1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /** 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include <sys/param.h> 42 #include <sys/arb.h> 43 #include <sys/module.h> 44 #include <sys/kernel.h> 45 #include <sys/libkern.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> 52 #include <sys/socket.h> 53 #include <sys/socketvar.h> 54 #include <sys/sysctl.h> 55 #include <sys/systm.h> 56 #ifdef STATS 57 #include <sys/qmath.h> 58 #include <sys/tree.h> 59 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 60 #endif 61 #include <sys/refcount.h> 62 #include <sys/queue.h> 63 #include <sys/eventhandler.h> 64 #include <sys/smp.h> 65 #include <sys/kthread.h> 66 #include <sys/lock.h> 67 #include <sys/mutex.h> 68 #include <sys/tim_filter.h> 69 #include <sys/time.h> 70 #include <sys/protosw.h> 71 #include <vm/uma.h> 72 #include <sys/kern_prefetch.h> 73 74 #include <net/route.h> 75 #include <net/route/nhop.h> 76 #include <net/vnet.h> 77 78 #define TCPSTATES /* for logging */ 79 80 #include <netinet/in.h> 81 #include <netinet/in_kdtrace.h> 82 #include <netinet/in_pcb.h> 83 #include <netinet/ip.h> 84 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 85 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 86 #include <netinet/ip_var.h> 87 #include <netinet/ip6.h> 88 #include <netinet6/in6_pcb.h> 89 #include <netinet6/ip6_var.h> 90 #define TCPOUTFLAGS 91 #include <netinet/tcp.h> 92 #include <netinet/tcp_fsm.h> 93 #include <netinet/tcp_seq.h> 94 #include <netinet/tcp_timer.h> 95 #include <netinet/tcp_var.h> 96 #include <netinet/tcpip.h> 97 #include <netinet/tcp_hpts.h> 98 #include <netinet/cc/cc.h> 99 #include <netinet/tcp_log_buf.h> 100 #include <netinet/tcp_ratelimit.h> 101 #include <netinet/tcp_lro.h> 102 #ifdef TCPDEBUG 103 #include <netinet/tcp_debug.h> 104 #endif /* TCPDEBUG */ 105 #ifdef TCP_OFFLOAD 106 #include <netinet/tcp_offload.h> 107 #endif 108 #ifdef INET6 109 #include <netinet6/tcp6_var.h> 110 #endif 111 #include <netinet/tcp_fastopen.h> 112 113 #include <netipsec/ipsec_support.h> 114 #include <net/if.h> 115 #include <net/if_var.h> 116 #include <net/ethernet.h> 117 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 119 #include <netipsec/ipsec.h> 120 #include <netipsec/ipsec6.h> 121 #endif /* IPSEC */ 122 123 #include <netinet/udp.h> 124 #include <netinet/udp_var.h> 125 #include <machine/in_cksum.h> 126 127 #ifdef MAC 128 #include <security/mac/mac_framework.h> 129 #endif 130 131 #include "sack_filter.h" 132 #include "tcp_bbr.h" 133 #include "rack_bbr_common.h" 134 uma_zone_t bbr_zone; 135 uma_zone_t bbr_pcb_zone; 136 137 struct sysctl_ctx_list bbr_sysctl_ctx; 138 struct sysctl_oid *bbr_sysctl_root; 139 140 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 141 (tv) = (value); \ 142 if ((u_long)(tv) < (u_long)(tvmin)) \ 143 (tv) = (tvmin); \ 144 if ((u_long)(tv) > (u_long)(tvmax)) \ 145 (tv) = (tvmax); \ 146 } while(0) 147 148 /*#define BBR_INVARIANT 1*/ 149 150 /* 151 * initial window 152 */ 153 static uint32_t bbr_def_init_win = 10; 154 static int32_t bbr_persist_min = 250000; /* 250ms */ 155 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 156 static int32_t bbr_cwnd_may_shrink = 0; 157 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 158 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 159 static int32_t bbr_hardware_pacing_limit = 8000; 160 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 161 static int32_t bbr_no_retran = 0; 162 163 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 164 static int32_t bbr_max_net_error_cnt = 10; 165 /* Should the following be dynamic too -- loss wise */ 166 static int32_t bbr_rtt_gain_thresh = 0; 167 /* Measurement controls */ 168 static int32_t bbr_use_google_algo = 1; 169 static int32_t bbr_ts_limiting = 1; 170 static int32_t bbr_ts_can_raise = 0; 171 static int32_t bbr_do_red = 600; 172 static int32_t bbr_red_scale = 20000; 173 static int32_t bbr_red_mul = 1; 174 static int32_t bbr_red_div = 2; 175 static int32_t bbr_red_growth_restrict = 1; 176 static int32_t bbr_target_is_bbunit = 0; 177 static int32_t bbr_drop_limit = 0; 178 /* 179 * How much gain do we need to see to 180 * stay in startup? 181 */ 182 static int32_t bbr_marks_rxt_sack_passed = 0; 183 static int32_t bbr_start_exit = 25; 184 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 185 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 186 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 187 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 188 * if we go back ever to where the pacer 189 * has priority over timers. 190 */ 191 static int32_t bbr_policer_call_from_rack_to = 0; 192 static int32_t bbr_policer_detection_enabled = 1; 193 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 194 * measurements before we are 195 * "good" note that 2 == 1. 196 * This is because we use a > 197 * comparison. This means if 198 * min_measure was 0, it takes 199 * num-measures > min(0) and 200 * you get 1 measurement and 201 * you are good. Set to 1, you 202 * have to have two 203 * measurements (this is done 204 * to prevent it from being ok 205 * to have no measurements). */ 206 static int32_t bbr_no_pacing_until = 4; 207 208 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 209 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 210 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 211 212 static int32_t bbr_target_cwnd_mult_limit = 8; 213 /* 214 * bbr_cwnd_min_val is the number of 215 * segments we hold to in the RTT probe 216 * state typically 4. 217 */ 218 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 219 220 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 221 222 static int32_t bbr_gain_to_target = 1; 223 static int32_t bbr_gain_gets_extra_too = 1; 224 /* 225 * bbr_high_gain is the 2/ln(2) value we need 226 * to double the sending rate in startup. This 227 * is used for both cwnd and hptsi gain's. 228 */ 229 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 230 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 231 static int32_t bbr_use_lower_gain_in_startup = 1; 232 233 /* thresholds for reduction on drain in sub-states/drain */ 234 static int32_t bbr_drain_rtt = BBR_SRTT; 235 static int32_t bbr_drain_floor = 88; 236 static int32_t google_allow_early_out = 1; 237 static int32_t google_consider_lost = 1; 238 static int32_t bbr_drain_drop_mul = 4; 239 static int32_t bbr_drain_drop_div = 5; 240 static int32_t bbr_rand_ot = 50; 241 static int32_t bbr_can_force_probertt = 0; 242 static int32_t bbr_can_adjust_probertt = 1; 243 static int32_t bbr_probertt_sets_rtt = 0; 244 static int32_t bbr_can_use_ts_for_rtt = 1; 245 static int32_t bbr_is_ratio = 0; 246 static int32_t bbr_sub_drain_app_limit = 1; 247 static int32_t bbr_prtt_slam_cwnd = 1; 248 static int32_t bbr_sub_drain_slam_cwnd = 1; 249 static int32_t bbr_slam_cwnd_in_main_drain = 1; 250 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 251 * hold */ 252 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 253 /* 254 * bbr_drain_gain is the reverse of the high_gain 255 * designed to drain back out the standing queue 256 * that is formed in startup by causing a larger 257 * hptsi gain and thus drainging the packets 258 * in flight. 259 */ 260 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 261 static int32_t bbr_rttprobe_gain = 192; 262 263 /* 264 * The cwnd_gain is the default cwnd gain applied when 265 * calculating a target cwnd. Note that the cwnd is 266 * a secondary factor in the way BBR works (see the 267 * paper and think about it, it will take some time). 268 * Basically the hptsi_gain spreads the packets out 269 * so you never get more than BDP to the peer even 270 * if the cwnd is high. In our implemenation that 271 * means in non-recovery/retransmission scenarios 272 * cwnd will never be reached by the flight-size. 273 */ 274 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 275 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 276 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 277 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 278 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 279 static int32_t bbr_ignore_data_after_close = 1; 280 static int16_t bbr_hptsi_gain[] = { 281 (BBR_UNIT *5 / 4), 282 (BBR_UNIT * 3 / 4), 283 BBR_UNIT, 284 BBR_UNIT, 285 BBR_UNIT, 286 BBR_UNIT, 287 BBR_UNIT, 288 BBR_UNIT 289 }; 290 int32_t bbr_use_rack_resend_cheat = 1; 291 int32_t bbr_sends_full_iwnd = 1; 292 293 #define BBR_HPTSI_GAIN_MAX 8 294 /* 295 * The BBR module incorporates a number of 296 * TCP ideas that have been put out into the IETF 297 * over the last few years: 298 * - Yuchung Cheng's RACK TCP (for which its named) that 299 * will stop us using the number of dup acks and instead 300 * use time as the gage of when we retransmit. 301 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 302 * of Dukkipati et.al. 303 * - Van Jacobson's et.al BBR. 304 * 305 * RACK depends on SACK, so if an endpoint arrives that 306 * cannot do SACK the state machine below will shuttle the 307 * connection back to using the "default" TCP stack that is 308 * in FreeBSD. 309 * 310 * To implement BBR and RACK the original TCP stack was first decomposed 311 * into a functional state machine with individual states 312 * for each of the possible TCP connection states. The do_segment 313 * functions role in life is to mandate the connection supports SACK 314 * initially and then assure that the RACK state matches the conenction 315 * state before calling the states do_segment function. Data processing 316 * of inbound segments also now happens in the hpts_do_segment in general 317 * with only one exception. This is so we can keep the connection on 318 * a single CPU. 319 * 320 * Each state is simplified due to the fact that the original do_segment 321 * has been decomposed and we *know* what state we are in (no 322 * switches on the state) and all tests for SACK are gone. This 323 * greatly simplifies what each state does. 324 * 325 * TCP output is also over-written with a new version since it 326 * must maintain the new rack scoreboard and has had hptsi 327 * integrated as a requirment. Still todo is to eliminate the 328 * use of the callout_() system and use the hpts for all 329 * timers as well. 330 */ 331 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 332 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 333 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 334 * free list */ 335 static int32_t bbr_tlp_thresh = 1; 336 static int32_t bbr_reorder_thresh = 2; 337 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 338 * 60,000,000 - 60 seconds */ 339 static int32_t bbr_pkt_delay = 1000; 340 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 341 static int32_t bbr_incr_timers = 1; 342 343 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 344 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 345 static int32_t bbr_exit_startup_at_loss = 1; 346 347 /* 348 * bbr_lt_bw_ratio is 1/8th 349 * bbr_lt_bw_diff is < 4 Kbit/sec 350 */ 351 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 352 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 353 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 354 * the lt_bw for */ 355 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 356 * lt_bw */ 357 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 358 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 359 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 360 361 static int32_t bbr_verbose_logging = 0; 362 /* 363 * Currently regular tcp has a rto_min of 30ms 364 * the backoff goes 12 times so that ends up 365 * being a total of 122.850 seconds before a 366 * connection is killed. 367 */ 368 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 369 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 370 371 /****************************************************/ 372 /* DEFAULT TSO SIZING (cpu performance impacting) */ 373 /****************************************************/ 374 /* What amount is our formula using to get TSO size */ 375 static int32_t bbr_hptsi_per_second = 1000; 376 377 /* 378 * For hptsi under bbr_cross_over connections what is delay 379 * target 7ms (in usec) combined with a seg_max of 2 380 * gets us close to identical google behavior in 381 * TSO size selection (possibly more 1MSS sends). 382 */ 383 static int32_t bbr_hptsi_segments_delay_tar = 7000; 384 385 /* Does pacing delay include overhead's in its time calculations? */ 386 static int32_t bbr_include_enet_oh = 0; 387 static int32_t bbr_include_ip_oh = 1; 388 static int32_t bbr_include_tcp_oh = 1; 389 static int32_t bbr_google_discount = 10; 390 391 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 392 static int32_t bbr_state_is_pkt_epoch = 0; 393 static int32_t bbr_state_drain_2_tar = 1; 394 /* What is the max the 0 - bbr_cross_over MBPS TSO target 395 * can reach using our delay target. Note that this 396 * value becomes the floor for the cross over 397 * algorithm. 398 */ 399 static int32_t bbr_hptsi_segments_max = 2; 400 static int32_t bbr_hptsi_segments_floor = 1; 401 static int32_t bbr_hptsi_utter_max = 0; 402 403 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 404 static int32_t bbr_hptsi_bytes_min = 1460; 405 static int32_t bbr_all_get_min = 0; 406 407 /* Cross over point from algo-a to algo-b */ 408 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 409 410 /* Do we deal with our restart state? */ 411 static int32_t bbr_uses_idle_restart = 0; 412 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 413 414 /* Do we allow hardware pacing? */ 415 static int32_t bbr_allow_hdwr_pacing = 0; 416 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 417 static int32_t bbr_hdwr_pace_floor = 1; 418 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 419 420 /****************************************************/ 421 static int32_t bbr_resends_use_tso = 0; 422 static int32_t bbr_tlp_max_resend = 2; 423 static int32_t bbr_sack_block_limit = 128; 424 425 #define BBR_MAX_STAT 19 426 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 427 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 428 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 429 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 430 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 431 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 432 counter_u64_t bbr_flows_whdwr_pacing; 433 counter_u64_t bbr_flows_nohdwr_pacing; 434 435 counter_u64_t bbr_nohdwr_pacing_enobuf; 436 counter_u64_t bbr_hdwr_pacing_enobuf; 437 438 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 439 440 /* 441 * Static defintions we need for forward declarations. 442 */ 443 static uint32_t 444 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 445 uint32_t useconds_time, uint64_t bw); 446 static uint32_t 447 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 448 static void 449 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 450 static void 451 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 452 static void 453 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 454 int dolog); 455 static uint32_t 456 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 457 static void 458 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 459 int32_t pkt_epoch, uint32_t losses); 460 static uint32_t 461 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, 462 struct bbr_sendmap *rsm); 463 static uint32_t 464 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 465 static uint32_t 466 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 467 struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); 468 static void 469 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 470 int32_t line); 471 static void 472 bbr_set_state_target(struct tcp_bbr *bbr, int line); 473 static void 474 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 475 static void 476 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, 477 int event, int line); 478 static void 479 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 480 static void 481 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 482 static void 483 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 484 uint32_t rtt, uint32_t line, uint8_t is_start, 485 uint16_t set); 486 static struct bbr_sendmap * 487 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 488 static __inline uint32_t 489 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 490 static void 491 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, 492 uint8_t which); 493 static void 494 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, 495 uint32_t time_since_sent, uint32_t srtt, 496 uint32_t thresh, uint32_t to); 497 static void 498 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 499 static void 500 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 501 uint32_t del_by, uint32_t cts, uint32_t sloton, 502 uint32_t prev_delay); 503 static void 504 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 505 int32_t line); 506 static void 507 bbr_stop_all_timers(struct tcpcb *tp); 508 static void 509 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 510 static void 511 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 512 static void 513 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 514 static void 515 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 516 uint32_t cts, uint32_t usecs, uint64_t bw, 517 uint32_t override, int mod); 518 static int 519 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt); 520 521 static inline uint8_t 522 bbr_state_val(struct tcp_bbr *bbr) 523 { 524 return(bbr->rc_bbr_substate); 525 } 526 527 static inline uint32_t 528 get_min_cwnd(struct tcp_bbr *bbr) 529 { 530 int mss; 531 532 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 533 bbr->r_ctl.rc_pace_max_segs); 534 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 535 return (bbr_cwnd_min_val_hs * mss); 536 else 537 return (bbr_cwnd_min_val * mss); 538 } 539 540 static uint32_t 541 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 542 { 543 uint64_t srtt, var; 544 uint64_t ret_val; 545 546 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 547 if (tp->t_srtt == 0) { 548 srtt = (uint64_t)BBR_INITIAL_RTO; 549 var = 0; 550 } else { 551 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 552 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 553 } 554 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 555 bbr_persist_min, bbr_persist_max); 556 return ((uint32_t)ret_val); 557 } 558 559 static uint32_t 560 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 561 { 562 /* 563 * Start the FR timer, we do this based on getting the first one in 564 * the rc_tmap. Note that if its NULL we must stop the timer. in all 565 * events we need to stop the running timer (if its running) before 566 * starting the new one. 567 */ 568 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 569 int32_t idx; 570 int32_t is_tlp_timer = 0; 571 struct bbr_sendmap *rsm; 572 573 if (bbr->rc_all_timers_stopped) { 574 /* All timers have been stopped none are to run */ 575 return (0); 576 } 577 if (bbr->rc_in_persist) { 578 /* We can't start any timer in persists */ 579 return (bbr_get_persists_timer_val(tp, bbr)); 580 } 581 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 582 if ((rsm == NULL) || 583 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 584 (tp->t_state < TCPS_ESTABLISHED)) { 585 /* Nothing on the send map */ 586 activate_rxt: 587 if (SEQ_LT(tp->snd_una, tp->snd_max) || 588 sbavail(&tptosocket(tp)->so_snd)) { 589 uint64_t tov; 590 591 time_since_sent = 0; 592 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 593 if (rsm) { 594 idx = rsm->r_rtr_cnt - 1; 595 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 596 tstmp_touse = rsm->r_tim_lastsent[idx]; 597 else 598 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 599 if (TSTMP_GT(tstmp_touse, cts)) 600 time_since_sent = cts - tstmp_touse; 601 } 602 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 603 if (tp->t_srtt == 0) 604 tov = BBR_INITIAL_RTO; 605 else 606 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 607 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 608 if (tp->t_rxtshift) 609 tov *= tcp_backoff[tp->t_rxtshift]; 610 if (tov > time_since_sent) 611 tov -= time_since_sent; 612 else 613 tov = bbr->r_ctl.rc_min_to; 614 TCPT_RANGESET_NOSLOP(to, tov, 615 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 616 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 617 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 618 return (to); 619 } 620 return (0); 621 } 622 if (rsm->r_flags & BBR_ACKED) { 623 rsm = bbr_find_lowest_rsm(bbr); 624 if (rsm == NULL) { 625 /* No lowest? */ 626 goto activate_rxt; 627 } 628 } 629 /* Convert from ms to usecs */ 630 if (rsm->r_flags & BBR_SACK_PASSED) { 631 if ((tp->t_flags & TF_SENTFIN) && 632 ((tp->snd_max - tp->snd_una) == 1) && 633 (rsm->r_flags & BBR_HAS_FIN)) { 634 /* 635 * We don't start a bbr rack timer if all we have is 636 * a FIN outstanding. 637 */ 638 goto activate_rxt; 639 } 640 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 641 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 642 idx = rsm->r_rtr_cnt - 1; 643 exp = rsm->r_tim_lastsent[idx] + thresh; 644 if (SEQ_GEQ(exp, cts)) { 645 to = exp - cts; 646 if (to < bbr->r_ctl.rc_min_to) { 647 to = bbr->r_ctl.rc_min_to; 648 } 649 } else { 650 to = bbr->r_ctl.rc_min_to; 651 } 652 } else { 653 /* Ok we need to do a TLP not RACK */ 654 if (bbr->rc_tlp_in_progress != 0) { 655 /* 656 * The previous send was a TLP. 657 */ 658 goto activate_rxt; 659 } 660 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 661 if (rsm == NULL) { 662 /* We found no rsm to TLP with. */ 663 goto activate_rxt; 664 } 665 if (rsm->r_flags & BBR_HAS_FIN) { 666 /* If its a FIN we don't do TLP */ 667 rsm = NULL; 668 goto activate_rxt; 669 } 670 time_since_sent = 0; 671 idx = rsm->r_rtr_cnt - 1; 672 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 673 tstmp_touse = rsm->r_tim_lastsent[idx]; 674 else 675 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 676 if (TSTMP_GT(tstmp_touse, cts)) 677 time_since_sent = cts - tstmp_touse; 678 is_tlp_timer = 1; 679 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 680 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 681 if (thresh > time_since_sent) 682 to = thresh - time_since_sent; 683 else 684 to = bbr->r_ctl.rc_min_to; 685 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 686 /* 687 * If the TLP time works out to larger than the max 688 * RTO lets not do TLP.. just RTO. 689 */ 690 goto activate_rxt; 691 } 692 if ((bbr->rc_tlp_rtx_out == 1) && 693 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 694 /* 695 * Second retransmit of the same TLP 696 * lets not. 697 */ 698 bbr->rc_tlp_rtx_out = 0; 699 goto activate_rxt; 700 } 701 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 702 /* 703 * The tail is no longer the last one I did a probe 704 * on 705 */ 706 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 707 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 708 } 709 } 710 if (is_tlp_timer == 0) { 711 BBR_STAT_INC(bbr_to_arm_rack); 712 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 713 } else { 714 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 715 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 716 /* 717 * We have exceeded how many times we can retran the 718 * current TLP timer, switch to the RTO timer. 719 */ 720 goto activate_rxt; 721 } else { 722 BBR_STAT_INC(bbr_to_arm_tlp); 723 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 724 } 725 } 726 return (to); 727 } 728 729 static inline int32_t 730 bbr_minseg(struct tcp_bbr *bbr) 731 { 732 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 733 } 734 735 static void 736 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 737 { 738 struct inpcb *inp = tptoinpcb(tp); 739 struct hpts_diag diag; 740 uint32_t delayed_ack = 0; 741 uint32_t left = 0; 742 uint32_t hpts_timeout; 743 uint8_t stopped; 744 int32_t delay_calc = 0; 745 uint32_t prev_delay = 0; 746 747 if (tcp_in_hpts(inp)) { 748 /* A previous call is already set up */ 749 return; 750 } 751 if ((tp->t_state == TCPS_CLOSED) || 752 (tp->t_state == TCPS_LISTEN)) { 753 return; 754 } 755 stopped = bbr->rc_tmr_stopped; 756 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 757 left = bbr->r_ctl.rc_timer_exp - cts; 758 } 759 bbr->r_ctl.rc_hpts_flags = 0; 760 bbr->r_ctl.rc_timer_exp = 0; 761 prev_delay = bbr->r_ctl.rc_last_delay_val; 762 if (bbr->r_ctl.rc_last_delay_val && 763 (slot == 0)) { 764 /* 765 * If a previous pacer delay was in place we 766 * are not coming from the output side (where 767 * we calculate a delay, more likely a timer). 768 */ 769 slot = bbr->r_ctl.rc_last_delay_val; 770 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 771 /* Compensate for time passed */ 772 delay_calc = cts - bbr->rc_pacer_started; 773 if (delay_calc <= slot) 774 slot -= delay_calc; 775 } 776 } 777 /* Do we have early to make up for by pushing out the pacing time? */ 778 if (bbr->r_agg_early_set) { 779 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 780 slot += bbr->r_ctl.rc_agg_early; 781 bbr->r_ctl.rc_agg_early = 0; 782 bbr->r_agg_early_set = 0; 783 } 784 /* Are we running a total debt that needs to be compensated for? */ 785 if (bbr->r_ctl.rc_hptsi_agg_delay) { 786 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 787 /* We nuke the delay */ 788 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 789 bbr->r_ctl.rc_hptsi_agg_delay = 0; 790 } else { 791 /* We nuke some of the delay, put in a minimal 100usecs */ 792 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 793 bbr->r_ctl.rc_last_delay_val = slot = 100; 794 } 795 } 796 bbr->r_ctl.rc_last_delay_val = slot; 797 hpts_timeout = bbr_timer_start(tp, bbr, cts); 798 if (tp->t_flags & TF_DELACK) { 799 if (bbr->rc_in_persist == 0) { 800 delayed_ack = bbr_delack_time; 801 } else { 802 /* 803 * We are in persists and have 804 * gotten a new data element. 805 */ 806 if (hpts_timeout > bbr_delack_time) { 807 /* 808 * Lets make the persists timer (which acks) 809 * be the smaller of hpts_timeout and bbr_delack_time. 810 */ 811 hpts_timeout = bbr_delack_time; 812 } 813 } 814 } 815 if (delayed_ack && 816 ((hpts_timeout == 0) || 817 (delayed_ack < hpts_timeout))) { 818 /* We need a Delayed ack timer */ 819 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 820 hpts_timeout = delayed_ack; 821 } 822 if (slot) { 823 /* Mark that we have a pacing timer up */ 824 BBR_STAT_INC(bbr_paced_segments); 825 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 826 } 827 /* 828 * If no timers are going to run and we will fall off thfe hptsi 829 * wheel, we resort to a keep-alive timer if its configured. 830 */ 831 if ((hpts_timeout == 0) && 832 (slot == 0)) { 833 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 834 (tp->t_state <= TCPS_CLOSING)) { 835 /* 836 * Ok we have no timer (persists, rack, tlp, rxt or 837 * del-ack), we don't have segments being paced. So 838 * all that is left is the keepalive timer. 839 */ 840 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 841 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 842 } else { 843 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 844 } 845 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 846 } 847 } 848 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 849 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 850 /* 851 * RACK, TLP, persists and RXT timers all are restartable 852 * based on actions input .. i.e we received a packet (ack 853 * or sack) and that changes things (rw, or snd_una etc). 854 * Thus we can restart them with a new value. For 855 * keep-alive, delayed_ack we keep track of what was left 856 * and restart the timer with a smaller value. 857 */ 858 if (left < hpts_timeout) 859 hpts_timeout = left; 860 } 861 if (bbr->r_ctl.rc_incr_tmrs && slot && 862 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 863 /* 864 * If configured to do so, and the timer is either 865 * the TLP or RXT timer, we need to increase the timeout 866 * by the pacing time. Consider the bottleneck at my 867 * machine as an example, we are sending something 868 * to start a TLP on. The last packet won't be emitted 869 * fully until the pacing time (the bottleneck will hold 870 * the data in place). Once the packet is emitted that 871 * is when we want to start waiting for the TLP. This 872 * is most evident with hardware pacing (where the nic 873 * is holding the packet(s) before emitting). But it 874 * can also show up in the network so we do it for all 875 * cases. Technically we would take off one packet from 876 * this extra delay but this is easier and being more 877 * conservative is probably better. 878 */ 879 hpts_timeout += slot; 880 } 881 if (hpts_timeout) { 882 /* 883 * Hack alert for now we can't time-out over 2147 seconds (a 884 * bit more than 35min) 885 */ 886 if (hpts_timeout > 0x7ffffffe) 887 hpts_timeout = 0x7ffffffe; 888 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 889 } else 890 bbr->r_ctl.rc_timer_exp = 0; 891 if ((slot) && 892 (bbr->rc_use_google || 893 bbr->output_error_seen || 894 (slot <= hpts_timeout)) ) { 895 /* 896 * Tell LRO that it can queue packets while 897 * we pace. 898 */ 899 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 900 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 901 (bbr->rc_cwnd_limited == 0)) { 902 /* 903 * If we are not cwnd limited and we 904 * are running a rack timer we put on 905 * the do not disturbe even for sack. 906 */ 907 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 908 } else 909 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 910 bbr->rc_pacer_started = cts; 911 912 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot), 913 __LINE__, &diag); 914 bbr->rc_timer_first = 0; 915 bbr->bbr_timer_src = frm; 916 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 917 bbr_log_hpts_diag(bbr, cts, &diag); 918 } else if (hpts_timeout) { 919 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout), 920 __LINE__, &diag); 921 /* 922 * We add the flag here as well if the slot is set, 923 * since hpts will call in to clear the queue first before 924 * calling the output routine (which does our timers). 925 * We don't want to set the flag if its just a timer 926 * else the arrival of data might (that causes us 927 * to send more) might get delayed. Imagine being 928 * on a keep-alive timer and a request comes in for 929 * more data. 930 */ 931 if (slot) 932 bbr->rc_pacer_started = cts; 933 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 934 (bbr->rc_cwnd_limited == 0)) { 935 /* 936 * For a rack timer, don't wake us even 937 * if a sack arrives as long as we are 938 * not cwnd limited. 939 */ 940 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 941 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 942 } else { 943 /* All other timers wake us up */ 944 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 945 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 946 } 947 bbr->bbr_timer_src = frm; 948 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 949 bbr_log_hpts_diag(bbr, cts, &diag); 950 bbr->rc_timer_first = 1; 951 } 952 bbr->rc_tmr_stopped = 0; 953 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 954 } 955 956 static void 957 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 958 { 959 /* 960 * We received an ack, and then did not call send or were bounced 961 * out due to the hpts was running. Now a timer is up as well, is it 962 * the right timer? 963 */ 964 struct inpcb *inp; 965 struct bbr_sendmap *rsm; 966 uint32_t hpts_timeout; 967 int tmr_up; 968 969 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 970 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 971 return; 972 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 973 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 974 (tmr_up == PACE_TMR_RXT)) { 975 /* Should be an RXT */ 976 return; 977 } 978 inp = bbr->rc_inp; 979 if (rsm == NULL) { 980 /* Nothing outstanding? */ 981 if (tp->t_flags & TF_DELACK) { 982 if (tmr_up == PACE_TMR_DELACK) 983 /* 984 * We are supposed to have delayed ack up 985 * and we do 986 */ 987 return; 988 } else if (sbavail(&inp->inp_socket->so_snd) && 989 (tmr_up == PACE_TMR_RXT)) { 990 /* 991 * if we hit enobufs then we would expect the 992 * possibility of nothing outstanding and the RXT up 993 * (and the hptsi timer). 994 */ 995 return; 996 } else if (((V_tcp_always_keepalive || 997 inp->inp_socket->so_options & SO_KEEPALIVE) && 998 (tp->t_state <= TCPS_CLOSING)) && 999 (tmr_up == PACE_TMR_KEEP) && 1000 (tp->snd_max == tp->snd_una)) { 1001 /* We should have keep alive up and we do */ 1002 return; 1003 } 1004 } 1005 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 1006 if ((tp->t_flags & TF_SENTFIN) && 1007 ((tp->snd_max - tp->snd_una) == 1) && 1008 (rsm->r_flags & BBR_HAS_FIN)) { 1009 /* needs to be a RXT */ 1010 if (tmr_up == PACE_TMR_RXT) 1011 return; 1012 else 1013 goto wrong_timer; 1014 } else if (tmr_up == PACE_TMR_RACK) 1015 return; 1016 else 1017 goto wrong_timer; 1018 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1019 /* Rack timer has priority if we have data out */ 1020 return; 1021 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1022 ((tmr_up == PACE_TMR_TLP) || 1023 (tmr_up == PACE_TMR_RXT))) { 1024 /* 1025 * Either a TLP or RXT is fine if no sack-passed is in place 1026 * and data is outstanding. 1027 */ 1028 return; 1029 } else if (tmr_up == PACE_TMR_DELACK) { 1030 /* 1031 * If the delayed ack was going to go off before the 1032 * rtx/tlp/rack timer were going to expire, then that would 1033 * be the timer in control. Note we don't check the time 1034 * here trusting the code is correct. 1035 */ 1036 return; 1037 } 1038 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1039 ((tmr_up == PACE_TMR_RXT) || 1040 (tmr_up == PACE_TMR_TLP) || 1041 (tmr_up == PACE_TMR_RACK))) { 1042 /* 1043 * We have outstanding data and 1044 * we *do* have a RACK, TLP or RXT 1045 * timer running. We won't restart 1046 * anything here since thats probably ok we 1047 * will get called with some timer here shortly. 1048 */ 1049 return; 1050 } 1051 /* 1052 * Ok the timer originally started is not what we want now. We will 1053 * force the hpts to be stopped if any, and restart with the slot 1054 * set to what was in the saved slot. 1055 */ 1056 wrong_timer: 1057 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1058 if (tcp_in_hpts(inp)) 1059 tcp_hpts_remove(inp); 1060 bbr_timer_cancel(bbr, __LINE__, cts); 1061 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1062 0); 1063 } else { 1064 /* 1065 * Output is hptsi so we just need to switch the type of 1066 * timer. We don't bother with keep-alive, since when we 1067 * jump through the output, it will start the keep-alive if 1068 * nothing is sent. 1069 * 1070 * We only need a delayed-ack added and or the hpts_timeout. 1071 */ 1072 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1073 if (tp->t_flags & TF_DELACK) { 1074 if (hpts_timeout == 0) { 1075 hpts_timeout = bbr_delack_time; 1076 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1077 } 1078 else if (hpts_timeout > bbr_delack_time) { 1079 hpts_timeout = bbr_delack_time; 1080 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1081 } 1082 } 1083 if (hpts_timeout) { 1084 if (hpts_timeout > 0x7ffffffe) 1085 hpts_timeout = 0x7ffffffe; 1086 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1087 } 1088 } 1089 } 1090 1091 int32_t bbr_clear_lost = 0; 1092 1093 /* 1094 * Considers the two time values now (cts) and earlier. 1095 * If cts is smaller than earlier, we could have 1096 * had a sequence wrap (our counter wraps every 1097 * 70 min or so) or it could be just clock skew 1098 * getting us two different time values. Clock skew 1099 * will show up within 10ms or so. So in such 1100 * a case (where cts is behind earlier time by 1101 * less than 10ms) we return 0. Otherwise we 1102 * return the true difference between them. 1103 */ 1104 static inline uint32_t 1105 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1106 /* 1107 * Given two timestamps, the current time stamp cts, and some other 1108 * time-stamp taken in theory earlier return the difference. The 1109 * trick is here sometimes locking will get the other timestamp 1110 * after the cts. If this occurs we need to return 0. 1111 */ 1112 if (TSTMP_GEQ(cts, earlier_time)) 1113 return (cts - earlier_time); 1114 /* 1115 * cts is behind earlier_time if its less than 10ms consider it 0. 1116 * If its more than 10ms difference then we had a time wrap. Else 1117 * its just the normal locking foo. I wonder if we should not go to 1118 * 64bit TS and get rid of this issue. 1119 */ 1120 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1121 return (0); 1122 /* 1123 * Ok the time must have wrapped. So we need to answer a large 1124 * amount of time, which the normal subtraction should do. 1125 */ 1126 return (cts - earlier_time); 1127 } 1128 1129 static int 1130 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1131 { 1132 uint32_t stat; 1133 int32_t error; 1134 1135 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1136 if (error || req->newptr == NULL) 1137 return error; 1138 1139 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1140 if (error) 1141 return (error); 1142 if (stat == 1) { 1143 #ifdef BBR_INVARIANTS 1144 printf("Clearing BBR lost counters\n"); 1145 #endif 1146 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1147 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1148 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1149 } else if (stat == 2) { 1150 #ifdef BBR_INVARIANTS 1151 printf("Clearing BBR option counters\n"); 1152 #endif 1153 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1154 } else if (stat == 3) { 1155 #ifdef BBR_INVARIANTS 1156 printf("Clearing BBR stats counters\n"); 1157 #endif 1158 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1159 } else if (stat == 4) { 1160 #ifdef BBR_INVARIANTS 1161 printf("Clearing BBR out-size counters\n"); 1162 #endif 1163 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1164 } 1165 bbr_clear_lost = 0; 1166 return (0); 1167 } 1168 1169 static void 1170 bbr_init_sysctls(void) 1171 { 1172 struct sysctl_oid *bbr_probertt; 1173 struct sysctl_oid *bbr_hptsi; 1174 struct sysctl_oid *bbr_measure; 1175 struct sysctl_oid *bbr_cwnd; 1176 struct sysctl_oid *bbr_timeout; 1177 struct sysctl_oid *bbr_states; 1178 struct sysctl_oid *bbr_startup; 1179 struct sysctl_oid *bbr_policer; 1180 1181 /* Probe rtt controls */ 1182 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1183 SYSCTL_CHILDREN(bbr_sysctl_root), 1184 OID_AUTO, 1185 "probertt", 1186 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1187 ""); 1188 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1189 SYSCTL_CHILDREN(bbr_probertt), 1190 OID_AUTO, "gain", CTLFLAG_RW, 1191 &bbr_rttprobe_gain, 192, 1192 "What is the filter gain drop in probe_rtt (0=disable)?"); 1193 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1194 SYSCTL_CHILDREN(bbr_probertt), 1195 OID_AUTO, "cwnd", CTLFLAG_RW, 1196 &bbr_rtt_probe_cwndtarg, 4, 1197 "How many mss's are outstanding during probe-rtt"); 1198 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1199 SYSCTL_CHILDREN(bbr_probertt), 1200 OID_AUTO, "int", CTLFLAG_RW, 1201 &bbr_rtt_probe_limit, 4000000, 1202 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1203 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1204 SYSCTL_CHILDREN(bbr_probertt), 1205 OID_AUTO, "mintime", CTLFLAG_RW, 1206 &bbr_rtt_probe_time, 200000, 1207 "How many microseconds in probe-rtt"); 1208 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1209 SYSCTL_CHILDREN(bbr_probertt), 1210 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1211 &bbr_filter_len_sec, 6, 1212 "How long in seconds does the rttProp filter run?"); 1213 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1214 SYSCTL_CHILDREN(bbr_probertt), 1215 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1216 &bbr_drain_rtt, BBR_SRTT, 1217 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1218 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1219 SYSCTL_CHILDREN(bbr_probertt), 1220 OID_AUTO, "can_force", CTLFLAG_RW, 1221 &bbr_can_force_probertt, 0, 1222 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1223 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1224 SYSCTL_CHILDREN(bbr_probertt), 1225 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1226 &bbr_probertt_sets_rtt, 0, 1227 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1228 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1229 SYSCTL_CHILDREN(bbr_probertt), 1230 OID_AUTO, "can_adjust", CTLFLAG_RW, 1231 &bbr_can_adjust_probertt, 1, 1232 "Can we dynamically adjust the probe-rtt limits and times?"); 1233 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1234 SYSCTL_CHILDREN(bbr_probertt), 1235 OID_AUTO, "is_ratio", CTLFLAG_RW, 1236 &bbr_is_ratio, 0, 1237 "is the limit to filter a ratio?"); 1238 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1239 SYSCTL_CHILDREN(bbr_probertt), 1240 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1241 &bbr_prtt_slam_cwnd, 0, 1242 "Should we set/recover cwnd?"); 1243 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1244 SYSCTL_CHILDREN(bbr_probertt), 1245 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1246 &bbr_can_use_ts_for_rtt, 1, 1247 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1248 1249 /* Pacing controls */ 1250 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1251 SYSCTL_CHILDREN(bbr_sysctl_root), 1252 OID_AUTO, 1253 "pacing", 1254 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1255 ""); 1256 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1257 SYSCTL_CHILDREN(bbr_hptsi), 1258 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1259 &bbr_allow_hdwr_pacing, 1, 1260 "Do we allow hardware pacing?"); 1261 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1262 SYSCTL_CHILDREN(bbr_hptsi), 1263 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1264 &bbr_hardware_pacing_limit, 4000, 1265 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1266 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1267 SYSCTL_CHILDREN(bbr_hptsi), 1268 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1269 &bbr_hdwr_pace_adjust, 2, 1270 "Multiplier to calculated tso size?"); 1271 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1272 SYSCTL_CHILDREN(bbr_hptsi), 1273 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1274 &bbr_hdwr_pace_floor, 1, 1275 "Do we invoke the hardware pacing floor?"); 1276 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1277 SYSCTL_CHILDREN(bbr_hptsi), 1278 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1279 &bbr_hdwr_pacing_delay_cnt, 10, 1280 "How many packets must be sent after hdwr pacing is enabled"); 1281 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1282 SYSCTL_CHILDREN(bbr_hptsi), 1283 OID_AUTO, "bw_cross", CTLFLAG_RW, 1284 &bbr_cross_over, 3000000, 1285 "What is the point where we cross over to linux like TSO size set"); 1286 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1287 SYSCTL_CHILDREN(bbr_hptsi), 1288 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1289 &bbr_hptsi_segments_delay_tar, 7000, 1290 "What is the worse case delay target for hptsi < 48Mbp connections"); 1291 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1292 SYSCTL_CHILDREN(bbr_hptsi), 1293 OID_AUTO, "enet_oh", CTLFLAG_RW, 1294 &bbr_include_enet_oh, 0, 1295 "Do we include the ethernet overhead in calculating pacing delay?"); 1296 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1297 SYSCTL_CHILDREN(bbr_hptsi), 1298 OID_AUTO, "ip_oh", CTLFLAG_RW, 1299 &bbr_include_ip_oh, 1, 1300 "Do we include the IP overhead in calculating pacing delay?"); 1301 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1302 SYSCTL_CHILDREN(bbr_hptsi), 1303 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1304 &bbr_include_tcp_oh, 0, 1305 "Do we include the TCP overhead in calculating pacing delay?"); 1306 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1307 SYSCTL_CHILDREN(bbr_hptsi), 1308 OID_AUTO, "google_discount", CTLFLAG_RW, 1309 &bbr_google_discount, 10, 1310 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1311 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1312 SYSCTL_CHILDREN(bbr_hptsi), 1313 OID_AUTO, "all_get_min", CTLFLAG_RW, 1314 &bbr_all_get_min, 0, 1315 "If you are less than a MSS do you just get the min?"); 1316 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1317 SYSCTL_CHILDREN(bbr_hptsi), 1318 OID_AUTO, "tso_min", CTLFLAG_RW, 1319 &bbr_hptsi_bytes_min, 1460, 1320 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1321 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1322 SYSCTL_CHILDREN(bbr_hptsi), 1323 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1324 &bbr_hptsi_segments_max, 6, 1325 "For 0 -> 24Mbps what is top number of segments for TSO"); 1326 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1327 SYSCTL_CHILDREN(bbr_hptsi), 1328 OID_AUTO, "seg_floor", CTLFLAG_RW, 1329 &bbr_hptsi_segments_floor, 1, 1330 "Minimum TSO size we will fall too in segments"); 1331 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1332 SYSCTL_CHILDREN(bbr_hptsi), 1333 OID_AUTO, "utter_max", CTLFLAG_RW, 1334 &bbr_hptsi_utter_max, 0, 1335 "The absolute maximum that any pacing (outside of hardware) can be"); 1336 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1337 SYSCTL_CHILDREN(bbr_hptsi), 1338 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1339 &bbr_hptsi_per_second, 100, 1340 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1341 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1342 SYSCTL_CHILDREN(bbr_hptsi), 1343 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1344 &bbr_hptsi_max_mul, 1, 1345 "The multiplier for pace len max"); 1346 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1347 SYSCTL_CHILDREN(bbr_hptsi), 1348 OID_AUTO, "srtt_div", CTLFLAG_RW, 1349 &bbr_hptsi_max_div, 2, 1350 "The divisor for pace len max"); 1351 /* Measurement controls */ 1352 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1353 SYSCTL_CHILDREN(bbr_sysctl_root), 1354 OID_AUTO, 1355 "measure", 1356 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1357 "Measurement controls"); 1358 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1359 SYSCTL_CHILDREN(bbr_measure), 1360 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1361 &bbr_initial_bw_bps, 62500, 1362 "Minimum initial b/w in bytes per second"); 1363 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1364 SYSCTL_CHILDREN(bbr_measure), 1365 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1366 &bbr_sack_not_required, 0, 1367 "Do we allow bbr to run on connections not supporting SACK?"); 1368 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1369 SYSCTL_CHILDREN(bbr_measure), 1370 OID_AUTO, "use_google", CTLFLAG_RW, 1371 &bbr_use_google_algo, 0, 1372 "Use has close to google V1.0 has possible?"); 1373 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1374 SYSCTL_CHILDREN(bbr_measure), 1375 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1376 &bbr_ts_limiting, 1, 1377 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1378 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1379 SYSCTL_CHILDREN(bbr_measure), 1380 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1381 &bbr_ts_can_raise, 0, 1382 "Can we raise the b/w via timestamp b/w calculation?"); 1383 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1384 SYSCTL_CHILDREN(bbr_measure), 1385 OID_AUTO, "ts_delta", CTLFLAG_RW, 1386 &bbr_min_usec_delta, 20000, 1387 "How long in usec between ts of our sends in ts validation code?"); 1388 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1389 SYSCTL_CHILDREN(bbr_measure), 1390 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1391 &bbr_min_peer_delta, 20, 1392 "What min numerical value should be between the peer deltas?"); 1393 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1394 SYSCTL_CHILDREN(bbr_measure), 1395 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1396 &bbr_delta_percent, 150, 1397 "What percentage (150 = 15.0) do we allow variance for?"); 1398 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1399 SYSCTL_CHILDREN(bbr_measure), 1400 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1401 &bbr_min_measurements_req, 1, 1402 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1403 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1404 SYSCTL_CHILDREN(bbr_measure), 1405 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1406 &bbr_no_pacing_until, 4, 1407 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1408 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1409 SYSCTL_CHILDREN(bbr_measure), 1410 OID_AUTO, "quanta", CTLFLAG_RW, 1411 &bbr_quanta, 2, 1412 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1413 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1414 SYSCTL_CHILDREN(bbr_measure), 1415 OID_AUTO, "noretran", CTLFLAG_RW, 1416 &bbr_no_retran, 0, 1417 "Should google mode not use retransmission measurements for the b/w estimation?"); 1418 /* State controls */ 1419 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1420 SYSCTL_CHILDREN(bbr_sysctl_root), 1421 OID_AUTO, 1422 "states", 1423 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1424 "State controls"); 1425 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1426 SYSCTL_CHILDREN(bbr_states), 1427 OID_AUTO, "idle_restart", CTLFLAG_RW, 1428 &bbr_uses_idle_restart, 0, 1429 "Do we use a new special idle_restart state to ramp back up quickly?"); 1430 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1431 SYSCTL_CHILDREN(bbr_states), 1432 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1433 &bbr_idle_restart_threshold, 100000, 1434 "How long must we be idle before we restart??"); 1435 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1436 SYSCTL_CHILDREN(bbr_states), 1437 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1438 &bbr_state_is_pkt_epoch, 0, 1439 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1440 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1441 SYSCTL_CHILDREN(bbr_states), 1442 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1443 &bbr_rtt_gain_thresh, 0, 1444 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1445 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1446 SYSCTL_CHILDREN(bbr_states), 1447 OID_AUTO, "drain_floor", CTLFLAG_RW, 1448 &bbr_drain_floor, 88, 1449 "What is the lowest we can drain (pg) too?"); 1450 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1451 SYSCTL_CHILDREN(bbr_states), 1452 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1453 &bbr_state_drain_2_tar, 1, 1454 "Do we drain to target in drain substate?"); 1455 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1456 SYSCTL_CHILDREN(bbr_states), 1457 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1458 &bbr_gain_to_target, 1, 1459 "Does probe bw gain to target??"); 1460 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1461 SYSCTL_CHILDREN(bbr_states), 1462 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1463 &bbr_gain_gets_extra_too, 1, 1464 "Does probe bw gain get the extra time too?"); 1465 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1466 SYSCTL_CHILDREN(bbr_states), 1467 OID_AUTO, "ld_div", CTLFLAG_RW, 1468 &bbr_drain_drop_div, 5, 1469 "Long drain drop divider?"); 1470 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1471 SYSCTL_CHILDREN(bbr_states), 1472 OID_AUTO, "ld_mul", CTLFLAG_RW, 1473 &bbr_drain_drop_mul, 4, 1474 "Long drain drop multiplier?"); 1475 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1476 SYSCTL_CHILDREN(bbr_states), 1477 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1478 &bbr_rand_ot, 50, 1479 "Random discount of the ot?"); 1480 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1481 SYSCTL_CHILDREN(bbr_states), 1482 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1483 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1484 "How many packet-epochs does the b/w delivery rate last?"); 1485 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1486 SYSCTL_CHILDREN(bbr_states), 1487 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1488 &bbr_sub_drain_app_limit, 0, 1489 "Does our sub-state drain invoke app limited if its long?"); 1490 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1491 SYSCTL_CHILDREN(bbr_states), 1492 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1493 &bbr_sub_drain_slam_cwnd, 0, 1494 "Should we set/recover cwnd for sub-state drain?"); 1495 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1496 SYSCTL_CHILDREN(bbr_states), 1497 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1498 &bbr_slam_cwnd_in_main_drain, 0, 1499 "Should we set/recover cwnd for main-state drain?"); 1500 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1501 SYSCTL_CHILDREN(bbr_states), 1502 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1503 &google_allow_early_out, 1, 1504 "Should we allow google probe-bw/drain to exit early at flight target?"); 1505 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1506 SYSCTL_CHILDREN(bbr_states), 1507 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1508 &google_consider_lost, 1, 1509 "Should we have losses exit gain of probebw in google mode??"); 1510 /* Startup controls */ 1511 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1512 SYSCTL_CHILDREN(bbr_sysctl_root), 1513 OID_AUTO, 1514 "startup", 1515 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1516 "Startup controls"); 1517 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1518 SYSCTL_CHILDREN(bbr_startup), 1519 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1520 &bbr_sends_full_iwnd, 1, 1521 "Do we not pace but burst out initial windows has our TSO size?"); 1522 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1523 SYSCTL_CHILDREN(bbr_startup), 1524 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1525 &bbr_startup_loss_thresh, 2000, 1526 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1527 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1528 SYSCTL_CHILDREN(bbr_startup), 1529 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1530 &bbr_use_lower_gain_in_startup, 1, 1531 "Should we use a lower hptsi gain if we see loss in startup?"); 1532 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1533 SYSCTL_CHILDREN(bbr_startup), 1534 OID_AUTO, "gain", CTLFLAG_RW, 1535 &bbr_start_exit, 25, 1536 "What gain percent do we need to see to stay in startup??"); 1537 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1538 SYSCTL_CHILDREN(bbr_startup), 1539 OID_AUTO, "low_gain", CTLFLAG_RW, 1540 &bbr_low_start_exit, 15, 1541 "What gain percent do we need to see to stay in the lower gain startup??"); 1542 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1543 SYSCTL_CHILDREN(bbr_startup), 1544 OID_AUTO, "loss_exit", CTLFLAG_RW, 1545 &bbr_exit_startup_at_loss, 1, 1546 "Should we exit startup at loss in an epoch if we are not gaining?"); 1547 /* CWND controls */ 1548 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1549 SYSCTL_CHILDREN(bbr_sysctl_root), 1550 OID_AUTO, 1551 "cwnd", 1552 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1553 "Cwnd controls"); 1554 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1555 SYSCTL_CHILDREN(bbr_cwnd), 1556 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1557 &bbr_cwndtarget_rtt_touse, 0, 1558 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1559 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1560 SYSCTL_CHILDREN(bbr_cwnd), 1561 OID_AUTO, "may_shrink", CTLFLAG_RW, 1562 &bbr_cwnd_may_shrink, 0, 1563 "Can the cwnd shrink if it would grow to more than the target?"); 1564 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1565 SYSCTL_CHILDREN(bbr_cwnd), 1566 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1567 &bbr_target_cwnd_mult_limit, 8, 1568 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1569 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1570 SYSCTL_CHILDREN(bbr_cwnd), 1571 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1572 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1573 "What is the high-speed min cwnd (rttProp under 1ms)"); 1574 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1575 SYSCTL_CHILDREN(bbr_cwnd), 1576 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1577 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1578 "What is the min cwnd (rttProp > 1ms)"); 1579 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1580 SYSCTL_CHILDREN(bbr_cwnd), 1581 OID_AUTO, "initwin", CTLFLAG_RW, 1582 &bbr_def_init_win, 10, 1583 "What is the BBR initial window, if 0 use tcp version"); 1584 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1585 SYSCTL_CHILDREN(bbr_cwnd), 1586 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1587 &bbr_do_red, 600, 1588 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1589 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1590 SYSCTL_CHILDREN(bbr_cwnd), 1591 OID_AUTO, "red_scale", CTLFLAG_RW, 1592 &bbr_red_scale, 20000, 1593 "What RTT do we scale with?"); 1594 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1595 SYSCTL_CHILDREN(bbr_cwnd), 1596 OID_AUTO, "red_growslow", CTLFLAG_RW, 1597 &bbr_red_growth_restrict, 1, 1598 "Do we restrict cwnd growth for whats in flight?"); 1599 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1600 SYSCTL_CHILDREN(bbr_cwnd), 1601 OID_AUTO, "red_div", CTLFLAG_RW, 1602 &bbr_red_div, 2, 1603 "If we reduce whats the divisor?"); 1604 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1605 SYSCTL_CHILDREN(bbr_cwnd), 1606 OID_AUTO, "red_mul", CTLFLAG_RW, 1607 &bbr_red_mul, 1, 1608 "If we reduce whats the mulitiplier?"); 1609 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1610 SYSCTL_CHILDREN(bbr_cwnd), 1611 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1612 &bbr_target_is_bbunit, 0, 1613 "Is the state target the pacing_gain or BBR_UNIT?"); 1614 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1615 SYSCTL_CHILDREN(bbr_cwnd), 1616 OID_AUTO, "drop_limit", CTLFLAG_RW, 1617 &bbr_drop_limit, 0, 1618 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1619 1620 /* Timeout controls */ 1621 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1622 SYSCTL_CHILDREN(bbr_sysctl_root), 1623 OID_AUTO, 1624 "timeout", 1625 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1626 "Time out controls"); 1627 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1628 SYSCTL_CHILDREN(bbr_timeout), 1629 OID_AUTO, "delack", CTLFLAG_RW, 1630 &bbr_delack_time, 100000, 1631 "BBR's delayed ack time"); 1632 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1633 SYSCTL_CHILDREN(bbr_timeout), 1634 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1635 &bbr_tlp_type_to_use, 3, 1636 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1637 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1638 SYSCTL_CHILDREN(bbr_timeout), 1639 OID_AUTO, "persmin", CTLFLAG_RW, 1640 &bbr_persist_min, 250000, 1641 "What is the minimum time in microseconds between persists"); 1642 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1643 SYSCTL_CHILDREN(bbr_timeout), 1644 OID_AUTO, "persmax", CTLFLAG_RW, 1645 &bbr_persist_max, 1000000, 1646 "What is the largest delay in microseconds between persists"); 1647 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1648 SYSCTL_CHILDREN(bbr_timeout), 1649 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1650 &bbr_tlp_min, 10000, 1651 "TLP Min timeout in usecs"); 1652 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1653 SYSCTL_CHILDREN(bbr_timeout), 1654 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1655 &bbr_delayed_ack_time, 200000, 1656 "TLP delayed ack compensation value"); 1657 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1658 SYSCTL_CHILDREN(bbr_sysctl_root), 1659 OID_AUTO, "minrto", CTLFLAG_RW, 1660 &bbr_rto_min_ms, 30, 1661 "Minimum RTO in ms"); 1662 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1663 SYSCTL_CHILDREN(bbr_timeout), 1664 OID_AUTO, "maxrto", CTLFLAG_RW, 1665 &bbr_rto_max_sec, 4, 1666 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1667 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1668 SYSCTL_CHILDREN(bbr_timeout), 1669 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1670 &bbr_tlp_max_resend, 2, 1671 "How many times does TLP retry a single segment or multiple with no ACK"); 1672 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1673 SYSCTL_CHILDREN(bbr_timeout), 1674 OID_AUTO, "minto", CTLFLAG_RW, 1675 &bbr_min_to, 1000, 1676 "Minimum rack timeout in useconds"); 1677 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1678 SYSCTL_CHILDREN(bbr_timeout), 1679 OID_AUTO, "pktdelay", CTLFLAG_RW, 1680 &bbr_pkt_delay, 1000, 1681 "Extra RACK time (in useconds) besides reordering thresh"); 1682 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1683 SYSCTL_CHILDREN(bbr_timeout), 1684 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1685 &bbr_incr_timers, 1, 1686 "Increase the RXT/TLP timer by the pacing time used?"); 1687 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1688 SYSCTL_CHILDREN(bbr_timeout), 1689 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1690 &bbr_marks_rxt_sack_passed, 0, 1691 "Mark sack passed on all those not ack'd when a RXT hits?"); 1692 /* Policer controls */ 1693 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1694 SYSCTL_CHILDREN(bbr_sysctl_root), 1695 OID_AUTO, 1696 "policer", 1697 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1698 "Policer controls"); 1699 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1700 SYSCTL_CHILDREN(bbr_policer), 1701 OID_AUTO, "detect_enable", CTLFLAG_RW, 1702 &bbr_policer_detection_enabled, 1, 1703 "Is policer detection enabled??"); 1704 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1705 