linux/net/ipv4/tcp_bbr.c
<<
>>
Prefs
   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: */
  79enum 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 */
  87struct 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        /* For tracking ACK aggregation: */
 120        u64     ack_epoch_mstamp;       /* start of ACK sampling epoch */
 121        u16     extra_acked[2];         /* max excess data ACKed in epoch */
 122        u32     ack_epoch_acked:20,     /* packets (S)ACKed in sampling epoch */
 123                extra_acked_win_rtts:5, /* age of extra_acked, in round trips */
 124                extra_acked_win_idx:1,  /* current index in extra_acked array */
 125                unused_c:6;
 126};
 127
 128#define CYCLE_LEN       8       /* number of phases in a pacing gain cycle */
 129
 130/* Window length of bw filter (in rounds): */
 131static const int bbr_bw_rtts = CYCLE_LEN + 2;
 132/* Window length of min_rtt filter (in sec): */
 133static const u32 bbr_min_rtt_win_sec = 10;
 134/* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
 135static const u32 bbr_probe_rtt_mode_ms = 200;
 136/* Skip TSO below the following bandwidth (bits/sec): */
 137static const int bbr_min_tso_rate = 1200000;
 138
 139/* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck.
 140 * In order to help drive the network toward lower queues and low latency while
 141 * maintaining high utilization, the average pacing rate aims to be slightly
 142 * lower than the estimated bandwidth. This is an important aspect of the
 143 * design.
 144 */
 145static const int bbr_pacing_margin_percent = 1;
 146
 147/* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
 148 * that will allow a smoothly increasing pacing rate that will double each RTT
 149 * and send the same number of packets per RTT that an un-paced, slow-starting
 150 * Reno or CUBIC flow would:
 151 */
 152static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
 153/* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
 154 * the queue created in BBR_STARTUP in a single round:
 155 */
 156static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
 157/* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
 158static const int bbr_cwnd_gain  = BBR_UNIT * 2;
 159/* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
 160static const int bbr_pacing_gain[] = {
 161        BBR_UNIT * 5 / 4,       /* probe for more available bw */
 162        BBR_UNIT * 3 / 4,       /* drain queue and/or yield bw to other flows */
 163        BBR_UNIT, BBR_UNIT, BBR_UNIT,   /* cruise at 1.0*bw to utilize pipe, */
 164        BBR_UNIT, BBR_UNIT, BBR_UNIT    /* without creating excess queue... */
 165};
 166/* Randomize the starting gain cycling phase over N phases: */
 167static const u32 bbr_cycle_rand = 7;
 168
 169/* Try to keep at least this many packets in flight, if things go smoothly. For
 170 * smooth functioning, a sliding window protocol ACKing every other packet
 171 * needs at least 4 packets in flight:
 172 */
 173static const u32 bbr_cwnd_min_target = 4;
 174
 175/* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
 176/* If bw has increased significantly (1.25x), there may be more bw available: */
 177static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
 178/* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
 179static const u32 bbr_full_bw_cnt = 3;
 180
 181/* "long-term" ("LT") bandwidth estimator parameters... */
 182/* The minimum number of rounds in an LT bw sampling interval: */
 183static const u32 bbr_lt_intvl_min_rtts = 4;
 184/* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
 185static const u32 bbr_lt_loss_thresh = 50;
 186/* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
 187static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
 188/* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
 189static const u32 bbr_lt_bw_diff = 4000 / 8;
 190/* If we estimate we're policed, use lt_bw for this many round trips: */
 191static const u32 bbr_lt_bw_max_rtts = 48;
 192
 193/* Gain factor for adding extra_acked to target cwnd: */
 194static const int bbr_extra_acked_gain = BBR_UNIT;
 195/* Window length of extra_acked window. */
 196static const u32 bbr_extra_acked_win_rtts = 5;
 197/* Max allowed val for ack_epoch_acked, after which sampling epoch is reset */
 198static const u32 bbr_ack_epoch_acked_reset_thresh = 1U << 20;
 199/* Time period for clamping cwnd increment due to ack aggregation */
 200static const u32 bbr_extra_acked_max_us = 100 * 1000;
 201
 202static void bbr_check_probe_rtt_done(struct sock *sk);
 203
 204/* Do we estimate that STARTUP filled the pipe? */
 205static bool bbr_full_bw_reached(const struct sock *sk)
 206{
 207        const struct bbr *bbr = inet_csk_ca(sk);
 208
 209        return bbr->full_bw_reached;
 210}
 211
 212/* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
 213static u32 bbr_max_bw(const struct sock *sk)
 214{
 215        struct bbr *bbr = inet_csk_ca(sk);
 216
 217        return minmax_get(&bbr->bw);
 218}
 219
 220/* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
 221static u32 bbr_bw(const struct sock *sk)
 222{
 223        struct bbr *bbr = inet_csk_ca(sk);
 224
 225        return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
 226}
 227
 228/* Return maximum extra acked in past k-2k round trips,
 229 * where k = bbr_extra_acked_win_rtts.
 230 */
 231static u16 bbr_extra_acked(const struct sock *sk)
 232{
 233        struct bbr *bbr = inet_csk_ca(sk);
 234
 235        return max(bbr->extra_acked[0], bbr->extra_acked[1]);
 236}
 237
 238/* Return rate in bytes per second, optionally with a gain.
