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/btf.h>
  60#include <linux/btf_ids.h>
  61#include <linux/module.h>
  62#include <net/tcp.h>
  63#include <linux/inet_diag.h>
  64#include <linux/inet.h>
  65#include <linux/random.h>
  66#include <linux/win_minmax.h>
  67
  68/* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth
  69 * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps.
  70 * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32.
  71 * Since the minimum window is >=4 packets, the lower bound isn't
  72 * an issue. The upper bound isn't an issue with existing technologies.
  73 */
  74#define BW_SCALE 24
  75#define BW_UNIT (1 << BW_SCALE)
  76
  77#define BBR_SCALE 8	/* scaling factor for fractions in BBR (e.g. gains) */
  78#define BBR_UNIT (1 << BBR_SCALE)
  79
  80/* BBR has the following modes for deciding how fast to send: */
  81enum bbr_mode {
  82	BBR_STARTUP,	/* ramp up sending rate rapidly to fill pipe */
  83	BBR_DRAIN,	/* drain any queue created during startup */
  84	BBR_PROBE_BW,	/* discover, share bw: pace around estimated bw */
  85	BBR_PROBE_RTT,	/* cut inflight to min to probe min_rtt */
  86};
  87
  88/* BBR congestion control block */
  89struct bbr {
  90	u32	min_rtt_us;	        /* min RTT in min_rtt_win_sec window */
  91	u32	min_rtt_stamp;	        /* timestamp of min_rtt_us */
  92	u32	probe_rtt_done_stamp;   /* end time for BBR_PROBE_RTT mode */
  93	struct minmax bw;	/* Max recent delivery rate in pkts/uS << 24 */
  94	u32	rtt_cnt;	    /* count of packet-timed rounds elapsed */
  95	u32     next_rtt_delivered; /* scb->tx.delivered at end of round */
  96	u64	cycle_mstamp;	     /* time of this cycle phase start */
  97	u32     mode:3,		     /* current bbr_mode in state machine */
  98		prev_ca_state:3,     /* CA state on previous ACK */
  99		packet_conservation:1,  /* use packet conservation? */
 100		round_start:1,	     /* start of packet-timed tx->ack round? */
 101		idle_restart:1,	     /* restarting after idle? */
 102		probe_rtt_round_done:1,  /* a BBR_PROBE_RTT round at 4 pkts? */
 103		unused:13,
 104		lt_is_sampling:1,    /* taking long-term ("LT") samples now? */
 105		lt_rtt_cnt:7,	     /* round trips in long-term interval */
 106		lt_use_bw:1;	     /* use lt_bw as our bw estimate? */
 107	u32	lt_bw;		     /* LT est delivery rate in pkts/uS << 24 */
 108	u32	lt_last_delivered;   /* LT intvl start: tp->delivered */
 109	u32	lt_last_stamp;	     /* LT intvl start: tp->delivered_mstamp */
 110	u32	lt_last_lost;	     /* LT intvl start: tp->lost */
 111	u32	pacing_gain:10,	/* current gain for setting pacing rate */
 112		cwnd_gain:10,	/* current gain for setting cwnd */
 113		full_bw_reached:1,   /* reached full bw in Startup? */
 114		full_bw_cnt:2,	/* number of rounds without large bw gains */
 115		cycle_idx:3,	/* current index in pacing_gain cycle array */
 116		has_seen_rtt:1, /* have we seen an RTT sample yet? */
 117		unused_b:5;
 118	u32	prior_cwnd;	/* prior cwnd upon entering loss recovery */
 119	u32	full_bw;	/* recent bw, to estimate if pipe is full */
 120
 121	/* For tracking ACK aggregation: */
 122	u64	ack_epoch_mstamp;	/* start of ACK sampling epoch */
 123	u16	extra_acked[2];		/* max excess data ACKed in epoch */
 124	u32	ack_epoch_acked:20,	/* packets (S)ACKed in sampling epoch */
 125		extra_acked_win_rtts:5,	/* age of extra_acked, in round trips */
 126		extra_acked_win_idx:1,	/* current index in extra_acked array */
 127		unused_c:6;
 128};
 129
 130#define CYCLE_LEN	8	/* number of phases in a pacing gain cycle */
 131
 132/* Window length of bw filter (in rounds): */
 133static const int bbr_bw_rtts = CYCLE_LEN + 2;
 134/* Window length of min_rtt filter (in sec): */
 135static const u32 bbr_min_rtt_win_sec = 10;
 136/* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
 137static const u32 bbr_probe_rtt_mode_ms = 200;
 138/* Skip TSO below the following bandwidth (bits/sec): */
 139static const int bbr_min_tso_rate = 1200000;
 140
 141/* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck.
