1 /* Bottleneck Bandwidth and RTT (BBR) congestion control 2 * 3 * BBR congestion control computes the sending rate based on the delivery 4 * rate (throughput) estimated from ACKs. In a nutshell: 5 * 6 * On each ACK, update our model of the network path: 7 * bottleneck_bandwidth = windowed_max(delivered / elapsed, 10 round trips) 8 * min_rtt = windowed_min(rtt, 10 seconds) 9 * pacing_rate = pacing_gain * bottleneck_bandwidth 10 * cwnd = max(cwnd_gain * bottleneck_bandwidth * min_rtt, 4) 11 * 12 * The core algorithm does not react directly to packet losses or delays, 13 * although BBR may adjust the size of next send per ACK when loss is 14 * observed, or adjust the sending rate if it estimates there is a 15 * traffic policer, in order to keep the drop rate reasonable. 16 * 17 * Here is a state transition diagram for BBR: 18 * 19 * | 20 * V 21 * +---> STARTUP ----+ 22 * | | | 23 * | V | 24 * | DRAIN ----+ 25 * | | | 26 * | V | 27 * +---> PROBE_BW ----+ 28 * | ^ | | 29 * | | | | 30 * | +----+ | 31 * | | 32 * +---- PROBE_RTT <--+ 33 * 34 * A BBR flow starts in STARTUP, and ramps up its sending rate quickly. 35 * When it estimates the pipe is full, it enters DRAIN to drain the queue. 36 * In steady state a BBR flow only uses PROBE_BW and PROBE_RTT. 37 * A long-lived BBR flow spends the vast majority of its time remaining 38 * (repeatedly) in PROBE_BW, fully probing and utilizing the pipe's bandwidth 39 * in a fair manner, with a small, bounded queue. *If* a flow has been 40 * continuously sending for the entire min_rtt window, and hasn't seen an RTT 41 * sample that matches or decreases its min_rtt estimate for 10 seconds, then 42 * it briefly enters PROBE_RTT to cut inflight to a minimum value to re-probe 43 * the path's two-way propagation delay (min_rtt). When exiting PROBE_RTT, if 44 * we estimated that we reached the full bw of the pipe then we enter PROBE_BW; 45 * otherwise we enter STARTUP to try to fill the pipe. 46 * 47 * BBR is described in detail in: 48 * "BBR: Congestion-Based Congestion Control", 49 * Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh, 50 * Van Jacobson. ACM Queue, Vol. 14 No. 5, September-October 2016. 51 * 52 * There is a public e-mail list for discussing BBR development and testing: 53 * https://groups.google.com/forum/#!forum/bbr-dev 54 * 55 * NOTE: BBR *must* be used with the fq qdisc ("man tc-fq") with pacing enabled, 56 * since pacing is integral to the BBR design and implementation. 57 * BBR without pacing would not function properly, and may incur unnecessary 58 * high packet loss rates. 59 */ 60 #include <linux/module.h> 61 #include <net/tcp.h> 62 #include <linux/inet_diag.h> 63 #include <linux/inet.h> 64 #include <linux/random.h> 65 #include <linux/win_minmax.h> 66 67 /* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth 68 * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps. 69 * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32. 70 * Since the minimum window is >=4 packets, the lower bound isn't 71 * an issue. The upper bound isn't an issue with existing technologies. 72 */ 73 #define BW_SCALE 24 74 #define BW_UNIT (1 << BW_SCALE) 75 76 #define BBR_SCALE 8 /* scaling factor for fractions in BBR (e.g. gains) */ 77 #define BBR_UNIT (1 << BBR_SCALE) 78 79 /* BBR has the following modes for deciding how fast to send: */ 80 enum bbr_mode { 81 BBR_STARTUP, /* ramp up sending rate rapidly to fill pipe */ 82 BBR_DRAIN, /* drain any queue created during startup */ 83 BBR_PROBE_BW, /* discover, share bw: pace around estimated bw */ 84 BBR_PROBE_RTT, /* cut inflight to min to probe min_rtt */ 85 }; 86 87 /* BBR congestion control block */ 88 struct bbr { 89 u32 min_rtt_us; /* min RTT in min_rtt_win_sec window */ 90 u32 min_rtt_stamp; /* timestamp of min_rtt_us */ 91 u32 probe_rtt_done_stamp; /* end time for BBR_PROBE_RTT mode */ 92 struct minmax bw; /* Max recent delivery rate in pkts/uS << 24 */ 93 u32 rtt_cnt; /* count of packet-timed rounds elapsed */ 94 u32 next_rtt_delivered; /* scb->tx.delivered at end of round */ 95 struct skb_mstamp cycle_mstamp; /* time of this cycle phase start */ 96 u32 mode:3, /* current bbr_mode in state machine */ 97 prev_ca_state:3, /* CA state on previous ACK */ 98 packet_conservation:1, /* use packet conservation? */ 99 restore_cwnd:1, /* decided to revert cwnd to old value */ 100 round_start:1, /* start of packet-timed tx->ack round? */ 101 