SYSCTL_CHILDREN(bbr_policer), 1706 OID_AUTO, "min_pes", CTLFLAG_RW, 1707 &bbr_lt_intvl_min_rtts, 4, 1708 "Minimum number of PE's?"); 1709 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1710 SYSCTL_CHILDREN(bbr_policer), 1711 OID_AUTO, "bwdiff", CTLFLAG_RW, 1712 &bbr_lt_bw_diff, (4000/8), 1713 "Minimal bw diff?"); 1714 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1715 SYSCTL_CHILDREN(bbr_policer), 1716 OID_AUTO, "bwratio", CTLFLAG_RW, 1717 &bbr_lt_bw_ratio, 8, 1718 "Minimal bw diff?"); 1719 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1720 SYSCTL_CHILDREN(bbr_policer), 1721 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1722 &bbr_policer_call_from_rack_to, 0, 1723 "Do we call the policer detection code from a rack-timeout?"); 1724 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1725 SYSCTL_CHILDREN(bbr_policer), 1726 OID_AUTO, "false_postive", CTLFLAG_RW, 1727 &bbr_lt_intvl_fp, 0, 1728 "What packet epoch do we do false-positive detection at (0=no)?"); 1729 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1730 SYSCTL_CHILDREN(bbr_policer), 1731 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1732 &bbr_lt_loss_thresh, 196, 1733 "Loss threshold 196 = 19.6%?"); 1734 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1735 SYSCTL_CHILDREN(bbr_policer), 1736 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1737 &bbr_lt_fd_thresh, 100, 1738 "What percentage is the false detection threshold (150=15.0)?"); 1739 /* All the rest */ 1740 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1741 SYSCTL_CHILDREN(bbr_sysctl_root), 1742 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1743 &bbr_use_rack_resend_cheat, 0, 1744 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1745 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1746 SYSCTL_CHILDREN(bbr_sysctl_root), 1747 OID_AUTO, "error_paceout", CTLFLAG_RW, 1748 &bbr_error_base_paceout, 10000, 1749 "When we hit an error what is the min to pace out in usec's?"); 1750 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1751 SYSCTL_CHILDREN(bbr_sysctl_root), 1752 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1753 &bbr_max_net_error_cnt, 10, 1754 "When we hit this many errors in a row, kill the session?"); 1755 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1756 SYSCTL_CHILDREN(bbr_sysctl_root), 1757 OID_AUTO, "data_after_close", CTLFLAG_RW, 1758 &bbr_ignore_data_after_close, 1, 1759 "Do we hold off sending a RST until all pending data is ack'd"); 1760 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1761 SYSCTL_CHILDREN(bbr_sysctl_root), 1762 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1763 &bbr_resends_use_tso, 0, 1764 "Can resends use TSO?"); 1765 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1766 SYSCTL_CHILDREN(bbr_sysctl_root), 1767 OID_AUTO, "sblklimit", CTLFLAG_RW, 1768 &bbr_sack_block_limit, 128, 1769 "When do we start ignoring small sack blocks"); 1770 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1771 SYSCTL_CHILDREN(bbr_sysctl_root), 1772 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1773 &bbr_verbose_logging, 0, 1774 "Should BBR black box logging be verbose"); 1775 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1776 SYSCTL_CHILDREN(bbr_sysctl_root), 1777 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1778 &bbr_reorder_thresh, 2, 1779 "What factor for rack will be added when seeing reordering (shift right)"); 1780 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1781 SYSCTL_CHILDREN(bbr_sysctl_root), 1782 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1783 &bbr_reorder_fade, 0, 1784 "Does reorder detection fade, if so how many ms (0 means never)"); 1785 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1786 SYSCTL_CHILDREN(bbr_sysctl_root), 1787 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1788 &bbr_tlp_thresh, 1, 1789 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1790 /* Stats and counters */ 1791 /* The pacing counters for hdwr/software can't be in the array */ 1792 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1793 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1794 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1795 SYSCTL_CHILDREN(bbr_sysctl_root), 1796 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1797 &bbr_hdwr_pacing_enobuf, 1798 "Total number of enobufs for hardware paced flows"); 1799 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1800 SYSCTL_CHILDREN(bbr_sysctl_root), 1801 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1802 &bbr_nohdwr_pacing_enobuf, 1803 "Total number of enobufs for non-hardware paced flows"); 1804 1805 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1806 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1807 SYSCTL_CHILDREN(bbr_sysctl_root), 1808 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1809 &bbr_flows_whdwr_pacing, 1810 "Total number of hardware paced flows"); 1811 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1812 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1813 SYSCTL_CHILDREN(bbr_sysctl_root), 1814 OID_AUTO, "software_pacing", CTLFLAG_RD, 1815 &bbr_flows_nohdwr_pacing, 1816 "Total number of software paced flows"); 1817 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1818 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1819 OID_AUTO, "stats", CTLFLAG_RD, 1820 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1821 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1822 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1823 OID_AUTO, "opts", CTLFLAG_RD, 1824 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1825 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1826 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1827 OID_AUTO, "lost", CTLFLAG_RD, 1828 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1829 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1830 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1831 OID_AUTO, "stateresend", CTLFLAG_RD, 1832 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1833 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1834 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1835 OID_AUTO, "statetime", CTLFLAG_RD, 1836 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1837 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1838 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1839 OID_AUTO, "outsize", CTLFLAG_RD, 1840 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1841 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1842 SYSCTL_CHILDREN(bbr_sysctl_root), 1843 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1844 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1845 } 1846 1847 static void 1848 bbr_counter_destroy(void) 1849 { 1850 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1851 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1852 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1853 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1854 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1855 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1856 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1857 counter_u64_free(bbr_hdwr_pacing_enobuf); 1858 counter_u64_free(bbr_flows_whdwr_pacing); 1859 counter_u64_free(bbr_flows_nohdwr_pacing); 1860 1861 } 1862 1863 static __inline void 1864 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1865 { 1866 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1867 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1868 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1869 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1870 l->bw_inuse = bbr_get_bw(bbr); 1871 l->inflight = ctf_flight_size(bbr->rc_tp, 1872 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1873 l->applimited = bbr->r_ctl.r_app_limited_until; 1874 l->delivered = bbr->r_ctl.rc_delivered; 1875 l->timeStamp = cts; 1876 l->lost = bbr->r_ctl.rc_lost; 1877 l->bbr_state = bbr->rc_bbr_state; 1878 l->bbr_substate = bbr_state_val(bbr); 1879 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1880 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1881 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1882 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1883 l->inhpts = tcp_in_hpts(bbr->rc_inp); 1884 l->use_lt_bw = bbr->rc_lt_use_bw; 1885 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1886 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1887 } 1888 1889 static void 1890 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1891 { 1892 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1893 union tcp_log_stackspecific log; 1894 1895 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1896 log.u_bbr.flex1 = 0; 1897 log.u_bbr.flex2 = 0; 1898 log.u_bbr.flex5 = 0; 1899 log.u_bbr.flex3 = 0; 1900 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1901 log.u_bbr.flex7 = reason; 1902 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1903 log.u_bbr.flex8 = 0; 1904 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1905 &bbr->rc_inp->inp_socket->so_rcv, 1906 &bbr->rc_inp->inp_socket->so_snd, 1907 BBR_LOG_BW_RED_EV, 0, 1908 0, &log, false, &bbr->rc_tv); 1909 } 1910 } 1911 1912 static void 1913 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1914 { 1915 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1916 union tcp_log_stackspecific log; 1917 1918 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1919 log.u_bbr.flex1 = seq; 1920 log.u_bbr.flex2 = count; 1921 log.u_bbr.flex8 = mode; 1922 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1923 &bbr->rc_inp->inp_socket->so_rcv, 1924 &bbr->rc_inp->inp_socket->so_snd, 1925 BBR_LOG_LOWGAIN, 0, 1926 0, &log, false, &bbr->rc_tv); 1927 } 1928 } 1929 1930 static void 1931 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1932 uint8_t reason, uint32_t p_maxseg, int len) 1933 { 1934 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1935 union tcp_log_stackspecific log; 1936 1937 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1938 log.u_bbr.flex1 = p_maxseg; 1939 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1940 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1941 log.u_bbr.flex4 = reason; 1942 log.u_bbr.flex5 = bbr->rc_in_persist; 1943 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1944 log.u_bbr.flex7 = p_maxseg; 1945 log.u_bbr.flex8 = bbr->rc_in_persist; 1946 log.u_bbr.pkts_out = 0; 1947 log.u_bbr.applimited = len; 1948 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1949 &bbr->rc_inp->inp_socket->so_rcv, 1950 &bbr->rc_inp->inp_socket->so_snd, 1951 BBR_LOG_JUSTRET, 0, 1952 tlen, &log, false, &bbr->rc_tv); 1953 } 1954 } 1955 1956 static void 1957 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1958 { 1959 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1960 union tcp_log_stackspecific log; 1961 1962 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1963 log.u_bbr.flex1 = seq; 1964 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1965 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1966 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1967 &bbr->rc_inp->inp_socket->so_rcv, 1968 &bbr->rc_inp->inp_socket->so_snd, 1969 BBR_LOG_ENTREC, 0, 1970 0, &log, false, &bbr->rc_tv); 1971 } 1972 } 1973 1974 static void 1975 bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts) 1976 { 1977 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 1978 union tcp_log_stackspecific log; 1979 1980 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1981 log.u_bbr.flex1 = tso; 1982 log.u_bbr.flex2 = maxseg; 1983 log.u_bbr.flex3 = mtu; 1984 log.u_bbr.flex4 = csum_flags; 1985 TCP_LOG_EVENTP(tp, NULL, 1986 &bbr->rc_inp->inp_socket->so_rcv, 1987 &bbr->rc_inp->inp_socket->so_snd, 1988 BBR_LOG_MSGSIZE, 0, 1989 0, &log, false, &bbr->rc_tv); 1990 } 1991 } 1992 1993 static void 1994 bbr_log_flowend(struct tcp_bbr *bbr) 1995 { 1996 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1997 union tcp_log_stackspecific log; 1998 struct sockbuf *r, *s; 1999 struct timeval tv; 2000 2001 if (bbr->rc_inp->inp_socket) { 2002 r = &bbr->rc_inp->inp_socket->so_rcv; 2003 s = &bbr->rc_inp->inp_socket->so_snd; 2004 } else { 2005 r = s = NULL; 2006 } 2007 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2008 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2009 r, s, 2010 TCP_LOG_FLOWEND, 0, 2011 0, &log, false, &tv); 2012 } 2013 } 2014 2015 static void 2016 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2017 uint32_t lost, uint32_t del) 2018 { 2019 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2020 union tcp_log_stackspecific log; 2021 2022 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2023 log.u_bbr.flex1 = lost; 2024 log.u_bbr.flex2 = del; 2025 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2026 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2027 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2028 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2029 log.u_bbr.flex7 = line; 2030 log.u_bbr.flex8 = 0; 2031 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2032 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2033 &bbr->rc_inp->inp_socket->so_rcv, 2034 &bbr->rc_inp->inp_socket->so_snd, 2035 BBR_LOG_PKT_EPOCH, 0, 2036 0, &log, false, &bbr->rc_tv); 2037 } 2038 } 2039 2040 static void 2041 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2042 { 2043 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2044 union tcp_log_stackspecific log; 2045 2046 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2047 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2048 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2049 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2050 log.u_bbr.flex7 = line; 2051 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2052 &bbr->rc_inp->inp_socket->so_rcv, 2053 &bbr->rc_inp->inp_socket->so_snd, 2054 BBR_LOG_TIME_EPOCH, 0, 2055 0, &log, false, &bbr->rc_tv); 2056 } 2057 } 2058 2059 static void 2060 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2061 { 2062 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2063 union tcp_log_stackspecific log; 2064 2065 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2066 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2067 log.u_bbr.flex2 = new_tar; 2068 log.u_bbr.flex3 = line; 2069 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2070 log.u_bbr.flex5 = bbr_quanta; 2071 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2072 log.u_bbr.flex7 = bbr->rc_last_options; 2073 log.u_bbr.flex8 = meth; 2074 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2075 &bbr->rc_inp->inp_socket->so_rcv, 2076 &bbr->rc_inp->inp_socket->so_snd, 2077 BBR_LOG_STATE_TARGET, 0, 2078 0, &log, false, &bbr->rc_tv); 2079 } 2080 2081 } 2082 2083 static void 2084 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2085 { 2086 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2087 union tcp_log_stackspecific log; 2088 2089 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2090 log.u_bbr.flex1 = line; 2091 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2092 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2093 if (bbr_state_is_pkt_epoch) 2094 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2095 else 2096 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2097 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2098 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2099 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2100 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2101 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2102 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2103 &bbr->rc_inp->inp_socket->so_rcv, 2104 &bbr->rc_inp->inp_socket->so_snd, 2105 BBR_LOG_STATE, 0, 2106 0, &log, false, &bbr->rc_tv); 2107 } 2108 } 2109 2110 static void 2111 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2112 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2113 { 2114 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2115 union tcp_log_stackspecific log; 2116 2117 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2118 log.u_bbr.flex1 = line; 2119 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2120 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2121 log.u_bbr.flex4 = applied; 2122 log.u_bbr.flex5 = rtt; 2123 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2124 log.u_bbr.flex7 = cond; 2125 log.u_bbr.flex8 = reas; 2126 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2127 &bbr->rc_inp->inp_socket->so_rcv, 2128 &bbr->rc_inp->inp_socket->so_snd, 2129 BBR_LOG_RTT_SHRINKS, 0, 2130 0, &log, false, &bbr->rc_tv); 2131 } 2132 } 2133 2134 static void 2135 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2136 { 2137 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2138 union tcp_log_stackspecific log; 2139 2140 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2141 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2142 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2143 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2144 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2145 &bbr->rc_inp->inp_socket->so_rcv, 2146 &bbr->rc_inp->inp_socket->so_snd, 2147 BBR_LOG_EXITREC, 0, 2148 0, &log, false, &bbr->rc_tv); 2149 } 2150 } 2151 2152 static void 2153 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2154 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2155 { 2156 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2157 union tcp_log_stackspecific log; 2158 2159 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2160 log.u_bbr.flex1 = line; 2161 log.u_bbr.flex2 = prev_acked; 2162 log.u_bbr.flex3 = bytes_this_ack; 2163 log.u_bbr.flex4 = chg; 2164 log.u_bbr.flex5 = th_ack; 2165 log.u_bbr.flex6 = target; 2166 log.u_bbr.flex8 = meth; 2167 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2168 &bbr->rc_inp->inp_socket->so_rcv, 2169 &bbr->rc_inp->inp_socket->so_snd, 2170 BBR_LOG_CWND, 0, 2171 0, &log, false, &bbr->rc_tv); 2172 } 2173 } 2174 2175 static void 2176 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2177 { 2178 /* 2179 * Log the rtt sample we are applying to the srtt algorithm in 2180 * useconds. 2181 */ 2182 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2183 union tcp_log_stackspecific log; 2184 2185 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2186 log.u_bbr.flex1 = rtt; 2187 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2188 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2189 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2190 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2191 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2192 log.u_bbr.flex6 = tsin; 2193 log.u_bbr.flex7 = 0; 2194 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2195 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2196 &bbr->rc_inp->inp_socket->so_rcv, 2197 &bbr->rc_inp->inp_socket->so_snd, 2198 TCP_LOG_RTT, 0, 2199 0, &log, false, &bbr->rc_tv); 2200 } 2201 } 2202 2203 static void 2204 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2205 { 2206 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2207 union tcp_log_stackspecific log; 2208 2209 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2210 log.u_bbr.flex1 = time_in; 2211 log.u_bbr.flex2 = line; 2212 log.u_bbr.flex8 = enter_exit; 2213 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2214 &bbr->rc_inp->inp_socket->so_rcv, 2215 &bbr->rc_inp->inp_socket->so_snd, 2216 BBR_LOG_PERSIST, 0, 2217 0, &log, false, &bbr->rc_tv); 2218 } 2219 } 2220 static void 2221 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2222 { 2223 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2224 union tcp_log_stackspecific log; 2225 2226 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2227 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2228 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2229 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2230 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2231 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2232 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2233 &bbr->rc_inp->inp_socket->so_rcv, 2234 &bbr->rc_inp->inp_socket->so_snd, 2235 BBR_LOG_ACKCLEAR, 0, 2236 0, &log, false, &bbr->rc_tv); 2237 } 2238 } 2239 2240 static void 2241 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2242 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2243 { 2244 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2245 union tcp_log_stackspecific log; 2246 struct timeval tv; 2247 2248 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2249 log.u_bbr.flex1 = nsegs; 2250 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2251 if (m) { 2252 struct timespec ts; 2253 2254 log.u_bbr.flex3 = m->m_flags; 2255 if (m->m_flags & M_TSTMP) { 2256 mbuf_tstmp2timespec(m, &ts); 2257 tv.tv_sec = ts.tv_sec; 2258 tv.tv_usec = ts.tv_nsec / 1000; 2259 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2260 } else { 2261 log.u_bbr.lt_epoch = 0; 2262 } 2263 if (m->m_flags & M_TSTMP_LRO) { 2264 mbuf_tstmp2timeval(m, &tv); 2265 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2266 } else { 2267 /* No arrival timestamp */ 2268 log.u_bbr.flex5 = 0; 2269 } 2270 2271 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2272 } else { 2273 log.u_bbr.flex3 = 0; 2274 log.u_bbr.flex5 = 0; 2275 log.u_bbr.flex6 = 0; 2276 log.u_bbr.pkts_out = 0; 2277 } 2278 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2279 log.u_bbr.flex7 = bbr->r_wanted_output; 2280 log.u_bbr.flex8 = bbr->rc_in_persist; 2281 TCP_LOG_EVENTP(bbr->rc_tp, th, 2282 &bbr->rc_inp->inp_socket->so_rcv, 2283 &bbr->rc_inp->inp_socket->so_snd, 2284 TCP_LOG_IN, 0, 2285 tlen, &log, true, &bbr->rc_tv); 2286 } 2287 } 2288 2289 static void 2290 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2291 { 2292 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2293 union tcp_log_stackspecific log; 2294 2295 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2296 log.u_bbr.flex1 = did_out; 2297 log.u_bbr.flex2 = nxt_pkt; 2298 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2299 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2300 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2301 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2302 log.u_bbr.flex7 = bbr->r_wanted_output; 2303 log.u_bbr.flex8 = bbr->rc_in_persist; 2304 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2305 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2306 &bbr->rc_inp->inp_socket->so_rcv, 2307 &bbr->rc_inp->inp_socket->so_snd, 2308 BBR_LOG_DOSEG_DONE, 0, 2309 0, &log, true, &bbr->rc_tv); 2310 } 2311 } 2312 2313 static void 2314 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2315 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2316 { 2317 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2318 union tcp_log_stackspecific log; 2319 2320 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2321 log.u_bbr.flex1 = line; 2322 log.u_bbr.flex2 = o_len; 2323 log.u_bbr.flex3 = segcnt; 2324 log.u_bbr.flex4 = segsiz; 2325 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2326 &bbr->rc_inp->inp_socket->so_rcv, 2327 &bbr->rc_inp->inp_socket->so_snd, 2328 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2329 len, &log, true, &bbr->rc_tv); 2330 } 2331 } 2332 2333 static void 2334 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2335 { 2336 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2337 union tcp_log_stackspecific log; 2338 2339 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2340 log.u_bbr.flex1 = timers; 2341 log.u_bbr.flex2 = ret; 2342 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2343 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2344 log.u_bbr.flex5 = cts; 2345 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2346 log.u_bbr.flex8 = hpts_calling; 2347 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2348 &bbr->rc_inp->inp_socket->so_rcv, 2349 &bbr->rc_inp->inp_socket->so_snd, 2350 BBR_LOG_TO_PROCESS, 0, 2351 0, &log, false, &bbr->rc_tv); 2352 } 2353 } 2354 2355 static void 2356 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2357 { 2358 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2359 union tcp_log_stackspecific log; 2360 uint64_t ar; 2361 2362 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2363 log.u_bbr.flex1 = bbr->bbr_timer_src; 2364 log.u_bbr.flex2 = 0; 2365 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2366 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2367 ar >>= 32; 2368 ar &= 0x00000000ffffffff; 2369 log.u_bbr.flex4 = (uint32_t)ar; 2370 ar = (uint64_t)bbr->r_ctl.rc_resend; 2371 ar &= 0x00000000ffffffff; 2372 log.u_bbr.flex5 = (uint32_t)ar; 2373 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2374 log.u_bbr.flex8 = to_num; 2375 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2376 &bbr->rc_inp->inp_socket->so_rcv, 2377 &bbr->rc_inp->inp_socket->so_snd, 2378 BBR_LOG_RTO, 0, 2379 0, &log, false, &bbr->rc_tv); 2380 } 2381 } 2382 2383 static void 2384 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2385 { 2386 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2387 union tcp_log_stackspecific log; 2388 2389 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2390 log.u_bbr.flex1 = flex1; 2391 log.u_bbr.flex2 = flex2; 2392 log.u_bbr.flex3 = flex3; 2393 log.u_bbr.flex4 = 0; 2394 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2395 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2396 log.u_bbr.flex8 = reason; 2397 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2398 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2399 &bbr->rc_inp->inp_socket->so_rcv, 2400 &bbr->rc_inp->inp_socket->so_snd, 2401 BBR_LOG_REDUCE, 0, 2402 0, &log, false, &bbr->rc_tv); 2403 } 2404 } 2405 2406 static void 2407 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2408 { 2409 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2410 union tcp_log_stackspecific log; 2411 2412 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2413 log.u_bbr.flex1 = diag->p_nxt_slot; 2414 log.u_bbr.flex2 = diag->p_cur_slot; 2415 log.u_bbr.flex3 = diag->slot_req; 2416 log.u_bbr.flex4 = diag->inp_hptsslot; 2417 log.u_bbr.flex5 = diag->slot_remaining; 2418 log.u_bbr.flex6 = diag->need_new_to; 2419 log.u_bbr.flex7 = diag->p_hpts_active; 2420 log.u_bbr.flex8 = diag->p_on_min_sleep; 2421 /* Hijack other fields as needed */ 2422 log.u_bbr.epoch = diag->have_slept; 2423 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2424 log.u_bbr.pkts_out = diag->co_ret; 2425 log.u_bbr.applimited = diag->hpts_sleep_time; 2426 log.u_bbr.delivered = diag->p_prev_slot; 2427 log.u_bbr.inflight = diag->p_runningslot; 2428 log.u_bbr.bw_inuse = diag->wheel_slot; 2429 log.u_bbr.rttProp = diag->wheel_cts; 2430 log.u_bbr.delRate = diag->maxslots; 2431 log.u_bbr.cur_del_rate = diag->p_curtick; 2432 log.u_bbr.cur_del_rate <<= 32; 2433 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2434 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2435 &bbr->rc_inp->inp_socket->so_rcv, 2436 &bbr->rc_inp->inp_socket->so_snd, 2437 BBR_LOG_HPTSDIAG, 0, 2438 0, &log, false, &bbr->rc_tv); 2439 } 2440 } 2441 2442 static void 2443 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2444 uint32_t thresh, uint32_t to) 2445 { 2446 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2447 union tcp_log_stackspecific log; 2448 2449 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2450 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2451 log.u_bbr.flex2 = time_since_sent; 2452 log.u_bbr.flex3 = srtt; 2453 log.u_bbr.flex4 = thresh; 2454 log.u_bbr.flex5 = to; 2455 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2456 log.u_bbr.flex8 = mode; 2457 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2458 &bbr->rc_inp->inp_socket->so_rcv, 2459 &bbr->rc_inp->inp_socket->so_snd, 2460 BBR_LOG_TIMERPREP, 0, 2461 0, &log, false, &bbr->rc_tv); 2462 } 2463 } 2464 2465 static void 2466 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2467 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2468 { 2469 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2470 union tcp_log_stackspecific log; 2471 2472 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2473 log.u_bbr.flex1 = usecs; 2474 log.u_bbr.flex2 = len; 2475 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2476 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2477 if (override) 2478 log.u_bbr.flex5 = (1 << 2); 2479 else 2480 log.u_bbr.flex5 = 0; 2481 log.u_bbr.flex6 = override; 2482 log.u_bbr.flex7 = gain; 2483 log.u_bbr.flex8 = mod; 2484 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2485 &bbr->rc_inp->inp_socket->so_rcv, 2486 &bbr->rc_inp->inp_socket->so_snd, 2487 BBR_LOG_HPTSI_CALC, 0, 2488 len, &log, false, &bbr->rc_tv); 2489 } 2490 } 2491 2492 static void 2493 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2494 { 2495 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2496 union tcp_log_stackspecific log; 2497 2498 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2499 2500 log.u_bbr.flex1 = bbr->bbr_timer_src; 2501 log.u_bbr.flex2 = to; 2502 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2503 log.u_bbr.flex4 = slot; 2504 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; 2505 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2506 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; 2507 log.u_bbr.flex8 = which; 2508 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2509 &bbr->rc_inp->inp_socket->so_rcv, 2510 &bbr->rc_inp->inp_socket->so_snd, 2511 BBR_LOG_TIMERSTAR, 0, 2512 0, &log, false, &bbr->rc_tv); 2513 } 2514 } 2515 2516 static void 2517 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) 2518 { 2519 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2520 union tcp_log_stackspecific log; 2521 2522 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2523 log.u_bbr.flex1 = thresh; 2524 log.u_bbr.flex2 = lro; 2525 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2526 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2527 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2528 log.u_bbr.flex6 = srtt; 2529 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2530 log.u_bbr.flex8 = frm; 2531 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2532 &bbr->rc_inp->inp_socket->so_rcv, 2533 &bbr->rc_inp->inp_socket->so_snd, 2534 BBR_LOG_THRESH_CALC, 0, 2535 0, &log, false, &bbr->rc_tv); 2536 } 2537 } 2538 2539 static void 2540 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2541 { 2542 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2543 union tcp_log_stackspecific log; 2544 2545 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2546 log.u_bbr.flex1 = line; 2547 log.u_bbr.flex2 = bbr->bbr_timer_src; 2548 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2549 log.u_bbr.flex4 = bbr->rc_in_persist; 2550 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2551 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2552 log.u_bbr.flex8 = hpts_removed; 2553 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2554 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2555 &bbr->rc_inp->inp_socket->so_rcv, 2556 &bbr->rc_inp->inp_socket->so_snd, 2557 BBR_LOG_TIMERCANC, 0, 2558 0, &log, false, &bbr->rc_tv); 2559 } 2560 } 2561 2562 static void 2563 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2564 { 2565 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2566 union tcp_log_stackspecific log; 2567 2568 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2569 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2570 log.u_bbr.flex2 = (peer_delta >> 32); 2571 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2572 log.u_bbr.flex4 = (delta >> 32); 2573 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2574 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2575 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2576 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2577 &bbr->rc_inp->inp_socket->so_rcv, 2578 &bbr->rc_inp->inp_socket->so_snd, 2579 BBR_LOG_TSTMP_VAL, 0, 2580 0, &log, false, &bbr->rc_tv); 2581 } 2582 } 2583 2584 static void 2585 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) 2586 { 2587 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2588 union tcp_log_stackspecific log; 2589 2590 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2591 log.u_bbr.flex1 = tsosz; 2592 log.u_bbr.flex2 = tls; 2593 log.u_bbr.flex3 = tcp_min_hptsi_time; 2594 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2595 log.u_bbr.flex5 = old_val; 2596 log.u_bbr.flex6 = maxseg; 2597 log.u_bbr.flex7 = bbr->rc_no_pacing; 2598 log.u_bbr.flex7 <<= 1; 2599 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2600 if (hdwr) 2601 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2602 else 2603 log.u_bbr.flex8 = bbr->rc_use_google; 2604 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2605 &bbr->rc_inp->inp_socket->so_rcv, 2606 &bbr->rc_inp->inp_socket->so_snd, 2607 BBR_LOG_BBRTSO, 0, 2608 0, &log, false, &bbr->rc_tv); 2609 } 2610 } 2611 2612 static void 2613 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2614 uint32_t flags, uint32_t line) 2615 { 2616 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2617 union tcp_log_stackspecific log; 2618 2619 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2620 log.u_bbr.flex1 = line; 2621 log.u_bbr.flex2 = rsm->r_start; 2622 log.u_bbr.flex3 = rsm->r_end; 2623 log.u_bbr.flex4 = rsm->r_delivered; 2624 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2625 log.u_bbr.flex6 = rsm->r_dupack; 2626 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2627 log.u_bbr.flex8 = rsm->r_flags; 2628 /* Hijack the pkts_out fids */ 2629 log.u_bbr.applimited = flags; 2630 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2631 &bbr->rc_inp->inp_socket->so_rcv, 2632 &bbr->rc_inp->inp_socket->so_snd, 2633 BBR_RSM_CLEARED, 0, 2634 0, &log, false, &bbr->rc_tv); 2635 } 2636 } 2637 2638 static void 2639 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2640 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2641 uint32_t flex6, uint32_t pkts_out, int flex7, 2642 uint32_t flex4, uint32_t flex1) 2643 { 2644 2645 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2646 union tcp_log_stackspecific log; 2647 2648 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2649 log.u_bbr.flex1 = flex1; 2650 log.u_bbr.flex2 = flex2; 2651 log.u_bbr.flex3 = flex3; 2652 log.u_bbr.flex4 = flex4; 2653 log.u_bbr.flex5 = flex5; 2654 log.u_bbr.flex6 = flex6; 2655 log.u_bbr.flex7 = flex7; 2656 /* Hijack the pkts_out fids */ 2657 log.u_bbr.pkts_out = pkts_out; 2658 log.u_bbr.flex8 = flex8; 2659 if (bbr->rc_ack_was_delayed) 2660 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2661 else 2662 log.u_bbr.epoch = 0; 2663 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2664 &bbr->rc_inp->inp_socket->so_rcv, 2665 &bbr->rc_inp->inp_socket->so_snd, 2666 BBR_LOG_BBRUPD, 0, 2667 flex2, &log, false, &bbr->rc_tv); 2668 } 2669 } 2670 2671 static void 2672 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2673 uint32_t newbw, uint32_t obw, uint32_t diff, 2674 uint32_t tim) 2675 { 2676 if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2677 union tcp_log_stackspecific log; 2678 2679 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2680 log.u_bbr.flex1 = reason; 2681 log.u_bbr.flex2 = newbw; 2682 log.u_bbr.flex3 = obw; 2683 log.u_bbr.flex4 = diff; 2684 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2685 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2686 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2687 log.u_bbr.pkts_out = tim; 2688 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2689 if (bbr->rc_lt_use_bw == 0) 2690 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2691 else 2692 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2693 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2694 &bbr->rc_inp->inp_socket->so_rcv, 2695 &bbr->rc_inp->inp_socket->so_snd, 2696 BBR_LOG_BWSAMP, 0, 2697 0, &log, false, &bbr->rc_tv); 2698 } 2699 } 2700 2701 static inline void 2702 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2703 { 2704 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2705 union tcp_log_stackspecific log; 2706 2707 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2708 log.u_bbr.flex1 = line; 2709 log.u_bbr.flex2 = tick; 2710 log.u_bbr.flex3 = tp->t_maxunacktime; 2711 log.u_bbr.flex4 = tp->t_acktime; 2712 log.u_bbr.flex8 = event; 2713 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2714 &bbr->rc_inp->inp_socket->so_rcv, 2715 &bbr->rc_inp->inp_socket->so_snd, 2716 BBR_LOG_PROGRESS, 0, 2717 0, &log, false, &bbr->rc_tv); 2718 } 2719 } 2720 2721 static void 2722 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2723 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2724 int error) 2725 { 2726 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2727 union tcp_log_stackspecific log; 2728 2729 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2730 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2731 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2732 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2733 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2734 log.u_bbr.bw_inuse = rate; 2735 log.u_bbr.flex5 = line; 2736 log.u_bbr.flex6 = error; 2737 log.u_bbr.flex8 = bbr->skip_gain; 2738 log.u_bbr.flex8 <<= 1; 2739 log.u_bbr.flex8 |= bbr->gain_is_limited; 2740 log.u_bbr.flex8 <<= 1; 2741 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2742 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2743 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2744 &bbr->rc_inp->inp_socket->so_rcv, 2745 &bbr->rc_inp->inp_socket->so_snd, 2746 BBR_LOG_HDWR_PACE, 0, 2747 0, &log, false, &bbr->rc_tv); 2748 } 2749 } 2750 2751 static void 2752 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) 2753 { 2754 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2755 union tcp_log_stackspecific log; 2756 2757 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2758 log.u_bbr.flex1 = slot; 2759 log.u_bbr.flex2 = del_by; 2760 log.u_bbr.flex3 = prev_delay; 2761 log.u_bbr.flex4 = line; 2762 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2763 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2764 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2765 log.u_bbr.flex8 = bbr->rc_in_persist; 2766 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2767 &bbr->rc_inp->inp_socket->so_rcv, 2768 &bbr->rc_inp->inp_socket->so_snd, 2769 BBR_LOG_BBRSND, 0, 2770 len, &log, false, &bbr->rc_tv); 2771 } 2772 } 2773 2774 static void 2775 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) 2776 { 2777 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2778 union tcp_log_stackspecific log; 2779 2780 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2781 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2782 log.u_bbr.flex2 = 0; 2783 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2784 log.u_bbr.flex4 = end; 2785 log.u_bbr.flex5 = seq; 2786 log.u_bbr.flex6 = t; 2787 log.u_bbr.flex7 = match; 2788 log.u_bbr.flex8 = flags; 2789 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2790 &bbr->rc_inp->inp_socket->so_rcv, 2791 &bbr->rc_inp->inp_socket->so_snd, 2792 BBR_LOG_BBRRTT, 0, 2793 0, &log, false, &bbr->rc_tv); 2794 } 2795 } 2796 2797 static void 2798 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2799 { 2800 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2801 union tcp_log_stackspecific log; 2802 2803 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2804 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2805 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2806 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2807 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2808 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2809 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2810 log.u_bbr.flex7 = 0; 2811 log.u_bbr.flex8 = entry_method; 2812 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2813 &bbr->rc_inp->inp_socket->so_rcv, 2814 &bbr->rc_inp->inp_socket->so_snd, 2815 BBR_LOG_EXIT_GAIN, 0, 2816 0, &log, false, &bbr->rc_tv); 2817 } 2818 } 2819 2820 static void 2821 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2822 { 2823 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2824 union tcp_log_stackspecific log; 2825 2826 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2827 /* R-HU */ 2828 log.u_bbr.flex1 = 0; 2829 log.u_bbr.flex2 = 0; 2830 log.u_bbr.flex3 = 0; 2831 log.u_bbr.flex4 = 0; 2832 log.u_bbr.flex7 = 0; 2833 log.u_bbr.flex8 = settings_desired; 2834 2835 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2836 &bbr->rc_inp->inp_socket->so_rcv, 2837 &bbr->rc_inp->inp_socket->so_snd, 2838 BBR_LOG_SETTINGS_CHG, 0, 2839 0, &log, false, &bbr->rc_tv); 2840 } 2841 } 2842 2843 /* 2844 * Returns the bw from the our filter. 2845 */ 2846 static inline uint64_t 2847 bbr_get_full_bw(struct tcp_bbr *bbr) 2848 { 2849 uint64_t bw; 2850 2851 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2852 2853 return (bw); 2854 } 2855 2856 static inline void 2857 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2858 { 2859 uint64_t calclr; 2860 uint32_t lost, del; 2861 2862 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2863 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2864 else 2865 lost = 0; 2866 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2867 if (lost == 0) { 2868 calclr = 0; 2869 } else if (del) { 2870 calclr = lost; 2871 calclr *= (uint64_t)1000; 2872 calclr /= (uint64_t)del; 2873 } else { 2874 /* Nothing delivered? 100.0% loss */ 2875 calclr = 1000; 2876 } 2877 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2878 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2879 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2880 bbr->r_ctl.rc_pkt_epoch++; 2881 if (bbr->rc_no_pacing && 2882 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2883 bbr->rc_no_pacing = 0; 2884 tcp_bbr_tso_size_check(bbr, cts); 2885 } 2886 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2887 bbr->r_ctl.rc_pkt_epoch_time = cts; 2888 /* What was our loss rate */ 2889 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2890 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2891 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2892 } 2893 2894 static inline void 2895 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2896 { 2897 uint32_t epoch_time; 2898 2899 /* Tick the RTT clock */ 2900 bbr->r_ctl.rc_rtt_epoch++; 2901 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2902 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2903 bbr->r_ctl.rc_rcv_epoch_start = cts; 2904 } 2905 2906 static inline void 2907 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2908 { 2909 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2910 bbr->rc_is_pkt_epoch_now = 1; 2911 } 2912 } 2913 2914 /* 2915 * Returns the bw from either the b/w filter 2916 * or from the lt_bw (if the connection is being 2917 * policed). 2918 */ 2919 static inline uint64_t 2920 __bbr_get_bw(struct tcp_bbr *bbr) 2921 { 2922 uint64_t bw, min_bw; 2923 uint64_t rtt; 2924 int gm_measure_cnt = 1; 2925 2926 /* 2927 * For startup we make, like google, a 2928 * minimum b/w. This is generated from the 2929 * IW and the rttProp. We do fall back to srtt 2930 * if for some reason (initial handshake) we don't 2931 * have a rttProp. We, in the worst case, fall back 2932 * to the configured min_bw (rc_initial_hptsi_bw). 2933 */ 2934 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2935 /* Attempt first to use rttProp */ 2936 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2937 if (rtt && (rtt < 0xffffffff)) { 2938 measure: 2939 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2940 ((uint64_t)1000000); 2941 min_bw /= rtt; 2942 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2943 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2944 } 2945 2946 } else if (bbr->rc_tp->t_srtt != 0) { 2947 /* No rttProp, use srtt? */ 2948 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2949 goto measure; 2950 } else { 2951 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2952 } 2953 } else 2954 min_bw = 0; 2955 2956 if ((bbr->rc_past_init_win == 0) && 2957 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2958 bbr->rc_past_init_win = 1; 2959 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2960 gm_measure_cnt = 0; 2961 if (gm_measure_cnt && 2962 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2963 (bbr->rc_past_init_win == 0))) { 2964 /* For google we use our guess rate until we get 1 measurement */ 2965 2966 use_initial_window: 2967 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2968 if (rtt && (rtt < 0xffffffff)) { 2969 /* 2970 * We have an RTT measurement. Use that in 2971 * combination with our initial window to calculate 2972 * a b/w. 2973 */ 2974 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2975 ((uint64_t)1000000); 2976 bw /= rtt; 2977 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2978 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2979 } 2980 } else { 2981 /* Drop back to the 40 and punt to a default */ 2982 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2983 } 2984 if (bw < 1) 2985 /* Probably should panic */ 2986 bw = 1; 2987 if (bw > min_bw) 2988 return (bw); 2989 else 2990 return (min_bw); 2991 } 2992 if (bbr->rc_lt_use_bw) 2993 bw = bbr->r_ctl.rc_lt_bw; 2994 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2995 bw = bbr->r_ctl.red_bw; 2996 else 2997 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2998 if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { 2999 /* 3000 * Enforce user set rate limit, keep in mind that 3001 * t_peakrate_thr is in B/s already 3002 */ 3003 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); 3004 } 3005 if (bw == 0) { 3006 /* We should not be at 0, go to the initial window then */ 3007 goto use_initial_window; 3008 } 3009 if (bw < 1) 3010 /* Probably should panic */ 3011 bw = 1; 3012 if (bw < min_bw) 3013 bw = min_bw; 3014 return (bw); 3015 } 3016 3017 static inline uint64_t 3018 bbr_get_bw(struct tcp_bbr *bbr) 3019 { 3020 uint64_t bw; 3021 3022 bw = __bbr_get_bw(bbr); 3023 return (bw); 3024 } 3025 3026 static inline void 3027 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3028 { 3029 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3030 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3031 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3032 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3033 } 3034 3035 static inline void 3036 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3037 { 3038 bbr->rc_lt_is_sampling = 0; 3039 bbr->rc_lt_use_bw = 0; 3040 bbr->r_ctl.rc_lt_bw = 0; 3041 bbr_reset_lt_bw_interval(bbr, cts); 3042 } 3043 3044 static inline void 3045 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3046 { 3047 uint64_t diff; 3048 3049 /* Do we have a previous sample? */ 3050 if (bbr->r_ctl.rc_lt_bw) { 3051 /* Get the diff in bytes per second */ 3052 if (bbr->r_ctl.rc_lt_bw > bw) 3053 diff = bbr->r_ctl.rc_lt_bw - bw; 3054 else 3055 diff = bw - bbr->r_ctl.rc_lt_bw; 3056 if ((diff <= bbr_lt_bw_diff) || 3057 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3058 /* Consider us policed */ 3059 uint32_t saved_bw; 3060 3061 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3062 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3063 bbr->rc_lt_use_bw = 1; 3064 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3065 /* 3066 * Use pkt based epoch for measuring length of 3067 * policer up 3068 */ 3069 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3070 /* 3071 * reason 4 is we need to start consider being 3072 * policed 3073 */ 3074 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3075 return; 3076 } 3077 } 3078 bbr->r_ctl.rc_lt_bw = bw; 3079 bbr_reset_lt_bw_interval(bbr, cts); 3080 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3081 } 3082 3083 static void 3084 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3085 { 3086 uint32_t ran, deduct; 3087 3088 ran = arc4random_uniform(bbr_rand_ot); 3089 if (ran) { 3090 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3091 bbr->r_ctl.rc_level_state_extra -= deduct; 3092 } 3093 } 3094 /* 3095 * Return randomly the starting state 3096 * to use in probebw. 3097 */ 3098 static uint8_t 3099 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3100 { 3101 uint32_t ran; 3102 uint8_t ret_val; 3103 3104 /* Initialize the offset to 0 */ 3105 bbr->r_ctl.rc_exta_time_gd = 0; 3106 bbr->rc_hit_state_1 = 0; 3107 bbr->r_ctl.rc_level_state_extra = 0; 3108 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3109 /* 3110 * The math works funny here :) the return value is used to set the 3111 * substate and then the state change is called which increments by 3112 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3113 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3114 * we return 1 - 7, so we dont return 0 and end up starting in 3115 * state 1 (DRAIN). 3116 */ 3117 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3118 /* Set an epoch */ 3119 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3120 bbr_set_epoch(bbr, cts, __LINE__); 3121 3122 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3123 return (ret_val); 3124 } 3125 3126 static void 3127 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3128 { 3129 uint32_t diff, d_time; 3130 uint64_t del_time, bw, lost, delivered; 3131 3132 if (bbr->r_use_policer == 0) 3133 return; 3134 if (bbr->rc_lt_use_bw) { 3135 /* We are using lt bw do we stop yet? */ 3136 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3137 if (diff > bbr_lt_bw_max_rtts) { 3138 /* Reset it all */ 3139 reset_all: 3140 bbr_reset_lt_bw_sampling(bbr, cts); 3141 if (bbr->rc_filled_pipe) { 3142 bbr_set_epoch(bbr, cts, __LINE__); 3143 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3144 bbr_substate_change(bbr, cts, __LINE__, 0); 3145 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3146 bbr_log_type_statechange(bbr, cts, __LINE__); 3147 } else { 3148 /* 3149 * This should not happen really 3150 * unless we remove the startup/drain 3151 * restrictions above. 3152 */ 3153 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3154 bbr_set_epoch(bbr, cts, __LINE__); 3155 bbr->r_ctl.rc_bbr_state_time = cts; 3156 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3157 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3158 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3159 bbr_set_state_target(bbr, __LINE__); 3160 bbr_log_type_statechange(bbr, cts, __LINE__); 3161 } 3162 /* reason 0 is to stop using lt-bw */ 3163 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3164 return; 3165 } 3166 if (bbr_lt_intvl_fp == 0) { 3167 /* Not doing false-positive detection */ 3168 return; 3169 } 3170 /* False positive detection */ 3171 if (diff == bbr_lt_intvl_fp) { 3172 /* At bbr_lt_intvl_fp we record the lost */ 3173 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3174 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3175 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3176 /* Now is our loss rate still high? */ 3177 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3178 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3179 if ((delivered == 0) || 3180 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3181 /* No still below our threshold */ 3182 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3183 } else { 3184 /* Yikes its still high, it must be a false positive */ 3185 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3186 goto reset_all; 3187 } 3188 } 3189 return; 3190 } 3191 /* 3192 * Wait for the first loss before sampling, to let the policer 3193 * exhaust its tokens and estimate the steady-state rate allowed by 3194 * the policer. Starting samples earlier includes bursts that 3195 * over-estimate the bw. 3196 */ 3197 if (bbr->rc_lt_is_sampling == 0) { 3198 /* reason 1 is to begin doing the sampling */ 3199 if (loss_detected == 0) 3200 return; 3201 bbr_reset_lt_bw_interval(bbr, cts); 3202 bbr->rc_lt_is_sampling = 1; 3203 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3204 return; 3205 } 3206 /* Now how long were we delivering long term last> */ 3207 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3208 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3209 else 3210 d_time = 0; 3211 3212 /* To avoid underestimates, reset sampling if we run out of data. */ 3213 if (bbr->r_ctl.r_app_limited_until) { 3214 /* Can not measure in app-limited state */ 3215 bbr_reset_lt_bw_sampling(bbr, cts); 3216 /* reason 2 is to reset sampling due to app limits */ 3217 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3218 return; 3219 } 3220 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3221 if (diff < bbr_lt_intvl_min_rtts) { 3222 /* 3223 * need more samples (we don't 3224 * start on a round like linux so 3225 * we need 1 more). 3226 */ 3227 /* 6 is not_enough time or no-loss */ 3228 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3229 return; 3230 } 3231 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3232 /* 3233 * For now if we wait too long, reset all sampling. We need 3234 * to do some research here, its possible that we should 3235 * base this on how much loss as occurred.. something like 3236 * if its under 10% (or some thresh) reset all otherwise 3237 * don't. Thats for phase II I guess. 3238 */ 3239 bbr_reset_lt_bw_sampling(bbr, cts); 3240 /* reason 3 is to reset sampling due too long of sampling */ 3241 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3242 return; 3243 } 3244 /* 3245 * End sampling interval when a packet is lost, so we estimate the 3246 * policer tokens were exhausted. Stopping the sampling before the 3247 * tokens are exhausted under-estimates the policed rate. 3248 */ 3249 if (loss_detected == 0) { 3250 /* 6 is not_enough time or no-loss */ 3251 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3252 return; 3253 } 3254 /* Calculate packets lost and delivered in sampling interval. */ 3255 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3256 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3257 if ((delivered == 0) || 3258 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3259 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3260 return; 3261 } 3262 if (d_time < 1000) { 3263 /* Not enough time. wait */ 3264 /* 6 is not_enough time or no-loss */ 3265 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3266 return; 3267 } 3268 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3269 /* Too long */ 3270 bbr_reset_lt_bw_sampling(bbr, cts); 3271 /* reason 3 is to reset sampling due too long of sampling */ 3272 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3273 return; 3274 } 3275 del_time = d_time; 3276 bw = delivered; 3277 bw *= (uint64_t)USECS_IN_SECOND; 3278 bw /= del_time; 3279 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3280 } 3281 3282 /* 3283 * Allocate a sendmap from our zone. 3284 */ 3285 static struct bbr_sendmap * 3286 bbr_alloc(struct tcp_bbr *bbr) 3287 { 3288 struct bbr_sendmap *rsm; 3289 3290 BBR_STAT_INC(bbr_to_alloc); 3291 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3292 if (rsm) { 3293 bbr->r_ctl.rc_num_maps_alloced++; 3294 return (rsm); 3295 } 3296 if (bbr->r_ctl.rc_free_cnt) { 3297 BBR_STAT_INC(bbr_to_alloc_emerg); 3298 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3299 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3300 bbr->r_ctl.rc_free_cnt--; 3301 return (rsm); 3302 } 3303 BBR_STAT_INC(bbr_to_alloc_failed); 3304 return (NULL); 3305 } 3306 3307 static struct bbr_sendmap * 3308 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3309 { 3310 if ((V_tcp_map_entries_limit > 0) && 3311 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3312 BBR_STAT_INC(bbr_alloc_limited); 3313 if (!bbr->alloc_limit_reported) { 3314 bbr->alloc_limit_reported = 1; 3315 BBR_STAT_INC(bbr_alloc_limited_conns); 3316 } 3317 return (NULL); 3318 } 3319 return (bbr_alloc(bbr)); 3320 } 3321 3322 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3323 static struct bbr_sendmap * 3324 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3325 { 3326 struct bbr_sendmap *rsm; 3327 3328 if (limit_type) { 3329 /* currently there is only one limit type */ 3330 if (V_tcp_map_split_limit > 0 && 3331 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3332 BBR_STAT_INC(bbr_split_limited); 3333 if (!bbr->alloc_limit_reported) { 3334 bbr->alloc_limit_reported = 1; 3335 BBR_STAT_INC(bbr_alloc_limited_conns); 3336 } 3337 return (NULL); 3338 } 3339 } 3340 3341 /* allocate and mark in the limit type, if set */ 3342 rsm = bbr_alloc(bbr); 3343 if (rsm != NULL && limit_type) { 3344 rsm->r_limit_type = limit_type; 3345 bbr->r_ctl.rc_num_split_allocs++; 3346 } 3347 return (rsm); 3348 } 3349 3350 static void 3351 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3352 { 3353 if (rsm->r_limit_type) { 3354 /* currently there is only one limit type */ 3355 bbr->r_ctl.rc_num_split_allocs--; 3356 } 3357 if (rsm->r_is_smallmap) 3358 bbr->r_ctl.rc_num_small_maps_alloced--; 3359 if (bbr->r_ctl.rc_tlp_send == rsm) 3360 bbr->r_ctl.rc_tlp_send = NULL; 3361 if (bbr->r_ctl.rc_resend == rsm) { 3362 bbr->r_ctl.rc_resend = NULL; 3363 } 3364 if (bbr->r_ctl.rc_next == rsm) 3365 bbr->r_ctl.rc_next = NULL; 3366 if (bbr->r_ctl.rc_sacklast == rsm) 3367 bbr->r_ctl.rc_sacklast = NULL; 3368 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3369 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3370 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3371 rsm->r_limit_type = 0; 3372 bbr->r_ctl.rc_free_cnt++; 3373 return; 3374 } 3375 bbr->r_ctl.rc_num_maps_alloced--; 3376 uma_zfree(bbr_zone, rsm); 3377 } 3378 3379 /* 3380 * Returns the BDP. 3381 */ 3382 static uint64_t 3383 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3384 /* 3385 * Calculate the bytes in flight needed given the bw (in bytes per 3386 * second) and the specifyed rtt in useconds. We need to put out the 3387 * returned value per RTT to match that rate. Gain will normally 3388 * raise it up from there. 3389 * 3390 * This should not overflow as long as the bandwidth is below 1 3391 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3392 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3393 */ 3394 uint64_t usec_per_sec; 3395 3396 usec_per_sec = USECS_IN_SECOND; 3397 return ((rtt * bw) / usec_per_sec); 3398 } 3399 3400 /* 3401 * Return the initial cwnd. 3402 */ 3403 static uint32_t 3404 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3405 { 3406 uint32_t i_cwnd; 3407 3408 if (bbr->rc_init_win) { 3409 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3410 } else if (V_tcp_initcwnd_segments) 3411 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3412 max(2 * tp->t_maxseg, 14600)); 3413 else if (V_tcp_do_rfc3390) 3414 i_cwnd = min(4 * tp->t_maxseg, 3415 max(2 * tp->t_maxseg, 4380)); 3416 else { 3417 /* Per RFC5681 Section 3.1 */ 3418 if (tp->t_maxseg > 2190) 3419 i_cwnd = 2 * tp->t_maxseg; 3420 else if (tp->t_maxseg > 1095) 3421 i_cwnd = 3 * tp->t_maxseg; 3422 else 3423 i_cwnd = 4 * tp->t_maxseg; 3424 } 3425 return (i_cwnd); 3426 } 3427 3428 /* 3429 * Given a specified gain, return the target 3430 * cwnd based on that gain. 3431 */ 3432 static uint32_t 3433 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3434 { 3435 uint64_t bdp, rtt; 3436 uint32_t cwnd; 3437 3438 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3439 (bbr_get_full_bw(bbr) == 0)) { 3440 /* No measurements yet */ 3441 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3442 } 3443 /* 3444 * Get bytes per RTT needed (rttProp is normally in 3445 * bbr_cwndtarget_rtt_touse) 3446 */ 3447 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3448 /* Get the bdp from the two values */ 3449 bdp = bbr_get_bw_delay_prod(rtt, bw); 3450 /* Now apply the gain */ 3451 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3452 3453 return (cwnd); 3454 } 3455 3456 static uint32_t 3457 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3458 { 3459 uint32_t cwnd, mss; 3460 3461 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3462 /* Get the base cwnd with gain rounded to a mss */ 3463 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3464 /* 3465 * Add in N (2 default since we do not have a 3466 * fq layer to trap packets in) quanta's per the I-D 3467 * section 4.2.3.2 quanta adjust. 3468 */ 3469 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3470 if (bbr->rc_use_google) { 3471 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3472 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3473 /* 3474 * The linux implementation adds 3475 * an extra 2 x mss in gain cycle which 3476 * is documented no-where except in the code. 3477 * so we add more for Neal undocumented feature 3478 */ 3479 cwnd += 2 * mss; 3480 } 3481 if ((cwnd / mss) & 0x1) { 3482 /* Round up for odd num mss */ 3483 cwnd += mss; 3484 } 3485 } 3486 /* Are we below the min cwnd? */ 3487 if (cwnd < get_min_cwnd(bbr)) 3488 return (get_min_cwnd(bbr)); 3489 return (cwnd); 3490 } 3491 3492 static uint16_t 3493 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3494 { 3495 if (gain < 1) 3496 gain = 1; 3497 return (gain); 3498 } 3499 3500 static uint32_t 3501 bbr_get_header_oh(struct tcp_bbr *bbr) 3502 { 3503 int seg_oh; 3504 3505 seg_oh = 0; 3506 if (bbr->r_ctl.rc_inc_tcp_oh) { 3507 /* Do we include TCP overhead? */ 3508 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3509 } 3510 if (bbr->r_ctl.rc_inc_ip_oh) { 3511 /* Do we include IP overhead? */ 3512 #ifdef INET6 3513 if (bbr->r_is_v6) { 3514 seg_oh += sizeof(struct ip6_hdr); 3515 } else 3516 #endif 3517 { 3518 3519 #ifdef INET 3520 seg_oh += sizeof(struct ip); 3521 #endif 3522 } 3523 } 3524 if (bbr->r_ctl.rc_inc_enet_oh) { 3525 /* Do we include the ethernet overhead? */ 3526 seg_oh += sizeof(struct ether_header); 3527 } 3528 return(seg_oh); 3529 } 3530 3531 static uint32_t 3532 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3533 { 3534 uint64_t divor, res, tim; 3535 3536 if (useconds_time == 0) 3537 return (0); 3538 gain = bbr_gain_adjust(bbr, gain); 3539 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3540 tim = useconds_time; 3541 res = (tim * bw * gain) / divor; 3542 if (res == 0) 3543 res = 1; 3544 return ((uint32_t)res); 3545 } 3546 3547 /* 3548 * Given a gain and a length return the delay in useconds that 3549 * should be used to evenly space out packets 3550 * on the connection (based on the gain factor). 