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 * $FreeBSD$ 26 */ 27 28#ifndef _NETINET_TCP_BBR_H_ 29#define _NETINET_TCP_BBR_H_ 30 31#define BBR_INITIAL_RTO 1000000 /* 1 second in micro-seconds */ 32/* Send map flags */ 33#define BBR_ACKED 0x0001 /* The remote endpoint acked this */ 34#define BBR_WAS_RENEGED 0x0002 /* The peer reneged the ack */ 35#define BBR_RXT_CLEARED 0x0004 /* ACK Cleared by the RXT timer */ 36#define BBR_OVERMAX 0x0008 /* We have more retran's then we can 37 * fit */ 38#define BBR_SACK_PASSED 0x0010 /* A sack was done above this block */ 39#define BBR_WAS_SACKPASS 0x0020 /* We retransmitted due to SACK pass */ 40#define BBR_HAS_FIN 0x0040 /* segment is sent with fin */ 41#define BBR_TLP 0x0080 /* segment sent as tail-loss-probe */ 42#define BBR_HAS_SYN 0x0100 /* segment has the syn */ 43#define BBR_MARKED_LOST 0x0200 /* 44 * This segments is lost and 45 * totaled into bbr->rc_ctl.rc_lost 46 */ 47#define BBR_RWND_COLLAPSED 0x0400 /* The peer collapsed the rwnd on the segment */ 48#define BBR_NUM_OF_RETRANS 7 49 50/* Defines for socket options to set pacing overheads */ 51#define BBR_INCL_ENET_OH 0x01 52#define BBR_INCL_IP_OH 0x02 53#define BBR_INCL_TCP_OH 0x03 54 55/* 56 * With the addition of both measurement algorithms 57 * I had to move over the size of a 58 * cache line (unfortunately). For now there is 59 * no way around this. We may be able to cut back 60 * at some point I hope. 61 */ 62struct bbr_sendmap { 63 TAILQ_ENTRY(bbr_sendmap) r_next; /* seq number arrayed next */ 64 TAILQ_ENTRY(bbr_sendmap) r_tnext; /* Time of tmit based next */ 65 uint32_t r_start; /* Sequence number of the segment */ 66 uint32_t r_end; /* End seq, this is 1 beyond actually */ 67 68 uint32_t r_rtr_bytes; /* How many bytes have been retransmitted */ 69 uint32_t r_delivered; /* Delivered amount at send */ 70 71 uint32_t r_del_time; /* The time of the last delivery update */ 72 uint8_t r_rtr_cnt:4, /* Retran count, index this -1 to get time 73 * sent */ 74 r_rtt_not_allowed:1, /* No rtt measurement allowed */ 75 r_is_drain:1, /* In a draining cycle */ 76 r_app_limited:1,/* We went app limited */ 77 r_ts_valid:1; /* Timestamp field is valid (r_del_ack_ts) */ 78 uint8_t r_dupack; /* Dup ack count */ 79 uint8_t r_in_tmap:1, /* Flag to see if its in the r_tnext array */ 80 r_is_smallmap:1,/* Was logged as a small-map send-map item */ 81 r_is_gain:1, /* Was in gain cycle */ 82 r_bbr_state:5; /* The BBR state at send */ 83 uint8_t r_limit_type; /* is this entry counted against a limit? */ 84 85 uint16_t r_flags; /* Flags as defined above */ 86 uint16_t r_spare16; 87 uint32_t r_del_ack_ts; /* At send what timestamp of peer was (if r_ts_valid set) */ 88 /****************Cache line*****************/ 89 uint32_t r_tim_lastsent[BBR_NUM_OF_RETRANS]; 90 /* 91 * Question, should we instead just grab the sending b/w 92 * from the filter with the gain and store it in a 93 * uint64_t instead? 94 */ 95 uint32_t r_first_sent_time; /* Time of first pkt in flight sent */ 96 uint32_t r_pacing_delay; /* pacing delay of this send */ 97 uint32_t r_flight_at_send; /* flight at the time of the send */ 98#ifdef _KERNEL 99} __aligned(CACHE_LINE_SIZE); 100#else 101}; 102#endif 103#define BBR_LIMIT_TYPE_SPLIT 1 104 105TAILQ_HEAD(bbr_head, bbr_sendmap); 106 107#define BBR_SEGMENT_TIME_SIZE 1500 /* How many bytes in time_between */ 108 109#define BBR_MIN_SEG 1460 /* MSS size */ 110#define BBR_MAX_GAIN_VALUE 0xffff 111 112#define BBR_TIMER_FUDGE 1500 /* 1.5ms in micro seconds */ 113 114/* BW twiddle secret codes */ 115#define BBR_RED_BW_CONGSIG 0 /* We enter recovery and set using b/w */ 116#define BBR_RED_BW_RATECAL 1 /* We are calculating the loss rate */ 117#define