10f8782eaSNeal Cardwell /* Bottleneck Bandwidth and RTT (BBR) congestion control 20f8782eaSNeal Cardwell * 30f8782eaSNeal Cardwell * BBR congestion control computes the sending rate based on the delivery 40f8782eaSNeal Cardwell * rate (throughput) estimated from ACKs. In a nutshell: 50f8782eaSNeal Cardwell * 60f8782eaSNeal Cardwell * On each ACK, update our model of the network path: 70f8782eaSNeal Cardwell * bottleneck_bandwidth = windowed_max(delivered / elapsed, 10 round trips) 80f8782eaSNeal Cardwell * min_rtt = windowed_min(rtt, 10 seconds) 90f8782eaSNeal Cardwell * pacing_rate = pacing_gain * bottleneck_bandwidth 100f8782eaSNeal Cardwell * cwnd = max(cwnd_gain * bottleneck_bandwidth * min_rtt, 4) 110f8782eaSNeal Cardwell * 120f8782eaSNeal Cardwell * The core algorithm does not react directly to packet losses or delays, 130f8782eaSNeal Cardwell * although BBR may adjust the size of next send per ACK when loss is 140f8782eaSNeal Cardwell * observed, or adjust the sending rate if it estimates there is a 150f8782eaSNeal Cardwell * traffic policer, in order to keep the drop rate reasonable. 160f8782eaSNeal Cardwell * 179b9375b5SNeal Cardwell * Here is a state transition diagram for BBR: 189b9375b5SNeal Cardwell * 199b9375b5SNeal Cardwell * | 209b9375b5SNeal Cardwell * V 219b9375b5SNeal Cardwell * +---> STARTUP ----+ 229b9375b5SNeal Cardwell * | | | 239b9375b5SNeal Cardwell * | V | 249b9375b5SNeal Cardwell * | DRAIN ----+ 259b9375b5SNeal Cardwell * | | | 269b9375b5SNeal Cardwell * | V | 279b9375b5SNeal Cardwell * +---> PROBE_BW ----+ 289b9375b5SNeal Cardwell * | ^ | | 299b9375b5SNeal Cardwell * | | | | 309b9375b5SNeal Cardwell * | +----+ | 319b9375b5SNeal Cardwell * | | 329b9375b5SNeal Cardwell * +---- PROBE_RTT <--+ 339b9375b5SNeal Cardwell * 349b9375b5SNeal Cardwell * A BBR flow starts in STARTUP, and ramps up its sending rate quickly. 359b9375b5SNeal Cardwell * When it estimates the pipe is full, it enters DRAIN to drain the queue. 369b9375b5SNeal Cardwell * In steady state a BBR flow only uses PROBE_BW and PROBE_RTT. 379b9375b5SNeal Cardwell * A long-lived BBR flow spends the vast majority of its time remaining 389b9375b5SNeal Cardwell * (repeatedly) in PROBE_BW, fully probing and utilizing the pipe's bandwidth 399b9375b5SNeal Cardwell * in a fair manner, with a small, bounded queue. *If* a flow has been 409b9375b5SNeal Cardwell * continuously sending for the entire min_rtt window, and hasn't seen an RTT 419b9375b5SNeal Cardwell * sample that matches or decreases its min_rtt estimate for 10 seconds, then 429b9375b5SNeal Cardwell * it briefly enters PROBE_RTT to cut inflight to a minimum value to re-probe 439b9375b5SNeal Cardwell * the path's two-way propagation delay (min_rtt). When exiting PROBE_RTT, if 449b9375b5SNeal Cardwell * we estimated that we reached the full bw of the pipe then we enter PROBE_BW; 459b9375b5SNeal Cardwell * otherwise we enter STARTUP to try to fill the pipe. 469b9375b5SNeal Cardwell * 470f8782eaSNeal Cardwell * BBR is described in detail in: 480f8782eaSNeal Cardwell * "BBR: Congestion-Based Congestion Control", 490f8782eaSNeal Cardwell * Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh, 500f8782eaSNeal Cardwell * Van Jacobson. ACM Queue, Vol. 14 No. 5, September-October 2016. 510f8782eaSNeal Cardwell * 520f8782eaSNeal Cardwell * There is a public e-mail list for discussing BBR development and testing: 530f8782eaSNeal Cardwell * https://groups.google.com/forum/#!forum/bbr-dev 540f8782eaSNeal Cardwell * 55218af599SEric Dumazet * NOTE: BBR might be used with the fq qdisc ("man tc-fq") with pacing enabled, 56218af599SEric Dumazet * otherwise TCP stack falls back to an internal pacing using one high 57218af599SEric Dumazet * resolution timer per TCP socket and may use more resources. 580f8782eaSNeal Cardwell */ 590f8782eaSNeal Cardwell #include <linux/module.h> 600f8782eaSNeal Cardwell #include <net/tcp.h> 610f8782eaSNeal Cardwell #include <linux/inet_diag.h> 620f8782eaSNeal Cardwell #include <linux/inet.h> 630f8782eaSNeal Cardwell #include <linux/random.h> 640f8782eaSNeal Cardwell #include <linux/win_minmax.h> 650f8782eaSNeal Cardwell 660f8782eaSNeal Cardwell /* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth 670f8782eaSNeal Cardwell * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps. 680f8782eaSNeal Cardwell * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32. 690f8782eaSNeal Cardwell * Since the minimum window is >=4 packets, the lower bound isn't 700f8782eaSNeal Cardwell * an issue. The upper bound isn't an issue with existing technologies. 710f8782eaSNeal Cardwell */ 720f8782eaSNeal Cardwell #define BW_SCALE 24 730f8782eaSNeal Cardwell #define BW_UNIT (1 << BW_SCALE) 740f8782eaSNeal Cardwell 750f8782eaSNeal Cardwell #define BBR_SCALE 8 /* scaling factor for fractions in BBR (e.g. gains) */ 760f8782eaSNeal Cardwell #define BBR_UNIT (1 << BBR_SCALE) 770f8782eaSNeal Cardwell 780f8782eaSNeal Cardwell /* BBR has the following modes for deciding how fast to send: */ 790f8782eaSNeal Cardwell enum bbr_mode { 800f8782eaSNeal Cardwell BBR_STARTUP, /* ramp up sending rate rapidly to fill pipe */ 810f8782eaSNeal Cardwell BBR_DRAIN, /* drain any queue created during startup */ 820f8782eaSNeal Cardwell BBR_PROBE_BW, /* discover, share bw: pace around estimated bw */ 839b9375b5SNeal Cardwell BBR_PROBE_RTT, /* cut inflight to min to probe min_rtt */ 840f8782eaSNeal Cardwell }; 850f8782eaSNeal Cardwell 860f8782eaSNeal Cardwell /* BBR congestion control block */ 870f8782eaSNeal Cardwell struct bbr { 880f8782eaSNeal Cardwell u32 min_rtt_us; /* min RTT in min_rtt_win_sec window */ 890f8782eaSNeal Cardwell u32 min_rtt_stamp; /* timestamp of min_rtt_us */ 900f8782eaSNeal Cardwell u32 probe_rtt_done_stamp; /* end time for BBR_PROBE_RTT mode */ 910f8782eaSNeal Cardwell struct minmax bw; /* Max recent delivery rate in pkts/uS << 24 */ 920f8782eaSNeal Cardwell u32 rtt_cnt; /* count of packet-timed rounds elapsed */ 930f8782eaSNeal Cardwell u32 next_rtt_delivered; /* scb->tx.delivered at end of round */ 949a568de4SEric Dumazet u64 cycle_mstamp; /* time of this cycle phase start */ 950f8782eaSNeal Cardwell u32 mode:3, /* current bbr_mode in state machine */ 960f8782eaSNeal Cardwell prev_ca_state:3, /* CA state on previous ACK */ 970f8782eaSNeal Cardwell packet_conservation:1, /* use packet conservation? */ 980f8782eaSNeal Cardwell round_start:1, /* start of packet-timed tx->ack round? */ 990f8782eaSNeal Cardwell idle_restart:1, /* restarting after idle? */ 1000f8782eaSNeal Cardwell probe_rtt_round_done:1, /* a BBR_PROBE_RTT round at 4 pkts? */ 101fb998862SKevin Yang unused:13, 1020f8782eaSNeal Cardwell lt_is_sampling:1, /* taking long-term ("LT") samples now? */ 1030f8782eaSNeal Cardwell lt_rtt_cnt:7, /* round trips in long-term interval */ 1040f8782eaSNeal Cardwell lt_use_bw:1; /* use lt_bw as our bw estimate? */ 1050f8782eaSNeal Cardwell u32 lt_bw; /* LT est delivery rate in pkts/uS << 24 */ 1060f8782eaSNeal Cardwell u32 lt_last_delivered; /* LT intvl start: tp->delivered */ 1070f8782eaSNeal Cardwell u32 lt_last_stamp; /* LT intvl start: tp->delivered_mstamp */ 1080f8782eaSNeal Cardwell u32 lt_last_lost; /* LT intvl start: tp->lost */ 1090f8782eaSNeal Cardwell u32 pacing_gain:10, /* current gain for setting pacing rate */ 1100f8782eaSNeal Cardwell cwnd_gain:10, /* current gain for setting cwnd */ 111c589e69bSNeal Cardwell full_bw_reached:1, /* reached full bw in Startup? */ 112c589e69bSNeal Cardwell full_bw_cnt:2, /* number of rounds without large bw gains */ 1130f8782eaSNeal Cardwell cycle_idx:3, /* current index in pacing_gain cycle array */ 11432984565SNeal Cardwell has_seen_rtt:1, /* have we seen an RTT sample yet? */ 