* 12530 * Decide if we can use TCP Segmentation Offloading (if supported by 12531 * hardware). 12532 * 12533 * TSO may only be used if we are in a pure bulk sending state. The 12534 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12535 * options prevent using TSO. With TSO the TCP header is the same 12536 * (except for the sequence number) for all generated packets. This 12537 * makes it impossible to transmit any options which vary per 12538 * generated segment or packet. 12539 * 12540 * IPv4 handling has a clear separation of ip options and ip header 12541 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12542 * does the right thing below to provide length of just ip options 12543 * and thus checking for ipoptlen is enough to decide if ip options 12544 * are present. 12545 */ 12546 #ifdef INET6 12547 if (isipv6) 12548 ipoptlen = ip6_optlen(inp); 12549 else 12550 #endif 12551 if (inp->inp_options) 12552 ipoptlen = inp->inp_options->m_len - 12553 offsetof(struct ipoption, ipopt_list); 12554 else 12555 ipoptlen = 0; 12556 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12557 /* 12558 * Pre-calculate here as we save another lookup into the darknesses 12559 * of IPsec that way and can actually decide if TSO is ok. 12560 */ 12561 #ifdef INET6 12562 if (isipv6 && IPSEC_ENABLED(ipv6)) 12563 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12564 #ifdef INET 12565 else 12566 #endif 12567 #endif /* INET6 */ 12568 #ifdef INET 12569 if (IPSEC_ENABLED(ipv4)) 12570 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12571 #endif /* INET */ 12572 #endif /* IPSEC */ 12573 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12574 ipoptlen += ipsec_optlen; 12575 #endif 12576 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12577 (len > maxseg) && 12578 (tp->t_port == 0) && 12579 ((tp->t_flags & TF_SIGNATURE) == 0) && 12580 tp->rcv_numsacks == 0 && 12581 ipoptlen == 0) 12582 tso = 1; 12583 12584 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12585 (long)TCP_MAXWIN << tp->rcv_scale); 12586 /* 12587 * Sender silly window avoidance. We transmit under the following 12588 * conditions when len is non-zero: 12589 * 12590 * - We have a full segment (or more with TSO) - This is the last 12591 * buffer in a write()/send() and we are either idle or running 12592 * NODELAY - we've timed out (e.g. persist timer) - we have more 12593 * then 1/2 the maximum send window's worth of data (receiver may be 12594 * limited the window size) - we need to retransmit 12595 */ 12596 if (rsm) 12597 goto send; 12598 if (len) { 12599 if (sack_rxmit) 12600 goto send; 12601 if (len >= p_maxseg) 12602 goto send; 12603 /* 12604 * NOTE! on localhost connections an 'ack' from the remote 12605 * end may occur synchronously with the output and cause us 12606 * to flush a buffer queued with moretocome. XXX 12607 * 12608 */ 12609 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12610 ((tp->t_flags & TF_NODELAY) || 12611 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12612 (tp->t_flags & TF_NOPUSH) == 0) { 12613 goto send; 12614 } 12615 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12616 goto send; 12617 } 12618 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12619 goto send; 12620 } 12621 } 12622 /* 12623 * Sending of standalone window updates. 12624 * 12625 * Window updates are important when we close our window due to a 12626 * full socket buffer and are opening it again after the application 12627 * reads data from it. Once the window has opened again and the 12628 * remote end starts to send again the ACK clock takes over and 12629 * provides the most current window information. 12630 * 12631 * We must avoid the silly window syndrome whereas every read from 12632 * the receive buffer, no matter how small, causes a window update 12633 * to be sent. We also should avoid sending a flurry of window 12634 * updates when the socket buffer had queued a lot of data and the 12635 * application is doing small reads. 12636 * 12637 * Prevent a flurry of pointless window updates by only sending an 12638 * update when we can increase the advertized window by more than 12639 * 1/4th of the socket buffer capacity. When the buffer is getting 12640 * full or is very small be more aggressive and send an update 12641 * whenever we can increase by two mss sized segments. In all other 12642 * situations the ACK's to new incoming data will carry further 12643 * window increases. 12644 * 12645 * Don't send an independent window update if a delayed ACK is 12646 * pending (it will get piggy-backed on it) or the remote side 12647 * already has done a half-close and won't send more data. Skip 12648 * this if the connection is in T/TCP half-open state. 12649 */ 12650 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12651 !(tp->t_flags & TF_DELACK) && 12652 !TCPS_HAVERCVDFIN(tp->t_state)) { 12653 /* Check to see if we should do a window update */ 12654 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12655 goto send; 12656 } 12657 /* 12658 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12659 * is also a catch-all for the retransmit timer timeout case. 12660 */ 12661 if (tp->t_flags & TF_ACKNOW) { 12662 goto send; 12663 } 12664 if (flags & TH_RST) { 12665 /* Always send a RST if one is due */ 12666 goto send; 12667 } 12668 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12669 goto send; 12670 } 12671 /* 12672 * If our state indicates that FIN should be sent and we have not 12673 * yet done so, then we need to send. 12674 */ 12675 if (flags & TH_FIN && 12676 ((tp->t_flags & TF_SENTFIN) == 0)) { 12677 goto send; 12678 } 12679 /* 12680 * No reason to send a segment, just return. 12681 */ 12682 just_return: 12683 SOCKBUF_UNLOCK(sb); 12684 just_return_nolock: 12685 if (tot_len) 12686 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12687 if (bbr->rc_no_pacing) 12688 slot = 0; 12689 if (tot_len == 0) { 12690 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12691 tp->snd_wnd) { 12692 BBR_STAT_INC(bbr_rwnd_limited); 12693 app_limited = BBR_JR_RWND_LIMITED; 12694 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12695 if ((bbr->rc_in_persist == 0) && 12696 TCPS_HAVEESTABLISHED(tp->t_state) && 12697 (tp->snd_max == tp->snd_una) && 12698 sbavail(&so->so_snd)) { 12699 /* No send window.. we must enter persist */ 12700 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12701 } 12702 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12703 BBR_STAT_INC(bbr_app_limited); 12704 app_limited = BBR_JR_APP_LIMITED; 12705 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12706 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12707 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12708 BBR_STAT_INC(bbr_cwnd_limited); 12709 app_limited = BBR_JR_CWND_LIMITED; 12710 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12711 bbr->r_ctl.rc_lost_bytes))); 12712 bbr->rc_cwnd_limited = 1; 12713 } else { 12714 BBR_STAT_INC(bbr_app_limited); 12715 app_limited = BBR_JR_APP_LIMITED; 12716 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12717 } 12718 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12719 bbr->r_agg_early_set = 0; 12720 bbr->r_ctl.rc_agg_early = 0; 12721 bbr->r_ctl.rc_last_delay_val = 0; 12722 } else if (bbr->rc_use_google == 0) 12723 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12724 /* Are we app limited? */ 12725 if ((app_limited == BBR_JR_APP_LIMITED) || 12726 (app_limited == BBR_JR_RWND_LIMITED)) { 12727 /** 12728 * We are application limited. 12729 */ 12730 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12731 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12732 } 12733 if (tot_len == 0) 12734 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12735 /* Dont update the time if we did not send */ 12736 bbr->r_ctl.rc_last_delay_val = 0; 12737 bbr->rc_output_starts_timer = 1; 12738 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12739 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12740 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12741 /* Make sure snd_nxt is drug up */ 12742 tp->snd_nxt = tp->snd_max; 12743 } 12744 return (error); 12745 12746 send: 12747 if (doing_tlp == 0) { 12748 /* 12749 * Data not a TLP, and its not the rxt firing. If it is the 12750 * rxt firing, we want to leave the tlp_in_progress flag on 12751 * so we don't send another TLP. It has to be a rack timer 12752 * or normal send (response to acked data) to clear the tlp 12753 * in progress flag. 12754 */ 12755 bbr->rc_tlp_in_progress = 0; 12756 bbr->rc_tlp_rtx_out = 0; 12757 } else { 12758 /* 12759 * Its a TLP. 12760 */ 12761 bbr->rc_tlp_in_progress = 1; 12762 } 12763 bbr_timer_cancel(bbr, __LINE__, cts); 12764 if (rsm == NULL) { 12765 if (sbused(sb) > 0) { 12766 /* 12767 * This is sub-optimal. We only send a stand alone 12768 * FIN on its own segment. 