3551 */ 3552 static uint32_t 3553 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3554 { 3555 uint64_t bw, lentim, res; 3556 uint32_t usecs, srtt, over = 0; 3557 uint32_t seg_oh, num_segs, maxseg; 3558 3559 if (len == 0) 3560 return (0); 3561 3562 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3563 num_segs = (len + maxseg - 1) / maxseg; 3564 if (bbr->rc_use_google == 0) { 3565 seg_oh = bbr_get_header_oh(bbr); 3566 len += (num_segs * seg_oh); 3567 } 3568 gain = bbr_gain_adjust(bbr, gain); 3569 bw = bbr_get_bw(bbr); 3570 if (bbr->rc_use_google) { 3571 uint64_t cbw; 3572 3573 /* 3574 * Reduce the b/w by the google discount 3575 * factor 10 = 1%. 3576 */ 3577 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3578 cbw /= (uint64_t)1000; 3579 /* We don't apply a discount if it results in 0 */ 3580 if (cbw > 0) 3581 bw = cbw; 3582 } 3583 lentim = ((uint64_t)len * 3584 (uint64_t)USECS_IN_SECOND * 3585 (uint64_t)BBR_UNIT); 3586 res = lentim / ((uint64_t)gain * bw); 3587 if (res == 0) 3588 res = 1; 3589 usecs = (uint32_t)res; 3590 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3591 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3592 (bbr->rc_use_google == 0) && 3593 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3594 /* 3595 * We cannot let the delay be more than 1/2 the srtt time. 3596 * Otherwise we cannot pace out or send properly. 3597 */ 3598 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3599 BBR_STAT_INC(bbr_hpts_min_time); 3600 } 3601 if (!nolog) 3602 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3603 return (usecs); 3604 } 3605 3606 static void 3607 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3608 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3609 { 3610 uint64_t bw; 3611 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3612 int32_t meth; 3613 3614 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3615 3616 #ifdef STATS 3617 if ((tp->t_flags & TF_GPUTINPROG) && 3618 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3619 /* 3620 * Strech acks and compressed acks will cause this to 3621 * oscillate but we are doing it the same way as the main 3622 * stack so it will be compariable (though possibly not 3623 * ideal). 3624 */ 3625 int32_t cgput; 3626 int64_t gput, time_stamp; 3627 3628 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3629 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3630 cgput = gput / time_stamp; 3631 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3632 cgput); 3633 if (tp->t_stats_gput_prev > 0) 3634 stats_voi_update_abs_s32(tp->t_stats, 3635 VOI_TCP_GPUT_ND, 3636 ((gput - tp->t_stats_gput_prev) * 100) / 3637 tp->t_stats_gput_prev); 3638 tp->t_flags &= ~TF_GPUTINPROG; 3639 tp->t_stats_gput_prev = cgput; 3640 } 3641 #endif 3642 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3643 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3644 /* We don't change anything in probe-rtt */ 3645 return; 3646 } 3647 maxseg = tp->t_maxseg - bbr->rc_last_options; 3648 saved_bytes = bytes_this_ack; 3649 bytes_this_ack += sack_changed; 3650 if (bytes_this_ack > prev_acked) { 3651 bytes_this_ack -= prev_acked; 3652 /* 3653 * A byte ack'd gives us a full mss 3654 * to be like linux i.e. they count packets. 3655 */ 3656 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3657 bytes_this_ack = maxseg; 3658 } else { 3659 /* Unlikely */ 3660 bytes_this_ack = 0; 3661 } 3662 cwnd = tp->snd_cwnd; 3663 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3664 if (bw) 3665 target_cwnd = bbr_get_target_cwnd(bbr, 3666 bw, 3667 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3668 else 3669 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3670 if (IN_RECOVERY(tp->t_flags) && 3671 (bbr->bbr_prev_in_rec == 0)) { 3672 /* 3673 * We are entering recovery and 3674 * thus packet conservation. 3675 */ 3676 bbr->pkt_conservation = 1; 3677 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3678 cwnd = ctf_flight_size(tp, 3679 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3680 bytes_this_ack; 3681 } 3682 if (IN_RECOVERY(tp->t_flags)) { 3683 uint32_t flight; 3684 3685 bbr->bbr_prev_in_rec = 1; 3686 if (cwnd > losses) { 3687 cwnd -= losses; 3688 if (cwnd < maxseg) 3689 cwnd = maxseg; 3690 } else 3691 cwnd = maxseg; 3692 flight = ctf_flight_size(tp, 3693 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3694 bbr_log_type_cwndupd(bbr, flight, 0, 3695 losses, 10, 0, 0, line); 3696 if (bbr->pkt_conservation) { 3697 uint32_t time_in; 3698 3699 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3700 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3701 else 3702 time_in = 0; 3703 3704 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3705 /* Clear packet conservation after an rttProp */ 3706 bbr->pkt_conservation = 0; 3707 } else { 3708 if ((flight + bytes_this_ack) > cwnd) 3709 cwnd = flight + bytes_this_ack; 3710 if (cwnd < get_min_cwnd(bbr)) 3711 cwnd = get_min_cwnd(bbr); 3712 tp->snd_cwnd = cwnd; 3713 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3714 prev_acked, 1, target_cwnd, th->th_ack, line); 3715 return; 3716 } 3717 } 3718 } else 3719 bbr->bbr_prev_in_rec = 0; 3720 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3721 bbr->r_ctl.restrict_growth--; 3722 if (bytes_this_ack > maxseg) 3723 bytes_this_ack = maxseg; 3724 } 3725 if (bbr->rc_filled_pipe) { 3726 /* 3727 * Here we have exited startup and filled the pipe. We will 3728 * thus allow the cwnd to shrink to the target. We hit here 3729 * mostly. 3730 */ 3731 uint32_t s_cwnd; 3732 3733 meth = 2; 3734 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3735 if (s_cwnd > cwnd) 3736 cwnd = s_cwnd; 3737 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3738 cwnd = s_cwnd; 3739 } else { 3740 /* 3741 * Here we are still in startup, we increase cwnd by what 3742 * has been acked. 3743 */ 3744 if ((cwnd < target_cwnd) || 3745 (bbr->rc_past_init_win == 0)) { 3746 meth = 3; 3747 cwnd += bytes_this_ack; 3748 } else { 3749 /* 3750 * Method 4 means we are at target so no gain in 3751 * startup and past the initial window. 3752 */ 3753 meth = 4; 3754 } 3755 } 3756 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3757 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3758 } 3759 3760 static void 3761 tcp_bbr_partialack(struct tcpcb *tp) 3762 { 3763 struct tcp_bbr *bbr; 3764 3765 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3766 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3767 if (ctf_flight_size(tp, 3768 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3769 tp->snd_cwnd) { 3770 bbr->r_wanted_output = 1; 3771 } 3772 } 3773 3774 static void 3775 bbr_post_recovery(struct tcpcb *tp) 3776 { 3777 struct tcp_bbr *bbr; 3778 uint32_t flight; 3779 3780 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3781 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3782 /* 3783 * Here we just exit recovery. 3784 */ 3785 EXIT_RECOVERY(tp->t_flags); 3786 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3787 bbr->r_recovery_bw = 0; 3788 tp->snd_recover = tp->snd_una; 3789 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3790 bbr->pkt_conservation = 0; 3791 if (bbr->rc_use_google == 0) { 3792 /* 3793 * For non-google mode lets 3794 * go ahead and make sure we clear 3795 * the recovery state so if we 3796 * bounce back in to recovery we 3797 * will do PC. 3798 */ 3799 bbr->bbr_prev_in_rec = 0; 3800 } 3801 bbr_log_type_exit_rec(bbr); 3802 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3803 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3804 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3805 } else { 3806 /* For probe-rtt case lets fix up its saved_cwnd */ 3807 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3808 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3809 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3810 } 3811 } 3812 flight = ctf_flight_size(tp, 3813 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3814 if ((bbr->rc_use_google == 0) && 3815 bbr_do_red) { 3816 uint64_t val, lr2use; 3817 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3818 uint32_t *cwnd_p; 3819 3820 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3821 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3822 val /= bbr_get_rtt(bbr, BBR_SRTT); 3823 ratio = (uint32_t)val; 3824 } else 3825 ratio = 1000; 3826 3827 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3828 bbr->r_ctl.recovery_lr, 21, 3829 ratio, 3830 bbr->r_ctl.rc_red_cwnd_pe, 3831 __LINE__); 3832 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3833 goto done; 3834 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3835 bbr_prtt_slam_cwnd) || 3836 (bbr_sub_drain_slam_cwnd && 3837 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3838 bbr->rc_hit_state_1 && 3839 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3840 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3841 bbr_slam_cwnd_in_main_drain)) { 3842 /* 3843 * Here we must poke at the saved cwnd 3844 * as well as the cwnd. 3845 */ 3846 cwnd = bbr->r_ctl.rc_saved_cwnd; 3847 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3848 } else { 3849 cwnd = tp->snd_cwnd; 3850 cwnd_p = &tp->snd_cwnd; 3851 } 3852 maxseg = tp->t_maxseg - bbr->rc_last_options; 3853 /* Add the overall lr with the recovery lr */ 3854 if (bbr->r_ctl.rc_lost == 0) 3855 lr2use = 0; 3856 else if (bbr->r_ctl.rc_delivered == 0) 3857 lr2use = 1000; 3858 else { 3859 lr2use = bbr->r_ctl.rc_lost * 1000; 3860 lr2use /= bbr->r_ctl.rc_delivered; 3861 } 3862 lr2use += bbr->r_ctl.recovery_lr; 3863 acks_inflight = (flight / (maxseg * 2)); 3864 if (bbr_red_scale) { 3865 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3866 lr2use /= bbr_red_scale; 3867 if ((bbr_red_growth_restrict) && 3868 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3869 bbr->r_ctl.restrict_growth += acks_inflight; 3870 } 3871 if (lr2use) { 3872 val = (uint64_t)cwnd * lr2use; 3873 val /= 1000; 3874 if (cwnd > val) 3875 newcwnd = roundup((cwnd - val), maxseg); 3876 else 3877 newcwnd = maxseg; 3878 } else { 3879 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3880 val /= (uint64_t)bbr_red_div; 3881 newcwnd = roundup((uint32_t)val, maxseg); 3882 } 3883 /* with standard delayed acks how many acks can I expect? */ 3884 if (bbr_drop_limit == 0) { 3885 /* 3886 * Anticpate how much we will 3887 * raise the cwnd based on the acks. 3888 */ 3889 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3890 /* We do enforce the min (with the acks) */ 3891 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3892 } 3893 } else { 3894 /* 3895 * A strict drop limit of N is inplace 3896 */ 3897 if (newcwnd < (bbr_drop_limit * maxseg)) { 3898 newcwnd = bbr_drop_limit * maxseg; 3899 } 3900 } 3901 /* For the next N acks do we restrict the growth */ 3902 *cwnd_p = newcwnd; 3903 if (tp->snd_cwnd > newcwnd) 3904 tp->snd_cwnd = newcwnd; 3905 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3906 (uint32_t)lr2use, 3907 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3908 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3909 } 3910 done: 3911 bbr->r_ctl.recovery_lr = 0; 3912 if (flight <= tp->snd_cwnd) { 3913 bbr->r_wanted_output = 1; 3914 } 3915 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3916 } 3917 3918 static void 3919 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3920 { 3921 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3922 /* Limit the drop in b/w to 1/2 our current filter. */ 3923 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3924 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3925 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3926 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3927 tcp_bbr_tso_size_check(bbr, cts); 3928 } 3929 3930 static void 3931 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3932 { 3933 struct tcp_bbr *bbr; 3934 3935 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3936 #ifdef STATS 3937 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3938 #endif 3939 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3940 switch (type) { 3941 case CC_NDUPACK: 3942 if (!IN_RECOVERY(tp->t_flags)) { 3943 tp->snd_recover = tp->snd_max; 3944 /* Start a new epoch */ 3945 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3946 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3947 /* 3948 * Move forward the lt epoch 3949 * so it won't count the truncated 3950 * epoch. 3951 */ 3952 bbr->r_ctl.rc_lt_epoch++; 3953 } 3954 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3955 /* 3956 * Just like the policer detection code 3957 * if we are in startup we must push 3958 * forward the last startup epoch 3959 * to hide the truncated PE. 3960 */ 3961 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3962 } 3963 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3964 ENTER_RECOVERY(tp->t_flags); 3965 bbr->rc_tlp_rtx_out = 0; 3966 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3967 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3968 if (tcp_in_hpts(bbr->rc_inp) && 3969 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3970 /* 3971 * When we enter recovery, we need to restart 3972 * any timers. This may mean we gain an agg 3973 * early, which will be made up for at the last 3974 * rxt out. 3975 */ 3976 bbr->rc_timer_first = 1; 3977 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3978 } 3979 /* 3980 * Calculate a new cwnd based on to the current 3981 * delivery rate with no gain. We get the bdp 3982 * without gaining it up like we normally would and 3983 * we use the last cur_del_rate. 3984 */ 3985 if ((bbr->rc_use_google == 0) && 3986 (bbr->r_ctl.bbr_rttprobe_gain_val || 3987 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3988 tp->snd_cwnd = ctf_flight_size(tp, 3989 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3990 (tp->t_maxseg - bbr->rc_last_options); 3991 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3992 /* We always gate to min cwnd */ 3993 tp->snd_cwnd = get_min_cwnd(bbr); 3994 } 3995 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3996 } 3997 bbr_log_type_enter_rec(bbr, rsm->r_start); 3998 } 3999 break; 4000 case CC_RTO_ERR: 4001 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 4002 /* RTO was unnecessary, so reset everything. */ 4003 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 4004 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 4005 tp->snd_cwnd = tp->snd_cwnd_prev; 4006 tp->snd_ssthresh = tp->snd_ssthresh_prev; 4007 tp->snd_recover = tp->snd_recover_prev; 4008 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 4009 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 4010 } 4011 tp->t_badrxtwin = 0; 4012 break; 4013 } 4014 } 4015 4016 /* 4017 * Indicate whether this ack should be delayed. We can delay the ack if 4018 * following conditions are met: 4019 * - There is no delayed ack timer in progress. 4020 * - Our last ack wasn't a 0-sized window. We never want to delay 4021 * the ack that opens up a 0-sized window. 4022 * - LRO wasn't used for this segment. We make sure by checking that the 4023 * segment size is not larger than the MSS. 4024 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4025 * connection. 4026 * - The data being acked is less than a full segment (a stretch ack 4027 * of more than a segment we should ack. 4028 * - nsegs is 1 (if its more than that we received more than 1 ack). 4029 */ 4030 #define DELAY_ACK(tp, bbr, nsegs) \ 4031 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4032 ((tp->t_flags & TF_DELACK) == 0) && \ 4033 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4034 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4035 4036 /* 4037 * Return the lowest RSM in the map of 4038 * packets still in flight that is not acked. 4039 * This should normally find on the first one 4040 * since we remove packets from the send 4041 * map after they are marked ACKED. 4042 */ 4043 static struct bbr_sendmap * 4044 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4045 { 4046 struct bbr_sendmap *rsm; 4047 4048 /* 4049 * Walk the time-order transmitted list looking for an rsm that is 4050 * not acked. This will be the one that was sent the longest time 4051 * ago that is still outstanding. 4052 */ 4053 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4054 if (rsm->r_flags & BBR_ACKED) { 4055 continue; 4056 } 4057 goto finish; 4058 } 4059 finish: 4060 return (rsm); 4061 } 4062 4063 static struct bbr_sendmap * 4064 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4065 { 4066 struct bbr_sendmap *prsm; 4067 4068 /* 4069 * Walk the sequence order list backward until we hit and arrive at 4070 * the highest seq not acked. In theory when this is called it 4071 * should be the last segment (which it was not). 4072 */ 4073 prsm = rsm; 4074 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4075 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4076 continue; 4077 } 4078 return (prsm); 4079 } 4080 return (NULL); 4081 } 4082 4083 /* 4084 * Returns to the caller the number of microseconds that 4085 * the packet can be outstanding before we think we 4086 * should have had an ack returned. 4087 */ 4088 static uint32_t 4089 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4090 { 4091 /* 4092 * lro is the flag we use to determine if we have seen reordering. 4093 * If it gets set we have seen reordering. The reorder logic either 4094 * works in one of two ways: 4095 * 4096 * If reorder-fade is configured, then we track the last time we saw 4097 * re-ordering occur. If we reach the point where enough time as 4098 * passed we no longer consider reordering has occuring. 4099 * 4100 * Or if reorder-face is 0, then once we see reordering we consider 4101 * the connection to alway be subject to reordering and just set lro 4102 * to 1. 4103 * 4104 * In the end if lro is non-zero we add the extra time for 4105 * reordering in. 4106 */ 4107 int32_t lro; 4108 uint32_t thresh, t_rxtcur; 4109 4110 if (srtt == 0) 4111 srtt = 1; 4112 if (bbr->r_ctl.rc_reorder_ts) { 4113 if (bbr->r_ctl.rc_reorder_fade) { 4114 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4115 lro = cts - bbr->r_ctl.rc_reorder_ts; 4116 if (lro == 0) { 4117 /* 4118 * No time as passed since the last 4119 * reorder, mark it as reordering. 4120 */ 4121 lro = 1; 4122 } 4123 } else { 4124 /* Negative time? */ 4125 lro = 0; 4126 } 4127 if (lro > bbr->r_ctl.rc_reorder_fade) { 4128 /* Turn off reordering seen too */ 4129 bbr->r_ctl.rc_reorder_ts = 0; 4130 lro = 0; 4131 } 4132 } else { 4133 /* Reodering does not fade */ 4134 lro = 1; 4135 } 4136 } else { 4137 lro = 0; 4138 } 4139 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4140 if (lro) { 4141 /* It must be set, if not you get 1/4 rtt */ 4142 if (bbr->r_ctl.rc_reorder_shift) 4143 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4144 else 4145 thresh += (srtt >> 2); 4146 } else { 4147 thresh += 1000; 4148 } 4149 /* We don't let the rack timeout be above a RTO */ 4150 if ((bbr->rc_tp)->t_srtt == 0) 4151 t_rxtcur = BBR_INITIAL_RTO; 4152 else 4153 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4154 if (thresh > t_rxtcur) { 4155 thresh = t_rxtcur; 4156 } 4157 /* And we don't want it above the RTO max either */ 4158 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4159 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4160 } 4161 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4162 return (thresh); 4163 } 4164 4165 /* 4166 * Return to the caller the amount of time in mico-seconds 4167 * that should be used for the TLP timer from the last 4168 * send time of this packet. 4169 */ 4170 static uint32_t 4171 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4172 struct bbr_sendmap *rsm, uint32_t srtt, 4173 uint32_t cts) 4174 { 4175 uint32_t thresh, len, maxseg, t_rxtcur; 4176 struct bbr_sendmap *prsm; 4177 4178 if (srtt == 0) 4179 srtt = 1; 4180 if (bbr->rc_tlp_threshold) 4181 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4182 else 4183 thresh = (srtt * 2); 4184 maxseg = tp->t_maxseg - bbr->rc_last_options; 4185 /* Get the previous sent packet, if any */ 4186 len = rsm->r_end - rsm->r_start; 4187 4188 /* 2.1 behavior */ 4189 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4190 if (prsm && (len <= maxseg)) { 4191 /* 4192 * Two packets outstanding, thresh should be (2*srtt) + 4193 * possible inter-packet delay (if any). 4194 */ 4195 uint32_t inter_gap = 0; 4196 int idx, nidx; 4197 4198 idx = rsm->r_rtr_cnt - 1; 4199 nidx = prsm->r_rtr_cnt - 1; 4200 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4201 /* Yes it was sent later (or at the same time) */ 4202 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4203 } 4204 thresh += inter_gap; 4205 } else if (len <= maxseg) { 4206 /* 4207 * Possibly compensate for delayed-ack. 4208 */ 4209 uint32_t alt_thresh; 4210 4211 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4212 if (alt_thresh > thresh) 4213 thresh = alt_thresh; 4214 } 4215 /* Not above the current RTO */ 4216 if (tp->t_srtt == 0) 4217 t_rxtcur = BBR_INITIAL_RTO; 4218 else 4219 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4220 4221 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4222 /* Not above an RTO */ 4223 if (thresh > t_rxtcur) { 4224 thresh = t_rxtcur; 4225 } 4226 /* Not above a RTO max */ 4227 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4228 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4229 } 4230 /* And now apply the user TLP min */ 4231 if (thresh < bbr_tlp_min) { 4232 thresh = bbr_tlp_min; 4233 } 4234 return (thresh); 4235 } 4236 4237 /* 4238 * Return one of three RTTs to use (in microseconds). 4239 */ 4240 static __inline uint32_t 4241 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4242 { 4243 uint32_t f_rtt; 4244 uint32_t srtt; 4245 4246 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4247 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4248 /* We have no rtt at all */ 4249 if (bbr->rc_tp->t_srtt == 0) 4250 f_rtt = BBR_INITIAL_RTO; 4251 else 4252 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4253 /* 4254 * Since we don't know how good the rtt is apply a 4255 * delayed-ack min 4256 */ 4257 if (f_rtt < bbr_delayed_ack_time) { 4258 f_rtt = bbr_delayed_ack_time; 4259 } 4260 } 4261 /* Take the filter version or last measured pkt-rtt */ 4262 if (rtt_type == BBR_RTT_PROP) { 4263 srtt = f_rtt; 4264 } else if (rtt_type == BBR_RTT_PKTRTT) { 4265 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4266 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4267 } else { 4268 /* No pkt rtt yet */ 4269 srtt = f_rtt; 4270 } 4271 } else if (rtt_type == BBR_RTT_RACK) { 4272 srtt = bbr->r_ctl.rc_last_rtt; 4273 /* We need to add in any internal delay for our timer */ 4274 if (bbr->rc_ack_was_delayed) 4275 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4276 } else if (rtt_type == BBR_SRTT) { 4277 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4278 } else { 4279 /* TSNH */ 4280 srtt = f_rtt; 4281 #ifdef BBR_INVARIANTS 4282 panic("Unknown rtt request type %d", rtt_type); 4283 #endif 4284 } 4285 return (srtt); 4286 } 4287 4288 static int 4289 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4290 { 4291 uint32_t thresh; 4292 4293 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4294 cts, rsm); 4295 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4296 /* It is lost (past time) */ 4297 return (1); 4298 } 4299 return (0); 4300 } 4301 4302 /* 4303 * Return a sendmap if we need to retransmit something. 4304 */ 4305 static struct bbr_sendmap * 4306 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4307 { 4308 /* 4309 * Check to see that we don't need to fall into recovery. We will 4310 * need to do so if our oldest transmit is past the time we should 4311 * have had an ack. 4312 */ 4313 4314 struct bbr_sendmap *rsm; 4315 int32_t idx; 4316 4317 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4318 /* Nothing outstanding that we know of */ 4319 return (NULL); 4320 } 4321 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4322 if (rsm == NULL) { 4323 /* Nothing in the transmit map */ 4324 return (NULL); 4325 } 4326 if (tp->t_flags & TF_SENTFIN) { 4327 /* Fin restricted, don't find anything once a fin is sent */ 4328 return (NULL); 4329 } 4330 if (rsm->r_flags & BBR_ACKED) { 4331 /* 4332 * Ok the first one is acked (this really should not happen 4333 * since we remove the from the tmap once they are acked) 4334 */ 4335 rsm = bbr_find_lowest_rsm(bbr); 4336 if (rsm == NULL) 4337 return (NULL); 4338 } 4339 idx = rsm->r_rtr_cnt - 1; 4340 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4341 /* Send timestamp is the same or less? can't be ready */ 4342 return (NULL); 4343 } 4344 /* Get our RTT time */ 4345 if (bbr_is_lost(bbr, rsm, cts) && 4346 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4347 (rsm->r_flags & BBR_SACK_PASSED))) { 4348 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4349 rsm->r_flags |= BBR_MARKED_LOST; 4350 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4351 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4352 } 4353 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4354 #ifdef BBR_INVARIANTS 4355 if ((rsm->r_end - rsm->r_start) == 0) 4356 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4357 #endif 4358 return (rsm); 4359 } 4360 return (NULL); 4361 } 4362 4363 /* 4364 * RACK Timer, here we simply do logging and house keeping. 4365 * the normal bbr_output_wtime() function will call the 4366 * appropriate thing to check if we need to do a RACK retransmit. 4367 * We return 1, saying don't proceed with bbr_output_wtime only 4368 * when all timers have been stopped (destroyed PCB?). 4369 */ 4370 static int 4371 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4372 { 4373 /* 4374 * This timer simply provides an internal trigger to send out data. 4375 * The check_recovery_mode call will see if there are needed 4376 * retransmissions, if so we will enter fast-recovery. The output 4377 * call may or may not do the same thing depending on sysctl 4378 * settings. 4379 */ 4380 uint32_t lost; 4381 4382 if (bbr->rc_all_timers_stopped) { 4383 return (1); 4384 } 4385 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4386 /* Its not time yet */ 4387 return (0); 4388 } 4389 BBR_STAT_INC(bbr_to_tot); 4390 lost = bbr->r_ctl.rc_lost; 4391 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4392 bbr_set_state(tp, bbr, 0); 4393 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4394 if (bbr->r_ctl.rc_resend == NULL) { 4395 /* Lets do the check here */ 4396 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4397 } 4398 if (bbr_policer_call_from_rack_to) 4399 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4400 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4401 return (0); 4402 } 4403 4404 static __inline void 4405 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4406 { 4407 int idx; 4408 4409 nrsm->r_start = start; 4410 nrsm->r_end = rsm->r_end; 4411 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4412 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4413 nrsm->r_flags = rsm->r_flags; 4414 /* We don't transfer forward the SYN flag */ 4415 nrsm->r_flags &= ~BBR_HAS_SYN; 4416 /* We move forward the FIN flag, not that this should happen */ 4417 rsm->r_flags &= ~BBR_HAS_FIN; 4418 nrsm->r_dupack = rsm->r_dupack; 4419 nrsm->r_rtr_bytes = 0; 4420 nrsm->r_is_gain = rsm->r_is_gain; 4421 nrsm->r_is_drain = rsm->r_is_drain; 4422 nrsm->r_delivered = rsm->r_delivered; 4423 nrsm->r_ts_valid = rsm->r_ts_valid; 4424 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4425 nrsm->r_del_time = rsm->r_del_time; 4426 nrsm->r_app_limited = rsm->r_app_limited; 4427 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4428 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4429 /* We split a piece the lower section looses any just_ret flag. */ 4430 nrsm->r_bbr_state = rsm->r_bbr_state; 4431 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4432 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4433 } 4434 rsm->r_end = nrsm->r_start; 4435 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4436 idx /= 8; 4437 /* Check if we got too small */ 4438 if ((rsm->r_is_smallmap == 0) && 4439 ((rsm->r_end - rsm->r_start) <= idx)) { 4440 bbr->r_ctl.rc_num_small_maps_alloced++; 4441 rsm->r_is_smallmap = 1; 4442 } 4443 /* Check the new one as well */ 4444 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4445 bbr->r_ctl.rc_num_small_maps_alloced++; 4446 nrsm->r_is_smallmap = 1; 4447 } 4448 } 4449 4450 static int 4451 bbr_sack_mergable(struct bbr_sendmap *at, 4452 uint32_t start, uint32_t end) 4453 { 4454 /* 4455 * Given a sack block defined by 4456 * start and end, and a current position 4457 * at. Return 1 if either side of at 4458 * would show that the block is mergable 4459 * to that side. A block to be mergable 4460 * must have overlap with the start/end 4461 * and be in the SACK'd state. 4462 */ 4463 struct bbr_sendmap *l_rsm; 4464 struct bbr_sendmap *r_rsm; 4465 4466 /* first get the either side blocks */ 4467 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4468 r_rsm = TAILQ_NEXT(at, r_next); 4469 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4470 /* Potentially mergeable */ 4471 if ((l_rsm->r_end == start) || 4472 (SEQ_LT(start, l_rsm->r_end) && 4473 SEQ_GT(end, l_rsm->r_end))) { 4474 /* 4475 * map blk |------| 4476 * sack blk |------| 4477 * <or> 4478 * map blk |------| 4479 * sack blk |------| 4480 */ 4481 return (1); 4482 } 4483 } 4484 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4485 /* Potentially mergeable */ 4486 if ((r_rsm->r_start == end) || 4487 (SEQ_LT(start, r_rsm->r_start) && 4488 SEQ_GT(end, r_rsm->r_start))) { 4489 /* 4490 * map blk |---------| 4491 * sack blk |----| 4492 * <or> 4493 * map blk |---------| 4494 * sack blk |-------| 4495 */ 4496 return (1); 4497 } 4498 } 4499 return (0); 4500 } 4501 4502 static struct bbr_sendmap * 4503 bbr_merge_rsm(struct tcp_bbr *bbr, 4504 struct bbr_sendmap *l_rsm, 4505 struct bbr_sendmap *r_rsm) 4506 { 4507 /* 4508 * We are merging two ack'd RSM's, 4509 * the l_rsm is on the left (lower seq 4510 * values) and the r_rsm is on the right 4511 * (higher seq value). The simplest way 4512 * to merge these is to move the right 4513 * one into the left. I don't think there 4514 * is any reason we need to try to find 4515 * the oldest (or last oldest retransmitted). 4516 */ 4517 l_rsm->r_end = r_rsm->r_end; 4518 if (l_rsm->r_dupack < r_rsm->r_dupack) 4519 l_rsm->r_dupack = r_rsm->r_dupack; 4520 if (r_rsm->r_rtr_bytes) 4521 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4522 if (r_rsm->r_in_tmap) { 4523 /* This really should not happen */ 4524 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4525 } 4526 if (r_rsm->r_app_limited) 4527 l_rsm->r_app_limited = r_rsm->r_app_limited; 4528 /* Now the flags */ 4529 if (r_rsm->r_flags & BBR_HAS_FIN) 4530 l_rsm->r_flags |= BBR_HAS_FIN; 4531 if (r_rsm->r_flags & BBR_TLP) 4532 l_rsm->r_flags |= BBR_TLP; 4533 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4534 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4535 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4536 /* This really should not happen */ 4537 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4538 } 4539 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4540 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4541 /* Transfer the split limit to the map we free */ 4542 r_rsm->r_limit_type = l_rsm->r_limit_type; 4543 l_rsm->r_limit_type = 0; 4544 } 4545 bbr_free(bbr, r_rsm); 4546 return(l_rsm); 4547 } 4548 4549 /* 4550 * TLP Timer, here we simply setup what segment we want to 4551 * have the TLP expire on, the normal bbr_output_wtime() will then 4552 * send it out. 4553 * 4554 * We return 1, saying don't proceed with bbr_output_wtime only 4555 * when all timers have been stopped (destroyed PCB?). 4556 */ 4557 static int 4558 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4559 { 4560 /* 4561 * Tail Loss Probe. 4562 */ 4563 struct bbr_sendmap *rsm = NULL; 4564 struct socket *so; 4565 uint32_t amm; 4566 uint32_t out, avail; 4567 uint32_t maxseg; 4568 int collapsed_win = 0; 4569 4570 if (bbr->rc_all_timers_stopped) { 4571 return (1); 4572 } 4573 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4574 /* Its not time yet */ 4575 return (0); 4576 } 4577 if (ctf_progress_timeout_check(tp, true)) { 4578 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4579 return (-ETIMEDOUT); /* tcp_drop() */ 4580 } 4581 /* Did we somehow get into persists? */ 4582 if (bbr->rc_in_persist) { 4583 return (0); 4584 } 4585 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4586 bbr_set_state(tp, bbr, 0); 4587 BBR_STAT_INC(bbr_tlp_tot); 4588 maxseg = tp->t_maxseg - bbr->rc_last_options; 4589 /* 4590 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4591 * need to figure out how to force a full MSS segment out. 4592 */ 4593 so = tptosocket(tp); 4594 avail = sbavail(&so->so_snd); 4595 out = ctf_outstanding(tp); 4596 if (out > tp->snd_wnd) { 4597 /* special case, we need a retransmission */ 4598 collapsed_win = 1; 4599 goto need_retran; 4600 } 4601 if (avail > out) { 4602 /* New data is available */ 4603 amm = avail - out; 4604 if (amm > maxseg) { 4605 amm = maxseg; 4606 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4607 /* not enough to fill a MTU and no-delay is off */ 4608 goto need_retran; 4609 } 4610 /* Set the send-new override */ 4611 if ((out + amm) <= tp->snd_wnd) { 4612 bbr->rc_tlp_new_data = 1; 4613 } else { 4614 goto need_retran; 4615 } 4616 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4617 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4618 bbr->r_ctl.rc_tlp_send = NULL; 4619 /* cap any slots */ 4620 BBR_STAT_INC(bbr_tlp_newdata); 4621 goto send; 4622 } 4623 need_retran: 4624 /* 4625 * Ok we need to arrange the last un-acked segment to be re-sent, or 4626 * optionally the first un-acked segment. 4627 */ 4628 if (collapsed_win == 0) { 4629 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4630 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4631 rsm = bbr_find_high_nonack(bbr, rsm); 4632 } 4633 if (rsm == NULL) { 4634 goto restore; 4635 } 4636 } else { 4637 /* 4638 * We must find the last segment 4639 * that was acceptable by the client. 4640 */ 4641 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4642 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4643 /* Found one */ 4644 break; 4645 } 4646 } 4647 if (rsm == NULL) { 4648 /* None? if so send the first */ 4649 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4650 if (rsm == NULL) 4651 goto restore; 4652 } 4653 } 4654 if ((rsm->r_end - rsm->r_start) > maxseg) { 4655 /* 4656 * We need to split this the last segment in two. 4657 */ 4658 struct bbr_sendmap *nrsm; 4659 4660 nrsm = bbr_alloc_full_limit(bbr); 4661 if (nrsm == NULL) { 4662 /* 4663 * We can't get memory to split, we can either just 4664 * not split it. Or retransmit the whole piece, lets 4665 * do the large send (BTLP :-) ). 4666 */ 4667 goto go_for_it; 4668 } 4669 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4670 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4671 if (rsm->r_in_tmap) { 4672 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4673 nrsm->r_in_tmap = 1; 4674 } 4675 rsm->r_flags &= (~BBR_HAS_FIN); 4676 rsm = nrsm; 4677 } 4678 go_for_it: 4679 bbr->r_ctl.rc_tlp_send = rsm; 4680 bbr->rc_tlp_rtx_out = 1; 4681 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4682 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4683 tp->t_rxtshift++; 4684 } else { 4685 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4686 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4687 } 4688 send: 4689 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4690 /* 4691 * Can't [re]/transmit a segment we have retransmitted the 4692 * max times. We need the retransmit timer to take over. 4693 */ 4694 restore: 4695 bbr->rc_tlp_new_data = 0; 4696 bbr->r_ctl.rc_tlp_send = NULL; 4697 if (rsm) 4698 rsm->r_flags &= ~BBR_TLP; 4699 BBR_STAT_INC(bbr_tlp_retran_fail); 4700 return (0); 4701 } else if (rsm) { 4702 rsm->r_flags |= BBR_TLP; 4703 } 4704 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4705 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4706 /* 4707 * We have retransmitted to many times for TLP. Switch to 4708 * the regular RTO timer 4709 */ 4710 goto restore; 4711 } 4712 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4713 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4714 return (0); 4715 } 4716 4717 /* 4718 * Delayed ack Timer, here we simply need to setup the 4719 * ACK_NOW flag and remove the DELACK flag. From there 4720 * the output routine will send the ack out. 4721 * 4722 * We only return 1, saying don't proceed, if all timers 4723 * are stopped (destroyed PCB?). 4724 */ 4725 static int 4726 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4727 { 4728 if (bbr->rc_all_timers_stopped) { 4729 return (1); 4730 } 4731 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4732 tp->t_flags &= ~TF_DELACK; 4733 tp->t_flags |= TF_ACKNOW; 4734 KMOD_TCPSTAT_INC(tcps_delack); 4735 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4736 return (0); 4737 } 4738 4739 /* 4740 * Here we send a KEEP-ALIVE like probe to the 4741 * peer, we do not send data. 4742 * 4743 * We only return 1, saying don't proceed, if all timers 4744 * are stopped (destroyed PCB?). 4745 */ 4746 static int 4747 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4748 { 4749 struct tcptemp *t_template; 4750 int32_t retval = 1; 4751 4752 if (bbr->rc_all_timers_stopped) { 4753 return (1); 4754 } 4755 if (bbr->rc_in_persist == 0) 4756 return (0); 4757 4758 /* 4759 * Persistence timer into zero window. Force a byte to be output, if 4760 * possible. 4761 */ 4762 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4763 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4764 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4765 /* 4766 * Have we exceeded the user specified progress time? 4767 */ 4768 if (ctf_progress_timeout_check(tp, true)) { 4769 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4770 return (-ETIMEDOUT); /* tcp_drop() */ 4771 } 4772 /* 4773 * Hack: if the peer is dead/unreachable, we do not time out if the 4774 * window is closed. After a full backoff, drop the connection if 4775 * the idle time (no responses to probes) reaches the maximum 4776 * backoff that we would use if retransmitting. 4777 */ 4778 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 4779 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4780 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4781 KMOD_TCPSTAT_INC(tcps_persistdrop); 4782 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4783 return (-ETIMEDOUT); /* tcp_drop() */ 4784 } 4785 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4786 tp->snd_una == tp->snd_max) { 4787 bbr_exit_persist(tp, bbr, cts, __LINE__); 4788 retval = 0; 4789 goto out; 4790 } 4791 /* 4792 * If the user has closed the socket then drop a persisting 4793 * connection after a much reduced timeout. 4794 */ 4795 if (tp->t_state > TCPS_CLOSE_WAIT && 4796 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4797 KMOD_TCPSTAT_INC(tcps_persistdrop); 4798 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4799 return (-ETIMEDOUT); /* tcp_drop() */ 4800 } 4801 t_template = tcpip_maketemplate(bbr->rc_inp); 4802 if (t_template) { 4803 tcp_respond(tp, t_template->tt_ipgen, 4804 &t_template->tt_t, (struct mbuf *)NULL, 4805 tp->rcv_nxt, tp->snd_una - 1, 0); 4806 /* This sends an ack */ 4807 if (tp->t_flags & TF_DELACK) 4808 tp->t_flags &= ~TF_DELACK; 4809 free(t_template, M_TEMP); 4810 } 4811 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 4812 tp->t_rxtshift++; 4813 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4814 out: 4815 return (retval); 4816 } 4817 4818 /* 4819 * If a keepalive goes off, we had no other timers 4820 * happening. We always return 1 here since this 4821 * routine either drops the connection or sends 4822 * out a segment with respond. 4823 */ 4824 static int 4825 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4826 { 4827 struct tcptemp *t_template; 4828 struct inpcb *inp = tptoinpcb(tp); 4829 4830 if (bbr->rc_all_timers_stopped) { 4831 return (1); 4832 } 4833 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4834 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4835 /* 4836 * Keep-alive timer went off; send something or drop connection if 4837 * idle for too long. 4838 */ 4839 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4840 if (tp->t_state < TCPS_ESTABLISHED) 4841 goto dropit; 4842 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4843 tp->t_state <= TCPS_CLOSING) { 4844 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4845 goto dropit; 4846 /* 4847 * Send a packet designed to force a response if the peer is 4848 * up and reachable: either an ACK if the connection is 4849 * still alive, or an RST if the peer has closed the 4850 * connection due to timeout or reboot. Using sequence 4851 * number tp->snd_una-1 causes the transmitted zero-length 4852 * segment to lie outside the receive window; by the 4853 * protocol spec, this requires the correspondent TCP to 4854 * respond. 4855 */ 4856 KMOD_TCPSTAT_INC(tcps_keepprobe); 4857 t_template = tcpip_maketemplate(inp); 4858 if (t_template) { 4859 tcp_respond(tp, t_template->tt_ipgen, 4860 &t_template->tt_t, (struct mbuf *)NULL, 4861 tp->rcv_nxt, tp->snd_una - 1, 0); 4862 free(t_template, M_TEMP); 4863 } 4864 } 4865 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4866 return (1); 4867 dropit: 4868 KMOD_TCPSTAT_INC(tcps_keepdrops); 4869 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4870 return (-ETIMEDOUT); /* tcp_drop() */ 4871 } 4872 4873 /* 4874 * Retransmit helper function, clear up all the ack 4875 * flags and take care of important book keeping. 4876 */ 4877 static void 4878 bbr_remxt_tmr(struct tcpcb *tp) 4879 { 4880 /* 4881 * The retransmit timer went off, all sack'd blocks must be 4882 * un-acked. 4883 */ 4884 struct bbr_sendmap *rsm, *trsm = NULL; 4885 struct tcp_bbr *bbr; 4886 uint32_t cts, lost; 4887 4888 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4889 cts = tcp_get_usecs(&bbr->rc_tv); 4890 lost = bbr->r_ctl.rc_lost; 4891 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4892 bbr_set_state(tp, bbr, 0); 4893 4894 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4895 if (rsm->r_flags & BBR_ACKED) { 4896 uint32_t old_flags; 4897 4898 rsm->r_dupack = 0; 4899 if (rsm->r_in_tmap == 0) { 4900 /* We must re-add it back to the tlist */ 4901 if (trsm == NULL) { 4902 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4903 } else { 4904 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4905 } 4906 rsm->r_in_tmap = 1; 4907 } 4908 old_flags = rsm->r_flags; 4909 rsm->r_flags |= BBR_RXT_CLEARED; 4910 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4911 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4912 } else { 4913 if ((tp->t_state < TCPS_ESTABLISHED) && 4914 (rsm->r_start == tp->snd_una)) { 4915 /* 4916 * Special case for TCP FO. Where 4917 * we sent more data beyond the snd_max. 4918 * We don't mark that as lost and stop here. 4919 */ 4920 break; 4921 } 4922 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4923 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4924 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4925 } 4926 if (bbr_marks_rxt_sack_passed) { 4927 /* 4928 * With this option, we will rack out 4929 * in 1ms increments the rest of the packets. 4930 */ 4931 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4932 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4933 } else { 4934 /* 4935 * With this option we only mark them lost 4936 * and remove all sack'd markings. We will run 4937 * another RXT or a TLP. This will cause 4938 * us to eventually send more based on what 4939 * ack's come in. 4940 */ 4941 rsm->r_flags |= BBR_MARKED_LOST; 4942 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4943 rsm->r_flags &= ~BBR_SACK_PASSED; 4944 } 4945 } 4946 trsm = rsm; 4947 } 4948 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4949 /* Clear the count (we just un-acked them) */ 4950 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4951 bbr->rc_tlp_new_data = 0; 4952 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4953 /* zap the behindness on a rxt */ 4954 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4955 bbr->r_agg_early_set = 0; 4956 bbr->r_ctl.rc_agg_early = 0; 4957 bbr->rc_tlp_rtx_out = 0; 4958 bbr->r_ctl.rc_sacked = 0; 4959 bbr->r_ctl.rc_sacklast = NULL; 4960 bbr->r_timer_override = 1; 4961 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4962 } 4963 4964 /* 4965 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4966 * we will setup to retransmit the lowest seq number outstanding. 4967 */ 4968 static int 4969 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4970 { 4971 struct inpcb *inp = tptoinpcb(tp); 4972 int32_t rexmt; 4973 int32_t retval = 0; 4974 bool isipv6; 4975 4976 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4977 if (bbr->rc_all_timers_stopped) { 4978 return (1); 4979 } 4980 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4981 (tp->snd_una == tp->snd_max)) { 4982 /* Nothing outstanding .. nothing to do */ 4983 return (0); 4984 } 4985 /* 4986 * Retransmission timer went off. Message has not been acked within 4987 * retransmit interval. Back off to a longer retransmit interval 4988 * and retransmit one segment. 4989 */ 4990 if (ctf_progress_timeout_check(tp, true)) { 4991 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4992 return (-ETIMEDOUT); /* tcp_drop() */ 4993 } 4994 bbr_remxt_tmr(tp); 4995 if ((bbr->r_ctl.rc_resend == NULL) || 4996 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 4997 /* 4998 * If the rwnd collapsed on 4999 * the one we are retransmitting 5000 * it does not count against the 5001 * rxt count. 5002 */ 5003 tp->t_rxtshift++; 5004 } 5005 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { 5006 tp->t_rxtshift = TCP_MAXRXTSHIFT; 5007 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 5008 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 5009 /* XXXGL: previously t_softerror was casted to uint16_t */ 5010 MPASS(tp->t_softerror >= 0); 5011 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 5012 return (retval); /* tcp_drop() */ 5013 } 5014 if (tp->t_state == TCPS_SYN_SENT) { 5015 /* 5016 * If the SYN was retransmitted, indicate CWND to be limited 5017 * to 1 segment in cc_conn_init(). 5018 */ 5019 tp->snd_cwnd = 1; 5020 } else if (tp->t_rxtshift == 1) { 5021 /* 5022 * first retransmit; record ssthresh and cwnd so they can be 5023 * recovered if this turns out to be a "bad" retransmit. A 5024 * retransmit is considered "bad" if an ACK for this segment 5025 * is received within RTT/2 interval; the assumption here is 5026 * that the ACK was already in flight. See "On Estimating 5027 * End-to-End Network Path Properties" by Allman and Paxson 5028 * for more details. 5029 */ 5030 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5031 if (!IN_RECOVERY(tp->t_flags)) { 5032 tp->snd_cwnd_prev = tp->snd_cwnd; 5033 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5034 tp->snd_recover_prev = tp->snd_recover; 5035 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5036 tp->t_flags |= TF_PREVVALID; 5037 } else { 5038 tp->t_flags &= ~TF_PREVVALID; 5039 } 5040 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5041 } else { 5042 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5043 tp->t_flags &= ~TF_PREVVALID; 5044 } 5045 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5046 if ((tp->t_state == TCPS_SYN_SENT) || 5047 (tp->t_state == TCPS_SYN_RECEIVED)) 5048 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5049 else 5050 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5051 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5052 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5053 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5054 /* 5055 * We enter the path for PLMTUD if connection is established or, if 5056 * connection is FIN_WAIT_1 status, reason for the last is that if 5057 * amount of data we send is very small, we could send it in couple 5058 * of packets and process straight to FIN. In that case we won't 5059 * catch ESTABLISHED state. 5060 */ 5061 #ifdef INET6 5062 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false; 5063 #else 5064 isipv6 = false; 5065 #endif 5066 if (((V_tcp_pmtud_blackhole_detect == 1) || 5067 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5068 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5069 ((tp->t_state == TCPS_ESTABLISHED) || 5070 (tp->t_state == TCPS_FIN_WAIT_1))) { 5071 /* 5072 * Idea here is that at each stage of mtu probe (usually, 5073 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5074 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5075 * should take care of that. 5076 */ 5077 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5078 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5079 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5080 tp->t_rxtshift % 2 == 0)) { 5081 /* 5082 * Enter Path MTU Black-hole Detection mechanism: - 5083 * Disable Path MTU Discovery (IP "DF" bit). - 5084 * Reduce MTU to lower value than what we negotiated 5085 * with peer. 5086 */ 5087 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5088 /* 5089 * Record that we may have found a black 5090 * hole. 5091 */ 5092 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5093 /* Keep track of previous MSS. */ 5094 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5095 } 5096 /* 5097 * Reduce the MSS to blackhole value or to the 5098 * default in an attempt to retransmit. 5099 */ 5100 #ifdef INET6 5101 isipv6 = bbr->r_is_v6; 5102 if (isipv6 && 5103 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5104 /* Use the sysctl tuneable blackhole MSS. */ 5105 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5106 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5107 } else if (isipv6) { 5108 /* Use the default MSS. */ 5109 tp->t_maxseg = V_tcp_v6mssdflt; 5110 /* 5111 * Disable Path MTU Discovery when we switch 5112 * to minmss. 5113 */ 5114 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5115 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5116 } 5117 #endif 5118 #if defined(INET6) && defined(INET) 5119 else 5120 #endif 5121 #ifdef INET 5122 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5123 /* Use the sysctl tuneable blackhole MSS. */ 5124 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5125 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5126 } else { 5127 /* Use the default MSS. */ 5128 tp->t_maxseg = V_tcp_mssdflt; 5129 /* 5130 * Disable Path MTU Discovery when we switch 5131 * to minmss. 5132 */ 5133 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5134 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5135 } 5136 #endif 5137 } else { 5138 /* 5139 * If further retransmissions are still unsuccessful 5140 * with a lowered MTU, maybe this isn't a blackhole 5141 * and we restore the previous MSS and blackhole 5142 * detection flags. The limit '6' is determined by 5143 * giving each probe stage (1448, 1188, 524) 2 5144 * chances to recover. 5145 */ 5146 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5147 (tp->t_rxtshift >= 6)) { 5148 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5149 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5150 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5151 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5152 } 5153 } 5154 } 5155 /* 5156 * Disable RFC1323 and SACK if we haven't got any response to our 5157 * third SYN to work-around some broken terminal servers (most of 5158 * which have hopefully been retired) that have bad VJ header 5159 * compression code which trashes TCP segments containing 5160 * unknown-to-them TCP options. 5161 */ 5162 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5163 (tp->t_rxtshift == 3)) 5164 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5165 /* 5166 * If we backed off this far, our srtt estimate is probably bogus. 5167 * Clobber it so we'll take the next rtt measurement as our srtt; 5168 * move the current srtt into rttvar to keep the current retransmit 5169 * times until then. 5170 */ 5171 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5172 #ifdef INET6 5173 if (bbr->r_is_v6) 5174 in6_losing(inp); 5175 else 5176 #endif 5177 in_losing(inp); 5178 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5179 tp->t_srtt = 0; 5180 } 5181 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5182 tp->snd_recover = tp->snd_max; 5183 tp->t_flags |= TF_ACKNOW; 5184 tp->t_rtttime = 0; 5185 5186 return (retval); 5187 } 5188 5189 static int 5190 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5191 { 5192 int32_t ret = 0; 5193 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5194 5195 if (timers == 0) { 5196 return (0); 5197 } 5198 if (tp->t_state == TCPS_LISTEN) { 5199 /* no timers on listen sockets */ 5200 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5201 return (0); 5202 return (1); 5203 } 5204 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5205 uint32_t left; 5206 5207 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5208 ret = -1; 5209 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5210 return (0); 5211 } 5212 if (hpts_calling == 0) { 5213 ret = -2; 5214 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5215 return (0); 5216 } 5217 /* 5218 * Ok our timer went off early and we are not paced false 5219 * alarm, go back to sleep. 5220 */ 5221 left = bbr->r_ctl.rc_timer_exp - cts; 5222 ret = -3; 5223 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5224 tcp_hpts_insert(tptoinpcb(tp), HPTS_USEC_TO_SLOTS(left)); 5225 return (1); 5226 } 5227 bbr->rc_tmr_stopped = 0; 5228 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5229 if (timers & PACE_TMR_DELACK) { 5230 ret = bbr_timeout_delack(tp, bbr, cts); 5231 } else if (timers & PACE_TMR_PERSIT) { 5232 ret = bbr_timeout_persist(tp, bbr, cts); 5233 } else if (timers & PACE_TMR_RACK) { 5234 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5235 ret = bbr_timeout_rack(tp, bbr, cts); 5236 } else if (timers & PACE_TMR_TLP) { 5237 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5238 ret = bbr_timeout_tlp(tp, bbr, cts); 5239 } else if (timers & PACE_TMR_RXT) { 5240 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5241 ret = bbr_timeout_rxt(tp, bbr, cts); 5242 } else if (timers & PACE_TMR_KEEP) { 5243 ret = bbr_timeout_keepalive(tp, bbr, cts); 5244 } 5245 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5246 return (ret); 5247 } 5248 5249 static void 5250 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5251 { 5252 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5253 uint8_t hpts_removed = 0; 5254 5255 if (tcp_in_hpts(bbr->rc_inp) && 5256 (bbr->rc_timer_first == 1)) { 5257 /* 5258 * If we are canceling timer's when we have the 5259 * timer ahead of the output being paced. We also 5260 * must remove ourselves from the hpts. 5261 */ 5262 hpts_removed = 1; 5263 tcp_hpts_remove(bbr->rc_inp); 5264 if (bbr->r_ctl.rc_last_delay_val) { 5265 /* Update the last hptsi delay too */ 5266 uint32_t time_since_send; 5267 5268 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5269 time_since_send = cts - bbr->rc_pacer_started; 5270 else 5271 time_since_send = 0; 5272 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5273 /* Cut down our slot time */ 5274 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5275 } else { 5276 bbr->r_ctl.rc_last_delay_val = 0; 5277 } 5278 bbr->rc_pacer_started = cts; 5279 } 5280 } 5281 bbr->rc_timer_first = 0; 5282 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5283 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5284 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5285 } 5286 } 5287 5288 static int 5289 bbr_stopall(struct tcpcb *tp) 5290 { 5291 struct tcp_bbr *bbr; 5292 5293 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5294 bbr->rc_all_timers_stopped = 1; 5295 return (0); 5296 } 5297 5298 static uint32_t 5299 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5300 { 5301 struct bbr_sendmap *rsm; 5302 5303 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5304 if ((rsm == NULL) || (u_rsm == rsm)) 5305 return (cts); 5306 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5307 } 5308 5309 static void 5310 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5311 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5312 { 5313 int32_t idx; 5314 5315 rsm->r_rtr_cnt++; 5316 rsm->r_dupack = 0; 5317 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5318 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5319 rsm->r_flags |= BBR_OVERMAX; 5320 } 5321 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5322 /* Take off the collapsed flag at rxt */ 5323 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5324 } 5325 if (rsm->r_flags & BBR_MARKED_LOST) { 5326 /* We have retransmitted, its no longer lost */ 5327 rsm->r_flags &= ~BBR_MARKED_LOST; 5328 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5329 } 5330 if (rsm->r_flags & BBR_RXT_CLEARED) { 5331 /* 5332 * We hit a RXT timer on it and 5333 * we cleared the "acked" flag. 5334 * We now have it going back into 5335 * flight, we can remove the cleared 5336 * flag and possibly do accounting on 5337 * this piece. 5338 */ 5339 rsm->r_flags &= ~BBR_RXT_CLEARED; 5340 } 5341 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5342 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5343 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5344 } 5345 idx = rsm->r_rtr_cnt - 1; 5346 rsm->r_tim_lastsent[idx] = cts; 5347 rsm->r_pacing_delay = pacing_time; 5348 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5349 rsm->r_ts_valid = bbr->rc_ts_valid; 5350 if (bbr->rc_ts_valid) 5351 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5352 if (bbr->r_ctl.r_app_limited_until) 5353 rsm->r_app_limited = 1; 5354 else 5355 rsm->r_app_limited = 0; 5356 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5357 rsm->r_bbr_state = bbr_state_val(bbr); 5358 else 5359 rsm->r_bbr_state = 8; 5360 if (rsm->r_flags & BBR_ACKED) { 5361 /* Problably MTU discovery messing with us */ 5362 uint32_t old_flags; 5363 5364 old_flags = rsm->r_flags; 5365 rsm->r_flags &= ~BBR_ACKED; 5366 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5367 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5368 if (bbr->r_ctl.rc_sacked == 0) 5369 bbr->r_ctl.rc_sacklast = NULL; 5370 } 5371 if (rsm->r_in_tmap) { 5372 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5373 } 5374 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5375 rsm->r_in_tmap = 1; 5376 if (rsm->r_flags & BBR_SACK_PASSED) { 5377 /* We have retransmitted due to the SACK pass */ 5378 rsm->r_flags &= ~BBR_SACK_PASSED; 5379 rsm->r_flags |= BBR_WAS_SACKPASS; 5380 } 5381 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5382 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5383 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5384 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5385 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5386 rsm->r_is_gain = 1; 5387 rsm->r_is_drain = 0; 5388 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5389 rsm->r_is_drain = 1; 5390 rsm->r_is_gain = 0; 5391 } else { 5392 rsm->r_is_drain = 0; 5393 rsm->r_is_gain = 0; 5394 } 5395 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5396 } 5397 5398 /* 5399 * Returns 0, or the sequence where we stopped 5400 * updating. We also update the lenp to be the amount 5401 * of data left. 5402 */ 5403 5404 static uint32_t 5405 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5406 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5407 { 5408 /* 5409 * We (re-)transmitted starting at rsm->r_start for some length 5410 * (possibly less than r_end. 5411 */ 5412 struct bbr_sendmap *nrsm; 5413 uint32_t c_end; 5414 int32_t len; 5415 5416 len = *lenp; 5417 c_end = rsm->r_start + len; 5418 if (SEQ_GEQ(c_end, rsm->r_end)) { 5419 /* 5420 * We retransmitted the whole piece or more than the whole 5421 * slopping into the next rsm. 5422 */ 5423 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5424 if (c_end == rsm->r_end) { 5425 *lenp = 0; 5426 return (0); 5427 } else { 5428 int32_t act_len; 5429 5430 /* Hangs over the end return whats left */ 5431 act_len = rsm->r_end - rsm->r_start; 5432 *lenp = (len - act_len); 5433 return (rsm->r_end); 5434 } 5435 /* We don't get out of this block. */ 5436 } 5437 /* 5438 * Here we retransmitted less than the whole thing which means we 5439 * have to split this into what was transmitted and what was not. 5440 */ 5441 nrsm = bbr_alloc_full_limit(bbr); 5442 if (nrsm == NULL) { 5443 *lenp = 0; 5444 return (0); 5445 } 5446 /* 5447 * So here we are going to take the original rsm and make it what we 5448 * retransmitted. nrsm will be the tail portion we did not 5449 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5450 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5451 * 1, 6 and the new piece will be 6, 11. 