11532984565SNeal Cardwell unused_b:5; 1160f8782eaSNeal Cardwell u32 prior_cwnd; /* prior cwnd upon entering loss recovery */ 1170f8782eaSNeal Cardwell u32 full_bw; /* recent bw, to estimate if pipe is full */ 1180f8782eaSNeal Cardwell }; 1190f8782eaSNeal Cardwell 1200f8782eaSNeal Cardwell #define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */ 1210f8782eaSNeal Cardwell 1220f8782eaSNeal Cardwell /* Window length of bw filter (in rounds): */ 1230f8782eaSNeal Cardwell static const int bbr_bw_rtts = CYCLE_LEN + 2; 1240f8782eaSNeal Cardwell /* Window length of min_rtt filter (in sec): */ 1250f8782eaSNeal Cardwell static const u32 bbr_min_rtt_win_sec = 10; 1260f8782eaSNeal Cardwell /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */ 1270f8782eaSNeal Cardwell static const u32 bbr_probe_rtt_mode_ms = 200; 1280f8782eaSNeal Cardwell /* Skip TSO below the following bandwidth (bits/sec): */ 1290f8782eaSNeal Cardwell static const int bbr_min_tso_rate = 1200000; 1300f8782eaSNeal Cardwell 131*1106a5adSNeal Cardwell /* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck. 132*1106a5adSNeal Cardwell * In order to help drive the network toward lower queues and low latency while 133*1106a5adSNeal Cardwell * maintaining high utilization, the average pacing rate aims to be slightly 134*1106a5adSNeal Cardwell * lower than the estimated bandwidth. This is an important aspect of the 135*1106a5adSNeal Cardwell * design. 136*1106a5adSNeal Cardwell */ 13797ec3eb3SNeal Cardwell static const int bbr_pacing_margin_percent = 1; 138ab408b6dSEric Dumazet 1390f8782eaSNeal Cardwell /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain 1400f8782eaSNeal Cardwell * that will allow a smoothly increasing pacing rate that will double each RTT 1410f8782eaSNeal Cardwell * and send the same number of packets per RTT that an un-paced, slow-starting 1420f8782eaSNeal Cardwell * Reno or CUBIC flow would: 1430f8782eaSNeal Cardwell */ 1440f8782eaSNeal Cardwell static const int bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 1450f8782eaSNeal Cardwell /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain 1460f8782eaSNeal Cardwell * the queue created in BBR_STARTUP in a single round: 1470f8782eaSNeal Cardwell */ 1480f8782eaSNeal Cardwell static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885; 1490f8782eaSNeal Cardwell /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */ 1500f8782eaSNeal Cardwell static const int bbr_cwnd_gain = BBR_UNIT * 2; 1510f8782eaSNeal Cardwell /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */ 1520f8782eaSNeal Cardwell static const int bbr_pacing_gain[] = { 1530f8782eaSNeal Cardwell BBR_UNIT * 5 / 4, /* probe for more available bw */ 1540f8782eaSNeal Cardwell BBR_UNIT * 3 / 4, /* drain queue and/or yield bw to other flows */ 1550f8782eaSNeal Cardwell BBR_UNIT, BBR_UNIT, BBR_UNIT, /* cruise at 1.0*bw to utilize pipe, */ 1560f8782eaSNeal Cardwell BBR_UNIT, BBR_UNIT, BBR_UNIT /* without creating excess queue... */ 1570f8782eaSNeal Cardwell }; 1580f8782eaSNeal Cardwell /* Randomize the starting gain cycling phase over N phases: */ 1590f8782eaSNeal Cardwell static const u32 bbr_cycle_rand = 7; 1600f8782eaSNeal Cardwell 1610f8782eaSNeal Cardwell /* Try to keep at least this many packets in flight, if things go smoothly. For 1620f8782eaSNeal Cardwell * smooth functioning, a sliding window protocol ACKing every other packet 1630f8782eaSNeal Cardwell * needs at least 4 packets in flight: 1640f8782eaSNeal Cardwell */ 1650f8782eaSNeal Cardwell static const u32 bbr_cwnd_min_target = 4; 1660f8782eaSNeal Cardwell 1670f8782eaSNeal Cardwell /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */ 1680f8782eaSNeal Cardwell /* If bw has increased significantly (1.25x), there may be more bw available: */ 1690f8782eaSNeal Cardwell static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4; 1700f8782eaSNeal Cardwell /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */ 1710f8782eaSNeal Cardwell static const u32 bbr_full_bw_cnt = 3; 1720f8782eaSNeal Cardwell 1730f8782eaSNeal Cardwell /* "long-term" ("LT") bandwidth estimator parameters... */ 1740f8782eaSNeal Cardwell /* The minimum number of rounds in an LT bw sampling interval: */ 1750f8782eaSNeal Cardwell static const u32 bbr_lt_intvl_min_rtts = 4; 1760f8782eaSNeal Cardwell /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */ 1770f8782eaSNeal Cardwell static const u32 bbr_lt_loss_thresh = 50; 1780f8782eaSNeal Cardwell /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */ 1790f8782eaSNeal Cardwell static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8; 1800f8782eaSNeal Cardwell /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */ 1810f8782eaSNeal Cardwell static const u32 bbr_lt_bw_diff = 4000 / 8; 1820f8782eaSNeal Cardwell /* If we estimate we're policed, use lt_bw for this many round trips: */ 1830f8782eaSNeal Cardwell static const u32 bbr_lt_bw_max_rtts = 48; 1840f8782eaSNeal Cardwell 1855490b32dSKevin Yang static void bbr_check_probe_rtt_done(struct sock *sk); 1865490b32dSKevin Yang 1870f8782eaSNeal Cardwell /* Do we estimate that STARTUP filled the pipe? */ 1880f8782eaSNeal Cardwell static bool bbr_full_bw_reached(const struct sock *sk) 1890f8782eaSNeal Cardwell { 1900f8782eaSNeal Cardwell const struct bbr *bbr = inet_csk_ca(sk); 1910f8782eaSNeal Cardwell 192c589e69bSNeal Cardwell return bbr->full_bw_reached; 1930f8782eaSNeal Cardwell } 1940f8782eaSNeal Cardwell 1950f8782eaSNeal Cardwell /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */ 1960f8782eaSNeal Cardwell static u32 bbr_max_bw(const struct sock *sk) 1970f8782eaSNeal Cardwell { 1980f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 1990f8782eaSNeal Cardwell 2000f8782eaSNeal Cardwell return minmax_get(&bbr->bw); 2010f8782eaSNeal Cardwell } 2020f8782eaSNeal Cardwell 2030f8782eaSNeal Cardwell /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */ 2040f8782eaSNeal Cardwell static u32 bbr_bw(const struct sock *sk) 2050f8782eaSNeal Cardwell { 2060f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 2070f8782eaSNeal Cardwell 2080f8782eaSNeal Cardwell return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk); 2090f8782eaSNeal Cardwell } 2100f8782eaSNeal Cardwell 2110f8782eaSNeal Cardwell /* Return rate in bytes per second, optionally with a gain. 2120f8782eaSNeal Cardwell * The order here is chosen carefully to avoid overflow of u64. This should 2130f8782eaSNeal Cardwell * work for input rates of up to 2.9Tbit/sec and gain of 2.89x. 2140f8782eaSNeal Cardwell */ 2150f8782eaSNeal Cardwell static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain) 2160f8782eaSNeal Cardwell { 217cadefe5fSEric Dumazet unsigned int mss = tcp_sk(sk)->mss_cache; 218cadefe5fSEric Dumazet 219cadefe5fSEric Dumazet rate *= mss; 2200f8782eaSNeal Cardwell rate *= gain; 2210f8782eaSNeal Cardwell rate >>= BBR_SCALE; 22297ec3eb3SNeal Cardwell rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent); 2230f8782eaSNeal Cardwell return rate >> BW_SCALE; 2240f8782eaSNeal Cardwell } 2250f8782eaSNeal Cardwell 226f19fd62dSNeal Cardwell /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */ 22776a9ebe8SEric Dumazet static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain) 228f19fd62dSNeal Cardwell { 229f19fd62dSNeal Cardwell u64 rate = bw; 230f19fd62dSNeal Cardwell 231f19fd62dSNeal Cardwell rate = bbr_rate_bytes_per_sec(sk, rate, gain); 232f19fd62dSNeal Cardwell rate = min_t(u64, rate, sk->sk_max_pacing_rate); 233f19fd62dSNeal Cardwell return rate; 234f19fd62dSNeal Cardwell } 235f19fd62dSNeal Cardwell 23679135b89SNeal Cardwell /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */ 23779135b89SNeal Cardwell static void bbr_init_pacing_rate_from_rtt(struct