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 might be used with the fq qdisc ("man tc-fq") with pacing enabled, 56 * otherwise TCP stack falls back to an internal pacing using one high 57 * resolution timer per TCP socket and may use more resources. 58 */ 59 #include <linux/module.h> 60 #include <net/tcp.h> 61 #include <linux/inet_diag.h> 62 #include <linux/inet.h> 63 #include <linux/random.h> 64 #include <linux/win_minmax.h> 65 66 /* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth 67 * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps. 68 * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32. 69 * Since the minimum window is >=4 packets, the lower bound isn't 70 * an issue. The upper bound isn't an issue with existing technologies. 71 */ 72 #define BW_SCALE 24 73 #define BW_UNIT (1 << BW_SCALE) 74 75 #define BBR_SCALE 8 /* scaling factor for fractions in BBR (e.g. gains) */ 76 #define BBR_UNIT (1 << BBR_SCALE) 77 78 /* BBR has the following modes for deciding how fast to send: */ 79 enum bbr_mode { 80 BBR_STARTUP, /* ramp up sending rate rapidly to fill pipe */ 81 BBR_DRAIN, /* drain any queue created during startup */ 82 BBR_PROBE_BW, /* discover, share bw: pace around estimated bw */ 83 BBR_PROBE_RTT, /* cut inflight to min to probe min_rtt */ 84 }; 85 86 /* BBR congestion control block */ 87 struct bbr { 88 u32 min_rtt_us; /* min RTT in min_rtt_win_sec window */ 89 u32 min_rtt_stamp; /* timestamp of min_rtt_us */ 90 u32 probe_rtt_done_stamp; /* end time for BBR_PROBE_RTT mode */ 91 struct minmax bw; /* Max recent delivery rate in pkts/uS << 24 */ 92 u32 rtt_cnt; /* count of packet-timed rounds elapsed */ 93 u32 next_rtt_delivered; /* scb->tx.delivered at end of round */ 94 u64 cycle_mstamp; /* time of this cycle phase start */ 95 u32 mode:3, /* current bbr_mode in state machine */ 96 prev_ca_state:3, /* CA state on previous ACK */ 97 packet_conservation:1, /* use packet conservation? */ 98 round_start:1, /* start of packet-timed tx->ack round? */ 99 idle_restart:1, /* restarting after idle? */ 100 probe_rtt_round_done:1, /* a BBR_PROBE_RTT round at 4 pkts? */ 101 unused:13, 102 lt_is_sampling:1, /* taking long-term ("LT") samples now? */ 103 lt_rtt_cnt:7, /* round trips in long-term interval */ 104 lt_use_bw:1; /* use lt_bw as our bw estimate? */ 105 u32 lt_bw; /* LT est delivery rate in pkts/uS << 24 */ 106 u32 lt_last_delivered; /* LT intvl start: tp->delivered */ 107 u32 lt_last_stamp; /* LT intvl start: tp->delivered_mstamp */ 108 u32 lt_last_lost; /* LT intvl start: tp->lost */ 109 u32 pacing_gain:10, /* current gain for setting pacing rate */ 110 cwnd_gain:10, /* current gain for setting cwnd */ 111 full_bw_reached:1, /* reached full bw in Startup? */ 112 full_bw_cnt:2, /* number of rounds without large bw gains */ 113 cycle_idx:3, /* current index in pacing_gain cycle array */ 114 has_seen_rtt:1, /* have we seen an RTT sample yet? */ 115 unused_b:5; 116 u32 prior_cwnd; /* prior cwnd upon entering loss recovery */ 117 u32 full_bw; /* recent bw, to estimate if pipe is full */ 118 }; 119 120 #define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */ 121 122 /* Window length of bw filter (in rounds): */ 123 static const int bbr_bw_rtts = CYCLE_LEN + 2; 124 /* Window length of min_rtt filter (in sec): */ 125 static const u32 bbr_min_rtt_win_sec = 10; 126 /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */ 127 static const u32 bbr_probe_rtt_mode_ms = 200; 128 /* Skip TSO below the following bandwidth (bits/sec): */ 129 static const int bbr_min_tso_rate = 1200000; 130 131 /* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck. */ 132 static const int bbr_pacing_margin_percent = 1; 133 134 /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain 135 * that will allow a smoothly increasing pacing rate that will double each RTT 136 * and send the same number of packets per RTT that an un-paced, slow-starting 137 * Reno or CUBIC flow would: 138 */ 139 static const int bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 140 /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain 141 * the queue created in BBR_STARTUP in a single round: 142 */ 143 static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885; 144 /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */ 145 static const int bbr_cwnd_gain = BBR_UNIT * 2; 146 /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */ 147 static const int bbr_pacing_gain[] = { 148 BBR_UNIT * 5 / 4, /* probe for more available bw */ 149 BBR_UNIT * 3 / 4, /* drain queue and/or yield bw to other flows */ 150 BBR_UNIT, BBR_UNIT, BBR_UNIT, /* cruise