BBR_RED_BW_USELRBW 2 /* We are dropping the lower b/w with 118 * cDR */ 119#define BBR_RED_BW_SETHIGHLOSS 3 /* We have set our highloss value at 120 * exit from probe-rtt */ 121#define BBR_RED_BW_PE_CLREARLY 4 /* We have decided to clear the 122 * reduction early */ 123#define BBR_RED_BW_PE_CLAFDEL 5 /* We are clearing it on schedule 124 * delayed */ 125#define BBR_RED_BW_REC_ENDCLL 6 /* Recover exits save high if needed 126 * an clear to start measuring */ 127#define BBR_RED_BW_PE_NOEARLY_OUT 7 /* Set pkt epoch judged that we do not 128 * get out of jail early */ 129/* For calculating a rate */ 130#define BBR_CALC_BW 1 131#define BBR_CALC_LOSS 2 132 133#define BBR_RTT_BY_TIMESTAMP 0 134#define BBR_RTT_BY_EXACTMATCH 1 135#define BBR_RTT_BY_EARLIER_RET 2 136#define BBR_RTT_BY_THIS_RETRAN 3 137#define BBR_RTT_BY_SOME_RETRAN 4 138#define BBR_RTT_BY_TSMATCHING 5 139 140/* Markers to track where we enter persists from */ 141#define BBR_PERSISTS_FROM_1 1 142#define BBR_PERSISTS_FROM_2 2 143#define BBR_PERSISTS_FROM_3 3 144#define BBR_PERSISTS_FROM_4 4 145#define BBR_PERSISTS_FROM_5 5 146 147/* magic cookies to ask for the RTT */ 148#define BBR_RTT_PROP 0 149#define BBR_RTT_RACK 1 150#define BBR_RTT_PKTRTT 2 151#define BBR_SRTT 3 152 153#define BBR_SACKED 0 154#define BBR_CUM_ACKED 1 155 156/* threshold in useconds where we consider we need a higher min cwnd */ 157#define BBR_HIGH_SPEED 1000 158#define BBR_HIGHSPEED_NUM_MSS 12 159 160#define MAX_REDUCE_RXT 3 /* What is the maximum times we are willing to 161 * reduce b/w in RTX's. Setting this has a 162 * multiplicative effect e.g. if we are 163 * reducing by 20% then setting it to 3 means 164 * you will have reduced the b/w estimate by > 165 * 60% before you stop. */ 166/* 167 * We use the rate sample structure to 168 * assist in single sack/ack rate and rtt 169 * calculation. In the future we will expand 170 * this in BBR to do forward rate sample 171 * b/w estimation. 172 */ 173#define BBR_RS_RTT_EMPTY 0x00000001 /* Nothing yet stored in RTT's */ 174#define BBR_RS_BW_EMPTY 0x00000002 /* Nothing yet stored in cDR */ 175#define BBR_RS_RTT_VALID 0x00000004 /* We have at least one valid RTT */ 176#define BBR_RS_BW_VAILD 0x00000008 /* We have a valid cDR */ 177#define BBR_RS_EMPTY (BBR_RS_RTT_EMPTY|BBR_RS_BW_EMPTY) 178struct bbr_rtt_sample { 179 uint32_t rs_flags; 180 uint32_t rs_rtt_lowest; 181 uint32_t rs_rtt_lowest_sendtime; 182 uint32_t rs_rtt_low_seq_start; 183 184 uint32_t rs_rtt_highest; 185 uint32_t rs_rtt_cnt; 186 187 uint64_t rs_rtt_tot; 188 uint32_t cur_rtt; 189 uint32_t cur_rtt_bytecnt; 190 191 uint32_t cur_rtt_rsmcnt; 192 uint32_t rc_crtt_set:1, 193 avail_bits:31; 194 uint64_t rs_cDR; 195}; 196 197/* RTT shrink reasons */ 198#define BBR_RTTS_INIT 0 199#define BBR_RTTS_NEWRTT 1 200#define BBR_RTTS_RTTPROBE 2 201#define BBR_RTTS_WASIDLE 3 202#define BBR_RTTS_PERSIST 4 203#define BBR_RTTS_REACHTAR 5 204#define BBR_RTTS_ENTERPROBE 6 205#define BBR_RTTS_SHRINK_PG 7 206#define BBR_RTTS_SHRINK_PG_FINAL 8 207#define BBR_RTTS_NEW_TARGET 9 208#define BBR_RTTS_LEAVE_DRAIN 10 209#define BBR_RTTS_RESETS_VALUES 11 210 211#define BBR_NUM_RATES 5 212/* Rate flags */ 213#define BBR_RT_FLAG_FREE 0x00 /* Is on the free list */ 214#define BBR_RT_FLAG_INUSE 0x01 /* Has been allocated */ 215#define BBR_RT_FLAG_READY 0x02 /* Ready to initiate a measurement. */ 216#define BBR_RT_FLAG_CAPPED_PRE 0x04 /* Ready to cap if we send the next segment */ 217#define BBR_RT_FLAG_CAPPED 0x08 /* Measurement is capped */ 218#define BBR_RT_FLAG_PASTFA 0x10 /* Past the first ack. */ 219#define BBR_RT_FLAG_LIMITED 0x20 /* Saw application/cwnd or rwnd limited period */ 220#define BBR_RT_SEEN_A_ACK 0x40 /* A ack has been