sock *sk) 23879135b89SNeal Cardwell { 23979135b89SNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 24032984565SNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 24179135b89SNeal Cardwell u64 bw; 24279135b89SNeal Cardwell u32 rtt_us; 24379135b89SNeal Cardwell 24479135b89SNeal Cardwell if (tp->srtt_us) { /* any RTT sample yet? */ 24579135b89SNeal Cardwell rtt_us = max(tp->srtt_us >> 3, 1U); 24632984565SNeal Cardwell bbr->has_seen_rtt = 1; 24779135b89SNeal Cardwell } else { /* no RTT sample yet */ 24879135b89SNeal Cardwell rtt_us = USEC_PER_MSEC; /* use nominal default RTT */ 24979135b89SNeal Cardwell } 25079135b89SNeal Cardwell bw = (u64)tp->snd_cwnd * BW_UNIT; 25179135b89SNeal Cardwell do_div(bw, rtt_us); 25279135b89SNeal Cardwell sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain); 25379135b89SNeal Cardwell } 25479135b89SNeal Cardwell 255*1106a5adSNeal Cardwell /* Pace using current bw estimate and a gain factor. */ 2560f8782eaSNeal Cardwell static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain) 2570f8782eaSNeal Cardwell { 25832984565SNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 25932984565SNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 26076a9ebe8SEric Dumazet unsigned long rate = bbr_bw_to_pacing_rate(sk, bw, gain); 2610f8782eaSNeal Cardwell 26232984565SNeal Cardwell if (unlikely(!bbr->has_seen_rtt && tp->srtt_us)) 26332984565SNeal Cardwell bbr_init_pacing_rate_from_rtt(sk); 2644aea287eSNeal Cardwell if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate) 2650f8782eaSNeal Cardwell sk->sk_pacing_rate = rate; 2660f8782eaSNeal Cardwell } 2670f8782eaSNeal Cardwell 268dcb8c9b4SEric Dumazet /* override sysctl_tcp_min_tso_segs */ 269dcb8c9b4SEric Dumazet static u32 bbr_min_tso_segs(struct sock *sk) 2700f8782eaSNeal Cardwell { 271dcb8c9b4SEric Dumazet return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2; 2720f8782eaSNeal Cardwell } 2730f8782eaSNeal Cardwell 27471abf467SEric Dumazet static u32 bbr_tso_segs_goal(struct sock *sk) 2750f8782eaSNeal Cardwell { 2760f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 277dcb8c9b4SEric Dumazet u32 segs, bytes; 2780f8782eaSNeal Cardwell 279dcb8c9b4SEric Dumazet /* Sort of tcp_tso_autosize() but ignoring 280dcb8c9b4SEric Dumazet * driver provided sk_gso_max_size. 281dcb8c9b4SEric Dumazet */ 28276a9ebe8SEric Dumazet bytes = min_t(unsigned long, sk->sk_pacing_rate >> sk->sk_pacing_shift, 283dcb8c9b4SEric Dumazet GSO_MAX_SIZE - 1 - MAX_TCP_HEADER); 284dcb8c9b4SEric Dumazet segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk)); 285dcb8c9b4SEric Dumazet 28671abf467SEric Dumazet return min(segs, 0x7FU); 2870f8782eaSNeal Cardwell } 2880f8782eaSNeal Cardwell 2890f8782eaSNeal Cardwell /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ 2900f8782eaSNeal Cardwell static void bbr_save_cwnd(struct sock *sk) 2910f8782eaSNeal Cardwell { 2920f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 2930f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 2940f8782eaSNeal Cardwell 2950f8782eaSNeal Cardwell if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT) 2960f8782eaSNeal Cardwell bbr->prior_cwnd = tp->snd_cwnd; /* this cwnd is good enough */ 2970f8782eaSNeal Cardwell else /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */ 2980f8782eaSNeal Cardwell bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd); 2990f8782eaSNeal Cardwell } 3000f8782eaSNeal Cardwell 3010f8782eaSNeal Cardwell static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event) 3020f8782eaSNeal Cardwell { 3030f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 3040f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 3050f8782eaSNeal Cardwell 3060f8782eaSNeal Cardwell if (event == CA_EVENT_TX_START && tp->app_limited) { 3070f8782eaSNeal Cardwell bbr->idle_restart = 1; 3080f8782eaSNeal Cardwell /* Avoid pointless buffer overflows: pace at est. bw if we don't 3090f8782eaSNeal Cardwell * need more speed (we're restarting from idle and app-limited). 3100f8782eaSNeal Cardwell */ 3110f8782eaSNeal Cardwell if (bbr->mode == BBR_PROBE_BW) 3120f8782eaSNeal Cardwell bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT); 3135490b32dSKevin Yang else if (bbr->mode == BBR_PROBE_RTT) 3145490b32dSKevin Yang bbr_check_probe_rtt_done(sk); 3150f8782eaSNeal Cardwell } 3160f8782eaSNeal Cardwell } 3170f8782eaSNeal Cardwell 3180f8782eaSNeal Cardwell /* Find target cwnd. Right-size the cwnd based on min RTT and the 3190f8782eaSNeal Cardwell * estimated bottleneck bandwidth: 3200f8782eaSNeal Cardwell * 3210f8782eaSNeal Cardwell * cwnd = bw * min_rtt * gain = BDP * gain 3220f8782eaSNeal Cardwell * 3230f8782eaSNeal Cardwell * The key factor, gain, controls the amount of queue. While a small gain 3240f8782eaSNeal Cardwell * builds a smaller queue, it becomes more vulnerable to noise in RTT 3250f8782eaSNeal Cardwell * measurements (e.g., delayed ACKs or other ACK compression effects). This 3260f8782eaSNeal Cardwell * noise may cause BBR to under-estimate the rate. 3270f8782eaSNeal Cardwell * 3280f8782eaSNeal Cardwell * To achieve full performance in high-speed paths, we budget enough cwnd to 3290f8782eaSNeal Cardwell * fit full-sized skbs in-flight on both end hosts to fully utilize the path: 3300f8782eaSNeal Cardwell * - one skb in sending host Qdisc, 3310f8782eaSNeal Cardwell * - one skb in sending host TSO/GSO engine 3320f8782eaSNeal Cardwell * - one skb being received by receiver host LRO/GRO/delayed-ACK engine 3330f8782eaSNeal Cardwell * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because 3340f8782eaSNeal Cardwell * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets, 3350f8782eaSNeal Cardwell * which allows 2 outstanding 2-packet sequences, to try to keep pipe 3360f8782eaSNeal Cardwell * full even with ACK-every-other-packet delayed ACKs. 3370f8782eaSNeal Cardwell */ 3380f8782eaSNeal Cardwell static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain) 3390f8782eaSNeal Cardwell { 3400f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 3410f8782eaSNeal Cardwell u32 cwnd; 3420f8782eaSNeal Cardwell u64 w; 3430f8782eaSNeal Cardwell 3440f8782eaSNeal Cardwell /* If we've never had a valid RTT sample, cap cwnd at the initial 3450f8782eaSNeal Cardwell * default. This should only happen when the connection is not using TCP 3460f8782eaSNeal Cardwell * timestamps and has retransmitted all of the SYN/SYNACK/data packets 3470f8782eaSNeal Cardwell * ACKed so far. In this case, an RTO can cut cwnd to 1, in which 3480f8782eaSNeal Cardwell * case we need to slow-start up toward something safe: TCP_INIT_CWND. 3490f8782eaSNeal Cardwell */ 3500f8782eaSNeal Cardwell if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */ 3510f8782eaSNeal Cardwell return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/ 3520f8782eaSNeal Cardwell 3530f8782eaSNeal Cardwell w = (u64)bw * bbr->min_rtt_us; 3540f8782eaSNeal Cardwell 3550f8782eaSNeal Cardwell /* Apply a gain to the given value, then remove the BW_SCALE shift. */ 3560f8782eaSNeal Cardwell cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT; 3570f8782eaSNeal Cardwell 3580f8782eaSNeal Cardwell /* Allow enough full-sized skbs in flight to utilize end systems. */ 35971abf467SEric Dumazet cwnd += 3 * bbr_tso_segs_goal(sk); 3600f8782eaSNeal Cardwell 3610f8782eaSNeal Cardwell /* Reduce delayed ACKs by rounding up cwnd to the next even number. */ 3620f8782eaSNeal Cardwell cwnd = (cwnd + 1) & ~1U; 3630f8782eaSNeal Cardwell 364383d4709SNeal Cardwell /* Ensure gain cycling gets inflight above BDP even for small BDPs. */ 365383d4709SNeal Cardwell if (bbr->mode == BBR_PROBE_BW && gain > BBR_UNIT) 366383d4709SNeal Cardwell cwnd += 2; 367383d4709SNeal Cardwell 