at 1.0*bw to utilize pipe, */ 151 BBR_UNIT, BBR_UNIT, BBR_UNIT /* without creating excess queue... */ 152 }; 153 /* Randomize the starting gain cycling phase over N phases: */ 154 static const u32 bbr_cycle_rand = 7; 155 156 /* Try to keep at least this many packets in flight, if things go smoothly. For 157 * smooth functioning, a sliding window protocol ACKing every other packet 158 * needs at least 4 packets in flight: 159 */ 160 static const u32 bbr_cwnd_min_target = 4; 161 162 /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */ 163 /* If bw has increased significantly (1.25x), there may be more bw available: */ 164 static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4; 165 /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */ 166 static const u32 bbr_full_bw_cnt = 3; 167 168 /* "long-term" ("LT") bandwidth estimator parameters... */ 169 /* The minimum number of rounds in an LT bw sampling interval: */ 170 static const u32 bbr_lt_intvl_min_rtts = 4; 171 /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */ 172 static const u32 bbr_lt_loss_thresh = 50; 173 /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */ 174 static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8; 175 /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */ 176 static const u32 bbr_lt_bw_diff = 4000 / 8; 177 /* If we estimate we're policed, use lt_bw for this many round trips: */ 178 static const u32 bbr_lt_bw_max_rtts = 48; 179 180 static void bbr_check_probe_rtt_done(struct sock *sk); 181 182 /* Do we estimate that STARTUP filled the pipe? */ 183 static bool bbr_full_bw_reached(const struct sock *sk) 184 { 185 const struct bbr *bbr = inet_csk_ca(sk); 186 187 return bbr->full_bw_reached; 188 } 189 190 /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */ 191 static u32 bbr_max_bw(const struct sock *sk) 192 { 193 struct bbr *bbr = inet_csk_ca(sk); 194 195 return minmax_get(&bbr->bw); 196 } 197 198 /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */ 199 static u32 bbr_bw(const struct sock *sk) 200 { 201 struct bbr *bbr = inet_csk_ca(sk); 202 203 return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk); 204 } 205 206 /* Return rate in bytes per second, optionally with a gain. 207 * The order here is chosen carefully to avoid overflow of u64. This should 208 * work for input rates of up to 2.9Tbit/sec and gain of 2.89x. 209 */ 210 static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain) 211 { 212 unsigned int mss = tcp_sk(sk)->mss_cache; 213 214 rate *= mss; 215 rate *= gain; 216 rate >>= BBR_SCALE; 217 rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent); 218 return rate >> BW_SCALE; 219 } 220 221 /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */ 222 static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain) 223 { 224 u64 rate = bw; 225 226 rate = bbr_rate_bytes_per_sec(sk, rate, gain); 227 rate = min_t(u64, rate, sk->sk_max_pacing_rate); 228 return rate; 229 } 230 231 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */ 232 static void bbr_init_pacing_rate_from_rtt(struct sock *sk) 233 { 234 struct tcp_sock *tp = tcp_sk(sk); 235 struct bbr *bbr = inet_csk_ca(sk); 236 u64 bw; 237 u32 rtt_us; 238 239 if (tp->srtt_us) { /* any RTT sample yet? */ 240 rtt_us = max(tp->srtt_us >> 3, 1U); 241 bbr->has_seen_rtt = 1; 242 } else { /* no RTT sample yet */ 243 rtt_us = USEC_PER_MSEC; /* use nominal default RTT */ 244 } 245 bw = (u64)tp->snd_cwnd * BW_UNIT; 246 do_div(bw, rtt_us); 247 sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain); 248 } 249 250 /* Pace using current bw estimate and a gain factor. In order to help drive the 251 * network toward lower queues while maintaining high utilization and low 252 * latency, the average pacing rate aims to be slightly (~1%) lower than the 253 * estimated bandwidth. This is an important aspect of the design. In this 254 * implementation this slightly lower pacing rate is achieved implicitly by not 255 * including link-layer headers in the packet size used for the pacing rate. 256 */ 257 static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain) 258 { 259 struct tcp_sock *tp = tcp_sk(sk); 260 struct bbr *bbr = inet_csk_ca(sk); 261 unsigned long rate = bbr_bw_to_pacing_rate(sk, bw, gain); 262 263 if (unlikely(!bbr->has_seen_rtt && tp->srtt_us)) 264 bbr_init_pacing_rate_from_rtt(sk); 265 if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate) 266 sk->sk_pacing_rate = rate; 267 } 268 269 /* override sysctl_tcp_min_tso_segs */ 270 static u32 bbr_min_tso_segs(struct sock *sk) 271 { 272 return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2; 