12769 */ 12770 if (flags & TH_FIN) { 12771 flags &= ~TH_FIN; 12772 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12773 /* Lets not send this */ 12774 slot = 0; 12775 goto just_return; 12776 } 12777 } 12778 } 12779 } else { 12780 /* 12781 * We do *not* send a FIN on a retransmit if it has data. 12782 * The if clause here where len > 1 should never come true. 12783 */ 12784 if ((len > 0) && 12785 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12786 (flags & TH_FIN))) { 12787 flags &= ~TH_FIN; 12788 len--; 12789 } 12790 } 12791 SOCKBUF_LOCK_ASSERT(sb); 12792 if (len > 0) { 12793 if ((tp->snd_una == tp->snd_max) && 12794 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12795 /* 12796 * This qualifies as a RTT_PROBE session since we 12797 * drop the data outstanding to nothing and waited 12798 * more than bbr_rtt_probe_time. 12799 */ 12800 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12801 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12802 } 12803 if (len >= maxseg) 12804 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12805 else 12806 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12807 } 12808 /* 12809 * Before ESTABLISHED, force sending of initial options unless TCP 12810 * set not to do any options. NOTE: we assume that the IP/TCP header 12811 * plus TCP options always fit in a single mbuf, leaving room for a 12812 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12813 * + optlen <= MCLBYTES 12814 */ 12815 optlen = 0; 12816 #ifdef INET6 12817 if (isipv6) 12818 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12819 else 12820 #endif 12821 hdrlen = sizeof(struct tcpiphdr); 12822 12823 /* 12824 * Compute options for segment. We only have to care about SYN and 12825 * established connection segments. Options for SYN-ACK segments 12826 * are handled in TCP syncache. 12827 */ 12828 to.to_flags = 0; 12829 local_options = 0; 12830 if ((tp->t_flags & TF_NOOPT) == 0) { 12831 /* Maximum segment size. */ 12832 if (flags & TH_SYN) { 12833 to.to_mss = tcp_mssopt(&inp->inp_inc); 12834 if (tp->t_port) 12835 to.to_mss -= V_tcp_udp_tunneling_overhead; 12836 to.to_flags |= TOF_MSS; 12837 /* 12838 * On SYN or SYN|ACK transmits on TFO connections, 12839 * only include the TFO option if it is not a 12840 * retransmit, as the presence of the TFO option may 12841 * have caused the original SYN or SYN|ACK to have 12842 * been dropped by a middlebox. 12843 */ 12844 if (IS_FASTOPEN(tp->t_flags) && 12845 (tp->t_rxtshift == 0)) { 12846 if (tp->t_state == TCPS_SYN_RECEIVED) { 12847 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12848 to.to_tfo_cookie = 12849 (u_int8_t *)&tp->t_tfo_cookie.server; 12850 to.to_flags |= TOF_FASTOPEN; 12851 wanted_cookie = 1; 12852 } else if (tp->t_state == TCPS_SYN_SENT) { 12853 to.to_tfo_len = 12854 tp->t_tfo_client_cookie_len; 12855 to.to_tfo_cookie = 12856 tp->t_tfo_cookie.client; 12857 to.to_flags |= TOF_FASTOPEN; 12858 wanted_cookie = 1; 12859 } 12860 } 12861 } 12862 /* Window scaling. */ 12863 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12864 to.to_wscale = tp->request_r_scale; 12865 to.to_flags |= TOF_SCALE; 12866 } 12867 /* Timestamps. */ 12868 if ((tp->t_flags & TF_RCVD_TSTMP) || 12869 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12870 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12871 to.to_tsecr = tp->ts_recent; 12872 to.to_flags |= TOF_TS; 12873 local_options += TCPOLEN_TIMESTAMP + 2; 12874 } 12875 /* Set receive buffer autosizing timestamp. */ 12876 if (tp->rfbuf_ts == 0 && 12877 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12878 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12879 /* Selective ACK's. */ 12880 if (flags & TH_SYN) 12881 to.to_flags |= TOF_SACKPERM; 12882 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12883 tp->rcv_numsacks > 0) { 12884 to.to_flags |= TOF_SACK; 12885 to.to_nsacks = tp->rcv_numsacks; 12886 to.to_sacks = (u_char *)tp->sackblks; 12887 } 12888 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 12889 /* TCP-MD5 (RFC2385). */ 12890 if (tp->t_flags & TF_SIGNATURE) 12891 to.to_flags |= TOF_SIGNATURE; 12892 #endif /* TCP_SIGNATURE */ 12893 12894 /* Processing the options. */ 12895 hdrlen += (optlen = tcp_addoptions(&to, opt)); 12896 /* 12897 * If we wanted a TFO option to be added, but it was unable 12898 * to fit, ensure no data is sent. 12899 */ 12900 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 12901 !(to.to_flags & TOF_FASTOPEN)) 12902 len = 0; 12903 } 12904 if (tp->t_port) { 12905 if (V_tcp_udp_tunneling_port == 0) { 12906 /* The port was removed?? */ 12907 SOCKBUF_UNLOCK(&so->so_snd); 12908 return (EHOSTUNREACH); 12909 } 12910 hdrlen += sizeof(struct udphdr); 12911 } 12912 #ifdef INET6 12913 if (isipv6) 12914 ipoptlen = ip6_optlen(inp); 12915 else 12916 #endif 12917 if (inp->inp_options) 12918 ipoptlen = inp->inp_options->m_len - 12919 offsetof(struct ipoption, ipopt_list); 12920 else 12921 ipoptlen = 0; 12922 ipoptlen = 0; 12923 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12924 ipoptlen += ipsec_optlen; 12925 #endif 12926 if (bbr->rc_last_options != local_options) { 12927 /* 12928 * Cache the options length this generally does not change 12929 * on a connection. We use this to calculate TSO. 12930 */ 12931 bbr->rc_last_options = local_options; 12932 } 12933 maxseg = tp->t_maxseg - (ipoptlen + optlen); 12934 p_maxseg = min(maxseg, pace_max_segs); 12935 /* 12936 * Adjust data length if insertion of options will bump the packet 12937 * length beyond the t_maxseg length. Clear the FIN bit because we 12938 * cut off the tail of the segment. 12939 */ 12940 if (len > maxseg) { 12941 if (len != 0 && (flags & TH_FIN)) { 12942 flags &= ~TH_FIN; 12943 } 12944 if (tso) { 12945 uint32_t moff; 12946 int32_t max_len; 12947 12948 /* extract TSO information */ 12949 if_hw_tsomax = tp->t_tsomax; 12950 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 12951 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 12952 KASSERT(ipoptlen == 0, 12953 ("%s: TSO can't do IP options", __func__)); 12954 12955 /* 12956 * Check if we should limit by maximum payload 12957 * length: 12958 */ 12959 if (if_hw_tsomax != 0) { 12960 /* compute maximum TSO length */ 12961 max_len = (if_hw_tsomax - hdrlen - 12962 max_linkhdr); 12963 if (max_len <= 0) { 12964 len = 0; 12965 } else if (len > max_len) { 12966 len = max_len; 12967 } 12968 } 12969 /* 12970 * Prevent the last segment from being fractional 12971 * unless the send sockbuf can be emptied: 12972 */ 12973 if ((sb_offset + len) < sbavail(sb)) { 12974 moff = len % (uint32_t)maxseg; 12975 if (moff != 0) { 12976 len -= moff; 12977 } 12978 } 12979 /* 12980 * In case there are too many small fragments don't 12981 * use TSO: 12982 */ 12983 if (len <= maxseg) { 12984 len = maxseg; 12985 tso = 0; 12986 } 12987 } else { 12988 /* Not doing TSO */ 12989 if (optlen + ipoptlen >= tp->t_maxseg) { 12990 /* 12991 * Since we don't have enough space to put 12992 * the IP header chain and the TCP header in 12993 * one packet as required by RFC 7112, don't 12994 * send it. Also ensure that at least one 12995 * byte of the payload can be put into the 12996 * TCP segment. 12997 */ 12998 SOCKBUF_UNLOCK(&so->so_snd); 12999 error = EMSGSIZE; 13000 sack_rxmit = 0; 13001 goto out; 13002 } 13003 len = maxseg; 13004 } 13005 } else { 13006 /* Not doing TSO */ 13007 if_hw_tsomaxsegcount = 0; 13008 tso = 0; 13009 } 13010 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13011 ("%s: len > IP_MAXPACKET", __func__)); 13012 #ifdef DIAGNOSTIC 13013 #ifdef INET6 13014 if (max_linkhdr + hdrlen > MCLBYTES) 13015 #else 13016 if (max_linkhdr + hdrlen > MHLEN) 13017 #endif 13018 panic("tcphdr too big"); 13019 #endif 13020 /* 13021 * This KASSERT is here to catch edge cases at a well defined place. 13022 * Before, those had triggered (random) panic conditions further 13023 * down. 13024 */ 13025 #ifdef BBR_INVARIANTS 13026 if (sack_rxmit) { 13027 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13028 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13029 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13030 } 13031 } 13032 #endif 13033 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13034 if ((len == 0) && 13035 (flags & TH_FIN) && 13036 (sbused(sb))) { 13037 /* 13038 * We have outstanding data, don't send a fin by itself!. 13039 */ 13040 slot = 0; 13041 goto just_return; 13042 } 13043 /* 13044 * Grab a header mbuf, attaching a copy of data to be transmitted, 13045 * and initialize the header from the template for sends on this 13046 * connection. 13047 */ 13048 if (len) { 13049 uint32_t moff; 13050 13051 /* 13052 * We place a limit on sending with hptsi. 13053 */ 13054 if ((rsm == NULL) && len > pace_max_segs) 13055 len = pace_max_segs; 13056 if (len <= maxseg) 13057 tso = 0; 13058 #ifdef INET6 13059 if (MHLEN < hdrlen + max_linkhdr) 13060 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13061 else 13062 #endif 13063 m = m_gethdr(M_NOWAIT, MT_DATA); 13064 13065 if (m == NULL) { 13066 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13067 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13068 SOCKBUF_UNLOCK(sb); 13069 error = ENOBUFS; 13070 sack_rxmit = 0; 13071 goto out; 13072 } 13073 m->m_data += max_linkhdr; 13074 m->m_len = hdrlen; 13075 /* 13076 * Start the m_copy functions from the closest mbuf to the 13077 * sb_offset in the socket buffer chain. 