 142 * In order to help drive the network toward lower queues and low latency while
 143 * maintaining high utilization, the average pacing rate aims to be slightly
 144 * lower than the estimated bandwidth. This is an important aspect of the
 145 * design.
 146 */
 147static const int bbr_pacing_margin_percent = 1;
 148
 149/* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
 150 * that will allow a smoothly increasing pacing rate that will double each RTT
 151 * and send the same number of packets per RTT that an un-paced, slow-starting
 152 * Reno or CUBIC flow would:
 153 */
 154static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
 155/* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
 156 * the queue created in BBR_STARTUP in a single round:
 157 */
 158static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
 159/* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
 160static const int bbr_cwnd_gain  = BBR_UNIT * 2;
 161/* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
 162static const int bbr_pacing_gain[] = {
 163	BBR_UNIT * 5 / 4,	/* probe for more available bw */
 164	BBR_UNIT * 3 / 4,	/* drain queue and/or yield bw to other flows */
 165	BBR_UNIT, BBR_UNIT, BBR_UNIT,	/* cruise at 1.0*bw to utilize pipe, */
 166	BBR_UNIT, BBR_UNIT, BBR_UNIT	/* without creating excess queue... */
 167};
 168/* Randomize the starting gain cycling phase over N phases: */
 169static const u32 bbr_cycle_rand = 7;
 170
 171/* Try to keep at least this many packets in flight, if things go smoothly. For
 172 * smooth functioning, a sliding window protocol ACKing every other packet
 173 * needs at least 4 packets in flight:
 174 */
 175static const u32 bbr_cwnd_min_target = 4;
 176
 177/* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
 178/* If bw has increased significantly (1.25x), there may be more bw available: */
 179static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
 180/* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
 181static const u32 bbr_full_bw_cnt = 3;
 182
 183/* "long-term" ("LT") bandwidth estimator parameters... */
 184/* The minimum number of rounds in an LT bw sampling interval: */
 185static const u32 bbr_lt_intvl_min_rtts = 4;
 186/* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
 187static const u32 bbr_lt_loss_thresh = 50;
 188/* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
 189static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
 190/* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
 191static const u32 bbr_lt_bw_diff = 4000 / 8;
 192/* If we estimate we're policed, use lt_bw for this many round trips: */
 193static const u32 bbr_lt_bw_max_rtts = 48;
 194
 195/* Gain factor for adding extra_acked to target cwnd: */
 196static const int bbr_extra_acked_gain = BBR_UNIT;
 197/* Window length of extra_acked window. */
 198static const u32 bbr_extra_acked_win_rtts = 5;
 199/* Max allowed val for ack_epoch_acked, after which sampling epoch is reset */
 200static const u32 bbr_ack_epoch_acked_reset_thresh = 1U << 20;
 201/* Time period for clamping cwnd increment due to ack aggregation */
 202static const u32 bbr_extra_acked_max_us = 100 * 1000;
 203
 204static void bbr_check_probe_rtt_done(struct sock *sk);
 205
 206/* Do we estimate that STARTUP filled the pipe? */
 207static bool bbr_full_bw_reached(const struct sock *sk)
 208{
 209	const struct bbr *bbr = inet_csk_ca(sk);
 210
 211	return bbr->full_bw_reached;
 212}
 213
 214/* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
 215static u32 bbr_max_bw(const struct sock *sk)
 216{
 217	struct bbr *bbr = inet_csk_ca(sk);
 218
 219	return minmax_get(&bbr->bw);
 220}
 221
 222/* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
 223static u32 bbr_bw(const struct sock *sk)
 224{
 225	struct bbr *bbr = inet_csk_ca(sk);
 226
 227	return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
 228}
 229
 230/* Return maximum extra acked in past k-2k round trips,
 231 * where k = bbr_extra_acked_win_rtts.