5452 */ 5453 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5454 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5455 nrsm->r_dupack = 0; 5456 if (rsm->r_in_tmap) { 5457 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5458 nrsm->r_in_tmap = 1; 5459 } 5460 rsm->r_flags &= (~BBR_HAS_FIN); 5461 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5462 *lenp = 0; 5463 return (0); 5464 } 5465 5466 static uint64_t 5467 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5468 { 5469 uint64_t bw; 5470 5471 bw = bbr_get_bw(bbr); 5472 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5473 bw /= (uint64_t)BBR_UNIT; 5474 return(bw); 5475 } 5476 5477 static void 5478 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5479 uint64_t act_rate, uint64_t rate_wanted) 5480 { 5481 /* 5482 * We could not get a full gains worth 5483 * of rate. 5484 */ 5485 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5486 /* we can't even get the real rate */ 5487 uint64_t red; 5488 5489 bbr->skip_gain = 1; 5490 bbr->gain_is_limited = 0; 5491 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5492 if (red) 5493 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5494 } else { 5495 /* We can use a lower gain */ 5496 bbr->skip_gain = 0; 5497 bbr->gain_is_limited = 1; 5498 } 5499 } 5500 5501 static void 5502 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5503 { 5504 const struct tcp_hwrate_limit_table *nrte; 5505 int error, rate = -1; 5506 5507 if (bbr->r_ctl.crte == NULL) 5508 return; 5509 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5510 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5511 /* Lost our routes? */ 5512 /* Clear the way for a re-attempt */ 5513 bbr->bbr_attempt_hdwr_pace = 0; 5514 lost_rate: 5515 bbr->gain_is_limited = 0; 5516 bbr->skip_gain = 0; 5517 bbr->bbr_hdrw_pacing = 0; 5518 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5519 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5520 tcp_bbr_tso_size_check(bbr, cts); 5521 return; 5522 } 5523 rate = bbr_get_hardware_rate(bbr); 5524 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5525 bbr->rc_tp, 5526 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5527 rate, 5528 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5529 &error, NULL); 5530 if (nrte == NULL) { 5531 goto lost_rate; 5532 } 5533 if (nrte != bbr->r_ctl.crte) { 5534 bbr->r_ctl.crte = nrte; 5535 if (error == 0) { 5536 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5537 if (bbr->r_ctl.crte->rate < rate) { 5538 /* We have a problem */ 5539 bbr_setup_less_of_rate(bbr, cts, 5540 bbr->r_ctl.crte->rate, rate); 5541 } else { 5542 /* We are good */ 5543 bbr->gain_is_limited = 0; 5544 bbr->skip_gain = 0; 5545 } 5546 } else { 5547 /* A failure should release the tag */ 5548 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5549 bbr->gain_is_limited = 0; 5550 bbr->skip_gain = 0; 5551 bbr->bbr_hdrw_pacing = 0; 5552 } 5553 bbr_type_log_hdwr_pacing(bbr, 5554 bbr->r_ctl.crte->ptbl->rs_ifp, 5555 rate, 5556 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5557 __LINE__, 5558 cts, 5559 error); 5560 } 5561 } 5562 5563 static void 5564 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5565 { 5566 /* 5567 * If we have hardware pacing support 5568 * we need to factor that in for our 5569 * TSO size. 5570 */ 5571 const struct tcp_hwrate_limit_table *rlp; 5572 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5573 5574 if ((bbr->bbr_hdrw_pacing == 0) || 5575 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5576 (bbr->r_ctl.crte == NULL)) 5577 return; 5578 if (bbr->hw_pacing_set == 0) { 5579 /* Not yet by the hdwr pacing count delay */ 5580 return; 5581 } 5582 if (bbr_hdwr_pace_adjust == 0) { 5583 /* No adjustment */ 5584 return; 5585 } 5586 rlp = bbr->r_ctl.crte; 5587 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5588 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5589 else 5590 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5591 /* 5592 * So lets first get the 5593 * time we will take between 5594 * TSO sized sends currently without 5595 * hardware help. 5596 */ 5597 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5598 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5599 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5600 hdwr_delay *= rlp->time_between; 5601 if (cur_delay > hdwr_delay) 5602 delta = cur_delay - hdwr_delay; 5603 else 5604 delta = 0; 5605 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5606 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5607 1); 5608 if (delta && 5609 (delta < (max(rlp->time_between, 5610 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5611 /* 5612 * Now lets divide by the pacing 5613 * time between each segment the 5614 * hardware sends rounding up and 5615 * derive a bytes from that. We multiply 5616 * that by bbr_hdwr_pace_adjust to get 5617 * more bang for our buck. 5618 * 5619 * The goal is to have the software pacer 5620 * waiting no more than an additional 5621 * pacing delay if we can (without the 5622 * compensation i.e. x bbr_hdwr_pace_adjust). 5623 */ 5624 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5625 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5626 seg_sz *= bbr_hdwr_pace_adjust; 5627 if (bbr_hdwr_pace_floor && 5628 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5629 /* Currently hardware paces 5630 * out rs_min_seg segments at a time. 5631 * We need to make sure we always send at least 5632 * a full burst of bbr_hdwr_pace_floor down. 5633 */ 5634 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5635 } 5636 seg_sz *= maxseg; 5637 } else if (delta == 0) { 5638 /* 5639 * The highest pacing rate is 5640 * above our b/w gained. This means 5641 * we probably are going quite fast at 5642 * the hardware highest rate. Lets just multiply 5643 * the calculated TSO size by the 5644 * multiplier factor (its probably 5645 * 4 segments in the default config for 5646 * mlx). 5647 */ 5648 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5649 if (bbr_hdwr_pace_floor && 5650 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5651 /* Currently hardware paces 5652 * out rs_min_seg segments at a time. 5653 * We need to make sure we always send at least 5654 * a full burst of bbr_hdwr_pace_floor down. 5655 */ 5656 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5657 } 5658 } else { 5659 /* 5660 * The pacing time difference is so 5661 * big that the hardware will 5662 * pace out more rapidly then we 5663 * really want and then we 5664 * will have a long delay. Lets just keep 5665 * the same TSO size so its as if 5666 * we were not using hdwr pacing (we 5667 * just gain a bit of spacing from the 5668 * hardware if seg_sz > 1). 5669 */ 5670 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5671 } 5672 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5673 new_tso = seg_sz; 5674 else 5675 new_tso = bbr->r_ctl.rc_pace_max_segs; 5676 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5677 new_tso = PACE_MAX_IP_BYTES - maxseg; 5678 5679 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5680 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5681 bbr->r_ctl.rc_pace_max_segs = new_tso; 5682 } 5683 } 5684 5685 static void 5686 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5687 { 5688 uint64_t bw; 5689 uint32_t old_tso = 0, new_tso; 5690 uint32_t maxseg, bytes; 5691 uint32_t tls_seg=0; 5692 /* 5693 * Google/linux uses the following algorithm to determine 5694 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5695 * 5696 * bytes = bw_in_bytes_per_second / 1000 5697 * bytes = min(bytes, 64k) 5698 * tso_segs = bytes / MSS 5699 * if (bw < 1.2Mbs) 5700 * min_tso_segs = 1 5701 * else 5702 * min_tso_segs = 2 5703 * tso_segs = max(tso_segs, min_tso_segs) 5704 * 5705 * * Note apply a device specific limit (we apply this in the 5706 * tcp_m_copym). 5707 * Note that before the initial measurement is made google bursts out 5708 * a full iwnd just like new-reno/cubic. 5709 * 5710 * We do not use this algorithm. Instead we 5711 * use a two phased approach: 5712 * 5713 * if ( bw <= per-tcb-cross-over) 5714 * goal_tso = calculate how much with this bw we 5715 * can send in goal-time seconds. 5716 * if (goal_tso > mss) 5717 * seg = goal_tso / mss 5718 * tso = seg * mss 5719 * else 5720 * tso = mss 5721 * if (tso > per-tcb-max) 5722 * tso = per-tcb-max 5723 * else if ( bw > 512Mbps) 5724 * tso = max-tso (64k/mss) 5725 * else 5726 * goal_tso = bw / per-tcb-divsor 5727 * seg = (goal_tso + mss-1)/mss 5728 * tso = seg * mss 5729 * 5730 * if (tso < per-tcb-floor) 5731 * tso = per-tcb-floor 5732 * if (tso > per-tcb-utter_max) 5733 * tso = per-tcb-utter_max 5734 * 5735 * Note the default per-tcb-divisor is 1000 (same as google). 5736 * the goal cross over is 30Mbps however. To recreate googles 5737 * algorithm you need to set: 5738 * 5739 * cross-over = 23,168,000 bps 5740 * goal-time = 18000 5741 * per-tcb-max = 2 5742 * per-tcb-divisor = 1000 5743 * per-tcb-floor = 1 5744 * 5745 * This will get you "google bbr" behavior with respect to tso size. 5746 * 5747 * Note we do set anything TSO size until we are past the initial 5748 * window. Before that we gnerally use either a single MSS 5749 * or we use the full IW size (so we burst a IW at a time) 5750 */ 5751 5752 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5753 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5754 } else { 5755 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5756 } 5757 old_tso = bbr->r_ctl.rc_pace_max_segs; 5758 if (bbr->rc_past_init_win == 0) { 5759 /* 5760 * Not enough data has been acknowledged to make a 5761 * judgement. Set up the initial TSO based on if we 5762 * are sending a full IW at once or not. 5763 */ 5764 if (bbr->rc_use_google) 5765 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5766 else if (bbr->bbr_init_win_cheat) 5767 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5768 else 5769 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5770 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5771 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5772 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5773 bbr->r_ctl.rc_pace_max_segs = maxseg; 5774 } 5775 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5776 bbr_adjust_for_hw_pacing(bbr, cts); 5777 return; 5778 } 5779 /** 5780 * Now lets set the TSO goal based on our delivery rate in 5781 * bytes per second. Note we only do this if 5782 * we have acked at least the initial cwnd worth of data. 5783 */ 5784 bw = bbr_get_bw(bbr); 5785 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5786 (bbr->rc_use_google == 0)) { 5787 /* We clamp to one MSS in recovery */ 5788 new_tso = maxseg; 5789 } else if (bbr->rc_use_google) { 5790 int min_tso_segs; 5791 5792 /* Google considers the gain too */ 5793 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5794 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5795 bw /= BBR_UNIT; 5796 } 5797 bytes = bw / 1024; 5798 if (bytes > (64 * 1024)) 5799 bytes = 64 * 1024; 5800 new_tso = bytes / maxseg; 5801 if (bw < ONE_POINT_TWO_MEG) 5802 min_tso_segs = 1; 5803 else 5804 min_tso_segs = 2; 5805 if (new_tso < min_tso_segs) 5806 new_tso = min_tso_segs; 5807 new_tso *= maxseg; 5808 } else if (bbr->rc_no_pacing) { 5809 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5810 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5811 /* 5812 * Calculate the worse case b/w TSO if we are inserting no 5813 * more than a delay_target number of TSO's. 5814 */ 5815 uint32_t tso_len, min_tso; 5816 5817 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5818 if (tso_len > maxseg) { 5819 new_tso = tso_len / maxseg; 5820 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5821 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5822 new_tso *= maxseg; 5823 } else { 5824 /* 5825 * less than a full sized frame yikes.. long rtt or 5826 * low bw? 5827 */ 5828 min_tso = bbr_minseg(bbr); 5829 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5830 new_tso = rounddown(tso_len, min_tso); 5831 else 5832 new_tso = min_tso; 5833 } 5834 } else if (bw > FIVETWELVE_MBPS) { 5835 /* 5836 * This guy is so fast b/w wise that we can TSO as large as 5837 * possible of segments that the NIC will allow. 5838 */ 5839 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5840 } else { 5841 /* 5842 * This formula is based on attempting to send a segment or 5843 * more every bbr_hptsi_per_second. The default is 1000 5844 * which means you are targeting what you can send every 1ms 5845 * based on the peers bw. 5846 * 5847 * If the number drops to say 500, then you are looking more 5848 * at 2ms and you will raise how much we send in a single 5849 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5850 * trade off of course is you will send more at once and 5851 * thus tend to clump up the sends into larger "bursts" 5852 * building a queue. 5853 */ 5854 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5855 new_tso = roundup(bw, (uint64_t)maxseg); 5856 /* 5857 * Gate the floor to match what our lower than 48Mbps 5858 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5859 * becomes the floor for this calculation. 5860 */ 5861 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5862 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5863 } 5864 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5865 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5866 if (new_tso > PACE_MAX_IP_BYTES) 5867 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5868 /* Enforce an utter maximum. */ 5869 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5870 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5871 } 5872 if (old_tso != new_tso) { 5873 /* Only log changes */ 5874 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5875 bbr->r_ctl.rc_pace_max_segs = new_tso; 5876 } 5877 /* We have hardware pacing! */ 5878 bbr_adjust_for_hw_pacing(bbr, cts); 5879 } 5880 5881 static void 5882 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5883 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5884 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5885 struct sockbuf *sb) 5886 { 5887 5888 struct bbr_sendmap *rsm, *nrsm; 5889 register uint32_t snd_max, snd_una; 5890 uint32_t pacing_time; 5891 /* 5892 * Add to the RACK log of packets in flight or retransmitted. If 5893 * there is a TS option we will use the TS echoed, if not we will 5894 * grab a TS. 5895 * 5896 * Retransmissions will increment the count and move the ts to its 5897 * proper place. Note that if options do not include TS's then we 5898 * won't be able to effectively use the ACK for an RTT on a retran. 5899 * 5900 * Notes about r_start and r_end. Lets consider a send starting at 5901 * sequence 1 for 10 bytes. In such an example the r_start would be 5902 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5903 * This means that r_end is actually the first sequence for the next 5904 * slot (11). 5905 * 5906 */ 5907 INP_WLOCK_ASSERT(tptoinpcb(tp)); 5908 if (err) { 5909 /* 5910 * We don't log errors -- we could but snd_max does not 5911 * advance in this case either. 5912 */ 5913 return; 5914 } 5915 if (th_flags & TH_RST) { 5916 /* 5917 * We don't log resets and we return immediately from 5918 * sending 5919 */ 5920 *abandon = 1; 5921 return; 5922 } 5923 snd_una = tp->snd_una; 5924 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5925 /* 5926 * The call to bbr_log_output is made before bumping 5927 * snd_max. This means we can record one extra byte on a SYN 5928 * or FIN if seq_out is adding more on and a FIN is present 5929 * (and we are not resending). 5930 */ 5931 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5932 len++; 5933 if (th_flags & TH_FIN) 5934 len++; 5935 } 5936 if (SEQ_LEQ((seq_out + len), snd_una)) { 5937 /* Are sending an old segment to induce an ack (keep-alive)? */ 5938 return; 5939 } 5940 if (SEQ_LT(seq_out, snd_una)) { 5941 /* huh? should we panic? */ 5942 uint32_t end; 5943 5944 end = seq_out + len; 5945 seq_out = snd_una; 5946 len = end - seq_out; 5947 } 5948 snd_max = tp->snd_max; 5949 if (len == 0) { 5950 /* We don't log zero window probes */ 5951 return; 5952 } 5953 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5954 /* First question is it a retransmission? */ 5955 if (seq_out == snd_max) { 5956 again: 5957 rsm = bbr_alloc(bbr); 5958 if (rsm == NULL) { 5959 return; 5960 } 5961 rsm->r_flags = 0; 5962 if (th_flags & TH_SYN) 5963 rsm->r_flags |= BBR_HAS_SYN; 5964 if (th_flags & TH_FIN) 5965 rsm->r_flags |= BBR_HAS_FIN; 5966 rsm->r_tim_lastsent[0] = cts; 5967 rsm->r_rtr_cnt = 1; 5968 rsm->r_rtr_bytes = 0; 5969 rsm->r_start = seq_out; 5970 rsm->r_end = rsm->r_start + len; 5971 rsm->r_dupack = 0; 5972 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5973 rsm->r_pacing_delay = pacing_time; 5974 rsm->r_ts_valid = bbr->rc_ts_valid; 5975 if (bbr->rc_ts_valid) 5976 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5977 rsm->r_del_time = bbr->r_ctl.rc_del_time; 5978 if (bbr->r_ctl.r_app_limited_until) 5979 rsm->r_app_limited = 1; 5980 else 5981 rsm->r_app_limited = 0; 5982 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5983 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5984 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5985 /* 5986 * Here we must also add in this rsm since snd_max 5987 * is updated after we return from a new send. 5988 */ 5989 rsm->r_flight_at_send += len; 5990 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 5991 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5992 rsm->r_in_tmap = 1; 5993 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5994 rsm->r_bbr_state = bbr_state_val(bbr); 5995 else 5996 rsm->r_bbr_state = 8; 5997 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5998 rsm->r_is_gain = 1; 5999 rsm->r_is_drain = 0; 6000 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 6001 rsm->r_is_drain = 1; 6002 rsm->r_is_gain = 0; 6003 } else { 6004 rsm->r_is_drain = 0; 6005 rsm->r_is_gain = 0; 6006 } 6007 return; 6008 } 6009 /* 6010 * If we reach here its a retransmission and we need to find it. 6011 */ 6012 more: 6013 if (hintrsm && (hintrsm->r_start == seq_out)) { 6014 rsm = hintrsm; 6015 hintrsm = NULL; 6016 } else if (bbr->r_ctl.rc_next) { 6017 /* We have a hint from a previous run */ 6018 rsm = bbr->r_ctl.rc_next; 6019 } else { 6020 /* No hints sorry */ 6021 rsm = NULL; 6022 } 6023 if ((rsm) && (rsm->r_start == seq_out)) { 6024 /* 6025 * We used rc_next or hintrsm to retransmit, hopefully the 6026 * likely case. 6027 */ 6028 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6029 if (len == 0) { 6030 return; 6031 } else { 6032 goto more; 6033 } 6034 } 6035 /* Ok it was not the last pointer go through it the hard way. */ 6036 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6037 if (rsm->r_start == seq_out) { 6038 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6039 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6040 if (len == 0) { 6041 return; 6042 } else { 6043 continue; 6044 } 6045 } 6046 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6047 /* Transmitted within this piece */ 6048 /* 6049 * Ok we must split off the front and then let the 6050 * update do the rest 6051 */ 6052 nrsm = bbr_alloc_full_limit(bbr); 6053 if (nrsm == NULL) { 6054 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6055 return; 6056 } 6057 /* 6058 * copy rsm to nrsm and then trim the front of rsm 6059 * to not include this part. 6060 */ 6061 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6062 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6063 if (rsm->r_in_tmap) { 6064 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6065 nrsm->r_in_tmap = 1; 6066 } 6067 rsm->r_flags &= (~BBR_HAS_FIN); 6068 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6069 if (len == 0) { 6070 return; 6071 } 6072 } 6073 } 6074 /* 6075 * Hmm not found in map did they retransmit both old and on into the 6076 * new? 6077 */ 6078 if (seq_out == tp->snd_max) { 6079 goto again; 6080 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6081 #ifdef BBR_INVARIANTS 6082 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6083 seq_out, len, tp->snd_una, tp->snd_max); 6084 printf("Starting Dump of all rack entries\n"); 6085 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6086 printf("rsm:%p start:%u end:%u\n", 6087 rsm, rsm->r_start, rsm->r_end); 6088 } 6089 printf("Dump complete\n"); 6090 panic("seq_out not found rack:%p tp:%p", 6091 bbr, tp); 6092 #endif 6093 } else { 6094 #ifdef BBR_INVARIANTS 6095 /* 6096 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6097 * flag) 6098 */ 6099 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6100 seq_out, len, tp->snd_max, tp); 6101 #endif 6102 } 6103 } 6104 6105 static void 6106 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6107 { 6108 /* 6109 * Collapse timeout back the cum-ack moved. 6110 */ 6111 tp->t_rxtshift = 0; 6112 tp->t_softerror = 0; 6113 } 6114 6115 static void 6116 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6117 { 6118 bbr->rtt_valid = 1; 6119 bbr->r_ctl.cur_rtt = rtt_usecs; 6120 bbr->r_ctl.ts_in = tsin; 6121 if (rsm_send_time) 6122 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6123 } 6124 6125 static void 6126 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6127 { 6128 /** 6129 * We have in our bbr control: 6130 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6131 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6132 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6133 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6134 * 6135 * Now we can calculate the time between the sends by doing: 6136 * 6137 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6138 * 6139 * And the peer's time between receiving them by doing: 6140 * 6141 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6142 * 6143 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6144 * We also may find that we can't use the timestamps if say we see 6145 * that the peer_delta indicates that though we may have taken 10ms to 6146 * pace out the data, it only saw 1ms between the two packets. This would 6147 * indicate that somewhere on the path is a batching entity that is giving 6148 * out time-slices of the actual b/w. This would mean we could not use 6149 * reliably the peers timestamps. 6150 * 6151 * We expect delta > peer_delta initially. Until we figure out the 6152 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6153 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6154 * then its 10ms vs our usec. If the peer is running a usec clock we would 6155 * put a 1 there. If the value is faster then ours, we will disable the 6156 * use of timestamps (though we could revist this later if we find it to be not 6157 * just an isolated one or two flows)). 6158 * 6159 * To detect the batching middle boxes we will come up with our compensation and 6160 * if with it in place, we find the peer is drastically off (by some margin) in 6161 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6162 * 6163 */ 6164 uint64_t delta, peer_delta, delta_up; 6165 6166 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6167 if (delta < bbr_min_usec_delta) { 6168 /* 6169 * Have not seen a min amount of time 6170 * between our send times so we can 6171 * make a determination of the timestamp 6172 * yet. 6173 */ 6174 return; 6175 } 6176 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6177 if (peer_delta < bbr_min_peer_delta) { 6178 /* 6179 * We may have enough in the form of 6180 * our delta but the peers number 6181 * has not changed that much. It could 6182 * be its clock ratio is such that 6183 * we need more data (10ms tick) or 6184 * there may be other compression scenarios 6185 * going on. In any event we need the 6186 * spread to be larger. 6187 */ 6188 return; 6189 } 6190 /* Ok lets first see which way our delta is going */ 6191 if (peer_delta > delta) { 6192 /* Very unlikely, the peer without 6193 * compensation shows that it saw 6194 * the two sends arrive further apart 6195 * then we saw then in micro-seconds. 6196 */ 6197 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6198 /* well it looks like the peer is a micro-second clock. */ 6199 bbr->rc_ts_clock_set = 1; 6200 bbr->r_ctl.bbr_peer_tsratio = 1; 6201 } else { 6202 bbr->rc_ts_cant_be_used = 1; 6203 bbr->rc_ts_clock_set = 1; 6204 } 6205 return; 6206 } 6207 /* Ok we know that the peer_delta is smaller than our send distance */ 6208 bbr->rc_ts_clock_set = 1; 6209 /* First question is it within the percentage that they are using usec time? */ 6210 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6211 if ((peer_delta + delta_up) >= delta) { 6212 /* Its a usec clock */ 6213 bbr->r_ctl.bbr_peer_tsratio = 1; 6214 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6215 return; 6216 } 6217 /* Ok if not usec, what about 10usec (though unlikely)? */ 6218 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6219 if (((peer_delta * 10) + delta_up) >= delta) { 6220 bbr->r_ctl.bbr_peer_tsratio = 10; 6221 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6222 return; 6223 } 6224 /* And what about 100usec (though again unlikely)? */ 6225 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6226 if (((peer_delta * 100) + delta_up) >= delta) { 6227 bbr->r_ctl.bbr_peer_tsratio = 100; 6228 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6229 return; 6230 } 6231 /* And how about 1 msec (the most likely one)? */ 6232 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6233 if (((peer_delta * 1000) + delta_up) >= delta) { 6234 bbr->r_ctl.bbr_peer_tsratio = 1000; 6235 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6236 return; 6237 } 6238 /* Ok if not msec could it be 10 msec? */ 6239 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6240 if (((peer_delta * 10000) + delta_up) >= delta) { 6241 bbr->r_ctl.bbr_peer_tsratio = 10000; 6242 return; 6243 } 6244 /* If we fall down here the clock tick so slowly we can't use it */ 6245 bbr->rc_ts_cant_be_used = 1; 6246 bbr->r_ctl.bbr_peer_tsratio = 0; 6247 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6248 } 6249 6250 /* 6251 * Collect new round-trip time estimate 6252 * and update averages and current timeout. 6253 */ 6254 static void 6255 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6256 { 6257 int32_t delta; 6258 uint32_t rtt, tsin; 6259 int32_t rtt_ticks; 6260 6261 if (bbr->rtt_valid == 0) 6262 /* No valid sample */ 6263 return; 6264 6265 rtt = bbr->r_ctl.cur_rtt; 6266 tsin = bbr->r_ctl.ts_in; 6267 if (bbr->rc_prtt_set_ts) { 6268 /* 6269 * We are to force feed the rttProp filter due 6270 * to an entry into PROBE_RTT. This assures 6271 * that the times are sync'd between when we 6272 * go into PROBE_RTT and the filter expiration. 6273 * 6274 * Google does not use a true filter, so they do 6275 * this implicitly since they only keep one value 6276 * and when they enter probe-rtt they update the 6277 * value to the newest rtt. 6278 */ 6279 uint32_t rtt_prop; 6280 6281 bbr->rc_prtt_set_ts = 0; 6282 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6283 if (rtt > rtt_prop) 6284 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6285 else 6286 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6287 } 6288 if (bbr->rc_ack_was_delayed) 6289 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6290 6291 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6292 bbr->r_ctl.rc_lowest_rtt = rtt; 6293 bbr_log_rtt_sample(bbr, rtt, tsin); 6294 if (bbr->r_init_rtt) { 6295 /* 6296 * The initial rtt is not-trusted, nuke it and lets get 6297 * our first valid measurement in. 6298 */ 6299 bbr->r_init_rtt = 0; 6300 tp->t_srtt = 0; 6301 } 6302 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6303 /* 6304 * So we have not yet figured out 6305 * what the peers TSTMP value is 6306 * in (most likely ms). We need a 6307 * series of cum-ack's to determine 6308 * this reliably. 6309 */ 6310 if (bbr->rc_ack_is_cumack) { 6311 if (bbr->rc_ts_data_set) { 6312 /* Lets attempt to determine the timestamp granularity. */ 6313 bbr_make_timestamp_determination(bbr); 6314 } else { 6315 bbr->rc_ts_data_set = 1; 6316 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6317 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6318 } 6319 } else { 6320 /* 6321 * We have to have consecutive acks 6322 * reset any "filled" state to none. 6323 */ 6324 bbr->rc_ts_data_set = 0; 6325 } 6326 } 6327 /* Round it up */ 6328 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6329 if (rtt_ticks == 0) 6330 rtt_ticks = 1; 6331 if (tp->t_srtt != 0) { 6332 /* 6333 * srtt is stored as fixed point with 5 bits after the 6334 * binary point (i.e., scaled by 8). The following magic is 6335 * equivalent to the smoothing algorithm in rfc793 with an 6336 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6337 * Adjust rtt to origin 0. 6338 */ 6339 6340 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6341 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6342 6343 tp->t_srtt += delta; 6344 if (tp->t_srtt <= 0) 6345 tp->t_srtt = 1; 6346 6347 /* 6348 * We accumulate a smoothed rtt variance (actually, a 6349 * smoothed mean difference), then set the retransmit timer 6350 * to smoothed rtt + 4 times the smoothed variance. rttvar 6351 * is stored as fixed point with 4 bits after the binary 6352 * point (scaled by 16). The following is equivalent to 6353 * rfc793 smoothing with an alpha of .75 (rttvar = 6354 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6355 * wired-in beta. 6356 */ 6357 if (delta < 0) 6358 delta = -delta; 6359 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6360 tp->t_rttvar += delta; 6361 if (tp->t_rttvar <= 0) 6362 tp->t_rttvar = 1; 6363 } else { 6364 /* 6365 * No rtt measurement yet - use the unsmoothed rtt. Set the 6366 * variance to half the rtt (so our first retransmit happens 6367 * at 3*rtt). 6368 */ 6369 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6370 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6371 } 6372 KMOD_TCPSTAT_INC(tcps_rttupdated); 6373 tp->t_rttupdated++; 6374 #ifdef STATS 6375 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6376 #endif 6377 /* 6378 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6379 * way we do the smoothing, srtt and rttvar will each average +1/2 6380 * tick of bias. When we compute the retransmit timer, we want 1/2 6381 * tick of rounding and 1 extra tick because of +-1/2 tick 6382 * uncertainty in the firing of the timer. The bias will give us 6383 * exactly the 1.5 tick we need. But, because the bias is 6384 * statistical, we have to test that we don't drop below the minimum 6385 * feasible timer (which is 2 ticks). 6386 */ 6387 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6388 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6389 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6390 6391 /* 6392 * We received an ack for a packet that wasn't retransmitted; it is 6393 * probably safe to discard any error indications we've received 6394 * recently. This isn't quite right, but close enough for now (a 6395 * route might have failed after we sent a segment, and the return 6396 * path might not be symmetrical). 6397 */ 6398 tp->t_softerror = 0; 6399 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6400 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6401 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6402 } 6403 6404 static void 6405 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6406 { 6407 bbr->r_ctl.rc_rtt_shrinks = cts; 6408 if (bbr_can_force_probertt && 6409 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6410 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6411 /* 6412 * We should enter probe-rtt its been too long 6413 * since we have been there. 6414 */ 6415 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6416 } else 6417 bbr_check_probe_rtt_limits(bbr, cts); 6418 } 6419 6420 static void 6421 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6422 { 6423 uint64_t orig_bw; 6424 6425 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6426 /* We never apply a zero measurement */ 6427 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6428 0, 0, 0, 0, 0, 0); 6429 return; 6430 } 6431 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6432 bbr->r_ctl.r_measurement_count++; 6433 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6434 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6435 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6436 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6437 0, 0, 0, 0, 0, 0); 6438 if (orig_bw && 6439 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6440 if (bbr->bbr_hdrw_pacing) { 6441 /* 6442 * Apply a new rate to the hardware 6443 * possibly. 6444 */ 6445 bbr_update_hardware_pacing_rate(bbr, cts); 6446 } 6447 bbr_set_state_target(bbr, __LINE__); 6448 tcp_bbr_tso_size_check(bbr, cts); 6449 if (bbr->r_recovery_bw) { 6450 bbr_setup_red_bw(bbr, cts); 6451 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6452 } 6453 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6454 tcp_bbr_tso_size_check(bbr, cts); 6455 } 6456 6457 static void 6458 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6459 { 6460 if (bbr->rc_in_persist == 0) { 6461 /* We log only when not in persist */ 6462 /* Translate to a Bytes Per Second */ 6463 uint64_t tim, bw, ts_diff, ts_bw; 6464 uint32_t delivered; 6465 6466 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6467 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6468 else 6469 tim = 1; 6470 /* 6471 * Now that we have processed the tim (skipping the sample 6472 * or possibly updating the time, go ahead and 6473 * calculate the cdr. 6474 */ 6475 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6476 bw = (uint64_t)delivered; 6477 bw *= (uint64_t)USECS_IN_SECOND; 6478 bw /= tim; 6479 if (bw == 0) { 6480 /* We must have a calculatable amount */ 6481 return; 6482 } 6483 /* 6484 * If we are using this b/w shove it in now so we 6485 * can see in the trace viewer if it gets over-ridden. 6486 */ 6487 if (rsm->r_ts_valid && 6488 bbr->rc_ts_valid && 6489 bbr->rc_ts_clock_set && 6490 (bbr->rc_ts_cant_be_used == 0) && 6491 bbr->rc_use_ts_limit) { 6492 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6493 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6494 if ((delivered == 0) || 6495 (rtt < 1000)) { 6496 /* Can't use the ts */ 6497 bbr_log_type_bbrupd(bbr, 61, cts, 6498 ts_diff, 6499 bbr->r_ctl.last_inbound_ts, 6500 rsm->r_del_ack_ts, 0, 6501 0, 0, 0, delivered); 6502 } else { 6503 ts_bw = (uint64_t)delivered; 6504 ts_bw *= (uint64_t)USECS_IN_SECOND; 6505 ts_bw /= ts_diff; 6506 bbr_log_type_bbrupd(bbr, 62, cts, 6507 (ts_bw >> 32), 6508 (ts_bw & 0xffffffff), 0, 0, 6509 0, 0, ts_diff, delivered); 6510 if ((bbr->ts_can_raise) && 6511 (ts_bw > bw)) { 6512 bbr_log_type_bbrupd(bbr, 8, cts, 6513 delivered, 6514 ts_diff, 6515 (bw >> 32), 6516 (bw & 0x00000000ffffffff), 6517 0, 0, 0, 0); 6518 bw = ts_bw; 6519 } else if (ts_bw && (ts_bw < bw)) { 6520 bbr_log_type_bbrupd(bbr, 7, cts, 6521 delivered, 6522 ts_diff, 6523 (bw >> 32), 6524 (bw & 0x00000000ffffffff), 6525 0, 0, 0, 0); 6526 bw = ts_bw; 6527 } 6528 } 6529 } 6530 if (rsm->r_first_sent_time && 6531 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6532 uint64_t sbw, sti; 6533 /* 6534 * We use what was in flight at the time of our 6535 * send and the size of this send to figure 6536 * out what we have been sending at (amount). 6537 * For the time we take from the time of 6538 * the send of the first send outstanding 6539 * until this send plus this sends pacing 6540 * time. This gives us a good calculation 6541 * as to the rate we have been sending at. 6542 */ 6543 6544 sbw = (uint64_t)(rsm->r_flight_at_send); 6545 sbw *= (uint64_t)USECS_IN_SECOND; 6546 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6547 sti += rsm->r_pacing_delay; 6548 sbw /= sti; 6549 if (sbw < bw) { 6550 bbr_log_type_bbrupd(bbr, 6, cts, 6551 delivered, 6552 (uint32_t)sti, 6553 (bw >> 32), 6554 (uint32_t)bw, 6555 rsm->r_first_sent_time, 0, (sbw >> 32), 6556 (uint32_t)sbw); 6557 bw = sbw; 6558 } 6559 } 6560 /* Use the google algorithm for b/w measurements */ 6561 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6562 if ((rsm->r_app_limited == 0) || 6563 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6564 tcp_bbr_commit_bw(bbr, cts); 6565 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6566 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6567 } 6568 } 6569 } 6570 6571 static void 6572 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6573 { 6574 if (bbr->rc_in_persist == 0) { 6575 /* We log only when not in persist */ 6576 /* Translate to a Bytes Per Second */ 6577 uint64_t tim, bw; 6578 uint32_t delivered; 6579 int no_apply = 0; 6580 6581 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6582 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6583 else 6584 tim = 1; 6585 /* 6586 * Now that we have processed the tim (skipping the sample 6587 * or possibly updating the time, go ahead and 6588 * calculate the cdr. 6589 */ 6590 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6591 bw = (uint64_t)delivered; 6592 bw *= (uint64_t)USECS_IN_SECOND; 6593 bw /= tim; 6594 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6595 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6596 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6597 6598 no_apply = 1; 6599 } 6600 /* 6601 * If we are using this b/w shove it in now so we 6602 * can see in the trace viewer if it gets over-ridden. 6603 */ 6604 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6605 /* Gate by the sending rate */ 6606 if (rsm->r_first_sent_time && 6607 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6608 uint64_t sbw, sti; 6609 /* 6610 * We use what was in flight at the time of our 6611 * send and the size of this send to figure 6612 * out what we have been sending at (amount). 6613 * For the time we take from the time of 6614 * the send of the first send outstanding 6615 * until this send plus this sends pacing 6616 * time. This gives us a good calculation 6617 * as to the rate we have been sending at. 6618 */ 6619 6620 sbw = (uint64_t)(rsm->r_flight_at_send); 6621 sbw *= (uint64_t)USECS_IN_SECOND; 6622 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6623 sti += rsm->r_pacing_delay; 6624 sbw /= sti; 6625 if (sbw < bw) { 6626 bbr_log_type_bbrupd(bbr, 6, cts, 6627 delivered, 6628 (uint32_t)sti, 6629 (bw >> 32), 6630 (uint32_t)bw, 6631 rsm->r_first_sent_time, 0, (sbw >> 32), 6632 (uint32_t)sbw); 6633 bw = sbw; 6634 } 6635 if ((sti > tim) && 6636 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6637 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6638 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6639 no_apply = 1; 6640 } else 6641 no_apply = 0; 6642 } 6643 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6644 if ((no_apply == 0) && 6645 ((rsm->r_app_limited == 0) || 6646 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6647 tcp_bbr_commit_bw(bbr, cts); 6648 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6649 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6650 } 6651 } 6652 } 6653 6654 static void 6655 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6656 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6657 { 6658 uint64_t old_rttprop; 6659 6660 /* Update our delivery time and amount */ 6661 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6662 bbr->r_ctl.rc_del_time = cts; 6663 if (rtt == 0) { 6664 /* 6665 * 0 means its a retransmit, for now we don't use these for 6666 * the rest of BBR. 6667 */ 6668 return; 6669 } 6670 if ((bbr->rc_use_google == 0) && 6671 (match != BBR_RTT_BY_EXACTMATCH) && 6672 (match != BBR_RTT_BY_TIMESTAMP)){ 6673 /* 6674 * We get a lot of rtt updates, lets not pay attention to 6675 * any that are not an exact match. That way we don't have 6676 * to worry about timestamps and the whole nonsense of 6677 * unsure if its a retransmission etc (if we ever had the 6678 * timestamp fixed to always have the last thing sent this 6679 * would not be a issue). 6680 */ 6681 return; 6682 } 6683 if ((bbr_no_retran && bbr->rc_use_google) && 6684 (match != BBR_RTT_BY_EXACTMATCH) && 6685 (match != BBR_RTT_BY_TIMESTAMP)){ 6686 /* 6687 * We only do measurements in google mode 6688 * with bbr_no_retran on for sure things. 6689 */ 6690 return; 6691 } 6692 /* Only update srtt if we know by exact match */ 6693 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6694 if (ack_type == BBR_CUM_ACKED) 6695 bbr->rc_ack_is_cumack = 1; 6696 else 6697 bbr->rc_ack_is_cumack = 0; 6698 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6699 /* 6700 * Note the following code differs to the original 6701 * BBR spec. It calls for <= not <. However after a 6702 * long discussion in email with Neal, he acknowledged 6703 * that it should be < than so that we will have flows 6704 * going into probe-rtt (we were seeing cases where that 6705 * did not happen and caused ugly things to occur). We 6706 * have added this agreed upon fix to our code base. 6707 */ 6708 if (rtt < old_rttprop) { 6709 /* Update when we last saw a rtt drop */ 6710 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6711 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6712 } 6713 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6714 match, rsm->r_start, rsm->r_flags); 6715 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6716 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6717 /* 6718 * The RTT-prop moved, reset the target (may be a 6719 * nop for some states). 6720 */ 6721 bbr_set_state_target(bbr, __LINE__); 6722 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6723 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6724 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6725 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6726 /* It went up */ 6727 bbr_check_probe_rtt_limits(bbr, cts); 6728 } 6729 if ((bbr->rc_use_google == 0) && 6730 (match == BBR_RTT_BY_TIMESTAMP)) { 6731 /* 6732 * We don't do b/w update with 6733 * these since they are not really 6734 * reliable. 6735 */ 6736 return; 6737 } 6738 if (bbr->r_ctl.r_app_limited_until && 6739 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6740 /* We are no longer app-limited */ 6741 bbr->r_ctl.r_app_limited_until = 0; 6742 } 6743 if (bbr->rc_use_google) { 6744 bbr_google_measurement(bbr, rsm, rtt, cts); 6745 } else { 6746 bbr_nf_measurement(bbr, rsm, rtt, cts); 6747 } 6748 } 6749 6750 /* 6751 * Convert a timestamp that the main stack 6752 * uses (milliseconds) into one that bbr uses 6753 * (microseconds). Return that converted timestamp. 6754 */ 6755 static uint32_t 6756 bbr_ts_convert(uint32_t cts) { 6757 uint32_t sec, msec; 6758 6759 sec = cts / MS_IN_USEC; 6760 msec = cts - (MS_IN_USEC * sec); 6761 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6762 } 6763 6764 /* 6765 * Return 0 if we did not update the RTT time, return 6766 * 1 if we did. 6767 */ 6768 static int 6769 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6770 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6771 { 6772 int32_t i; 6773 uint32_t t, uts = 0; 6774 6775 if ((rsm->r_flags & BBR_ACKED) || 6776 (rsm->r_flags & BBR_WAS_RENEGED) || 6777 (rsm->r_flags & BBR_RXT_CLEARED)) { 6778 /* Already done */ 6779 return (0); 6780 } 6781 if (rsm->r_rtt_not_allowed) { 6782 /* Not allowed */ 6783 return (0); 6784 } 6785 if (rsm->r_rtr_cnt == 1) { 6786 /* 6787 * Only one transmit. Hopefully the normal case. 6788 */ 6789 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6790 t = cts - rsm->r_tim_lastsent[0]; 6791 else 6792 t = 1; 6793 if ((int)t <= 0) 6794 t = 1; 6795 bbr->r_ctl.rc_last_rtt = t; 6796 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6797 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6798 return (1); 6799 } 6800 /* Convert to usecs */ 6801 if ((bbr_can_use_ts_for_rtt == 1) && 6802 (bbr->rc_use_google == 1) && 6803 (ack_type == BBR_CUM_ACKED) && 6804 (to->to_flags & TOF_TS) && 6805 (to->to_tsecr != 0)) { 6806 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6807 if (t < 1) 6808 t = 1; 6809 t *= MS_IN_USEC; 6810 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6811 BBR_RTT_BY_TIMESTAMP, 6812 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6813 ack_type, to); 6814 return (1); 6815 } 6816 uts = bbr_ts_convert(to->to_tsecr); 6817 if ((to->to_flags & TOF_TS) && 6818 (to->to_tsecr != 0) && 6819 (ack_type == BBR_CUM_ACKED) && 6820 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6821 /* 6822 * Now which timestamp does it match? In this block the ACK 6823 * may be coming from a previous transmission. 6824 */ 6825 uint32_t fudge; 6826 6827 fudge = BBR_TIMER_FUDGE; 6828 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6829 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6830 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6831 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6832 t = cts - rsm->r_tim_lastsent[i]; 6833 else 6834 t = 1; 6835 if ((int)t <= 0) 6836 t = 1; 6837 bbr->r_ctl.rc_last_rtt = t; 6838 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6839 rsm->r_tim_lastsent[i], ack_type, to); 6840 if ((i + 1) < rsm->r_rtr_cnt) { 6841 /* Likely */ 6842 return (0); 6843 } else if (rsm->r_flags & BBR_TLP) { 6844 bbr->rc_tlp_rtx_out = 0; 6845 } 6846 return (1); 6847 } 6848 } 6849 /* Fall through if we can't find a matching timestamp */ 6850 } 6851 /* 6852 * Ok its a SACK block that we retransmitted. or a windows 6853 * machine without timestamps. We can tell nothing from the 6854 * time-stamp since its not there or the time the peer last 6855 * recieved a segment that moved forward its cum-ack point. 6856 * 6857 * Lets look at the last retransmit and see what we can tell 6858 * (with BBR for space we only keep 2 note we have to keep 6859 * at least 2 so the map can not be condensed more). 6860 */ 6861 i = rsm->r_rtr_cnt - 1; 6862 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6863 t = cts - rsm->r_tim_lastsent[i]; 6864 else 6865 goto not_sure; 6866 if (t < bbr->r_ctl.rc_lowest_rtt) { 6867 /* 6868 * We retransmitted and the ack came back in less 6869 * than the smallest rtt we have observed in the 6870 * windowed rtt. We most likey did an improper 6871 * retransmit as outlined in 4.2 Step 3 point 2 in 6872 * the rack-draft. 6873 * 6874 * Use the prior transmission to update all the 6875 * information as long as there is only one prior 6876 * transmission. 6877 */ 6878 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6879 #ifdef BBR_INVARIANTS 6880 if (rsm->r_rtr_cnt == 1) 6881 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6882 #endif 6883 i = rsm->r_rtr_cnt - 2; 6884 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6885 t = cts - rsm->r_tim_lastsent[i]; 6886 else 6887 t = 1; 6888 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6889 rsm->r_tim_lastsent[i], ack_type, to); 6890 return (0); 6891 } else { 6892 /* 6893 * Too many prior transmissions, just 6894 * updated BBR delivered 6895 */ 6896 not_sure: 6897 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6898 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6899 } 6900 } else { 6901 /* 6902 * We retransmitted it and the retransmit did the 6903 * job. 6904 */ 6905 if (rsm->r_flags & BBR_TLP) 6906 bbr->rc_tlp_rtx_out = 0; 6907 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6908 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6909 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6910 else 6911 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6912 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6913 return (1); 6914 } 6915 return (0); 6916 } 6917 6918 /* 6919 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6920 */ 6921 static void 6922 bbr_log_sack_passed(struct tcpcb *tp, 6923 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6924 { 6925 struct bbr_sendmap *nrsm; 6926 6927 nrsm = rsm; 6928 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6929 bbr_head, r_tnext) { 6930 if (nrsm == rsm) { 6931 /* Skip orginal segment he is acked */ 6932 continue; 6933 } 6934 if (nrsm->r_flags & BBR_ACKED) { 6935 /* Skip ack'd segments */ 6936 continue; 6937 } 6938 if (nrsm->r_flags & BBR_SACK_PASSED) { 6939 /* 6940 * We found one that is already marked 6941 * passed, we have been here before and 6942 * so all others below this are marked. 6943 */ 6944 break; 6945 } 6946 BBR_STAT_INC(bbr_sack_passed); 6947 nrsm->r_flags |= BBR_SACK_PASSED; 6948 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6949 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6950 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6951 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6952 nrsm->r_flags |= BBR_MARKED_LOST; 6953 } 6954 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6955 } 6956 } 6957 6958 /* 6959 * Returns the number of bytes that were 6960 * newly ack'd by sack blocks. 6961 */ 6962 static uint32_t 6963 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6964 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6965 { 6966 int32_t times = 0; 6967 uint32_t start, end, changed = 0; 6968 struct bbr_sendmap *rsm, *nrsm; 6969 int32_t used_ref = 1; 6970 uint8_t went_back = 0, went_fwd = 0; 6971 6972 start = sack->start; 6973 end = sack->end; 6974 rsm = *prsm; 6975 if (rsm == NULL) 6976 used_ref = 0; 6977 6978 /* Do we locate the block behind where we last were? */ 6979 if (rsm && SEQ_LT(start, rsm->r_start)) { 6980 went_back = 1; 6981 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 6982 if (SEQ_GEQ(start, rsm->r_start) && 6983 SEQ_LT(start, rsm->r_end)) { 6984 goto do_rest_ofb; 6985 } 6986 } 6987 } 6988 start_at_beginning: 6989 went_fwd = 1; 6990 /* 6991 * Ok lets locate the block where this guy is fwd from rsm (if its 6992 * set) 6993 */ 6994 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 6995 if (SEQ_GEQ(start, rsm->r_start) && 6996 SEQ_LT(start, rsm->r_end)) { 6997 break; 6998 } 6999 } 7000 do_rest_ofb: 7001 if (rsm == NULL) { 7002 /* 7003 * This happens when we get duplicate sack blocks with the 7004 * same end. For example SACK 4: 100 SACK 3: 100 The sort 7005 * will not change there location so we would just start at 7006 * the end of the first one and get lost. 7007 */ 7008 if (tp->t_flags & TF_SENTFIN) { 7009 /* 7010 * Check to see if we have not logged the FIN that 7011 * went out. 7012 */ 7013 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7014 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7015 /* 7016 * Ok we did not get the FIN logged. 7017 */ 7018 nrsm->r_end++; 7019 rsm = nrsm; 7020 goto do_rest_ofb; 7021 } 7022 } 7023 if (times == 1) { 7024 #ifdef BBR_INVARIANTS 7025 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7026 tp, bbr, sack, to, prsm); 7027 #else 7028 goto out; 7029 #endif 7030 } 7031 times++; 7032 BBR_STAT_INC(bbr_sack_proc_restart); 7033 rsm = NULL; 7034 goto start_at_beginning; 7035 } 7036 /* Ok we have an ACK for some piece of rsm */ 7037 if (rsm->r_start != start) { 7038 /* 7039 * Need to split this in two pieces the before and after. 7040 */ 7041 if (bbr_sack_mergable(rsm, start, end)) 7042 nrsm = bbr_alloc_full_limit(bbr); 7043 else 7044 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7045 if (nrsm == NULL) { 7046 /* We could not allocate ignore the sack */ 7047 struct sackblk blk; 7048 7049 blk.start = start; 7050 blk.end = end; 7051 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7052 goto out; 7053 } 7054 bbr_clone_rsm(bbr, nrsm, rsm, start); 7055 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7056 if (rsm->r_in_tmap) { 7057 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7058 nrsm->r_in_tmap = 1; 7059 } 7060 rsm->r_flags &= (~BBR_HAS_FIN); 7061 rsm = nrsm; 7062 } 7063 if (SEQ_GEQ(end, rsm->r_end)) { 7064 /* 7065 * The end of this block is either beyond this guy or right 7066 * at this guy. 7067 */ 7068 if ((rsm->r_flags & BBR_ACKED) == 0) { 7069 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7070 changed += (rsm->r_end - rsm->r_start); 7071 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7072 bbr_log_sack_passed(tp, bbr, rsm); 7073 if (rsm->r_flags & BBR_MARKED_LOST) { 7074 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7075 } 7076 /* Is Reordering occuring? */ 7077 if (rsm->r_flags & BBR_SACK_PASSED) { 7078 BBR_STAT_INC(bbr_reorder_seen); 7079 bbr->r_ctl.rc_reorder_ts = cts; 7080 if (rsm->r_flags & BBR_MARKED_LOST) { 7081 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7082 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7083 /* LT sampling also needs adjustment */ 7084 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7085 } 7086 } 7087 rsm->r_flags |= BBR_ACKED; 7088 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7089 if (rsm->r_in_tmap) { 7090 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7091 rsm->r_in_tmap = 0; 7092 } 7093 } 7094 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7095 if (end == rsm->r_end) { 7096 /* This block only - done */ 7097 goto out; 7098 } 7099 /* There is more not coverend by this rsm move on */ 7100 start = rsm->r_end; 7101 nrsm = TAILQ_NEXT(rsm, r_next); 7102 rsm = nrsm; 7103 times = 0; 7104 goto do_rest_ofb; 7105 } 7106 if (rsm->r_flags & BBR_ACKED) { 7107 /* Been here done that */ 7108 goto out; 7109 } 7110 /* Ok we need to split off this one at the tail */ 7111 if (bbr_sack_mergable(rsm, start, end)) 7112 nrsm = bbr_alloc_full_limit(bbr); 7113 else 7114 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7115 if (nrsm == NULL) { 7116 /* failed XXXrrs what can we do but loose the sack info? */ 7117 struct sackblk blk; 7118 7119 blk.start = start; 7120 blk.end = end; 7121 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7122 goto out; 7123 } 7124 /* Clone it */ 7125 bbr_clone_rsm(bbr, nrsm, rsm, end); 7126 /* The sack block does not cover this guy fully */ 7127 rsm->r_flags &= (~BBR_HAS_FIN); 7128 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7129 if (rsm->r_in_tmap) { 7130 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7131 nrsm->r_in_tmap = 1; 7132 } 7133 nrsm->r_dupack = 0; 7134 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7135 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7136 changed += (rsm->r_end - rsm->r_start); 7137 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7138 bbr_log_sack_passed(tp, bbr, rsm); 7139 /* Is Reordering occuring? */ 7140 if (rsm->r_flags & BBR_MARKED_LOST) { 7141 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7142 } 7143 if (rsm->r_flags & BBR_SACK_PASSED) { 7144 BBR_STAT_INC(bbr_reorder_seen); 7145 bbr->r_ctl.rc_reorder_ts = cts; 7146 if (rsm->r_flags & BBR_MARKED_LOST) { 7147 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7148 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7149 /* LT sampling also needs adjustment */ 7150 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7151 } 7152 } 7153 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7154 rsm->r_flags |= BBR_ACKED; 7155 if (rsm->r_in_tmap) { 7156 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7157 rsm->r_in_tmap = 0; 7158 } 7159 out: 7160 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7161 /* 7162 * Now can we merge this newly acked 7163 * block with either the previous or 7164 * next block? 7165 */ 7166 nrsm = TAILQ_NEXT(rsm, r_next); 7167 if (nrsm && 7168 (nrsm->r_flags & BBR_ACKED)) { 7169 /* yep this and next can be merged */ 7170 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7171 } 7172 /* Now what about the previous? */ 7173 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7174 if (nrsm && 7175 (nrsm->r_flags & BBR_ACKED)) { 7176 /* yep the previous and this can be merged */ 7177 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7178 } 7179 } 7180 if (used_ref == 0) { 7181 BBR_STAT_INC(bbr_sack_proc_all); 7182 } else { 7183 BBR_STAT_INC(bbr_sack_proc_short); 7184 } 7185 if (went_fwd && went_back) { 7186 BBR_STAT_INC(bbr_sack_search_both); 7187 } else if (went_fwd) { 7188 BBR_STAT_INC(bbr_sack_search_fwd); 7189 } else if (went_back) { 7190 BBR_STAT_INC(bbr_sack_search_back); 7191 } 7192 /* Save off where the next seq is */ 7193 if (rsm) 7194 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7195 else 7196 bbr->r_ctl.rc_sacklast = NULL; 7197 *prsm = rsm; 7198 return (changed); 7199 } 7200 7201 static void inline 7202 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7203 { 7204 struct bbr_sendmap *tmap; 7205 7206 BBR_STAT_INC(bbr_reneges_seen); 7207 tmap = NULL; 7208 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7209 /* Its no longer sacked, mark it so */ 7210 uint32_t oflags; 7211 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7212 #ifdef BBR_INVARIANTS 7213 if (rsm->r_in_tmap) { 7214 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7215 bbr, rsm, rsm->r_flags); 7216 } 7217 #endif 7218 oflags = rsm->r_flags; 7219 if (rsm->r_flags & BBR_MARKED_LOST) { 7220 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7221 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7222 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7223 /* LT sampling also needs adjustment */ 7224 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7225 } 7226 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7227 rsm->r_flags |= BBR_WAS_RENEGED; 7228 rsm->r_flags |= BBR_RXT_CLEARED; 7229 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7230 /* Rebuild it into our tmap */ 7231 if (tmap == NULL) { 7232 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7233 tmap = rsm; 7234 } else { 7235 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7236 tmap = rsm; 7237 } 7238 tmap->r_in_tmap = 1; 7239 /* 7240 * XXXrrs Delivered? Should we do anything here? 7241 * 7242 * Of course we don't on a rxt timeout so maybe its ok that 7243 * we don't? 7244 * 7245 * For now lets not. 7246 */ 7247 rsm = TAILQ_NEXT(rsm, r_next); 7248 } 7249 /* 7250 * Now lets possibly clear the sack filter so we start recognizing 7251 * sacks that cover this area. 7252 */ 7253 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7254 } 7255 7256 static void 7257 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7258 { 7259 struct tcp_bbr *bbr; 7260 struct bbr_sendmap *rsm; 7261 uint32_t cts; 7262 7263 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7264 cts = bbr->r_ctl.rc_rcvtime; 7265 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7266 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7267 if ((rsm->r_end - rsm->r_start) <= 1) { 7268 /* Log out the SYN completely */ 7269 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7270 rsm->r_rtr_bytes = 0; 7271 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7272 if (rsm->r_in_tmap) { 7273 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7274 rsm->r_in_tmap = 0; 7275 } 7276 if (bbr->r_ctl.rc_next == rsm) { 7277 /* scoot along the marker */ 7278 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7279 } 7280 if (to != NULL) 7281 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7282 bbr_free(bbr, rsm); 7283 } else { 7284 /* There is more (Fast open)? strip out SYN. */ 7285 rsm->r_flags &= ~BBR_HAS_SYN; 7286 rsm->r_start++; 7287 } 7288 } 7289 } 7290 7291 /* 7292 * Returns the number of bytes that were 7293 * acknowledged by SACK blocks. 7294 */ 7295 7296 static uint32_t 7297 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7298 uint32_t *prev_acked) 7299 { 7300 uint32_t changed, last_seq, entered_recovery = 0; 7301 struct tcp_bbr *bbr; 7302 struct bbr_sendmap *rsm; 7303 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7304 register uint32_t th_ack; 7305 int32_t i, j, k, new_sb, num_sack_blks = 0; 7306 uint32_t cts, acked, ack_point, sack_changed = 0; 7307 uint32_t p_maxseg, maxseg, p_acked = 0; 7308 7309 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7310 if (tcp_get_flags(th) & TH_RST) { 7311 /* We don't log resets */ 7312 return (0); 7313 } 7314 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7315 cts = bbr->r_ctl.rc_rcvtime; 7316 7317 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7318 changed = 0; 7319 maxseg = tp->t_maxseg - bbr->rc_last_options; 7320 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7321 th_ack = th->th_ack; 7322 if (SEQ_GT(th_ack, tp->snd_una)) { 7323 acked = th_ack - tp->snd_una; 7324 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7325 bbr->rc_tp->t_acktime = ticks; 7326 } else 7327 acked = 0; 7328 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7329 /* Only sent here for sack processing */ 7330 goto proc_sack; 7331 } 7332 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7333 changed = th_ack - rsm->r_start; 7334 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7335 /* 7336 * For the SYN incoming case we will not have called 7337 * tcp_output for the sending of the SYN, so there will be 7338 * no map. All other cases should probably be a panic. 7339 */ 7340 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7341 /* 7342 * We have a timestamp that can be used to generate 7343 * an initial RTT. 7344 */ 7345 uint32_t ts, now, rtt; 7346 7347 ts = bbr_ts_convert(to->to_tsecr); 7348 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7349 rtt = now - ts; 7350 if (rtt < 1) 7351 rtt = 1; 7352 bbr_log_type_bbrrttprop(bbr, rtt, 7353 tp->iss, 0, cts, 7354 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7355 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7356 changed = 1; 7357 bbr->r_wanted_output = 1; 7358 goto out; 7359 } 7360 goto proc_sack; 7361 } else if (rsm == NULL) { 7362 goto out; 7363 } 7364 if (changed) { 7365 /* 7366 * The ACK point is advancing to th_ack, we must drop off 7367 * the packets in the rack log and calculate any eligble 7368 * RTT's. 7369 */ 7370 bbr->r_wanted_output = 1; 7371 more: 7372 if (rsm == NULL) { 7373 if (tp->t_flags & TF_SENTFIN) { 7374 /* if we send a FIN we will not hav a map */ 7375 goto proc_sack; 7376 } 7377 #ifdef BBR_INVARIANTS 7378 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7379 tp, 7380 th, tp->t_state, bbr, 7381 tp->snd_una, tp->snd_max, changed); 7382 #endif 7383 goto proc_sack; 7384 } 7385 } 7386 if (SEQ_LT(th_ack, rsm->r_start)) { 7387 /* Huh map is missing this */ 7388 #ifdef BBR_INVARIANTS 7389 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7390 rsm->r_start, 7391 th_ack, tp->t_state, 7392 bbr->r_state, bbr); 7393 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7394 #endif 7395 goto proc_sack; 7396 } else if (th_ack == rsm->r_start) { 7397 /* None here to ack */ 7398 goto proc_sack; 7399 } 7400 /* 7401 * Clear the dup ack counter, it will 7402 * either be freed or if there is some 7403 * remaining we need to start it at zero. 7404 */ 7405 rsm->r_dupack = 0; 7406 /* Now do we consume the whole thing? */ 7407 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7408 /* Its all consumed. */ 7409 uint32_t left; 7410 7411 if (rsm->r_flags & BBR_ACKED) { 7412 /* 7413 * It was acked on the scoreboard -- remove it from 7414 * total 7415 */ 7416 p_acked += (rsm->r_end - rsm->r_start); 7417 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7418 if (bbr->r_ctl.rc_sacked == 0) 7419 bbr->r_ctl.rc_sacklast = NULL; 7420 } else { 7421 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7422 if (rsm->r_flags & BBR_MARKED_LOST) { 7423 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7424 } 7425 if (rsm->r_flags & BBR_SACK_PASSED) { 7426 /* 7427 * There are acked segments ACKED on the 7428 * scoreboard further up. We are seeing 7429 * reordering. 