13078 */ 13079 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13080 #ifdef BBR_INVARIANTS 13081 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13082 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13083 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13084 doing_retran_from, 13085 picked_up_retran, 13086 doing_tlp); 13087 13088 #endif 13089 /* 13090 * In this messed up situation we have two choices, 13091 * a) pretend the send worked, and just start timers 13092 * and what not (not good since that may lead us 13093 * back here a lot). <or> b) Send the lowest segment 13094 * in the map. <or> c) Drop the connection. Lets do 13095 * <b> which if it continues to happen will lead to 13096 * <c> via timeouts. 13097 */ 13098 BBR_STAT_INC(bbr_offset_recovery); 13099 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13100 sb_offset = 0; 13101 if (rsm == NULL) { 13102 sack_rxmit = 0; 13103 len = sbavail(sb); 13104 } else { 13105 sack_rxmit = 1; 13106 if (rsm->r_start != tp->snd_una) { 13107 /* 13108 * Things are really messed up, <c> 13109 * is the only thing to do. 13110 */ 13111 BBR_STAT_INC(bbr_offset_drop); 13112 SOCKBUF_UNLOCK(sb); 13113 (void)m_free(m); 13114 return (-EFAULT); /* tcp_drop() */ 13115 } 13116 len = rsm->r_end - rsm->r_start; 13117 } 13118 if (len > sbavail(sb)) 13119 len = sbavail(sb); 13120 if (len > maxseg) 13121 len = maxseg; 13122 } 13123 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13124 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13125 m_copydata(mb, moff, (int)len, 13126 mtod(m, caddr_t)+hdrlen); 13127 if (rsm == NULL) 13128 sbsndptr_adv(sb, mb, len); 13129 m->m_len += len; 13130 } else { 13131 struct sockbuf *msb; 13132 13133 if (rsm) 13134 msb = NULL; 13135 else 13136 msb = sb; 13137 #ifdef BBR_INVARIANTS 13138 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13139 if (rsm) { 13140 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 ", 13141 tp, bbr, len, moff, 13142 sbavail(sb), rsm, 13143 tp->snd_una, rsm->r_flags, rsm->r_start, 13144 doing_retran_from, 13145 picked_up_retran, 13146 doing_tlp, sack_rxmit); 13147 } else { 13148 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13149 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13150 } 13151 } 13152 #endif 13153 m->m_next = tcp_m_copym( 13154 mb, moff, &len, 13155 if_hw_tsomaxsegcount, 13156 if_hw_tsomaxsegsize, msb, 13157 ((rsm == NULL) ? hw_tls : 0) 13158 #ifdef NETFLIX_COPY_ARGS 13159 , NULL, NULL 13160 #endif 13161 ); 13162 if (len <= maxseg) { 13163 /* 13164 * Must have ran out of mbufs for the copy 13165 * shorten it to no longer need tso. Lets 13166 * not put on sendalot since we are low on 13167 * mbufs. 13168 */ 13169 tso = 0; 13170 } 13171 if (m->m_next == NULL) { 13172 SOCKBUF_UNLOCK(sb); 13173 (void)m_free(m); 13174 error = ENOBUFS; 13175 sack_rxmit = 0; 13176 goto out; 13177 } 13178 } 13179 #ifdef BBR_INVARIANTS 13180 if (tso && len < maxseg) { 13181 panic("tp:%p tso on, but len:%d < maxseg:%d", 13182 tp, len, maxseg); 13183 } 13184 if (tso && if_hw_tsomaxsegcount) { 13185 int32_t seg_cnt = 0; 13186 struct mbuf *foo; 13187 13188 foo = m; 13189 while (foo) { 13190 seg_cnt++; 13191 foo = foo->m_next; 13192 } 13193 if (seg_cnt > if_hw_tsomaxsegcount) { 13194 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13195 } 13196 } 13197 #endif 13198 /* 13199 * If we're sending everything we've got, set PUSH. (This 13200 * will keep happy those implementations which only give 13201 * data to the user when a buffer fills or a PUSH comes in.) 13202 */ 13203 if (sb_offset + len == sbused(sb) && 13204 sbused(sb) && 13205 !(flags & TH_SYN)) { 13206 flags |= TH_PUSH; 13207 } 13208 SOCKBUF_UNLOCK(sb); 13209 } else { 13210 SOCKBUF_UNLOCK(sb); 13211 if (tp->t_flags & TF_ACKNOW) 13212 KMOD_TCPSTAT_INC(tcps_sndacks); 13213 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13214 KMOD_TCPSTAT_INC(tcps_sndctrl); 13215 else 13216 KMOD_TCPSTAT_INC(tcps_sndwinup); 13217 13218 m = m_gethdr(M_NOWAIT, MT_DATA); 13219 if (m == NULL) { 13220 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13221 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13222 error = ENOBUFS; 13223 /* Fudge the send time since we could not send */ 13224 sack_rxmit = 0; 13225 goto out; 13226 } 13227 #ifdef INET6 13228 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13229 MHLEN >= hdrlen) { 13230 M_ALIGN(m, hdrlen); 13231 } else 13232 #endif 13233 m->m_data += max_linkhdr; 13234 m->m_len = hdrlen; 13235 } 13236 SOCKBUF_UNLOCK_ASSERT(sb); 13237 m->m_pkthdr.rcvif = (struct ifnet *)0; 13238 #ifdef MAC 13239 mac_inpcb_create_mbuf(inp, m); 13240 #endif 13241 #ifdef INET6 13242 if (isipv6) { 13243 ip6 = mtod(m, struct ip6_hdr *); 13244 if (tp->t_port) { 13245 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13246 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13247 udp->uh_dport = tp->t_port; 13248 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13249 udp->uh_ulen = htons(ulen); 13250 th = (struct tcphdr *)(udp + 1); 13251 } else { 13252 th = (struct tcphdr *)(ip6 + 1); 13253 } 13254 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13255 } else 13256 #endif /* INET6 */ 13257 { 13258 ip = mtod(m, struct ip *); 13259 if (tp->t_port) { 13260 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13261 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13262 udp->uh_dport = tp->t_port; 13263 ulen = hdrlen + len - sizeof(struct ip); 13264 udp->uh_ulen = htons(ulen); 13265 th = (struct tcphdr *)(udp + 1); 13266 } else { 13267 th = (struct tcphdr *)(ip + 1); 13268 } 13269 tcpip_fillheaders(inp, tp->t_port, ip, th); 13270 } 13271 /* 13272 * If we are doing retransmissions, then snd_nxt will not reflect 13273 * the first unsent octet. For ACK only packets, we do not want the 13274 * sequence number of the retransmitted packet, we want the sequence 13275 * number of the next unsent octet. So, if there is no data (and no 13276 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13277 * ti_seq. But if we are in persist state, snd_max might reflect 13278 * one byte beyond the right edge of the window, so use snd_nxt in 13279 * that case, since we know we aren't doing a retransmission. 13280 * (retransmit and persist are mutually exclusive...) 13281 */ 13282 if (sack_rxmit == 0) { 13283 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13284 /* New data (including new persists) */ 13285 th->th_seq = htonl(tp->snd_max); 13286 bbr_seq = tp->snd_max; 13287 } else if (flags & TH_SYN) { 13288 /* Syn's always send from iss */ 13289 th->th_seq = htonl(tp->iss); 13290 bbr_seq = tp->iss; 13291 } else if (flags & TH_FIN) { 13292 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13293 /* 13294 * If we sent the fin already its 1 minus 13295 * snd_max 13296 */ 13297 th->th_seq = (htonl(tp->snd_max - 1)); 13298 bbr_seq = (tp->snd_max - 1); 13299 } else { 13300 /* First time FIN use snd_max */ 13301 th->th_seq = htonl(tp->snd_max); 13302 bbr_seq = tp->snd_max; 13303 } 13304 } else { 13305 /* 13306 * len == 0 and not persist we use snd_max, sending 13307 * an ack unless we have sent the fin then its 1 13308 * minus. 13309 */ 13310 /* 13311 * XXXRRS Question if we are in persists and we have 13312 * nothing outstanding to send and we have not sent 13313 * a FIN, we will send an ACK. In such a case it 13314 * might be better to send (tp->snd_una - 1) which 13315 * would force the peer to ack. 13316 */ 13317 if (tp->t_flags & TF_SENTFIN) { 13318 th->th_seq = htonl(tp->snd_max - 1); 13319 bbr_seq = (tp->snd_max - 1); 13320 } else { 13321 th->th_seq = htonl(tp->snd_max); 13322 bbr_seq = tp->snd_max; 13323 } 13324 } 13325 } else { 13326 /* All retransmits use the rsm to guide the send */ 13327 th->th_seq = htonl(rsm->r_start); 13328 bbr_seq = rsm->r_start; 13329 } 13330 th->th_ack = htonl(tp->rcv_nxt); 13331 if (optlen) { 13332 bcopy(opt, th + 1, optlen); 13333 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13334 } 13335 tcp_set_flags(th, flags); 13336 /* 13337 * Calculate receive window. Don't shrink window, but avoid silly 13338 * window syndrome. 13339 */ 13340 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13341 recwin < maxseg))) 13342 recwin = 0; 13343 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13344 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13345 recwin = (tp->rcv_adv - tp->rcv_nxt); 13346 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13347 recwin = TCP_MAXWIN << tp->rcv_scale; 13348 13349 /* 13350 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13351 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13352 * handled in syncache. 