3680f8782eaSNeal Cardwell return cwnd; 3690f8782eaSNeal Cardwell } 3700f8782eaSNeal Cardwell 371a87c83d5SNeal Cardwell /* With pacing at lower layers, there's often less data "in the network" than 372a87c83d5SNeal Cardwell * "in flight". With TSQ and departure time pacing at lower layers (e.g. fq), 373a87c83d5SNeal Cardwell * we often have several skbs queued in the pacing layer with a pre-scheduled 374a87c83d5SNeal Cardwell * earliest departure time (EDT). BBR adapts its pacing rate based on the 375a87c83d5SNeal Cardwell * inflight level that it estimates has already been "baked in" by previous 376a87c83d5SNeal Cardwell * departure time decisions. We calculate a rough estimate of the number of our 377a87c83d5SNeal Cardwell * packets that might be in the network at the earliest departure time for the 378a87c83d5SNeal Cardwell * next skb scheduled: 379a87c83d5SNeal Cardwell * in_network_at_edt = inflight_at_edt - (EDT - now) * bw 380a87c83d5SNeal Cardwell * If we're increasing inflight, then we want to know if the transmit of the 381a87c83d5SNeal Cardwell * EDT skb will push inflight above the target, so inflight_at_edt includes 382a87c83d5SNeal Cardwell * bbr_tso_segs_goal() from the skb departing at EDT. If decreasing inflight, 383a87c83d5SNeal Cardwell * then estimate if inflight will sink too low just before the EDT transmit. 384a87c83d5SNeal Cardwell */ 385a87c83d5SNeal Cardwell static u32 bbr_packets_in_net_at_edt(struct sock *sk, u32 inflight_now) 386a87c83d5SNeal Cardwell { 387a87c83d5SNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 388a87c83d5SNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 389a87c83d5SNeal Cardwell u64 now_ns, edt_ns, interval_us; 390a87c83d5SNeal Cardwell u32 interval_delivered, inflight_at_edt; 391a87c83d5SNeal Cardwell 392a87c83d5SNeal Cardwell now_ns = tp->tcp_clock_cache; 393a87c83d5SNeal Cardwell edt_ns = max(tp->tcp_wstamp_ns, now_ns); 394a87c83d5SNeal Cardwell interval_us = div_u64(edt_ns - now_ns, NSEC_PER_USEC); 395a87c83d5SNeal Cardwell interval_delivered = (u64)bbr_bw(sk) * interval_us >> BW_SCALE; 396a87c83d5SNeal Cardwell inflight_at_edt = inflight_now; 397a87c83d5SNeal Cardwell if (bbr->pacing_gain > BBR_UNIT) /* increasing inflight */ 398a87c83d5SNeal Cardwell inflight_at_edt += bbr_tso_segs_goal(sk); /* include EDT skb */ 399a87c83d5SNeal Cardwell if (interval_delivered >= inflight_at_edt) 400a87c83d5SNeal Cardwell return 0; 401a87c83d5SNeal Cardwell return inflight_at_edt - interval_delivered; 402a87c83d5SNeal Cardwell } 403a87c83d5SNeal Cardwell 4040f8782eaSNeal Cardwell /* An optimization in BBR to reduce losses: On the first round of recovery, we 4050f8782eaSNeal Cardwell * follow the packet conservation principle: send P packets per P packets acked. 4060f8782eaSNeal Cardwell * After that, we slow-start and send at most 2*P packets per P packets acked. 4070f8782eaSNeal Cardwell * After recovery finishes, or upon undo, we restore the cwnd we had when 4080f8782eaSNeal Cardwell * recovery started (capped by the target cwnd based on estimated BDP). 4090f8782eaSNeal Cardwell * 4100f8782eaSNeal Cardwell * TODO(ycheng/ncardwell): implement a rate-based approach. 4110f8782eaSNeal Cardwell */ 4120f8782eaSNeal Cardwell static bool bbr_set_cwnd_to_recover_or_restore( 4130f8782eaSNeal Cardwell struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd) 4140f8782eaSNeal Cardwell { 4150f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 4160f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 4170f8782eaSNeal Cardwell u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state; 4180f8782eaSNeal Cardwell u32 cwnd = tp->snd_cwnd; 4190f8782eaSNeal Cardwell 4200f8782eaSNeal Cardwell /* An ACK for P pkts should release at most 2*P packets. We do this 4210f8782eaSNeal Cardwell * in two steps. First, here we deduct the number of lost packets. 4220f8782eaSNeal Cardwell * Then, in bbr_set_cwnd() we slow start up toward the target cwnd. 4230f8782eaSNeal Cardwell */ 4240f8782eaSNeal Cardwell if (rs->losses > 0) 4250f8782eaSNeal Cardwell cwnd = max_t(s32, cwnd - rs->losses, 1); 4260f8782eaSNeal Cardwell 4270f8782eaSNeal Cardwell if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) { 4280f8782eaSNeal Cardwell /* Starting 1st round of Recovery, so do packet conservation. */ 4290f8782eaSNeal Cardwell bbr->packet_conservation = 1; 4300f8782eaSNeal Cardwell bbr->next_rtt_delivered = tp->delivered; /* start round now */ 4310f8782eaSNeal Cardwell /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */ 4320f8782eaSNeal Cardwell cwnd = tcp_packets_in_flight(tp) + acked; 4330f8782eaSNeal Cardwell } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) { 4340f8782eaSNeal Cardwell /* Exiting loss recovery; restore cwnd saved before recovery. */ 435fb998862SKevin Yang cwnd = max(cwnd, bbr->prior_cwnd); 4360f8782eaSNeal Cardwell bbr->packet_conservation = 0; 4370f8782eaSNeal Cardwell } 4380f8782eaSNeal Cardwell bbr->prev_ca_state = state; 4390f8782eaSNeal Cardwell 4400f8782eaSNeal Cardwell if (bbr->packet_conservation) { 4410f8782eaSNeal Cardwell *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked); 4420f8782eaSNeal Cardwell return true; /* yes, using packet conservation */ 4430f8782eaSNeal Cardwell } 4440f8782eaSNeal Cardwell *new_cwnd = cwnd; 4450f8782eaSNeal Cardwell return false; 4460f8782eaSNeal Cardwell } 4470f8782eaSNeal Cardwell 4480f8782eaSNeal Cardwell /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss 4490f8782eaSNeal Cardwell * has drawn us down below target), or snap down to target if we're above it. 4500f8782eaSNeal Cardwell */ 4510f8782eaSNeal Cardwell static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs, 4520f8782eaSNeal Cardwell u32 acked, u32 bw, int gain) 4530f8782eaSNeal Cardwell { 4540f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 4550f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 4568e995bf1SKevin Yang u32 cwnd = tp->snd_cwnd, target_cwnd = 0; 4570f8782eaSNeal Cardwell 4580f8782eaSNeal Cardwell if (!acked) 4598e995bf1SKevin Yang goto done; /* no packet fully ACKed; just apply caps */ 4600f8782eaSNeal Cardwell 4610f8782eaSNeal Cardwell if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd)) 4620f8782eaSNeal Cardwell goto done; 4630f8782eaSNeal Cardwell 4640f8782eaSNeal Cardwell /* If we're below target cwnd, slow start cwnd toward target cwnd. */ 4650f8782eaSNeal Cardwell target_cwnd = bbr_target_cwnd(sk, bw, gain); 4660f8782eaSNeal Cardwell if (bbr_full_bw_reached(sk)) /* only cut cwnd if we filled the pipe */ 4670f8782eaSNeal Cardwell cwnd = min(cwnd + acked, target_cwnd); 4680f8782eaSNeal Cardwell else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND) 4690f8782eaSNeal Cardwell cwnd = cwnd + acked; 4700f8782eaSNeal Cardwell cwnd = max(cwnd, bbr_cwnd_min_target); 4710f8782eaSNeal Cardwell 4720f8782eaSNeal Cardwell done: 4730f8782eaSNeal Cardwell tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); /* apply global cap */ 4740f8782eaSNeal Cardwell if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */ 4750f8782eaSNeal Cardwell tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target); 4760f8782eaSNeal Cardwell } 4770f8782eaSNeal Cardwell 4780f8782eaSNeal Cardwell /* End cycle phase if it's time and/or we hit the phase's in-flight target. */ 4790f8782eaSNeal Cardwell static bool bbr_is_next_cycle_phase(struct sock *sk, 4800f8782eaSNeal Cardwell const struct rate_sample *rs) 4810f8782eaSNeal Cardwell { 4820f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 4830f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 4840f8782eaSNeal Cardwell bool is_full_length = 4859a568de4SEric Dumazet tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) > 4860f8782eaSNeal Cardwell bbr->min_rtt_us; 4870f8782eaSNeal Cardwell u32 inflight, bw; 4880f8782eaSNeal Cardwell 4890f8782eaSNeal Cardwell /* The pacing_gain of 1.0 paces at the estimated bw to try to fully 4900f8782eaSNeal Cardwell * use the pipe without increasing the queue. 