7430 */ 7431 BBR_STAT_INC(bbr_reorder_seen); 7432 bbr->r_ctl.rc_reorder_ts = cts; 7433 if (rsm->r_flags & BBR_MARKED_LOST) { 7434 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7435 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7436 /* LT sampling also needs adjustment */ 7437 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7438 } 7439 } 7440 rsm->r_flags &= ~BBR_MARKED_LOST; 7441 } 7442 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7443 rsm->r_rtr_bytes = 0; 7444 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7445 if (rsm->r_in_tmap) { 7446 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7447 rsm->r_in_tmap = 0; 7448 } 7449 if (bbr->r_ctl.rc_next == rsm) { 7450 /* scoot along the marker */ 7451 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7452 } 7453 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7454 /* Adjust the packet counts */ 7455 left = th_ack - rsm->r_end; 7456 /* Free back to zone */ 7457 bbr_free(bbr, rsm); 7458 if (left) { 7459 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7460 goto more; 7461 } 7462 goto proc_sack; 7463 } 7464 if (rsm->r_flags & BBR_ACKED) { 7465 /* 7466 * It was acked on the scoreboard -- remove it from total 7467 * for the part being cum-acked. 7468 */ 7469 p_acked += (rsm->r_end - rsm->r_start); 7470 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7471 if (bbr->r_ctl.rc_sacked == 0) 7472 bbr->r_ctl.rc_sacklast = NULL; 7473 } else { 7474 /* 7475 * It was acked up to th_ack point for the first time 7476 */ 7477 struct bbr_sendmap lrsm; 7478 7479 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7480 lrsm.r_end = th_ack; 7481 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7482 } 7483 if ((rsm->r_flags & BBR_MARKED_LOST) && 7484 ((rsm->r_flags & BBR_ACKED) == 0)) { 7485 /* 7486 * It was marked lost and partly ack'd now 7487 * for the first time. We lower the rc_lost_bytes 7488 * and still leave it MARKED. 7489 */ 7490 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7491 } 7492 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7493 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7494 rsm->r_rtr_bytes = 0; 7495 /* adjust packet count */ 7496 rsm->r_start = th_ack; 7497 proc_sack: 7498 /* Check for reneging */ 7499 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7500 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7501 /* 7502 * The peer has moved snd_una up to the edge of this send, 7503 * i.e. one that it had previously acked. The only way that 7504 * can be true if the peer threw away data (space issues) 7505 * that it had previously sacked (else it would have given 7506 * us snd_una up to (rsm->r_end). We need to undo the acked 7507 * markings here. 7508 * 7509 * Note we have to look to make sure th_ack is our 7510 * rsm->r_start in case we get an old ack where th_ack is 7511 * behind snd_una. 7512 */ 7513 bbr_peer_reneges(bbr, rsm, th->th_ack); 7514 } 7515 if ((to->to_flags & TOF_SACK) == 0) { 7516 /* We are done nothing left to log */ 7517 goto out; 7518 } 7519 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7520 if (rsm) { 7521 last_seq = rsm->r_end; 7522 } else { 7523 last_seq = tp->snd_max; 7524 } 7525 /* Sack block processing */ 7526 if (SEQ_GT(th_ack, tp->snd_una)) 7527 ack_point = th_ack; 7528 else 7529 ack_point = tp->snd_una; 7530 for (i = 0; i < to->to_nsacks; i++) { 7531 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7532 &sack, sizeof(sack)); 7533 sack.start = ntohl(sack.start); 7534 sack.end = ntohl(sack.end); 7535 if (SEQ_GT(sack.end, sack.start) && 7536 SEQ_GT(sack.start, ack_point) && 7537 SEQ_LT(sack.start, tp->snd_max) && 7538 SEQ_GT(sack.end, ack_point) && 7539 SEQ_LEQ(sack.end, tp->snd_max)) { 7540 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7541 (SEQ_LT(sack.end, last_seq)) && 7542 ((sack.end - sack.start) < (p_maxseg / 8))) { 7543 /* 7544 * Not the last piece and its smaller than 7545 * 1/8th of a p_maxseg. We ignore this. 7546 */ 7547 BBR_STAT_INC(bbr_runt_sacks); 7548 continue; 7549 } 7550 sack_blocks[num_sack_blks] = sack; 7551 num_sack_blks++; 7552 } else if (SEQ_LEQ(sack.start, th_ack) && 7553 SEQ_LEQ(sack.end, th_ack)) { 7554 /* 7555 * Its a D-SACK block. 7556 */ 7557 tcp_record_dsack(tp, sack.start, sack.end, 0); 7558 } 7559 } 7560 if (num_sack_blks == 0) 7561 goto out; 7562 /* 7563 * Sort the SACK blocks so we can update the rack scoreboard with 7564 * just one pass. 7565 */ 7566 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, 7567 num_sack_blks, th->th_ack); 7568 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7569 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7570 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7571 num_sack_blks = new_sb; 7572 if (num_sack_blks < 2) { 7573 goto do_sack_work; 7574 } 7575 /* Sort the sacks */ 7576 for (i = 0; i < num_sack_blks; i++) { 7577 for (j = i + 1; j < num_sack_blks; j++) { 7578 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7579 sack = sack_blocks[i]; 7580 sack_blocks[i] = sack_blocks[j]; 7581 sack_blocks[j] = sack; 7582 } 7583 } 7584 } 7585 /* 7586 * Now are any of the sack block ends the same (yes some 7587 * implememtations send these)? 7588 */ 7589 again: 7590 if (num_sack_blks > 1) { 7591 for (i = 0; i < num_sack_blks; i++) { 7592 for (j = i + 1; j < num_sack_blks; j++) { 7593 if (sack_blocks[i].end == sack_blocks[j].end) { 7594 /* 7595 * Ok these two have the same end we 7596 * want the smallest end and then 7597 * throw away the larger and start 7598 * again. 7599 */ 7600 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7601 /* 7602 * The second block covers 7603 * more area use that 7604 */ 7605 sack_blocks[i].start = sack_blocks[j].start; 7606 } 7607 /* 7608 * Now collapse out the dup-sack and 7609 * lower the count 7610 */ 7611 for (k = (j + 1); k < num_sack_blks; k++) { 7612 sack_blocks[j].start = sack_blocks[k].start; 7613 sack_blocks[j].end = sack_blocks[k].end; 7614 j++; 7615 } 7616 num_sack_blks--; 7617 goto again; 7618 } 7619 } 7620 } 7621 } 7622 do_sack_work: 7623 rsm = bbr->r_ctl.rc_sacklast; 7624 for (i = 0; i < num_sack_blks; i++) { 7625 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7626 if (acked) { 7627 bbr->r_wanted_output = 1; 7628 changed += acked; 7629 sack_changed += acked; 7630 } 7631 } 7632 out: 7633 *prev_acked = p_acked; 7634 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7635 /* 7636 * Ok we have a high probability that we need to go in to 7637 * recovery since we have data sack'd 7638 */ 7639 struct bbr_sendmap *rsm; 7640 7641 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7642 if (rsm) { 7643 /* Enter recovery */ 7644 entered_recovery = 1; 7645 bbr->r_wanted_output = 1; 7646 /* 7647 * When we enter recovery we need to assure we send 7648 * one packet. 7649 */ 7650 if (bbr->r_ctl.rc_resend == NULL) { 7651 bbr->r_ctl.rc_resend = rsm; 7652 } 7653 } 7654 } 7655 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7656 /* 7657 * See if we need to rack-retransmit anything if so set it 7658 * up as the thing to resend assuming something else is not 7659 * already in that position. 7660 */ 7661 if (bbr->r_ctl.rc_resend == NULL) { 7662 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7663 } 7664 } 7665 /* 7666 * We return the amount that changed via sack, this is used by the 7667 * ack-received code to augment what was changed between th_ack <-> 7668 * snd_una. 7669 */ 7670 return (sack_changed); 7671 } 7672 7673 static void 7674 bbr_strike_dupack(struct tcp_bbr *bbr) 7675 { 7676 struct bbr_sendmap *rsm; 7677 7678 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7679 if (rsm && (rsm->r_dupack < 0xff)) { 7680 rsm->r_dupack++; 7681 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7682 bbr->r_wanted_output = 1; 7683 } 7684 } 7685 7686 /* 7687 * Return value of 1, we do not need to call bbr_process_data(). 7688 * return value of 0, bbr_process_data can be called. 7689 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7690 * its unlocked and probably unsafe to touch the TCB. 7691 */ 7692 static int 7693 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7694 struct tcpcb *tp, struct tcpopt *to, 7695 uint32_t tiwin, int32_t tlen, 7696 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7697 { 7698 int32_t ourfinisacked = 0; 7699 int32_t acked_amount; 7700 uint16_t nsegs; 7701 int32_t acked; 7702 uint32_t lost, sack_changed = 0; 7703 struct mbuf *mfree; 7704 struct tcp_bbr *bbr; 7705 uint32_t prev_acked = 0; 7706 7707 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7708 lost = bbr->r_ctl.rc_lost; 7709 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7710 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7711 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7712 bbr->r_wanted_output = 1; 7713 return (1); 7714 } 7715 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7716 /* Process the ack */ 7717 if (bbr->rc_in_persist) 7718 tp->t_rxtshift = 0; 7719 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7720 bbr_strike_dupack(bbr); 7721 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7722 } 7723 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7724 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7725 /* 7726 * Old ack, behind the last one rcv'd or a duplicate ack 7727 * with SACK info. 7728 */ 7729 if (th->th_ack == tp->snd_una) { 7730 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7731 if (bbr->r_state == TCPS_SYN_SENT) { 7732 /* 7733 * Special case on where we sent SYN. When 7734 * the SYN-ACK is processed in syn_sent 7735 * state it bumps the snd_una. This causes 7736 * us to hit here even though we did ack 1 7737 * byte. 7738 * 7739 * Go through the nothing left case so we 7740 * send data. 7741 */ 7742 goto nothing_left; 7743 } 7744 } 7745 return (0); 7746 } 7747 /* 7748 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7749 * something we sent. 7750 */ 7751 if (tp->t_flags & TF_NEEDSYN) { 7752 /* 7753 * T/TCP: Connection was half-synchronized, and our SYN has 7754 * been ACK'd (so connection is now fully synchronized). Go 7755 * to non-starred state, increment snd_una for ACK of SYN, 7756 * and check if we can do window scaling. 7757 */ 7758 tp->t_flags &= ~TF_NEEDSYN; 7759 tp->snd_una++; 7760 /* Do window scaling? */ 7761 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7762 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7763 tp->rcv_scale = tp->request_r_scale; 7764 /* Send window already scaled. */ 7765 } 7766 } 7767 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7768 7769 acked = BYTES_THIS_ACK(tp, th); 7770 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7771 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7772 7773 /* 7774 * If we just performed our first retransmit, and the ACK arrives 7775 * within our recovery window, then it was a mistake to do the 7776 * retransmit in the first place. Recover our original cwnd and 7777 * ssthresh, and proceed to transmit where we left off. 7778 */ 7779 if (tp->t_flags & TF_PREVVALID) { 7780 tp->t_flags &= ~TF_PREVVALID; 7781 if (tp->t_rxtshift == 1 && 7782 (int)(ticks - tp->t_badrxtwin) < 0) 7783 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7784 } 7785 SOCKBUF_LOCK(&so->so_snd); 7786 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7787 tp->snd_wnd -= acked_amount; 7788 mfree = sbcut_locked(&so->so_snd, acked_amount); 7789 /* NB: sowwakeup_locked() does an implicit unlock. */ 7790 sowwakeup_locked(so); 7791 m_freem(mfree); 7792 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7793 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7794 } 7795 tp->snd_una = th->th_ack; 7796 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7797 if (IN_RECOVERY(tp->t_flags)) { 7798 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7799 (SEQ_LT(th->th_ack, tp->snd_max))) { 7800 tcp_bbr_partialack(tp); 7801 } else { 7802 bbr_post_recovery(tp); 7803 } 7804 } 7805 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7806 tp->snd_recover = tp->snd_una; 7807 } 7808 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7809 tp->snd_nxt = tp->snd_max; 7810 } 7811 if (tp->snd_una == tp->snd_max) { 7812 /* Nothing left outstanding */ 7813 nothing_left: 7814 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7815 if (sbavail(&so->so_snd) == 0) 7816 bbr->rc_tp->t_acktime = 0; 7817 if ((sbused(&so->so_snd) == 0) && 7818 (tp->t_flags & TF_SENTFIN)) { 7819 ourfinisacked = 1; 7820 } 7821 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7822 if (bbr->rc_in_persist == 0) { 7823 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7824 } 7825 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7826 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7827 /* 7828 * We invalidate the last ack here since we 7829 * don't want to transfer forward the time 7830 * for our sum's calculations. 7831 */ 7832 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7833 (sbavail(&so->so_snd) == 0) && 7834 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7835 /* 7836 * The socket was gone and the peer sent data, time 7837 * to reset him. 7838 */ 7839 *ret_val = 1; 7840 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7841 /* tcp_close will kill the inp pre-log the Reset */ 7842 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7843 tp = tcp_close(tp); 7844 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7845 BBR_STAT_INC(bbr_dropped_af_data); 7846 return (1); 7847 } 7848 /* Set need output so persist might get set */ 7849 bbr->r_wanted_output = 1; 7850 } 7851 if (ofia) 7852 *ofia = ourfinisacked; 7853 return (0); 7854 } 7855 7856 static void 7857 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7858 { 7859 if (bbr->rc_in_persist == 0) { 7860 bbr_timer_cancel(bbr, __LINE__, cts); 7861 bbr->r_ctl.rc_last_delay_val = 0; 7862 tp->t_rxtshift = 0; 7863 bbr->rc_in_persist = 1; 7864 bbr->r_ctl.rc_went_idle_time = cts; 7865 /* We should be capped when rw went to 0 but just in case */ 7866 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7867 /* Time freezes for the state, so do the accounting now */ 7868 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7869 uint32_t time_in; 7870 7871 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7872 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7873 int32_t idx; 7874 7875 idx = bbr_state_val(bbr); 7876 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7877 } else { 7878 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7879 } 7880 } 7881 bbr->r_ctl.rc_bbr_state_time = cts; 7882 } 7883 } 7884 7885 static void 7886 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7887 { 7888 /* 7889 * Note that if idle time does not exceed our 7890 * threshold, we do nothing continuing the state 7891 * transitions we were last walking through. 7892 */ 7893 if (idle_time >= bbr_idle_restart_threshold) { 7894 if (bbr->rc_use_idle_restart) { 7895 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7896 /* 7897 * Set our target using BBR_UNIT, so 7898 * we increase at a dramatic rate but 7899 * we stop when we get the pipe 7900 * full again for our current b/w estimate. 7901 */ 7902 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7903 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7904 bbr_set_state_target(bbr, __LINE__); 7905 /* Now setup our gains to ramp up */ 7906 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7907 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7908 bbr_log_type_statechange(bbr, cts, __LINE__); 7909 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7910 bbr_substate_change(bbr, cts, __LINE__, 1); 7911 } 7912 } 7913 } 7914 7915 static void 7916 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7917 { 7918 uint32_t idle_time; 7919 7920 if (bbr->rc_in_persist == 0) 7921 return; 7922 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7923 bbr->rc_in_persist = 0; 7924 bbr->rc_hit_state_1 = 0; 7925 bbr->r_ctl.rc_del_time = cts; 7926 /* 7927 * We invalidate the last ack here since we 7928 * don't want to transfer forward the time 7929 * for our sum's calculations. 7930 */ 7931 if (tcp_in_hpts(bbr->rc_inp)) { 7932 tcp_hpts_remove(bbr->rc_inp); 7933 bbr->rc_timer_first = 0; 7934 bbr->r_ctl.rc_hpts_flags = 0; 7935 bbr->r_ctl.rc_last_delay_val = 0; 7936 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7937 bbr->r_agg_early_set = 0; 7938 bbr->r_ctl.rc_agg_early = 0; 7939 } 7940 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7941 if (idle_time >= bbr_rtt_probe_time) { 7942 /* 7943 * This qualifies as a RTT_PROBE session since we drop the 7944 * data outstanding to nothing and waited more than 7945 * bbr_rtt_probe_time. 7946 */ 7947 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7948 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7949 } 7950 tp->t_rxtshift = 0; 7951 /* 7952 * If in probeBW and we have persisted more than an RTT lets do 7953 * special handling. 7954 */ 7955 /* Force a time based epoch */ 7956 bbr_set_epoch(bbr, cts, __LINE__); 7957 /* 7958 * Setup the lost so we don't count anything against the guy 7959 * we have been stuck with during persists. 7960 */ 7961 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7962 /* Time un-freezes for the state */ 7963 bbr->r_ctl.rc_bbr_state_time = cts; 7964 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7965 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7966 /* 7967 * If we are going back to probe-bw 7968 * or probe_rtt, we may need to possibly 7969 * do a fast restart. 7970 */ 7971 bbr_restart_after_idle(bbr, cts, idle_time); 7972 } 7973 } 7974 7975 static void 7976 bbr_collapsed_window(struct tcp_bbr *bbr) 7977 { 7978 /* 7979 * Now we must walk the 7980 * send map and divide the 7981 * ones left stranded. These 7982 * guys can't cause us to abort 7983 * the connection and are really 7984 * "unsent". However if a buggy 7985 * client actually did keep some 7986 * of the data i.e. collapsed the win 7987 * and refused to ack and then opened 7988 * the win and acked that data. We would 7989 * get into an ack war, the simplier 7990 * method then of just pretending we 7991 * did not send those segments something 7992 * won't work. 7993 */ 7994 struct bbr_sendmap *rsm, *nrsm; 7995 tcp_seq max_seq; 7996 uint32_t maxseg; 7997 int can_split = 0; 7998 int fnd = 0; 7999 8000 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 8001 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 8002 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 8003 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 8004 /* Find the first seq past or at maxseq */ 8005 if (rsm->r_flags & BBR_RWND_COLLAPSED) 8006 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8007 if (SEQ_GEQ(max_seq, rsm->r_start) && 8008 SEQ_GEQ(rsm->r_end, max_seq)) { 8009 fnd = 1; 8010 break; 8011 } 8012 } 8013 bbr->rc_has_collapsed = 0; 8014 if (!fnd) { 8015 /* Nothing to do strange */ 8016 return; 8017 } 8018 /* 8019 * Now can we split? 8020 * 8021 * We don't want to split if splitting 8022 * would generate too many small segments 8023 * less we let an attacker fragment our 8024 * send_map and leave us out of memory. 8025 */ 8026 if ((max_seq != rsm->r_start) && 8027 (max_seq != rsm->r_end)){ 8028 /* can we split? */ 8029 int res1, res2; 8030 8031 res1 = max_seq - rsm->r_start; 8032 res2 = rsm->r_end - max_seq; 8033 if ((res1 >= (maxseg/8)) && 8034 (res2 >= (maxseg/8))) { 8035 /* No small pieces here */ 8036 can_split = 1; 8037 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8038 /* We are under the limit */ 8039 can_split = 1; 8040 } 8041 } 8042 /* Ok do we need to split this rsm? */ 8043 if (max_seq == rsm->r_start) { 8044 /* It's this guy no split required */ 8045 nrsm = rsm; 8046 } else if (max_seq == rsm->r_end) { 8047 /* It's the next one no split required. */ 8048 nrsm = TAILQ_NEXT(rsm, r_next); 8049 if (nrsm == NULL) { 8050 /* Huh? */ 8051 return; 8052 } 8053 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8054 /* yep we need to split it */ 8055 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8056 if (nrsm == NULL) { 8057 /* failed XXXrrs what can we do mark the whole? */ 8058 nrsm = rsm; 8059 goto no_split; 8060 } 8061 /* Clone it */ 8062 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8063 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8064 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8065 if (rsm->r_in_tmap) { 8066 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8067 nrsm->r_in_tmap = 1; 8068 } 8069 } else { 8070 /* 8071 * Split not allowed just start here just 8072 * use this guy. 8073 */ 8074 nrsm = rsm; 8075 } 8076 no_split: 8077 BBR_STAT_INC(bbr_collapsed_win); 8078 /* reuse fnd as a count */ 8079 fnd = 0; 8080 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8081 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8082 fnd++; 8083 bbr->rc_has_collapsed = 1; 8084 } 8085 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8086 } 8087 8088 static void 8089 bbr_un_collapse_window(struct tcp_bbr *bbr) 8090 { 8091 struct bbr_sendmap *rsm; 8092 int cleared = 0; 8093 8094 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8095 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8096 /* Clear the flag */ 8097 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8098 cleared++; 8099 } else 8100 break; 8101 } 8102 bbr_log_type_rwnd_collapse(bbr, 8103 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8104 bbr->rc_has_collapsed = 0; 8105 } 8106 8107 /* 8108 * Return value of 1, the TCB is unlocked and most 8109 * likely gone, return value of 0, the TCB is still 8110 * locked. 8111 */ 8112 static int 8113 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8114 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8115 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8116 { 8117 /* 8118 * Update window information. Don't look at window if no ACK: TAC's 8119 * send garbage on first SYN. 8120 */ 8121 uint16_t nsegs; 8122 int32_t tfo_syn; 8123 struct tcp_bbr *bbr; 8124 8125 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8126 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8127 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8128 if ((thflags & TH_ACK) && 8129 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8130 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8131 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8132 /* keep track of pure window updates */ 8133 if (tlen == 0 && 8134 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8135 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8136 tp->snd_wnd = tiwin; 8137 tp->snd_wl1 = th->th_seq; 8138 tp->snd_wl2 = th->th_ack; 8139 if (tp->snd_wnd > tp->max_sndwnd) 8140 tp->max_sndwnd = tp->snd_wnd; 8141 bbr->r_wanted_output = 1; 8142 } else if (thflags & TH_ACK) { 8143 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8144 tp->snd_wnd = tiwin; 8145 tp->snd_wl1 = th->th_seq; 8146 tp->snd_wl2 = th->th_ack; 8147 } 8148 } 8149 if (tp->snd_wnd < ctf_outstanding(tp)) 8150 /* The peer collapsed its window on us */ 8151 bbr_collapsed_window(bbr); 8152 else if (bbr->rc_has_collapsed) 8153 bbr_un_collapse_window(bbr); 8154 /* Was persist timer active and now we have window space? */ 8155 if ((bbr->rc_in_persist != 0) && 8156 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8157 bbr_minseg(bbr)))) { 8158 /* 8159 * Make the rate persist at end of persist mode if idle long 8160 * enough 8161 */ 8162 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8163 8164 /* Make sure we output to start the timer */ 8165 bbr->r_wanted_output = 1; 8166 } 8167 /* Do we need to enter persist? */ 8168 if ((bbr->rc_in_persist == 0) && 8169 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8170 TCPS_HAVEESTABLISHED(tp->t_state) && 8171 (tp->snd_max == tp->snd_una) && 8172 sbavail(&so->so_snd) && 8173 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8174 /* No send window.. we must enter persist */ 8175 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8176 } 8177 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8178 m_freem(m); 8179 return (0); 8180 } 8181 /* 8182 * We don't support urgent data but 8183 * drag along the up just to make sure 8184 * if there is a stack switch no one 8185 * is surprised. 8186 */ 8187 tp->rcv_up = tp->rcv_nxt; 8188 8189 /* 8190 * Process the segment text, merging it into the TCP sequencing 8191 * queue, and arranging for acknowledgment of receipt if necessary. 8192 * This process logically involves adjusting tp->rcv_wnd as data is 8193 * presented to the user (this happens in tcp_usrreq.c, case 8194 * PRU_RCVD). If a FIN has already been received on this connection 8195 * then we just ignore the text. 8196 */ 8197 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8198 IS_FASTOPEN(tp->t_flags)); 8199 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8200 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8201 tcp_seq save_start = th->th_seq; 8202 tcp_seq save_rnxt = tp->rcv_nxt; 8203 int save_tlen = tlen; 8204 8205 m_adj(m, drop_hdrlen); /* delayed header drop */ 8206 /* 8207 * Insert segment which includes th into TCP reassembly 8208 * queue with control block tp. Set thflags to whether 8209 * reassembly now includes a segment with FIN. This handles 8210 * the common case inline (segment is the next to be 8211 * received on an established connection, and the queue is 8212 * empty), avoiding linkage into and removal from the queue 8213 * and repetition of various conversions. Set DELACK for 8214 * segments received in order, but ack immediately when 8215 * segments are out of order (so fast retransmit can work). 8216 */ 8217 if (th->th_seq == tp->rcv_nxt && 8218 SEGQ_EMPTY(tp) && 8219 (TCPS_HAVEESTABLISHED(tp->t_state) || 8220 tfo_syn)) { 8221 #ifdef NETFLIX_SB_LIMITS 8222 u_int mcnt, appended; 8223 8224 if (so->so_rcv.sb_shlim) { 8225 mcnt = m_memcnt(m); 8226 appended = 0; 8227 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8228 CFO_NOSLEEP, NULL) == false) { 8229 counter_u64_add(tcp_sb_shlim_fails, 1); 8230 m_freem(m); 8231 return (0); 8232 } 8233 } 8234 8235 #endif 8236 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8237 bbr->bbr_segs_rcvd += max(1, nsegs); 8238 tp->t_flags |= TF_DELACK; 8239 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8240 } else { 8241 bbr->r_wanted_output = 1; 8242 tp->t_flags |= TF_ACKNOW; 8243 } 8244 tp->rcv_nxt += tlen; 8245 if (tlen && 8246 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8247 (tp->t_fbyte_in == 0)) { 8248 tp->t_fbyte_in = ticks; 8249 if (tp->t_fbyte_in == 0) 8250 tp->t_fbyte_in = 1; 8251 if (tp->t_fbyte_out && tp->t_fbyte_in) 8252 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8253 } 8254 thflags = tcp_get_flags(th) & TH_FIN; 8255 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8256 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8257 SOCKBUF_LOCK(&so->so_rcv); 8258 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8259 m_freem(m); 8260 else 8261 #ifdef NETFLIX_SB_LIMITS 8262 appended = 8263 #endif 8264 sbappendstream_locked(&so->so_rcv, m, 0); 8265 /* NB: sorwakeup_locked() does an implicit unlock. */ 8266 sorwakeup_locked(so); 8267 #ifdef NETFLIX_SB_LIMITS 8268 if (so->so_rcv.sb_shlim && appended != mcnt) 8269 counter_fo_release(so->so_rcv.sb_shlim, 8270 mcnt - appended); 8271 #endif 8272 8273 } else { 8274 /* 8275 * XXX: Due to the header drop above "th" is 8276 * theoretically invalid by now. Fortunately 8277 * m_adj() doesn't actually frees any mbufs when 8278 * trimming from the head. 8279 */ 8280 tcp_seq temp = save_start; 8281 8282 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8283 tp->t_flags |= TF_ACKNOW; 8284 if (tp->t_flags & TF_WAKESOR) { 8285 tp->t_flags &= ~TF_WAKESOR; 8286 /* NB: sorwakeup_locked() does an implicit unlock. */ 8287 sorwakeup_locked(so); 8288 } 8289 } 8290 if ((tp->t_flags & TF_SACK_PERMIT) && 8291 (save_tlen > 0) && 8292 TCPS_HAVEESTABLISHED(tp->t_state)) { 8293 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8294 /* 8295 * DSACK actually handled in the fastpath 8296 * above. 8297 */ 8298 tcp_update_sack_list(tp, save_start, 8299 save_start + save_tlen); 8300 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8301 if ((tp->rcv_numsacks >= 1) && 8302 (tp->sackblks[0].end == save_start)) { 8303 /* 8304 * Partial overlap, recorded at todrop 8305 * above. 8306 */ 8307 tcp_update_sack_list(tp, 8308 tp->sackblks[0].start, 8309 tp->sackblks[0].end); 8310 } else { 8311 tcp_update_dsack_list(tp, save_start, 8312 save_start + save_tlen); 8313 } 8314 } else if (tlen >= save_tlen) { 8315 /* Update of sackblks. */ 8316 tcp_update_dsack_list(tp, save_start, 8317 save_start + save_tlen); 8318 } else if (tlen > 0) { 8319 tcp_update_dsack_list(tp, save_start, 8320 save_start + tlen); 8321 } 8322 } 8323 } else { 8324 m_freem(m); 8325 thflags &= ~TH_FIN; 8326 } 8327 8328 /* 8329 * If FIN is received ACK the FIN and let the user know that the 8330 * connection is closing. 8331 */ 8332 if (thflags & TH_FIN) { 8333 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8334 /* The socket upcall is handled by socantrcvmore. */ 8335 socantrcvmore(so); 8336 /* 8337 * If connection is half-synchronized (ie NEEDSYN 8338 * flag on) then delay ACK, so it may be piggybacked 8339 * when SYN is sent. Otherwise, since we received a 8340 * FIN then no more input can be expected, send ACK 8341 * now. 8342 */ 8343 if (tp->t_flags & TF_NEEDSYN) { 8344 tp->t_flags |= TF_DELACK; 8345 bbr_timer_cancel(bbr, 8346 __LINE__, bbr->r_ctl.rc_rcvtime); 8347 } else { 8348 tp->t_flags |= TF_ACKNOW; 8349 } 8350 tp->rcv_nxt++; 8351 } 8352 switch (tp->t_state) { 8353 /* 8354 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8355 * CLOSE_WAIT state. 8356 */ 8357 case TCPS_SYN_RECEIVED: 8358 tp->t_starttime = ticks; 8359 /* FALLTHROUGH */ 8360 case TCPS_ESTABLISHED: 8361 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8362 break; 8363 8364 /* 8365 * If still in FIN_WAIT_1 STATE FIN has not been 8366 * acked so enter the CLOSING state. 8367 */ 8368 case TCPS_FIN_WAIT_1: 8369 tcp_state_change(tp, TCPS_CLOSING); 8370 break; 8371 8372 /* 8373 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8374 * starting the time-wait timer, turning off the 8375 * other standard timers. 8376 */ 8377 case TCPS_FIN_WAIT_2: 8378 bbr->rc_timer_first = 1; 8379 bbr_timer_cancel(bbr, 8380 __LINE__, bbr->r_ctl.rc_rcvtime); 8381 tcp_twstart(tp); 8382 return (1); 8383 } 8384 } 8385 /* 8386 * Return any desired output. 8387 */ 8388 if ((tp->t_flags & TF_ACKNOW) || 8389 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8390 bbr->r_wanted_output = 1; 8391 } 8392 return (0); 8393 } 8394 8395 /* 8396 * Here nothing is really faster, its just that we 8397 * have broken out the fast-data path also just like 8398 * the fast-ack. Return 1 if we processed the packet 8399 * return 0 if you need to take the "slow-path". 8400 */ 8401 static int 8402 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8403 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8404 uint32_t tiwin, int32_t nxt_pkt) 8405 { 8406 uint16_t nsegs; 8407 int32_t newsize = 0; /* automatic sockbuf scaling */ 8408 struct tcp_bbr *bbr; 8409 #ifdef NETFLIX_SB_LIMITS 8410 u_int mcnt, appended; 8411 #endif 8412 #ifdef TCPDEBUG 8413 /* 8414 * The size of tcp_saveipgen must be the size of the max ip header, 8415 * now IPv6. 8416 */ 8417 u_char tcp_saveipgen[IP6_HDR_LEN]; 8418 struct tcphdr tcp_savetcp; 8419 short ostate = 0; 8420 8421 #endif 8422 /* On the hpts and we would have called output */ 8423 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8424 8425 /* 8426 * If last ACK falls within this segment's sequence numbers, record 8427 * the timestamp. NOTE that the test is modified according to the 8428 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8429 */ 8430 if (bbr->r_ctl.rc_resend != NULL) { 8431 return (0); 8432 } 8433 if (tiwin && tiwin != tp->snd_wnd) { 8434 return (0); 8435 } 8436 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8437 return (0); 8438 } 8439 if (__predict_false((to->to_flags & TOF_TS) && 8440 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8441 return (0); 8442 } 8443 if (__predict_false((th->th_ack != tp->snd_una))) { 8444 return (0); 8445 } 8446 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8447 return (0); 8448 } 8449 if ((to->to_flags & TOF_TS) != 0 && 8450 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8451 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8452 tp->ts_recent = to->to_tsval; 8453 } 8454 /* 8455 * This is a pure, in-sequence data packet with nothing on the 8456 * reassembly queue and we have enough buffer space to take it. 8457 */ 8458 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8459 8460 #ifdef NETFLIX_SB_LIMITS 8461 if (so->so_rcv.sb_shlim) { 8462 mcnt = m_memcnt(m); 8463 appended = 0; 8464 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8465 CFO_NOSLEEP, NULL) == false) { 8466 counter_u64_add(tcp_sb_shlim_fails, 1); 8467 m_freem(m); 8468 return (1); 8469 } 8470 } 8471 #endif 8472 /* Clean receiver SACK report if present */ 8473 if (tp->rcv_numsacks) 8474 tcp_clean_sackreport(tp); 8475 KMOD_TCPSTAT_INC(tcps_preddat); 8476 tp->rcv_nxt += tlen; 8477 if (tlen && 8478 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8479 (tp->t_fbyte_in == 0)) { 8480 tp->t_fbyte_in = ticks; 8481 if (tp->t_fbyte_in == 0) 8482 tp->t_fbyte_in = 1; 8483 if (tp->t_fbyte_out && tp->t_fbyte_in) 8484 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8485 } 8486 /* 8487 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8488 */ 8489 tp->snd_wl1 = th->th_seq; 8490 /* 8491 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8492 */ 8493 tp->rcv_up = tp->rcv_nxt; 8494 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8495 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8496 #ifdef TCPDEBUG 8497 if (so->so_options & SO_DEBUG) 8498 tcp_trace(TA_INPUT, ostate, tp, 8499 (void *)tcp_saveipgen, &tcp_savetcp, 0); 8500 #endif 8501 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8502 8503 /* Add data to socket buffer. */ 8504 SOCKBUF_LOCK(&so->so_rcv); 8505 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8506 m_freem(m); 8507 } else { 8508 /* 8509 * Set new socket buffer size. Give up when limit is 8510 * reached. 8511 */ 8512 if (newsize) 8513 if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) 8514 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8515 m_adj(m, drop_hdrlen); /* delayed header drop */ 8516 8517 #ifdef NETFLIX_SB_LIMITS 8518 appended = 8519 #endif 8520 sbappendstream_locked(&so->so_rcv, m, 0); 8521 ctf_calc_rwin(so, tp); 8522 } 8523 /* NB: sorwakeup_locked() does an implicit unlock. */ 8524 sorwakeup_locked(so); 8525 #ifdef NETFLIX_SB_LIMITS 8526 if (so->so_rcv.sb_shlim && mcnt != appended) 8527 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8528 #endif 8529 if (DELAY_ACK(tp, bbr, nsegs)) { 8530 bbr->bbr_segs_rcvd += max(1, nsegs); 8531 tp->t_flags |= TF_DELACK; 8532 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8533 } else { 8534 bbr->r_wanted_output = 1; 8535 tp->t_flags |= TF_ACKNOW; 8536 } 8537 return (1); 8538 } 8539 8540 /* 8541 * This subfunction is used to try to highly optimize the 8542 * fast path. We again allow window updates that are 8543 * in sequence to remain in the fast-path. We also add 8544 * in the __predict's to attempt to help the compiler. 8545 * Note that if we return a 0, then we can *not* process 8546 * it and the caller should push the packet into the 8547 * slow-path. If we return 1, then all is well and 8548 * the packet is fully processed. 8549 */ 8550 static int 8551 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8552 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8553 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8554 { 8555 int32_t acked; 8556 uint16_t nsegs; 8557 uint32_t sack_changed; 8558 #ifdef TCPDEBUG 8559 /* 8560 * The size of tcp_saveipgen must be the size of the max ip header, 8561 * now IPv6. 8562 */ 8563 u_char tcp_saveipgen[IP6_HDR_LEN]; 8564 struct tcphdr tcp_savetcp; 8565 short ostate = 0; 8566 8567 #endif 8568 uint32_t prev_acked = 0; 8569 struct tcp_bbr *bbr; 8570 8571 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8572 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8573 return (0); 8574 } 8575 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8576 /* Above what we have sent? */ 8577 return (0); 8578 } 8579 if (__predict_false(tiwin == 0)) { 8580 /* zero window */ 8581 return (0); 8582 } 8583 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8584 /* We need a SYN or a FIN, unlikely.. */ 8585 return (0); 8586 } 8587 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8588 /* Timestamp is behind .. old ack with seq wrap? */ 8589 return (0); 8590 } 8591 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8592 /* Still recovering */ 8593 return (0); 8594 } 8595 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8596 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8597 /* We are retransmitting */ 8598 return (0); 8599 } 8600 if (__predict_false(bbr->rc_in_persist != 0)) { 8601 /* In persist mode */ 8602 return (0); 8603 } 8604 if (bbr->r_ctl.rc_sacked) { 8605 /* We have sack holes on our scoreboard */ 8606 return (0); 8607 } 8608 /* Ok if we reach here, we can process a fast-ack */ 8609 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8610 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8611 /* 8612 * We never detect loss in fast ack [we can't 8613 * have a sack and can't be in recovery so 8614 * we always pass 0 (nothing detected)]. 8615 */ 8616 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8617 /* Did the window get updated? */ 8618 if (tiwin != tp->snd_wnd) { 8619 tp->snd_wnd = tiwin; 8620 tp->snd_wl1 = th->th_seq; 8621 if (tp->snd_wnd > tp->max_sndwnd) 8622 tp->max_sndwnd = tp->snd_wnd; 8623 } 8624 /* Do we need to exit persists? */ 8625 if ((bbr->rc_in_persist != 0) && 8626 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8627 bbr_minseg(bbr)))) { 8628 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8629 bbr->r_wanted_output = 1; 8630 } 8631 /* Do we need to enter persists? */ 8632 if ((bbr->rc_in_persist == 0) && 8633 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8634 TCPS_HAVEESTABLISHED(tp->t_state) && 8635 (tp->snd_max == tp->snd_una) && 8636 sbavail(&so->so_snd) && 8637 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8638 /* No send window.. we must enter persist */ 8639 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8640 } 8641 /* 8642 * If last ACK falls within this segment's sequence numbers, record 8643 * the timestamp. NOTE that the test is modified according to the 8644 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8645 */ 8646 if ((to->to_flags & TOF_TS) != 0 && 8647 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8648 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8649 tp->ts_recent = to->to_tsval; 8650 } 8651 /* 8652 * This is a pure ack for outstanding data. 8653 */ 8654 KMOD_TCPSTAT_INC(tcps_predack); 8655 8656 /* 8657 * "bad retransmit" recovery. 8658 */ 8659 if (tp->t_flags & TF_PREVVALID) { 8660 tp->t_flags &= ~TF_PREVVALID; 8661 if (tp->t_rxtshift == 1 && 8662 (int)(ticks - tp->t_badrxtwin) < 0) 8663 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8664 } 8665 /* 8666 * Recalculate the transmit timer / rtt. 8667 * 8668 * Some boxes send broken timestamp replies during the SYN+ACK 8669 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8670 * and blow up the retransmit timer. 8671 */ 8672 acked = BYTES_THIS_ACK(tp, th); 8673 8674 #ifdef TCP_HHOOK 8675 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8676 hhook_run_tcp_est_in(tp, th, to); 8677 #endif 8678 8679 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8680 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8681 sbdrop(&so->so_snd, acked); 8682 8683 if (SEQ_GT(th->th_ack, tp->snd_una)) 8684 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8685 tp->snd_una = th->th_ack; 8686 if (tp->snd_wnd < ctf_outstanding(tp)) 8687 /* The peer collapsed its window on us */ 8688 bbr_collapsed_window(bbr); 8689 else if (bbr->rc_has_collapsed) 8690 bbr_un_collapse_window(bbr); 8691 8692 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8693 tp->snd_recover = tp->snd_una; 8694 } 8695 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8696 /* 8697 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8698 */ 8699 tp->snd_wl2 = th->th_ack; 8700 m_freem(m); 8701 /* 8702 * If all outstanding data are acked, stop retransmit timer, 8703 * otherwise restart timer using current (possibly backed-off) 8704 * value. If process is waiting for space, wakeup/selwakeup/signal. 8705 * If data are ready to send, let tcp_output decide between more 8706 * output or persist. 8707 */ 8708 #ifdef TCPDEBUG 8709 if (so->so_options & SO_DEBUG) 8710 tcp_trace(TA_INPUT, ostate, tp, 8711 (void *)tcp_saveipgen, 8712 &tcp_savetcp, 0); 8713 #endif 8714 /* Wake up the socket if we have room to write more */ 8715 sowwakeup(so); 8716 if (tp->snd_una == tp->snd_max) { 8717 /* Nothing left outstanding */ 8718 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8719 if (sbavail(&so->so_snd) == 0) 8720 bbr->rc_tp->t_acktime = 0; 8721 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8722 if (bbr->rc_in_persist == 0) { 8723 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8724 } 8725 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8726 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8727 /* 8728 * We invalidate the last ack here since we 8729 * don't want to transfer forward the time 8730 * for our sum's calculations. 8731 */ 8732 bbr->r_wanted_output = 1; 8733 } 8734 if (sbavail(&so->so_snd)) { 8735 bbr->r_wanted_output = 1; 8736 } 8737 return (1); 8738 } 8739 8740 /* 8741 * Return value of 1, the TCB is unlocked and most 8742 * likely gone, return value of 0, the TCB is still 8743 * locked. 8744 */ 8745 static int 8746 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8747 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8748 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8749 { 8750 int32_t todrop; 8751 int32_t ourfinisacked = 0; 8752 struct tcp_bbr *bbr; 8753 int32_t ret_val = 0; 8754 8755 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8756 8757 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8758 ctf_calc_rwin(so, tp); 8759 /* 8760 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8761 * SYN, drop the input. if seg contains a RST, then drop the 8762 * connection. if seg does not contain SYN, then drop it. Otherwise 8763 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8764 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8765 * not support ECN so we will not say we are capable. if SYN has 8766 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8767 * segment to be acked (eventually) continue processing rest of 8768 * data/controls, beginning with URG 8769 */ 8770 if ((thflags & TH_ACK) && 8771 (SEQ_LEQ(th->th_ack, tp->iss) || 8772 SEQ_GT(th->th_ack, tp->snd_max))) { 8773 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8774 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8775 return (1); 8776 } 8777 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8778 TCP_PROBE5(connect__refused, NULL, tp, 8779 mtod(m, const char *), tp, th); 8780 tp = tcp_drop(tp, ECONNREFUSED); 8781 ctf_do_drop(m, tp); 8782 return (1); 8783 } 8784 if (thflags & TH_RST) { 8785 ctf_do_drop(m, tp); 8786 return (1); 8787 } 8788 if (!(thflags & TH_SYN)) { 8789 ctf_do_drop(m, tp); 8790 return (1); 8791 } 8792 tp->irs = th->th_seq; 8793 tcp_rcvseqinit(tp); 8794 if (thflags & TH_ACK) { 8795 int tfo_partial = 0; 8796 8797 KMOD_TCPSTAT_INC(tcps_connects); 8798 soisconnected(so); 8799 #ifdef MAC 8800 mac_socketpeer_set_from_mbuf(m, so); 8801 #endif 8802 /* Do window scaling on this connection? */ 8803 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8804 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8805 tp->rcv_scale = tp->request_r_scale; 8806 } 8807 tp->rcv_adv += min(tp->rcv_wnd, 8808 TCP_MAXWIN << tp->rcv_scale); 8809 /* 8810 * If not all the data that was sent in the TFO SYN 8811 * has been acked, resend the remainder right away. 8812 */ 8813 if (IS_FASTOPEN(tp->t_flags) && 8814 (tp->snd_una != tp->snd_max)) { 8815 tp->snd_nxt = th->th_ack; 8816 tfo_partial = 1; 8817 } 8818 /* 8819 * If there's data, delay ACK; if there's also a FIN ACKNOW 8820 * will be turned on later. 8821 */ 8822 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8823 bbr->bbr_segs_rcvd += 1; 8824 tp->t_flags |= TF_DELACK; 8825 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8826 } else { 8827 bbr->r_wanted_output = 1; 8828 tp->t_flags |= TF_ACKNOW; 8829 } 8830 if (SEQ_GT(th->th_ack, tp->iss)) { 8831 /* 8832 * The SYN is acked 8833 * handle it specially. 8834 */ 8835 bbr_log_syn(tp, to); 8836 } 8837 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8838 /* 8839 * We advance snd_una for the 8840 * fast open case. If th_ack is 8841 * acknowledging data beyond 8842 * snd_una we can't just call 8843 * ack-processing since the 8844 * data stream in our send-map 8845 * will start at snd_una + 1 (one 8846 * beyond the SYN). If its just 8847 * equal we don't need to do that 8848 * and there is no send_map. 8849 */ 8850 tp->snd_una++; 8851 } 8852 /* 8853 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8854 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8855 */ 8856 tp->t_starttime = ticks; 8857 if (tp->t_flags & TF_NEEDFIN) { 8858 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8859 tp->t_flags &= ~TF_NEEDFIN; 8860 thflags &= ~TH_SYN; 8861 } else { 8862 tcp_state_change(tp, TCPS_ESTABLISHED); 8863 TCP_PROBE5(connect__established, NULL, tp, 8864 mtod(m, const char *), tp, th); 8865 cc_conn_init(tp); 8866 } 8867 } else { 8868 /* 8869 * Received initial SYN in SYN-SENT[*] state => simultaneous 8870 * open. If segment contains CC option and there is a 8871 * cached CC, apply TAO test. If it succeeds, connection is * 8872 * half-synchronized. Otherwise, do 3-way handshake: 8873 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8874 * there was no CC option, clear cached CC value. 8875 */ 8876 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); 8877 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8878 } 8879 /* 8880 * Advance th->th_seq to correspond to first data byte. If data, 8881 * trim to stay within window, dropping FIN if necessary. 8882 */ 8883 th->th_seq++; 8884 if (tlen > tp->rcv_wnd) { 8885 todrop = tlen - tp->rcv_wnd; 8886 m_adj(m, -todrop); 8887 tlen = tp->rcv_wnd; 8888 thflags &= ~TH_FIN; 8889 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8890 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8891 } 8892 tp->snd_wl1 = th->th_seq - 1; 8893 tp->rcv_up = th->th_seq; 8894 /* 8895 * Client side of transaction: already sent SYN and data. If the 8896 * remote host used T/TCP to validate the SYN, our data will be 8897 * ACK'd; if so, enter normal data segment processing in the middle 8898 * of step 5, ack processing. Otherwise, goto step 6. 8899 */ 8900 if (thflags & TH_ACK) { 8901 if ((to->to_flags & TOF_TS) != 0) { 8902 uint32_t t, rtt; 8903 8904 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8905 if (TSTMP_GEQ(t, to->to_tsecr)) { 8906 rtt = t - to->to_tsecr; 8907 if (rtt == 0) { 8908 rtt = 1; 8909 } 8910 rtt *= MS_IN_USEC; 8911 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8912 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8913 rtt, bbr->r_ctl.rc_rcvtime); 8914 } 8915 } 8916 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8917 return (ret_val); 8918 /* We may have changed to FIN_WAIT_1 above */ 8919 if (tp->t_state == TCPS_FIN_WAIT_1) { 8920 /* 8921 * In FIN_WAIT_1 STATE in addition to the processing 8922 * for the ESTABLISHED state if our FIN is now 8923 * acknowledged then enter FIN_WAIT_2. 8924 */ 8925 if (ourfinisacked) { 8926 /* 8927 * If we can't receive any more data, then 8928 * closing user can proceed. Starting the 8929 * timer is contrary to the specification, 8930 * but if we don't get a FIN we'll hang 8931 * forever. 8932 * 8933 * XXXjl: we should release the tp also, and 8934 * use a compressed state. 8935 */ 8936 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8937 soisdisconnected(so); 8938 tcp_timer_activate(tp, TT_2MSL, 8939 (tcp_fast_finwait2_recycle ? 8940 tcp_finwait2_timeout : 8941 TP_MAXIDLE(tp))); 8942 } 8943 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8944 } 8945 } 8946 } 8947 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8948 tiwin, thflags, nxt_pkt)); 8949 } 8950 8951 /* 8952 * Return value of 1, the TCB is unlocked and most 8953 * likely gone, return value of 0, the TCB is still 8954 * locked. 8955 */ 8956 static int 8957 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8958 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8959 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8960 { 8961 int32_t ourfinisacked = 0; 8962 int32_t ret_val; 8963 struct tcp_bbr *bbr; 8964 8965 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8966 8967 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8968 ctf_calc_rwin(so, tp); 8969 if ((thflags & TH_ACK) && 8970 (SEQ_LEQ(th->th_ack, tp->snd_una) || 8971 SEQ_GT(th->th_ack, tp->snd_max))) { 8972 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8973 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8974 return (1); 8975 } 8976 if (IS_FASTOPEN(tp->t_flags)) { 8977 /* 8978 * When a TFO connection is in SYN_RECEIVED, the only valid 8979 * packets are the initial SYN, a retransmit/copy of the 8980 * initial SYN (possibly with a subset of the original 8981 * data), a valid ACK, a FIN, or a RST. 8982 */ 8983 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 8984 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8985 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8986 return (1); 8987 } else if (thflags & TH_SYN) { 8988 /* non-initial SYN is ignored */ 8989 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 8990 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 8991 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 8992 ctf_do_drop(m, NULL); 8993 return (0); 8994 } 8995 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 8996 ctf_do_drop(m, NULL); 8997 return (0); 8998 } 8999 } 9000 if ((thflags & TH_RST) || 9001 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9002 return (ctf_process_rst(m, th, so, tp)); 9003 /* 9004 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9005 * it's less than ts_recent, drop it. 9006 */ 9007 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9008 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9009 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9010 return (ret_val); 9011 } 9012 /* 9013 * In the SYN-RECEIVED state, validate that the packet belongs to 9014 * this connection before trimming the data to fit the receive 9015 * window. Check the sequence number versus IRS since we know the 9016 * sequence numbers haven't wrapped. This is a partial fix for the 9017 * "LAND" DoS attack. 9018 */ 9019 if (SEQ_LT(th->th_seq, tp->irs)) { 9020 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9021 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9022 return (1); 9023 } 9024 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9025 return (ret_val); 9026 } 9027 /* 9028 * If last ACK falls within this segment's sequence numbers, record 9029 * its timestamp. NOTE: 1) That the test incorporates suggestions 9030 * from the latest proposal of the tcplw@cray.com list (Braden 9031 * 1993/04/26). 2) That updating only on newer timestamps interferes 9032 * with our earlier PAWS tests, so this check should be solely 9033 * predicated on the sequence space of this segment. 3) That we 9034 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9035 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9036 * SEG.Len, This modified check allows us to overcome RFC1323's 9037 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9038 * p.869. In such cases, we can still calculate the RTT correctly 9039 * when RCV.NXT == Last.ACK.Sent. 9040 */ 9041 if ((to->to_flags & TOF_TS) != 0 && 9042 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9043 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9044 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9045 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9046 tp->ts_recent = to->to_tsval; 9047 } 9048 tp->snd_wnd = tiwin; 9049 /* 9050 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9051 * is on (half-synchronized state), then queue data for later 9052 * processing; else drop segment and return. 9053 */ 9054 if ((thflags & TH_ACK) == 0) { 9055 if (IS_FASTOPEN(tp->t_flags)) { 9056 cc_conn_init(tp); 9057 } 9058 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9059 tiwin, thflags, nxt_pkt)); 9060 } 9061 KMOD_TCPSTAT_INC(tcps_connects); 9062 if (tp->t_flags & TF_SONOTCONN) { 9063 tp->t_flags &= ~TF_SONOTCONN; 9064 soisconnected(so); 9065 } 9066 /* Do window scaling? */ 9067 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9068 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9069 tp->rcv_scale = tp->request_r_scale; 9070 } 9071 /* 9072 * ok for the first time in lets see if we can use the ts to figure 9073 * out what the initial RTT was. 9074 */ 9075 if ((to->to_flags & TOF_TS) != 0) { 9076 uint32_t t, rtt; 9077 9078 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9079 if (TSTMP_GEQ(t, to->to_tsecr)) { 9080 rtt = t - to->to_tsecr; 9081 if (rtt == 0) { 9082 rtt = 1; 9083 } 9084 rtt *= MS_IN_USEC; 9085 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9086 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9087 } 9088 } 9089 /* Drop off any SYN in the send map (probably not there) */ 9090 if (thflags & TH_ACK) 9091 bbr_log_syn(tp, to); 9092 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 9093 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9094 tp->t_tfo_pending = NULL; 9095 } 9096 /* 9097 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9098 * FIN-WAIT-1 9099 */ 9100 tp->t_starttime = ticks; 9101 if (tp->t_flags & TF_NEEDFIN) { 9102 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9103 tp->t_flags &= ~TF_NEEDFIN; 9104 } else { 9105 tcp_state_change(tp, TCPS_ESTABLISHED); 9106 TCP_PROBE5(accept__established, NULL, tp, 9107 mtod(m, const char *), tp, th); 9108 /* 9109 * TFO connections call cc_conn_init() during SYN 9110 * processing. Calling it again here for such connections 9111 * is not harmless as it would undo the snd_cwnd reduction 9112 * that occurs when a TFO SYN|ACK is retransmitted. 9113 */ 9114 if (!IS_FASTOPEN(tp->t_flags)) 9115 cc_conn_init(tp); 9116 } 9117 /* 9118 * Account for the ACK of our SYN prior to 9119 * regular ACK processing below, except for 9120 * simultaneous SYN, which is handled later. 9121 */ 9122 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9123 tp->snd_una++; 9124 /* 9125 * If segment contains data or ACK, will call tcp_reass() later; if 9126 * not, do so now to pass queued data to user. 9127 */ 9128 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9129 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9130 (struct mbuf *)0); 9131 if (tp->t_flags & TF_WAKESOR) { 9132 tp->t_flags &= ~TF_WAKESOR; 9133 /* NB: sorwakeup_locked() does an implicit unlock. */ 9134 sorwakeup_locked(so); 9135 } 9136 } 9137 tp->snd_wl1 = th->th_seq - 1; 9138 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9139 return (ret_val); 9140 } 9141 if (tp->t_state == TCPS_FIN_WAIT_1) { 9142 /* We could have went to FIN_WAIT_1 (or EST) above */ 9143 /* 9144 * In FIN_WAIT_1 STATE in addition to the processing for the 9145 * ESTABLISHED state if our FIN is now acknowledged then 9146 * enter FIN_WAIT_2. 9147 */ 9148 if (ourfinisacked) { 9149 /* 9150 * If we can't receive any more data, then closing 9151 * user can proceed. Starting the timer is contrary 9152 * to the specification, but if we don't get a FIN 9153 * we'll hang forever. 9154 * 9155 * XXXjl: we should release the tp also, and use a 9156 * compressed state. 9157 */ 9158 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9159 soisdisconnected(so); 9160 tcp_timer_activate(tp, TT_2MSL, 9161 (tcp_fast_finwait2_recycle ? 9162 tcp_finwait2_timeout : 9163 TP_MAXIDLE(tp))); 9164 } 9165 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9166 } 9167 } 9168 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9169 tiwin, thflags, nxt_pkt)); 9170 } 9171 9172 /* 9173 * Return value of 1, the TCB is unlocked and most 9174 * likely gone, return value of 0, the TCB is still 9175 * locked. 9176 */ 9177 static int 9178 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9179 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9180 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9181 { 9182 struct tcp_bbr *bbr; 9183 int32_t ret_val; 9184 9185 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9186 9187 /* 9188 * Header prediction: check for the two common cases of a 9189 * uni-directional data xfer. If the packet has no control flags, 9190 * is in-sequence, the window didn't change and we're not 9191 * retransmitting, it's a candidate. If the length is zero and the 9192 * ack moved forward, we're the sender side of the xfer. Just free 9193 * the data acked & wake any higher level process that was blocked 9194 * waiting for space. If the length is non-zero and the ack didn't 9195 * move, we're the receiver side. If we're getting packets in-order 9196 * (the reassembly queue is empty), add the data toc The socket 9197 * buffer and note that we need a delayed ack. Make sure that the 9198 * hidden state-flags are also off. Since we check for 9199 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9200 */ 9201 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9202 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9203 /* 9204 * If we have delived under 4 segments increase the initial 9205 * window if raised by the peer. We use this to determine 9206 * dynamic and static rwnd's at the end of a connection. 9207 */ 9208 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9209 } 9210 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9211 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9212 __predict_true(SEGQ_EMPTY(tp)) && 9213 __predict_true(th->th_seq == tp->rcv_nxt)) { 9214 if (tlen == 0) { 9215 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9216 tiwin, nxt_pkt, iptos)) { 9217 return (0); 9218 } 9219 } else { 9220 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9221 tiwin, nxt_pkt)) { 9222 return (0); 9223 } 9224 } 9225 } 9226 ctf_calc_rwin(so, tp); 9227 9228 if ((thflags & TH_RST) || 9229 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9230 return (ctf_process_rst(m, th, so, tp)); 9231 /* 9232 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9233 * synchronized state. 9234 */ 9235 if (thflags & TH_SYN) { 9236 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9237 return (ret_val); 9238 } 9239 /* 9240 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9241 * it's less than ts_recent, drop it. 9242 */ 9243 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9244 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9245 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9246 return (ret_val); 9247 } 9248 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9249 return (ret_val); 9250 } 9251 /* 9252 * If last ACK falls within this segment's sequence numbers, record 9253 * its timestamp. NOTE: 1) That the test incorporates suggestions 9254 * from the latest proposal of the tcplw@cray.com list (Braden 9255 * 1993/04/26). 2) That updating only on newer timestamps interferes 9256 * with our earlier PAWS tests, so this check should be solely 9257 * predicated on the sequence space of this segment. 3) That we 9258 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9259 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9260 * SEG.Len, This modified check allows us to overcome RFC1323's 9261 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9262 * p.869. In such cases, we can still calculate the RTT correctly 9263 * when RCV.NXT == Last.ACK.Sent. 9264 */ 9265 if ((to->to_flags & TOF_TS) != 0 && 9266 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9267 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9268 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9269 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9270 tp->ts_recent = to->to_tsval; 9271 } 9272 /* 9273 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9274 * is on (half-synchronized state), then queue data for later 9275 * processing; else drop segment and return. 