13353 */ 13354 if (flags & TH_SYN) 13355 th->th_win = htons((u_short) 13356 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13357 else { 13358 /* Avoid shrinking window with window scaling. */ 13359 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13360 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13361 } 13362 /* 13363 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13364 * window. This may cause the remote transmitter to stall. This 13365 * flag tells soreceive() to disable delayed acknowledgements when 13366 * draining the buffer. This can occur if the receiver is 13367 * attempting to read more data than can be buffered prior to 13368 * transmitting on the connection. 13369 */ 13370 if (th->th_win == 0) { 13371 tp->t_sndzerowin++; 13372 tp->t_flags |= TF_RXWIN0SENT; 13373 } else 13374 tp->t_flags &= ~TF_RXWIN0SENT; 13375 /* 13376 * We don't support urgent data, but drag along 13377 * the pointer in case of a stack switch. 13378 */ 13379 tp->snd_up = tp->snd_una; 13380 13381 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13382 if (to.to_flags & TOF_SIGNATURE) { 13383 /* 13384 * Calculate MD5 signature and put it into the place 13385 * determined before. NOTE: since TCP options buffer doesn't 13386 * point into mbuf's data, calculate offset and use it. 13387 */ 13388 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13389 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13390 /* 13391 * Do not send segment if the calculation of MD5 13392 * digest has failed. 13393 */ 13394 goto out; 13395 } 13396 } 13397 #endif 13398 13399 /* 13400 * Put TCP length in extended header, and then checksum extended 13401 * header and data. 13402 */ 13403 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13404 #ifdef INET6 13405 if (isipv6) { 13406 /* 13407 * ip6_plen is not need to be filled now, and will be filled 13408 * in ip6_output. 13409 */ 13410 if (tp->t_port) { 13411 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13412 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13413 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13414 th->th_sum = htons(0); 13415 UDPSTAT_INC(udps_opackets); 13416 } else { 13417 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13418 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13419 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13420 optlen + len, IPPROTO_TCP, 0); 13421 } 13422 } 13423 #endif 13424 #if defined(INET6) && defined(INET) 13425 else 13426 #endif 13427 #ifdef INET 13428 { 13429 if (tp->t_port) { 13430 m->m_pkthdr.csum_flags = CSUM_UDP; 13431 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13432 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13433 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13434 th->th_sum = htons(0); 13435 UDPSTAT_INC(udps_opackets); 13436 } else { 13437 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13438 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13439 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13440 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13441 IPPROTO_TCP + len + optlen)); 13442 } 13443 /* IP version must be set here for ipv4/ipv6 checking later */ 13444 KASSERT(ip->ip_v == IPVERSION, 13445 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13446 } 13447 #endif 13448 13449 /* 13450 * Enable TSO and specify the size of the segments. The TCP pseudo 13451 * header checksum is always provided. XXX: Fixme: This is currently 13452 * not the case for IPv6. 13453 */ 13454 if (tso) { 13455 KASSERT(len > maxseg, 13456 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13457 m->m_pkthdr.csum_flags |= CSUM_TSO; 13458 csum_flags |= CSUM_TSO; 13459 m->m_pkthdr.tso_segsz = maxseg; 13460 } 13461 KASSERT(len + hdrlen == m_length(m, NULL), 13462 ("%s: mbuf chain different than expected: %d + %u != %u", 13463 __func__, len, hdrlen, m_length(m, NULL))); 13464 13465 #ifdef TCP_HHOOK 13466 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13467 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13468 #endif 13469 13470 /* Log to the black box */ 13471 if (tcp_bblogging_on(tp)) { 13472 union tcp_log_stackspecific log; 13473 13474 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13475 /* Record info on type of transmission */ 13476 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13477 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13478 log.u_bbr.flex3 = maxseg; 13479 log.u_bbr.flex4 = delay_calc; 13480 log.u_bbr.flex5 = bbr->rc_past_init_win; 13481 log.u_bbr.flex5 <<= 1; 13482 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13483 log.u_bbr.flex5 <<= 29; 13484 log.u_bbr.flex5 |= tp->t_maxseg; 13485 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13486 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13487 /* lets poke in the low and the high here for debugging */ 13488 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13489 if (rsm || sack_rxmit) { 13490 if (doing_tlp) 13491 log.u_bbr.flex8 = 2; 13492 else 13493 log.u_bbr.flex8 = 1; 13494 } else { 13495 log.u_bbr.flex8 = 0; 13496 } 13497 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13498 len, &log, false, NULL, NULL, 0, tv); 13499 } else { 13500 lgb = NULL; 13501 } 13502 /* 13503 * Fill in IP length and desired time to live and send to IP level. 13504 * There should be a better way to handle ttl and tos; we could keep 13505 * them in the template, but need a way to checksum without them. 13506 */ 13507 /* 13508 * m->m_pkthdr.len should have been set before cksum calcuration, 13509 * because in6_cksum() need it. 13510 */ 13511 #ifdef INET6 13512 if (isipv6) { 13513 /* 13514 * we separately set hoplimit for every segment, since the 13515 * user might want to change the value via setsockopt. Also, 13516 * desired default hop limit might be changed via Neighbor 13517 * Discovery. 13518 */ 13519 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13520 13521 /* 13522 * Set the packet size here for the benefit of DTrace 13523 * probes. ip6_output() will set it properly; it's supposed 13524 * to include the option header lengths as well. 13525 */ 13526 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13527 13528 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13529 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13530 else 13531 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13532 13533 if (tp->t_state == TCPS_SYN_SENT) 13534 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13535 13536 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13537 /* TODO: IPv6 IP6TOS_ECT bit on */ 13538 error = ip6_output(m, inp->in6p_outputopts, 13539 &inp->inp_route6, 13540 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13541 NULL, NULL, inp); 13542 13543 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13544 mtu = inp->inp_route6.ro_nh->nh_mtu; 13545 } 13546 #endif /* INET6 */ 13547 #if defined(INET) && defined(INET6) 13548 else 13549 #endif 13550 #ifdef INET 13551 { 13552 ip->ip_len = htons(m->m_pkthdr.len); 13553 #ifdef INET6 13554 if (isipv6) 13555 ip->ip_ttl = in6_selecthlim(inp, NULL); 13556 #endif /* INET6 */ 13557 /* 13558 * If we do path MTU discovery, then we set DF on every 13559 * packet. This might not be the best thing to do according 13560 * to RFC3390 Section 2. However the tcp hostcache migitates 13561 * the problem so it affects only the first tcp connection 13562 * with a host. 13563 * 13564 * NB: Don't set DF on small MTU/MSS to have a safe 13565 * fallback. 13566 */ 13567 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13568 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13569 if (tp->t_port == 0 || len < V_tcp_minmss) { 13570 ip->ip_off |= htons(IP_DF); 13571 } 13572 } else { 13573 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13574 } 13575 13576 if (tp->t_state == TCPS_SYN_SENT) 13577 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13578 13579 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13580 13581 error = ip_output(m, inp->inp_options, &inp->inp_route, 13582 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13583 inp); 13584 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13585 mtu = inp->inp_route.ro_nh->nh_mtu; 13586 } 13587 #endif /* INET */ 13588 out: 13589 13590 if (lgb) { 13591 lgb->tlb_errno = error; 13592 lgb = NULL; 13593 } 13594 /* 13595 * In transmit state, time the transmission and arrange for the 13596 * retransmit. In persist state, just set snd_max. 13597 */ 13598 if (error == 0) { 13599 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13600 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13601 (tp->t_flags & TF_SACK_PERMIT) && 13602 tp->rcv_numsacks > 0) 13603 tcp_clean_dsack_blocks(tp); 13604 /* We sent an ack clear the bbr_segs_rcvd count */ 13605 bbr->output_error_seen = 0; 13606 bbr->oerror_cnt = 0; 13607 bbr->bbr_segs_rcvd = 0; 13608 if (len == 0) 13609 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13610 /* Do accounting for new sends */ 13611 if ((len > 0) && (rsm == NULL)) { 13612 int idx; 13613 if (tp->snd_una == tp->snd_max) { 13614 /* 13615 * Special case to match google, when 13616 * nothing is in flight the delivered 13617 * time does get updated to the current 13618 * time (see tcp_rate_bsd.c). 