 239 * The order here is chosen carefully to avoid overflow of u64. This should
 240 * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
 241 */
 242static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
 243{
 244        unsigned int mss = tcp_sk(sk)->mss_cache;
 245
 246        rate *= mss;
 247        rate *= gain;
 248        rate >>= BBR_SCALE;
 249        rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent);
 250        return rate >> BW_SCALE;
 251}
 252
 253/* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
 254static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
 255{
 256        u64 rate = bw;
 257
 258        rate = bbr_rate_bytes_per_sec(sk, rate, gain);
 259        rate = min_t(u64, rate, sk->sk_max_pacing_rate);
 260        return rate;
 261}
 262
 263/* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
 264static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
 265{
 266        struct tcp_sock *tp = tcp_sk(sk);
 267        struct bbr *bbr = inet_csk_ca(sk);
 268        u64 bw;
 269        u32 rtt_us;
 270
 271        if (tp->srtt_us) {              /* any RTT sample yet? */
 272                rtt_us = max(tp->srtt_us >> 3, 1U);
 273                bbr->has_seen_rtt = 1;
 274        } else {                         /* no RTT sample yet */
 275                rtt_us = USEC_PER_MSEC;  /* use nominal default RTT */
 276        }
 277        bw = (u64)tp->snd_cwnd * BW_UNIT;
 278        do_div(bw, rtt_us);
 279        sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain);
 280}
 281
 282/* Pace using current bw estimate and a gain factor. */
 283static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
 284{
 285        struct tcp_sock *tp = tcp_sk(sk);
 286        struct bbr *bbr = inet_csk_ca(sk);
 287        unsigned long rate = bbr_bw_to_pacing_rate(sk, bw, gain);
 288
 289        if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
 290                bbr_init_pacing_rate_from_rtt(sk);
 291        if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate)
 292                sk->sk_pacing_rate = rate;
 293}
 294
 295/* override sysctl_tcp_min_tso_segs */
 296static u32 bbr_min_tso_segs(struct sock *sk)
 297{
 298        return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
 299}
 300
 301static u32 bbr_tso_segs_goal(struct sock *sk)
 302{
 303        struct tcp_sock *tp = tcp_sk(sk);
 304        u32 segs, bytes;
 305
 306        /* Sort of tcp_tso_autosize() but ignoring
 307         * driver provided sk_gso_max_size.
 308         */
 309        bytes = min_t(unsigned long,
 310                      sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
 311                      GSO_MAX_SIZE - 1 - MAX_TCP_HEADER);
 312        segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk));
 313
 314        return min(segs, 0x7FU);
 315}
 316
 317/* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
 318static void bbr_save_cwnd(struct sock *sk)
 319{
 320        struct tcp_sock *tp = tcp_sk(sk);
 321        struct bbr *bbr = inet_csk_ca(sk);
 322
 323        if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
 324                bbr->prior_cwnd = tp->snd_cwnd;  /* this cwnd is good enough */
 325        else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
 326                bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
 327}
 328
 329static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 330{
 331        struct tcp_sock *tp = tcp_sk(sk);
 332        struct bbr *bbr = inet_csk_ca(sk);
 333
 334        if (event == CA_EVENT_TX_START && tp->app_limited) {
 335                bbr->idle_restart = 1;
 336                bbr->ack_epoch_mstamp = tp->tcp_mstamp;
 337                bbr->ack_epoch_acked = 0;
 338                /* Avoid pointless buffer overflows: pace at est. bw if we don't
 339                 * need more speed (we're restarting from idle and app-limited).
 340                 */
 341                if (bbr->mode == BBR_PROBE_BW)
 342                        bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
 343                else if (bbr->mode == BBR_PROBE_RTT)
 344                        bbr_check_probe_rtt_done(sk);
 345        }
 346}
 347
 348/* Calculate bdp based on min RTT and the estimated bottleneck bandwidth:
 349 *
 350 * bdp = ceil(bw * min_rtt * gain)
 351 *
 352 * The key factor, gain, controls the amount of queue. While a small gain
 353 * builds a smaller queue, it becomes more vulnerable to noise in RTT
 354 * measurements (e.g., delayed ACKs or other ACK compression effects). This
 355 * noise may cause BBR to under-estimate the rate.
 356 */
 357static u32 bbr_bdp(struct sock *sk, u32 bw, int gain)
 358{
 359        struct bbr *bbr = inet_csk_ca(sk);
 360        u32 bdp;
 361        u64 w;
 362
 363        /* If we've never had a valid RTT sample, cap cwnd at the initial
 364         * default. This should only happen when the connection is not using TCP
 365         * timestamps and has retransmitted all of the SYN/SYNACK/data packets
 366         * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
 367         * case we need to slow-start up toward something safe: TCP_INIT_CWND.
 368         */
 369        if (unlikely(bbr->min_rtt_us == ~0U))    /* no valid RTT samples yet? */
 370                return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
 371
 372        w = (u64)bw * bbr->min_rtt_us;
 373
 374        /* Apply a gain to the given value, remove the BW_SCALE shift, and
 375         * round the value up to avoid a negative feedback loop.