saved */ 221#define BBR_RT_PREV_RTT_SET 0x80 /* There was a RTT set in */ 222#define BBR_RT_PREV_SEND_TIME 0x100 /* 223 *There was a RTT send time set that can be used 224 * no snd_limits 225 */ 226#define BBR_RT_SET_GRADIENT 0x200 227#define BBR_RT_TS_VALID 0x400 228 229struct bbr_log { 230 union { 231 struct bbr_sendmap *rsm; /* For alloc/free */ 232 uint64_t sb_acc; /* For out/ack or t-o */ 233 }; 234 struct tcpcb *tp; 235 uint32_t t_flags; 236 uint32_t th_seq; 237 uint32_t th_ack; 238 uint32_t snd_una; 239 uint32_t snd_nxt; 240 uint32_t snd_max; 241 uint32_t snd_cwnd; 242 uint32_t snd_wnd; 243 uint32_t rc_lost; 244 uint32_t target_cwnd; /* UU */ 245 uint32_t inflight; /* UU */ 246 uint32_t applimited; /* UU */ 247 /* Things for BBR */ 248 uint32_t delivered; /* UU */ 249 uint64_t cur_del_rate; /* UU */ 250 uint64_t delRate; /* UU */ 251 uint64_t rttProp; /* UU */ 252 uint64_t lt_bw; /* UU */ 253 uint32_t timeStamp; 254 uint32_t time; 255 uint32_t slot; /* UU */ 256 uint32_t delayed_by; 257 uint32_t exp_del; 258 uint32_t pkts_out; 259 uint32_t new_win; 260 uint32_t hptsi_gain; /* UU */ 261 uint32_t cwnd_gain; /* UU */ 262 uint32_t epoch; /* UU */ 263 uint32_t lt_epoch; /* UU */ 264 /* Sack fun */ 265 uint32_t blk_start[4]; /* xx */ 266 uint32_t blk_end[4]; 267 uint32_t len; /* Timeout T3=1, TLP=2, RACK=3 */ 268 uint8_t type; 269 uint8_t n_sackblks; 270 uint8_t applied; /* UU */ 271 uint8_t inhpts; /* UU */ 272 uint8_t ininput; /* UU */ 273 uint8_t use_lt_bw; /* UU */ 274}; 275 276struct bbr_log_sysctl_out { 277 uint32_t bbr_log_at; 278 uint32_t bbr_log_max; 279 struct bbr_log entries[0]; 280}; 281 282/* 283 * Magic numbers for logging timeout events if the 284 * logging is enabled. 285 */ 286#define BBR_TO_FRM_TMR 1 287#define BBR_TO_FRM_TLP 2 288#define BBR_TO_FRM_RACK 3 289#define BBR_TO_FRM_KEEP 4 290#define BBR_TO_FRM_PERSIST 5 291#define BBR_TO_FRM_DELACK 6 292 293#define BBR_SEES_STRETCH_ACK 1 294#define BBR_SEES_COMPRESSED_ACKS 2 295 296/* 297 * As we get each SACK we wade through the 298 * rc_map and mark off what is acked. 299 * We also increment rc_sacked as well. 300 * 301 * We also pay attention to missing entries 302 * based on the time and possibly mark them 303 * for retransmit. If we do and we are not already 304 * in recovery we enter recovery. In doing 305 * so we claer prr_delivered/holes_rxt and prr_sent_dur_rec. 306 * We also setup rc_next/rc_snd_nxt/rc_send_end so 307 * we will know where to send from. When not in 308 * recovery rc_next will be NULL and rc_snd_nxt should 309 * equal snd_max. 310 * 311 * Whenever we retransmit from recovery we increment 312 * rc_holes_rxt as we retran a block and mark it as retransmitted 313 * with the time it was sent. During non-recovery sending we 314 * add to our map and note the time down of any send expanding 315 * the rc_map at the tail and moving rc_snd_nxt up with snd_max. 316 * 317 * In recovery during SACK/ACK processing if a chunk has 318 * been retransmitted and it is now acked, we decrement rc_holes_rxt. 319 * When we retransmit from the scoreboard we use 320 * rc_next and rc_snd_nxt/rc_send_end to help us 321 * find what needs to be retran. 322 * 323 * To calculate pipe we simply take (snd_max - snd_una) + rc_holes_rxt 324 * This gets us the effect of RFC6675 pipe, counting twice for 325 * bytes retransmitted. 