13619 */ 13620 bbr->r_ctl.rc_del_time = cts; 13621 } 13622 if (len >= maxseg) { 13623 idx = (len / maxseg) + 3; 13624 if (idx >= TCP_MSS_ACCT_ATIMER) 13625 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13626 else 13627 counter_u64_add(bbr_out_size[idx], 1); 13628 } else { 13629 /* smaller than a MSS */ 13630 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13631 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13632 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13633 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13634 } 13635 } 13636 } 13637 abandon = 0; 13638 /* 13639 * We must do the send accounting before we log the output, 13640 * otherwise the state of the rsm could change and we account to the 13641 * wrong bucket. 13642 */ 13643 if (len > 0) { 13644 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13645 if (error == 0) { 13646 if (tp->snd_una == tp->snd_max) 13647 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13648 } 13649 } 13650 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13651 cts, mb, &abandon, rsm, 0, sb); 13652 if (abandon) { 13653 /* 13654 * If bbr_log_output destroys the TCB or sees a TH_RST being 13655 * sent we should hit this condition. 13656 */ 13657 return (0); 13658 } 13659 if (bbr->rc_in_persist == 0) { 13660 /* 13661 * Advance snd_nxt over sequence space of this segment. 13662 */ 13663 if (error) 13664 /* We don't log or do anything with errors */ 13665 goto skip_upd; 13666 13667 if (tp->snd_una == tp->snd_max && 13668 (len || (flags & (TH_SYN | TH_FIN)))) { 13669 /* 13670 * Update the time we just added data since none was 13671 * outstanding. 13672 */ 13673 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13674 bbr->rc_tp->t_acktime = ticks; 13675 } 13676 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13677 if (flags & TH_SYN) { 13678 /* 13679 * Smack the snd_max to iss + 1 13680 * if its a FO we will add len below. 13681 */ 13682 tp->snd_max = tp->iss + 1; 13683 } 13684 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13685 tp->snd_max++; 13686 tp->t_flags |= TF_SENTFIN; 13687 } 13688 } 13689 if (sack_rxmit == 0) 13690 tp->snd_max += len; 13691 skip_upd: 13692 if ((error == 0) && len) 13693 tot_len += len; 13694 } else { 13695 /* Persists case */ 13696 int32_t xlen = len; 13697 13698 if (error) 13699 goto nomore; 13700 13701 if (flags & TH_SYN) 13702 ++xlen; 13703 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13704 ++xlen; 13705 tp->t_flags |= TF_SENTFIN; 13706 } 13707 if (xlen && (tp->snd_una == tp->snd_max)) { 13708 /* 13709 * Update the time we just added data since none was 13710 * outstanding. 13711 */ 13712 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13713 bbr->rc_tp->t_acktime = ticks; 13714 } 13715 if (sack_rxmit == 0) 13716 tp->snd_max += xlen; 13717 tot_len += (len + optlen + ipoptlen); 13718 } 13719 nomore: 13720 if (error) { 13721 /* 13722 * Failures do not advance the seq counter above. For the 13723 * case of ENOBUFS we will fall out and become ack-clocked. 13724 * capping the cwnd at the current flight. 13725 * Everything else will just have to retransmit with the timer 13726 * (no pacer). 13727 */ 13728 SOCKBUF_UNLOCK_ASSERT(sb); 13729 BBR_STAT_INC(bbr_saw_oerr); 13730 /* Clear all delay/early tracks */ 13731 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13732 bbr->r_ctl.rc_agg_early = 0; 13733 bbr->r_agg_early_set = 0; 13734 bbr->output_error_seen = 1; 13735 if (bbr->oerror_cnt < 0xf) 13736 bbr->oerror_cnt++; 13737 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13738 /* drop the session */ 13739 return (-ENETDOWN); 13740 } 13741 switch (error) { 13742 case ENOBUFS: 13743 /* 13744 * Make this guy have to get ack's to send 13745 * more but lets make sure we don't 13746 * slam him below a T-O (1MSS). 13747 */ 13748 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13749 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13750 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13751 if (tp->snd_cwnd < maxseg) 13752 tp->snd_cwnd = maxseg; 13753 } 13754 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13755 BBR_STAT_INC(bbr_saw_enobuf); 13756 if (bbr->bbr_hdrw_pacing) 13757 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13758 else 13759 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13760 /* 13761 * Here even in the enobuf's case we want to do our 13762 * state update. The reason being we may have been 13763 * called by the input function. If so we have had 13764 * things change. 13765 */ 13766 error = 0; 13767 goto enobufs; 13768 case EMSGSIZE: 13769 /* 13770 * For some reason the interface we used initially 13771 * to send segments changed to another or lowered 13772 * its MTU. If TSO was active we either got an 13773 * interface without TSO capabilits or TSO was 13774 * turned off. If we obtained mtu from ip_output() 13775 * then update it and try again. 13776 */ 13777 /* Turn on tracing (or try to) */ 13778 { 13779 int old_maxseg; 13780 13781 old_maxseg = tp->t_maxseg; 13782 BBR_STAT_INC(bbr_saw_emsgsiz); 13783 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13784 if (mtu != 0) 13785 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13786 if (old_maxseg <= tp->t_maxseg) { 13787 /* Huh it did not shrink? */ 13788 tp->t_maxseg = old_maxseg - 40; 13789 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13790 } 13791 /* 13792 * Nuke all other things that can interfere 13793 * with slot 13794 */ 13795 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13796 slot = bbr_get_pacing_delay(bbr, 13797 bbr->r_ctl.rc_bbr_hptsi_gain, 13798 (tot_len + len), cts, 0); 13799 if (slot < bbr_error_base_paceout) 13800 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13801 } else 13802 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13803 bbr->rc_output_starts_timer = 1; 13804 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13805 tot_len); 13806 return (error); 13807 } 13808 case EPERM: 13809 tp->t_softerror = error; 13810 /* Fall through */ 13811 case EHOSTDOWN: 13812 case EHOSTUNREACH: 13813 case ENETDOWN: 13814 case ENETUNREACH: 13815 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13816 tp->t_softerror = error; 13817 } 13818 /* FALLTHROUGH */ 13819 default: 13820 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13821 bbr->rc_output_starts_timer = 1; 13822 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13823 return (error); 13824 } 13825 #ifdef STATS 13826 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13827 len && 13828 (rsm == NULL) && 13829 (bbr->rc_in_persist == 0)) { 13830 tp->gput_seq = bbr_seq; 13831 tp->gput_ack = bbr_seq + 13832 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13833 tp->gput_ts = cts; 13834 tp->t_flags |= TF_GPUTINPROG; 13835 #endif 13836 } 13837 KMOD_TCPSTAT_INC(tcps_sndtotal); 13838 if ((bbr->bbr_hdw_pace_ena) && 13839 (bbr->bbr_attempt_hdwr_pace == 0) && 13840 (bbr->rc_past_init_win) && 13841 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13842 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13843 (inp->inp_route.ro_nh && 13844 inp->inp_route.ro_nh->nh_ifp)) { 13845 /* 13846 * We are past the initial window and 13847 * have at least one measurement so we 13848 * could use hardware pacing if its available. 13849 * We have an interface and we have not attempted 13850 * to setup hardware pacing, lets try to now. 13851 */ 13852 uint64_t rate_wanted; 13853 int err = 0; 13854 13855 rate_wanted = bbr_get_hardware_rate(bbr); 13856 bbr->bbr_attempt_hdwr_pace = 1; 13857 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 13858 inp->inp_route.ro_nh->nh_ifp, 13859 rate_wanted, 13860 (RS_PACING_GEQ|RS_PACING_SUB_OK), 13861 &err, NULL); 13862 if (bbr->r_ctl.crte) { 13863 bbr_type_log_hdwr_pacing(bbr, 13864 bbr->r_ctl.crte->ptbl->rs_ifp, 13865 rate_wanted, 13866 bbr->r_ctl.crte->rate, 13867 __LINE__, cts, err); 13868 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 13869 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 13870 counter_u64_add(bbr_flows_whdwr_pacing, 1); 13871 bbr->bbr_hdrw_pacing = 1; 13872 /* Now what is our gain status? */ 13873 if (bbr->r_ctl.crte->rate < rate_wanted) { 13874 /* We have a problem */ 13875 bbr_setup_less_of_rate(bbr, cts, 13876 bbr->r_ctl.crte->rate, rate_wanted); 13877 } else { 13878 /* We are good */ 13879 bbr->gain_is_limited = 0; 13880 bbr->skip_gain = 0; 13881 } 13882 tcp_bbr_tso_size_check(bbr, cts); 13883 } else { 13884 bbr_type_log_hdwr_pacing(bbr, 13885 inp->inp_route.ro_nh->nh_ifp, 13886 rate_wanted, 13887 0, 13888 __LINE__, cts, err); 13889 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 13890 } 13891 } 13892 if (bbr->bbr_hdrw_pacing) { 13893 /* 13894 * Worry about cases where the route 13895 * changes or something happened that we 13896 * lost our hardware pacing possibly during 13897 * the last ip_output call. 13898 */ 13899 if (inp->inp_snd_tag == NULL) { 13900 /* A change during ip output disabled hw pacing? */ 13901 bbr->bbr_hdrw_pacing = 0; 13902 } else if ((inp->inp_route.ro_nh == NULL) || 13903 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 13904 /* 13905 * We had an interface or route change, 13906 * detach from the current hdwr pacing 13907 * and setup to re-attempt next go 13908 * round. 