273 } 274 275 static u32 bbr_tso_segs_goal(struct sock *sk) 276 { 277 struct tcp_sock *tp = tcp_sk(sk); 278 u32 segs, bytes; 279 280 /* Sort of tcp_tso_autosize() but ignoring 281 * driver provided sk_gso_max_size. 282 */ 283 bytes = min_t(unsigned long, sk->sk_pacing_rate >> sk->sk_pacing_shift, 284 GSO_MAX_SIZE - 1 - MAX_TCP_HEADER); 285 segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk)); 286 287 return min(segs, 0x7FU); 288 } 289 290 /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ 291 static void bbr_save_cwnd(struct sock *sk) 292 { 293 struct tcp_sock *tp = tcp_sk(sk); 294 struct bbr *bbr = inet_csk_ca(sk); 295 296 if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT) 297 bbr->prior_cwnd = tp->snd_cwnd; /* this cwnd is good enough */ 298 else /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */ 299 bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd); 300 } 301 302 static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event) 303 { 304 struct tcp_sock *tp = tcp_sk(sk); 305 struct bbr *bbr = inet_csk_ca(sk); 306 307 if (event == CA_EVENT_TX_START && tp->app_limited) { 308 bbr->idle_restart = 1; 309 /* Avoid pointless buffer overflows: pace at est. bw if we don't 310 * need more speed (we're restarting from idle and app-limited). 311 */ 312 if (bbr->mode == BBR_PROBE_BW) 313 bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT); 314 else if (bbr->mode == BBR_PROBE_RTT) 315 bbr_check_probe_rtt_done(sk); 316 } 317 } 318 319 /* Find target cwnd. Right-size the cwnd based on min RTT and the 320 * estimated bottleneck bandwidth: 321 * 322 * cwnd = bw * min_rtt * gain = BDP * gain 323 * 324 * The key factor, gain, controls the amount of queue. While a small gain 325 * builds a smaller queue, it becomes more vulnerable to noise in RTT 326 * measurements (e.g., delayed ACKs or other ACK compression effects). This 327 * noise may cause BBR to under-estimate the rate. 328 * 329 * To achieve full performance in high-speed paths, we budget enough cwnd to 330 * fit full-sized skbs in-flight on both end hosts to fully utilize the path: 331 * - one skb in sending host Qdisc, 332 * - one skb in sending host TSO/GSO engine 333 * - one skb being received by receiver host LRO/GRO/delayed-ACK engine 334 * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because 335 * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets, 336 * which allows 2 outstanding 2-packet sequences, to try to keep pipe 337 * full even with ACK-every-other-packet delayed ACKs. 338 */ 339 static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain) 340 { 341 struct bbr *bbr = inet_csk_ca(sk); 342 u32 cwnd; 343 u64 w; 344 345 /* If we've never had a valid RTT sample, cap cwnd at the initial 346 * default. This should only happen when the connection is not using TCP 347 * timestamps and has retransmitted all of the SYN/SYNACK/data packets 348 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which 349 * case we need to slow-start up toward something safe: TCP_INIT_CWND. 350 */ 351 if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */ 352 return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/ 353 354 w = (u64)bw * bbr->min_rtt_us; 355 356 /* Apply a gain to the given value, then remove the BW_SCALE shift. */ 357 cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT; 358 359 /* Allow enough full-sized skbs in flight to utilize end systems. */ 360 cwnd += 3 * bbr_tso_segs_goal(sk); 361 362 /* Reduce delayed ACKs by rounding up cwnd to the next even number. */ 363 cwnd = (cwnd + 1) & ~1U; 364 365 /* Ensure gain cycling gets inflight above BDP even for small BDPs. */ 366 if (bbr->mode == BBR_PROBE_BW && gain > BBR_UNIT) 367 cwnd += 2; 368 369 return cwnd; 370 } 371 372 /* With pacing at lower layers, there's often less data "in the network" than 373 * "in flight". With TSQ and departure time pacing at lower layers (e.g. fq), 374 * we often have several skbs queued in the pacing layer with a pre-scheduled 375 * earliest departure time (EDT). BBR adapts its pacing rate based on the 376 * inflight level that it estimates has already been "baked in" by previous 377 * departure time decisions. We calculate a rough estimate of the number of our 378 * packets that might be in the network at the earliest departure time for the 379 * next skb scheduled: 380 * in_network_at_edt = inflight_at_edt - (EDT - now) * bw 381 * If we're increasing inflight, then we want to know if the transmit of the 382 * EDT skb will push inflight above the target, so inflight_at_edt includes 383 * bbr_tso_segs_goal() from the skb departing at EDT. If decreasing inflight, 384 * then estimate if inflight will sink too low just before the EDT transmit. 