 376         */
 377        bdp = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
 378
 379        return bdp;
 380}
 381
 382/* To achieve full performance in high-speed paths, we budget enough cwnd to
 383 * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
 384 *   - one skb in sending host Qdisc,
 385 *   - one skb in sending host TSO/GSO engine
 386 *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
 387 * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
 388 * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
 389 * which allows 2 outstanding 2-packet sequences, to try to keep pipe
 390 * full even with ACK-every-other-packet delayed ACKs.
 391 */
 392static u32 bbr_quantization_budget(struct sock *sk, u32 cwnd)
 393{
 394        struct bbr *bbr = inet_csk_ca(sk);
 395
 396        /* Allow enough full-sized skbs in flight to utilize end systems. */
 397        cwnd += 3 * bbr_tso_segs_goal(sk);
 398
 399        /* Reduce delayed ACKs by rounding up cwnd to the next even number. */
 400        cwnd = (cwnd + 1) & ~1U;
 401
 402        /* Ensure gain cycling gets inflight above BDP even for small BDPs. */
 403        if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == 0)
 404                cwnd += 2;
 405
 406        return cwnd;
 407}
 408
 409/* Find inflight based on min RTT and the estimated bottleneck bandwidth. */
 410static u32 bbr_inflight(struct sock *sk, u32 bw, int gain)
 411{
 412        u32 inflight;
 413
 414        inflight = bbr_bdp(sk, bw, gain);
 415        inflight = bbr_quantization_budget(sk, inflight);
 416
 417        return inflight;
 418}
 419
 420/* With pacing at lower layers, there's often less data "in the network" than
 421 * "in flight". With TSQ and departure time pacing at lower layers (e.g. fq),
 422 * we often have several skbs queued in the pacing layer with a pre-scheduled
 423 * earliest departure time (EDT). BBR adapts its pacing rate based on the
 424 * inflight level that it estimates has already been "baked in" by previous
 425 * departure time decisions. We calculate a rough estimate of the number of our
 426 * packets that might be in the network at the earliest departure time for the
 427 * next skb scheduled:
 428 *   in_network_at_edt = inflight_at_edt - (EDT - now) * bw
 429 * If we're increasing inflight, then we want to know if the transmit of the
 430 * EDT skb will push inflight above the target, so inflight_at_edt includes
 431 * bbr_tso_segs_goal() from the skb departing at EDT. If decreasing inflight,
 432 * then estimate if inflight will sink too low just before the EDT transmit.
 433 */
 434static u32 bbr_packets_in_net_at_edt(struct sock *sk, u32 inflight_now)
 435{
 436        struct tcp_sock *tp = tcp_sk(sk);
 437        struct bbr *bbr = inet_csk_ca(sk);
 438        u64 now_ns, edt_ns, interval_us;
 439        u32 interval_delivered, inflight_at_edt;
 440
 441        now_ns = tp->tcp_clock_cache;
 442        edt_ns = max(tp->tcp_wstamp_ns, now_ns);
 443        interval_us = div_u64(edt_ns - now_ns, NSEC_PER_USEC);
 444        interval_delivered = (u64)bbr_bw(sk) * interval_us >> BW_SCALE;
 445        inflight_at_edt = inflight_now;
 446        if (bbr->pacing_gain > BBR_UNIT)              /* increasing inflight */
 447                inflight_at_edt += bbr_tso_segs_goal(sk);  /* include EDT skb */
 448        if (interval_delivered >= inflight_at_edt)
 449                return 0;
 450        return inflight_at_edt - interval_delivered;
 451}
 452
 453/* Find the cwnd increment based on estimate of ack aggregation */
 454static u32 bbr_ack_aggregation_cwnd(struct sock *sk)
 455{
 456        u32 max_aggr_cwnd, aggr_cwnd = 0;
 457
 458        if (bbr_extra_acked_gain && bbr_full_bw_reached(sk)) {
 459                max_aggr_cwnd = ((u64)bbr_bw(sk) * bbr_extra_acked_max_us)
 460                                / BW_UNIT;
 461                aggr_cwnd = (bbr_extra_acked_gain * bbr_extra_acked(sk))
 462                             >> BBR_SCALE;
 463                aggr_cwnd = min(aggr_cwnd, max_aggr_cwnd);
 464        }
 465
 466        return aggr_cwnd;
 467}
 468
 469/* An optimization in BBR to reduce losses: On the first round of recovery, we
 470 * follow the packet conservation principle: send P packets per P packets acked.
 471 * After that, we slow-start and send at most 2*P packets per P packets acked.
 472 * After recovery finishes, or upon undo, we restore the cwnd we had when
 473 * recovery started (capped by the target cwnd based on estimated BDP).
 474 *
 475 * TODO(ycheng/ncardwell): implement a rate-based approach.
 476 */
 477static bool bbr_set_cwnd_to_recover_or_restore(
 478        struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
 479{
 480        struct tcp_sock *tp = tcp_sk(sk);
 481        struct bbr *bbr = inet_csk_ca(sk);
 482        u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
 483        u32 cwnd = tp->snd_cwnd;
 484
 485        /* An ACK for P pkts should release at most 2*P packets. We do this
 486         * in two steps. First, here we deduct the number of lost packets.