4910f8782eaSNeal Cardwell */ 4920f8782eaSNeal Cardwell if (bbr->pacing_gain == BBR_UNIT) 4930f8782eaSNeal Cardwell return is_full_length; /* just use wall clock time */ 4940f8782eaSNeal Cardwell 495a87c83d5SNeal Cardwell inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight); 4960f8782eaSNeal Cardwell bw = bbr_max_bw(sk); 4970f8782eaSNeal Cardwell 4980f8782eaSNeal Cardwell /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at 4990f8782eaSNeal Cardwell * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is 5000f8782eaSNeal Cardwell * small (e.g. on a LAN). We do not persist if packets are lost, since 5010f8782eaSNeal Cardwell * a path with small buffers may not hold that much. 5020f8782eaSNeal Cardwell */ 5030f8782eaSNeal Cardwell if (bbr->pacing_gain > BBR_UNIT) 5040f8782eaSNeal Cardwell return is_full_length && 5050f8782eaSNeal Cardwell (rs->losses || /* perhaps pacing_gain*BDP won't fit */ 5060f8782eaSNeal Cardwell inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain)); 5070f8782eaSNeal Cardwell 5080f8782eaSNeal Cardwell /* A pacing_gain < 1.0 tries to drain extra queue we added if bw 5090f8782eaSNeal Cardwell * probing didn't find more bw. If inflight falls to match BDP then we 5100f8782eaSNeal Cardwell * estimate queue is drained; persisting would underutilize the pipe. 5110f8782eaSNeal Cardwell */ 5120f8782eaSNeal Cardwell return is_full_length || 5130f8782eaSNeal Cardwell inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT); 5140f8782eaSNeal Cardwell } 5150f8782eaSNeal Cardwell 5160f8782eaSNeal Cardwell static void bbr_advance_cycle_phase(struct sock *sk) 5170f8782eaSNeal Cardwell { 5180f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 5190f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5200f8782eaSNeal Cardwell 5210f8782eaSNeal Cardwell bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1); 5220f8782eaSNeal Cardwell bbr->cycle_mstamp = tp->delivered_mstamp; 5230f8782eaSNeal Cardwell } 5240f8782eaSNeal Cardwell 5250f8782eaSNeal Cardwell /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */ 5260f8782eaSNeal Cardwell static void bbr_update_cycle_phase(struct sock *sk, 5270f8782eaSNeal Cardwell const struct rate_sample *rs) 5280f8782eaSNeal Cardwell { 5290f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5300f8782eaSNeal Cardwell 5313aff3b4bSNeal Cardwell if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs)) 5320f8782eaSNeal Cardwell bbr_advance_cycle_phase(sk); 5330f8782eaSNeal Cardwell } 5340f8782eaSNeal Cardwell 5350f8782eaSNeal Cardwell static void bbr_reset_startup_mode(struct sock *sk) 5360f8782eaSNeal Cardwell { 5370f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5380f8782eaSNeal Cardwell 5390f8782eaSNeal Cardwell bbr->mode = BBR_STARTUP; 5400f8782eaSNeal Cardwell } 5410f8782eaSNeal Cardwell 5420f8782eaSNeal Cardwell static void bbr_reset_probe_bw_mode(struct sock *sk) 5430f8782eaSNeal Cardwell { 5440f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5450f8782eaSNeal Cardwell 5460f8782eaSNeal Cardwell bbr->mode = BBR_PROBE_BW; 5470f8782eaSNeal Cardwell bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand); 5480f8782eaSNeal Cardwell bbr_advance_cycle_phase(sk); /* flip to next phase of gain cycle */ 5490f8782eaSNeal Cardwell } 5500f8782eaSNeal Cardwell 5510f8782eaSNeal Cardwell static void bbr_reset_mode(struct sock *sk) 5520f8782eaSNeal Cardwell { 5530f8782eaSNeal Cardwell if (!bbr_full_bw_reached(sk)) 5540f8782eaSNeal Cardwell bbr_reset_startup_mode(sk); 5550f8782eaSNeal Cardwell else 5560f8782eaSNeal Cardwell bbr_reset_probe_bw_mode(sk); 5570f8782eaSNeal Cardwell } 5580f8782eaSNeal Cardwell 5590f8782eaSNeal Cardwell /* Start a new long-term sampling interval. */ 5600f8782eaSNeal Cardwell static void bbr_reset_lt_bw_sampling_interval(struct sock *sk) 5610f8782eaSNeal Cardwell { 5620f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 5630f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5640f8782eaSNeal Cardwell 5659a568de4SEric Dumazet bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC); 5660f8782eaSNeal Cardwell bbr->lt_last_delivered = tp->delivered; 5670f8782eaSNeal Cardwell bbr->lt_last_lost = tp->lost; 5680f8782eaSNeal Cardwell bbr->lt_rtt_cnt = 0; 5690f8782eaSNeal Cardwell } 5700f8782eaSNeal Cardwell 5710f8782eaSNeal Cardwell /* Completely reset long-term bandwidth sampling. */ 5720f8782eaSNeal Cardwell static void bbr_reset_lt_bw_sampling(struct sock *sk) 5730f8782eaSNeal Cardwell { 5740f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5750f8782eaSNeal Cardwell 5760f8782eaSNeal Cardwell bbr->lt_bw = 0; 5770f8782eaSNeal Cardwell bbr->lt_use_bw = 0; 5780f8782eaSNeal Cardwell bbr->lt_is_sampling = false; 5790f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling_interval(sk); 5800f8782eaSNeal Cardwell } 5810f8782eaSNeal Cardwell 5820f8782eaSNeal Cardwell /* Long-term bw sampling interval is done. Estimate whether we're policed. */ 5830f8782eaSNeal Cardwell static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw) 5840f8782eaSNeal Cardwell { 5850f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 5860f8782eaSNeal Cardwell u32 diff; 5870f8782eaSNeal Cardwell 5880f8782eaSNeal Cardwell if (bbr->lt_bw) { /* do we have bw from a previous interval? */ 5890f8782eaSNeal Cardwell /* Is new bw close to the lt_bw from the previous interval? */ 5900f8782eaSNeal Cardwell diff = abs(bw - bbr->lt_bw); 5910f8782eaSNeal Cardwell if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) || 5920f8782eaSNeal Cardwell (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <= 5930f8782eaSNeal Cardwell bbr_lt_bw_diff)) { 5940f8782eaSNeal Cardwell /* All criteria are met; estimate we're policed. */ 5950f8782eaSNeal Cardwell bbr->lt_bw = (bw + bbr->lt_bw) >> 1; /* avg 2 intvls */ 5960f8782eaSNeal Cardwell bbr->lt_use_bw = 1; 5970f8782eaSNeal Cardwell bbr->pacing_gain = BBR_UNIT; /* try to avoid drops */ 5980f8782eaSNeal Cardwell bbr->lt_rtt_cnt = 0; 5990f8782eaSNeal Cardwell return; 6000f8782eaSNeal Cardwell } 6010f8782eaSNeal Cardwell } 6020f8782eaSNeal Cardwell bbr->lt_bw = bw; 6030f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling_interval(sk); 6040f8782eaSNeal Cardwell } 6050f8782eaSNeal Cardwell 6060f8782eaSNeal Cardwell /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of 6070f8782eaSNeal Cardwell * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and 6080f8782eaSNeal Cardwell * explicitly models their policed rate, to reduce unnecessary losses. We 6090f8782eaSNeal Cardwell * estimate that we're policed if we see 2 consecutive sampling intervals with 6100f8782eaSNeal Cardwell * consistent throughput and high packet loss. If we think we're being policed, 6110f8782eaSNeal Cardwell * set lt_bw to the "long-term" average delivery rate from those 2 intervals. 6120f8782eaSNeal Cardwell */ 6130f8782eaSNeal Cardwell static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs) 6140f8782eaSNeal Cardwell { 6150f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 6160f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 6170f8782eaSNeal Cardwell u32 lost, delivered; 6180f8782eaSNeal Cardwell u64 bw; 6199a568de4SEric Dumazet u32 t; 6200f8782eaSNeal Cardwell 6210f8782eaSNeal Cardwell if (bbr->lt_use_bw) { /* already using long-term rate, lt_bw? */ 6220f8782eaSNeal Cardwell if (bbr->mode == BBR_PROBE_BW && bbr->round_start && 6230f8782eaSNeal Cardwell ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) { 6240f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling(sk); /* stop using lt_bw */ 6250f8782eaSNeal Cardwell bbr_reset_probe_bw_mode(sk); /* restart gain cycling */ 6260f8782eaSNeal Cardwell } 6270f8782eaSNeal Cardwell return; 6280f8782eaSNeal Cardwell } 6290f8782eaSNeal Cardwell 6300f8782eaSNeal Cardwell /* Wait for the first loss before sampling, to let the policer exhaust 6310f8782eaSNeal Cardwell * its tokens and estimate the steady-state rate allowed by the policer. 