326 */ 327 328#define TT_BBR_FR_TMR 0x2001 329 330#define BBR_SCALE 8 331#define BBR_UNIT (1 << BBR_SCALE) 332 333#define BBR_NUM_RTTS_FOR_DEL_LIMIT 8 /* How many pkt-rtts do we keep 334 * Delivery rate for */ 335#define BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT 10 /* How many pkt-rtts do we keep 336 * Delivery rate for google */ 337 338#define BBR_SECONDS_NO_RTT 10 /* 10 seconds with no RTT shrinkage */ 339#define BBR_PROBERTT_MAX 200 /* 200ms */ 340#define BBR_PROBERTT_NUM_MSS 4 341#define BBR_STARTUP_EPOCHS 3 342#define USECS_IN_MSEC 1000 343#define BBR_TIME_TO_SECONDS(a) (a / USECS_IN_SECOND) 344#define BBR_TIME_TO_MILLI(a) (a / MS_IN_USEC) 345 346/* BBR keeps time in usec's so we divide by 1000 and round up */ 347#define BBR_TS_TO_MS(t) ((t+999)/MS_IN_USEC) 348 349/* 350 * Locking for the rack control block. 351 * a) Locked by INP_WLOCK 352 * b) Locked by the hpts-mutex 353 * 354 */ 355#define BBR_STATE_STARTUP 0x01 356#define BBR_STATE_DRAIN 0x02 357#define BBR_STATE_PROBE_BW 0x03 358#define BBR_STATE_PROBE_RTT 0x04 359#define BBR_STATE_IDLE_EXIT 0x05 360 361/* Substate defines for STATE == PROBE_BW */ 362#define BBR_SUB_GAIN 0 /* State 0 where we are 5/4 BBR_UNIT */ 363#define BBR_SUB_DRAIN 1 /* State 1 where we are at 3/4 BBR_UNIT */ 364#define BBR_SUB_LEVEL1 2 /* State 1 first BBR_UNIT */ 365#define BBR_SUB_LEVEL2 3 /* State 2nd BBR_UNIT */ 366#define BBR_SUB_LEVEL3 4 /* State 3rd BBR_UNIT */ 367#define BBR_SUB_LEVEL4 5 /* State 4th BBR_UNIT */ 368#define BBR_SUB_LEVEL5 6 /* State 5th BBR_UNIT */ 369#define BBR_SUB_LEVEL6 7 /* State last BBR_UNIT */ 370#define BBR_SUBSTATE_COUNT 8 371 372/* Single remaining reduce log */ 373#define BBR_REDUCE_AT_FR 5 374 375#define BBR_BIG_LOG_SIZE 300000 376 377struct bbr_stats { 378 uint64_t bbr_badfr; /* 0 */ 379 uint64_t bbr_badfr_bytes; /* 1 */ 380 uint64_t bbr_saw_oerr; /* 2 */ 381 uint64_t bbr_saw_emsgsiz; /* 3 */ 382 uint64_t bbr_reorder_seen; /* 4 */ 383 uint64_t bbr_tlp_tot; /* 5 */ 384 uint64_t bbr_tlp_newdata; /* 6 */ 385 uint64_t bbr_offset_recovery; /* 7 */ 386 uint64_t bbr_tlp_retran_fail; /* 8 */ 387 uint64_t bbr_to_tot; /* 9 */ 388 uint64_t bbr_to_arm_rack; /* 10 */ 389 uint64_t bbr_enter_probertt; /* 11 */ 390 uint64_t bbr_tlp_set; /* 12 */ 391 uint64_t bbr_resends_set; /* 13 */ 392 uint64_t bbr_force_output; /* 14 */ 393 uint64_t bbr_to_arm_tlp; /* 15 */ 394 uint64_t bbr_paced_segments; /* 16 */ 395 uint64_t bbr_saw_enobuf; /* 17 */ 396 uint64_t bbr_to_alloc_failed; /* 18 */ 397 uint64_t bbr_to_alloc_emerg; /* 19 */ 398 uint64_t bbr_sack_proc_all; /* 20 */ 399 uint64_t bbr_sack_proc_short; /* 21 */ 400 uint64_t bbr_sack_proc_restart; /* 22 */ 401 uint64_t bbr_to_alloc; /* 23 */ 402 uint64_t bbr_offset_drop; /* 24 */ 403 uint64_t bbr_runt_sacks; /* 25 */ 404 uint64_t bbr_sack_passed; /* 26 */ 405 uint64_t bbr_rlock_left_ret0; /* 27 */ 406 uint64_t bbr_rlock_left_ret1; /* 28 */ 407 uint64_t bbr_dynamic_rwnd; /* 29 */ 408 uint64_t bbr_static_rwnd; /* 30 */ 409 uint64_t bbr_sack_blocks; /* 31 */ 410 uint64_t bbr_sack_blocks_skip; /* 32 */ 411 uint64_t bbr_sack_search_both; /* 33 */ 412 uint64_t bbr_sack_search_fwd; /* 34 */ 413 uint64_t bbr_sack_search_back; /* 35 */ 414 uint64_t bbr_plain_acks; /* 36 */ 415 uint64_t bbr_acks_with_sacks; /* 37 */ 416 uint64_t bbr_progress_drops; /* 38 */ 417 uint64_t bbr_early; /* 39 */ 418 uint64_t bbr_reneges_seen; /* 40 */ 419 uint64_t bbr_persist_reneg; /* 41 */ 420 uint64_t bbr_dropped_af_data; /* 42 */ 421 uint64_t bbr_failed_mbuf_aloc; /* 43 */ 422 uint64_t bbr_cwnd_limited; /* 44 */ 423 uint64_t bbr_rwnd_limited; /* 45 */ 424 uint64_t bbr_app_limited; /* 46 */ 425 uint64_t bbr_force_timer_start; /* 47 */ 426 uint64_t bbr_hpts_min_time; /* 48 */ 427 uint64_t bbr_meets_tso_thresh; /* 49 */ 428 uint64_t bbr_miss_tso_rwnd; /* 50 */ 429 uint64_t bbr_miss_tso_cwnd; /* 51 */ 430 uint64_t bbr_miss_tso_app; /* 52 */ 431 uint64_t bbr_miss_retran; /* 53 */ 432 uint64_t bbr_miss_tlp; /* 54 */ 433 uint64_t bbr_miss_unknown; /* 55 */ 434 uint64_t bbr_hdwr_rl_add_ok; /* 56 */ 435 uint64_t bbr_hdwr_rl_add_fail; /* 57 */ 436 uint64_t bbr_hdwr_rl_mod_ok; /* 58 */ 437 uint64_t bbr_hdwr_rl_mod_fail; /* 59 */ 438 uint64_t bbr_collapsed_win; /* 60 */ 439 uint64_t bbr_alloc_limited; /* 61 */ 440 uint64_t bbr_alloc_limited_conns; /* 62 */ 441 uint64_t bbr_split_limited; /* 63 */ 442}; 443 444/* 445 * The structure bbr_opt_stats is a simple 446 * way to see how many options are being 447 * changed in the stack. 