 487         * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
 488         */
 489        if (rs->losses > 0)
 490                cwnd = max_t(s32, cwnd - rs->losses, 1);
 491
 492        if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
 493                /* Starting 1st round of Recovery, so do packet conservation. */
 494                bbr->packet_conservation = 1;
 495                bbr->next_rtt_delivered = tp->delivered;  /* start round now */
 496                /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
 497                cwnd = tcp_packets_in_flight(tp) + acked;
 498        } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
 499                /* Exiting loss recovery; restore cwnd saved before recovery. */
 500                cwnd = max(cwnd, bbr->prior_cwnd);
 501                bbr->packet_conservation = 0;
 502        }
 503        bbr->prev_ca_state = state;
 504
 505        if (bbr->packet_conservation) {
 506                *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
 507                return true;    /* yes, using packet conservation */
 508        }
 509        *new_cwnd = cwnd;
 510        return false;
 511}
 512
 513/* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
 514 * has drawn us down below target), or snap down to target if we're above it.
 515 */
 516static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
 517                         u32 acked, u32 bw, int gain)
 518{
 519        struct tcp_sock *tp = tcp_sk(sk);
 520        struct bbr *bbr = inet_csk_ca(sk);
 521        u32 cwnd = tp->snd_cwnd, target_cwnd = 0;
 522
 523        if (!acked)
 524                goto done;  /* no packet fully ACKed; just apply caps */
 525
 526        if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
 527                goto done;
 528
 529        target_cwnd = bbr_bdp(sk, bw, gain);
 530
 531        /* Increment the cwnd to account for excess ACKed data that seems
 532         * due to aggregation (of data and/or ACKs) visible in the ACK stream.
 533         */
 534        target_cwnd += bbr_ack_aggregation_cwnd(sk);
 535        target_cwnd = bbr_quantization_budget(sk, target_cwnd);
 536
 537        /* If we're below target cwnd, slow start cwnd toward target cwnd. */
 538        if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
 539                cwnd = min(cwnd + acked, target_cwnd);
 540        else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
 541                cwnd = cwnd + acked;
 542        cwnd = max(cwnd, bbr_cwnd_min_target);
 543
 544done:
 545        tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);   /* apply global cap */
 546        if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
 547                tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
 548}
 549
 550/* End cycle phase if it's time and/or we hit the phase's in-flight target. */
 551static bool bbr_is_next_cycle_phase(struct sock *sk,
 552                                    const struct rate_sample *rs)
 553{
 554        struct tcp_sock *tp = tcp_sk(sk);
 555        struct bbr *bbr = inet_csk_ca(sk);
 556        bool is_full_length =
 557                tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
 558                bbr->min_rtt_us;
 559        u32 inflight, bw;
 560
 561        /* The pacing_gain of 1.0 paces at the estimated bw to try to fully
 562         * use the pipe without increasing the queue.
 563         */
 564        if (bbr->pacing_gain == BBR_UNIT)
 565                return is_full_length;          /* just use wall clock time */
 566
 567        inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight);
 568        bw = bbr_max_bw(sk);
 569
 570        /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
 571         * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
 572         * small (e.g. on a LAN). We do not persist if packets are lost, since
 573         * a path with small buffers may not hold that much.
 574         */
 575        if (bbr->pacing_gain > BBR_UNIT)
 576                return is_full_length &&
 577                        (rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
 578                         inflight >= bbr_inflight(sk, bw, bbr->pacing_gain));
 579
 580        /* A pacing_gain < 1.0 tries to drain extra queue we added if bw
 581         * probing didn't find more bw. If inflight falls to match BDP then we
 582         * estimate queue is drained; persisting would underutilize the pipe.
 583         */
 584        return is_full_length ||
 585                inflight <= bbr_inflight(sk, bw, BBR_UNIT);
 586}
 587
 588static void bbr_advance_cycle_phase(struct sock *sk)
 589{
 590        struct tcp_sock *tp = tcp_sk(sk);
 591        struct bbr *bbr = inet_csk_ca(sk);
 592
 593        bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
 594        bbr->cycle_mstamp = tp->delivered_mstamp;
 595}
 596
 597/* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
 598static void bbr_update_cycle_phase(struct sock *sk,
 599                                   const struct rate_sample *rs)
 600{
 601        struct bbr *bbr = inet_csk_ca(sk);
 602
 603        if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
 604                bbr_advance_cycle_phase(sk);
 605}
 606
 607static void bbr_reset_startup_mode(struct sock *sk)
 608{
 609        struct bbr *bbr = inet_csk_ca(sk);
 610
 611        bbr->mode = BBR_STARTUP;
 612}
 613
 614static void bbr_reset_probe_bw_mode(struct sock *sk)
 615{
 616        struct bbr *bbr = inet_csk_ca(sk);
 617
 618        bbr->mode = BBR_PROBE_BW;
 619        bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
 620        bbr_advance_cycle_phase(sk);    /* flip to next phase of gain cycle */
 621}
 622
 623static void bbr_reset_mode(struct sock *sk)
 624{
 625        if (!bbr_full_bw_reached(sk))
 626                bbr_reset_startup_mode(sk);
 627        else
 628                bbr_reset_probe_bw_mode(sk);
 629}
 630
 631/* Start a new long-term sampling interval. */
 632static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
 633{
 634        struct tcp_sock *tp = tcp_sk(sk);
 635        struct bbr *bbr = inet_csk_ca(sk);
 636
 637        bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