tso_segs_goal:7, /* segments we want in each skb we send */ 102 idle_restart:1, /* restarting after idle? */ 103 probe_rtt_round_done:1, /* a BBR_PROBE_RTT round at 4 pkts? */ 104 unused:5, 105 lt_is_sampling:1, /* taking long-term ("LT") samples now? */ 106 lt_rtt_cnt:7, /* round trips in long-term interval */ 107 lt_use_bw:1; /* use lt_bw as our bw estimate? */ 108 u32 lt_bw; /* LT est delivery rate in pkts/uS << 24 */ 109 u32 lt_last_delivered; /* LT intvl start: tp->delivered */ 110 u32 lt_last_stamp; /* LT intvl start: tp->delivered_mstamp */ 111 u32 lt_last_lost; /* LT intvl start: tp->lost */ 112 u32 pacing_gain:10, /* current gain for setting pacing rate */ 113 cwnd_gain:10, /* current gain for setting cwnd */ 114 full_bw_cnt:3, /* number of rounds without large bw gains */ 115 cycle_idx:3, /* current index in pacing_gain cycle array */ 116 unused_b:6; 117 u32 prior_cwnd; /* prior cwnd upon entering loss recovery */ 118 u32 full_bw; /* recent bw, to estimate if pipe is full */ 119 }; 120 121 #define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */ 122 123 /* Window length of bw filter (in rounds): */ 124 static const int bbr_bw_rtts = CYCLE_LEN + 2; 125 /* Window length of min_rtt filter (in sec): */ 126 static const u32 bbr_min_rtt_win_sec = 10; 127 /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */ 128 static const u32 bbr_probe_rtt_mode_ms = 200; 129 /* Skip TSO below the following bandwidth (bits/sec): */ 130 static const int bbr_min_tso_rate = 1200000; 131 132 /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain 133 * that will allow a smoothly increasing pacing rate that will double each RTT 134 * and send the same number of packets per RTT that an un-paced, slow-starting 135 * Reno or CUBIC flow would: 136 */ 137 static const int bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 138 /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain 139 * the queue created in BBR_STARTUP in a single round: 140 */ 141 static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885; 142 /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */ 143 static const int bbr_cwnd_gain = BBR_UNIT * 2; 144 /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */ 145 static const int bbr_pacing_gain[] = { 146 BBR_UNIT * 5 / 4, /* probe for more available bw */ 147 BBR_UNIT * 3 / 4, /* drain queue and/or yield bw to other flows */ 148 BBR_UNIT, BBR_UNIT, BBR_UNIT, /* cruise at 1.0*bw to utilize pipe, */ 149 BBR_UNIT, BBR_UNIT, BBR_UNIT /* without creating excess queue... */ 150 }; 151 /* Randomize the starting gain cycling phase over N phases: */ 152 static const u32 bbr_cycle_rand = 7; 153 154 /* Try to keep at least this many packets in flight, if things go smoothly. For 155 * smooth functioning, a sliding window protocol ACKing every other packet 156 * needs at least 4 packets in flight: 157 */ 158 static const u32 bbr_cwnd_min_target = 4; 159 160 /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */ 161 /* If bw has increased significantly (1.25x), there may be more bw available: */ 162 static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4; 163 /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */ 164 static const u32 bbr_full_bw_cnt = 3; 165 166 /* "long-term" ("LT") bandwidth estimator parameters... */ 167 /* The minimum number of rounds in an LT bw sampling interval: */ 168 static const u32 bbr_lt_intvl_min_rtts = 4; 169 /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */ 170 static const u32 bbr_lt_loss_thresh = 50; 171 /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */ 172 static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8; 173 /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */ 174 static const u32 bbr_lt_bw_diff = 4000 / 8; 175 /* If we estimate we're policed, use lt_bw for this many round trips: */ 176 static const u32 bbr_lt_bw_max_rtts = 48; 177 178 /* Do we estimate that STARTUP filled the pipe? */ 179 static bool bbr_full_bw_reached(const struct sock *sk) 180 { 181 const struct bbr *bbr = inet_csk_ca(sk); 182 183 return bbr->full_bw_cnt >= bbr_full_bw_cnt; 184 } 185 186 /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */ 187 static u32 bbr_max_bw(const struct sock *sk) 188 { 189 struct bbr *bbr = inet_csk_ca(sk); 190 191 return minmax_get(&bbr->bw); 192 } 193 194 /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */ 195 static u32 bbr_bw(const