448 */ 449struct bbr_opts_stats { 450 uint64_t tcp_bbr_pace_per_sec; 451 uint64_t tcp_bbr_pace_del_tar; 452 uint64_t tcp_bbr_pace_seg_max; 453 uint64_t tcp_bbr_pace_seg_min; 454 uint64_t tcp_bbr_pace_cross; 455 uint64_t tcp_bbr_drain_inc_extra; 456 uint64_t tcp_bbr_unlimited; 457 uint64_t tcp_bbr_iwintso; 458 uint64_t tcp_bbr_rec_over_hpts; 459 uint64_t tcp_bbr_recforce; 460 uint64_t tcp_bbr_startup_pg; 461 uint64_t tcp_bbr_drain_pg; 462 uint64_t tcp_bbr_rwnd_is_app; 463 uint64_t tcp_bbr_probe_rtt_int; 464 uint64_t tcp_bbr_one_retran; 465 uint64_t tcp_bbr_startup_loss_exit; 466 uint64_t tcp_bbr_use_lowgain; 467 uint64_t tcp_bbr_lowgain_thresh; 468 uint64_t tcp_bbr_lowgain_half; 469 uint64_t tcp_bbr_lowgain_fd; 470 uint64_t tcp_bbr_usedel_rate; 471 uint64_t tcp_bbr_min_rto; 472 uint64_t tcp_bbr_max_rto; 473 uint64_t tcp_rack_pace_max_seg; 474 uint64_t tcp_rack_min_to; 475 uint64_t tcp_rack_reord_thresh; 476 uint64_t tcp_rack_reord_fade; 477 uint64_t tcp_rack_tlp_thresh; 478 uint64_t tcp_rack_pkt_delay; 479 uint64_t tcp_bbr_startup_exit_epoch; 480 uint64_t tcp_bbr_ack_comp_alg; 481 uint64_t tcp_rack_cheat; 482 uint64_t tcp_iwnd_tso; 483 uint64_t tcp_utter_max_tso; 484 uint64_t tcp_hdwr_pacing; 485 uint64_t tcp_extra_state; 486 uint64_t tcp_floor_min_tso; 487 /* New */ 488 uint64_t tcp_bbr_algorithm; 489 uint64_t tcp_bbr_tslimits; 490 uint64_t tcp_bbr_probertt_len; 491 uint64_t tcp_bbr_probertt_gain; 492 uint64_t tcp_bbr_topaceout; 493 uint64_t tcp_use_rackcheat; 494 uint64_t tcp_delack; 495 uint64_t tcp_maxpeak; 496 uint64_t tcp_retran_wtso; 497 uint64_t tcp_data_ac; 498 uint64_t tcp_ts_raises; 499 uint64_t tcp_pacing_oh_tmr; 500 uint64_t tcp_pacing_oh; 501 uint64_t tcp_policer_det; 502}; 503 504#ifdef _KERNEL 505#define BBR_STAT_SIZE (sizeof(struct bbr_stats)/sizeof(uint64_t)) 506extern counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 507#define BBR_STAT_ADD(name, amm) counter_u64_add(bbr_stat_arry[(offsetof(struct bbr_stats, name)/sizeof(uint64_t))], (amm)) 508#define BBR_STAT_INC(name) BBR_STAT_ADD(name, 1) 509#define BBR_OPTS_SIZE (sizeof(struct bbr_stats)/sizeof(uint64_t)) 510extern counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 511#define BBR_OPTS_ADD(name, amm) counter_u64_add(bbr_opts_arry[(offsetof(struct bbr_opts_stats, name)/sizeof(uint64_t))], (amm)) 512#define BBR_OPTS_INC(name) BBR_OPTS_ADD(name, 1) 513#endif 514 515#define BBR_NUM_LOSS_RATES 3 516#define BBR_NUM_BW_RATES 3 517 518#define BBR_RECOVERY_LOWRTT 1 519#define BBR_RECOVERY_MEDRTT 2 520#define BBR_RECOVERY_HIGHRTT 3 521#define BBR_RECOVERY_EXTREMERTT 4 522 523struct bbr_control { 524 /*******************************/ 525 /* Cache line 2 from bbr start */ 526 /*******************************/ 527 struct bbr_head rc_map; /* List of all segments Lock(a) */ 528 struct bbr_head rc_tmap; /* List in transmit order Lock(a) */ 529 struct bbr_sendmap *rc_resend; /* something we have been asked to 530 * resend */ 531 uint32_t rc_last_delay_val; /* How much we expect to delay Lock(a) */ 532 uint32_t rc_bbr_hptsi_gain:16, /* Current hptsi gain Lock(a) */ 533 rc_hpts_flags:16; /* flags on whats on the pacer wheel */ 534 535 uint32_t rc_delivered; /* BRR delivered amount Lock(a) */ 536 uint32_t rc_hptsi_agg_delay; /* How much time are we behind */ 537 