 638        bbr->lt_last_delivered = tp->delivered;
 639        bbr->lt_last_lost = tp->lost;
 640        bbr->lt_rtt_cnt = 0;
 641}
 642
 643/* Completely reset long-term bandwidth sampling. */
 644static void bbr_reset_lt_bw_sampling(struct sock *sk)
 645{
 646        struct bbr *bbr = inet_csk_ca(sk);
 647
 648        bbr->lt_bw = 0;
 649        bbr->lt_use_bw = 0;
 650        bbr->lt_is_sampling = false;
 651        bbr_reset_lt_bw_sampling_interval(sk);
 652}
 653
 654/* Long-term bw sampling interval is done. Estimate whether we're policed. */
 655static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
 656{
 657        struct bbr *bbr = inet_csk_ca(sk);
 658        u32 diff;
 659
 660        if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
 661                /* Is new bw close to the lt_bw from the previous interval? */
 662                diff = abs(bw - bbr->lt_bw);
 663                if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
 664                    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
 665                     bbr_lt_bw_diff)) {
 666                        /* All criteria are met; estimate we're policed. */
 667                        bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
 668                        bbr->lt_use_bw = 1;
 669                        bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
 670                        bbr->lt_rtt_cnt = 0;
 671                        return;
 672                }
 673        }
 674        bbr->lt_bw = bw;
 675        bbr_reset_lt_bw_sampling_interval(sk);
 676}
 677
 678/* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
 679 * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
 680 * explicitly models their policed rate, to reduce unnecessary losses. We
 681 * estimate that we're policed if we see 2 consecutive sampling intervals with
 682 * consistent throughput and high packet loss. If we think we're being policed,
 683 * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
 684 */
 685static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
 686{
 687        struct tcp_sock *tp = tcp_sk(sk);
 688        struct bbr *bbr = inet_csk_ca(sk);
 689        u32 lost, delivered;
 690        u64 bw;
 691        u32 t;
 692
 693        if (bbr->lt_use_bw) {   /* already using long-term rate, lt_bw? */
 694                if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
 695                    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
 696                        bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
 697                        bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
 698                }
 699                return;
 700        }
 701
 702        /* Wait for the first loss before sampling, to let the policer exhaust
 703         * its tokens and estimate the steady-state rate allowed by the policer.
 704         * Starting samples earlier includes bursts that over-estimate the bw.
 705         */
 706        if (!bbr->lt_is_sampling) {
 707                if (!rs->losses)
 708                        return;
 709                bbr_reset_lt_bw_sampling_interval(sk);
 710                bbr->lt_is_sampling = true;
 711        }
 712
 713        /* To avoid underestimates, reset sampling if we run out of data. */
 714        if (rs->is_app_limited) {
 715                bbr_reset_lt_bw_sampling(sk);
 716                return;
 717        }
 718
 719        if (bbr->round_start)
 720                bbr->lt_rtt_cnt++;      /* count round trips in this interval */
 721        if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
 722                return;         /* sampling interval needs to be longer */
 723        if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
 724                bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
 725                return;
 726        }
 727
 728        /* End sampling interval when a packet is lost, so we estimate the
 729         * policer tokens were exhausted. Stopping the sampling before the
 730         * tokens are exhausted under-estimates the policed rate.
 731         */
 732        if (!rs->losses)
 733                return;
 734
 735        /* Calculate packets lost and delivered in sampling interval. */
 736        lost = tp->lost - bbr->lt_last_lost;
 737        delivered = tp->delivered - bbr->lt_last_delivered;
 738        /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
 739        if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
 740                return;
 741
 742        /* Find average delivery rate in this sampling interval. */
 743        t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
 744        if ((s32)t < 1)
 745                return;         /* interval is less than one ms, so wait */
 746        /* Check if can multiply without overflow */
 747        if (t >= ~0U / USEC_PER_MSEC) {
 748                bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
 749                return;
 750        }
 751        t *= USEC_PER_MSEC;
 752        bw = (u64)delivered * BW_UNIT;
 753        do_div(bw, t);
 754        bbr_lt_bw_interval_done(sk, bw);
 755}
 756
 757/* Estimate the bandwidth based on how fast packets are delivered */
 758static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
 759{
 760        struct tcp_sock *tp = tcp_sk(sk);
 761        struct bbr *bbr = inet_csk_ca(sk);
 762        u64 bw;
 763
 764        bbr->round_start = 0;
 765        if (rs->delivered < 0 || rs->interval_us <= 0)
 766                return; /* Not a valid observation */
 767
 768        /* See if we've reached the next RTT */
 769        if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
 770                bbr->next_rtt_delivered = tp->delivered;
 771                bbr->rtt_cnt++;
 772                bbr->round_start = 1;
 773                bbr->packet_conservation = 0;
 774        }
 775
 776        bbr_lt_bw_sampling(sk, rs);
 777
 778        /* Divide delivered by the interval to find a (lower bound) bottleneck
 779         * bandwidth sample. Delivered is in packets and interval_us in uS and
 780         * ratio will be <<1 for most connections. So delivered is first scaled.