struct sock *sk) 196 { 197 struct bbr *bbr = inet_csk_ca(sk); 198 199 return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk); 200 } 201 202 /* Return rate in bytes per second, optionally with a gain. 203 * The order here is chosen carefully to avoid overflow of u64. This should 204 * work for input rates of up to 2.9Tbit/sec and gain of 2.89x. 205 */ 206 static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain) 207 { 208 rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache); 209 rate *= gain; 210 rate >>= BBR_SCALE; 211 rate *= USEC_PER_SEC; 212 return rate >> BW_SCALE; 213 } 214 215 /* Pace using current bw estimate and a gain factor. In order to help drive the 216 * network toward lower queues while maintaining high utilization and low 217 * latency, the average pacing rate aims to be slightly (~1%) lower than the 218 * estimated bandwidth. This is an important aspect of the design. In this 219 * implementation this slightly lower pacing rate is achieved implicitly by not 220 * including link-layer headers in the packet size used for the pacing rate. 221 */ 222 static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain) 223 { 224 struct bbr *bbr = inet_csk_ca(sk); 225 u64 rate = bw; 226 227 rate = bbr_rate_bytes_per_sec(sk, rate, gain); 228 rate = min_t(u64, rate, sk->sk_max_pacing_rate); 229 if (bbr->mode != BBR_STARTUP || rate > sk->sk_pacing_rate) 230 sk->sk_pacing_rate = rate; 231 } 232 233 /* Return count of segments we want in the skbs we send, or 0 for default. */ 234 static u32 bbr_tso_segs_goal(struct sock *sk) 235 { 236 struct bbr *bbr = inet_csk_ca(sk); 237 238 return bbr->tso_segs_goal; 239 } 240 241 static void bbr_set_tso_segs_goal(struct sock *sk) 242 { 243 struct tcp_sock *tp = tcp_sk(sk); 244 struct bbr *bbr = inet_csk_ca(sk); 245 u32 min_segs; 246 247 min_segs = sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2; 248 bbr->tso_segs_goal = min(tcp_tso_autosize(sk, tp->mss_cache, min_segs), 249 0x7FU); 250 } 251 252 /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ 253 static void bbr_save_cwnd(struct sock *sk) 254 { 255 struct tcp_sock *tp = tcp_sk(sk); 256 struct bbr *bbr = inet_csk_ca(sk); 257 258 if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT) 259 bbr->prior_cwnd = tp->snd_cwnd; /* this cwnd is good enough */ 260 else /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */ 261 bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd); 262 } 263 264 static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event) 265 { 266 struct tcp_sock *tp = tcp_sk(sk); 267 struct bbr *bbr = inet_csk_ca(sk); 268 269 if (event == CA_EVENT_TX_START && tp->app_limited) { 270 bbr->idle_restart = 1; 271 /* Avoid pointless buffer overflows: pace at est. bw if we don't 272 * need more speed (we're restarting from idle and app-limited). 273 */ 274 if (bbr->mode == BBR_PROBE_BW) 275 bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT); 276 } 277 } 278 279 /* Find target cwnd. Right-size the cwnd based on min RTT and the 280 * estimated bottleneck bandwidth: 281 * 282 * cwnd = bw * min_rtt * gain = BDP * gain 283 * 284 * The key factor, gain, controls the amount of queue. While a small gain 285 * builds a smaller queue, it becomes more vulnerable to noise in RTT 286 * measurements (e.g., delayed ACKs or other ACK compression effects). This 287 * noise may cause BBR to under-estimate the rate. 288 * 289 * To achieve full performance in high-speed paths, we budget enough cwnd to 290 * fit full-sized skbs in-flight on both end hosts to fully utilize the path: 291 * - one skb in sending host Qdisc, 292 * - one skb in sending host TSO/GSO engine 293 * - one skb being received by receiver host LRO/GRO/delayed-ACK engine 294 * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because 295 * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets, 296 * which allows 2 outstanding 2-packet sequences, to try to keep pipe 297 * full even with ACK-every-other-packet delayed ACKs. 298 */ 299 static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain) 300 { 301 struct bbr *bbr = inet_csk_ca(sk); 302 u32 cwnd; 303 u64 w; 304 305 /* If we've never had a valid RTT sample, cap cwnd at the initial 306 * default. This should only happen when the connection is not using TCP 307 * timestamps and has retransmitted all of the SYN/SYNACK/data packets 308 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which 