385 */ 386 static u32 bbr_packets_in_net_at_edt(struct sock *sk, u32 inflight_now) 387 { 388 struct tcp_sock *tp = tcp_sk(sk); 389 struct bbr *bbr = inet_csk_ca(sk); 390 u64 now_ns, edt_ns, interval_us; 391 u32 interval_delivered, inflight_at_edt; 392 393 now_ns = tp->tcp_clock_cache; 394 edt_ns = max(tp->tcp_wstamp_ns, now_ns); 395 interval_us = div_u64(edt_ns - now_ns, NSEC_PER_USEC); 396 interval_delivered = (u64)bbr_bw(sk) * interval_us >> BW_SCALE; 397 inflight_at_edt = inflight_now; 398 if (bbr->pacing_gain > BBR_UNIT) /* increasing inflight */ 399 inflight_at_edt += bbr_tso_segs_goal(sk); /* include EDT skb */ 400 if (interval_delivered >= inflight_at_edt) 401 return 0; 402 return inflight_at_edt - interval_delivered; 403 } 404 405 /* An optimization in BBR to reduce losses: On the first round of recovery, we 406 * follow the packet conservation principle: send P packets per P packets acked. 407 * After that, we slow-start and send at most 2*P packets per P packets acked. 408 * After recovery finishes, or upon undo, we restore the cwnd we had when 409 * recovery started (capped by the target cwnd based on estimated BDP). 410 * 411 * TODO(ycheng/ncardwell): implement a rate-based approach. 412 */ 413 static bool bbr_set_cwnd_to_recover_or_restore( 414 struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd) 415 { 416 struct tcp_sock *tp = tcp_sk(sk); 417 struct bbr *bbr = inet_csk_ca(sk); 418 u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state; 419 u32 cwnd = tp->snd_cwnd; 420 421 /* An ACK for P pkts should release at most 2*P packets. We do this 422 * in two steps. First, here we deduct the number of lost packets. 423 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd. 424 */ 425 if (rs->losses > 0) 426 cwnd = max_t(s32, cwnd - rs->losses, 1); 427 428 if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) { 429 /* Starting 1st round of Recovery, so do packet conservation. */ 430 bbr->packet_conservation = 1; 431 bbr->next_rtt_delivered = tp->delivered; /* start round now */ 432 /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */ 433 cwnd = tcp_packets_in_flight(tp) + acked; 434 } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) { 435 /* Exiting loss recovery; restore cwnd saved before recovery. */ 436 cwnd = max(cwnd, bbr->prior_cwnd); 437 bbr->packet_conservation = 0; 438 } 439 bbr->prev_ca_state = state; 440 441 if (bbr->packet_conservation) { 442 *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked); 443 return true; /* yes, using packet conservation */ 444 } 445 *new_cwnd = cwnd; 446 return false; 447 } 448 449 /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss 450 * has drawn us down below target), or snap down to target if we're above it. 451 */ 452 static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs, 453 u32 acked, u32 bw, int gain) 454 { 455 struct tcp_sock *tp = tcp_sk(sk); 456 struct bbr *bbr = inet_csk_ca(sk); 457 u32 cwnd = tp->snd_cwnd, target_cwnd = 0; 458 459 if (!acked) 460 goto done; /* no packet fully ACKed; just apply caps */ 461 462 if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd)) 463 goto done; 464 465 /* If we're below target cwnd, slow start cwnd toward target cwnd. */ 466 target_cwnd = bbr_target_cwnd(sk, bw, gain); 467 if (bbr_full_bw_reached(sk)) /* only cut cwnd if we filled the pipe */ 468 cwnd = min(cwnd + acked, target_cwnd); 469 else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND) 470 cwnd = cwnd + acked; 471 cwnd = max(cwnd, bbr_cwnd_min_target); 472 473 done: 474 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); /* apply global cap */ 475 if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */ 476 tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target); 477 } 478 479 /* End cycle phase if it's time and/or we hit the phase's in-flight target. */ 480 static bool bbr_is_next_cycle_phase(struct sock *sk, 481 const struct rate_sample *rs) 482 { 483 struct tcp_sock *tp = tcp_sk(sk); 484 struct bbr *bbr = inet_csk_ca(sk); 485 bool is_full_length = 486 tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) > 487 bbr->min_rtt_us; 488 u32 inflight, bw; 489 490 /* The pacing_gain of 1.0 paces at the estimated bw to try to fully 491 * use the pipe without increasing the queue. 