6320f8782eaSNeal Cardwell * Starting samples earlier includes bursts that over-estimate the bw. 6330f8782eaSNeal Cardwell */ 6340f8782eaSNeal Cardwell if (!bbr->lt_is_sampling) { 6350f8782eaSNeal Cardwell if (!rs->losses) 6360f8782eaSNeal Cardwell return; 6370f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling_interval(sk); 6380f8782eaSNeal Cardwell bbr->lt_is_sampling = true; 6390f8782eaSNeal Cardwell } 6400f8782eaSNeal Cardwell 6410f8782eaSNeal Cardwell /* To avoid underestimates, reset sampling if we run out of data. */ 6420f8782eaSNeal Cardwell if (rs->is_app_limited) { 6430f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling(sk); 6440f8782eaSNeal Cardwell return; 6450f8782eaSNeal Cardwell } 6460f8782eaSNeal Cardwell 6470f8782eaSNeal Cardwell if (bbr->round_start) 6480f8782eaSNeal Cardwell bbr->lt_rtt_cnt++; /* count round trips in this interval */ 6490f8782eaSNeal Cardwell if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts) 6500f8782eaSNeal Cardwell return; /* sampling interval needs to be longer */ 6510f8782eaSNeal Cardwell if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) { 6520f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling(sk); /* interval is too long */ 6530f8782eaSNeal Cardwell return; 6540f8782eaSNeal Cardwell } 6550f8782eaSNeal Cardwell 6560f8782eaSNeal Cardwell /* End sampling interval when a packet is lost, so we estimate the 6570f8782eaSNeal Cardwell * policer tokens were exhausted. Stopping the sampling before the 6580f8782eaSNeal Cardwell * tokens are exhausted under-estimates the policed rate. 6590f8782eaSNeal Cardwell */ 6600f8782eaSNeal Cardwell if (!rs->losses) 6610f8782eaSNeal Cardwell return; 6620f8782eaSNeal Cardwell 6630f8782eaSNeal Cardwell /* Calculate packets lost and delivered in sampling interval. */ 6640f8782eaSNeal Cardwell lost = tp->lost - bbr->lt_last_lost; 6650f8782eaSNeal Cardwell delivered = tp->delivered - bbr->lt_last_delivered; 6660f8782eaSNeal Cardwell /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */ 6670f8782eaSNeal Cardwell if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered) 6680f8782eaSNeal Cardwell return; 6690f8782eaSNeal Cardwell 6700f8782eaSNeal Cardwell /* Find average delivery rate in this sampling interval. */ 6719a568de4SEric Dumazet t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp; 6729a568de4SEric Dumazet if ((s32)t < 1) 6739a568de4SEric Dumazet return; /* interval is less than one ms, so wait */ 6749a568de4SEric Dumazet /* Check if can multiply without overflow */ 6759a568de4SEric Dumazet if (t >= ~0U / USEC_PER_MSEC) { 6760f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */ 6770f8782eaSNeal Cardwell return; 6780f8782eaSNeal Cardwell } 6799a568de4SEric Dumazet t *= USEC_PER_MSEC; 6800f8782eaSNeal Cardwell bw = (u64)delivered * BW_UNIT; 6810f8782eaSNeal Cardwell do_div(bw, t); 6820f8782eaSNeal Cardwell bbr_lt_bw_interval_done(sk, bw); 6830f8782eaSNeal Cardwell } 6840f8782eaSNeal Cardwell 6850f8782eaSNeal Cardwell /* Estimate the bandwidth based on how fast packets are delivered */ 6860f8782eaSNeal Cardwell static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs) 6870f8782eaSNeal Cardwell { 6880f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 6890f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 6900f8782eaSNeal Cardwell u64 bw; 6910f8782eaSNeal Cardwell 6920f8782eaSNeal Cardwell bbr->round_start = 0; 6930f8782eaSNeal Cardwell if (rs->delivered < 0 || rs->interval_us <= 0) 6940f8782eaSNeal Cardwell return; /* Not a valid observation */ 6950f8782eaSNeal Cardwell 6960f8782eaSNeal Cardwell /* See if we've reached the next RTT */ 6970f8782eaSNeal Cardwell if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) { 6980f8782eaSNeal Cardwell bbr->next_rtt_delivered = tp->delivered; 6990f8782eaSNeal Cardwell bbr->rtt_cnt++; 7000f8782eaSNeal Cardwell bbr->round_start = 1; 7010f8782eaSNeal Cardwell bbr->packet_conservation = 0; 7020f8782eaSNeal Cardwell } 7030f8782eaSNeal Cardwell 7040f8782eaSNeal Cardwell bbr_lt_bw_sampling(sk, rs); 7050f8782eaSNeal Cardwell 7060f8782eaSNeal Cardwell /* Divide delivered by the interval to find a (lower bound) bottleneck 7070f8782eaSNeal Cardwell * bandwidth sample. Delivered is in packets and interval_us in uS and 7080f8782eaSNeal Cardwell * ratio will be <<1 for most connections. So delivered is first scaled. 7090f8782eaSNeal Cardwell */ 7100f8782eaSNeal Cardwell bw = (u64)rs->delivered * BW_UNIT; 7110f8782eaSNeal Cardwell do_div(bw, rs->interval_us); 7120f8782eaSNeal Cardwell 7130f8782eaSNeal Cardwell /* If this sample is application-limited, it is likely to have a very 7140f8782eaSNeal Cardwell * low delivered count that represents application behavior rather than 7150f8782eaSNeal Cardwell * the available network rate. Such a sample could drag down estimated 7160f8782eaSNeal Cardwell * bw, causing needless slow-down. Thus, to continue to send at the 7170f8782eaSNeal Cardwell * last measured network rate, we filter out app-limited samples unless 7180f8782eaSNeal Cardwell * they describe the path bw at least as well as our bw model. 7190f8782eaSNeal Cardwell * 7200f8782eaSNeal Cardwell * So the goal during app-limited phase is to proceed with the best 7210f8782eaSNeal Cardwell * network rate no matter how long. We automatically leave this 7220f8782eaSNeal Cardwell * phase when app writes faster than the network can deliver :) 7230f8782eaSNeal Cardwell */ 7240f8782eaSNeal Cardwell if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) { 7250f8782eaSNeal Cardwell /* Incorporate new sample into our max bw filter. */ 7260f8782eaSNeal Cardwell minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw); 7270f8782eaSNeal Cardwell } 7280f8782eaSNeal Cardwell } 7290f8782eaSNeal Cardwell 7300f8782eaSNeal Cardwell /* Estimate when the pipe is full, using the change in delivery rate: BBR 7310f8782eaSNeal Cardwell * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by 7320f8782eaSNeal Cardwell * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited 7330f8782eaSNeal Cardwell * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the 7340f8782eaSNeal Cardwell * higher rwin, 3: we get higher delivery rate samples. Or transient 7350f8782eaSNeal Cardwell * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar 7360f8782eaSNeal Cardwell * design goal, but uses delay and inter-ACK spacing instead of bandwidth. 