309 * case we need to slow-start up toward something safe: TCP_INIT_CWND. 310 */ 311 if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */ 312 return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/ 313 314 w = (u64)bw * bbr->min_rtt_us; 315 316 /* Apply a gain to the given value, then remove the BW_SCALE shift. */ 317 cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT; 318 319 /* Allow enough full-sized skbs in flight to utilize end systems. */ 320 cwnd += 3 * bbr->tso_segs_goal; 321 322 /* Reduce delayed ACKs by rounding up cwnd to the next even number. */ 323 cwnd = (cwnd + 1) & ~1U; 324 325 return cwnd; 326 } 327 328 /* An optimization in BBR to reduce losses: On the first round of recovery, we 329 * follow the packet conservation principle: send P packets per P packets acked. 330 * After that, we slow-start and send at most 2*P packets per P packets acked. 331 * After recovery finishes, or upon undo, we restore the cwnd we had when 332 * recovery started (capped by the target cwnd based on estimated BDP). 333 * 334 * TODO(ycheng/ncardwell): implement a rate-based approach. 335 */ 336 static bool bbr_set_cwnd_to_recover_or_restore( 337 struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd) 338 { 339 struct tcp_sock *tp = tcp_sk(sk); 340 struct bbr *bbr = inet_csk_ca(sk); 341 u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state; 342 u32 cwnd = tp->snd_cwnd; 343 344 /* An ACK for P pkts should release at most 2*P packets. We do this 345 * in two steps. First, here we deduct the number of lost packets. 346 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd. 347 */ 348 if (rs->losses > 0) 349 cwnd = max_t(s32, cwnd - rs->losses, 1); 350 351 if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) { 352 /* Starting 1st round of Recovery, so do packet conservation. */ 353 bbr->packet_conservation = 1; 354 bbr->next_rtt_delivered = tp->delivered; /* start round now */ 355 /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */ 356 cwnd = tcp_packets_in_flight(tp) + acked; 357 } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) { 358 /* Exiting loss recovery; restore cwnd saved before recovery. */ 359 bbr->restore_cwnd = 1; 360 bbr->packet_conservation = 0; 361 } 362 bbr->prev_ca_state = state; 363 364 if (bbr->restore_cwnd) { 365 /* Restore cwnd after exiting loss recovery or PROBE_RTT. */ 366 cwnd = max(cwnd, bbr->prior_cwnd); 367 bbr->restore_cwnd = 0; 368 } 369 370 if (bbr->packet_conservation) { 371 *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked); 372 return true; /* yes, using packet conservation */ 373 } 374 *new_cwnd = cwnd; 375 return false; 376 } 377 378 /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss 379 * has drawn us down below target), or snap down to target if we're above it. 380 */ 381 static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs, 382 u32 acked, u32 bw, int gain) 383 { 384 struct tcp_sock *tp = tcp_sk(sk); 385 struct bbr *bbr = inet_csk_ca(sk); 386 u32 cwnd = 0, target_cwnd = 0; 387 388 if (!acked) 389 return; 390 391 if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd)) 392 goto done; 393 394 /* If we're below target cwnd, slow start cwnd toward target cwnd. */ 395 target_cwnd = bbr_target_cwnd(sk, bw, gain); 396 if (bbr_full_bw_reached(sk)) /* only cut cwnd if we filled the pipe */ 397 cwnd = min(cwnd + acked, target_cwnd); 398 else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND) 399 cwnd = cwnd + acked; 400 cwnd = max(cwnd, bbr_cwnd_min_target); 401 402 done: 403 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); /* apply global cap */ 404 if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */ 405 tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target); 406 } 407 408 /* End cycle phase if it's time and/or we hit the phase's in-flight target. */ 409 static bool bbr_is_next_cycle_phase(struct sock *sk, 410 const struct rate_sample *rs) 411 { 412 struct tcp_sock *tp = tcp_sk(sk); 413 struct bbr *bbr = inet_csk_ca(sk); 414 bool is_full_length = 415 skb_mstamp_us_delta(&tp->delivered_mstamp, &bbr->cycle_mstamp) > 416 bbr->min_rtt_us; 417 u32 inflight, bw; 418 419 /* The pacing_gain of 1.0 paces at the estimated bw to try to fully 420 * use the pipe without increasing the queue. 