492 */ 493 if (bbr->pacing_gain == BBR_UNIT) 494 return is_full_length; /* just use wall clock time */ 495 496 inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight); 497 bw = bbr_max_bw(sk); 498 499 /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at 500 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is 501 * small (e.g. on a LAN). We do not persist if packets are lost, since 502 * a path with small buffers may not hold that much. 503 */ 504 if (bbr->pacing_gain > BBR_UNIT) 505 return is_full_length && 506 (rs->losses || /* perhaps pacing_gain*BDP won't fit */ 507 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain)); 508 509 /* A pacing_gain < 1.0 tries to drain extra queue we added if bw 510 * probing didn't find more bw. If inflight falls to match BDP then we 511 * estimate queue is drained; persisting would underutilize the pipe. 512 */ 513 return is_full_length || 514 inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT); 515 } 516 517 static void bbr_advance_cycle_phase(struct sock *sk) 518 { 519 struct tcp_sock *tp = tcp_sk(sk); 520 struct bbr *bbr = inet_csk_ca(sk); 521 522 bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1); 523 bbr->cycle_mstamp = tp->delivered_mstamp; 524 } 525 526 /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */ 527 static void bbr_update_cycle_phase(struct sock *sk, 528 const struct rate_sample *rs) 529 { 530 struct bbr *bbr = inet_csk_ca(sk); 531 532 if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs)) 533 bbr_advance_cycle_phase(sk); 534 } 535 536 static void bbr_reset_startup_mode(struct sock *sk) 537 { 538 struct bbr *bbr = inet_csk_ca(sk); 539 540 bbr->mode = BBR_STARTUP; 541 } 542 543 static void bbr_reset_probe_bw_mode(struct sock *sk) 544 { 545 struct bbr *bbr = inet_csk_ca(sk); 546 547 bbr->mode = BBR_PROBE_BW; 548 bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand); 549 bbr_advance_cycle_phase(sk); /* flip to next phase of gain cycle */ 550 } 551 552 static void bbr_reset_mode(struct sock *sk) 553 { 554 if (!bbr_full_bw_reached(sk)) 555 bbr_reset_startup_mode(sk); 556 else 557 bbr_reset_probe_bw_mode(sk); 558 } 559 560 /* Start a new long-term sampling interval. */ 561 static void bbr_reset_lt_bw_sampling_interval(struct sock *sk) 562 { 563 struct tcp_sock *tp = tcp_sk(sk); 564 struct bbr *bbr = inet_csk_ca(sk); 565 566 bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC); 567 bbr->lt_last_delivered = tp->delivered; 568 bbr->lt_last_lost = tp->lost; 569 bbr->lt_rtt_cnt = 0; 570 } 571 572 /* Completely reset long-term bandwidth sampling. */ 573 static void bbr_reset_lt_bw_sampling(struct sock *sk) 574 { 575 struct bbr *bbr = inet_csk_ca(sk); 576 577 bbr->lt_bw = 0; 578 bbr->lt_use_bw = 0; 579 bbr->lt_is_sampling = false; 580 bbr_reset_lt_bw_sampling_interval(sk); 581 } 582 583 /* Long-term bw sampling interval is done. Estimate whether we're policed. */ 584 static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw) 585 { 586 struct bbr *bbr = inet_csk_ca(sk); 587 u32 diff; 588 589 if (bbr->lt_bw) { /* do we have bw from a previous interval? */ 590 /* Is new bw close to the lt_bw from the previous interval? */ 591 diff = abs(bw - bbr->lt_bw); 592 if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) || 593 (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <= 594 bbr_lt_bw_diff)) { 595 /* All criteria are met; estimate we're policed. */ 596 bbr->lt_bw = (bw + bbr->lt_bw) >> 1; /* avg 2 intvls */ 597 bbr->lt_use_bw = 1; 598 bbr->pacing_gain = BBR_UNIT; /* try to avoid drops */ 599 bbr->lt_rtt_cnt = 0; 600 return; 601 } 602 } 603 bbr->lt_bw = bw; 604 bbr_reset_lt_bw_sampling_interval(sk); 605 } 606 607 /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of 608 * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and 609 * explicitly models their policed rate, to reduce unnecessary losses. We 610 * estimate that we're policed if we see 2 consecutive sampling intervals with 611 * consistent throughput and high packet loss. If we think we're being policed, 612 * set lt_bw to the "long-term" average delivery rate from those 2 intervals. 613 */ 614 static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs) 615 { 616 struct tcp_sock *tp = tcp_sk(sk); 617 struct bbr *bbr = inet_csk_ca(sk); 618 u32 lost, delivered; 619 u64 bw; 620 u32 t; 621 622 if (bbr->lt_use_bw) { /* already using long-term rate, lt_bw? */ 623 if (bbr->mode == BBR_PROBE_BW && bbr->round_start && 624 ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) { 625 bbr_reset_lt_bw_sampling(sk); /* stop using lt_bw */ 626 bbr_reset_probe_bw_mode(sk); /* restart gain cycling */ 627 } 628 return; 629 } 630 631 /* Wait for the first loss before sampling, to let the policer exhaust 632 * its tokens and estimate the steady-state rate allowed by the policer. 