 232 */
 233static u16 bbr_extra_acked(const struct sock *sk)
 234{
 235	struct bbr *bbr = inet_csk_ca(sk);
 236
 237	return max(bbr->extra_acked[0], bbr->extra_acked[1]);
 238}
 239
 240/* Return rate in bytes per second, optionally with a gain.
 241 * The order here is chosen carefully to avoid overflow of u64. This should
 242 * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
 243 */
 244static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
 245{
 246	unsigned int mss = tcp_sk(sk)->mss_cache;
 247
 248	rate *= mss;
 249	rate *= gain;
 250	rate >>= BBR_SCALE;
 251	rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent);
 252	return rate >> BW_SCALE;
 253}
 254
 255/* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
 256static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
 257{
 258	u64 rate = bw;
 259
 260	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
 261	rate = min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate));
 262	return rate;
 263}
 264
 265/* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
 266static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
 267{
 268	struct tcp_sock *tp = tcp_sk(sk);
 269	struct bbr *bbr = inet_csk_ca(sk);
 270	u64 bw;
 271	u32 rtt_us;
 272
 273	if (tp->srtt_us) {		/* any RTT sample yet? */
 274		rtt_us = max(tp->srtt_us >> 3, 1U);
 275		bbr->has_seen_rtt = 1;
 276	} else {			 /* no RTT sample yet */
 277		rtt_us = USEC_PER_MSEC;	 /* use nominal default RTT */
 278	}
 279	bw = (u64)tcp_snd_cwnd(tp) * BW_UNIT;
 280	do_div(bw, rtt_us);
 281	WRITE_ONCE(sk->sk_pacing_rate,
 282		   bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain));
 283}
 284
 285/* Pace using current bw estimate and a gain factor. */
 286static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
 287{
 288	struct tcp_sock *tp = tcp_sk(sk);
 289	struct bbr *bbr = inet_csk_ca(sk);
 290	unsigned long rate = bbr_bw_to_pacing_rate(sk, bw, gain);
 291
 292	if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
 293		bbr_init_pacing_rate_from_rtt(sk);
 294	if (bbr_full_bw_reached(sk) || rate > READ_ONCE(sk->sk_pacing_rate))
 295		WRITE_ONCE(sk->sk_pacing_rate, rate);
 296}
 297
 298/* override sysctl_tcp_min_tso_segs */
 299__bpf_kfunc static u32 bbr_min_tso_segs(struct sock *sk)
 300{
 301	return READ_ONCE(sk->sk_pacing_rate) < (bbr_min_tso_rate >> 3) ? 1 : 2;
 302}
 303
 304static u32 bbr_tso_segs_goal(struct sock *sk)
 305{
 306	struct tcp_sock *tp = tcp_sk(sk);
 307	u32 segs, bytes;
 308
 309	/* Sort of tcp_tso_autosize() but ignoring
 310	 * driver provided sk_gso_max_size.
 311	 */
 312	bytes = min_t(unsigned long,
 313		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift),
 314		      GSO_LEGACY_MAX_SIZE - 1 - MAX_TCP_HEADER);
 315	segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk));
 316
 317	return min(segs, 0x7FU);
 318}
 319
 320/* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
 321static void bbr_save_cwnd(struct sock *sk)
 322{
 323	struct tcp_sock *tp = tcp_sk(sk);
 324	struct bbr *bbr = inet_csk_ca(sk);
 325
 326	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
 327		bbr->prior_cwnd = tcp_snd_cwnd(tp);  /* this cwnd is good enough */
 328	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
 329		bbr->prior_cwnd = max(bbr->prior_cwnd, tcp_snd_cwnd(tp));
 330}
 331
 332__bpf_kfunc static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 333{
 334	struct tcp_sock *tp = tcp_sk(sk);
 335	struct bbr *bbr = inet_csk_ca(sk);
 336
 337	if (event == CA_EVENT_TX_START && tp->app_limited) {
 338		bbr->idle_restart = 1;
 339		bbr->ack_epoch_mstamp = tp->tcp_mstamp;
 340		bbr->ack_epoch_acked = 0;
 341		/* Avoid pointless buffer overflows: pace at est. bw if we don't
 342		 * need more speed (we're restarting from idle and app-limited).