9276 */ 9277 if ((thflags & TH_ACK) == 0) { 9278 if (tp->t_flags & TF_NEEDSYN) { 9279 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9280 tiwin, thflags, nxt_pkt)); 9281 } else if (tp->t_flags & TF_ACKNOW) { 9282 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9283 bbr->r_wanted_output = 1; 9284 return (ret_val); 9285 } else { 9286 ctf_do_drop(m, NULL); 9287 return (0); 9288 } 9289 } 9290 /* 9291 * Ack processing. 9292 */ 9293 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9294 return (ret_val); 9295 } 9296 if (sbavail(&so->so_snd)) { 9297 if (ctf_progress_timeout_check(tp, true)) { 9298 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9299 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9300 return (1); 9301 } 9302 } 9303 /* State changes only happen in bbr_process_data() */ 9304 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9305 tiwin, thflags, nxt_pkt)); 9306 } 9307 9308 /* 9309 * Return value of 1, the TCB is unlocked and most 9310 * likely gone, return value of 0, the TCB is still 9311 * locked. 9312 */ 9313 static int 9314 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9315 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9316 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9317 { 9318 struct tcp_bbr *bbr; 9319 int32_t ret_val; 9320 9321 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9322 9323 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9324 ctf_calc_rwin(so, tp); 9325 if ((thflags & TH_RST) || 9326 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9327 return (ctf_process_rst(m, th, so, tp)); 9328 /* 9329 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9330 * synchronized state. 9331 */ 9332 if (thflags & TH_SYN) { 9333 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9334 return (ret_val); 9335 } 9336 /* 9337 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9338 * it's less than ts_recent, drop it. 9339 */ 9340 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9341 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9342 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9343 return (ret_val); 9344 } 9345 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9346 return (ret_val); 9347 } 9348 /* 9349 * If last ACK falls within this segment's sequence numbers, record 9350 * its timestamp. NOTE: 1) That the test incorporates suggestions 9351 * from the latest proposal of the tcplw@cray.com list (Braden 9352 * 1993/04/26). 2) That updating only on newer timestamps interferes 9353 * with our earlier PAWS tests, so this check should be solely 9354 * predicated on the sequence space of this segment. 3) That we 9355 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9356 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9357 * SEG.Len, This modified check allows us to overcome RFC1323's 9358 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9359 * p.869. In such cases, we can still calculate the RTT correctly 9360 * when RCV.NXT == Last.ACK.Sent. 9361 */ 9362 if ((to->to_flags & TOF_TS) != 0 && 9363 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9364 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9365 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9366 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9367 tp->ts_recent = to->to_tsval; 9368 } 9369 /* 9370 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9371 * is on (half-synchronized state), then queue data for later 9372 * processing; else drop segment and return. 9373 */ 9374 if ((thflags & TH_ACK) == 0) { 9375 if (tp->t_flags & TF_NEEDSYN) { 9376 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9377 tiwin, thflags, nxt_pkt)); 9378 } else if (tp->t_flags & TF_ACKNOW) { 9379 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9380 bbr->r_wanted_output = 1; 9381 return (ret_val); 9382 } else { 9383 ctf_do_drop(m, NULL); 9384 return (0); 9385 } 9386 } 9387 /* 9388 * Ack processing. 9389 */ 9390 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9391 return (ret_val); 9392 } 9393 if (sbavail(&so->so_snd)) { 9394 if (ctf_progress_timeout_check(tp, true)) { 9395 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9396 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9397 return (1); 9398 } 9399 } 9400 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9401 tiwin, thflags, nxt_pkt)); 9402 } 9403 9404 static int 9405 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9406 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9407 { 9408 9409 if (bbr->rc_allow_data_af_clo == 0) { 9410 close_now: 9411 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9412 /* tcp_close will kill the inp pre-log the Reset */ 9413 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9414 tp = tcp_close(tp); 9415 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9416 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9417 return (1); 9418 } 9419 if (sbavail(&so->so_snd) == 0) 9420 goto close_now; 9421 /* Ok we allow data that is ignored and a followup reset */ 9422 tp->rcv_nxt = th->th_seq + *tlen; 9423 tp->t_flags2 |= TF2_DROP_AF_DATA; 9424 bbr->r_wanted_output = 1; 9425 *tlen = 0; 9426 return (0); 9427 } 9428 9429 /* 9430 * Return value of 1, the TCB is unlocked and most 9431 * likely gone, return value of 0, the TCB is still 9432 * locked. 9433 */ 9434 static int 9435 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9436 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9437 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9438 { 9439 int32_t ourfinisacked = 0; 9440 int32_t ret_val; 9441 struct tcp_bbr *bbr; 9442 9443 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9444 9445 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9446 ctf_calc_rwin(so, tp); 9447 if ((thflags & TH_RST) || 9448 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9449 return (ctf_process_rst(m, th, so, tp)); 9450 /* 9451 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9452 * synchronized state. 9453 */ 9454 if (thflags & TH_SYN) { 9455 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9456 return (ret_val); 9457 } 9458 /* 9459 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9460 * it's less than ts_recent, drop it. 9461 */ 9462 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9463 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9464 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9465 return (ret_val); 9466 } 9467 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9468 return (ret_val); 9469 } 9470 /* 9471 * If new data are received on a connection after the user processes 9472 * are gone, then RST the other end. 9473 * We call a new function now so we might continue and setup 9474 * to reset at all data being ack'd. 9475 */ 9476 if ((tp->t_flags & TF_CLOSED) && tlen && 9477 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9478 return (1); 9479 /* 9480 * If last ACK falls within this segment's sequence numbers, record 9481 * its timestamp. NOTE: 1) That the test incorporates suggestions 9482 * from the latest proposal of the tcplw@cray.com list (Braden 9483 * 1993/04/26). 2) That updating only on newer timestamps interferes 9484 * with our earlier PAWS tests, so this check should be solely 9485 * predicated on the sequence space of this segment. 3) That we 9486 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9487 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9488 * SEG.Len, This modified check allows us to overcome RFC1323's 9489 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9490 * p.869. In such cases, we can still calculate the RTT correctly 9491 * when RCV.NXT == Last.ACK.Sent. 9492 */ 9493 if ((to->to_flags & TOF_TS) != 0 && 9494 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9495 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9496 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9497 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9498 tp->ts_recent = to->to_tsval; 9499 } 9500 /* 9501 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9502 * is on (half-synchronized state), then queue data for later 9503 * processing; else drop segment and return. 9504 */ 9505 if ((thflags & TH_ACK) == 0) { 9506 if (tp->t_flags & TF_NEEDSYN) { 9507 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9508 tiwin, thflags, nxt_pkt)); 9509 } else if (tp->t_flags & TF_ACKNOW) { 9510 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9511 bbr->r_wanted_output = 1; 9512 return (ret_val); 9513 } else { 9514 ctf_do_drop(m, NULL); 9515 return (0); 9516 } 9517 } 9518 /* 9519 * Ack processing. 9520 */ 9521 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9522 return (ret_val); 9523 } 9524 if (ourfinisacked) { 9525 /* 9526 * If we can't receive any more data, then closing user can 9527 * proceed. Starting the timer is contrary to the 9528 * specification, but if we don't get a FIN we'll hang 9529 * forever. 9530 * 9531 * XXXjl: we should release the tp also, and use a 9532 * compressed state. 9533 */ 9534 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9535 soisdisconnected(so); 9536 tcp_timer_activate(tp, TT_2MSL, 9537 (tcp_fast_finwait2_recycle ? 9538 tcp_finwait2_timeout : 9539 TP_MAXIDLE(tp))); 9540 } 9541 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9542 } 9543 if (sbavail(&so->so_snd)) { 9544 if (ctf_progress_timeout_check(tp, true)) { 9545 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9546 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9547 return (1); 9548 } 9549 } 9550 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9551 tiwin, thflags, nxt_pkt)); 9552 } 9553 9554 /* 9555 * Return value of 1, the TCB is unlocked and most 9556 * likely gone, return value of 0, the TCB is still 9557 * locked. 9558 */ 9559 static int 9560 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9561 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9562 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9563 { 9564 int32_t ourfinisacked = 0; 9565 int32_t ret_val; 9566 struct tcp_bbr *bbr; 9567 9568 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9569 9570 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9571 ctf_calc_rwin(so, tp); 9572 if ((thflags & TH_RST) || 9573 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9574 return (ctf_process_rst(m, th, so, tp)); 9575 /* 9576 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9577 * synchronized state. 9578 */ 9579 if (thflags & TH_SYN) { 9580 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9581 return (ret_val); 9582 } 9583 /* 9584 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9585 * it's less than ts_recent, drop it. 9586 */ 9587 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9588 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9589 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9590 return (ret_val); 9591 } 9592 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9593 return (ret_val); 9594 } 9595 /* 9596 * If new data are received on a connection after the user processes 9597 * are gone, then RST the other end. 9598 * We call a new function now so we might continue and setup 9599 * to reset at all data being ack'd. 9600 */ 9601 if ((tp->t_flags & TF_CLOSED) && tlen && 9602 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9603 return (1); 9604 /* 9605 * If last ACK falls within this segment's sequence numbers, record 9606 * its timestamp. NOTE: 1) That the test incorporates suggestions 9607 * from the latest proposal of the tcplw@cray.com list (Braden 9608 * 1993/04/26). 2) That updating only on newer timestamps interferes 9609 * with our earlier PAWS tests, so this check should be solely 9610 * predicated on the sequence space of this segment. 3) That we 9611 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9612 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9613 * SEG.Len, This modified check allows us to overcome RFC1323's 9614 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9615 * p.869. In such cases, we can still calculate the RTT correctly 9616 * when RCV.NXT == Last.ACK.Sent. 9617 */ 9618 if ((to->to_flags & TOF_TS) != 0 && 9619 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9620 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9621 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9622 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9623 tp->ts_recent = to->to_tsval; 9624 } 9625 /* 9626 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9627 * is on (half-synchronized state), then queue data for later 9628 * processing; else drop segment and return. 9629 */ 9630 if ((thflags & TH_ACK) == 0) { 9631 if (tp->t_flags & TF_NEEDSYN) { 9632 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9633 tiwin, thflags, nxt_pkt)); 9634 } else if (tp->t_flags & TF_ACKNOW) { 9635 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9636 bbr->r_wanted_output = 1; 9637 return (ret_val); 9638 } else { 9639 ctf_do_drop(m, NULL); 9640 return (0); 9641 } 9642 } 9643 /* 9644 * Ack processing. 9645 */ 9646 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9647 return (ret_val); 9648 } 9649 if (ourfinisacked) { 9650 tcp_twstart(tp); 9651 m_freem(m); 9652 return (1); 9653 } 9654 if (sbavail(&so->so_snd)) { 9655 if (ctf_progress_timeout_check(tp, true)) { 9656 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9657 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9658 return (1); 9659 } 9660 } 9661 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9662 tiwin, thflags, nxt_pkt)); 9663 } 9664 9665 /* 9666 * Return value of 1, the TCB is unlocked and most 9667 * likely gone, return value of 0, the TCB is still 9668 * locked. 9669 */ 9670 static int 9671 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9672 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9673 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9674 { 9675 int32_t ourfinisacked = 0; 9676 int32_t ret_val; 9677 struct tcp_bbr *bbr; 9678 9679 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9680 9681 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9682 ctf_calc_rwin(so, tp); 9683 if ((thflags & TH_RST) || 9684 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9685 return (ctf_process_rst(m, th, so, tp)); 9686 /* 9687 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9688 * synchronized state. 9689 */ 9690 if (thflags & TH_SYN) { 9691 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9692 return (ret_val); 9693 } 9694 /* 9695 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9696 * it's less than ts_recent, drop it. 9697 */ 9698 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9699 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9700 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9701 return (ret_val); 9702 } 9703 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9704 return (ret_val); 9705 } 9706 /* 9707 * If new data are received on a connection after the user processes 9708 * are gone, then RST the other end. 9709 * We call a new function now so we might continue and setup 9710 * to reset at all data being ack'd. 9711 */ 9712 if ((tp->t_flags & TF_CLOSED) && tlen && 9713 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9714 return (1); 9715 /* 9716 * If last ACK falls within this segment's sequence numbers, record 9717 * its timestamp. NOTE: 1) That the test incorporates suggestions 9718 * from the latest proposal of the tcplw@cray.com list (Braden 9719 * 1993/04/26). 2) That updating only on newer timestamps interferes 9720 * with our earlier PAWS tests, so this check should be solely 9721 * predicated on the sequence space of this segment. 3) That we 9722 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9723 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9724 * SEG.Len, This modified check allows us to overcome RFC1323's 9725 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9726 * p.869. In such cases, we can still calculate the RTT correctly 9727 * when RCV.NXT == Last.ACK.Sent. 9728 */ 9729 if ((to->to_flags & TOF_TS) != 0 && 9730 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9731 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9732 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9733 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9734 tp->ts_recent = to->to_tsval; 9735 } 9736 /* 9737 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9738 * is on (half-synchronized state), then queue data for later 9739 * processing; else drop segment and return. 9740 */ 9741 if ((thflags & TH_ACK) == 0) { 9742 if (tp->t_flags & TF_NEEDSYN) { 9743 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9744 tiwin, thflags, nxt_pkt)); 9745 } else if (tp->t_flags & TF_ACKNOW) { 9746 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9747 bbr->r_wanted_output = 1; 9748 return (ret_val); 9749 } else { 9750 ctf_do_drop(m, NULL); 9751 return (0); 9752 } 9753 } 9754 /* 9755 * case TCPS_LAST_ACK: Ack processing. 9756 */ 9757 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9758 return (ret_val); 9759 } 9760 if (ourfinisacked) { 9761 tp = tcp_close(tp); 9762 ctf_do_drop(m, tp); 9763 return (1); 9764 } 9765 if (sbavail(&so->so_snd)) { 9766 if (ctf_progress_timeout_check(tp, true)) { 9767 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9768 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9769 return (1); 9770 } 9771 } 9772 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9773 tiwin, thflags, nxt_pkt)); 9774 } 9775 9776 /* 9777 * Return value of 1, the TCB is unlocked and most 9778 * likely gone, return value of 0, the TCB is still 9779 * locked. 9780 */ 9781 static int 9782 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9783 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9784 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9785 { 9786 int32_t ourfinisacked = 0; 9787 int32_t ret_val; 9788 struct tcp_bbr *bbr; 9789 9790 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9791 9792 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9793 ctf_calc_rwin(so, tp); 9794 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9795 if ((thflags & TH_RST) || 9796 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9797 return (ctf_process_rst(m, th, so, tp)); 9798 9799 /* 9800 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9801 * synchronized state. 9802 */ 9803 if (thflags & TH_SYN) { 9804 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9805 return (ret_val); 9806 } 9807 /* 9808 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9809 * it's less than ts_recent, drop it. 9810 */ 9811 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9812 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9813 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9814 return (ret_val); 9815 } 9816 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9817 return (ret_val); 9818 } 9819 /* 9820 * If new data are received on a connection after the user processes 9821 * are gone, then we may RST the other end depending on the outcome 9822 * of bbr_check_data_after_close. 9823 * We call a new function now so we might continue and setup 9824 * to reset at all data being ack'd. 9825 */ 9826 if ((tp->t_flags & TF_CLOSED) && tlen && 9827 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9828 return (1); 9829 /* 9830 * If last ACK falls within this segment's sequence numbers, record 9831 * its timestamp. NOTE: 1) That the test incorporates suggestions 9832 * from the latest proposal of the tcplw@cray.com list (Braden 9833 * 1993/04/26). 2) That updating only on newer timestamps interferes 9834 * with our earlier PAWS tests, so this check should be solely 9835 * predicated on the sequence space of this segment. 3) That we 9836 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9837 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9838 * SEG.Len, This modified check allows us to overcome RFC1323's 9839 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9840 * p.869. In such cases, we can still calculate the RTT correctly 9841 * when RCV.NXT == Last.ACK.Sent. 9842 */ 9843 if ((to->to_flags & TOF_TS) != 0 && 9844 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9845 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9846 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9847 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9848 tp->ts_recent = to->to_tsval; 9849 } 9850 /* 9851 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9852 * is on (half-synchronized state), then queue data for later 9853 * processing; else drop segment and return. 9854 */ 9855 if ((thflags & TH_ACK) == 0) { 9856 if (tp->t_flags & TF_NEEDSYN) { 9857 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9858 tiwin, thflags, nxt_pkt)); 9859 } else if (tp->t_flags & TF_ACKNOW) { 9860 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9861 bbr->r_wanted_output = 1; 9862 return (ret_val); 9863 } else { 9864 ctf_do_drop(m, NULL); 9865 return (0); 9866 } 9867 } 9868 /* 9869 * Ack processing. 9870 */ 9871 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9872 return (ret_val); 9873 } 9874 if (sbavail(&so->so_snd)) { 9875 if (ctf_progress_timeout_check(tp, true)) { 9876 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9877 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9878 return (1); 9879 } 9880 } 9881 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9882 tiwin, thflags, nxt_pkt)); 9883 } 9884 9885 static void 9886 bbr_stop_all_timers(struct tcpcb *tp) 9887 { 9888 struct tcp_bbr *bbr; 9889 9890 /* 9891 * Assure no timers are running. 9892 */ 9893 if (tcp_timer_active(tp, TT_PERSIST)) { 9894 /* We enter in persists, set the flag appropriately */ 9895 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9896 bbr->rc_in_persist = 1; 9897 } 9898 } 9899 9900 static void 9901 bbr_google_mode_on(struct tcp_bbr *bbr) 9902 { 9903 bbr->rc_use_google = 1; 9904 bbr->rc_no_pacing = 0; 9905 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9906 bbr->r_use_policer = bbr_policer_detection_enabled; 9907 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9908 bbr->bbr_use_rack_cheat = 0; 9909 bbr->r_ctl.rc_incr_tmrs = 0; 9910 bbr->r_ctl.rc_inc_tcp_oh = 0; 9911 bbr->r_ctl.rc_inc_ip_oh = 0; 9912 bbr->r_ctl.rc_inc_enet_oh = 0; 9913 reset_time(&bbr->r_ctl.rc_delrate, 9914 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9915 reset_time_small(&bbr->r_ctl.rc_rttprop, 9916 (11 * USECS_IN_SECOND)); 9917 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9918 } 9919 9920 static void 9921 bbr_google_mode_off(struct tcp_bbr *bbr) 9922 { 9923 bbr->rc_use_google = 0; 9924 bbr->r_ctl.bbr_google_discount = 0; 9925 bbr->no_pacing_until = bbr_no_pacing_until; 9926 bbr->r_use_policer = 0; 9927 if (bbr->no_pacing_until) 9928 bbr->rc_no_pacing = 1; 9929 else 9930 bbr->rc_no_pacing = 0; 9931 if (bbr_use_rack_resend_cheat) 9932 bbr->bbr_use_rack_cheat = 1; 9933 else 9934 bbr->bbr_use_rack_cheat = 0; 9935 if (bbr_incr_timers) 9936 bbr->r_ctl.rc_incr_tmrs = 1; 9937 else 9938 bbr->r_ctl.rc_incr_tmrs = 0; 9939 if (bbr_include_tcp_oh) 9940 bbr->r_ctl.rc_inc_tcp_oh = 1; 9941 else 9942 bbr->r_ctl.rc_inc_tcp_oh = 0; 9943 if (bbr_include_ip_oh) 9944 bbr->r_ctl.rc_inc_ip_oh = 1; 9945 else 9946 bbr->r_ctl.rc_inc_ip_oh = 0; 9947 if (bbr_include_enet_oh) 9948 bbr->r_ctl.rc_inc_enet_oh = 1; 9949 else 9950 bbr->r_ctl.rc_inc_enet_oh = 0; 9951 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9952 reset_time(&bbr->r_ctl.rc_delrate, 9953 bbr_num_pktepo_for_del_limit); 9954 reset_time_small(&bbr->r_ctl.rc_rttprop, 9955 (bbr_filter_len_sec * USECS_IN_SECOND)); 9956 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9957 } 9958 /* 9959 * Return 0 on success, non-zero on failure 9960 * which indicates the error (usually no memory). 9961 */ 9962 static int 9963 bbr_init(struct tcpcb *tp) 9964 { 9965 struct inpcb *inp = tptoinpcb(tp); 9966 struct tcp_bbr *bbr = NULL; 9967 uint32_t cts; 9968 9969 tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 9970 if (tp->t_fb_ptr == NULL) { 9971 /* 9972 * We need to allocate memory but cant. The INP and INP_INFO 9973 * locks and they are recursive (happens during setup. So a 9974 * scheme to drop the locks fails :( 9975 * 9976 */ 9977 return (ENOMEM); 9978 } 9979 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9980 bbr->rtt_valid = 0; 9981 inp->inp_flags2 |= INP_CANNOT_DO_ECN; 9982 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; 9983 TAILQ_INIT(&bbr->r_ctl.rc_map); 9984 TAILQ_INIT(&bbr->r_ctl.rc_free); 9985 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 9986 bbr->rc_tp = tp; 9987 bbr->rc_inp = inp; 9988 cts = tcp_get_usecs(&bbr->rc_tv); 9989 tp->t_acktime = 0; 9990 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 9991 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 9992 bbr->rc_tlp_threshold = bbr_tlp_thresh; 9993 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 9994 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 9995 bbr->r_ctl.rc_min_to = bbr_min_to; 9996 bbr->rc_bbr_state = BBR_STATE_STARTUP; 9997 bbr->r_ctl.bbr_lost_at_state = 0; 9998 bbr->r_ctl.rc_lost_at_startup = 0; 9999 bbr->rc_all_timers_stopped = 0; 10000 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 10001 bbr->r_ctl.rc_pkt_epoch_del = 0; 10002 bbr->r_ctl.rc_pkt_epoch = 0; 10003 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 10004 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 10005 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 10006 bbr->r_ctl.rc_went_idle_time = cts; 10007 bbr->rc_pacer_started = cts; 10008 bbr->r_ctl.rc_pkt_epoch_time = cts; 10009 bbr->r_ctl.rc_rcvtime = cts; 10010 bbr->r_ctl.rc_bbr_state_time = cts; 10011 bbr->r_ctl.rc_del_time = cts; 10012 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 10013 bbr->r_ctl.last_in_probertt = cts; 10014 bbr->skip_gain = 0; 10015 bbr->gain_is_limited = 0; 10016 bbr->no_pacing_until = bbr_no_pacing_until; 10017 if (bbr->no_pacing_until) 10018 bbr->rc_no_pacing = 1; 10019 if (bbr_use_google_algo) { 10020 bbr->rc_no_pacing = 0; 10021 bbr->rc_use_google = 1; 10022 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 10023 bbr->r_use_policer = bbr_policer_detection_enabled; 10024 } else { 10025 bbr->rc_use_google = 0; 10026 bbr->r_ctl.bbr_google_discount = 0; 10027 bbr->r_use_policer = 0; 10028 } 10029 if (bbr_ts_limiting) 10030 bbr->rc_use_ts_limit = 1; 10031 else 10032 bbr->rc_use_ts_limit = 0; 10033 if (bbr_ts_can_raise) 10034 bbr->ts_can_raise = 1; 10035 else 10036 bbr->ts_can_raise = 0; 10037 if (V_tcp_delack_enabled == 1) 10038 tp->t_delayed_ack = 2; 10039 else if (V_tcp_delack_enabled == 0) 10040 tp->t_delayed_ack = 0; 10041 else if (V_tcp_delack_enabled < 100) 10042 tp->t_delayed_ack = V_tcp_delack_enabled; 10043 else 10044 tp->t_delayed_ack = 2; 10045 if (bbr->rc_use_google == 0) 10046 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10047 else 10048 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10049 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10050 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10051 bbr->rc_init_win = bbr_def_init_win; 10052 if (tp->t_flags & TF_REQ_TSTMP) 10053 bbr->rc_last_options = TCP_TS_OVERHEAD; 10054 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10055 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10056 bbr->r_init_rtt = 1; 10057 10058 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10059 if (bbr_allow_hdwr_pacing) 10060 bbr->bbr_hdw_pace_ena = 1; 10061 else 10062 bbr->bbr_hdw_pace_ena = 0; 10063 if (bbr_sends_full_iwnd) 10064 bbr->bbr_init_win_cheat = 1; 10065 else 10066 bbr->bbr_init_win_cheat = 0; 10067 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10068 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10069 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10070 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10071 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10072 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10073 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10074 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10075 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10076 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10077 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10078 bbr->r_ctl.rc_rtt_shrinks = cts; 10079 if (bbr->rc_use_google) { 10080 setup_time_filter(&bbr->r_ctl.rc_delrate, 10081 FILTER_TYPE_MAX, 10082 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10083 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10084 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10085 } else { 10086 setup_time_filter(&bbr->r_ctl.rc_delrate, 10087 FILTER_TYPE_MAX, 10088 bbr_num_pktepo_for_del_limit); 10089 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10090 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10091 } 10092 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10093 if (bbr_uses_idle_restart) 10094 bbr->rc_use_idle_restart = 1; 10095 else 10096 bbr->rc_use_idle_restart = 0; 10097 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10098 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10099 if (bbr_resends_use_tso) 10100 bbr->rc_resends_use_tso = 1; 10101 #ifdef NETFLIX_PEAKRATE 10102 tp->t_peakrate_thr = tp->t_maxpeakrate; 10103 #endif 10104 if (tp->snd_una != tp->snd_max) { 10105 /* Create a send map for the current outstanding data */ 10106 struct bbr_sendmap *rsm; 10107 10108 rsm = bbr_alloc(bbr); 10109 if (rsm == NULL) { 10110 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10111 tp->t_fb_ptr = NULL; 10112 return (ENOMEM); 10113 } 10114 rsm->r_rtt_not_allowed = 1; 10115 rsm->r_tim_lastsent[0] = cts; 10116 rsm->r_rtr_cnt = 1; 10117 rsm->r_rtr_bytes = 0; 10118 rsm->r_start = tp->snd_una; 10119 rsm->r_end = tp->snd_max; 10120 rsm->r_dupack = 0; 10121 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10122 rsm->r_ts_valid = 0; 10123 rsm->r_del_ack_ts = tp->ts_recent; 10124 rsm->r_del_time = cts; 10125 if (bbr->r_ctl.r_app_limited_until) 10126 rsm->r_app_limited = 1; 10127 else 10128 rsm->r_app_limited = 0; 10129 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10130 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10131 rsm->r_in_tmap = 1; 10132 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10133 rsm->r_bbr_state = bbr_state_val(bbr); 10134 else 10135 rsm->r_bbr_state = 8; 10136 } 10137 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10138 bbr->bbr_use_rack_cheat = 1; 10139 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10140 bbr->r_ctl.rc_incr_tmrs = 1; 10141 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10142 bbr->r_ctl.rc_inc_tcp_oh = 1; 10143 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10144 bbr->r_ctl.rc_inc_ip_oh = 1; 10145 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10146 bbr->r_ctl.rc_inc_enet_oh = 1; 10147 10148 bbr_log_type_statechange(bbr, cts, __LINE__); 10149 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10150 (tp->t_srtt)) { 10151 uint32_t rtt; 10152 10153 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10154 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10155 } 10156 /* announce the settings and state */ 10157 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10158 tcp_bbr_tso_size_check(bbr, cts); 10159 /* 10160 * Now call the generic function to start a timer. This will place 10161 * the TCB on the hptsi wheel if a timer is needed with appropriate 10162 * flags. 10163 */ 10164 bbr_stop_all_timers(tp); 10165 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10166 return (0); 10167 } 10168 10169 /* 10170 * Return 0 if we can accept the connection. Return 10171 * non-zero if we can't handle the connection. A EAGAIN 10172 * means you need to wait until the connection is up. 10173 * a EADDRNOTAVAIL means we can never handle the connection 10174 * (no SACK). 10175 */ 10176 static int 10177 bbr_handoff_ok(struct tcpcb *tp) 10178 { 10179 if ((tp->t_state == TCPS_CLOSED) || 10180 (tp->t_state == TCPS_LISTEN)) { 10181 /* Sure no problem though it may not stick */ 10182 return (0); 10183 } 10184 if ((tp->t_state == TCPS_SYN_SENT) || 10185 (tp->t_state == TCPS_SYN_RECEIVED)) { 10186 /* 10187 * We really don't know you have to get to ESTAB or beyond 10188 * to tell. 10189 */ 10190 return (EAGAIN); 10191 } 10192 if (tp->t_flags & TF_SENTFIN) 10193 return (EINVAL); 10194 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10195 return (0); 10196 } 10197 /* 10198 * If we reach here we don't do SACK on this connection so we can 10199 * never do rack. 10200 */ 10201 return (EINVAL); 10202 } 10203 10204 static void 10205 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10206 { 10207 if (tp->t_fb_ptr) { 10208 struct inpcb *inp = tptoinpcb(tp); 10209 uint32_t calc; 10210 struct tcp_bbr *bbr; 10211 struct bbr_sendmap *rsm; 10212 10213 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10214 if (bbr->r_ctl.crte) 10215 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10216 bbr_log_flowend(bbr); 10217 bbr->rc_tp = NULL; 10218 /* Backout any flags2 we applied */ 10219 inp->inp_flags2 &= ~INP_CANNOT_DO_ECN; 10220 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; 10221 inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 10222 if (bbr->bbr_hdrw_pacing) 10223 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10224 else 10225 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10226 if (bbr->r_ctl.crte != NULL) { 10227 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10228 bbr->r_ctl.crte = NULL; 10229 } 10230 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10231 while (rsm) { 10232 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10233 uma_zfree(bbr_zone, rsm); 10234 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10235 } 10236 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10237 while (rsm) { 10238 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10239 uma_zfree(bbr_zone, rsm); 10240 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10241 } 10242 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10243 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10244 BBR_STAT_INC(bbr_dynamic_rwnd); 10245 else 10246 BBR_STAT_INC(bbr_static_rwnd); 10247 bbr->r_ctl.rc_free_cnt = 0; 10248 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10249 tp->t_fb_ptr = NULL; 10250 } 10251 /* Make sure snd_nxt is correctly set */ 10252 tp->snd_nxt = tp->snd_max; 10253 } 10254 10255 static void 10256 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10257 { 10258 switch (tp->t_state) { 10259 case TCPS_SYN_SENT: 10260 bbr->r_state = TCPS_SYN_SENT; 10261 bbr->r_substate = bbr_do_syn_sent; 10262 break; 10263 case TCPS_SYN_RECEIVED: 10264 bbr->r_state = TCPS_SYN_RECEIVED; 10265 bbr->r_substate = bbr_do_syn_recv; 10266 break; 10267 case TCPS_ESTABLISHED: 10268 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10269 bbr->r_state = TCPS_ESTABLISHED; 10270 bbr->r_substate = bbr_do_established; 10271 break; 10272 case TCPS_CLOSE_WAIT: 10273 bbr->r_state = TCPS_CLOSE_WAIT; 10274 bbr->r_substate = bbr_do_close_wait; 10275 break; 10276 case TCPS_FIN_WAIT_1: 10277 bbr->r_state = TCPS_FIN_WAIT_1; 10278 bbr->r_substate = bbr_do_fin_wait_1; 10279 break; 10280 case TCPS_CLOSING: 10281 bbr->r_state = TCPS_CLOSING; 10282 bbr->r_substate = bbr_do_closing; 10283 break; 10284 case TCPS_LAST_ACK: 10285 bbr->r_state = TCPS_LAST_ACK; 10286 bbr->r_substate = bbr_do_lastack; 10287 break; 10288 case TCPS_FIN_WAIT_2: 10289 bbr->r_state = TCPS_FIN_WAIT_2; 10290 bbr->r_substate = bbr_do_fin_wait_2; 10291 break; 10292 case TCPS_LISTEN: 10293 case TCPS_CLOSED: 10294 case TCPS_TIME_WAIT: 10295 default: 10296 break; 10297 }; 10298 } 10299 10300 static void 10301 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10302 { 10303 /* 10304 * Now what state are we going into now? Is there adjustments 10305 * needed? 10306 */ 10307 int32_t old_state; 10308 10309 old_state = bbr_state_val(bbr); 10310 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10311 /* Save the lowest srtt we saw in our end of the sub-state */ 10312 bbr->rc_hit_state_1 = 0; 10313 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10314 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10315 } 10316 bbr->rc_bbr_substate++; 10317 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10318 /* Cycle back to first state-> gain */ 10319 bbr->rc_bbr_substate = 0; 10320 } 10321 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10322 /* 10323 * We enter the gain(5/4) cycle (possibly less if 10324 * shallow buffer detection is enabled) 10325 */ 10326 if (bbr->skip_gain) { 10327 /* 10328 * Hardware pacing has set our rate to 10329 * the max and limited our b/w just 10330 * do level i.e. no gain. 10331 */ 10332 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10333 } else if (bbr->gain_is_limited && 10334 bbr->bbr_hdrw_pacing && 10335 bbr->r_ctl.crte) { 10336 /* 10337 * We can't gain above the hardware pacing 10338 * rate which is less than our rate + the gain 10339 * calculate the gain needed to reach the hardware 10340 * pacing rate.. 10341 */ 10342 uint64_t bw, rate, gain_calc; 10343 10344 bw = bbr_get_bw(bbr); 10345 rate = bbr->r_ctl.crte->rate; 10346 if ((rate > bw) && 10347 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10348 gain_calc = (rate * BBR_UNIT) / bw; 10349 if (gain_calc < BBR_UNIT) 10350 gain_calc = BBR_UNIT; 10351 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10352 } else { 10353 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10354 } 10355 } else 10356 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10357 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10358 bbr->r_ctl.rc_bbr_state_atflight = cts; 10359 } else 10360 bbr->r_ctl.rc_bbr_state_atflight = 0; 10361 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10362 bbr->rc_hit_state_1 = 1; 10363 bbr->r_ctl.rc_exta_time_gd = 0; 10364 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10365 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10366 if (bbr_state_drain_2_tar) { 10367 bbr->r_ctl.rc_bbr_state_atflight = 0; 10368 } else 10369 bbr->r_ctl.rc_bbr_state_atflight = cts; 10370 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10371 } else { 10372 /* All other cycles hit here 2-7 */ 10373 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10374 if (bbr_sub_drain_slam_cwnd && 10375 (bbr->rc_use_google == 0) && 10376 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10377 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10378 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10379 } 10380 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10381 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10382 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10383 else 10384 bbr->r_ctl.rc_exta_time_gd = 0; 10385 if (bbr->r_ctl.rc_exta_time_gd) { 10386 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10387 /* Now chop up the time for each state (div by 7) */ 10388 bbr->r_ctl.rc_level_state_extra /= 7; 10389 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10390 /* Add a randomization */ 10391 bbr_randomize_extra_state_time(bbr); 10392 } 10393 } 10394 } 10395 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10396 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10397 } 10398 if (bbr->rc_use_google) { 10399 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10400 } 10401 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10402 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10403 if (dolog) 10404 bbr_log_type_statechange(bbr, cts, line); 10405 10406 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10407 uint32_t time_in; 10408 10409 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10410 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10411 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10412 } else { 10413 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10414 } 10415 } 10416 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10417 bbr_set_state_target(bbr, __LINE__); 10418 if (bbr_sub_drain_slam_cwnd && 10419 (bbr->rc_use_google == 0) && 10420 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10421 /* Slam down the cwnd */ 10422 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10423 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10424 if (bbr_sub_drain_app_limit) { 10425 /* Go app limited if we are on a long drain */ 10426 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10427 ctf_flight_size(bbr->rc_tp, 10428 (bbr->r_ctl.rc_sacked + 10429 bbr->r_ctl.rc_lost_bytes))); 10430 } 10431 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10432 } 10433 if (bbr->rc_lt_use_bw) { 10434 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10435 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10436 } 10437 /* Google changes TSO size every cycle */ 10438 if (bbr->rc_use_google) 10439 tcp_bbr_tso_size_check(bbr, cts); 10440 bbr->r_ctl.gain_epoch = cts; 10441 bbr->r_ctl.rc_bbr_state_time = cts; 10442 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10443 } 10444 10445 static void 10446 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10447 { 10448 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10449 (google_allow_early_out == 1) && 10450 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10451 /* We have reached out target flight size possibly early */ 10452 goto change_state; 10453 } 10454 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10455 return; 10456 } 10457 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10458 /* 10459 * Must be a rttProp movement forward before 10460 * we can change states. 10461 */ 10462 return; 10463 } 10464 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10465 /* 10466 * The needed time has passed but for 10467 * the gain cycle extra rules apply: 10468 * 1) If we have seen loss, we exit 10469 * 2) If we have not reached the target 10470 * we stay in GAIN (gain-to-target). 10471 */ 10472 if (google_consider_lost && losses) 10473 goto change_state; 10474 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10475 return; 10476 } 10477 } 10478 change_state: 10479 /* For gain we must reach our target, all others last 1 rttProp */ 10480 bbr_substate_change(bbr, cts, __LINE__, 1); 10481 } 10482 10483 static void 10484 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10485 { 10486 uint32_t flight, bbr_cur_cycle_time; 10487 10488 if (bbr->rc_use_google) { 10489 bbr_set_probebw_google_gains(bbr, cts, losses); 10490 return; 10491 } 10492 if (cts == 0) { 10493 /* 10494 * Never alow cts to be 0 we 10495 * do this so we can judge if 10496 * we have set a timestamp. 10497 */ 10498 cts = 1; 10499 } 10500 if (bbr_state_is_pkt_epoch) 10501 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10502 else 10503 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10504 10505 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10506 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10507 flight = ctf_flight_size(bbr->rc_tp, 10508 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10509 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10510 /* Keep it slam down */ 10511 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10512 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10513 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10514 } 10515 if (bbr_sub_drain_app_limit) { 10516 /* Go app limited if we are on a long drain */ 10517 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10518 } 10519 } 10520 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10521 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10522 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10523 /* 10524 * Still here after the same time as 10525 * the gain. We need to drain harder 10526 * for the next srtt. Reduce by a set amount 10527 * the gain drop is capped at DRAIN states 10528 * value (88). 10529 */ 10530 bbr->r_ctl.flightsize_at_drain = flight; 10531 if (bbr_drain_drop_mul && 10532 bbr_drain_drop_div && 10533 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10534 /* Use your specific drop value (def 4/5 = 20%) */ 10535 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10536 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10537 } else { 10538 /* You get drop of 20% */ 10539 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10540 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10541 } 10542 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10543 /* Reduce our gain again to the bottom */ 10544 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10545 } 10546 bbr_log_exit_gain(bbr, cts, 4); 10547 /* 10548 * Extend out so we wait another 10549 * epoch before dropping again. 10550 */ 10551 bbr->r_ctl.gain_epoch = cts; 10552 } 10553 if (flight <= bbr->r_ctl.rc_target_at_state) { 10554 if (bbr_sub_drain_slam_cwnd && 10555 (bbr->rc_use_google == 0) && 10556 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10557 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10558 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10559 } 10560 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10561 bbr_log_exit_gain(bbr, cts, 3); 10562 } 10563 } else { 10564 /* Its a gain */ 10565 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10566 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10567 goto change_state; 10568 } 10569 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10570 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10571 bbr->rc_tp->snd_wnd)) { 10572 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10573 bbr_log_exit_gain(bbr, cts, 2); 10574 } 10575 } 10576 /** 10577 * We fall through and return always one of two things has 10578 * occurred. 10579 * 1) We are still not at target 10580 * <or> 10581 * 2) We reached the target and set rc_bbr_state_atflight 10582 * which means we no longer hit this block 10583 * next time we are called. 10584 */ 10585 return; 10586 } 10587 change_state: 10588 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10589 return; 10590 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10591 /* Less than a full time-period has passed */ 10592 return; 10593 } 10594 if (bbr->r_ctl.rc_level_state_extra && 10595 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10596 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10597 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10598 /* Less than a full time-period + extra has passed */ 10599 return; 10600 } 10601 if (bbr_gain_gets_extra_too && 10602 bbr->r_ctl.rc_level_state_extra && 10603 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10604 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10605 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10606 /* Less than a full time-period + extra has passed */ 10607 return; 10608 } 10609 bbr_substate_change(bbr, cts, __LINE__, 1); 10610 } 10611 10612 static uint32_t 10613 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10614 { 10615 uint32_t mss, tar; 10616 10617 if (bbr->rc_use_google) { 10618 /* Google just uses the cwnd target */ 10619 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10620 } else { 10621 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10622 bbr->r_ctl.rc_pace_max_segs); 10623 /* Get the base cwnd with gain rounded to a mss */ 10624 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10625 gain), mss); 10626 /* Make sure it is within our min */ 10627 if (tar < get_min_cwnd(bbr)) 10628 return (get_min_cwnd(bbr)); 10629 } 10630 return (tar); 10631 } 10632 10633 static void 10634 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10635 { 10636 uint32_t tar, meth; 10637 10638 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10639 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10640 /* Special case using old probe-rtt method */ 10641 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10642 meth = 1; 10643 } else { 10644 /* Non-probe-rtt case and reduced probe-rtt */ 10645 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10646 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10647 /* For gain cycle we use the hptsi gain */ 10648 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10649 meth = 2; 10650 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10651 /* 10652 * If configured, or for google all other states 10653 * get BBR_UNIT. 10654 */ 10655 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10656 meth = 3; 10657 } else { 10658 /* 10659 * Or we set a target based on the pacing gain 10660 * for non-google mode and default (non-configured). 10661 * Note we don't set a target goal below drain (192). 10662 */ 10663 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10664 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10665 meth = 4; 10666 } else { 10667 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10668 meth = 5; 10669 } 10670 } 10671 } 10672 bbr_log_set_of_state_target(bbr, tar, line, meth); 10673 bbr->r_ctl.rc_target_at_state = tar; 10674 } 10675 10676 static void 10677 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10678 { 10679 /* Change to probe_rtt */ 10680 uint32_t time_in; 10681 10682 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10683 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10684 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10685 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10686 + bbr->r_ctl.rc_delivered); 10687 /* Setup so we force feed the filter */ 10688 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10689 bbr->rc_prtt_set_ts = 1; 10690 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10691 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10692 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10693 } 10694 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10695 bbr->r_ctl.rc_rtt_shrinks = cts; 10696 bbr->r_ctl.last_in_probertt = cts; 10697 bbr->r_ctl.rc_probertt_srttchktim = cts; 10698 bbr->r_ctl.rc_bbr_state_time = cts; 10699 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10700 /* We need to force the filter to update */ 10701 10702 if ((bbr_sub_drain_slam_cwnd) && 10703 bbr->rc_hit_state_1 && 10704 (bbr->rc_use_google == 0) && 10705 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10706 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10707 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10708 } else 10709 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10710 /* Update the lost */ 10711 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10712 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10713 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10714 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10715 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10716 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10717 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10718 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10719 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10720 } else { 10721 /* 10722 * We bring it down slowly by using a hptsi gain that is 10723 * probably 75%. This will slowly float down our outstanding 10724 * without tampering with the cwnd. 10725 */ 10726 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10727 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10728 bbr_set_state_target(bbr, __LINE__); 10729 if (bbr_prtt_slam_cwnd && 10730 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10731 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10732 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10733 } 10734 } 10735 if (ctf_flight_size(bbr->rc_tp, 10736 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10737 bbr->r_ctl.rc_target_at_state) { 10738 /* We are at target */ 10739 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10740 } else { 10741 /* We need to come down to reach target before our time begins */ 10742 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10743 } 10744 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10745 BBR_STAT_INC(bbr_enter_probertt); 10746 bbr_log_exit_gain(bbr, cts, 0); 10747 bbr_log_type_statechange(bbr, cts, line); 10748 } 10749 10750 static void 10751 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10752 { 10753 /* 10754 * Sanity check on probe-rtt intervals. 10755 * In crazy situations where we are competing 10756 * against new-reno flows with huge buffers 10757 * our rtt-prop interval could come to dominate 10758 * things if we can't get through a full set 10759 * of cycles, we need to adjust it. 10760 */ 10761 if (bbr_can_adjust_probertt && 10762 (bbr->rc_use_google == 0)) { 10763 uint16_t val = 0; 10764 uint32_t cur_rttp, fval, newval, baseval; 10765 10766 /* Are we to small and go into probe-rtt to often? */ 10767 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10768 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10769 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10770 if (bbr_is_ratio == 0) { 10771 if (fval > bbr_rtt_probe_limit) 10772 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10773 else 10774 newval = cur_rttp; 10775 } else { 10776 int mul; 10777 10778 mul = fval / bbr_rtt_probe_limit; 10779 newval = cur_rttp * mul; 10780 } 10781 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10782 bbr->r_ctl.rc_probertt_int = cur_rttp; 10783 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10784 val = 1; 10785 } else { 10786 /* 10787 * No adjustments were made 10788 * do we need to shrink it? 10789 */ 10790 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10791 if (cur_rttp <= bbr_rtt_probe_limit) { 10792 /* 10793 * Things have calmed down lets 10794 * shrink all the way to default 10795 */ 10796 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10797 reset_time_small(&bbr->r_ctl.rc_rttprop, 10798 (bbr_filter_len_sec * USECS_IN_SECOND)); 10799 cur_rttp = bbr_rtt_probe_limit; 10800 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10801 val = 2; 10802 } else { 10803 /* 10804 * Well does some adjustment make sense? 10805 */ 10806 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10807 /* We can reduce interval time some */ 10808 bbr->r_ctl.rc_probertt_int = cur_rttp; 10809 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10810 val = 3; 10811 } 10812 } 10813 } 10814 } 10815 if (val) 10816 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10817 } 10818 } 10819 10820 static void 10821 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10822 { 10823 /* Exit probe-rtt */ 10824 10825 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10826 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10827 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10828 } 10829 bbr_log_exit_gain(bbr, cts, 1); 10830 bbr->rc_hit_state_1 = 0; 10831 bbr->r_ctl.rc_rtt_shrinks = cts; 10832 bbr->r_ctl.last_in_probertt = cts; 10833 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10834 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10835 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10836 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10837 bbr->r_ctl.rc_delivered); 10838 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10839 uint32_t time_in; 10840 10841 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10842 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10843 } 10844 if (bbr->rc_filled_pipe) { 10845 /* Switch to probe_bw */ 10846 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10847 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10848 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10849 bbr_substate_change(bbr, cts, __LINE__, 0); 10850 bbr_log_type_statechange(bbr, cts, __LINE__); 10851 } else { 10852 /* Back to startup */ 10853 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10854 bbr->r_ctl.rc_bbr_state_time = cts; 10855 /* 10856 * We don't want to give a complete free 3 10857 * measurements until we exit, so we use 10858 * the number of pe's we were in probe-rtt 10859 * to add to the startup_epoch. That way 10860 * we will still retain the old state. 10861 */ 10862 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10863 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10864 /* Make sure to use the lower pg when shifting back in */ 10865 if (bbr->r_ctl.rc_lost && 10866 bbr_use_lower_gain_in_startup && 10867 (bbr->rc_use_google == 0)) 10868 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10869 else 10870 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10871 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10872 /* Probably not needed but set it anyway */ 10873 bbr_set_state_target(bbr, __LINE__); 10874 bbr_log_type_statechange(bbr, cts, __LINE__); 10875 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10876 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10877 } 10878 bbr_check_probe_rtt_limits(bbr, cts); 10879 } 10880 10881 static int32_t inline 10882 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10883 { 10884 if ((bbr->rc_past_init_win == 1) && 10885 (bbr->rc_in_persist == 0) && 10886 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10887 return (1); 10888 } 10889 if (bbr_can_force_probertt && 10890 (bbr->rc_in_persist == 0) && 10891 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10892 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10893 return (1); 10894 } 10895 return (0); 10896 } 10897 10898 static int32_t 10899 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10900 { 10901 uint64_t btlbw, gain; 10902 if (pkt_epoch == 0) { 10903 /* 10904 * Need to be on a pkt-epoch to continue. 10905 */ 10906 return (0); 10907 } 10908 btlbw = bbr_get_full_bw(bbr); 10909 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10910 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10911 if (btlbw >= gain) { 10912 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10913 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10914 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10915 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10916 } 10917 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10918 return (1); 10919 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10920 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10921 return(0); 10922 } 10923 10924 static int32_t inline 10925 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10926 { 10927 /* Have we gained 25% in the last 3 packet based epoch's? */ 10928 uint64_t btlbw, gain; 10929 int do_exit; 10930 int delta, rtt_gain; 10931 10932 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10933 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10934 /* 10935 * This qualifies as a RTT_PROBE session since we drop the 10936 * data outstanding to nothing and waited more than 10937 * bbr_rtt_probe_time. 10938 */ 10939 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10940 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10941 } 10942 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10943 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10944 return (0); 10945 } 10946 if (bbr->rc_use_google) 10947 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10948 10949 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10950 (bbr_use_lower_gain_in_startup)) { 10951 /* Drop to a lower gain 1.5 x since we saw loss */ 10952 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10953 } 10954 if (pkt_epoch == 0) { 10955 /* 10956 * Need to be on a pkt-epoch to continue. 10957 */ 10958 return (0); 10959 } 10960 if (bbr_rtt_gain_thresh) { 10961 /* 10962 * Do we allow a flow to stay 10963 * in startup with no loss and no 10964 * gain in rtt over a set threshold? 10965 */ 10966 if (bbr->r_ctl.rc_pkt_epoch_rtt && 10967 bbr->r_ctl.startup_last_srtt && 10968 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 10969 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 10970 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 10971 } else 10972 rtt_gain = 0; 10973 if ((bbr->r_ctl.startup_last_srtt == 0) || 10974 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 10975 /* First time or new lower value */ 10976 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 10977 10978 if ((bbr->r_ctl.rc_lost == 0) && 10979 (rtt_gain < bbr_rtt_gain_thresh)) { 10980 /* 10981 * No loss, and we are under 10982 * our gain threhold for 10983 * increasing RTT. 10984 */ 10985 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10986 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10987 bbr_log_startup_event(bbr, cts, rtt_gain, 10988 delta, bbr->r_ctl.startup_last_srtt, 10); 10989 return (0); 10990 } 10991 } 10992 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 10993 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 10994 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 10995 /* 10996 * We only assess if we have a new measurement when 10997 * we have no loss and are not in recovery. 10998 * Drag up by one our last_startup epoch so we will hold 10999 * the number of non-gain we have already accumulated. 