633 * Starting samples earlier includes bursts that over-estimate the bw. 634 */ 635 if (!bbr->lt_is_sampling) { 636 if (!rs->losses) 637 return; 638 bbr_reset_lt_bw_sampling_interval(sk); 639 bbr->lt_is_sampling = true; 640 } 641 642 /* To avoid underestimates, reset sampling if we run out of data. */ 643 if (rs->is_app_limited) { 644 bbr_reset_lt_bw_sampling(sk); 645 return; 646 } 647 648 if (bbr->round_start) 649 bbr->lt_rtt_cnt++; /* count round trips in this interval */ 650 if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts) 651 return; /* sampling interval needs to be longer */ 652 if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) { 653 bbr_reset_lt_bw_sampling(sk); /* interval is too long */ 654 return; 655 } 656 657 /* End sampling interval when a packet is lost, so we estimate the 658 * policer tokens were exhausted. Stopping the sampling before the 659 * tokens are exhausted under-estimates the policed rate. 660 */ 661 if (!rs->losses) 662 return; 663 664 /* Calculate packets lost and delivered in sampling interval. */ 665 lost = tp->lost - bbr->lt_last_lost; 666 delivered = tp->delivered - bbr->lt_last_delivered; 667 /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */ 668 if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered) 669 return; 670 671 /* Find average delivery rate in this sampling interval. */ 672 t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp; 673 if ((s32)t < 1) 674 return; /* interval is less than one ms, so wait */ 675 /* Check if can multiply without overflow */ 676 if (t >= ~0U / USEC_PER_MSEC) { 677 bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */ 678 return; 679 } 680 t *= USEC_PER_MSEC; 681 bw = (u64)delivered * BW_UNIT; 682 do_div(bw, t); 683 bbr_lt_bw_interval_done(sk, bw); 684 } 685 686 /* Estimate the bandwidth based on how fast packets are delivered */ 687 static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs) 688 { 689 struct tcp_sock *tp = tcp_sk(sk); 690 struct bbr *bbr = inet_csk_ca(sk); 691 u64 bw; 692 693 bbr->round_start = 0; 694 if (rs->delivered < 0 || rs->interval_us <= 0) 695 return; /* Not a valid observation */ 696 697 /* See if we've reached the next RTT */ 698 if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) { 699 bbr->next_rtt_delivered = tp->delivered; 700 bbr->rtt_cnt++; 701 bbr->round_start = 1; 702 bbr->packet_conservation = 0; 703 } 704 705 bbr_lt_bw_sampling(sk, rs); 706 707 /* Divide delivered by the interval to find a (lower bound) bottleneck 708 * bandwidth sample. Delivered is in packets and interval_us in uS and 709 * ratio will be <<1 for most connections. So delivered is first scaled. 710 */ 711 bw = (u64)rs->delivered * BW_UNIT; 712 do_div(bw, rs->interval_us); 713 714 /* If this sample is application-limited, it is likely to have a very 715 * low delivered count that represents application behavior rather than 716 * the available network rate. Such a sample could drag down estimated 717 * bw, causing needless slow-down. Thus, to continue to send at the 718 * last measured network rate, we filter out app-limited samples unless 719 * they describe the path bw at least as well as our bw model. 720 * 721 * So the goal during app-limited phase is to proceed with the best 722 * network rate no matter how long. We automatically leave this 723 * phase when app writes faster than the network can deliver :) 724 */ 725 if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) { 726 /* Incorporate new sample into our max bw filter. */ 727 minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw); 728 } 729 } 730 731 /* Estimate when the pipe is full, using the change in delivery rate: BBR 732 * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by 733 * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited 734 * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the 735 * higher rwin, 3: we get higher delivery rate samples. Or transient 736 * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar 737 * design goal, but uses delay and inter-ACK spacing instead of bandwidth. 