421 */ 422 if (bbr->pacing_gain == BBR_UNIT) 423 return is_full_length; /* just use wall clock time */ 424 425 inflight = rs->prior_in_flight; /* what was in-flight before ACK? */ 426 bw = bbr_max_bw(sk); 427 428 /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at 429 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is 430 * small (e.g. on a LAN). We do not persist if packets are lost, since 431 * a path with small buffers may not hold that much. 432 */ 433 if (bbr->pacing_gain > BBR_UNIT) 434 return is_full_length && 435 (rs->losses || /* perhaps pacing_gain*BDP won't fit */ 436 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain)); 437 438 /* A pacing_gain < 1.0 tries to drain extra queue we added if bw 439 * probing didn't find more bw. If inflight falls to match BDP then we 440 * estimate queue is drained; persisting would underutilize the pipe. 441 */ 442 return is_full_length || 443 inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT); 444 } 445 446 static void bbr_advance_cycle_phase(struct sock *sk) 447 { 448 struct tcp_sock *tp = tcp_sk(sk); 449 struct bbr *bbr = inet_csk_ca(sk); 450 451 bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1); 452 bbr->cycle_mstamp = tp->delivered_mstamp; 453 bbr->pacing_gain = bbr_pacing_gain[bbr->cycle_idx]; 454 } 455 456 /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */ 457 static void bbr_update_cycle_phase(struct sock *sk, 458 const struct rate_sample *rs) 459 { 460 struct bbr *bbr = inet_csk_ca(sk); 461 462 if ((bbr->mode == BBR_PROBE_BW) && !bbr->lt_use_bw && 463 bbr_is_next_cycle_phase(sk, rs)) 464 bbr_advance_cycle_phase(sk); 465 } 466 467 static void bbr_reset_startup_mode(struct sock *sk) 468 { 469 struct bbr *bbr = inet_csk_ca(sk); 470 471 bbr->mode = BBR_STARTUP; 472 bbr->pacing_gain = bbr_high_gain; 473 bbr->cwnd_gain = bbr_high_gain; 474 } 475 476 static void bbr_reset_probe_bw_mode(struct sock *sk) 477 { 478 struct bbr *bbr = inet_csk_ca(sk); 479 480 bbr->mode = BBR_PROBE_BW; 481 bbr->pacing_gain = BBR_UNIT; 482 bbr->cwnd_gain = bbr_cwnd_gain; 483 bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand); 484 bbr_advance_cycle_phase(sk); /* flip to next phase of gain cycle */ 485 } 486 487 static void bbr_reset_mode(struct sock *sk) 488 { 489 if (!bbr_full_bw_reached(sk)) 490 bbr_reset_startup_mode(sk); 491 else 492 bbr_reset_probe_bw_mode(sk); 493 } 494 495 /* Start a new long-term sampling interval. */ 496 static void bbr_reset_lt_bw_sampling_interval(struct sock *sk) 497 { 498 struct tcp_sock *tp = tcp_sk(sk); 499 struct bbr *bbr = inet_csk_ca(sk); 500 501 bbr->lt_last_stamp = tp->delivered_mstamp.stamp_jiffies; 502 bbr->lt_last_delivered = tp->delivered; 503 bbr->lt_last_lost = tp->lost; 504 bbr->lt_rtt_cnt = 0; 505 } 506 507 /* Completely reset long-term bandwidth sampling. */ 508 static void bbr_reset_lt_bw_sampling(struct sock *sk) 509 { 510 struct bbr *bbr = inet_csk_ca(sk); 511 512 bbr->lt_bw = 0; 513 bbr->lt_use_bw = 0; 514 bbr->lt_is_sampling = false; 515 bbr_reset_lt_bw_sampling_interval(sk); 516 } 517 518 /* Long-term bw sampling interval is done. Estimate whether we're policed. */ 519 static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw) 520 { 521 struct bbr *bbr = inet_csk_ca(sk); 522 u32 diff; 523 524 if (bbr->lt_bw) { /* do we have bw from a previous interval? */ 525 /* Is new bw close to the lt_bw from the previous interval? */ 526 diff = abs(bw - bbr->lt_bw); 527 if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) || 528 (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <= 529 bbr_lt_bw_diff)) { 530 /* All criteria are met; estimate we're policed. */ 531 bbr->lt_bw = (bw + bbr->lt_bw) >> 1; /* avg 2 intvls */ 532 bbr->lt_use_bw = 1; 533 bbr->pacing_gain = BBR_UNIT; /* try to avoid drops */ 534 bbr->lt_rtt_cnt = 0; 535 return; 536 } 537 } 538 bbr->lt_bw = bw; 539 bbr_reset_lt_bw_sampling_interval(sk); 540 } 541 542 /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of 543 * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and 544 * explicitly models their policed rate, to reduce unnecessary losses. We 545 * estimate that we're policed if we see 2 consecutive sampling intervals with 546 * consistent throughput and high packet loss. If we think we're being policed, 547 * set lt_bw to the "long-term" average delivery rate from those 2 intervals. 