13909 */ 13910 bbr->bbr_hdrw_pacing = 0; 13911 bbr->bbr_attempt_hdwr_pace = 0; 13912 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 13913 tcp_bbr_tso_size_check(bbr, cts); 13914 } 13915 } 13916 /* 13917 * Data sent (as far as we can tell). If this advertises a larger 13918 * window than any other segment, then remember the size of the 13919 * advertised window. Any pending ACK has now been sent. 13920 */ 13921 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 13922 tp->rcv_adv = tp->rcv_nxt + recwin; 13923 13924 tp->last_ack_sent = tp->rcv_nxt; 13925 if ((error == 0) && 13926 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 13927 (doing_tlp == 0) && 13928 (tso == 0) && 13929 (len > 0) && 13930 ((flags & TH_RST) == 0) && 13931 ((flags & TH_SYN) == 0) && 13932 (IN_RECOVERY(tp->t_flags) == 0) && 13933 (bbr->rc_in_persist == 0) && 13934 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 13935 /* 13936 * For non-tso we need to goto again until we have sent out 13937 * enough data to match what we are hptsi out every hptsi 13938 * interval. 13939 */ 13940 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 13941 /* Make sure snd_nxt is drug up */ 13942 tp->snd_nxt = tp->snd_max; 13943 } 13944 if (rsm != NULL) { 13945 rsm = NULL; 13946 goto skip_again; 13947 } 13948 rsm = NULL; 13949 sack_rxmit = 0; 13950 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13951 goto again; 13952 } 13953 skip_again: 13954 if ((error == 0) && (flags & TH_FIN)) 13955 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 13956 if ((error == 0) && (flags & TH_RST)) 13957 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 13958 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 13959 /* 13960 * Calculate/Re-Calculate the hptsi slot in usecs based on 13961 * what we have sent so far 13962 */ 13963 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 13964 if (bbr->rc_no_pacing) 13965 slot = 0; 13966 } 13967 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13968 enobufs: 13969 if (bbr->rc_use_google == 0) 13970 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 13971 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13972 bbr->r_ctl.rc_lost_bytes))); 13973 bbr->rc_output_starts_timer = 1; 13974 if (bbr->bbr_use_rack_cheat && 13975 (more_to_rxt || 13976 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 13977 /* Rack cheats and shotguns out all rxt's 1ms apart */ 13978 if (slot > 1000) 13979 slot = 1000; 13980 } 13981 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 13982 /* 13983 * We don't change the tso size until some number of sends 13984 * to give the hardware commands time to get down 13985 * to the interface. 13986 */ 13987 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 13988 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 13989 bbr->hw_pacing_set = 1; 13990 tcp_bbr_tso_size_check(bbr, cts); 13991 } 13992 } 13993 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 13994 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 13995 /* Make sure snd_nxt is drug up */ 13996 tp->snd_nxt = tp->snd_max; 13997 } 13998 return (error); 13999 14000 } 14001 14002 /* 14003 * See bbr_output_wtime() for return values. 14004 */ 14005 static int 14006 bbr_output(struct tcpcb *tp) 14007 { 14008 int32_t ret; 14009 struct timeval tv; 14010 14011 NET_EPOCH_ASSERT(); 14012 14013 INP_WLOCK_ASSERT(tptoinpcb(tp)); 14014 (void)tcp_get_usecs(&tv); 14015 ret = bbr_output_wtime(tp, &tv); 14016 return (ret); 14017 } 14018 14019 static void 14020 bbr_mtu_chg(struct tcpcb *tp) 14021 { 14022 struct tcp_bbr *bbr; 14023 struct bbr_sendmap *rsm, *frsm = NULL; 14024 uint32_t maxseg; 14025 14026 /* 14027 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14028 * over the current size as SACK_PASS so a retransmit will occur. 14029 */ 14030 14031 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14032 maxseg = tp->t_maxseg - bbr->rc_last_options; 14033 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14034 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14035 /* Don't mess with ones acked (by sack?) */ 14036 if (rsm->r_flags & BBR_ACKED) 14037 continue; 14038 if ((rsm->r_end - rsm->r_start) > maxseg) { 14039 /* 14040 * We mark sack-passed on all the previous large 14041 * sends we did. This will force them to retransmit. 14042 */ 14043 rsm->r_flags |= BBR_SACK_PASSED; 14044 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14045 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14046 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14047 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14048 rsm->r_flags |= BBR_MARKED_LOST; 14049 } 14050 if (frsm == NULL) 14051 frsm = rsm; 14052 } 14053 } 14054 if (frsm) { 14055 bbr->r_ctl.rc_resend = frsm; 14056 } 14057 } 14058 14059 static int 14060 bbr_pru_options(struct tcpcb *tp, int flags) 14061 { 14062 if (flags & PRUS_OOB) 14063 return (EOPNOTSUPP); 14064 return (0); 14065 } 14066 14067 static void 14068 bbr_switch_failed(struct tcpcb *tp) 14069 { 14070 /* 14071 * If a switch fails we only need to 14072 * make sure mbuf_queuing is still in place. 14073 * We also need to make sure we are still in 14074 * ticks granularity (though we should probably 14075 * change bbr to go to USECs). 14076 * 14077 * For timers we need to see if we are still in the 14078 * pacer (if our flags are up) if so we are good, if 14079 * not we need to get back into the pacer. 14080 */ 14081 struct timeval tv; 14082 uint32_t cts; 14083 uint32_t toval; 14084 struct tcp_bbr *bbr; 14085 struct hpts_diag diag; 14086 14087 tp->t_flags2 |= TF2_CANNOT_DO_ECN; 14088 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ; 14089 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS); 14090 if (tp->t_in_hpts > IHPTS_NONE) { 14091 return; 14092 } 14093 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14094 cts = tcp_get_usecs(&tv); 14095 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 14096 if (TSTMP_GT(bbr->rc_pacer_started, cts)) { 14097 toval = bbr->rc_pacer_started - cts; 14098 } else { 14099 /* one slot please */ 14100 toval = HPTS_TICKS_PER_SLOT; 14101 } 14102 } else if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 14103 if (TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 14104 toval = bbr->r_ctl.rc_timer_exp - cts; 14105 } else { 14106 /* one slot please */ 14107 toval = HPTS_TICKS_PER_SLOT; 14108 } 14109 } else 14110 toval = HPTS_TICKS_PER_SLOT; 14111 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval), 14112 __LINE__, &diag); 14113 bbr_log_hpts_diag(bbr, cts, &diag); 14114 } 14115 14116 struct tcp_function_block __tcp_bbr = { 14117 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14118 .tfb_tcp_output = bbr_output, 14119 .tfb_do_queued_segments = ctf_do_queued_segments, 14120 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14121 .tfb_tcp_do_segment = bbr_do_segment, 14122 .tfb_tcp_ctloutput = bbr_ctloutput, 14123 .tfb_tcp_fb_init = bbr_init, 14124 .tfb_tcp_fb_fini = bbr_fini, 14125 .tfb_tcp_timer_stop_all = bbr_stopall, 14126 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14127 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14128 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14129 .tfb_pru_options = bbr_pru_options, 14130 .tfb_switch_failed = bbr_switch_failed, 14131 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14132 }; 14133 14134 /* 14135 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14136 * socket option arguments. When it re-acquires the lock after the copy, it 14137 * has to revalidate that the connection is still valid for the socket 14138 * option. 