738 */ 739 static void bbr_check_full_bw_reached(struct sock *sk, 740 const struct rate_sample *rs) 741 { 742 struct bbr *bbr = inet_csk_ca(sk); 743 u32 bw_thresh; 744 745 if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited) 746 return; 747 748 bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE; 749 if (bbr_max_bw(sk) >= bw_thresh) { 750 bbr->full_bw = bbr_max_bw(sk); 751 bbr->full_bw_cnt = 0; 752 return; 753 } 754 ++bbr->full_bw_cnt; 755 bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt; 756 } 757 758 /* If pipe is probably full, drain the queue and then enter steady-state. */ 759 static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs) 760 { 761 struct bbr *bbr = inet_csk_ca(sk); 762 763 if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) { 764 bbr->mode = BBR_DRAIN; /* drain queue we created */ 765 tcp_sk(sk)->snd_ssthresh = 766 bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT); 767 } /* fall through to check if in-flight is already small: */ 768 if (bbr->mode == BBR_DRAIN && 769 bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <= 770 bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT)) 771 bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */ 772 } 773 774 static void bbr_check_probe_rtt_done(struct sock *sk) 775 { 776 struct tcp_sock *tp = tcp_sk(sk); 777 struct bbr *bbr = inet_csk_ca(sk); 778 779 if (!(bbr->probe_rtt_done_stamp && 780 after(tcp_jiffies32, bbr->probe_rtt_done_stamp))) 781 return; 782 783 bbr->min_rtt_stamp = tcp_jiffies32; /* wait a while until PROBE_RTT */ 784 tp->snd_cwnd = max(tp->snd_cwnd, bbr->prior_cwnd); 785 bbr_reset_mode(sk); 786 } 787 788 /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and 789 * periodically drain the bottleneck queue, to converge to measure the true 790 * min_rtt (unloaded propagation delay). This allows the flows to keep queues 791 * small (reducing queuing delay and packet loss) and achieve fairness among 792 * BBR flows. 793 * 794 * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires, 795 * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets. 796 * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed 797 * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and 798 * re-enter the previous mode. BBR uses 200ms to approximately bound the 799 * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s). 800 * 801 * Note that flows need only pay 2% if they are busy sending over the last 10 802 * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have 803 * natural silences or low-rate periods within 10 seconds where the rate is low 804 * enough for long enough to drain its queue in the bottleneck. We pick up 805 * these min RTT measurements opportunistically with our min_rtt filter. :-) 806 */ 807 static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs) 808 { 809 struct tcp_sock *tp = tcp_sk(sk); 810 struct bbr *bbr = inet_csk_ca(sk); 811 bool filter_expired; 812 813 /* Track min RTT seen in the min_rtt_win_sec filter window: */ 814 filter_expired = after(tcp_jiffies32, 815 bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ); 816 if (rs->rtt_us >= 0 && 817 (rs->rtt_us <= bbr->min_rtt_us || 818 (filter_expired && !rs->is_ack_delayed))) { 819 bbr->min_rtt_us = rs->rtt_us; 820 bbr->min_rtt_stamp = tcp_jiffies32; 821 } 822 823 if (bbr_probe_rtt_mode_ms > 0 && filter_expired && 824 !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) { 825 bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */ 826 bbr_save_cwnd(sk); /* note cwnd so we can restore it */ 827 bbr->probe_rtt_done_stamp = 0; 828 } 829 830 if (bbr->mode == BBR_PROBE_RTT) { 831 /* Ignore low rate samples during this mode. */ 832 tp->app_limited = 833 (tp->delivered + tcp_packets_in_flight(tp)) ? : 1; 834 /* Maintain min packets in flight for max(200 ms, 1 round). */ 835 if (!bbr->probe_rtt_done_stamp && 836 tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) { 837 bbr->probe_rtt_done_stamp = tcp_jiffies32 + 838 msecs_to_jiffies(bbr_probe_rtt_mode_ms); 839 bbr->probe_rtt_round_done = 0; 840 bbr->next_rtt_delivered = tp->delivered; 841 } else if (bbr->probe_rtt_done_stamp) { 842 if (bbr->round_start) 843 bbr->probe_rtt_round_done = 1; 844 if (bbr->probe_rtt_round_done) 845 bbr_check_probe_rtt_done(sk); 846 } 847 } 848 /* Restart after idle ends only once we process a new S/ACK for data */ 849 if (rs->delivered > 0) 850 bbr->idle_restart = 0; 851 } 852 853 static void bbr_update_gains(struct sock *sk) 854 { 855 struct bbr *bbr = inet_csk_ca(sk); 856 857 switch (bbr->mode) { 858 case BBR_STARTUP: 859 bbr->pacing_gain = bbr_high_gain; 860 bbr->cwnd_gain = bbr_high_gain; 861 break; 862 case BBR_DRAIN: 863 bbr->pacing_gain = bbr_drain_gain; /* slow, to drain */ 864 bbr->cwnd_gain = bbr_high_gain; /* keep cwnd */ 865 break; 866 case BBR_PROBE_BW: 867 bbr->pacing_gain = (bbr->lt_use_bw ? 