538 uint32_t rc_flight_at_input; 539 uint32_t rc_lost_bytes; /* Total bytes currently marked lost */ 540 /*******************************/ 541 /* Cache line 3 from bbr start */ 542 /*******************************/ 543 struct time_filter rc_delrate; 544 /*******************************/ 545 /* Cache line 4 from bbr start */ 546 /*******************************/ 547 struct bbr_head rc_free; /* List of Free map entries Lock(a) */ 548 struct bbr_sendmap *rc_tlp_send; /* something we have been 549 * asked to resend */ 550 uint32_t rc_del_time; 551 uint32_t rc_target_at_state; /* Target for a state */ 552 553 uint16_t rc_free_cnt; /* Number of free entries on the rc_free list 554 * Lock(a) */ 555 uint16_t rc_startup_pg; 556 557 uint32_t cur_rtt; /* Last RTT from ack */ 558 559 uint32_t rc_went_idle_time; /* Used for persits to see if its 560 * probe-rtt qualified */ 561 uint32_t rc_pace_max_segs:17, /* How much in any single TSO we send Lock(a) */ 562 rc_pace_min_segs:15; /* The minimum single segment size before we enter persists */ 563 564 uint32_t rc_rtt_shrinks; /* Time of last rtt shrinkage Lock(a) */ 565 uint32_t r_app_limited_until; 566 uint32_t rc_timer_exp; /* If a timer ticks of expiry */ 567 uint32_t rc_rcv_epoch_start; /* Start time of the Epoch Lock(a) */ 568 569 /*******************************/ 570 /* Cache line 5 from bbr start */ 571 /*******************************/ 572 573 uint32_t rc_lost_at_pktepoch; /* what the lost value was at the last 574 * pkt-epoch */ 575 uint32_t r_measurement_count; /* count of measurement applied lock(a) */ 576 577 uint32_t rc_last_tlp_seq; /* Last tlp sequence Lock(a) */ 578 uint16_t rc_reorder_shift; /* Socket option value Lock(a) */ 579 uint16_t rc_pkt_delay; /* Socket option value Lock(a) */ 580 581 struct bbr_sendmap *rc_sacklast; /* sack remembered place 582 * Lock(a) */ 583 struct bbr_sendmap *rc_next; /* remembered place where we next 584 * retransmit at Lock(a) */ 585 586 uint32_t rc_sacked; /* Tot sacked on scoreboard Lock(a) */ 587 uint32_t rc_holes_rxt; /* Tot retraned from scoreboard Lock(a) */ 588 589 uint32_t rc_reorder_ts; /* Last time we saw reordering Lock(a) */ 590 uint32_t rc_init_rwnd; /* Initial rwnd when we transitioned */ 591 /*- --- 592 * used only initial and close 593 */ 594 uint32_t rc_high_rwnd; /* Highest rwnd seen */ 595 uint32_t rc_lowest_rtt; /* Smallest RTT we have seen */ 596 597 uint32_t rc_last_rtt; /* Last valid measured RTT that ack'd data */ 598 uint32_t bbr_cross_over; 599 600 /*******************************/ 601 /* Cache line 6 from bbr start */ 602 /*******************************/ 603 struct sack_filter bbr_sf; 604 605 /*******************************/ 606 /* Cache line 7 from bbr start */ 607 /*******************************/ 608 struct time_filter_small rc_rttprop; 609 uint32_t last_inbound_ts; /* Peers last timestamp */ 610 611 uint32_t rc_inc_tcp_oh: 1, 612 rc_inc_ip_oh: 1, 613 rc_inc_enet_oh:1, 614 rc_incr_tmrs:1, 615 restrict_growth:28; 616 uint32_t rc_lt_epoch_use; /* When we started lt-bw use Lock(a) */ 617 618 uint32_t rc_recovery_start; /* Time we start recovery Lock(a) */ 619 uint32_t rc_lt_del; /* Delivered at lt bw sampling start Lock(a) */ 620 621 uint64_t rc_bbr_cur_del_rate; /* Current measured delivery rate 622 * Lock(a) */ 623 624 /*******************************/ 625 /* Cache line 8 from bbr start */ 626 /*******************************/ 627 uint32_t rc_cwnd_on_ent; /* On entry to recovery the cwnd 628 * Lock(a) */ 629 uint32_t rc_agg_early; /* aggregate amount early */ 630 631 uint32_t rc_rcvtime; /* When we last received data Lock(a) */ 632 uint32_t rc_pkt_epoch_del; /* seq num that we need for RTT epoch */ 633 634 uint32_t