14139 */ 14140 static int 14141 bbr_set_sockopt(struct tcpcb *tp, struct sockopt *sopt) 14142 { 14143 struct epoch_tracker et; 14144 struct inpcb *inp = tptoinpcb(tp); 14145 struct tcp_bbr *bbr; 14146 int32_t error = 0, optval; 14147 14148 switch (sopt->sopt_level) { 14149 case IPPROTO_IPV6: 14150 case IPPROTO_IP: 14151 return (tcp_default_ctloutput(tp, sopt)); 14152 } 14153 14154 switch (sopt->sopt_name) { 14155 case TCP_RACK_PACE_MAX_SEG: 14156 case TCP_RACK_MIN_TO: 14157 case TCP_RACK_REORD_THRESH: 14158 case TCP_RACK_REORD_FADE: 14159 case TCP_RACK_TLP_THRESH: 14160 case TCP_RACK_PKT_DELAY: 14161 case TCP_BBR_ALGORITHM: 14162 case TCP_BBR_TSLIMITS: 14163 case TCP_BBR_IWINTSO: 14164 case TCP_BBR_RECFORCE: 14165 case TCP_BBR_STARTUP_PG: 14166 case TCP_BBR_DRAIN_PG: 14167 case TCP_BBR_RWND_IS_APP: 14168 case TCP_BBR_PROBE_RTT_INT: 14169 case TCP_BBR_PROBE_RTT_GAIN: 14170 case TCP_BBR_PROBE_RTT_LEN: 14171 case TCP_BBR_STARTUP_LOSS_EXIT: 14172 case TCP_BBR_USEDEL_RATE: 14173 case TCP_BBR_MIN_RTO: 14174 case TCP_BBR_MAX_RTO: 14175 case TCP_BBR_PACE_PER_SEC: 14176 case TCP_DELACK: 14177 case TCP_BBR_PACE_DEL_TAR: 14178 case TCP_BBR_SEND_IWND_IN_TSO: 14179 case TCP_BBR_EXTRA_STATE: 14180 case TCP_BBR_UTTER_MAX_TSO: 14181 case TCP_BBR_MIN_TOPACEOUT: 14182 case TCP_BBR_FLOOR_MIN_TSO: 14183 case TCP_BBR_TSTMP_RAISES: 14184 case TCP_BBR_POLICER_DETECT: 14185 case TCP_BBR_USE_RACK_CHEAT: 14186 case TCP_DATA_AFTER_CLOSE: 14187 case TCP_BBR_HDWR_PACE: 14188 case TCP_BBR_PACE_SEG_MAX: 14189 case TCP_BBR_PACE_SEG_MIN: 14190 case TCP_BBR_PACE_CROSS: 14191 case TCP_BBR_PACE_OH: 14192 case TCP_BBR_TMR_PACE_OH: 14193 case TCP_BBR_RACK_RTT_USE: 14194 case TCP_BBR_RETRAN_WTSO: 14195 break; 14196 default: 14197 return (tcp_default_ctloutput(tp, sopt)); 14198 break; 14199 } 14200 INP_WUNLOCK(inp); 14201 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14202 if (error) 14203 return (error); 14204 INP_WLOCK(inp); 14205 if (inp->inp_flags & INP_DROPPED) { 14206 INP_WUNLOCK(inp); 14207 return (ECONNRESET); 14208 } 14209 if (tp->t_fb != &__tcp_bbr) { 14210 INP_WUNLOCK(inp); 14211 return (ENOPROTOOPT); 14212 } 14213 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14214 switch (sopt->sopt_name) { 14215 case TCP_BBR_PACE_PER_SEC: 14216 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14217 bbr->r_ctl.bbr_hptsi_per_second = optval; 14218 break; 14219 case TCP_BBR_PACE_DEL_TAR: 14220 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14221 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14222 break; 14223 case TCP_BBR_PACE_SEG_MAX: 14224 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14225 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14226 break; 14227 case TCP_BBR_PACE_SEG_MIN: 14228 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14229 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14230 break; 14231 case TCP_BBR_PACE_CROSS: 14232 BBR_OPTS_INC(tcp_bbr_pace_cross); 14233 bbr->r_ctl.bbr_cross_over = optval; 14234 break; 14235 case TCP_BBR_ALGORITHM: 14236 BBR_OPTS_INC(tcp_bbr_algorithm); 14237 if (optval && (bbr->rc_use_google == 0)) { 14238 /* Turn on the google mode */ 14239 bbr_google_mode_on(bbr); 14240 if ((optval > 3) && (optval < 500)) { 14241 /* 14242 * Must be at least greater than .3% 14243 * and must be less than 50.0%. 14244 */ 14245 bbr->r_ctl.bbr_google_discount = optval; 14246 } 14247 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14248 /* Turn off the google mode */ 14249 bbr_google_mode_off(bbr); 14250 } 14251 break; 14252 case TCP_BBR_TSLIMITS: 14253 BBR_OPTS_INC(tcp_bbr_tslimits); 14254 if (optval == 1) 14255 bbr->rc_use_ts_limit = 1; 14256 else if (optval == 0) 14257 bbr->rc_use_ts_limit = 0; 14258 else 14259 error = EINVAL; 14260 break; 14261 14262 case TCP_BBR_IWINTSO: 14263 BBR_OPTS_INC(tcp_bbr_iwintso); 14264 if ((optval >= 0) && (optval < 128)) { 14265 uint32_t twin; 14266 14267 bbr->rc_init_win = optval; 14268 twin = bbr_initial_cwnd(bbr, tp); 14269 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14270 tp->snd_cwnd = twin; 14271 else 14272 error = EBUSY; 14273 } else 14274 error = EINVAL; 14275 break; 14276 case TCP_BBR_STARTUP_PG: 14277 BBR_OPTS_INC(tcp_bbr_startup_pg); 14278 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14279 bbr->r_ctl.rc_startup_pg = optval; 14280 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14281 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14282 } 14283 } else 14284 error = EINVAL; 14285 break; 14286 case TCP_BBR_DRAIN_PG: 14287 BBR_OPTS_INC(tcp_bbr_drain_pg); 14288 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14289 bbr->r_ctl.rc_drain_pg = optval; 14290 else 14291 error = EINVAL; 14292 break; 14293 case TCP_BBR_PROBE_RTT_LEN: 14294 BBR_OPTS_INC(tcp_bbr_probertt_len); 14295 if (optval <= 1) 14296 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14297 else 14298 error = EINVAL; 14299 break; 14300 case TCP_BBR_PROBE_RTT_GAIN: 14301 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14302 if (optval <= BBR_UNIT) 14303 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14304 else 14305 error = EINVAL; 14306 break; 14307 case TCP_BBR_PROBE_RTT_INT: 14308 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14309 if (optval > 1000) 14310 bbr->r_ctl.rc_probertt_int = optval; 14311 else 14312 error = EINVAL; 14313 break; 14314 case TCP_BBR_MIN_TOPACEOUT: 14315 BBR_OPTS_INC(tcp_bbr_topaceout); 14316 if (optval == 0) { 14317 bbr->no_pacing_until = 0; 14318 bbr->rc_no_pacing = 0; 14319 } else if (optval <= 0x00ff) { 14320 bbr->no_pacing_until = optval; 14321 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14322 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14323 /* Turn on no pacing */ 14324 bbr->rc_no_pacing = 1; 14325 } 14326 } else 14327 error = EINVAL; 14328 break; 14329 case TCP_BBR_STARTUP_LOSS_EXIT: 14330 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14331 bbr->rc_loss_exit = optval; 14332 break; 14333 case TCP_BBR_USEDEL_RATE: 14334 error = EINVAL; 14335 break; 14336 case TCP_BBR_MIN_RTO: 14337 BBR_OPTS_INC(tcp_bbr_min_rto); 14338 bbr->r_ctl.rc_min_rto_ms = optval; 14339 break; 14340 case TCP_BBR_MAX_RTO: 14341 BBR_OPTS_INC(tcp_bbr_max_rto); 14342 bbr->rc_max_rto_sec = optval; 14343 break; 14344 case TCP_RACK_MIN_TO: 14345 /* Minimum time between rack t-o's in ms */ 14346 BBR_OPTS_INC(tcp_rack_min_to); 14347 bbr->r_ctl.rc_min_to = optval; 14348 break; 14349 case TCP_RACK_REORD_THRESH: 14350 /* RACK reorder threshold (shift amount) */ 14351 BBR_OPTS_INC(tcp_rack_reord_thresh); 14352 if ((optval > 0) && (optval < 31)) 14353 bbr->r_ctl.rc_reorder_shift = optval; 14354 else 14355 error = EINVAL; 14356 break; 14357 case TCP_RACK_REORD_FADE: 14358 /* Does reordering fade after ms time */ 14359 BBR_OPTS_INC(tcp_rack_reord_fade); 14360 bbr->r_ctl.rc_reorder_fade = optval; 14361 break; 14362 case TCP_RACK_TLP_THRESH: 14363 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14364 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14365 if (optval) 14366 bbr->rc_tlp_threshold = optval; 14367 else 14368 error = EINVAL; 14369 break; 14370 case TCP_BBR_USE_RACK_CHEAT: 14371 BBR_OPTS_INC(tcp_use_rackcheat); 14372 if (bbr->rc_use_google) { 14373 error = EINVAL; 14374 break; 14375 } 14376 BBR_OPTS_INC(tcp_rack_cheat); 14377 if (optval) 14378 bbr->bbr_use_rack_cheat = 1; 14379 else 14380 bbr->bbr_use_rack_cheat = 0; 14381 break; 14382 case TCP_BBR_FLOOR_MIN_TSO: 14383 BBR_OPTS_INC(tcp_utter_max_tso); 14384 if ((optval >= 0) && (optval < 40)) 14385 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14386 else 14387 error = EINVAL; 14388 break; 14389 case TCP_BBR_UTTER_MAX_TSO: 14390 BBR_OPTS_INC(tcp_utter_max_tso); 14391 if ((optval >= 0) && (optval < 0xffff)) 14392 bbr->r_ctl.bbr_utter_max = optval; 14393 else 14394 error = EINVAL; 14395 break; 14396 14397 case TCP_BBR_EXTRA_STATE: 14398 BBR_OPTS_INC(tcp_extra_state); 14399 if (optval) 14400 bbr->rc_use_idle_restart = 1; 14401 else 14402 bbr->rc_use_idle_restart = 0; 14403 break; 14404 case TCP_BBR_SEND_IWND_IN_TSO: 14405 BBR_OPTS_INC(tcp_iwnd_tso); 14406 if (optval) { 14407 bbr->bbr_init_win_cheat = 1; 14408 if (bbr->rc_past_init_win == 0) { 14409 uint32_t cts; 14410 cts = tcp_get_usecs(&bbr->rc_tv); 14411 tcp_bbr_tso_size_check(bbr, cts); 14412 } 14413 } else 14414 bbr->bbr_init_win_cheat = 0; 14415 break; 14416 case TCP_BBR_HDWR_PACE: 14417 BBR_OPTS_INC(tcp_hdwr_pacing); 14418 if (optval){ 14419 bbr->bbr_hdw_pace_ena = 1; 14420 bbr->bbr_attempt_hdwr_pace = 0; 14421 } else { 14422 bbr->bbr_hdw_pace_ena = 0; 14423 #ifdef RATELIMIT 14424 if (bbr->r_ctl.crte != NULL) { 14425 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14426 bbr->r_ctl.crte = NULL; 14427 } 14428 #endif 14429 } 14430 break; 14431 14432 case TCP_DELACK: 14433 BBR_OPTS_INC(tcp_delack); 14434 if (optval < 100) { 14435 if (optval == 0) /* off */ 14436 tp->t_delayed_ack = 0; 14437 else if (optval == 1) /* on which is 2 */ 14438 tp->t_delayed_ack = 2; 14439 else /* higher than 2 and less than 100 */ 14440 tp->t_delayed_ack = optval; 14441 if (tp->t_flags & TF_DELACK) { 14442 tp->t_flags &= ~TF_DELACK; 14443 tp->t_flags |= TF_ACKNOW; 14444 NET_EPOCH_ENTER(et); 14445 bbr_output(tp); 14446 NET_EPOCH_EXIT(et); 14447 } 14448 } else 14449 error = EINVAL; 14450 break; 14451 case TCP_RACK_PKT_DELAY: 14452 /* RACK added ms i.e. rack-rtt + reord + N */ 14453 BBR_OPTS_INC(tcp_rack_pkt_delay); 14454 bbr->r_ctl.rc_pkt_delay = optval; 14455 break; 14456 14457 case TCP_BBR_RETRAN_WTSO: 14458 BBR_OPTS_INC(tcp_retran_wtso); 14459 if (optval) 14460 bbr->rc_resends_use_tso = 1; 14461 else 14462 bbr->rc_resends_use_tso = 0; 14463 break; 14464 case TCP_DATA_AFTER_CLOSE: 14465 BBR_OPTS_INC(tcp_data_ac); 14466 if (optval) 14467 bbr->rc_allow_data_af_clo = 1; 14468 else 14469 bbr->rc_allow_data_af_clo = 0; 14470 break; 14471 case TCP_BBR_POLICER_DETECT: 14472 BBR_OPTS_INC(tcp_policer_det); 14473 if (bbr->rc_use_google == 0) 14474 error = EINVAL; 14475 else if (optval) 14476 bbr->r_use_policer = 1; 14477 else 14478 bbr->r_use_policer = 0; 14479 break; 14480 14481 case TCP_BBR_TSTMP_RAISES: 14482 BBR_OPTS_INC(tcp_ts_raises); 14483 if (optval) 14484 bbr->ts_can_raise = 1; 14485 else 14486 bbr->ts_can_raise = 0; 14487 break; 14488 case TCP_BBR_TMR_PACE_OH: 14489 