 781         */
 782        bw = div64_long((u64)rs->delivered * BW_UNIT, rs->interval_us);
 783
 784        /* If this sample is application-limited, it is likely to have a very
 785         * low delivered count that represents application behavior rather than
 786         * the available network rate. Such a sample could drag down estimated
 787         * bw, causing needless slow-down. Thus, to continue to send at the
 788         * last measured network rate, we filter out app-limited samples unless
 789         * they describe the path bw at least as well as our bw model.
 790         *
 791         * So the goal during app-limited phase is to proceed with the best
 792         * network rate no matter how long. We automatically leave this
 793         * phase when app writes faster than the network can deliver :)
 794         */
 795        if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
 796                /* Incorporate new sample into our max bw filter. */
 797                minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
 798        }
 799}
 800
 801/* Estimates the windowed max degree of ack aggregation.
 802 * This is used to provision extra in-flight data to keep sending during
 803 * inter-ACK silences.
 804 *
 805 * Degree of ack aggregation is estimated as extra data acked beyond expected.
 806 *
 807 * max_extra_acked = "maximum recent excess data ACKed beyond max_bw * interval"
 808 * cwnd += max_extra_acked
 809 *
 810 * Max extra_acked is clamped by cwnd and bw * bbr_extra_acked_max_us (100 ms).
 811 * Max filter is an approximate sliding window of 5-10 (packet timed) round
 812 * trips.
 813 */
 814static void bbr_update_ack_aggregation(struct sock *sk,
 815                                       const struct rate_sample *rs)
 816{
 817        u32 epoch_us, expected_acked, extra_acked;
 818        struct bbr *bbr = inet_csk_ca(sk);
 819        struct tcp_sock *tp = tcp_sk(sk);
 820
 821        if (!bbr_extra_acked_gain || rs->acked_sacked <= 0 ||
 822            rs->delivered < 0 || rs->interval_us <= 0)
 823                return;
 824
 825        if (bbr->round_start) {
 826                bbr->extra_acked_win_rtts = min(0x1F,
 827                                                bbr->extra_acked_win_rtts + 1);
 828                if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
 829                        bbr->extra_acked_win_rtts = 0;
 830                        bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
 831                                                   0 : 1;
 832                        bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
 833                }
 834        }
 835
 836        /* Compute how many packets we expected to be delivered over epoch. */
 837        epoch_us = tcp_stamp_us_delta(tp->delivered_mstamp,
 838                                      bbr->ack_epoch_mstamp);
 839        expected_acked = ((u64)bbr_bw(sk) * epoch_us) / BW_UNIT;
 840
 841        /* Reset the aggregation epoch if ACK rate is below expected rate or
 842         * significantly large no. of ack received since epoch (potentially
 843         * quite old epoch).
 844         */
 845        if (bbr->ack_epoch_acked <= expected_acked ||
 846            (bbr->ack_epoch_acked + rs->acked_sacked >=
 847             bbr_ack_epoch_acked_reset_thresh)) {
 848                bbr->ack_epoch_acked = 0;
 849                bbr->ack_epoch_mstamp = tp->delivered_mstamp;
 850                expected_acked = 0;
 851        }
 852
 853        /* Compute excess data delivered, beyond what was expected. */
 854        bbr->ack_epoch_acked = min_t(u32, 0xFFFFF,
 855                                     bbr->ack_epoch_acked + rs->acked_sacked);
 856        extra_acked = bbr->ack_epoch_acked - expected_acked;
 857        extra_acked = min(extra_acked, tp->snd_cwnd);
 858        if (extra_acked > bbr->extra_acked[bbr->extra_acked_win_idx])
 859                bbr->extra_acked[bbr->extra_acked_win_idx] = extra_acked;
 860}
 861
 862/* Estimate when the pipe is full, using the change in delivery rate: BBR
 863 * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
 864 * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
 865 * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
 866 * higher rwin, 3: we get higher delivery rate samples. Or transient
 867 * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
 868 * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
 869 */
 870static void bbr_check_full_bw_reached(struct sock *sk,
 871                                      const struct rate_sample *rs)
 872{
 873        struct bbr *bbr = inet_csk_ca(sk);
 874        u32 bw_thresh;
 875
 876        if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
 877                return;
 878
 879        bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
 880        if (bbr_max_bw(sk) >= bw_thresh) {
 881                bbr->full_bw = bbr_max_bw(sk);
 882                bbr->full_bw_cnt = 0;
 883                return;
 884        }
 885        ++bbr->full_bw_cnt;
 886        bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
 887}
 888
 889/* If pipe is probably full, drain the queue and then enter steady-state. */
 890static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
 891{
 892        struct bbr *bbr = inet_csk_ca(sk);
 893
 894        if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
 895                bbr->mode = BBR_DRAIN;  /* drain queue we created */
 896                tcp_sk(sk)->snd_ssthresh =
 897                                bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT);
 898        }       /* fall through to check if in-flight is already small: */
 899        if (bbr->mode == BBR_DRAIN &&
 900            bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <=
 901            bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT))
 902                bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
 903}
 904
 905static void bbr_check_probe_rtt_done(struct sock *sk)
 906{
 907        struct tcp_sock *tp = tcp_sk(sk);
 908        struct bbr *bbr = inet_csk_ca(sk);
 909
 910        if (!(bbr->probe_rtt_done_stamp &&
 911              after(tcp_jiffies32, bbr->probe_rtt_done_stamp)))
 912                return;
 913
 914        bbr->min_rtt_stamp = tcp_jiffies32;  /* wait a while until PROBE_RTT */
 915        tp->snd_cwnd = max(tp->snd_cwnd, bbr->prior_cwnd);
 916        bbr_reset_mode(sk);
 917}
 918
 919/* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
 920 * periodically drain the bottleneck queue, to converge to measure the true
 921 * min_rtt (unloaded propagation delay). This allows the flows to keep queues
 922 * small (reducing queuing delay and packet loss) and achieve fairness among
 923 * BBR flows.