548 */ 549 static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs) 550 { 551 struct tcp_sock *tp = tcp_sk(sk); 552 struct bbr *bbr = inet_csk_ca(sk); 553 u32 lost, delivered; 554 u64 bw; 555 s32 t; 556 557 if (bbr->lt_use_bw) { /* already using long-term rate, lt_bw? */ 558 if (bbr->mode == BBR_PROBE_BW && bbr->round_start && 559 ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) { 560 bbr_reset_lt_bw_sampling(sk); /* stop using lt_bw */ 561 bbr_reset_probe_bw_mode(sk); /* restart gain cycling */ 562 } 563 return; 564 } 565 566 /* Wait for the first loss before sampling, to let the policer exhaust 567 * its tokens and estimate the steady-state rate allowed by the policer. 568 * Starting samples earlier includes bursts that over-estimate the bw. 569 */ 570 if (!bbr->lt_is_sampling) { 571 if (!rs->losses) 572 return; 573 bbr_reset_lt_bw_sampling_interval(sk); 574 bbr->lt_is_sampling = true; 575 } 576 577 /* To avoid underestimates, reset sampling if we run out of data. */ 578 if (rs->is_app_limited) { 579 bbr_reset_lt_bw_sampling(sk); 580 return; 581 } 582 583 if (bbr->round_start) 584 bbr->lt_rtt_cnt++; /* count round trips in this interval */ 585 if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts) 586 return; /* sampling interval needs to be longer */ 587 if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) { 588 bbr_reset_lt_bw_sampling(sk); /* interval is too long */ 589 return; 590 } 591 592 /* End sampling interval when a packet is lost, so we estimate the 593 * policer tokens were exhausted. Stopping the sampling before the 594 * tokens are exhausted under-estimates the policed rate. 595 */ 596 if (!rs->losses) 597 return; 598 599 /* Calculate packets lost and delivered in sampling interval. */ 600 lost = tp->lost - bbr->lt_last_lost; 601 delivered = tp->delivered - bbr->lt_last_delivered; 602 /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */ 603 if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered) 604 return; 605 606 /* Find average delivery rate in this sampling interval. */ 607 t = (s32)(tp->delivered_mstamp.stamp_jiffies - bbr->lt_last_stamp); 608 if (t < 1) 609 return; /* interval is less than one jiffy, so wait */ 610 t = jiffies_to_usecs(t); 611 /* Interval long enough for jiffies_to_usecs() to return a bogus 0? */ 612 if (t < 1) { 613 bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */ 614 return; 615 } 616 bw = (u64)delivered * BW_UNIT; 617 do_div(bw, t); 618 bbr_lt_bw_interval_done(sk, bw); 619 } 620 621 /* Estimate the bandwidth based on how fast packets are delivered */ 622 static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs) 623 { 624 struct tcp_sock *tp = tcp_sk(sk); 625 struct bbr *bbr = inet_csk_ca(sk); 626 u64 bw; 627 628 bbr->round_start = 0; 629 if (rs->delivered < 0 || rs->interval_us <= 0) 630 return; /* Not a valid observation */ 631 632 /* See if we've reached the next RTT */ 633 if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) { 634 bbr->next_rtt_delivered = tp->delivered; 635 bbr->rtt_cnt++; 636 bbr->round_start = 1; 637 bbr->packet_conservation = 0; 638 } 639 640 bbr_lt_bw_sampling(sk, rs); 641 642 /* Divide delivered by the interval to find a (lower bound) bottleneck 643 * bandwidth sample. Delivered is in packets and interval_us in uS and 644 * ratio will be <<1 for most connections. So delivered is first scaled. 645 */ 646 bw = (u64)rs->delivered * BW_UNIT; 647 do_div(bw, rs->interval_us); 648 649 /* If this sample is application-limited, it is likely to have a very 650 * low delivered count that represents application behavior rather than 651 * the available network rate. Such a sample could drag down estimated 652 * bw, causing needless slow-down. Thus, to continue to send at the 653 * last measured network rate, we filter out app-limited samples unless 654 * they describe the path bw at least as well as our bw model. 655 * 656 * So the goal during app-limited phase is to proceed with the best 657 * network rate no matter how long. We automatically leave this 658 * phase when app writes faster than the network can deliver :) 659 */ 660 if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) { 661 /* Incorporate new sample into our max bw filter. */ 662 minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw); 663 } 664 } 665 666 /* Estimate when the pipe is full, using the change in delivery rate: BBR 667 * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by 668 * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited 669 * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the 670 * higher rwin, 3: we get higher delivery rate samples. Or transient 671 * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar 672 * design goal, but uses delay and inter-ACK spacing instead of bandwidth. 