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14490 if (bbr->rc_use_google) { 14491 error = EINVAL; 14492 } else { 14493 if (optval) 14494 bbr->r_ctl.rc_incr_tmrs = 1; 14495 else 14496 bbr->r_ctl.rc_incr_tmrs = 0; 14497 } 14498 break; 14499 case TCP_BBR_PACE_OH: 14500 BBR_OPTS_INC(tcp_pacing_oh); 14501 if (bbr->rc_use_google) { 14502 error = EINVAL; 14503 } else { 14504 if (optval > (BBR_INCL_TCP_OH| 14505 BBR_INCL_IP_OH| 14506 BBR_INCL_ENET_OH)) { 14507 error = EINVAL; 14508 break; 14509 } 14510 if (optval & BBR_INCL_TCP_OH) 14511 bbr->r_ctl.rc_inc_tcp_oh = 1; 14512 else 14513 bbr->r_ctl.rc_inc_tcp_oh = 0; 14514 if (optval & BBR_INCL_IP_OH) 14515 bbr->r_ctl.rc_inc_ip_oh = 1; 14516 else 14517 bbr->r_ctl.rc_inc_ip_oh = 0; 14518 if (optval & BBR_INCL_ENET_OH) 14519 bbr->r_ctl.rc_inc_enet_oh = 1; 14520 else 14521 bbr->r_ctl.rc_inc_enet_oh = 0; 14522 } 14523 break; 14524 default: 14525 return (tcp_default_ctloutput(tp, sopt)); 14526 break; 14527 } 14528 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14529 INP_WUNLOCK(inp); 14530 return (error); 14531 } 14532 14533 /* 14534 * return 0 on success, error-num on failure 14535 */ 14536 static int 14537 bbr_get_sockopt(struct tcpcb *tp, struct sockopt *sopt) 14538 { 14539 struct inpcb *inp = tptoinpcb(tp); 14540 struct tcp_bbr *bbr; 14541 int32_t error, optval; 14542 14543 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14544 if (bbr == NULL) { 14545 INP_WUNLOCK(inp); 14546 return (EINVAL); 14547 } 14548 /* 14549 * Because all our options are either boolean or an int, we can just 14550 * pull everything into optval and then unlock and copy. If we ever 14551 * add a option that is not a int, then this will have quite an 14552 * impact to this routine. 14553 */ 14554 switch (sopt->sopt_name) { 14555 case TCP_BBR_PACE_PER_SEC: 14556 optval = bbr->r_ctl.bbr_hptsi_per_second; 14557 break; 14558 case TCP_BBR_PACE_DEL_TAR: 14559 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14560 break; 14561 case TCP_BBR_PACE_SEG_MAX: 14562 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14563 break; 14564 case TCP_BBR_MIN_TOPACEOUT: 14565 optval = bbr->no_pacing_until; 14566 break; 14567 case TCP_BBR_PACE_SEG_MIN: 14568 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14569 break; 14570 case TCP_BBR_PACE_CROSS: 14571 optval = bbr->r_ctl.bbr_cross_over; 14572 break; 14573 case TCP_BBR_ALGORITHM: 14574 optval = bbr->rc_use_google; 14575 break; 14576 case TCP_BBR_TSLIMITS: 14577 optval = bbr->rc_use_ts_limit; 14578 break; 14579 case TCP_BBR_IWINTSO: 14580 optval = bbr->rc_init_win; 14581 break; 14582 case TCP_BBR_STARTUP_PG: 14583 optval = bbr->r_ctl.rc_startup_pg; 14584 break; 14585 case TCP_BBR_DRAIN_PG: 14586 optval = bbr->r_ctl.rc_drain_pg; 14587 break; 14588 case TCP_BBR_PROBE_RTT_INT: 14589 optval = bbr->r_ctl.rc_probertt_int; 14590 break; 14591 case TCP_BBR_PROBE_RTT_LEN: 14592 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14593 break; 14594 case TCP_BBR_PROBE_RTT_GAIN: 14595 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14596 break; 14597 case TCP_BBR_STARTUP_LOSS_EXIT: 14598 optval = bbr->rc_loss_exit; 14599 break; 14600 case TCP_BBR_USEDEL_RATE: 14601 error = EINVAL; 14602 break; 14603 case TCP_BBR_MIN_RTO: 14604 optval = bbr->r_ctl.rc_min_rto_ms; 14605 break; 14606 case TCP_BBR_MAX_RTO: 14607 optval = bbr->rc_max_rto_sec; 14608 break; 14609 case TCP_RACK_PACE_MAX_SEG: 14610 /* Max segments in a pace */ 14611 optval = bbr->r_ctl.rc_pace_max_segs; 14612 break; 14613 case TCP_RACK_MIN_TO: 14614 /* Minimum time between rack t-o's in ms */ 14615 optval = bbr->r_ctl.rc_min_to; 14616 break; 14617 case TCP_RACK_REORD_THRESH: 14618 /* RACK reorder threshold (shift amount) */ 14619 optval = bbr->r_ctl.rc_reorder_shift; 14620 break; 14621 case TCP_RACK_REORD_FADE: 14622 /* Does reordering fade after ms time */ 14623 optval = bbr->r_ctl.rc_reorder_fade; 14624 break; 14625 case TCP_BBR_USE_RACK_CHEAT: 14626 /* Do we use the rack cheat for rxt */ 14627 optval = bbr->bbr_use_rack_cheat; 14628 break; 14629 case TCP_BBR_FLOOR_MIN_TSO: 14630 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14631 break; 14632 case TCP_BBR_UTTER_MAX_TSO: 14633 optval = bbr->r_ctl.bbr_utter_max; 14634 break; 14635 case TCP_BBR_SEND_IWND_IN_TSO: 14636 /* Do we send TSO size segments initially */ 14637 optval = bbr->bbr_init_win_cheat; 14638 break; 14639 case TCP_BBR_EXTRA_STATE: 14640 optval = bbr->rc_use_idle_restart; 14641 break; 14642 case TCP_RACK_TLP_THRESH: 14643 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14644 optval = bbr->rc_tlp_threshold; 14645 break; 14646 case TCP_RACK_PKT_DELAY: 14647 /* RACK added ms i.e. rack-rtt + reord + N */ 14648 optval = bbr->r_ctl.rc_pkt_delay; 14649 break; 14650 case TCP_BBR_RETRAN_WTSO: 14651 optval = bbr->rc_resends_use_tso; 14652 break; 14653 case TCP_DATA_AFTER_CLOSE: 14654 optval = bbr->rc_allow_data_af_clo; 14655 break; 14656 case TCP_DELACK: 14657 optval = tp->t_delayed_ack; 14658 break; 14659 case TCP_BBR_HDWR_PACE: 14660 optval = bbr->bbr_hdw_pace_ena; 14661 break; 14662 case TCP_BBR_POLICER_DETECT: 14663 optval = bbr->r_use_policer; 14664 break; 14665 case TCP_BBR_TSTMP_RAISES: 14666 optval = bbr->ts_can_raise; 14667 break; 14668 case TCP_BBR_TMR_PACE_OH: 14669 optval = bbr->r_ctl.rc_incr_tmrs; 14670 break; 14671 case TCP_BBR_PACE_OH: 14672 optval = 0; 14673 if (bbr->r_ctl.rc_inc_tcp_oh) 14674 optval |= BBR_INCL_TCP_OH; 14675 if (bbr->r_ctl.rc_inc_ip_oh) 14676 optval |= BBR_INCL_IP_OH; 14677 if (bbr->r_ctl.rc_inc_enet_oh) 14678 optval |= BBR_INCL_ENET_OH; 14679 break; 14680 default: 14681 return (tcp_default_ctloutput(tp, sopt)); 14682 break; 14683 } 14684 INP_WUNLOCK(inp); 14685 error = sooptcopyout(sopt, &optval, sizeof optval); 14686 return (error); 14687 } 14688 14689 /* 14690 * return 0 on success, error-num on failure 14691 */ 14692 static int 14693 bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt) 14694 { 14695 if (sopt->sopt_dir == SOPT_SET) { 14696 return (bbr_set_sockopt(tp, sopt)); 14697 } else if (sopt->sopt_dir == SOPT_GET) { 14698 return (bbr_get_sockopt(tp, sopt)); 14699 } else { 14700 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14701 } 14702 } 14703 14704 static const char *bbr_stack_names[] = { 14705 __XSTRING(STACKNAME), 14706 #ifdef STACKALIAS 14707 __XSTRING(STACKALIAS), 14708 #endif 14709 }; 14710 14711 static bool bbr_mod_inited = false; 14712 14713 static int 14714 tcp_addbbr(module_t mod, int32_t type, void *data) 14715 { 14716 int32_t err = 0; 14717 int num_stacks; 14718 14719 switch (type) { 14720 case MOD_LOAD: 14721 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14722 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14723 sizeof(struct bbr_sendmap), 14724 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14725 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14726 sizeof(struct tcp_bbr), 14727 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14728 sysctl_ctx_init(&bbr_sysctl_ctx); 14729 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14730 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14731 OID_AUTO, 14732 #ifdef STACKALIAS 14733 __XSTRING(STACKALIAS), 14734 #else 14735 __XSTRING(STACKNAME), 14736 #endif 14737 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14738 ""); 14739 if (bbr_sysctl_root == NULL) { 14740 printf("Failed to add sysctl node\n"); 14741 err = EFAULT; 14742 goto free_uma; 14743 } 14744 bbr_init_sysctls(); 14745 num_stacks = nitems(bbr_stack_names); 14746 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14747 bbr_stack_names, &num_stacks); 14748 if (err) { 14749 printf("Failed to register %s stack name for " 14750 "%s module\n", bbr_stack_names[num_stacks], 14751 __XSTRING(MODNAME)); 14752 sysctl_ctx_free(&bbr_sysctl_ctx); 14753 free_uma: 14754 uma_zdestroy(bbr_zone); 14755 uma_zdestroy(bbr_pcb_zone); 14756 bbr_counter_destroy(); 14757 printf("Failed to register " __XSTRING(MODNAME) 14758 " module err:%d\n", err); 14759 return (err); 14760 } 14761 tcp_lro_reg_mbufq(); 14762 bbr_mod_inited = true; 14763 printf(__XSTRING(MODNAME) " is now available\n"); 14764 break; 14765 case MOD_QUIESCE: 14766 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14767 break; 14768 case MOD_UNLOAD: 14769 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14770 if (err == EBUSY) 14771 break; 14772 if (bbr_mod_inited) { 14773 uma_zdestroy(bbr_zone); 14774 uma_zdestroy(bbr_pcb_zone); 14775 sysctl_ctx_free(&bbr_sysctl_ctx); 14776 bbr_counter_destroy(); 14777 printf(__XSTRING(MODNAME) 14778 " is now no longer available\n"); 14779 bbr_mod_inited = false; 14780 } 14781 tcp_lro_dereg_mbufq(); 14782 err = 0; 14783 break; 14784 default: 14785 return (EOPNOTSUPP); 14786 } 14787 return (err); 14788 } 14789 14790 static moduledata_t tcp_bbr = { 14791 .name = __XSTRING(MODNAME), 14792 .evhand = tcp_addbbr, 14793 .priv = 0 14794 }; 14795 14796 MODULE_VERSION(MODNAME, 1); 14797 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14798 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14799