7370f8782eaSNeal Cardwell */ 7380f8782eaSNeal Cardwell static void bbr_check_full_bw_reached(struct sock *sk, 7390f8782eaSNeal Cardwell const struct rate_sample *rs) 7400f8782eaSNeal Cardwell { 7410f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 7420f8782eaSNeal Cardwell u32 bw_thresh; 7430f8782eaSNeal Cardwell 7440f8782eaSNeal Cardwell if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited) 7450f8782eaSNeal Cardwell return; 7460f8782eaSNeal Cardwell 7470f8782eaSNeal Cardwell bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE; 7480f8782eaSNeal Cardwell if (bbr_max_bw(sk) >= bw_thresh) { 7490f8782eaSNeal Cardwell bbr->full_bw = bbr_max_bw(sk); 7500f8782eaSNeal Cardwell bbr->full_bw_cnt = 0; 7510f8782eaSNeal Cardwell return; 7520f8782eaSNeal Cardwell } 7530f8782eaSNeal Cardwell ++bbr->full_bw_cnt; 754c589e69bSNeal Cardwell bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt; 7550f8782eaSNeal Cardwell } 7560f8782eaSNeal Cardwell 7570f8782eaSNeal Cardwell /* If pipe is probably full, drain the queue and then enter steady-state. */ 7580f8782eaSNeal Cardwell static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs) 7590f8782eaSNeal Cardwell { 7600f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 7610f8782eaSNeal Cardwell 7620f8782eaSNeal Cardwell if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) { 7630f8782eaSNeal Cardwell bbr->mode = BBR_DRAIN; /* drain queue we created */ 76453794570SYousuk Seung tcp_sk(sk)->snd_ssthresh = 76553794570SYousuk Seung bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT); 7660f8782eaSNeal Cardwell } /* fall through to check if in-flight is already small: */ 7670f8782eaSNeal Cardwell if (bbr->mode == BBR_DRAIN && 768a87c83d5SNeal Cardwell bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <= 7690f8782eaSNeal Cardwell bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT)) 7700f8782eaSNeal Cardwell bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */ 7710f8782eaSNeal Cardwell } 7720f8782eaSNeal Cardwell 773fb998862SKevin Yang static void bbr_check_probe_rtt_done(struct sock *sk) 774fb998862SKevin Yang { 775fb998862SKevin Yang struct tcp_sock *tp = tcp_sk(sk); 776fb998862SKevin Yang struct bbr *bbr = inet_csk_ca(sk); 777fb998862SKevin Yang 778fb998862SKevin Yang if (!(bbr->probe_rtt_done_stamp && 779fb998862SKevin Yang after(tcp_jiffies32, bbr->probe_rtt_done_stamp))) 780fb998862SKevin Yang return; 781fb998862SKevin Yang 782fb998862SKevin Yang bbr->min_rtt_stamp = tcp_jiffies32; /* wait a while until PROBE_RTT */ 783fb998862SKevin Yang tp->snd_cwnd = max(tp->snd_cwnd, bbr->prior_cwnd); 784fb998862SKevin Yang bbr_reset_mode(sk); 785fb998862SKevin Yang } 786fb998862SKevin Yang 7870f8782eaSNeal Cardwell /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and 7880f8782eaSNeal Cardwell * periodically drain the bottleneck queue, to converge to measure the true 7890f8782eaSNeal Cardwell * min_rtt (unloaded propagation delay). This allows the flows to keep queues 7900f8782eaSNeal Cardwell * small (reducing queuing delay and packet loss) and achieve fairness among 7910f8782eaSNeal Cardwell * BBR flows. 7920f8782eaSNeal Cardwell * 7930f8782eaSNeal Cardwell * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires, 7940f8782eaSNeal Cardwell * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets. 7950f8782eaSNeal Cardwell * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed 7960f8782eaSNeal Cardwell * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and 7970f8782eaSNeal Cardwell * re-enter the previous mode. BBR uses 200ms to approximately bound the 7980f8782eaSNeal Cardwell * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s). 7990f8782eaSNeal Cardwell * 8000f8782eaSNeal Cardwell * Note that flows need only pay 2% if they are busy sending over the last 10 8010f8782eaSNeal Cardwell * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have 8020f8782eaSNeal Cardwell * natural silences or low-rate periods within 10 seconds where the rate is low 8030f8782eaSNeal Cardwell * enough for long enough to drain its queue in the bottleneck. We pick up 8040f8782eaSNeal Cardwell * these min RTT measurements opportunistically with our min_rtt filter. :-) 8050f8782eaSNeal Cardwell */ 8060f8782eaSNeal Cardwell static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs) 8070f8782eaSNeal Cardwell { 8080f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 8090f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 8100f8782eaSNeal Cardwell bool filter_expired; 8110f8782eaSNeal Cardwell 8120f8782eaSNeal Cardwell /* Track min RTT seen in the min_rtt_win_sec filter window: */ 8132660bfa8SEric Dumazet filter_expired = after(tcp_jiffies32, 8140f8782eaSNeal Cardwell bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ); 8150f8782eaSNeal Cardwell if (rs->rtt_us >= 0 && 816e4286603SYuchung Cheng (rs->rtt_us <= bbr->min_rtt_us || 817e4286603SYuchung Cheng (filter_expired && !rs->is_ack_delayed))) { 8180f8782eaSNeal Cardwell bbr->min_rtt_us = rs->rtt_us; 8192660bfa8SEric Dumazet bbr->min_rtt_stamp = tcp_jiffies32; 8200f8782eaSNeal Cardwell } 8210f8782eaSNeal Cardwell 8220f8782eaSNeal Cardwell if (bbr_probe_rtt_mode_ms > 0 && filter_expired && 8230f8782eaSNeal Cardwell !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) { 8240f8782eaSNeal Cardwell bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */ 8250f8782eaSNeal Cardwell bbr_save_cwnd(sk); /* note cwnd so we can restore it */ 8260f8782eaSNeal Cardwell bbr->probe_rtt_done_stamp = 0; 8270f8782eaSNeal Cardwell } 8280f8782eaSNeal Cardwell 8290f8782eaSNeal Cardwell if (bbr->mode == BBR_PROBE_RTT) { 8300f8782eaSNeal Cardwell /* Ignore low rate samples during this mode. */ 8310f8782eaSNeal Cardwell tp->app_limited = 8320f8782eaSNeal Cardwell (tp->delivered + tcp_packets_in_flight(tp)) ? : 1; 8330f8782eaSNeal Cardwell /* Maintain min packets in flight for max(200 ms, 1 round). */ 8340f8782eaSNeal Cardwell if (!bbr->probe_rtt_done_stamp && 8350f8782eaSNeal Cardwell tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) { 8362660bfa8SEric Dumazet bbr->probe_rtt_done_stamp = tcp_jiffies32 + 8370f8782eaSNeal Cardwell msecs_to_jiffies(bbr_probe_rtt_mode_ms); 8380f8782eaSNeal Cardwell bbr->probe_rtt_round_done = 0; 8390f8782eaSNeal Cardwell bbr->next_rtt_delivered = tp->delivered; 8400f8782eaSNeal Cardwell } else if (bbr->probe_rtt_done_stamp) { 8410f8782eaSNeal Cardwell if (bbr->round_start) 8420f8782eaSNeal Cardwell bbr->probe_rtt_round_done = 1; 843fb998862SKevin Yang if (bbr->probe_rtt_round_done) 844fb998862SKevin Yang bbr_check_probe_rtt_done(sk); 8450f8782eaSNeal Cardwell } 8460f8782eaSNeal Cardwell } 847e6e6a278SNeal Cardwell /* Restart after idle ends only once we process a new S/ACK for data */ 848e6e6a278SNeal Cardwell if (rs->delivered > 0) 8490f8782eaSNeal Cardwell bbr->idle_restart = 0; 8500f8782eaSNeal Cardwell } 8510f8782eaSNeal Cardwell 852cf33e25cSNeal Cardwell static void bbr_update_gains(struct sock *sk) 853cf33e25cSNeal Cardwell { 854cf33e25cSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 855cf33e25cSNeal Cardwell 856cf33e25cSNeal Cardwell switch (bbr->mode) { 857cf33e25cSNeal Cardwell case BBR_STARTUP: 858cf33e25cSNeal Cardwell bbr->pacing_gain = bbr_high_gain; 859cf33e25cSNeal Cardwell bbr->cwnd_gain = bbr_high_gain; 860cf33e25cSNeal Cardwell break; 861cf33e25cSNeal Cardwell case BBR_DRAIN: 862cf33e25cSNeal Cardwell bbr->pacing_gain = bbr_drain_gain; /* slow, to drain */ 863cf33e25cSNeal Cardwell bbr->cwnd_gain = bbr_high_gain; /* keep cwnd */ 864cf33e25cSNeal Cardwell break; 865cf33e25cSNeal Cardwell case BBR_PROBE_BW: 866cf33e25cSNeal Cardwell bbr->pacing_gain = (bbr->lt_use_bw ? 