 924 *
 925 * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
 926 * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
 927 * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
 928 * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
 929 * re-enter the previous mode. BBR uses 200ms to approximately bound the
 930 * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
 931 *
 932 * Note that flows need only pay 2% if they are busy sending over the last 10
 933 * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
 934 * natural silences or low-rate periods within 10 seconds where the rate is low
 935 * enough for long enough to drain its queue in the bottleneck. We pick up
 936 * these min RTT measurements opportunistically with our min_rtt filter. :-)
 937 */
 938static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
 939{
 940        struct tcp_sock *tp = tcp_sk(sk);
 941        struct bbr *bbr = inet_csk_ca(sk);
 942        bool filter_expired;
 943
 944        /* Track min RTT seen in the min_rtt_win_sec filter window: */
 945        filter_expired = after(tcp_jiffies32,
 946                               bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
 947        if (rs->rtt_us >= 0 &&
 948            (rs->rtt_us <= bbr->min_rtt_us ||
 949             (filter_expired && !rs->is_ack_delayed))) {
 950                bbr->min_rtt_us = rs->rtt_us;
 951                bbr->min_rtt_stamp = tcp_jiffies32;
 952        }
 953
 954        if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
 955            !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
 956                bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
 957                bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
 958                bbr->probe_rtt_done_stamp = 0;
 959        }
 960
 961        if (bbr->mode == BBR_PROBE_RTT) {
 962                /* Ignore low rate samples during this mode. */
 963                tp->app_limited =
 964                        (tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
 965                /* Maintain min packets in flight for max(200 ms, 1 round). */
 966                if (!bbr->probe_rtt_done_stamp &&
 967                    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
 968                        bbr->probe_rtt_done_stamp = tcp_jiffies32 +
 969                                msecs_to_jiffies(bbr_probe_rtt_mode_ms);
 970                        bbr->probe_rtt_round_done = 0;
 971                        bbr->next_rtt_delivered = tp->delivered;
 972                } else if (bbr->probe_rtt_done_stamp) {
 973                        if (bbr->round_start)
 974                                bbr->probe_rtt_round_done = 1;
 975                        if (bbr->probe_rtt_round_done)
 976                                bbr_check_probe_rtt_done(sk);
 977                }
 978        }
 979        /* Restart after idle ends only once we process a new S/ACK for data */
 980        if (rs->delivered > 0)
 981                bbr->idle_restart = 0;
 982}
 983
 984static void bbr_update_gains(struct sock *sk)
 985{
 986        struct bbr *bbr = inet_csk_ca(sk);
 987
 988        switch (bbr->mode) {
 989        case BBR_STARTUP:
 990                bbr->pacing_gain = bbr_high_gain;
 991                bbr->cwnd_gain   = bbr_high_gain;
 992                break;
 993        case BBR_DRAIN:
 994                bbr->pacing_gain = bbr_drain_gain;      /* slow, to drain */
 995                bbr->cwnd_gain   = bbr_high_gain;       /* keep cwnd */
 996                break;
 997        case BBR_PROBE_BW:
 998                bbr->pacing_gain = (bbr->lt_use_bw ?