rc_pkt_epoch; /* Epoch based on packet RTTs */ 635 uint32_t rc_pkt_epoch_time; /* Time we started the pkt epoch */ 636 637 uint32_t rc_pkt_epoch_rtt; /* RTT using the packet epoch */ 638 uint32_t rc_rtt_epoch; /* Current RTT epoch, it ticks every rttProp 639 * Lock(a) */ 640 uint32_t lowest_rtt; 641 uint32_t bbr_smallest_srtt_this_state; 642 643 uint32_t rc_lt_epoch; /* LT epoch start of bw_sampling */ 644 uint32_t rc_lost_at_startup; 645 646 uint32_t rc_bbr_state_atflight; 647 uint32_t rc_bbr_last_startup_epoch; /* Last startup epoch where we 648 * increased 20% */ 649 uint32_t rc_bbr_enters_probertt; /* Timestamp we entered 650 * probertt Lock(a) */ 651 uint32_t rc_lt_time; /* Time of lt sampling start Lock(a) */ 652 653 /*******************************/ 654 /* Cache line 9 from bbr start */ 655 /*******************************/ 656 uint64_t rc_lt_bw; /* LT bw calculated Lock(a) */ 657 uint64_t rc_bbr_lastbtlbw; /* For startup, what was last btlbw I 658 * saw to check the 20% gain Lock(a) */ 659 660 uint32_t rc_bbr_cwnd_gain; /* Current cwnd gain Lock(a) */ 661 uint32_t rc_pkt_epoch_loss_rate; /* pkt-epoch loss rate */ 662 663 uint32_t rc_saved_cwnd; /* Saved cwnd during Probe-rtt drain Lock(a) */ 664 uint32_t substate_pe; 665 666 uint32_t rc_lost; /* Number of bytes lost Lock(a) */ 667 uint32_t rc_exta_time_gd; /* How much extra time we got in d/g */ 668 669 uint32_t rc_lt_lost; /* Number of lt bytes lost at sampling start 670 * Lock(a) */ 671 uint32_t rc_bbr_state_time; 672 673 uint32_t rc_min_to; /* Socket option value Lock(a) */ 674 uint32_t rc_initial_hptsi_bw; /* Our initial startup bw Lock(a) */ 675 676 uint32_t bbr_lost_at_state; /* Temp counter debug lost value as we 677 * enter a state */ 678 /*******************************/ 679 /* Cache line 10 from bbr start */ 680 /*******************************/ 681 uint32_t rc_level_state_extra; 682 uint32_t rc_red_cwnd_pe; 683 const struct tcp_hwrate_limit_table *crte; 684 uint64_t red_bw; 685 686 uint32_t rc_probertt_int; 687 uint32_t rc_probertt_srttchktim; /* Time we last did a srtt 688 * check */ 689 uint32_t gain_epoch; /* Epoch we should be out of gain */ 690 uint32_t rc_min_rto_ms; 691 692 uint32_t rc_reorder_fade; /* Socket option value Lock(a) */ 693 uint32_t last_startup_measure; 694 695 int32_t bbr_hptsi_per_second; 696 int32_t bbr_hptsi_segments_delay_tar; 697 698 int32_t bbr_hptsi_segments_max; 699 uint32_t bbr_rttprobe_gain_val; 700 /*******************************/ 701 /* Cache line 11 from bbr start */ 702 /*******************************/ 703 uint32_t cur_rtt_send_time; /* Time we sent our rtt measured packet */ 704 uint32_t bbr_peer_tsratio; /* Our calculated ts ratio to multply */ 705 uint32_t bbr_ts_check_tstmp; /* When we filled it the TS that came on the ack */ 706 uint32_t bbr_ts_check_our_cts; /* When we filled it the cts of the send */ 707 uint32_t rc_tlp_rxt_last_time; 708 uint32_t bbr_smallest_srtt_state2; 709 uint32_t bbr_hdwr_cnt_noset_snt; /* count of hw pacing sends during delay */ 710 uint32_t startup_last_srtt; 711 uint32_t rc_ack_hdwr_delay; 712 uint32_t highest_hdwr_delay; /* Largest delay we have seen from hardware */ 713 uint32_t non_gain_extra; 714 uint32_t recovery_lr; /* The sum of the loss rate from the pe's during recovery */ 715 uint32_t last_in_probertt; 716 uint32_t flightsize_at_drain; /* In draining what was the last marked flight size */ 717 uint32_t rc_pe_of_prtt; /* PE we went into probe-rtt */ 718 uint32_t ts_in; /* ts that went with the last rtt */ 719 720 uint16_t rc_tlp_seg_send_cnt; /* Number of times we have TLP sent 721 * rc_last_tlp_seq Lock(a) */ 722 uint16_t rc_drain_pg; 723 uint32_t rc_num_maps_alloced; /* num send map entries allocated */ 724 uint32_t rc_num_split_allocs; /* num split map entries allocated */ 725 uint16_t rc_num_small_maps_alloced; /* Number of sack blocks 726 * allocated */ 727 uint16_t bbr_hptsi_bytes_min; 728 729 uint16_t bbr_hptsi_segments_floor; 730 uint16_t bbr_utter_max; 731 uint16_t bbr_google_discount; 732 733}; 734 735struct socket; 736struct tcp_bbr { 737 /* First cache line 0x00 */ 738 int32_t(*r_substate) (struct mbuf *, struct tcphdr *, 739 struct socket *, struct tcpcb *, struct tcpopt *, 740 int32_t, int32_t, uint32_t, int32_t, int32_t, uint8_t); /* Lock(a) */ 741 struct tcpcb *rc_tp; /* The tcpcb Lock(a) */ 742 struct inpcb *rc_inp; /* The inpcb Lock(a) */ 743 struct timeval rc_tv; 744 uint32_t rc_pacer_started; /* Time we started the pacer */ 745 uint16_t no_pacing_until:8, /* No pacing until N packet epochs */ 746 ts_can_raise:1,/* TS b/w calculations can raise the bw higher */ 747 skip_gain:1, /* Skip the gain cycle (hardware pacing) */ 748 gain_is_limited:1, /* With hardware pacing we are limiting gain */ 749 output_error_seen:1, 750 oerror_cnt:4, 751 hw_pacing_set:1; /* long enough has passed for us to start pacing */ 752 uint16_t xxx_r_ack_count; /* During recovery count of ack's received 753 * that added data since output */ 754 uint16_t bbr_segs_rcvd; /* In Segment count since we sent a ack */ 755 756 uint8_t bbr_timer_src:4, /* Used for debugging Lock(a) */ 757 bbr_use_rack_cheat:1, /* Use the rack cheat */ 758 bbr_init_win_cheat:1, /* Send full IW for TSO */ 759 bbr_attempt_hdwr_pace:1,/* Try to do hardware pacing */ 760 bbr_hdrw_pacing:1; /* Hardware pacing is available */ 761 uint8_t bbr_hdw_pace_ena:1, /* Does the connection allow hardware pacing to be attempted */ 762 bbr_prev_in_rec:1, /* We were previously in recovery */ 763 pkt_conservation:1, 764 use_policer_detection:1, 765 xxx_bbr_hdw_pace_idx:4; /* If hardware pacing is on, index to slot in pace tbl */ 766 uint16_t r_wanted_output:1, 767 rtt_valid:1, 768 rc_timer_first:1, 769 rc_output_starts_timer:1, 770 rc_resends_use_tso:1, 771 rc_all_timers_stopped:1, 772 rc_loss_exit:1, 773 rc_ack_was_delayed:1, 774 rc_lt_is_sampling:1, 775 rc_filled_pipe:1, 776 rc_tlp_new_data:1, 777 rc_hit_state_1:1, 778 rc_ts_valid:1, 779 rc_prtt_set_ts:1, 780 rc_is_pkt_epoch_now:1, 781 rc_has_collapsed:1; 782 783 uint8_t r_state:4, /* Current bbr state Lock(a) */ 784 r_agg_early_set:1, /* Did we get called early */ 785 r_init_rtt:1, 786 r_use_policer:1, /* For google mode only */ 787 r_recovery_bw:1; 788 uint8_t r_timer_override:1, /* pacer override Lock(a) 0/1 */ 789 rc_in_persist:1, 790 rc_lt_use_bw:1, 791 rc_allow_data_af_clo:1, 792 rc_tlp_rtx_out:1, /* A TLP is in flight */ 793 rc_tlp_in_progress:1, /* a TLP timer is running needed? */ 794 rc_use_idle_restart:1; /* Do we restart fast after idle (persist or applim) */ 795 uint8_t rc_bbr_state:3, /* What is the major BBR state */ 796 rc_bbr_substate:3, /* For probeBW state */ 797 r_is_v6:1, 798 rc_past_init_win:1; 799 uint8_t rc_last_options; 800 uint8_t rc_tlp_threshold; /* Socket option value Lock(a) */ 801 uint8_t rc_max_rto_sec; 802 uint8_t rc_cwnd_limited:1, /* We are cwnd limited */ 803 rc_tmr_stopped:7; /* What timers have been stopped */ 804 uint8_t rc_use_google:1, 805 rc_use_ts_limit:1, 806 rc_ts_data_set:1, /* We have filled a set point to determine */ 807 rc_ts_clock_set:1, /* We have determined the ts type */ 808 rc_ts_cant_be_used:1, /* We determined we can't use ts values */ 809 rc_ack_is_cumack:1, 810 rc_no_pacing:1, 811 alloc_limit_reported:1; 812 uint8_t rc_init_win; 813 /* Cache line 2 0x40 */ 814 struct bbr_control r_ctl; 815#ifdef _KERNEL 816} __aligned(CACHE_LINE_SIZE); 817#else 818}; 819#endif 820 821#endif 822