673 */ 674 static void bbr_check_full_bw_reached(struct sock *sk, 675 const struct rate_sample *rs) 676 { 677 struct bbr *bbr = inet_csk_ca(sk); 678 u32 bw_thresh; 679 680 if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited) 681 return; 682 683 bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE; 684 if (bbr_max_bw(sk) >= bw_thresh) { 685 bbr->full_bw = bbr_max_bw(sk); 686 bbr->full_bw_cnt = 0; 687 return; 688 } 689 ++bbr->full_bw_cnt; 690 } 691 692 /* If pipe is probably full, drain the queue and then enter steady-state. */ 693 static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs) 694 { 695 struct bbr *bbr = inet_csk_ca(sk); 696 697 if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) { 698 bbr->mode = BBR_DRAIN; /* drain queue we created */ 699 bbr->pacing_gain = bbr_drain_gain; /* pace slow to drain */ 700 bbr->cwnd_gain = bbr_high_gain; /* maintain cwnd */ 701 } /* fall through to check if in-flight is already small: */ 702 if (bbr->mode == BBR_DRAIN && 703 tcp_packets_in_flight(tcp_sk(sk)) <= 704 bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT)) 705 bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */ 706 } 707 708 /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and 709 * periodically drain the bottleneck queue, to converge to measure the true 710 * min_rtt (unloaded propagation delay). This allows the flows to keep queues 711 * small (reducing queuing delay and packet loss) and achieve fairness among 712 * BBR flows. 713 * 714 * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires, 715 * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets. 716 * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed 717 * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and 718 * re-enter the previous mode. BBR uses 200ms to approximately bound the 719 * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s). 720 * 721 * Note that flows need only pay 2% if they are busy sending over the last 10 722 * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have 723 * natural silences or low-rate periods within 10 seconds where the rate is low 724 * enough for long enough to drain its queue in the bottleneck. We pick up 725 * these min RTT measurements opportunistically with our min_rtt filter. :-) 726 */ 727 static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs) 728 { 729 struct tcp_sock *tp = tcp_sk(sk); 730 struct bbr *bbr = inet_csk_ca(sk); 731 bool filter_expired; 732 733 /* Track min RTT seen in the min_rtt_win_sec filter window: */ 734 filter_expired = after(tcp_time_stamp, 735 bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ); 736 if (rs->rtt_us >= 0 && 737 (rs->rtt_us <= bbr->min_rtt_us || filter_expired)) { 738 bbr->min_rtt_us = rs->rtt_us; 739 bbr->min_rtt_stamp = tcp_time_stamp; 740 } 741 742 if (bbr_probe_rtt_mode_ms > 0 && filter_expired && 743 !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) { 744 bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */ 745 bbr->pacing_gain = BBR_UNIT; 746 bbr->cwnd_gain = BBR_UNIT; 747 bbr_save_cwnd(sk); /* note cwnd so we can restore it */ 748 bbr->probe_rtt_done_stamp = 0; 749 } 750 751 if (bbr->mode == BBR_PROBE_RTT) { 752 /* Ignore low rate samples during this mode. */ 753 tp->app_limited = 754 (tp->delivered + tcp_packets_in_flight(tp)) ? : 1; 755 /* Maintain min packets in flight for max(200 ms, 1 round). */ 756 if (!bbr->probe_rtt_done_stamp && 757 tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) { 758 bbr->probe_rtt_done_stamp = tcp_time_stamp + 759 msecs_to_jiffies(bbr_probe_rtt_mode_ms); 760 bbr->probe_rtt_round_done = 0; 761 bbr->next_rtt_delivered = tp->delivered; 762 } else if (bbr->probe_rtt_done_stamp) { 763 if (bbr->round_start) 764 bbr->probe_rtt_round_done = 1; 765 if (bbr->probe_rtt_round_done && 766 after(tcp_time_stamp, bbr->probe_rtt_done_stamp)) { 767 bbr->min_rtt_stamp = tcp_time_stamp; 768 bbr->restore_cwnd = 1; /* snap to prior_cwnd */ 769 bbr_reset_mode(sk); 770 } 771 } 772 } 773 bbr->idle_restart = 0; 774 } 775 776 static void bbr_update_model(struct sock *sk, const struct rate_sample *rs) 777 { 778 bbr_update_bw(sk, rs); 779 bbr_update_cycle_phase(sk, rs); 780 bbr_check_full_bw_reached(sk, rs); 