11000 */ 11001 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 11002 bbr->r_ctl.rc_bbr_last_startup_epoch++; 11003 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11004 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); 11005 return (0); 11006 } 11007 /* Case where we reduced the lost (bad retransmit) */ 11008 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 11009 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11010 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 11011 btlbw = bbr_get_full_bw(bbr); 11012 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 11013 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11014 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11015 else 11016 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11017 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11018 do_exit = 0; 11019 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 11020 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 11021 if (btlbw >= gain) { 11022 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 11023 /* Update the lost so we won't exit in next set of tests */ 11024 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11025 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11026 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 11027 } 11028 if ((bbr->rc_loss_exit && 11029 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 11030 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 11031 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 11032 /* 11033 * If we had no gain, we had loss and that loss was above 11034 * our threshould, the rwnd is not constrained, and we have 11035 * had at least 3 packet epochs exit. Note that this is 11036 * switched off by sysctl. Google does not do this by the 11037 * way. 11038 */ 11039 if ((ctf_flight_size(bbr->rc_tp, 11040 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11041 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11042 do_exit = 1; 11043 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11044 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11045 } else { 11046 /* Just record an updated loss value */ 11047 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11048 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11049 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11050 } 11051 } else 11052 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11053 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11054 do_exit) { 11055 /* Return 1 to exit the startup state. */ 11056 return (1); 11057 } 11058 /* Stay in startup */ 11059 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11060 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11061 return (0); 11062 } 11063 11064 static void 11065 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11066 { 11067 /* 11068 * A tick occurred in the rtt epoch do we need to do anything? 11069 */ 11070 #ifdef BBR_INVARIANTS 11071 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11072 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11073 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11074 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11075 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11076 /* Debug code? */ 11077 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11078 } 11079 #endif 11080 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11081 /* Do we exit the startup state? */ 11082 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11083 uint32_t time_in; 11084 11085 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11086 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11087 bbr->rc_filled_pipe = 1; 11088 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11089 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11090 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11091 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11092 } else 11093 time_in = 0; 11094 if (bbr->rc_no_pacing) 11095 bbr->rc_no_pacing = 0; 11096 bbr->r_ctl.rc_bbr_state_time = cts; 11097 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11098 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11099 bbr_set_state_target(bbr, __LINE__); 11100 if ((bbr->rc_use_google == 0) && 11101 bbr_slam_cwnd_in_main_drain) { 11102 /* Here we don't have to worry about probe-rtt */ 11103 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11104 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11105 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11106 } 11107 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11108 bbr_log_type_statechange(bbr, cts, __LINE__); 11109 if (ctf_flight_size(bbr->rc_tp, 11110 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11111 bbr->r_ctl.rc_target_at_state) { 11112 /* 11113 * Switch to probe_bw if we are already 11114 * there 11115 */ 11116 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11117 bbr_substate_change(bbr, cts, __LINE__, 0); 11118 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11119 bbr_log_type_statechange(bbr, cts, __LINE__); 11120 } 11121 } 11122 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11123 uint32_t inflight; 11124 struct tcpcb *tp; 11125 11126 tp = bbr->rc_tp; 11127 inflight = ctf_flight_size(tp, 11128 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11129 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11130 /* We have reached a flight of the cwnd target */ 11131 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11132 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11133 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11134 bbr_set_state_target(bbr, __LINE__); 11135 /* 11136 * Rig it so we don't do anything crazy and 11137 * start fresh with a new randomization. 11138 */ 11139 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11140 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11141 bbr_substate_change(bbr, cts, __LINE__, 1); 11142 } 11143 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11144 /* Has in-flight reached the bdp (or less)? */ 11145 uint32_t inflight; 11146 struct tcpcb *tp; 11147 11148 tp = bbr->rc_tp; 11149 inflight = ctf_flight_size(tp, 11150 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11151 if ((bbr->rc_use_google == 0) && 11152 bbr_slam_cwnd_in_main_drain && 11153 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11154 /* 11155 * Here we don't have to worry about probe-rtt 11156 * re-slam it, but keep it slammed down. 11157 */ 11158 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11159 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11160 } 11161 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11162 /* We have drained */ 11163 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11164 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11165 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11166 uint32_t time_in; 11167 11168 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11169 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11170 } 11171 if ((bbr->rc_use_google == 0) && 11172 bbr_slam_cwnd_in_main_drain && 11173 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11174 /* Restore the cwnd */ 11175 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11176 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11177 } 11178 /* Setup probe-rtt has being done now RRS-HERE */ 11179 bbr->r_ctl.rc_rtt_shrinks = cts; 11180 bbr->r_ctl.last_in_probertt = cts; 11181 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11182 /* Randomly pick a sub-state */ 11183 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11184 bbr_substate_change(bbr, cts, __LINE__, 0); 11185 bbr_log_type_statechange(bbr, cts, __LINE__); 11186 } 11187 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11188 uint32_t flight; 11189 11190 flight = ctf_flight_size(bbr->rc_tp, 11191 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11192 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11193 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11194 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11195 /* 11196 * We must keep cwnd at the desired MSS. 11197 */ 11198 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11199 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11200 } else if ((bbr_prtt_slam_cwnd) && 11201 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11202 /* Re-slam it */ 11203 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11204 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11205 } 11206 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11207 /* Has outstanding reached our target? */ 11208 if (flight <= bbr->r_ctl.rc_target_at_state) { 11209 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11210 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11211 /* If time is exactly 0, be 1usec off */ 11212 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11213 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11214 if (bbr->rc_use_google == 0) { 11215 /* 11216 * Restore any lowering that as occurred to 11217 * reach here 11218 */ 11219 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11220 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11221 else 11222 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11223 } 11224 } 11225 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11226 (bbr->rc_use_google == 0) && 11227 bbr->r_ctl.bbr_rttprobe_gain_val && 11228 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11229 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11230 /* 11231 * We have doddled with our current hptsi 11232 * gain an srtt and have still not made it 11233 * to target, or we have increased our flight. 11234 * Lets reduce the gain by xx% 11235 * flooring the reduce at DRAIN (based on 11236 * mul/div) 11237 */ 11238 int red; 11239 11240 bbr->r_ctl.flightsize_at_drain = flight; 11241 bbr->r_ctl.rc_probertt_srttchktim = cts; 11242 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11243 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11244 /* Reduce our gain again */ 11245 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11246 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11247 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11248 /* one more chance before we give up */ 11249 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11250 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11251 } else { 11252 /* At the very bottom */ 11253 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11254 } 11255 } 11256 } 11257 if (bbr->r_ctl.rc_bbr_enters_probertt && 11258 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11259 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11260 /* Time to exit probe RTT normally */ 11261 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11262 } 11263 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11264 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11265 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11266 /* 11267 * This qualifies as a RTT_PROBE session since we 11268 * drop the data outstanding to nothing and waited 11269 * more than bbr_rtt_probe_time. 11270 */ 11271 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11272 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11273 } 11274 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11275 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11276 } else { 11277 bbr_set_probebw_gains(bbr, cts, losses); 11278 } 11279 } 11280 } 11281 11282 static void 11283 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11284 { 11285 int32_t epoch = 0; 11286 11287 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11288 bbr_set_epoch(bbr, cts, line); 11289 /* At each epoch doe lt bw sampling */ 11290 epoch = 1; 11291 } 11292 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11293 } 11294 11295 static int 11296 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, 11297 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 11298 int32_t nxt_pkt, struct timeval *tv) 11299 { 11300 struct inpcb *inp = tptoinpcb(tp); 11301 int32_t thflags, retval; 11302 uint32_t cts, lcts; 11303 uint32_t tiwin; 11304 struct tcpopt to; 11305 struct tcp_bbr *bbr; 11306 struct bbr_sendmap *rsm; 11307 struct timeval ltv; 11308 int32_t did_out = 0; 11309 uint16_t nsegs; 11310 int32_t prev_state; 11311 uint32_t lost; 11312 11313 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11314 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11315 /* add in our stats */ 11316 kern_prefetch(bbr, &prev_state); 11317 prev_state = 0; 11318 thflags = tcp_get_flags(th); 11319 /* 11320 * If this is either a state-changing packet or current state isn't 11321 * established, we require a write lock on tcbinfo. Otherwise, we 11322 * allow the tcbinfo to be in either alocked or unlocked, as the 11323 * caller may have unnecessarily acquired a write lock due to a 11324 * race. 11325 */ 11326 INP_WLOCK_ASSERT(tptoinpcb(tp)); 11327 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11328 __func__)); 11329 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11330 __func__)); 11331 11332 tp->t_rcvtime = ticks; 11333 /* 11334 * Unscale the window into a 32-bit value. For the SYN_SENT state 11335 * the scale is zero. 11336 */ 11337 tiwin = th->th_win << tp->snd_scale; 11338 #ifdef STATS 11339 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11340 #endif 11341 11342 if (m->m_flags & M_TSTMP) { 11343 /* Prefer the hardware timestamp if present */ 11344 struct timespec ts; 11345 11346 mbuf_tstmp2timespec(m, &ts); 11347 bbr->rc_tv.tv_sec = ts.tv_sec; 11348 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11349 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11350 } else if (m->m_flags & M_TSTMP_LRO) { 11351 /* Next the arrival timestamp */ 11352 struct timespec ts; 11353 11354 mbuf_tstmp2timespec(m, &ts); 11355 bbr->rc_tv.tv_sec = ts.tv_sec; 11356 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11357 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11358 } else { 11359 /* 11360 * Ok just get the current time. 11361 */ 11362 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11363 } 11364 /* 11365 * Parse options on any incoming segment. 11366 */ 11367 tcp_dooptions(&to, (u_char *)(th + 1), 11368 (th->th_off << 2) - sizeof(struct tcphdr), 11369 (thflags & TH_SYN) ? TO_SYN : 0); 11370 11371 /* 11372 * If timestamps were negotiated during SYN/ACK and a 11373 * segment without a timestamp is received, silently drop 11374 * the segment, unless it is a RST segment or missing timestamps are 11375 * tolerated. 11376 * See section 3.2 of RFC 7323. 11377 */ 11378 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11379 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11380 retval = 0; 11381 m_freem(m); 11382 goto done_with_input; 11383 } 11384 /* 11385 * If echoed timestamp is later than the current time, fall back to 11386 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11387 * were used when this connection was established. 11388 */ 11389 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11390 to.to_tsecr -= tp->ts_offset; 11391 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11392 to.to_tsecr = 0; 11393 } 11394 /* 11395 * If its the first time in we need to take care of options and 11396 * verify we can do SACK for rack! 11397 */ 11398 if (bbr->r_state == 0) { 11399 /* 11400 * Process options only when we get SYN/ACK back. The SYN 11401 * case for incoming connections is handled in tcp_syncache. 11402 * According to RFC1323 the window field in a SYN (i.e., a 11403 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11404 * this is traditional behavior, may need to be cleaned up. 11405 */ 11406 if (bbr->rc_inp == NULL) { 11407 bbr->rc_inp = inp; 11408 } 11409 /* 11410 * We need to init rc_inp here since its not init'd when 11411 * bbr_init is called 11412 */ 11413 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11414 if ((to.to_flags & TOF_SCALE) && 11415 (tp->t_flags & TF_REQ_SCALE)) { 11416 tp->t_flags |= TF_RCVD_SCALE; 11417 tp->snd_scale = to.to_wscale; 11418 } else 11419 tp->t_flags &= ~TF_REQ_SCALE; 11420 /* 11421 * Initial send window. It will be updated with the 11422 * next incoming segment to the scaled value. 11423 */ 11424 tp->snd_wnd = th->th_win; 11425 if ((to.to_flags & TOF_TS) && 11426 (tp->t_flags & TF_REQ_TSTMP)) { 11427 tp->t_flags |= TF_RCVD_TSTMP; 11428 tp->ts_recent = to.to_tsval; 11429 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11430 } else 11431 tp->t_flags &= ~TF_REQ_TSTMP; 11432 if (to.to_flags & TOF_MSS) 11433 tcp_mss(tp, to.to_mss); 11434 if ((tp->t_flags & TF_SACK_PERMIT) && 11435 (to.to_flags & TOF_SACKPERM) == 0) 11436 tp->t_flags &= ~TF_SACK_PERMIT; 11437 if (IS_FASTOPEN(tp->t_flags)) { 11438 if (to.to_flags & TOF_FASTOPEN) { 11439 uint16_t mss; 11440 11441 if (to.to_flags & TOF_MSS) 11442 mss = to.to_mss; 11443 else 11444 if ((inp->inp_vflag & INP_IPV6) != 0) 11445 mss = TCP6_MSS; 11446 else 11447 mss = TCP_MSS; 11448 tcp_fastopen_update_cache(tp, mss, 11449 to.to_tfo_len, to.to_tfo_cookie); 11450 } else 11451 tcp_fastopen_disable_path(tp); 11452 } 11453 } 11454 /* 11455 * At this point we are at the initial call. Here we decide 11456 * if we are doing RACK or not. We do this by seeing if 11457 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11458 * we switch to the default code. 11459 */ 11460 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11461 /* Bail */ 11462 tcp_switch_back_to_default(tp); 11463 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 11464 tlen, iptos); 11465 return (1); 11466 } 11467 /* Set the flag */ 11468 bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0; 11469 tcp_set_hpts(inp); 11470 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11471 } 11472 if (thflags & TH_ACK) { 11473 /* Track ack types */ 11474 if (to.to_flags & TOF_SACK) 11475 BBR_STAT_INC(bbr_acks_with_sacks); 11476 else 11477 BBR_STAT_INC(bbr_plain_acks); 11478 } 11479 /* 11480 * This is the one exception case where we set the rack state 11481 * always. All other times (timers etc) we must have a rack-state 11482 * set (so we assure we have done the checks above for SACK). 11483 */ 11484 if (thflags & TH_FIN) 11485 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11486 if (bbr->r_state != tp->t_state) 11487 bbr_set_state(tp, bbr, tiwin); 11488 11489 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11490 kern_prefetch(rsm, &prev_state); 11491 prev_state = bbr->r_state; 11492 bbr->rc_ack_was_delayed = 0; 11493 lost = bbr->r_ctl.rc_lost; 11494 bbr->rc_is_pkt_epoch_now = 0; 11495 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11496 /* Get the real time into lcts and figure the real delay */ 11497 lcts = tcp_get_usecs(<v); 11498 if (TSTMP_GT(lcts, cts)) { 11499 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11500 bbr->rc_ack_was_delayed = 1; 11501 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11502 bbr->r_ctl.highest_hdwr_delay)) 11503 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11504 } else { 11505 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11506 bbr->rc_ack_was_delayed = 0; 11507 } 11508 } else { 11509 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11510 bbr->rc_ack_was_delayed = 0; 11511 } 11512 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11513 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11514 retval = 0; 11515 m_freem(m); 11516 goto done_with_input; 11517 } 11518 /* 11519 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11520 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11521 */ 11522 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11523 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11524 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11525 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11526 return (1); 11527 } 11528 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11529 bbr->r_ctl.rc_high_rwnd = tiwin; 11530 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11531 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11532 bbr->rtt_valid = 0; 11533 if (to.to_flags & TOF_TS) { 11534 bbr->rc_ts_valid = 1; 11535 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11536 } else { 11537 bbr->rc_ts_valid = 0; 11538 bbr->r_ctl.last_inbound_ts = 0; 11539 } 11540 retval = (*bbr->r_substate) (m, th, so, 11541 tp, &to, drop_hdrlen, 11542 tlen, tiwin, thflags, nxt_pkt, iptos); 11543 if (nxt_pkt == 0) 11544 BBR_STAT_INC(bbr_rlock_left_ret0); 11545 else 11546 BBR_STAT_INC(bbr_rlock_left_ret1); 11547 if (retval == 0) { 11548 /* 11549 * If retval is 1 the tcb is unlocked and most likely the tp 11550 * is gone. 11551 */ 11552 INP_WLOCK_ASSERT(inp); 11553 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11554 if (bbr->rc_is_pkt_epoch_now) 11555 bbr_set_pktepoch(bbr, cts, __LINE__); 11556 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11557 if (nxt_pkt == 0) { 11558 if (bbr->r_wanted_output != 0) { 11559 bbr->rc_output_starts_timer = 0; 11560 did_out = 1; 11561 if (tcp_output(tp) < 0) 11562 return (1); 11563 } else 11564 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11565 } 11566 if ((nxt_pkt == 0) && 11567 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11568 (SEQ_GT(tp->snd_max, tp->snd_una) || 11569 (tp->t_flags & TF_DELACK) || 11570 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11571 (tp->t_state <= TCPS_CLOSING)))) { 11572 /* 11573 * We could not send (probably in the hpts but 11574 * stopped the timer)? 11575 */ 11576 if ((tp->snd_max == tp->snd_una) && 11577 ((tp->t_flags & TF_DELACK) == 0) && 11578 (tcp_in_hpts(bbr->rc_inp)) && 11579 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11580 /* 11581 * keep alive not needed if we are hptsi 11582 * output yet 11583 */ 11584 ; 11585 } else { 11586 if (tcp_in_hpts(bbr->rc_inp)) { 11587 tcp_hpts_remove(bbr->rc_inp); 11588 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11589 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11590 uint32_t del; 11591 11592 del = lcts - bbr->rc_pacer_started; 11593 if (bbr->r_ctl.rc_last_delay_val > del) { 11594 BBR_STAT_INC(bbr_force_timer_start); 11595 bbr->r_ctl.rc_last_delay_val -= del; 11596 bbr->rc_pacer_started = lcts; 11597 } else { 11598 /* We are late */ 11599 bbr->r_ctl.rc_last_delay_val = 0; 11600 BBR_STAT_INC(bbr_force_output); 11601 if (tcp_output(tp) < 0) 11602 return (1); 11603 } 11604 } 11605 } 11606 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11607 0); 11608 } 11609 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11610 /* Do we have the correct timer running? */ 11611 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11612 } 11613 /* Do we have a new state */ 11614 if (bbr->r_state != tp->t_state) 11615 bbr_set_state(tp, bbr, tiwin); 11616 done_with_input: 11617 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11618 if (did_out) 11619 bbr->r_wanted_output = 0; 11620 } 11621 return (retval); 11622 } 11623 11624 static void 11625 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 11626 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11627 { 11628 struct timeval tv; 11629 int retval; 11630 11631 /* First lets see if we have old packets */ 11632 if (tp->t_in_pkt) { 11633 if (ctf_do_queued_segments(so, tp, 1)) { 11634 m_freem(m); 11635 return; 11636 } 11637 } 11638 if (m->m_flags & M_TSTMP_LRO) { 11639 mbuf_tstmp2timeval(m, &tv); 11640 } else { 11641 /* Should not be should we kassert instead? */ 11642 tcp_get_usecs(&tv); 11643 } 11644 retval = bbr_do_segment_nounlock(m, th, so, tp, 11645 drop_hdrlen, tlen, iptos, 0, &tv); 11646 if (retval == 0) { 11647 INP_WUNLOCK(tptoinpcb(tp)); 11648 } 11649 } 11650 11651 /* 11652 * Return how much data can be sent without violating the 11653 * cwnd or rwnd. 11654 */ 11655 11656 static inline uint32_t 11657 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11658 uint32_t avail, int32_t sb_offset, uint32_t cts) 11659 { 11660 uint32_t len; 11661 11662 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11663 /* We never want to go over our peers rcv-window */ 11664 len = 0; 11665 } else { 11666 uint32_t flight; 11667 11668 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11669 if (flight >= sendwin) { 11670 /* 11671 * We have in flight what we are allowed by cwnd (if 11672 * it was rwnd blocking it would have hit above out 11673 * >= tp->snd_wnd). 11674 */ 11675 return (0); 11676 } 11677 len = sendwin - flight; 11678 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11679 /* We would send too much (beyond the rwnd) */ 11680 len = tp->snd_wnd - ctf_outstanding(tp); 11681 } 11682 if ((len + sb_offset) > avail) { 11683 /* 11684 * We don't have that much in the SB, how much is 11685 * there? 11686 */ 11687 len = avail - sb_offset; 11688 } 11689 } 11690 return (len); 11691 } 11692 11693 static inline void 11694 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11695 { 11696 #ifdef NETFLIX_STATS 11697 KMOD_TCPSTAT_INC(tcps_sndpack_error); 11698 KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); 11699 #endif 11700 } 11701 11702 static inline void 11703 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11704 { 11705 if (error) { 11706 bbr_do_error_accounting(tp, bbr, rsm, len, error); 11707 return; 11708 } 11709 if (rsm) { 11710 if (rsm->r_flags & BBR_TLP) { 11711 /* 11712 * TLP should not count in retran count, but in its 11713 * own bin 11714 */ 11715 #ifdef NETFLIX_STATS 11716 KMOD_TCPSTAT_INC(tcps_tlpresends); 11717 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11718 #endif 11719 } else { 11720 /* Retransmit */ 11721 tp->t_sndrexmitpack++; 11722 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11723 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11724 #ifdef STATS 11725 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11726 len); 11727 #endif 11728 } 11729 /* 11730 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11731 * sub-state 11732 */ 11733 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11734 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11735 /* Non probe_bw log in 1, 2, or 4. */ 11736 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11737 } else { 11738 /* 11739 * Log our probe state 3, and log also 5-13 to show 11740 * us the recovery sub-state for the send. This 11741 * means that 3 == (5+6+7+8+9+10+11+12+13) 11742 */ 11743 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11744 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11745 } 11746 /* Place in both 16's the totals of retransmitted */ 11747 counter_u64_add(bbr_state_lost[16], len); 11748 counter_u64_add(bbr_state_resend[16], len); 11749 /* Place in 17's the total sent */ 11750 counter_u64_add(bbr_state_resend[17], len); 11751 counter_u64_add(bbr_state_lost[17], len); 11752 11753 } else { 11754 /* New sends */ 11755 KMOD_TCPSTAT_INC(tcps_sndpack); 11756 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11757 /* Place in 17's the total sent */ 11758 counter_u64_add(bbr_state_resend[17], len); 11759 counter_u64_add(bbr_state_lost[17], len); 11760 #ifdef STATS 11761 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11762 len); 11763 #endif 11764 } 11765 } 11766 11767 static void 11768 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11769 { 11770 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11771 /* 11772 * Limit the cwnd to not be above N x the target plus whats 11773 * is outstanding. The target is based on the current b/w 11774 * estimate. 11775 */ 11776 uint32_t target; 11777 11778 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11779 target += ctf_outstanding(tp); 11780 target *= bbr_target_cwnd_mult_limit; 11781 if (tp->snd_cwnd > target) 11782 tp->snd_cwnd = target; 11783 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11784 } 11785 } 11786 11787 static int 11788 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11789 { 11790 /* 11791 * "adv" is the amount we could increase the window, taking into 11792 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11793 */ 11794 int32_t adv; 11795 int32_t oldwin; 11796 11797 adv = recwin; 11798 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11799 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11800 if (adv > oldwin) 11801 adv -= oldwin; 11802 else { 11803 /* We can't increase the window */ 11804 adv = 0; 11805 } 11806 } else 11807 oldwin = 0; 11808 11809 /* 11810 * If the new window size ends up being the same as or less 11811 * than the old size when it is scaled, then don't force 11812 * a window update. 11813 */ 11814 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11815 return (0); 11816 11817 if (adv >= (2 * maxseg) && 11818 (adv >= (so->so_rcv.sb_hiwat / 4) || 11819 recwin <= (so->so_rcv.sb_hiwat / 8) || 11820 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11821 return (1); 11822 } 11823 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11824 return (1); 11825 return (0); 11826 } 11827 11828 /* 11829 * Return 0 on success and a errno on failure to send. 11830 * Note that a 0 return may not mean we sent anything 11831 * if the TCB was on the hpts. A non-zero return 11832 * does indicate the error we got from ip[6]_output. 11833 */ 11834 static int 11835 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11836 { 11837 struct socket *so; 11838 int32_t len; 11839 uint32_t cts; 11840 uint32_t recwin, sendwin; 11841 int32_t sb_offset; 11842 int32_t flags, abandon, error = 0; 11843 struct tcp_log_buffer *lgb = NULL; 11844 struct mbuf *m; 11845 struct mbuf *mb; 11846 uint32_t if_hw_tsomaxsegcount = 0; 11847 uint32_t if_hw_tsomaxsegsize = 0; 11848 uint32_t if_hw_tsomax = 0; 11849 struct ip *ip = NULL; 11850 #ifdef TCPDEBUG 11851 struct ipovly *ipov = NULL; 11852 #endif 11853 struct tcp_bbr *bbr; 11854 struct tcphdr *th; 11855 struct udphdr *udp = NULL; 11856 u_char opt[TCP_MAXOLEN]; 11857 unsigned ipoptlen, optlen, hdrlen; 11858 unsigned ulen; 11859 uint32_t bbr_seq; 11860 uint32_t delay_calc=0; 11861 uint8_t doing_tlp = 0; 11862 uint8_t local_options; 11863 #ifdef BBR_INVARIANTS 11864 uint8_t doing_retran_from = 0; 11865 uint8_t picked_up_retran = 0; 11866 #endif 11867 uint8_t wanted_cookie = 0; 11868 uint8_t more_to_rxt=0; 11869 int32_t prefetch_so_done = 0; 11870 int32_t prefetch_rsm = 0; 11871 uint32_t tot_len = 0; 11872 uint32_t maxseg, pace_max_segs, p_maxseg; 11873 int32_t csum_flags = 0; 11874 int32_t hw_tls; 11875 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11876 unsigned ipsec_optlen = 0; 11877 11878 #endif 11879 volatile int32_t sack_rxmit; 11880 struct bbr_sendmap *rsm = NULL; 11881 int32_t tso, mtu; 11882 struct tcpopt to; 11883 int32_t slot = 0; 11884 struct inpcb *inp; 11885 struct sockbuf *sb; 11886 uint32_t hpts_calling; 11887 #ifdef INET6 11888 struct ip6_hdr *ip6 = NULL; 11889 int32_t isipv6; 11890 #endif 11891 uint8_t app_limited = BBR_JR_SENT_DATA; 11892 uint8_t filled_all = 0; 11893 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11894 /* We take a cache hit here */ 11895 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11896 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11897 inp = bbr->rc_inp; 11898 so = inp->inp_socket; 11899 sb = &so->so_snd; 11900 if (sb->sb_flags & SB_TLS_IFNET) 11901 hw_tls = 1; 11902 else 11903 hw_tls = 0; 11904 kern_prefetch(sb, &maxseg); 11905 maxseg = tp->t_maxseg - bbr->rc_last_options; 11906 if (bbr_minseg(bbr) < maxseg) { 11907 tcp_bbr_tso_size_check(bbr, cts); 11908 } 11909 /* Remove any flags that indicate we are pacing on the inp */ 11910 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11911 p_maxseg = min(maxseg, pace_max_segs); 11912 INP_WLOCK_ASSERT(inp); 11913 #ifdef TCP_OFFLOAD 11914 if (tp->t_flags & TF_TOE) 11915 return (tcp_offload_output(tp)); 11916 #endif 11917 11918 #ifdef INET6 11919 if (bbr->r_state) { 11920 /* Use the cache line loaded if possible */ 11921 isipv6 = bbr->r_is_v6; 11922 } else { 11923 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11924 } 11925 #endif 11926 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11927 tcp_in_hpts(inp)) { 11928 /* 11929 * We are on the hpts for some timer but not hptsi output. 11930 * Possibly remove from the hpts so we can send/recv etc. 11931 */ 11932 if ((tp->t_flags & TF_ACKNOW) == 0) { 11933 /* 11934 * No immediate demand right now to send an ack, but 11935 * the user may have read, making room for new data 11936 * (a window update). If so we may want to cancel 11937 * whatever timer is running (KEEP/DEL-ACK?) and 11938 * continue to send out a window update. Or we may 11939 * have gotten more data into the socket buffer to 11940 * send. 11941 */ 11942 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 11943 (long)TCP_MAXWIN << tp->rcv_scale); 11944 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 11945 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 11946 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 11947 (tp->snd_max - tp->snd_una))) { 11948 /* 11949 * Nothing new to send and no window update 11950 * is needed to send. Lets just return and 11951 * let the timer-run off. 11952 */ 11953 return (0); 11954 } 11955 } 11956 tcp_hpts_remove(inp); 11957 bbr_timer_cancel(bbr, __LINE__, cts); 11958 } 11959 if (bbr->r_ctl.rc_last_delay_val) { 11960 /* Calculate a rough delay for early escape to sending */ 11961 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11962 delay_calc = cts - bbr->rc_pacer_started; 11963 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11964 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11965 else 11966 delay_calc = 0; 11967 } 11968 /* Mark that we have called bbr_output(). */ 11969 if ((bbr->r_timer_override) || 11970 (tp->t_state < TCPS_ESTABLISHED)) { 11971 /* Timeouts or early states are exempt */ 11972 if (tcp_in_hpts(inp)) 11973 tcp_hpts_remove(inp); 11974 } else if (tcp_in_hpts(inp)) { 11975 if ((bbr->r_ctl.rc_last_delay_val) && 11976 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11977 delay_calc) { 11978 /* 11979 * We were being paced for output and the delay has 11980 * already exceeded when we were supposed to be 11981 * called, lets go ahead and pull out of the hpts 11982 * and call output. 11983 */ 11984 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 11985 bbr->r_ctl.rc_last_delay_val = 0; 11986 tcp_hpts_remove(inp); 11987 } else if (tp->t_state == TCPS_CLOSED) { 11988 bbr->r_ctl.rc_last_delay_val = 0; 11989 tcp_hpts_remove(inp); 11990 } else { 11991 /* 11992 * On the hpts, you shall not pass! even if ACKNOW 11993 * is on, we will when the hpts fires, unless of 11994 * course we are overdue. 11995 */ 11996 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 11997 return (0); 11998 } 11999 } 12000 bbr->rc_cwnd_limited = 0; 12001 if (bbr->r_ctl.rc_last_delay_val) { 12002 /* recalculate the real delay and deal with over/under */ 12003 if (SEQ_GT(cts, bbr->rc_pacer_started)) 12004 delay_calc = cts - bbr->rc_pacer_started; 12005 else 12006 delay_calc = 0; 12007 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 12008 /* Setup the delay which will be added in */ 12009 delay_calc -= bbr->r_ctl.rc_last_delay_val; 12010 else { 12011 /* 12012 * We are early setup to adjust 12013 * our slot time. 12014 */ 12015 uint64_t merged_val; 12016 12017 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 12018 bbr->r_agg_early_set = 1; 12019 if (bbr->r_ctl.rc_hptsi_agg_delay) { 12020 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 12021 /* Nope our previous late cancels out the early */ 12022 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 12023 bbr->r_agg_early_set = 0; 12024 bbr->r_ctl.rc_agg_early = 0; 12025 } else { 12026 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 12027 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12028 } 12029 } 12030 merged_val = bbr->rc_pacer_started; 12031 merged_val <<= 32; 12032 merged_val |= bbr->r_ctl.rc_last_delay_val; 12033 bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, 12034 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 12035 bbr->r_agg_early_set, 3); 12036 bbr->r_ctl.rc_last_delay_val = 0; 12037 BBR_STAT_INC(bbr_early); 12038 delay_calc = 0; 12039 } 12040 } else { 12041 /* We were not delayed due to hptsi */ 12042 if (bbr->r_agg_early_set) 12043 bbr->r_ctl.rc_agg_early = 0; 12044 bbr->r_agg_early_set = 0; 12045 delay_calc = 0; 12046 } 12047 if (delay_calc) { 12048 /* 12049 * We had a hptsi delay which means we are falling behind on 12050 * sending at the expected rate. Calculate an extra amount 12051 * of data we can send, if any, to put us back on track. 12052 */ 12053 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12054 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12055 else 12056 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12057 } 12058 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12059 if ((tp->snd_una == tp->snd_max) && 12060 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12061 (sbavail(sb))) { 12062 /* 12063 * Ok we have been idle with nothing outstanding 12064 * we possibly need to start fresh with either a new 12065 * suite of states or a fast-ramp up. 12066 */ 12067 bbr_restart_after_idle(bbr, 12068 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12069 } 12070 /* 12071 * Now was there a hptsi delay where we are behind? We only count 12072 * being behind if: a) We are not in recovery. b) There was a delay. 12073 * <and> c) We had room to send something. 12074 * 12075 */ 12076 hpts_calling = inp->inp_hpts_calls; 12077 inp->inp_hpts_calls = 0; 12078 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12079 int retval; 12080 12081 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12082 if (retval != 0) { 12083 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12084 /* 12085 * If timers want tcp_drop(), then pass error out, 12086 * otherwise suppress it. 12087 */ 12088 return (retval < 0 ? retval : 0); 12089 } 12090 } 12091 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 12092 if (hpts_calling && 12093 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12094 bbr->r_ctl.rc_last_delay_val = 0; 12095 } 12096 bbr->r_timer_override = 0; 12097 bbr->r_wanted_output = 0; 12098 /* 12099 * For TFO connections in SYN_RECEIVED, only allow the initial 12100 * SYN|ACK and those sent by the retransmit timer. 12101 */ 12102 if (IS_FASTOPEN(tp->t_flags) && 12103 ((tp->t_state == TCPS_SYN_RECEIVED) || 12104 (tp->t_state == TCPS_SYN_SENT)) && 12105 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12106 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12107 len = 0; 12108 goto just_return_nolock; 12109 } 12110 /* 12111 * Before sending anything check for a state update. For hpts 12112 * calling without input this is important. If its input calling 12113 * then this was already done. 12114 */ 12115 if (bbr->rc_use_google == 0) 12116 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12117 again: 12118 /* 12119 * If we've recently taken a timeout, snd_max will be greater than 12120 * snd_max. BBR in general does not pay much attention to snd_nxt 12121 * for historic reasons the persist timer still uses it. This means 12122 * we have to look at it. All retransmissions that are not persits 12123 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12124 * end of this routine we pull snd_nxt always up to snd_max. 12125 */ 12126 doing_tlp = 0; 12127 #ifdef BBR_INVARIANTS 12128 doing_retran_from = picked_up_retran = 0; 12129 #endif 12130 error = 0; 12131 tso = 0; 12132 slot = 0; 12133 mtu = 0; 12134 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12135 sb_offset = tp->snd_max - tp->snd_una; 12136 flags = tcp_outflags[tp->t_state]; 12137 sack_rxmit = 0; 12138 len = 0; 12139 rsm = NULL; 12140 if (flags & TH_RST) { 12141 SOCKBUF_LOCK(sb); 12142 goto send; 12143 } 12144 recheck_resend: 12145 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12146 /* We need to always have one in reserve */ 12147 rsm = bbr_alloc(bbr); 12148 if (rsm == NULL) { 12149 error = ENOMEM; 12150 /* Lie to get on the hpts */ 12151 tot_len = tp->t_maxseg; 12152 if (hpts_calling) 12153 /* Retry in a ms */ 12154 slot = 1001; 12155 goto just_return_nolock; 12156 } 12157 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12158 bbr->r_ctl.rc_free_cnt++; 12159 rsm = NULL; 12160 } 12161 /* What do we send, a resend? */ 12162 if (bbr->r_ctl.rc_resend == NULL) { 12163 /* Check for rack timeout */ 12164 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12165 if (bbr->r_ctl.rc_resend) { 12166 #ifdef BBR_INVARIANTS 12167 picked_up_retran = 1; 12168 #endif 12169 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12170 } 12171 } 12172 if (bbr->r_ctl.rc_resend) { 12173 rsm = bbr->r_ctl.rc_resend; 12174 #ifdef BBR_INVARIANTS 12175 doing_retran_from = 1; 12176 #endif 12177 /* Remove any TLP flags its a RACK or T-O */ 12178 rsm->r_flags &= ~BBR_TLP; 12179 bbr->r_ctl.rc_resend = NULL; 12180 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12181 #ifdef BBR_INVARIANTS 12182 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12183 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12184 goto recheck_resend; 12185 #else 12186 /* TSNH */ 12187 rsm = NULL; 12188 goto recheck_resend; 12189 #endif 12190 } 12191 if (rsm->r_flags & BBR_HAS_SYN) { 12192 /* Only retransmit a SYN by itself */ 12193 len = 0; 12194 if ((flags & TH_SYN) == 0) { 12195 /* Huh something is wrong */ 12196 rsm->r_start++; 12197 if (rsm->r_start == rsm->r_end) { 12198 /* Clean it up, somehow we missed the ack? */ 12199 bbr_log_syn(tp, NULL); 12200 } else { 12201 /* TFO with data? */ 12202 rsm->r_flags &= ~BBR_HAS_SYN; 12203 len = rsm->r_end - rsm->r_start; 12204 } 12205 } else { 12206 /* Retransmitting SYN */ 12207 rsm = NULL; 12208 SOCKBUF_LOCK(sb); 12209 goto send; 12210 } 12211 } else 12212 len = rsm->r_end - rsm->r_start; 12213 if ((bbr->rc_resends_use_tso == 0) && 12214 (len > maxseg)) { 12215 len = maxseg; 12216 more_to_rxt = 1; 12217 } 12218 sb_offset = rsm->r_start - tp->snd_una; 12219 if (len > 0) { 12220 sack_rxmit = 1; 12221 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12222 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12223 min(len, maxseg)); 12224 } else { 12225 /* I dont think this can happen */ 12226 rsm = NULL; 12227 goto recheck_resend; 12228 } 12229 BBR_STAT_INC(bbr_resends_set); 12230 } else if (bbr->r_ctl.rc_tlp_send) { 12231 /* 12232 * Tail loss probe 12233 */ 12234 doing_tlp = 1; 12235 rsm = bbr->r_ctl.rc_tlp_send; 12236 bbr->r_ctl.rc_tlp_send = NULL; 12237 sack_rxmit = 1; 12238 len = rsm->r_end - rsm->r_start; 12239 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12240 len = maxseg; 12241 12242 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12243 #ifdef BBR_INVARIANTS 12244 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12245 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12246 #else 12247 /* TSNH */ 12248 rsm = NULL; 12249 goto recheck_resend; 12250 #endif 12251 } 12252 sb_offset = rsm->r_start - tp->snd_una; 12253 BBR_STAT_INC(bbr_tlp_set); 12254 } 12255 /* 12256 * Enforce a connection sendmap count limit if set 12257 * as long as we are not retransmiting. 12258 */ 12259 if ((rsm == NULL) && 12260 (V_tcp_map_entries_limit > 0) && 12261 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12262 BBR_STAT_INC(bbr_alloc_limited); 12263 if (!bbr->alloc_limit_reported) { 12264 bbr->alloc_limit_reported = 1; 12265 BBR_STAT_INC(bbr_alloc_limited_conns); 12266 } 12267 goto just_return_nolock; 12268 } 12269 #ifdef BBR_INVARIANTS 12270 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12271 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12272 tp, bbr, rsm, sb_offset, len); 12273 } 12274 #endif 12275 /* 12276 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12277 * state flags. 12278 */ 12279 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12280 flags |= TH_FIN; 12281 if (tp->t_flags & TF_NEEDSYN) 12282 flags |= TH_SYN; 12283 12284 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12285 /* we are retransmitting the fin */ 12286 len--; 12287 if (len) { 12288 /* 12289 * When retransmitting data do *not* include the 12290 * FIN. This could happen from a TLP probe if we 12291 * allowed data with a FIN. 12292 */ 12293 flags &= ~TH_FIN; 12294 } 12295 } else if (rsm) { 12296 if (flags & TH_FIN) 12297 flags &= ~TH_FIN; 12298 } 12299 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12300 void *end_rsm; 12301 12302 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12303 if (end_rsm) 12304 kern_prefetch(end_rsm, &prefetch_rsm); 12305 prefetch_rsm = 1; 12306 } 12307 SOCKBUF_LOCK(sb); 12308 /* 12309 * If snd_nxt == snd_max and we have transmitted a FIN, the 12310 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12311 * negative length. This can also occur when TCP opens up its 12312 * congestion window while receiving additional duplicate acks after 12313 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12314 * the fast-retransmit. 12315 * 12316 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12317 * set to snd_una, the sb_offset will be 0, and the length may wind 12318 * up 0. 12319 * 12320 * If sack_rxmit is true we are retransmitting from the scoreboard 12321 * in which case len is already set. 12322 */ 12323 if (sack_rxmit == 0) { 12324 uint32_t avail; 12325 12326 avail = sbavail(sb); 12327 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12328 sb_offset = tp->snd_max - tp->snd_una; 12329 else 12330 sb_offset = 0; 12331 if (bbr->rc_tlp_new_data) { 12332 /* TLP is forcing out new data */ 12333 uint32_t tlplen; 12334 12335 doing_tlp = 1; 12336 tlplen = maxseg; 12337 12338 if (tlplen > (uint32_t)(avail - sb_offset)) { 12339 tlplen = (uint32_t)(avail - sb_offset); 12340 } 12341 if (tlplen > tp->snd_wnd) { 12342 len = tp->snd_wnd; 12343 } else { 12344 len = tlplen; 12345 } 12346 bbr->rc_tlp_new_data = 0; 12347 } else { 12348 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12349 if ((len < p_maxseg) && 12350 (bbr->rc_in_persist == 0) && 12351 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12352 ((avail - sb_offset) >= p_maxseg)) { 12353 /* 12354 * We are not completing whats in the socket 12355 * buffer (i.e. there is at least a segment 12356 * waiting to send) and we have 2 or more 12357 * segments outstanding. There is no sense 12358 * of sending a little piece. Lets defer and 12359 * and wait until we can send a whole 12360 * segment. 12361 */ 12362 len = 0; 12363 } 12364 if (bbr->rc_in_persist) { 12365 /* 12366 * We are in persists, figure out if 12367 * a retransmit is available (maybe the previous 12368 * persists we sent) or if we have to send new 12369 * data. 12370 */ 12371 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12372 if (rsm) { 12373 len = rsm->r_end - rsm->r_start; 12374 if (rsm->r_flags & BBR_HAS_FIN) 12375 len--; 12376 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12377 len = maxseg; 12378 if (len > 1) 12379 BBR_STAT_INC(bbr_persist_reneg); 12380 /* 12381 * XXXrrs we could force the len to 12382 * 1 byte here to cause the chunk to 12383 * split apart.. but that would then 12384 * mean we always retransmit it as 12385 * one byte even after the window 12386 * opens. 12387 */ 12388 sack_rxmit = 1; 12389 sb_offset = rsm->r_start - tp->snd_una; 12390 } else { 12391 /* 12392 * First time through in persists or peer 12393 * acked our one byte. Though we do have 12394 * to have something in the sb. 12395 */ 12396 len = 1; 12397 sb_offset = 0; 12398 if (avail == 0) 12399 len = 0; 12400 } 12401 } 12402 } 12403 } 12404 if (prefetch_so_done == 0) { 12405 kern_prefetch(so, &prefetch_so_done); 12406 prefetch_so_done = 1; 12407 } 12408 /* 12409 * Lop off SYN bit if it has already been sent. However, if this is 12410 * SYN-SENT state and if segment contains data and if we don't know 12411 * that foreign host supports TAO, suppress sending segment. 12412 */ 12413 if ((flags & TH_SYN) && (rsm == NULL) && 12414 SEQ_GT(tp->snd_max, tp->snd_una)) { 12415 if (tp->t_state != TCPS_SYN_RECEIVED) 12416 flags &= ~TH_SYN; 12417 /* 12418 * When sending additional segments following a TFO SYN|ACK, 12419 * do not include the SYN bit. 12420 */ 12421 if (IS_FASTOPEN(tp->t_flags) && 12422 (tp->t_state == TCPS_SYN_RECEIVED)) 12423 flags &= ~TH_SYN; 12424 sb_offset--, len++; 12425 if (sbavail(sb) == 0) 12426 len = 0; 12427 } else if ((flags & TH_SYN) && rsm) { 12428 /* 12429 * Subtract one from the len for the SYN being 12430 * retransmitted. 12431 */ 12432 len--; 12433 } 12434 /* 12435 * Be careful not to send data and/or FIN on SYN segments. This 12436 * measure is needed to prevent interoperability problems with not 12437 * fully conformant TCP implementations. 12438 */ 12439 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12440 len = 0; 12441 flags &= ~TH_FIN; 12442 } 12443 /* 12444 * On TFO sockets, ensure no data is sent in the following cases: 12445 * 12446 * - When retransmitting SYN|ACK on a passively-created socket 12447 * - When retransmitting SYN on an actively created socket 12448 * - When sending a zero-length cookie (cookie request) on an 12449 * actively created socket 12450 * - When the socket is in the CLOSED state (RST is being sent) 12451 */ 12452 if (IS_FASTOPEN(tp->t_flags) && 12453 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12454 ((tp->t_state == TCPS_SYN_SENT) && 12455 (tp->t_tfo_client_cookie_len == 0)) || 12456 (flags & TH_RST))) { 12457 len = 0; 12458 sack_rxmit = 0; 12459 rsm = NULL; 12460 } 12461 /* Without fast-open there should never be data sent on a SYN */ 12462 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 12463 len = 0; 12464 if (len <= 0) { 12465 /* 12466 * If FIN has been sent but not acked, but we haven't been 12467 * called to retransmit, len will be < 0. Otherwise, window 12468 * shrank after we sent into it. If window shrank to 0, 12469 * cancel pending retransmit, pull snd_nxt back to (closed) 12470 * window, and set the persist timer if it isn't already 12471 * going. If the window didn't close completely, just wait 12472 * for an ACK. 12473 * 12474 * We also do a general check here to ensure that we will 12475 * set the persist timer when we have data to send, but a 12476 * 0-byte window. This makes sure the persist timer is set 12477 * even if the packet hits one of the "goto send" lines 12478 * below. 12479 */ 12480 len = 0; 12481 if ((tp->snd_wnd == 0) && 12482 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12483 (tp->snd_una == tp->snd_max) && 12484 (sb_offset < (int)sbavail(sb))) { 12485 /* 12486 * Not enough room in the rwnd to send 12487 * a paced segment out. 12488 */ 12489 bbr_enter_persist(tp, bbr, cts, __LINE__); 12490 } 12491 } else if ((rsm == NULL) && 12492 (doing_tlp == 0) && 12493 (len < bbr->r_ctl.rc_pace_max_segs)) { 12494 /* 12495 * We are not sending a full segment for 12496 * some reason. Should we not send anything (think 12497 * sws or persists)? 12498 */ 12499 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12500 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12501 (len < (int)(sbavail(sb) - sb_offset))) { 12502 /* 12503 * Here the rwnd is less than 12504 * the pacing size, this is not a retransmit, 12505 * we are established and 12506 * the send is not the last in the socket buffer 12507 * lets not send, and possibly enter persists. 12508 */ 12509 len = 0; 12510 if (tp->snd_max == tp->snd_una) 12511 bbr_enter_persist(tp, bbr, cts, __LINE__); 12512 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12513 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12514 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12515 (len < (int)(sbavail(sb) - sb_offset)) && 12516 (len < bbr_minseg(bbr))) { 12517 /* 12518 * Here we are not retransmitting, and 12519 * the cwnd is not so small that we could 12520 * not send at least a min size (rxt timer 12521 * not having gone off), We have 2 segments or 12522 * more already in flight, its not the tail end 12523 * of the socket buffer and the cwnd is blocking 12524 * us from sending out minimum pacing segment size. 12525 * Lets not send anything. 12526 */ 12527 bbr->rc_cwnd_limited = 1; 12528 len = 0; 12529 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12530 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12531 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12532 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12533 (len < (int)(sbavail(sb) - sb_offset)) && 12534 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12535 /* 12536 * Here we have a send window but we have 12537 * filled it up and we can't send another pacing segment. 12538 * We also have in flight more than 2 segments 12539 * and we are not completing the sb i.e. we allow 12540 * the last bytes of the sb to go out even if 12541 * its not a full pacing segment. 12542 */ 12543 len = 0; 12544 } 12545 } 12546 /* len will be >= 0 after this point. */ 12547 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12548 tcp_sndbuf_autoscale(tp, so, sendwin); 12549 /* 12550 * 12551 */ 12552 if (bbr->rc_in_persist && 12553 len && 12554 (rsm == NULL) && 12555 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12556 /* 12557 * We are in persist, not doing a retransmit and don't have enough space 12558 * yet to send a full TSO. So is it at the end of the sb 12559 * if so we need to send else nuke to 0 and don't send. 12560 */ 12561 int sbleft; 12562 if (sbavail(sb) > sb_offset) 12563 sbleft = sbavail(sb) - sb_offset; 12564 else 12565 sbleft = 0; 12566 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12567 /* not at end of sb lets not send */ 12568 len = 0; 12569 } 12570 } 12571 /* 12572 * Decide if we can use TCP Segmentation Offloading (if supported by 12573 * hardware). 12574 * 12575 * TSO may only be used if we are in a pure bulk sending state. The 12576 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12577 * options prevent using TSO. With TSO the TCP header is the same 12578 * (except for the sequence number) for all generated packets. This 12579 * makes it impossible to transmit any options which vary per 12580 * generated segment or packet. 12581 * 12582 * IPv4 handling has a clear separation of ip options and ip header 12583 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12584 * does the right thing below to provide length of just ip options 12585 * and thus checking for ipoptlen is enough to decide if ip options 12586 * are present. 12587 */ 12588 #ifdef INET6 12589 if (isipv6) 12590 ipoptlen = ip6_optlen(inp); 12591 else 12592 #endif 12593 if (inp->inp_options) 12594 ipoptlen = inp->inp_options->m_len - 12595 offsetof(struct ipoption, ipopt_list); 12596 else 12597 ipoptlen = 0; 12598 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12599 /* 12600 * Pre-calculate here as we save another lookup into the darknesses 12601 * of IPsec that way and can actually decide if TSO is ok. 12602 */ 12603 #ifdef INET6 12604 if (isipv6 && IPSEC_ENABLED(ipv6)) 12605 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12606 #ifdef INET 12607 else 12608 #endif 12609 #endif /* INET6 */ 12610 #ifdef INET 12611 if (IPSEC_ENABLED(ipv4)) 12612 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12613 #endif /* INET */ 12614 #endif /* IPSEC */ 12615 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12616 ipoptlen += ipsec_optlen; 12617 #endif 12618 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12619 (len > maxseg) && 12620 (tp->t_port == 0) && 12621 ((tp->t_flags & TF_SIGNATURE) == 0) && 12622 tp->rcv_numsacks == 0 && 12623 ipoptlen == 0) 12624 tso = 1; 12625 12626 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12627 (long)TCP_MAXWIN << tp->rcv_scale); 12628 /* 12629 * Sender silly window avoidance. We transmit under the following 12630 * conditions when len is non-zero: 12631 * 12632 * - We have a full segment (or more with TSO) - This is the last 12633 * buffer in a write()/send() and we are either idle or running 12634 * NODELAY - we've timed out (e.g. persist timer) - we have more 12635 * then 1/2 the maximum send window's worth of data (receiver may be 12636 * limited the window size) - we need to retransmit 12637 */ 12638 if (rsm) 12639 goto send; 12640 if (len) { 12641 if (sack_rxmit) 12642 goto send; 12643 if (len >= p_maxseg) 12644 goto send; 12645 /* 12646 * NOTE! on localhost connections an 'ack' from the remote 12647 * end may occur synchronously with the output and cause us 12648 * to flush a buffer queued with moretocome. XXX 12649 * 12650 */ 12651 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12652 ((tp->t_flags & TF_NODELAY) || 12653 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12654 (tp->t_flags & TF_NOPUSH) == 0) { 12655 goto send; 12656 } 12657 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12658 goto send; 12659 } 12660 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12661 goto send; 12662 } 12663 } 12664 /* 12665 * Sending of standalone window updates. 12666 * 12667 * Window updates are important when we close our window due to a 12668 * full socket buffer and are opening it again after the application 12669 * reads data from it. Once the window has opened again and the 12670 * remote end starts to send again the ACK clock takes over and 12671 * provides the most current window information. 12672 * 12673 * We must avoid the silly window syndrome whereas every read from 12674 * the receive buffer, no matter how small, causes a window update 12675 * to be sent. We also should avoid sending a flurry of window 12676 * updates when the socket buffer had queued a lot of data and the 12677 * application is doing small reads. 12678 * 12679 * Prevent a flurry of pointless window updates by only sending an 12680 * update when we can increase the advertized window by more than 12681 * 1/4th of the socket buffer capacity. When the buffer is getting 12682 * full or is very small be more aggressive and send an update 12683 * whenever we can increase by two mss sized segments. In all other 12684 * situations the ACK's to new incoming data will carry further 12685 * window increases. 12686 * 12687 * Don't send an independent window update if a delayed ACK is 12688 * pending (it will get piggy-backed on it) or the remote side 12689 * already has done a half-close and won't send more data. Skip 12690 * this if the connection is in T/TCP half-open state. 12691 */ 12692 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12693 !(tp->t_flags & TF_DELACK) && 12694 !TCPS_HAVERCVDFIN(tp->t_state)) { 12695 /* Check to see if we should do a window update */ 12696 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12697 goto send; 12698 } 12699 /* 12700 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12701 * is also a catch-all for the retransmit timer timeout case. 12702 */ 12703 if (tp->t_flags & TF_ACKNOW) { 12704 goto send; 12705 } 12706 if (flags & TH_RST) { 12707 /* Always send a RST if one is due */ 12708 goto send; 12709 } 12710 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12711 goto send; 12712 } 12713 /* 12714 * If our state indicates that FIN should be sent and we have not 12715 * yet done so, then we need to send. 12716 */ 12717 if (flags & TH_FIN && 12718 ((tp->t_flags & TF_SENTFIN) == 0)) { 12719 goto send; 12720 } 12721 /* 12722 * No reason to send a segment, just return. 12723 */ 12724 just_return: 12725 SOCKBUF_UNLOCK(sb); 12726 just_return_nolock: 12727 if (tot_len) 12728 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12729 if (bbr->rc_no_pacing) 12730 slot = 0; 12731 if (tot_len == 0) { 12732 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12733 tp->snd_wnd) { 12734 BBR_STAT_INC(bbr_rwnd_limited); 12735 app_limited = BBR_JR_RWND_LIMITED; 12736 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12737 if ((bbr->rc_in_persist == 0) && 12738 TCPS_HAVEESTABLISHED(tp->t_state) && 12739 (tp->snd_max == tp->snd_una) && 12740 sbavail(&so->so_snd)) { 12741 /* No send window.. we must enter persist */ 12742 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12743 } 12744 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12745 BBR_STAT_INC(bbr_app_limited); 12746 app_limited = BBR_JR_APP_LIMITED; 12747 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12748 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12749 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12750 BBR_STAT_INC(bbr_cwnd_limited); 12751 app_limited = BBR_JR_CWND_LIMITED; 12752 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12753 bbr->r_ctl.rc_lost_bytes))); 12754 bbr->rc_cwnd_limited = 1; 12755 } else { 12756 BBR_STAT_INC(bbr_app_limited); 12757 app_limited = BBR_JR_APP_LIMITED; 12758 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12759 } 12760 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12761 bbr->r_agg_early_set = 0; 12762 bbr->r_ctl.rc_agg_early = 0; 12763 bbr->r_ctl.rc_last_delay_val = 0; 12764 } else if (bbr->rc_use_google == 0) 12765 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12766 /* Are we app limited? */ 12767 if ((app_limited == BBR_JR_APP_LIMITED) || 12768 (app_limited == BBR_JR_RWND_LIMITED)) { 12769 /** 12770 * We are application limited. 12771 */ 12772 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12773 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12774 } 12775 if (tot_len == 0) 12776 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12777 /* Dont update the time if we did not send */ 12778 bbr->r_ctl.rc_last_delay_val = 0; 12779 bbr->rc_output_starts_timer = 1; 12780 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12781 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12782 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12783 /* Make sure snd_nxt is drug up */ 12784 tp->snd_nxt = tp->snd_max; 12785 } 12786 return (error); 12787 12788 send: 12789 if (doing_tlp == 0) { 12790 /* 12791 * Data not a TLP, and its not the rxt firing. If it is the 12792 * rxt firing, we want to leave the tlp_in_progress flag on 12793 * so we don't send another TLP. It has to be a rack timer 12794 * or normal send (response to acked data) to clear the tlp 12795 * in progress flag. 12796 */ 12797 bbr->rc_tlp_in_progress = 0; 12798 bbr->rc_tlp_rtx_out = 0; 12799 } else { 12800 /* 12801 * Its a TLP. 12802 */ 12803 bbr->rc_tlp_in_progress = 1; 12804 } 12805 bbr_timer_cancel(bbr, __LINE__, cts); 12806 if (rsm == NULL) { 12807 if (sbused(sb) > 0) { 12808 /* 12809 * This is sub-optimal. We only send a stand alone 12810 * FIN on its own segment. 