868 BBR_UNIT : 869 bbr_pacing_gain[bbr->cycle_idx]); 870 bbr->cwnd_gain = bbr_cwnd_gain; 871 break; 872 case BBR_PROBE_RTT: 873 bbr->pacing_gain = BBR_UNIT; 874 bbr->cwnd_gain = BBR_UNIT; 875 break; 876 default: 877 WARN_ONCE(1, "BBR bad mode: %u\n", bbr->mode); 878 break; 879 } 880 } 881 882 static void bbr_update_model(struct sock *sk, const struct rate_sample *rs) 883 { 884 bbr_update_bw(sk, rs); 885 bbr_update_cycle_phase(sk, rs); 886 bbr_check_full_bw_reached(sk, rs); 887 bbr_check_drain(sk, rs); 888 bbr_update_min_rtt(sk, rs); 889 bbr_update_gains(sk); 890 } 891 892 static void bbr_main(struct sock *sk, const struct rate_sample *rs) 893 { 894 struct bbr *bbr = inet_csk_ca(sk); 895 u32 bw; 896 897 bbr_update_model(sk, rs); 898 899 bw = bbr_bw(sk); 900 bbr_set_pacing_rate(sk, bw, bbr->pacing_gain); 901 bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain); 902 } 903 904 static void bbr_init(struct sock *sk) 905 { 906 struct tcp_sock *tp = tcp_sk(sk); 907 struct bbr *bbr = inet_csk_ca(sk); 908 909 bbr->prior_cwnd = 0; 910 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 911 bbr->rtt_cnt = 0; 912 bbr->next_rtt_delivered = 0; 913 bbr->prev_ca_state = TCP_CA_Open; 914 bbr->packet_conservation = 0; 915 916 bbr->probe_rtt_done_stamp = 0; 917 bbr->probe_rtt_round_done = 0; 918 bbr->min_rtt_us = tcp_min_rtt(tp); 919 bbr->min_rtt_stamp = tcp_jiffies32; 920 921 minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */ 922 923 bbr->has_seen_rtt = 0; 924 bbr_init_pacing_rate_from_rtt(sk); 925 926 bbr->round_start = 0; 927 bbr->idle_restart = 0; 928 bbr->full_bw_reached = 0; 929 bbr->full_bw = 0; 930 bbr->full_bw_cnt = 0; 931 bbr->cycle_mstamp = 0; 932 bbr->cycle_idx = 0; 933 bbr_reset_lt_bw_sampling(sk); 934 bbr_reset_startup_mode(sk); 935 936 cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED); 937 } 938 939 static u32 bbr_sndbuf_expand(struct sock *sk) 940 { 941 /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ 942 return 3; 943 } 944 945 /* In theory BBR does not need to undo the cwnd since it does not 946 * always reduce cwnd on losses (see bbr_main()). Keep it for now. 947 */ 948 static u32 bbr_undo_cwnd(struct sock *sk) 949 { 950 struct bbr *bbr = inet_csk_ca(sk); 951 952 bbr->full_bw = 0; /* spurious slow-down; reset full pipe detection */ 953 bbr->full_bw_cnt = 0; 954 bbr_reset_lt_bw_sampling(sk); 955 return tcp_sk(sk)->snd_cwnd; 956 } 957 958 /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */ 959 static u32 bbr_ssthresh(struct sock *sk) 960 { 961 bbr_save_cwnd(sk); 962 return tcp_sk(sk)->snd_ssthresh; 963 } 964 965 static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr, 966 union tcp_cc_info *info) 967 { 968 if (ext & (1 << (INET_DIAG_BBRINFO - 1)) || 969 ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 970 struct tcp_sock *tp = tcp_sk(sk); 971 struct bbr *bbr = inet_csk_ca(sk); 972 u64 bw = bbr_bw(sk); 973 974 bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE; 975 memset(&info->bbr, 0, sizeof(info->bbr)); 976 info->bbr.bbr_bw_lo = (u32)bw; 977 info->bbr.bbr_bw_hi = (u32)(bw >> 32); 978 info->bbr.bbr_min_rtt = bbr->min_rtt_us; 979 info->bbr.bbr_pacing_gain = bbr->pacing_gain; 980 info->bbr.bbr_cwnd_gain = bbr->cwnd_gain; 981 *attr = INET_DIAG_BBRINFO; 982 return sizeof(info->bbr); 983 } 984 return 0; 985 } 986 987 static void bbr_set_state(struct sock *sk, u8 new_state) 988 { 989 struct bbr *bbr = inet_csk_ca(sk); 990 991 if (new_state == TCP_CA_Loss) { 992 struct rate_sample rs = { .losses = 1 }; 993 994 bbr->prev_ca_state = TCP_CA_Loss; 995 bbr->full_bw = 0; 996 bbr->round_start = 1; /* treat RTO like end of a round */ 997 bbr_lt_bw_sampling(sk, &rs); 998 } 999 } 1000 1001 static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { 1002 .flags = TCP_CONG_NON_RESTRICTED, 1003 .name = "bbr", 1004 .owner = THIS_MODULE, 1005 .init = bbr_init, 1006 .cong_control = bbr_main, 1007 .sndbuf_expand = bbr_sndbuf_expand, 1008 .undo_cwnd = bbr_undo_cwnd, 1009 .cwnd_event = bbr_cwnd_event, 1010 .ssthresh = bbr_ssthresh, 1011 .min_tso_segs = bbr_min_tso_segs, 1012 .get_info = bbr_get_info, 1013 .set_state = bbr_set_state, 1014 }; 1015 1016 static int __init bbr_register(void) 1017 { 1018 BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE); 1019 return tcp_register_congestion_control(&tcp_bbr_cong_ops); 1020 } 1021 1022 static void __exit bbr_unregister(void) 1023 { 1024 tcp_unregister_congestion_control(&tcp_bbr_cong_ops); 1025 } 1026 1027 module_init(bbr_register); 1028 module_exit(bbr_unregister); 1029 1030 MODULE_AUTHOR("Van Jacobson <vanj@google.com>"); 1031 MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>"); 1032 MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>"); 1033 MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>"); 1034 MODULE_LICENSE("Dual BSD/GPL"); 1035 MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)"); 1036