781 bbr_check_drain(sk, rs); 782 bbr_update_min_rtt(sk, rs); 783 } 784 785 static void bbr_main(struct sock *sk, const struct rate_sample *rs) 786 { 787 struct bbr *bbr = inet_csk_ca(sk); 788 u32 bw; 789 790 bbr_update_model(sk, rs); 791 792 bw = bbr_bw(sk); 793 bbr_set_pacing_rate(sk, bw, bbr->pacing_gain); 794 bbr_set_tso_segs_goal(sk); 795 bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain); 796 } 797 798 static void bbr_init(struct sock *sk) 799 { 800 struct tcp_sock *tp = tcp_sk(sk); 801 struct bbr *bbr = inet_csk_ca(sk); 802 u64 bw; 803 804 bbr->prior_cwnd = 0; 805 bbr->tso_segs_goal = 0; /* default segs per skb until first ACK */ 806 bbr->rtt_cnt = 0; 807 bbr->next_rtt_delivered = 0; 808 bbr->prev_ca_state = TCP_CA_Open; 809 bbr->packet_conservation = 0; 810 811 bbr->probe_rtt_done_stamp = 0; 812 bbr->probe_rtt_round_done = 0; 813 bbr->min_rtt_us = tcp_min_rtt(tp); 814 bbr->min_rtt_stamp = tcp_time_stamp; 815 816 minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */ 817 818 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */ 819 bw = (u64)tp->snd_cwnd * BW_UNIT; 820 do_div(bw, (tp->srtt_us >> 3) ? : USEC_PER_MSEC); 821 sk->sk_pacing_rate = 0; /* force an update of sk_pacing_rate */ 822 bbr_set_pacing_rate(sk, bw, bbr_high_gain); 823 824 bbr->restore_cwnd = 0; 825 bbr->round_start = 0; 826 bbr->idle_restart = 0; 827 bbr->full_bw = 0; 828 bbr->full_bw_cnt = 0; 829 bbr->cycle_mstamp.v64 = 0; 830 bbr->cycle_idx = 0; 831 bbr_reset_lt_bw_sampling(sk); 832 bbr_reset_startup_mode(sk); 833 } 834 835 static u32 bbr_sndbuf_expand(struct sock *sk) 836 { 837 /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ 838 return 3; 839 } 840 841 /* In theory BBR does not need to undo the cwnd since it does not 842 * always reduce cwnd on losses (see bbr_main()). Keep it for now. 843 */ 844 static u32 bbr_undo_cwnd(struct sock *sk) 845 { 846 return tcp_sk(sk)->snd_cwnd; 847 } 848 849 /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */ 850 static u32 bbr_ssthresh(struct sock *sk) 851 { 852 bbr_save_cwnd(sk); 853 return TCP_INFINITE_SSTHRESH; /* BBR does not use ssthresh */ 854 } 855 856 static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr, 857 union tcp_cc_info *info) 858 { 859 if (ext & (1 << (INET_DIAG_BBRINFO - 1)) || 860 ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 861 struct tcp_sock *tp = tcp_sk(sk); 862 struct bbr *bbr = inet_csk_ca(sk); 863 u64 bw = bbr_bw(sk); 864 865 bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE; 866 memset(&info->bbr, 0, sizeof(info->bbr)); 867 info->bbr.bbr_bw_lo = (u32)bw; 868 info->bbr.bbr_bw_hi = (u32)(bw >> 32); 869 info->bbr.bbr_min_rtt = bbr->min_rtt_us; 870 info->bbr.bbr_pacing_gain = bbr->pacing_gain; 871 info->bbr.bbr_cwnd_gain = bbr->cwnd_gain; 872 *attr = INET_DIAG_BBRINFO; 873 return sizeof(info->bbr); 874 } 875 return 0; 876 } 877 878 static void bbr_set_state(struct sock *sk, u8 new_state) 879 { 880 struct bbr *bbr = inet_csk_ca(sk); 881 882 if (new_state == TCP_CA_Loss) { 883 struct rate_sample rs = { .losses = 1 }; 884 885 bbr->prev_ca_state = TCP_CA_Loss; 886 bbr->full_bw = 0; 887 bbr->round_start = 1; /* treat RTO like end of a round */ 888 bbr_lt_bw_sampling(sk, &rs); 889 } 890 } 891 892 static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { 893 .flags = TCP_CONG_NON_RESTRICTED, 894 .name = "bbr", 895 .owner = THIS_MODULE, 896 .init = bbr_init, 897 .cong_control = bbr_main, 898 .sndbuf_expand = bbr_sndbuf_expand, 899 .undo_cwnd = bbr_undo_cwnd, 900 .cwnd_event = bbr_cwnd_event, 901 .ssthresh = bbr_ssthresh, 902 .tso_segs_goal = bbr_tso_segs_goal, 903 .get_info = bbr_get_info, 904 .set_state = bbr_set_state, 905 }; 906 907 static int __init bbr_register(void) 908 { 909 BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE); 910 return tcp_register_congestion_control(&tcp_bbr_cong_ops); 911 } 912 913 static void __exit bbr_unregister(void) 914 { 915 tcp_unregister_congestion_control(&tcp_bbr_cong_ops); 916 } 917 918 module_init(bbr_register); 919 module_exit(bbr_unregister); 920 921 MODULE_AUTHOR("Van Jacobson <vanj@google.com>"); 922 MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>"); 923 MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>"); 924 MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>"); 925 MODULE_LICENSE("Dual BSD/GPL"); 926 MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)"); 927