867cf33e25cSNeal Cardwell BBR_UNIT : 868cf33e25cSNeal Cardwell bbr_pacing_gain[bbr->cycle_idx]); 869cf33e25cSNeal Cardwell bbr->cwnd_gain = bbr_cwnd_gain; 870cf33e25cSNeal Cardwell break; 871cf33e25cSNeal Cardwell case BBR_PROBE_RTT: 872cf33e25cSNeal Cardwell bbr->pacing_gain = BBR_UNIT; 873cf33e25cSNeal Cardwell bbr->cwnd_gain = BBR_UNIT; 874cf33e25cSNeal Cardwell break; 875cf33e25cSNeal Cardwell default: 876cf33e25cSNeal Cardwell WARN_ONCE(1, "BBR bad mode: %u\n", bbr->mode); 877cf33e25cSNeal Cardwell break; 878cf33e25cSNeal Cardwell } 879cf33e25cSNeal Cardwell } 880cf33e25cSNeal Cardwell 8810f8782eaSNeal Cardwell static void bbr_update_model(struct sock *sk, const struct rate_sample *rs) 8820f8782eaSNeal Cardwell { 8830f8782eaSNeal Cardwell bbr_update_bw(sk, rs); 8840f8782eaSNeal Cardwell bbr_update_cycle_phase(sk, rs); 8850f8782eaSNeal Cardwell bbr_check_full_bw_reached(sk, rs); 8860f8782eaSNeal Cardwell bbr_check_drain(sk, rs); 8870f8782eaSNeal Cardwell bbr_update_min_rtt(sk, rs); 888cf33e25cSNeal Cardwell bbr_update_gains(sk); 8890f8782eaSNeal Cardwell } 8900f8782eaSNeal Cardwell 8910f8782eaSNeal Cardwell static void bbr_main(struct sock *sk, const struct rate_sample *rs) 8920f8782eaSNeal Cardwell { 8930f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 8940f8782eaSNeal Cardwell u32 bw; 8950f8782eaSNeal Cardwell 8960f8782eaSNeal Cardwell bbr_update_model(sk, rs); 8970f8782eaSNeal Cardwell 8980f8782eaSNeal Cardwell bw = bbr_bw(sk); 8990f8782eaSNeal Cardwell bbr_set_pacing_rate(sk, bw, bbr->pacing_gain); 9000f8782eaSNeal Cardwell bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain); 9010f8782eaSNeal Cardwell } 9020f8782eaSNeal Cardwell 9030f8782eaSNeal Cardwell static void bbr_init(struct sock *sk) 9040f8782eaSNeal Cardwell { 9050f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 9060f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 9070f8782eaSNeal Cardwell 9080f8782eaSNeal Cardwell bbr->prior_cwnd = 0; 90953794570SYousuk Seung tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 9100f8782eaSNeal Cardwell bbr->rtt_cnt = 0; 9110f8782eaSNeal Cardwell bbr->next_rtt_delivered = 0; 9120f8782eaSNeal Cardwell bbr->prev_ca_state = TCP_CA_Open; 9130f8782eaSNeal Cardwell bbr->packet_conservation = 0; 9140f8782eaSNeal Cardwell 9150f8782eaSNeal Cardwell bbr->probe_rtt_done_stamp = 0; 9160f8782eaSNeal Cardwell bbr->probe_rtt_round_done = 0; 9170f8782eaSNeal Cardwell bbr->min_rtt_us = tcp_min_rtt(tp); 9182660bfa8SEric Dumazet bbr->min_rtt_stamp = tcp_jiffies32; 9190f8782eaSNeal Cardwell 9200f8782eaSNeal Cardwell minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */ 9210f8782eaSNeal Cardwell 92232984565SNeal Cardwell bbr->has_seen_rtt = 0; 92379135b89SNeal Cardwell bbr_init_pacing_rate_from_rtt(sk); 9240f8782eaSNeal Cardwell 9250f8782eaSNeal Cardwell bbr->round_start = 0; 9260f8782eaSNeal Cardwell bbr->idle_restart = 0; 927c589e69bSNeal Cardwell bbr->full_bw_reached = 0; 9280f8782eaSNeal Cardwell bbr->full_bw = 0; 9290f8782eaSNeal Cardwell bbr->full_bw_cnt = 0; 9309a568de4SEric Dumazet bbr->cycle_mstamp = 0; 9310f8782eaSNeal Cardwell bbr->cycle_idx = 0; 9320f8782eaSNeal Cardwell bbr_reset_lt_bw_sampling(sk); 9330f8782eaSNeal Cardwell bbr_reset_startup_mode(sk); 934218af599SEric Dumazet 935218af599SEric Dumazet cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED); 9360f8782eaSNeal Cardwell } 9370f8782eaSNeal Cardwell 9380f8782eaSNeal Cardwell static u32 bbr_sndbuf_expand(struct sock *sk) 9390f8782eaSNeal Cardwell { 9400f8782eaSNeal Cardwell /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ 9410f8782eaSNeal Cardwell return 3; 9420f8782eaSNeal Cardwell } 9430f8782eaSNeal Cardwell 9440f8782eaSNeal Cardwell /* In theory BBR does not need to undo the cwnd since it does not 9450f8782eaSNeal Cardwell * always reduce cwnd on losses (see bbr_main()). Keep it for now. 9460f8782eaSNeal Cardwell */ 9470f8782eaSNeal Cardwell static u32 bbr_undo_cwnd(struct sock *sk) 9480f8782eaSNeal Cardwell { 9492f6c498eSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 9502f6c498eSNeal Cardwell 9512f6c498eSNeal Cardwell bbr->full_bw = 0; /* spurious slow-down; reset full pipe detection */ 9522f6c498eSNeal Cardwell bbr->full_bw_cnt = 0; 953600647d4SNeal Cardwell bbr_reset_lt_bw_sampling(sk); 9540f8782eaSNeal Cardwell return tcp_sk(sk)->snd_cwnd; 9550f8782eaSNeal Cardwell } 9560f8782eaSNeal Cardwell 9570f8782eaSNeal Cardwell /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */ 9580f8782eaSNeal Cardwell static u32 bbr_ssthresh(struct sock *sk) 9590f8782eaSNeal Cardwell { 9600f8782eaSNeal Cardwell bbr_save_cwnd(sk); 96153794570SYousuk Seung return tcp_sk(sk)->snd_ssthresh; 9620f8782eaSNeal Cardwell } 9630f8782eaSNeal Cardwell 9640f8782eaSNeal Cardwell static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr, 9650f8782eaSNeal Cardwell union tcp_cc_info *info) 9660f8782eaSNeal Cardwell { 9670f8782eaSNeal Cardwell if (ext & (1 << (INET_DIAG_BBRINFO - 1)) || 9680f8782eaSNeal Cardwell ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 9690f8782eaSNeal Cardwell struct tcp_sock *tp = tcp_sk(sk); 9700f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 9710f8782eaSNeal Cardwell u64 bw = bbr_bw(sk); 9720f8782eaSNeal Cardwell 9730f8782eaSNeal Cardwell bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE; 9740f8782eaSNeal Cardwell memset(&info->bbr, 0, sizeof(info->bbr)); 9750f8782eaSNeal Cardwell info->bbr.bbr_bw_lo = (u32)bw; 9760f8782eaSNeal Cardwell info->bbr.bbr_bw_hi = (u32)(bw >> 32); 9770f8782eaSNeal Cardwell info->bbr.bbr_min_rtt = bbr->min_rtt_us; 9780f8782eaSNeal Cardwell info->bbr.bbr_pacing_gain = bbr->pacing_gain; 9790f8782eaSNeal Cardwell info->bbr.bbr_cwnd_gain = bbr->cwnd_gain; 9800f8782eaSNeal Cardwell *attr = INET_DIAG_BBRINFO; 9810f8782eaSNeal Cardwell return sizeof(info->bbr); 9820f8782eaSNeal Cardwell } 9830f8782eaSNeal Cardwell return 0; 9840f8782eaSNeal Cardwell } 9850f8782eaSNeal Cardwell 9860f8782eaSNeal Cardwell static void bbr_set_state(struct sock *sk, u8 new_state) 9870f8782eaSNeal Cardwell { 9880f8782eaSNeal Cardwell struct bbr *bbr = inet_csk_ca(sk); 9890f8782eaSNeal Cardwell 9900f8782eaSNeal Cardwell if (new_state == TCP_CA_Loss) { 9910f8782eaSNeal Cardwell struct rate_sample rs = { .losses = 1 }; 9920f8782eaSNeal Cardwell 9930f8782eaSNeal Cardwell bbr->prev_ca_state = TCP_CA_Loss; 9940f8782eaSNeal Cardwell bbr->full_bw = 0; 9950f8782eaSNeal Cardwell bbr->round_start = 1; /* treat RTO like end of a round */ 9960f8782eaSNeal Cardwell bbr_lt_bw_sampling(sk, &rs); 9970f8782eaSNeal Cardwell } 9980f8782eaSNeal Cardwell } 9990f8782eaSNeal Cardwell 10000f8782eaSNeal Cardwell static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { 10010f8782eaSNeal Cardwell .flags = TCP_CONG_NON_RESTRICTED, 10020f8782eaSNeal Cardwell .name = "bbr", 10030f8782eaSNeal Cardwell .owner = THIS_MODULE, 10040f8782eaSNeal Cardwell .init = bbr_init, 10050f8782eaSNeal Cardwell .cong_control = bbr_main, 10060f8782eaSNeal Cardwell .sndbuf_expand = bbr_sndbuf_expand, 10070f8782eaSNeal Cardwell .undo_cwnd = bbr_undo_cwnd, 10080f8782eaSNeal Cardwell .cwnd_event = bbr_cwnd_event, 10090f8782eaSNeal Cardwell .ssthresh = bbr_ssthresh, 1010dcb8c9b4SEric Dumazet .min_tso_segs = bbr_min_tso_segs, 10110f8782eaSNeal Cardwell .get_info = bbr_get_info, 10120f8782eaSNeal Cardwell .set_state = bbr_set_state, 10130f8782eaSNeal Cardwell }; 10140f8782eaSNeal Cardwell 10150f8782eaSNeal Cardwell static int __init bbr_register(void) 10160f8782eaSNeal Cardwell { 10170f8782eaSNeal Cardwell BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE); 10180f8782eaSNeal Cardwell return tcp_register_congestion_control(&tcp_bbr_cong_ops); 10190f8782eaSNeal Cardwell } 10200f8782eaSNeal Cardwell 10210f8782eaSNeal Cardwell static void __exit bbr_unregister(void) 10220f8782eaSNeal Cardwell { 10230f8782eaSNeal Cardwell tcp_unregister_congestion_control(&tcp_bbr_cong_ops); 10240f8782eaSNeal Cardwell } 10250f8782eaSNeal Cardwell 10260f8782eaSNeal Cardwell module_init(bbr_register); 10270f8782eaSNeal Cardwell module_exit(bbr_unregister); 10280f8782eaSNeal Cardwell 10290f8782eaSNeal Cardwell MODULE_AUTHOR("Van Jacobson <vanj@google.com>"); 10300f8782eaSNeal Cardwell MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>"); 10310f8782eaSNeal Cardwell MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>"); 10320f8782eaSNeal Cardwell MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>"); 10330f8782eaSNeal Cardwell MODULE_LICENSE("Dual BSD/GPL"); 10340f8782eaSNeal Cardwell MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)"); 1035