 999                                    BBR_UNIT :
1000                                    bbr_pacing_gain[bbr->cycle_idx]);
1001                bbr->cwnd_gain   = bbr_cwnd_gain;
1002                break;
1003        case BBR_PROBE_RTT:
1004                bbr->pacing_gain = BBR_UNIT;
1005                bbr->cwnd_gain   = BBR_UNIT;
1006                break;
1007        default:
1008                WARN_ONCE(1, "BBR bad mode: %u\n", bbr->mode);
1009                break;
1010        }
1011}
1012
1013static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
1014{
1015        bbr_update_bw(sk, rs);
1016        bbr_update_ack_aggregation(sk, rs);
1017        bbr_update_cycle_phase(sk, rs);
1018        bbr_check_full_bw_reached(sk, rs);
1019        bbr_check_drain(sk, rs);
1020        bbr_update_min_rtt(sk, rs);
1021        bbr_update_gains(sk);
1022}
1023
1024static void bbr_main(struct sock *sk, const struct rate_sample *rs)
1025{
1026        struct bbr *bbr = inet_csk_ca(sk);
1027        u32 bw;
1028
1029        bbr_update_model(sk, rs);
1030
1031        bw = bbr_bw(sk);
1032        bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
1033        bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
1034}
1035
1036static void bbr_init(struct sock *sk)
1037{
1038        struct tcp_sock *tp = tcp_sk(sk);
1039        struct bbr *bbr = inet_csk_ca(sk);
1040
1041        bbr->prior_cwnd = 0;
1042        tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
1043        bbr->rtt_cnt = 0;
1044        bbr->next_rtt_delivered = 0;
1045        bbr->prev_ca_state = TCP_CA_Open;
1046        bbr->packet_conservation = 0;
1047
1048        bbr->probe_rtt_done_stamp = 0;
1049        bbr->probe_rtt_round_done = 0;
1050        bbr->min_rtt_us = tcp_min_rtt(tp);
1051        bbr->min_rtt_stamp = tcp_jiffies32;
1052
1053        minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
1054
1055        bbr->has_seen_rtt = 0;
1056        bbr_init_pacing_rate_from_rtt(sk);
1057
1058        bbr->round_start = 0;
1059        bbr->idle_restart = 0;
1060        bbr->full_bw_reached = 0;
1061        bbr->full_bw = 0;
1062        bbr->full_bw_cnt = 0;
1063        bbr->cycle_mstamp = 0;
1064        bbr->cycle_idx = 0;
1065        bbr_reset_lt_bw_sampling(sk);
1066        bbr_reset_startup_mode(sk);
1067
1068        bbr->ack_epoch_mstamp = tp->tcp_mstamp;
1069        bbr->ack_epoch_acked = 0;
1070        bbr->extra_acked_win_rtts = 0;
1071        bbr->extra_acked_win_idx = 0;
1072        bbr->extra_acked[0] = 0;
1073        bbr->extra_acked[1] = 0;
1074
1075        cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
1076}
1077
1078static u32 bbr_sndbuf_expand(struct sock *sk)
1079{
1080        /* Provision 3 * cwnd since BBR may slow-start even during recovery. */
1081        return 3;
1082}
1083
1084/* In theory BBR does not need to undo the cwnd since it does not
1085 * always reduce cwnd on losses (see bbr_main()). Keep it for now.
1086 */
1087static u32 bbr_undo_cwnd(struct sock *sk)
1088{
1089        struct bbr *bbr = inet_csk_ca(sk);
1090
1091        bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
1092        bbr->full_bw_cnt = 0;
1093        bbr_reset_lt_bw_sampling(sk);
1094        return tcp_sk(sk)->snd_cwnd;
1095}
1096
1097/* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
1098static u32 bbr_ssthresh(struct sock *sk)
1099{
1100        bbr_save_cwnd(sk);
1101        return tcp_sk(sk)->snd_ssthresh;
1102}
1103
1104static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
1105                           union tcp_cc_info *info)
1106{
1107        if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
1108            ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
1109                struct tcp_sock *tp = tcp_sk(sk);
1110                struct bbr *bbr = inet_csk_ca(sk);
1111                u64 bw = bbr_bw(sk);
1112
1113                bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
1114                memset(&info->bbr, 0, sizeof(info->bbr));
1115                info->bbr.bbr_bw_lo             = (u32)bw;
1116                info->bbr.bbr_bw_hi             = (u32)(bw >> 32);
1117                info->bbr.bbr_min_rtt           = bbr->min_rtt_us;
1118                info->bbr.bbr_pacing_gain       = bbr->pacing_gain;
1119                info->bbr.bbr_cwnd_gain         = bbr->cwnd_gain;
1120                *attr = INET_DIAG_BBRINFO;
1121                return sizeof(info->bbr);
1122        }
1123        return 0;
1124}
1125
1126static void bbr_set_state(struct sock *sk, u8 new_state)
1127{
1128        struct bbr *bbr = inet_csk_ca(sk);
1129
1130        if (new_state == TCP_CA_Loss) {
1131                struct rate_sample rs = { .losses = 1 };
1132
1133                bbr->prev_ca_state = TCP_CA_Loss;
1134                bbr->full_bw = 0;
1135                bbr->round_start = 1;   /* treat RTO like end of a round */
1136                bbr_lt_bw_sampling(sk, &rs);
1137        }
1138}
1139
1140static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
1141        .flags          = TCP_CONG_NON_RESTRICTED,
1142        .name           = "bbr",
1143        .owner          = THIS_MODULE,
1144        .init           = bbr_init,
1145        .cong_control   = bbr_main,
1146        .sndbuf_expand  = bbr_sndbuf_expand,
1147        .undo_cwnd      = bbr_undo_cwnd,
1148        .cwnd_event     = bbr_cwnd_event,
1149        .ssthresh       = bbr_ssthresh,
1150        .min_tso_segs   = bbr_min_tso_segs,
1151        .get_info       = bbr_get_info,
1152        .set_state      = bbr_set_state,
1153};
1154
1155static int __init bbr_register(void)
1156{
1157        BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
1158        return tcp_register_congestion_control(&tcp_bbr_cong_ops);
1159}
1160
1161static void __exit bbr_unregister(void)
1162{
1163        tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
1164}
1165
1166module_init(bbr_register);
1167module_exit(bbr_unregister);
1168
1169MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
1170MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
1171MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
1172MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
1173MODULE_LICENSE("Dual BSD/GPL");
1174MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");
1175