12811 */ 12812 if (flags & TH_FIN) { 12813 flags &= ~TH_FIN; 12814 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12815 /* Lets not send this */ 12816 slot = 0; 12817 goto just_return; 12818 } 12819 } 12820 } 12821 } else { 12822 /* 12823 * We do *not* send a FIN on a retransmit if it has data. 12824 * The if clause here where len > 1 should never come true. 12825 */ 12826 if ((len > 0) && 12827 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12828 (flags & TH_FIN))) { 12829 flags &= ~TH_FIN; 12830 len--; 12831 } 12832 } 12833 SOCKBUF_LOCK_ASSERT(sb); 12834 if (len > 0) { 12835 if ((tp->snd_una == tp->snd_max) && 12836 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12837 /* 12838 * This qualifies as a RTT_PROBE session since we 12839 * drop the data outstanding to nothing and waited 12840 * more than bbr_rtt_probe_time. 12841 */ 12842 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12843 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12844 } 12845 if (len >= maxseg) 12846 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12847 else 12848 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12849 } 12850 /* 12851 * Before ESTABLISHED, force sending of initial options unless TCP 12852 * set not to do any options. NOTE: we assume that the IP/TCP header 12853 * plus TCP options always fit in a single mbuf, leaving room for a 12854 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12855 * + optlen <= MCLBYTES 12856 */ 12857 optlen = 0; 12858 #ifdef INET6 12859 if (isipv6) 12860 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12861 else 12862 #endif 12863 hdrlen = sizeof(struct tcpiphdr); 12864 12865 /* 12866 * Compute options for segment. We only have to care about SYN and 12867 * established connection segments. Options for SYN-ACK segments 12868 * are handled in TCP syncache. 12869 */ 12870 to.to_flags = 0; 12871 local_options = 0; 12872 if ((tp->t_flags & TF_NOOPT) == 0) { 12873 /* Maximum segment size. */ 12874 if (flags & TH_SYN) { 12875 to.to_mss = tcp_mssopt(&inp->inp_inc); 12876 if (tp->t_port) 12877 to.to_mss -= V_tcp_udp_tunneling_overhead; 12878 to.to_flags |= TOF_MSS; 12879 /* 12880 * On SYN or SYN|ACK transmits on TFO connections, 12881 * only include the TFO option if it is not a 12882 * retransmit, as the presence of the TFO option may 12883 * have caused the original SYN or SYN|ACK to have 12884 * been dropped by a middlebox. 12885 */ 12886 if (IS_FASTOPEN(tp->t_flags) && 12887 (tp->t_rxtshift == 0)) { 12888 if (tp->t_state == TCPS_SYN_RECEIVED) { 12889 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12890 to.to_tfo_cookie = 12891 (u_int8_t *)&tp->t_tfo_cookie.server; 12892 to.to_flags |= TOF_FASTOPEN; 12893 wanted_cookie = 1; 12894 } else if (tp->t_state == TCPS_SYN_SENT) { 12895 to.to_tfo_len = 12896 tp->t_tfo_client_cookie_len; 12897 to.to_tfo_cookie = 12898 tp->t_tfo_cookie.client; 12899 to.to_flags |= TOF_FASTOPEN; 12900 wanted_cookie = 1; 12901 } 12902 } 12903 } 12904 /* Window scaling. */ 12905 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12906 to.to_wscale = tp->request_r_scale; 12907 to.to_flags |= TOF_SCALE; 12908 } 12909 /* Timestamps. */ 12910 if ((tp->t_flags & TF_RCVD_TSTMP) || 12911 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12912 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12913 to.to_tsecr = tp->ts_recent; 12914 to.to_flags |= TOF_TS; 12915 local_options += TCPOLEN_TIMESTAMP + 2; 12916 } 12917 /* Set receive buffer autosizing timestamp. */ 12918 if (tp->rfbuf_ts == 0 && 12919 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12920 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12921 /* Selective ACK's. */ 12922 if (flags & TH_SYN) 12923 to.to_flags |= TOF_SACKPERM; 12924 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12925 tp->rcv_numsacks > 0) { 12926 to.to_flags |= TOF_SACK; 12927 to.to_nsacks = tp->rcv_numsacks; 12928 to.to_sacks = (u_char *)tp->sackblks; 12929 } 12930 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 12931 /* TCP-MD5 (RFC2385). */ 12932 if (tp->t_flags & TF_SIGNATURE) 12933 to.to_flags |= TOF_SIGNATURE; 12934 #endif /* TCP_SIGNATURE */ 12935 12936 /* Processing the options. */ 12937 hdrlen += (optlen = tcp_addoptions(&to, opt)); 12938 /* 12939 * If we wanted a TFO option to be added, but it was unable 12940 * to fit, ensure no data is sent. 12941 */ 12942 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 12943 !(to.to_flags & TOF_FASTOPEN)) 12944 len = 0; 12945 } 12946 if (tp->t_port) { 12947 if (V_tcp_udp_tunneling_port == 0) { 12948 /* The port was removed?? */ 12949 SOCKBUF_UNLOCK(&so->so_snd); 12950 return (EHOSTUNREACH); 12951 } 12952 hdrlen += sizeof(struct udphdr); 12953 } 12954 #ifdef INET6 12955 if (isipv6) 12956 ipoptlen = ip6_optlen(inp); 12957 else 12958 #endif 12959 if (inp->inp_options) 12960 ipoptlen = inp->inp_options->m_len - 12961 offsetof(struct ipoption, ipopt_list); 12962 else 12963 ipoptlen = 0; 12964 ipoptlen = 0; 12965 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12966 ipoptlen += ipsec_optlen; 12967 #endif 12968 if (bbr->rc_last_options != local_options) { 12969 /* 12970 * Cache the options length this generally does not change 12971 * on a connection. We use this to calculate TSO. 12972 */ 12973 bbr->rc_last_options = local_options; 12974 } 12975 maxseg = tp->t_maxseg - (ipoptlen + optlen); 12976 p_maxseg = min(maxseg, pace_max_segs); 12977 /* 12978 * Adjust data length if insertion of options will bump the packet 12979 * length beyond the t_maxseg length. Clear the FIN bit because we 12980 * cut off the tail of the segment. 12981 */ 12982 if (len > maxseg) { 12983 if (len != 0 && (flags & TH_FIN)) { 12984 flags &= ~TH_FIN; 12985 } 12986 if (tso) { 12987 uint32_t moff; 12988 int32_t max_len; 12989 12990 /* extract TSO information */ 12991 if_hw_tsomax = tp->t_tsomax; 12992 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 12993 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 12994 KASSERT(ipoptlen == 0, 12995 ("%s: TSO can't do IP options", __func__)); 12996 12997 /* 12998 * Check if we should limit by maximum payload 12999 * length: 13000 */ 13001 if (if_hw_tsomax != 0) { 13002 /* compute maximum TSO length */ 13003 max_len = (if_hw_tsomax - hdrlen - 13004 max_linkhdr); 13005 if (max_len <= 0) { 13006 len = 0; 13007 } else if (len > max_len) { 13008 len = max_len; 13009 } 13010 } 13011 /* 13012 * Prevent the last segment from being fractional 13013 * unless the send sockbuf can be emptied: 13014 */ 13015 if ((sb_offset + len) < sbavail(sb)) { 13016 moff = len % (uint32_t)maxseg; 13017 if (moff != 0) { 13018 len -= moff; 13019 } 13020 } 13021 /* 13022 * In case there are too many small fragments don't 13023 * use TSO: 13024 */ 13025 if (len <= maxseg) { 13026 len = maxseg; 13027 tso = 0; 13028 } 13029 } else { 13030 /* Not doing TSO */ 13031 if (optlen + ipoptlen >= tp->t_maxseg) { 13032 /* 13033 * Since we don't have enough space to put 13034 * the IP header chain and the TCP header in 13035 * one packet as required by RFC 7112, don't 13036 * send it. Also ensure that at least one 13037 * byte of the payload can be put into the 13038 * TCP segment. 13039 */ 13040 SOCKBUF_UNLOCK(&so->so_snd); 13041 error = EMSGSIZE; 13042 sack_rxmit = 0; 13043 goto out; 13044 } 13045 len = maxseg; 13046 } 13047 } else { 13048 /* Not doing TSO */ 13049 if_hw_tsomaxsegcount = 0; 13050 tso = 0; 13051 } 13052 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13053 ("%s: len > IP_MAXPACKET", __func__)); 13054 #ifdef DIAGNOSTIC 13055 #ifdef INET6 13056 if (max_linkhdr + hdrlen > MCLBYTES) 13057 #else 13058 if (max_linkhdr + hdrlen > MHLEN) 13059 #endif 13060 panic("tcphdr too big"); 13061 #endif 13062 /* 13063 * This KASSERT is here to catch edge cases at a well defined place. 13064 * Before, those had triggered (random) panic conditions further 13065 * down. 13066 */ 13067 #ifdef BBR_INVARIANTS 13068 if (sack_rxmit) { 13069 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13070 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13071 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13072 } 13073 } 13074 #endif 13075 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13076 if ((len == 0) && 13077 (flags & TH_FIN) && 13078 (sbused(sb))) { 13079 /* 13080 * We have outstanding data, don't send a fin by itself!. 13081 */ 13082 slot = 0; 13083 goto just_return; 13084 } 13085 /* 13086 * Grab a header mbuf, attaching a copy of data to be transmitted, 13087 * and initialize the header from the template for sends on this 13088 * connection. 13089 */ 13090 if (len) { 13091 uint32_t moff; 13092 13093 /* 13094 * We place a limit on sending with hptsi. 13095 */ 13096 if ((rsm == NULL) && len > pace_max_segs) 13097 len = pace_max_segs; 13098 if (len <= maxseg) 13099 tso = 0; 13100 #ifdef INET6 13101 if (MHLEN < hdrlen + max_linkhdr) 13102 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13103 else 13104 #endif 13105 m = m_gethdr(M_NOWAIT, MT_DATA); 13106 13107 if (m == NULL) { 13108 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13109 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13110 SOCKBUF_UNLOCK(sb); 13111 error = ENOBUFS; 13112 sack_rxmit = 0; 13113 goto out; 13114 } 13115 m->m_data += max_linkhdr; 13116 m->m_len = hdrlen; 13117 /* 13118 * Start the m_copy functions from the closest mbuf to the 13119 * sb_offset in the socket buffer chain. 13120 */ 13121 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13122 #ifdef BBR_INVARIANTS 13123 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13124 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13125 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13126 doing_retran_from, 13127 picked_up_retran, 13128 doing_tlp); 13129 13130 #endif 13131 /* 13132 * In this messed up situation we have two choices, 13133 * a) pretend the send worked, and just start timers 13134 * and what not (not good since that may lead us 13135 * back here a lot). <or> b) Send the lowest segment 13136 * in the map. <or> c) Drop the connection. Lets do 13137 * <b> which if it continues to happen will lead to 13138 * <c> via timeouts. 13139 */ 13140 BBR_STAT_INC(bbr_offset_recovery); 13141 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13142 sb_offset = 0; 13143 if (rsm == NULL) { 13144 sack_rxmit = 0; 13145 len = sbavail(sb); 13146 } else { 13147 sack_rxmit = 1; 13148 if (rsm->r_start != tp->snd_una) { 13149 /* 13150 * Things are really messed up, <c> 13151 * is the only thing to do. 13152 */ 13153 BBR_STAT_INC(bbr_offset_drop); 13154 SOCKBUF_UNLOCK(sb); 13155 (void)m_free(m); 13156 return (-EFAULT); /* tcp_drop() */ 13157 } 13158 len = rsm->r_end - rsm->r_start; 13159 } 13160 if (len > sbavail(sb)) 13161 len = sbavail(sb); 13162 if (len > maxseg) 13163 len = maxseg; 13164 } 13165 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13166 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13167 m_copydata(mb, moff, (int)len, 13168 mtod(m, caddr_t)+hdrlen); 13169 if (rsm == NULL) 13170 sbsndptr_adv(sb, mb, len); 13171 m->m_len += len; 13172 } else { 13173 struct sockbuf *msb; 13174 13175 if (rsm) 13176 msb = NULL; 13177 else 13178 msb = sb; 13179 #ifdef BBR_INVARIANTS 13180 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13181 if (rsm) { 13182 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 ", 13183 tp, bbr, len, moff, 13184 sbavail(sb), rsm, 13185 tp->snd_una, rsm->r_flags, rsm->r_start, 13186 doing_retran_from, 13187 picked_up_retran, 13188 doing_tlp, sack_rxmit); 13189 } else { 13190 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13191 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13192 } 13193 } 13194 #endif 13195 m->m_next = tcp_m_copym( 13196 mb, moff, &len, 13197 if_hw_tsomaxsegcount, 13198 if_hw_tsomaxsegsize, msb, 13199 ((rsm == NULL) ? hw_tls : 0) 13200 #ifdef NETFLIX_COPY_ARGS 13201 , &filled_all 13202 #endif 13203 ); 13204 if (len <= maxseg) { 13205 /* 13206 * Must have ran out of mbufs for the copy 13207 * shorten it to no longer need tso. Lets 13208 * not put on sendalot since we are low on 13209 * mbufs. 13210 */ 13211 tso = 0; 13212 } 13213 if (m->m_next == NULL) { 13214 SOCKBUF_UNLOCK(sb); 13215 (void)m_free(m); 13216 error = ENOBUFS; 13217 sack_rxmit = 0; 13218 goto out; 13219 } 13220 } 13221 #ifdef BBR_INVARIANTS 13222 if (tso && len < maxseg) { 13223 panic("tp:%p tso on, but len:%d < maxseg:%d", 13224 tp, len, maxseg); 13225 } 13226 if (tso && if_hw_tsomaxsegcount) { 13227 int32_t seg_cnt = 0; 13228 struct mbuf *foo; 13229 13230 foo = m; 13231 while (foo) { 13232 seg_cnt++; 13233 foo = foo->m_next; 13234 } 13235 if (seg_cnt > if_hw_tsomaxsegcount) { 13236 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13237 } 13238 } 13239 #endif 13240 /* 13241 * If we're sending everything we've got, set PUSH. (This 13242 * will keep happy those implementations which only give 13243 * data to the user when a buffer fills or a PUSH comes in.) 13244 */ 13245 if (sb_offset + len == sbused(sb) && 13246 sbused(sb) && 13247 !(flags & TH_SYN)) { 13248 flags |= TH_PUSH; 13249 } 13250 SOCKBUF_UNLOCK(sb); 13251 } else { 13252 SOCKBUF_UNLOCK(sb); 13253 if (tp->t_flags & TF_ACKNOW) 13254 KMOD_TCPSTAT_INC(tcps_sndacks); 13255 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13256 KMOD_TCPSTAT_INC(tcps_sndctrl); 13257 else 13258 KMOD_TCPSTAT_INC(tcps_sndwinup); 13259 13260 m = m_gethdr(M_NOWAIT, MT_DATA); 13261 if (m == NULL) { 13262 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13263 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13264 error = ENOBUFS; 13265 /* Fudge the send time since we could not send */ 13266 sack_rxmit = 0; 13267 goto out; 13268 } 13269 #ifdef INET6 13270 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13271 MHLEN >= hdrlen) { 13272 M_ALIGN(m, hdrlen); 13273 } else 13274 #endif 13275 m->m_data += max_linkhdr; 13276 m->m_len = hdrlen; 13277 } 13278 SOCKBUF_UNLOCK_ASSERT(sb); 13279 m->m_pkthdr.rcvif = (struct ifnet *)0; 13280 #ifdef MAC 13281 mac_inpcb_create_mbuf(inp, m); 13282 #endif 13283 #ifdef INET6 13284 if (isipv6) { 13285 ip6 = mtod(m, struct ip6_hdr *); 13286 if (tp->t_port) { 13287 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13288 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13289 udp->uh_dport = tp->t_port; 13290 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13291 udp->uh_ulen = htons(ulen); 13292 th = (struct tcphdr *)(udp + 1); 13293 } else { 13294 th = (struct tcphdr *)(ip6 + 1); 13295 } 13296 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13297 } else 13298 #endif /* INET6 */ 13299 { 13300 ip = mtod(m, struct ip *); 13301 #ifdef TCPDEBUG 13302 ipov = (struct ipovly *)ip; 13303 #endif 13304 if (tp->t_port) { 13305 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13306 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13307 udp->uh_dport = tp->t_port; 13308 ulen = hdrlen + len - sizeof(struct ip); 13309 udp->uh_ulen = htons(ulen); 13310 th = (struct tcphdr *)(udp + 1); 13311 } else { 13312 th = (struct tcphdr *)(ip + 1); 13313 } 13314 tcpip_fillheaders(inp, tp->t_port, ip, th); 13315 } 13316 /* 13317 * If we are doing retransmissions, then snd_nxt will not reflect 13318 * the first unsent octet. For ACK only packets, we do not want the 13319 * sequence number of the retransmitted packet, we want the sequence 13320 * number of the next unsent octet. So, if there is no data (and no 13321 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13322 * ti_seq. But if we are in persist state, snd_max might reflect 13323 * one byte beyond the right edge of the window, so use snd_nxt in 13324 * that case, since we know we aren't doing a retransmission. 13325 * (retransmit and persist are mutually exclusive...) 13326 */ 13327 if (sack_rxmit == 0) { 13328 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13329 /* New data (including new persists) */ 13330 th->th_seq = htonl(tp->snd_max); 13331 bbr_seq = tp->snd_max; 13332 } else if (flags & TH_SYN) { 13333 /* Syn's always send from iss */ 13334 th->th_seq = htonl(tp->iss); 13335 bbr_seq = tp->iss; 13336 } else if (flags & TH_FIN) { 13337 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13338 /* 13339 * If we sent the fin already its 1 minus 13340 * snd_max 13341 */ 13342 th->th_seq = (htonl(tp->snd_max - 1)); 13343 bbr_seq = (tp->snd_max - 1); 13344 } else { 13345 /* First time FIN use snd_max */ 13346 th->th_seq = htonl(tp->snd_max); 13347 bbr_seq = tp->snd_max; 13348 } 13349 } else { 13350 /* 13351 * len == 0 and not persist we use snd_max, sending 13352 * an ack unless we have sent the fin then its 1 13353 * minus. 13354 */ 13355 /* 13356 * XXXRRS Question if we are in persists and we have 13357 * nothing outstanding to send and we have not sent 13358 * a FIN, we will send an ACK. In such a case it 13359 * might be better to send (tp->snd_una - 1) which 13360 * would force the peer to ack. 13361 */ 13362 if (tp->t_flags & TF_SENTFIN) { 13363 th->th_seq = htonl(tp->snd_max - 1); 13364 bbr_seq = (tp->snd_max - 1); 13365 } else { 13366 th->th_seq = htonl(tp->snd_max); 13367 bbr_seq = tp->snd_max; 13368 } 13369 } 13370 } else { 13371 /* All retransmits use the rsm to guide the send */ 13372 th->th_seq = htonl(rsm->r_start); 13373 bbr_seq = rsm->r_start; 13374 } 13375 th->th_ack = htonl(tp->rcv_nxt); 13376 if (optlen) { 13377 bcopy(opt, th + 1, optlen); 13378 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13379 } 13380 tcp_set_flags(th, flags); 13381 /* 13382 * Calculate receive window. Don't shrink window, but avoid silly 13383 * window syndrome. 13384 */ 13385 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13386 recwin < maxseg))) 13387 recwin = 0; 13388 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13389 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13390 recwin = (tp->rcv_adv - tp->rcv_nxt); 13391 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13392 recwin = TCP_MAXWIN << tp->rcv_scale; 13393 13394 /* 13395 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13396 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13397 * handled in syncache. 13398 */ 13399 if (flags & TH_SYN) 13400 th->th_win = htons((u_short) 13401 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13402 else { 13403 /* Avoid shrinking window with window scaling. */ 13404 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13405 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13406 } 13407 /* 13408 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13409 * window. This may cause the remote transmitter to stall. This 13410 * flag tells soreceive() to disable delayed acknowledgements when 13411 * draining the buffer. This can occur if the receiver is 13412 * attempting to read more data than can be buffered prior to 13413 * transmitting on the connection. 13414 */ 13415 if (th->th_win == 0) { 13416 tp->t_sndzerowin++; 13417 tp->t_flags |= TF_RXWIN0SENT; 13418 } else 13419 tp->t_flags &= ~TF_RXWIN0SENT; 13420 /* 13421 * We don't support urgent data, but drag along 13422 * the pointer in case of a stack switch. 13423 */ 13424 tp->snd_up = tp->snd_una; 13425 13426 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13427 if (to.to_flags & TOF_SIGNATURE) { 13428 /* 13429 * Calculate MD5 signature and put it into the place 13430 * determined before. NOTE: since TCP options buffer doesn't 13431 * point into mbuf's data, calculate offset and use it. 13432 */ 13433 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13434 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13435 /* 13436 * Do not send segment if the calculation of MD5 13437 * digest has failed. 13438 */ 13439 goto out; 13440 } 13441 } 13442 #endif 13443 13444 /* 13445 * Put TCP length in extended header, and then checksum extended 13446 * header and data. 13447 */ 13448 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13449 #ifdef INET6 13450 if (isipv6) { 13451 /* 13452 * ip6_plen is not need to be filled now, and will be filled 13453 * in ip6_output. 13454 */ 13455 if (tp->t_port) { 13456 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13457 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13458 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13459 th->th_sum = htons(0); 13460 UDPSTAT_INC(udps_opackets); 13461 } else { 13462 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13463 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13464 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13465 optlen + len, IPPROTO_TCP, 0); 13466 } 13467 } 13468 #endif 13469 #if defined(INET6) && defined(INET) 13470 else 13471 #endif 13472 #ifdef INET 13473 { 13474 if (tp->t_port) { 13475 m->m_pkthdr.csum_flags = CSUM_UDP; 13476 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13477 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13478 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13479 th->th_sum = htons(0); 13480 UDPSTAT_INC(udps_opackets); 13481 } else { 13482 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13483 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13484 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13485 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13486 IPPROTO_TCP + len + optlen)); 13487 } 13488 /* IP version must be set here for ipv4/ipv6 checking later */ 13489 KASSERT(ip->ip_v == IPVERSION, 13490 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13491 } 13492 #endif 13493 13494 /* 13495 * Enable TSO and specify the size of the segments. The TCP pseudo 13496 * header checksum is always provided. XXX: Fixme: This is currently 13497 * not the case for IPv6. 13498 */ 13499 if (tso) { 13500 KASSERT(len > maxseg, 13501 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13502 m->m_pkthdr.csum_flags |= CSUM_TSO; 13503 csum_flags |= CSUM_TSO; 13504 m->m_pkthdr.tso_segsz = maxseg; 13505 } 13506 KASSERT(len + hdrlen == m_length(m, NULL), 13507 ("%s: mbuf chain different than expected: %d + %u != %u", 13508 __func__, len, hdrlen, m_length(m, NULL))); 13509 13510 #ifdef TCP_HHOOK 13511 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13512 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13513 #endif 13514 #ifdef TCPDEBUG 13515 /* 13516 * Trace. 13517 */ 13518 if (so->so_options & SO_DEBUG) { 13519 u_short save = 0; 13520 13521 #ifdef INET6 13522 if (!isipv6) 13523 #endif 13524 { 13525 save = ipov->ih_len; 13526 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 13527 * (th->th_off << 2) */ ); 13528 } 13529 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 13530 #ifdef INET6 13531 if (!isipv6) 13532 #endif 13533 ipov->ih_len = save; 13534 } 13535 #endif /* TCPDEBUG */ 13536 13537 /* Log to the black box */ 13538 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 13539 union tcp_log_stackspecific log; 13540 13541 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13542 /* Record info on type of transmission */ 13543 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13544 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13545 log.u_bbr.flex3 = maxseg; 13546 log.u_bbr.flex4 = delay_calc; 13547 /* Encode filled_all into the upper flex5 bit */ 13548 log.u_bbr.flex5 = bbr->rc_past_init_win; 13549 log.u_bbr.flex5 <<= 1; 13550 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13551 log.u_bbr.flex5 <<= 29; 13552 if (filled_all) 13553 log.u_bbr.flex5 |= 0x80000000; 13554 log.u_bbr.flex5 |= tp->t_maxseg; 13555 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13556 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13557 /* lets poke in the low and the high here for debugging */ 13558 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13559 if (rsm || sack_rxmit) { 13560 if (doing_tlp) 13561 log.u_bbr.flex8 = 2; 13562 else 13563 log.u_bbr.flex8 = 1; 13564 } else { 13565 log.u_bbr.flex8 = 0; 13566 } 13567 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13568 len, &log, false, NULL, NULL, 0, tv); 13569 } else { 13570 lgb = NULL; 13571 } 13572 /* 13573 * Fill in IP length and desired time to live and send to IP level. 13574 * There should be a better way to handle ttl and tos; we could keep 13575 * them in the template, but need a way to checksum without them. 13576 */ 13577 /* 13578 * m->m_pkthdr.len should have been set before cksum calcuration, 13579 * because in6_cksum() need it. 13580 */ 13581 #ifdef INET6 13582 if (isipv6) { 13583 /* 13584 * we separately set hoplimit for every segment, since the 13585 * user might want to change the value via setsockopt. Also, 13586 * desired default hop limit might be changed via Neighbor 13587 * Discovery. 13588 */ 13589 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13590 13591 /* 13592 * Set the packet size here for the benefit of DTrace 13593 * probes. ip6_output() will set it properly; it's supposed 13594 * to include the option header lengths as well. 13595 */ 13596 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13597 13598 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13599 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13600 else 13601 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13602 13603 if (tp->t_state == TCPS_SYN_SENT) 13604 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13605 13606 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13607 /* TODO: IPv6 IP6TOS_ECT bit on */ 13608 error = ip6_output(m, inp->in6p_outputopts, 13609 &inp->inp_route6, 13610 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13611 NULL, NULL, inp); 13612 13613 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13614 mtu = inp->inp_route6.ro_nh->nh_mtu; 13615 } 13616 #endif /* INET6 */ 13617 #if defined(INET) && defined(INET6) 13618 else 13619 #endif 13620 #ifdef INET 13621 { 13622 ip->ip_len = htons(m->m_pkthdr.len); 13623 #ifdef INET6 13624 if (isipv6) 13625 ip->ip_ttl = in6_selecthlim(inp, NULL); 13626 #endif /* INET6 */ 13627 /* 13628 * If we do path MTU discovery, then we set DF on every 13629 * packet. This might not be the best thing to do according 13630 * to RFC3390 Section 2. However the tcp hostcache migitates 13631 * the problem so it affects only the first tcp connection 13632 * with a host. 13633 * 13634 * NB: Don't set DF on small MTU/MSS to have a safe 13635 * fallback. 13636 */ 13637 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13638 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13639 if (tp->t_port == 0 || len < V_tcp_minmss) { 13640 ip->ip_off |= htons(IP_DF); 13641 } 13642 } else { 13643 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13644 } 13645 13646 if (tp->t_state == TCPS_SYN_SENT) 13647 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13648 13649 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13650 13651 error = ip_output(m, inp->inp_options, &inp->inp_route, 13652 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13653 inp); 13654 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13655 mtu = inp->inp_route.ro_nh->nh_mtu; 13656 } 13657 #endif /* INET */ 13658 out: 13659 13660 if (lgb) { 13661 lgb->tlb_errno = error; 13662 lgb = NULL; 13663 } 13664 /* 13665 * In transmit state, time the transmission and arrange for the 13666 * retransmit. In persist state, just set snd_max. 13667 */ 13668 if (error == 0) { 13669 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13670 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13671 (tp->t_flags & TF_SACK_PERMIT) && 13672 tp->rcv_numsacks > 0) 13673 tcp_clean_dsack_blocks(tp); 13674 /* We sent an ack clear the bbr_segs_rcvd count */ 13675 bbr->output_error_seen = 0; 13676 bbr->oerror_cnt = 0; 13677 bbr->bbr_segs_rcvd = 0; 13678 if (len == 0) 13679 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13680 /* Do accounting for new sends */ 13681 if ((len > 0) && (rsm == NULL)) { 13682 int idx; 13683 if (tp->snd_una == tp->snd_max) { 13684 /* 13685 * Special case to match google, when 13686 * nothing is in flight the delivered 13687 * time does get updated to the current 13688 * time (see tcp_rate_bsd.c). 13689 */ 13690 bbr->r_ctl.rc_del_time = cts; 13691 } 13692 if (len >= maxseg) { 13693 idx = (len / maxseg) + 3; 13694 if (idx >= TCP_MSS_ACCT_ATIMER) 13695 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13696 else 13697 counter_u64_add(bbr_out_size[idx], 1); 13698 } else { 13699 /* smaller than a MSS */ 13700 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13701 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13702 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13703 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13704 } 13705 } 13706 } 13707 abandon = 0; 13708 /* 13709 * We must do the send accounting before we log the output, 13710 * otherwise the state of the rsm could change and we account to the 13711 * wrong bucket. 13712 */ 13713 if (len > 0) { 13714 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13715 if (error == 0) { 13716 if (tp->snd_una == tp->snd_max) 13717 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13718 } 13719 } 13720 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13721 cts, mb, &abandon, rsm, 0, sb); 13722 if (abandon) { 13723 /* 13724 * If bbr_log_output destroys the TCB or sees a TH_RST being 13725 * sent we should hit this condition. 13726 */ 13727 return (0); 13728 } 13729 if (bbr->rc_in_persist == 0) { 13730 /* 13731 * Advance snd_nxt over sequence space of this segment. 13732 */ 13733 if (error) 13734 /* We don't log or do anything with errors */ 13735 goto skip_upd; 13736 13737 if (tp->snd_una == tp->snd_max && 13738 (len || (flags & (TH_SYN | TH_FIN)))) { 13739 /* 13740 * Update the time we just added data since none was 13741 * outstanding. 13742 */ 13743 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13744 bbr->rc_tp->t_acktime = ticks; 13745 } 13746 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13747 if (flags & TH_SYN) { 13748 /* 13749 * Smack the snd_max to iss + 1 13750 * if its a FO we will add len below. 13751 */ 13752 tp->snd_max = tp->iss + 1; 13753 } 13754 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13755 tp->snd_max++; 13756 tp->t_flags |= TF_SENTFIN; 13757 } 13758 } 13759 if (sack_rxmit == 0) 13760 tp->snd_max += len; 13761 skip_upd: 13762 if ((error == 0) && len) 13763 tot_len += len; 13764 } else { 13765 /* Persists case */ 13766 int32_t xlen = len; 13767 13768 if (error) 13769 goto nomore; 13770 13771 if (flags & TH_SYN) 13772 ++xlen; 13773 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13774 ++xlen; 13775 tp->t_flags |= TF_SENTFIN; 13776 } 13777 if (xlen && (tp->snd_una == tp->snd_max)) { 13778 /* 13779 * Update the time we just added data since none was 13780 * outstanding. 13781 */ 13782 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13783 bbr->rc_tp->t_acktime = ticks; 13784 } 13785 if (sack_rxmit == 0) 13786 tp->snd_max += xlen; 13787 tot_len += (len + optlen + ipoptlen); 13788 } 13789 nomore: 13790 if (error) { 13791 /* 13792 * Failures do not advance the seq counter above. For the 13793 * case of ENOBUFS we will fall out and become ack-clocked. 13794 * capping the cwnd at the current flight. 13795 * Everything else will just have to retransmit with the timer 13796 * (no pacer). 13797 */ 13798 SOCKBUF_UNLOCK_ASSERT(sb); 13799 BBR_STAT_INC(bbr_saw_oerr); 13800 /* Clear all delay/early tracks */ 13801 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13802 bbr->r_ctl.rc_agg_early = 0; 13803 bbr->r_agg_early_set = 0; 13804 bbr->output_error_seen = 1; 13805 if (bbr->oerror_cnt < 0xf) 13806 bbr->oerror_cnt++; 13807 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13808 /* drop the session */ 13809 return (-ENETDOWN); 13810 } 13811 switch (error) { 13812 case ENOBUFS: 13813 /* 13814 * Make this guy have to get ack's to send 13815 * more but lets make sure we don't 13816 * slam him below a T-O (1MSS). 13817 */ 13818 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13819 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13820 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13821 if (tp->snd_cwnd < maxseg) 13822 tp->snd_cwnd = maxseg; 13823 } 13824 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13825 BBR_STAT_INC(bbr_saw_enobuf); 13826 if (bbr->bbr_hdrw_pacing) 13827 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13828 else 13829 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13830 /* 13831 * Here even in the enobuf's case we want to do our 13832 * state update. The reason being we may have been 13833 * called by the input function. If so we have had 13834 * things change. 13835 */ 13836 error = 0; 13837 goto enobufs; 13838 case EMSGSIZE: 13839 /* 13840 * For some reason the interface we used initially 13841 * to send segments changed to another or lowered 13842 * its MTU. If TSO was active we either got an 13843 * interface without TSO capabilits or TSO was 13844 * turned off. If we obtained mtu from ip_output() 13845 * then update it and try again. 13846 */ 13847 /* Turn on tracing (or try to) */ 13848 { 13849 int old_maxseg; 13850 13851 old_maxseg = tp->t_maxseg; 13852 BBR_STAT_INC(bbr_saw_emsgsiz); 13853 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13854 if (mtu != 0) 13855 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13856 if (old_maxseg <= tp->t_maxseg) { 13857 /* Huh it did not shrink? */ 13858 tp->t_maxseg = old_maxseg - 40; 13859 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13860 } 13861 /* 13862 * Nuke all other things that can interfere 13863 * with slot 13864 */ 13865 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13866 slot = bbr_get_pacing_delay(bbr, 13867 bbr->r_ctl.rc_bbr_hptsi_gain, 13868 (tot_len + len), cts, 0); 13869 if (slot < bbr_error_base_paceout) 13870 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13871 } else 13872 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13873 bbr->rc_output_starts_timer = 1; 13874 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13875 tot_len); 13876 return (error); 13877 } 13878 case EPERM: 13879 tp->t_softerror = error; 13880 /* Fall through */ 13881 case EHOSTDOWN: 13882 case EHOSTUNREACH: 13883 case ENETDOWN: 13884 case ENETUNREACH: 13885 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13886 tp->t_softerror = error; 13887 } 13888 /* FALLTHROUGH */ 13889 default: 13890 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13891 bbr->rc_output_starts_timer = 1; 13892 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13893 return (error); 13894 } 13895 #ifdef STATS 13896 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13897 len && 13898 (rsm == NULL) && 13899 (bbr->rc_in_persist == 0)) { 13900 tp->gput_seq = bbr_seq; 13901 tp->gput_ack = bbr_seq + 13902 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13903 tp->gput_ts = cts; 13904 tp->t_flags |= TF_GPUTINPROG; 13905 #endif 13906 } 13907 KMOD_TCPSTAT_INC(tcps_sndtotal); 13908 if ((bbr->bbr_hdw_pace_ena) && 13909 (bbr->bbr_attempt_hdwr_pace == 0) && 13910 (bbr->rc_past_init_win) && 13911 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13912 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13913 (inp->inp_route.ro_nh && 13914 inp->inp_route.ro_nh->nh_ifp)) { 13915 /* 13916 * We are past the initial window and 13917 * have at least one measurement so we 13918 * could use hardware pacing if its available. 13919 * We have an interface and we have not attempted 13920 * to setup hardware pacing, lets try to now. 13921 */ 13922 uint64_t rate_wanted; 13923 int err = 0; 13924 13925 rate_wanted = bbr_get_hardware_rate(bbr); 13926 bbr->bbr_attempt_hdwr_pace = 1; 13927 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 13928 inp->inp_route.ro_nh->nh_ifp, 13929 rate_wanted, 13930 (RS_PACING_GEQ|RS_PACING_SUB_OK), 13931 &err, NULL); 13932 if (bbr->r_ctl.crte) { 13933 bbr_type_log_hdwr_pacing(bbr, 13934 bbr->r_ctl.crte->ptbl->rs_ifp, 13935 rate_wanted, 13936 bbr->r_ctl.crte->rate, 13937 __LINE__, cts, err); 13938 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 13939 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 13940 counter_u64_add(bbr_flows_whdwr_pacing, 1); 13941 bbr->bbr_hdrw_pacing = 1; 13942 /* Now what is our gain status? */ 13943 if (bbr->r_ctl.crte->rate < rate_wanted) { 13944 /* We have a problem */ 13945 bbr_setup_less_of_rate(bbr, cts, 13946 bbr->r_ctl.crte->rate, rate_wanted); 13947 } else { 13948 /* We are good */ 13949 bbr->gain_is_limited = 0; 13950 bbr->skip_gain = 0; 13951 } 13952 tcp_bbr_tso_size_check(bbr, cts); 13953 } else { 13954 bbr_type_log_hdwr_pacing(bbr, 13955 inp->inp_route.ro_nh->nh_ifp, 13956 rate_wanted, 13957 0, 13958 __LINE__, cts, err); 13959 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 13960 } 13961 } 13962 if (bbr->bbr_hdrw_pacing) { 13963 /* 13964 * Worry about cases where the route 13965 * changes or something happened that we 13966 * lost our hardware pacing possibly during 13967 * the last ip_output call. 13968 */ 13969 if (inp->inp_snd_tag == NULL) { 13970 /* A change during ip output disabled hw pacing? */ 13971 bbr->bbr_hdrw_pacing = 0; 13972 } else if ((inp->inp_route.ro_nh == NULL) || 13973 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 13974 /* 13975 * We had an interface or route change, 13976 * detach from the current hdwr pacing 13977 * and setup to re-attempt next go 13978 * round. 13979 */ 13980 bbr->bbr_hdrw_pacing = 0; 13981 bbr->bbr_attempt_hdwr_pace = 0; 13982 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 13983 tcp_bbr_tso_size_check(bbr, cts); 13984 } 13985 } 13986 /* 13987 * Data sent (as far as we can tell). If this advertises a larger 13988 * window than any other segment, then remember the size of the 13989 * advertised window. Any pending ACK has now been sent. 13990 */ 13991 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 13992 tp->rcv_adv = tp->rcv_nxt + recwin; 13993 13994 tp->last_ack_sent = tp->rcv_nxt; 13995 if ((error == 0) && 13996 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 13997 (doing_tlp == 0) && 13998 (tso == 0) && 13999 (len > 0) && 14000 ((flags & TH_RST) == 0) && 14001 ((flags & TH_SYN) == 0) && 14002 (IN_RECOVERY(tp->t_flags) == 0) && 14003 (bbr->rc_in_persist == 0) && 14004 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 14005 /* 14006 * For non-tso we need to goto again until we have sent out 14007 * enough data to match what we are hptsi out every hptsi 14008 * interval. 14009 */ 14010 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14011 /* Make sure snd_nxt is drug up */ 14012 tp->snd_nxt = tp->snd_max; 14013 } 14014 if (rsm != NULL) { 14015 rsm = NULL; 14016 goto skip_again; 14017 } 14018 rsm = NULL; 14019 sack_rxmit = 0; 14020 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14021 goto again; 14022 } 14023 skip_again: 14024 if ((error == 0) && (flags & TH_FIN)) 14025 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 14026 if ((error == 0) && (flags & TH_RST)) 14027 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 14028 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 14029 /* 14030 * Calculate/Re-Calculate the hptsi slot in usecs based on 14031 * what we have sent so far 14032 */ 14033 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 14034 if (bbr->rc_no_pacing) 14035 slot = 0; 14036 } 14037 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14038 enobufs: 14039 if (bbr->rc_use_google == 0) 14040 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 14041 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 14042 bbr->r_ctl.rc_lost_bytes))); 14043 bbr->rc_output_starts_timer = 1; 14044 if (bbr->bbr_use_rack_cheat && 14045 (more_to_rxt || 14046 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 14047 /* Rack cheats and shotguns out all rxt's 1ms apart */ 14048 if (slot > 1000) 14049 slot = 1000; 14050 } 14051 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 14052 /* 14053 * We don't change the tso size until some number of sends 14054 * to give the hardware commands time to get down 14055 * to the interface. 14056 */ 14057 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 14058 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 14059 bbr->hw_pacing_set = 1; 14060 tcp_bbr_tso_size_check(bbr, cts); 14061 } 14062 } 14063 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 14064 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14065 /* Make sure snd_nxt is drug up */ 14066 tp->snd_nxt = tp->snd_max; 14067 } 14068 return (error); 14069 14070 } 14071 14072 /* 14073 * See bbr_output_wtime() for return values. 14074 */ 14075 static int 14076 bbr_output(struct tcpcb *tp) 14077 { 14078 int32_t ret; 14079 struct timeval tv; 14080 14081 NET_EPOCH_ASSERT(); 14082 14083 INP_WLOCK_ASSERT(tptoinpcb(tp)); 14084 (void)tcp_get_usecs(&tv); 14085 ret = bbr_output_wtime(tp, &tv); 14086 return (ret); 14087 } 14088 14089 static void 14090 bbr_mtu_chg(struct tcpcb *tp) 14091 { 14092 struct tcp_bbr *bbr; 14093 struct bbr_sendmap *rsm, *frsm = NULL; 14094 uint32_t maxseg; 14095 14096 /* 14097 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14098 * over the current size as SACK_PASS so a retransmit will occur. 14099 */ 14100 14101 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14102 maxseg = tp->t_maxseg - bbr->rc_last_options; 14103 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14104 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14105 /* Don't mess with ones acked (by sack?) */ 14106 if (rsm->r_flags & BBR_ACKED) 14107 continue; 14108 if ((rsm->r_end - rsm->r_start) > maxseg) { 14109 /* 14110 * We mark sack-passed on all the previous large 14111 * sends we did. This will force them to retransmit. 14112 */ 14113 rsm->r_flags |= BBR_SACK_PASSED; 14114 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14115 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14116 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14117 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14118 rsm->r_flags |= BBR_MARKED_LOST; 14119 } 14120 if (frsm == NULL) 14121 frsm = rsm; 14122 } 14123 } 14124 if (frsm) { 14125 bbr->r_ctl.rc_resend = frsm; 14126 } 14127 } 14128 14129 static int 14130 bbr_pru_options(struct tcpcb *tp, int flags) 14131 { 14132 if (flags & PRUS_OOB) 14133 return (EOPNOTSUPP); 14134 return (0); 14135 } 14136 14137 struct tcp_function_block __tcp_bbr = { 14138 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14139 .tfb_tcp_output = bbr_output, 14140 .tfb_do_queued_segments = ctf_do_queued_segments, 14141 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14142 .tfb_tcp_do_segment = bbr_do_segment, 14143 .tfb_tcp_ctloutput = bbr_ctloutput, 14144 .tfb_tcp_fb_init = bbr_init, 14145 .tfb_tcp_fb_fini = bbr_fini, 14146 .tfb_tcp_timer_stop_all = bbr_stopall, 14147 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14148 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14149 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14150 .tfb_pru_options = bbr_pru_options, 14151 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14152 }; 14153 14154 /* 14155 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14156 * socket option arguments. When it re-acquires the lock after the copy, it 14157 * has to revalidate that the connection is still valid for the socket 14158 * option. 14159 */ 14160 static int 14161 bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt) 14162 { 14163 struct epoch_tracker et; 14164 struct tcpcb *tp; 14165 struct tcp_bbr *bbr; 14166 int32_t error = 0, optval; 14167 14168 switch (sopt->sopt_level) { 14169 case IPPROTO_IPV6: 14170 case IPPROTO_IP: 14171 return (tcp_default_ctloutput(inp, sopt)); 14172 } 14173 14174 switch (sopt->sopt_name) { 14175 case TCP_RACK_PACE_MAX_SEG: 14176 case TCP_RACK_MIN_TO: 14177 case TCP_RACK_REORD_THRESH: 14178 case TCP_RACK_REORD_FADE: 14179 case TCP_RACK_TLP_THRESH: 14180 case TCP_RACK_PKT_DELAY: 14181 case TCP_BBR_ALGORITHM: 14182 case TCP_BBR_TSLIMITS: 14183 case TCP_BBR_IWINTSO: 14184 case TCP_BBR_RECFORCE: 14185 case TCP_BBR_STARTUP_PG: 14186 case TCP_BBR_DRAIN_PG: 14187 case TCP_BBR_RWND_IS_APP: 14188 case TCP_BBR_PROBE_RTT_INT: 14189 case TCP_BBR_PROBE_RTT_GAIN: 14190 case TCP_BBR_PROBE_RTT_LEN: 14191 case TCP_BBR_STARTUP_LOSS_EXIT: 14192 case TCP_BBR_USEDEL_RATE: 14193 case TCP_BBR_MIN_RTO: 14194 case TCP_BBR_MAX_RTO: 14195 case TCP_BBR_PACE_PER_SEC: 14196 case TCP_DELACK: 14197 case TCP_BBR_PACE_DEL_TAR: 14198 case TCP_BBR_SEND_IWND_IN_TSO: 14199 case TCP_BBR_EXTRA_STATE: 14200 case TCP_BBR_UTTER_MAX_TSO: 14201 case TCP_BBR_MIN_TOPACEOUT: 14202 case TCP_BBR_FLOOR_MIN_TSO: 14203 case TCP_BBR_TSTMP_RAISES: 14204 case TCP_BBR_POLICER_DETECT: 14205 case TCP_BBR_USE_RACK_CHEAT: 14206 case TCP_DATA_AFTER_CLOSE: 14207 case TCP_BBR_HDWR_PACE: 14208 case TCP_BBR_PACE_SEG_MAX: 14209 case TCP_BBR_PACE_SEG_MIN: 14210 case TCP_BBR_PACE_CROSS: 14211 case TCP_BBR_PACE_OH: 14212 #ifdef NETFLIX_PEAKRATE 14213 case TCP_MAXPEAKRATE: 14214 #endif 14215 case TCP_BBR_TMR_PACE_OH: 14216 case TCP_BBR_RACK_RTT_USE: 14217 case TCP_BBR_RETRAN_WTSO: 14218 break; 14219 default: 14220 return (tcp_default_ctloutput(inp, sopt)); 14221 break; 14222 } 14223 INP_WUNLOCK(inp); 14224 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14225 if (error) 14226 return (error); 14227 INP_WLOCK(inp); 14228 if (inp->inp_flags & INP_DROPPED) { 14229 INP_WUNLOCK(inp); 14230 return (ECONNRESET); 14231 } 14232 tp = intotcpcb(inp); 14233 if (tp->t_fb != &__tcp_bbr) { 14234 INP_WUNLOCK(inp); 14235 return (ENOPROTOOPT); 14236 } 14237 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14238 switch (sopt->sopt_name) { 14239 case TCP_BBR_PACE_PER_SEC: 14240 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14241 bbr->r_ctl.bbr_hptsi_per_second = optval; 14242 break; 14243 case TCP_BBR_PACE_DEL_TAR: 14244 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14245 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14246 break; 14247 case TCP_BBR_PACE_SEG_MAX: 14248 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14249 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14250 break; 14251 case TCP_BBR_PACE_SEG_MIN: 14252 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14253 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14254 break; 14255 case TCP_BBR_PACE_CROSS: 14256 BBR_OPTS_INC(tcp_bbr_pace_cross); 14257 bbr->r_ctl.bbr_cross_over = optval; 14258 break; 14259 case TCP_BBR_ALGORITHM: 14260 BBR_OPTS_INC(tcp_bbr_algorithm); 14261 if (optval && (bbr->rc_use_google == 0)) { 14262 /* Turn on the google mode */ 14263 bbr_google_mode_on(bbr); 14264 if ((optval > 3) && (optval < 500)) { 14265 /* 14266 * Must be at least greater than .3% 14267 * and must be less than 50.0%. 14268 */ 14269 bbr->r_ctl.bbr_google_discount = optval; 14270 } 14271 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14272 /* Turn off the google mode */ 14273 bbr_google_mode_off(bbr); 14274 } 14275 break; 14276 case TCP_BBR_TSLIMITS: 14277 BBR_OPTS_INC(tcp_bbr_tslimits); 14278 if (optval == 1) 14279 bbr->rc_use_ts_limit = 1; 14280 else if (optval == 0) 14281 bbr->rc_use_ts_limit = 0; 14282 else 14283 error = EINVAL; 14284 break; 14285 14286 case TCP_BBR_IWINTSO: 14287 BBR_OPTS_INC(tcp_bbr_iwintso); 14288 if ((optval >= 0) && (optval < 128)) { 14289 uint32_t twin; 14290 14291 bbr->rc_init_win = optval; 14292 twin = bbr_initial_cwnd(bbr, tp); 14293 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14294 tp->snd_cwnd = twin; 14295 else 14296 error = EBUSY; 14297 } else 14298 error = EINVAL; 14299 break; 14300 case TCP_BBR_STARTUP_PG: 14301 BBR_OPTS_INC(tcp_bbr_startup_pg); 14302 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14303 bbr->r_ctl.rc_startup_pg = optval; 14304 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14305 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14306 } 14307 } else 14308 error = EINVAL; 14309 break; 14310 case TCP_BBR_DRAIN_PG: 14311 BBR_OPTS_INC(tcp_bbr_drain_pg); 14312 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14313 bbr->r_ctl.rc_drain_pg = optval; 14314 else 14315 error = EINVAL; 14316 break; 14317 case TCP_BBR_PROBE_RTT_LEN: 14318 BBR_OPTS_INC(tcp_bbr_probertt_len); 14319 if (optval <= 1) 14320 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14321 else 14322 error = EINVAL; 14323 break; 14324 case TCP_BBR_PROBE_RTT_GAIN: 14325 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14326 if (optval <= BBR_UNIT) 14327 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14328 else 14329 error = EINVAL; 14330 break; 14331 case TCP_BBR_PROBE_RTT_INT: 14332 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14333 if (optval > 1000) 14334 bbr->r_ctl.rc_probertt_int = optval; 14335 else 14336 error = EINVAL; 14337 break; 14338 case TCP_BBR_MIN_TOPACEOUT: 14339 BBR_OPTS_INC(tcp_bbr_topaceout); 14340 if (optval == 0) { 14341 bbr->no_pacing_until = 0; 14342 bbr->rc_no_pacing = 0; 14343 } else if (optval <= 0x00ff) { 14344 bbr->no_pacing_until = optval; 14345 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14346 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14347 /* Turn on no pacing */ 14348 bbr->rc_no_pacing = 1; 14349 } 14350 } else 14351 error = EINVAL; 14352 break; 14353 case TCP_BBR_STARTUP_LOSS_EXIT: 14354 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14355 bbr->rc_loss_exit = optval; 14356 break; 14357 case TCP_BBR_USEDEL_RATE: 14358 error = EINVAL; 14359 break; 14360 case TCP_BBR_MIN_RTO: 14361 BBR_OPTS_INC(tcp_bbr_min_rto); 14362 bbr->r_ctl.rc_min_rto_ms = optval; 14363 break; 14364 case TCP_BBR_MAX_RTO: 14365 BBR_OPTS_INC(tcp_bbr_max_rto); 14366 bbr->rc_max_rto_sec = optval; 14367 break; 14368 case TCP_RACK_MIN_TO: 14369 /* Minimum time between rack t-o's in ms */ 14370 BBR_OPTS_INC(tcp_rack_min_to); 14371 bbr->r_ctl.rc_min_to = optval; 14372 break; 14373 case TCP_RACK_REORD_THRESH: 14374 /* RACK reorder threshold (shift amount) */ 14375 BBR_OPTS_INC(tcp_rack_reord_thresh); 14376 if ((optval > 0) && (optval < 31)) 14377 bbr->r_ctl.rc_reorder_shift = optval; 14378 else 14379 error = EINVAL; 14380 break; 14381 case TCP_RACK_REORD_FADE: 14382 /* Does reordering fade after ms time */ 14383 BBR_OPTS_INC(tcp_rack_reord_fade); 14384 bbr->r_ctl.rc_reorder_fade = optval; 14385 break; 14386 case TCP_RACK_TLP_THRESH: 14387 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14388 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14389 if (optval) 14390 bbr->rc_tlp_threshold = optval; 14391 else 14392 error = EINVAL; 14393 break; 14394 case TCP_BBR_USE_RACK_CHEAT: 14395 BBR_OPTS_INC(tcp_use_rackcheat); 14396 if (bbr->rc_use_google) { 14397 error = EINVAL; 14398 break; 14399 } 14400 BBR_OPTS_INC(tcp_rack_cheat); 14401 if (optval) 14402 bbr->bbr_use_rack_cheat = 1; 14403 else 14404 bbr->bbr_use_rack_cheat = 0; 14405 break; 14406 case TCP_BBR_FLOOR_MIN_TSO: 14407 BBR_OPTS_INC(tcp_utter_max_tso); 14408 if ((optval >= 0) && (optval < 40)) 14409 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14410 else 14411 error = EINVAL; 14412 break; 14413 case TCP_BBR_UTTER_MAX_TSO: 14414 BBR_OPTS_INC(tcp_utter_max_tso); 14415 if ((optval >= 0) && (optval < 0xffff)) 14416 bbr->r_ctl.bbr_utter_max = optval; 14417 else 14418 error = EINVAL; 14419 break; 14420 14421 case TCP_BBR_EXTRA_STATE: 14422 BBR_OPTS_INC(tcp_extra_state); 14423 if (optval) 14424 bbr->rc_use_idle_restart = 1; 14425 else 14426 bbr->rc_use_idle_restart = 0; 14427 break; 14428 case TCP_BBR_SEND_IWND_IN_TSO: 14429 BBR_OPTS_INC(tcp_iwnd_tso); 14430 if (optval) { 14431 bbr->bbr_init_win_cheat = 1; 14432 if (bbr->rc_past_init_win == 0) { 14433 uint32_t cts; 14434 cts = tcp_get_usecs(&bbr->rc_tv); 14435 tcp_bbr_tso_size_check(bbr, cts); 14436 } 14437 } else 14438 bbr->bbr_init_win_cheat = 0; 14439 break; 14440 case TCP_BBR_HDWR_PACE: 14441 BBR_OPTS_INC(tcp_hdwr_pacing); 14442 if (optval){ 14443 bbr->bbr_hdw_pace_ena = 1; 14444 bbr->bbr_attempt_hdwr_pace = 0; 14445 } else { 14446 bbr->bbr_hdw_pace_ena = 0; 14447 #ifdef RATELIMIT 14448 if (bbr->r_ctl.crte != NULL) { 14449 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14450 bbr->r_ctl.crte = NULL; 14451 } 14452 #endif 14453 } 14454 break; 14455 14456 case TCP_DELACK: 14457 BBR_OPTS_INC(tcp_delack); 14458 if (optval < 100) { 14459 if (optval == 0) /* off */ 14460 tp->t_delayed_ack = 0; 14461 else if (optval == 1) /* on which is 2 */ 14462 tp->t_delayed_ack = 2; 14463 else /* higher than 2 and less than 100 */ 14464 tp->t_delayed_ack = optval; 14465 if (tp->t_flags & TF_DELACK) { 14466 tp->t_flags &= ~TF_DELACK; 14467 tp->t_flags |= TF_ACKNOW; 14468 NET_EPOCH_ENTER(et); 14469 bbr_output(tp); 14470 NET_EPOCH_EXIT(et); 14471 } 14472 } else 14473 error = EINVAL; 14474 break; 14475 case TCP_RACK_PKT_DELAY: 14476 /* RACK added ms i.e. rack-rtt + reord + N */ 14477 BBR_OPTS_INC(tcp_rack_pkt_delay); 14478 bbr->r_ctl.rc_pkt_delay = optval; 14479 break; 14480 #ifdef NETFLIX_PEAKRATE 14481 case TCP_MAXPEAKRATE: 14482 BBR_OPTS_INC(tcp_maxpeak); 14483 error = tcp_set_maxpeakrate(tp, optval); 14484 if (!error) 14485 tp->t_peakrate_thr = tp->t_maxpeakrate; 14486 break; 14487 #endif 14488 case TCP_BBR_RETRAN_WTSO: 14489 BBR_OPTS_INC(tcp_retran_wtso); 14490 if (optval) 14491 bbr->rc_resends_use_tso = 1; 14492 else 14493 bbr->rc_resends_use_tso = 0; 14494 break; 14495 case TCP_DATA_AFTER_CLOSE: 14496 BBR_OPTS_INC(tcp_data_ac); 14497 if (optval) 14498 bbr->rc_allow_data_af_clo = 1; 14499 else 14500 bbr->rc_allow_data_af_clo = 0; 14501 break; 14502 case TCP_BBR_POLICER_DETECT: 14503 BBR_OPTS_INC(tcp_policer_det); 14504 if (bbr->rc_use_google == 0) 14505 error = EINVAL; 14506 else if (optval) 14507 bbr->r_use_policer = 1; 14508 else 14509 bbr->r_use_policer = 0; 14510 break; 14511 14512 case TCP_BBR_TSTMP_RAISES: 14513 BBR_OPTS_INC(tcp_ts_raises); 14514 if (optval) 14515 bbr->ts_can_raise = 1; 14516 else 14517 bbr->ts_can_raise = 0; 14518 break; 14519 case TCP_BBR_TMR_PACE_OH: 14520 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14521 if (bbr->rc_use_google) { 14522 error = EINVAL; 14523 } else { 14524 if (optval) 14525 bbr->r_ctl.rc_incr_tmrs = 1; 14526 else 14527 bbr->r_ctl.rc_incr_tmrs = 0; 14528 } 14529 break; 14530 case TCP_BBR_PACE_OH: 14531 BBR_OPTS_INC(tcp_pacing_oh); 14532 if (bbr->rc_use_google) { 14533 error = EINVAL; 14534 } else { 14535 if (optval > (BBR_INCL_TCP_OH| 14536 BBR_INCL_IP_OH| 14537 BBR_INCL_ENET_OH)) { 14538 error = EINVAL; 14539 break; 14540 } 14541 if (optval & BBR_INCL_TCP_OH) 14542 bbr->r_ctl.rc_inc_tcp_oh = 1; 14543 else 14544 bbr->r_ctl.rc_inc_tcp_oh = 0; 14545 if (optval & BBR_INCL_IP_OH) 14546 bbr->r_ctl.rc_inc_ip_oh = 1; 14547 else 14548 bbr->r_ctl.rc_inc_ip_oh = 0; 14549 if (optval & BBR_INCL_ENET_OH) 14550 bbr->r_ctl.rc_inc_enet_oh = 1; 14551 else 14552 bbr->r_ctl.rc_inc_enet_oh = 0; 14553 } 14554 break; 14555 default: 14556 return (tcp_default_ctloutput(inp, sopt)); 14557 break; 14558 } 14559 #ifdef NETFLIX_STATS 14560 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14561 #endif 14562 INP_WUNLOCK(inp); 14563 return (error); 14564 } 14565 14566 /* 14567 * return 0 on success, error-num on failure 14568 */ 14569 static int 14570 bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt) 14571 { 14572 struct tcpcb *tp; 14573 struct tcp_bbr *bbr; 14574 int32_t error, optval; 14575 14576 tp = intotcpcb(inp); 14577 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14578 if (bbr == NULL) { 14579 INP_WUNLOCK(inp); 14580 return (EINVAL); 14581 } 14582 /* 14583 * Because all our options are either boolean or an int, we can just 14584 * pull everything into optval and then unlock and copy. If we ever 14585 * add a option that is not a int, then this will have quite an 14586 * impact to this routine. 14587 */ 14588 switch (sopt->sopt_name) { 14589 case TCP_BBR_PACE_PER_SEC: 14590 optval = bbr->r_ctl.bbr_hptsi_per_second; 14591 break; 14592 case TCP_BBR_PACE_DEL_TAR: 14593 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14594 break; 14595 case TCP_BBR_PACE_SEG_MAX: 14596 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14597 break; 14598 case TCP_BBR_MIN_TOPACEOUT: 14599 optval = bbr->no_pacing_until; 14600 break; 14601 case TCP_BBR_PACE_SEG_MIN: 14602 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14603 break; 14604 case TCP_BBR_PACE_CROSS: 14605 optval = bbr->r_ctl.bbr_cross_over; 14606 break; 14607 case TCP_BBR_ALGORITHM: 14608 optval = bbr->rc_use_google; 14609 break; 14610 case TCP_BBR_TSLIMITS: 14611 optval = bbr->rc_use_ts_limit; 14612 break; 14613 case TCP_BBR_IWINTSO: 14614 optval = bbr->rc_init_win; 14615 break; 14616 case TCP_BBR_STARTUP_PG: 14617 optval = bbr->r_ctl.rc_startup_pg; 14618 break; 14619 case TCP_BBR_DRAIN_PG: 14620 optval = bbr->r_ctl.rc_drain_pg; 14621 break; 14622 case TCP_BBR_PROBE_RTT_INT: 14623 optval = bbr->r_ctl.rc_probertt_int; 14624 break; 14625 case TCP_BBR_PROBE_RTT_LEN: 14626 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14627 break; 14628 case TCP_BBR_PROBE_RTT_GAIN: 14629 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14630 break; 14631 case TCP_BBR_STARTUP_LOSS_EXIT: 14632 optval = bbr->rc_loss_exit; 14633 break; 14634 case TCP_BBR_USEDEL_RATE: 14635 error = EINVAL; 14636 break; 14637 case TCP_BBR_MIN_RTO: 14638 optval = bbr->r_ctl.rc_min_rto_ms; 14639 break; 14640 case TCP_BBR_MAX_RTO: 14641 optval = bbr->rc_max_rto_sec; 14642 break; 14643 case TCP_RACK_PACE_MAX_SEG: 14644 /* Max segments in a pace */ 14645 optval = bbr->r_ctl.rc_pace_max_segs; 14646 break; 14647 case TCP_RACK_MIN_TO: 14648 /* Minimum time between rack t-o's in ms */ 14649 optval = bbr->r_ctl.rc_min_to; 14650 break; 14651 case TCP_RACK_REORD_THRESH: 14652 /* RACK reorder threshold (shift amount) */ 14653 optval = bbr->r_ctl.rc_reorder_shift; 14654 break; 14655 case TCP_RACK_REORD_FADE: 14656 /* Does reordering fade after ms time */ 14657 optval = bbr->r_ctl.rc_reorder_fade; 14658 break; 14659 case TCP_BBR_USE_RACK_CHEAT: 14660 /* Do we use the rack cheat for rxt */ 14661 optval = bbr->bbr_use_rack_cheat; 14662 break; 14663 case TCP_BBR_FLOOR_MIN_TSO: 14664 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14665 break; 14666 case TCP_BBR_UTTER_MAX_TSO: 14667 optval = bbr->r_ctl.bbr_utter_max; 14668 break; 14669 case TCP_BBR_SEND_IWND_IN_TSO: 14670 /* Do we send TSO size segments initially */ 14671 optval = bbr->bbr_init_win_cheat; 14672 break; 14673 case TCP_BBR_EXTRA_STATE: 14674 optval = bbr->rc_use_idle_restart; 14675 break; 14676 case TCP_RACK_TLP_THRESH: 14677 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14678 optval = bbr->rc_tlp_threshold; 14679 break; 14680 case TCP_RACK_PKT_DELAY: 14681 /* RACK added ms i.e. rack-rtt + reord + N */ 14682 optval = bbr->r_ctl.rc_pkt_delay; 14683 break; 14684 case TCP_BBR_RETRAN_WTSO: 14685 optval = bbr->rc_resends_use_tso; 14686 break; 14687 case TCP_DATA_AFTER_CLOSE: 14688 optval = bbr->rc_allow_data_af_clo; 14689 break; 14690 case TCP_DELACK: 14691 optval = tp->t_delayed_ack; 14692 break; 14693 case TCP_BBR_HDWR_PACE: 14694 optval = bbr->bbr_hdw_pace_ena; 14695 break; 14696 case TCP_BBR_POLICER_DETECT: 14697 optval = bbr->r_use_policer; 14698 break; 14699 case TCP_BBR_TSTMP_RAISES: 14700 optval = bbr->ts_can_raise; 14701 break; 14702 case TCP_BBR_TMR_PACE_OH: 14703 optval = bbr->r_ctl.rc_incr_tmrs; 14704 break; 14705 case TCP_BBR_PACE_OH: 14706 optval = 0; 14707 if (bbr->r_ctl.rc_inc_tcp_oh) 14708 optval |= BBR_INCL_TCP_OH; 14709 if (bbr->r_ctl.rc_inc_ip_oh) 14710 optval |= BBR_INCL_IP_OH; 14711 if (bbr->r_ctl.rc_inc_enet_oh) 14712 optval |= BBR_INCL_ENET_OH; 14713 break; 14714 default: 14715 return (tcp_default_ctloutput(inp, sopt)); 14716 break; 14717 } 14718 INP_WUNLOCK(inp); 14719 error = sooptcopyout(sopt, &optval, sizeof optval); 14720 return (error); 14721 } 14722 14723 /* 14724 * return 0 on success, error-num on failure 14725 */ 14726 static int 14727 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt) 14728 { 14729 if (sopt->sopt_dir == SOPT_SET) { 14730 return (bbr_set_sockopt(inp, sopt)); 14731 } else if (sopt->sopt_dir == SOPT_GET) { 14732 return (bbr_get_sockopt(inp, sopt)); 14733 } else { 14734 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14735 } 14736 } 14737 14738 static const char *bbr_stack_names[] = { 14739 __XSTRING(STACKNAME), 14740 #ifdef STACKALIAS 14741 __XSTRING(STACKALIAS), 14742 #endif 14743 }; 14744 14745 static bool bbr_mod_inited = false; 14746 14747 static int 14748 tcp_addbbr(module_t mod, int32_t type, void *data) 14749 { 14750 int32_t err = 0; 14751 int num_stacks; 14752 14753 switch (type) { 14754 case MOD_LOAD: 14755 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14756 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14757 sizeof(struct bbr_sendmap), 14758 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14759 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14760 sizeof(struct tcp_bbr), 14761 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14762 sysctl_ctx_init(&bbr_sysctl_ctx); 14763 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14764 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14765 OID_AUTO, 14766 #ifdef STACKALIAS 14767 __XSTRING(STACKALIAS), 14768 #else 14769 __XSTRING(STACKNAME), 14770 #endif 14771 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14772 ""); 14773 if (bbr_sysctl_root == NULL) { 14774 printf("Failed to add sysctl node\n"); 14775 err = EFAULT; 14776 goto free_uma; 14777 } 14778 bbr_init_sysctls(); 14779 num_stacks = nitems(bbr_stack_names); 14780 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14781 bbr_stack_names, &num_stacks); 14782 if (err) { 14783 printf("Failed to register %s stack name for " 14784 "%s module\n", bbr_stack_names[num_stacks], 14785 __XSTRING(MODNAME)); 14786 sysctl_ctx_free(&bbr_sysctl_ctx); 14787 free_uma: 14788 uma_zdestroy(bbr_zone); 14789 uma_zdestroy(bbr_pcb_zone); 14790 bbr_counter_destroy(); 14791 printf("Failed to register " __XSTRING(MODNAME) 14792 " module err:%d\n", err); 14793 return (err); 14794 } 14795 tcp_lro_reg_mbufq(); 14796 bbr_mod_inited = true; 14797 printf(__XSTRING(MODNAME) " is now available\n"); 14798 break; 14799 case MOD_QUIESCE: 14800 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14801 break; 14802 case MOD_UNLOAD: 14803 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14804 if (err == EBUSY) 14805 break; 14806 if (bbr_mod_inited) { 14807 uma_zdestroy(bbr_zone); 14808 uma_zdestroy(bbr_pcb_zone); 14809 sysctl_ctx_free(&bbr_sysctl_ctx); 14810 bbr_counter_destroy(); 14811 printf(__XSTRING(MODNAME) 14812 " is now no longer available\n"); 14813 bbr_mod_inited = false; 14814 } 14815 tcp_lro_dereg_mbufq(); 14816 err = 0; 14817 break; 14818 default: 14819 return (EOPNOTSUPP); 14820 } 14821 return (err); 14822 } 14823 14824 static moduledata_t tcp_bbr = { 14825 .name = __XSTRING(MODNAME), 14826 .evhand = tcp_addbbr, 14827 .priv = 0 14828 }; 14829 14830 MODULE_VERSION(MODNAME, 1); 14831 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14832 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14833