 343		 */
 344		if (bbr->mode == BBR_PROBE_BW)
 345			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
 346		else if (bbr->mode == BBR_PROBE_RTT)
 347			bbr_check_probe_rtt_done(sk);
 348	}
 349}
 350
 351/* Calculate bdp based on min RTT and the estimated bottleneck bandwidth:
 352 *
 353 * bdp = ceil(bw * min_rtt * gain)
 354 *
 355 * The key factor, gain, controls the amount of queue. While a small gain
 356 * builds a smaller queue, it becomes more vulnerable to noise in RTT
 357 * measurements (e.g., delayed ACKs or other ACK compression effects). This
 358 * noise may cause BBR to under-estimate the rate.
 359 */
 360static u32 bbr_bdp(struct sock *sk, u32 bw, int gain)
 361{
 362	struct bbr *bbr = inet_csk_ca(sk);
 363	u32 bdp;
 364	u64 w;
 365
 366	/* If we've never had a valid RTT sample, cap cwnd at the initial
 367	 * default. This should only happen when the connection is not using TCP
 368	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
 369	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
 370	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
 371	 */
 372	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
 373		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
 374
 375	w = (u64)bw * bbr->min_rtt_us;
 376
 377	/* Apply a gain to the given value, remove the BW_SCALE shift, and
 378	 * round the value up to avoid a negative feedback loop.
 379	 */
 380	bdp = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
 381
 382	return bdp;
 383}
 384
 385/* To achieve full performance in high-speed paths, we budget enough cwnd to
 386 * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
 387 *   - one skb in sending host Qdisc,
 388 *   - one skb in sending host TSO/GSO engine
 389 *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
 390 * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
 391 * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
 392 * which allows 2 outstanding 2-packet sequences, to try to keep pipe
 393 * full even with ACK-every-other-packet delayed ACKs.
 394 */
 395static u32 bbr_quantization_budget(struct sock *sk, u32 cwnd)
 396{
 397	struct bbr *bbr = inet_csk_ca(sk);
 398
 399	/* Allow enough full-sized skbs in flight to utilize end systems. */
 400	cwnd += 3 * bbr_tso_segs_goal(sk);
 401
 402	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
 403	cwnd = (cwnd + 1) & ~1U;
 404
 405	/* Ensure gain cycling gets inflight above BDP even for small BDPs. */
 406	if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == 0)
 407		cwnd += 2;
 408
 409	return cwnd;
 410}
 411
 412/* Find inflight based on min RTT and the estimated bottleneck bandwidth. */
 413static u32 bbr_inflight(struct sock *sk, u32 bw, int gain)
 414{
 415	u32 inflight;
 416
 417	inflight = bbr_bdp(sk, bw, gain);
 418	inflight = bbr_quantization_budget(sk, inflight);
 419
 420	return inflight;
 421}
 422
 423/* With pacing at lower layers, there's often less data "in the network" than
 424 * "in flight". With TSQ and departure time pacing at lower layers (e.g. fq),
 425 * we often have several skbs queued in the pacing layer with a pre-scheduled
 426 * earliest departure time (EDT). BBR adapts its pacing rate based on the
 427 * inflight level that it estimates has already been "baked in" by previous
 428 * departure time decisions. We calculate a rough estimate of the number of our
 429 * packets that might be in the network at the earliest departure time for the
 430 * next skb scheduled:
 431 *   in_network_at_edt = inflight_at_edt - (EDT - now) * bw
 432 * If we're increasing inflight, then we want to know if the transmit of the
 433 * EDT skb will push inflight above the target, so inflight_at_edt includes
 434 * bbr_tso_segs_goal() from the skb departing at EDT. If decreasing inflight,
 435 * then estimate if inflight will sink too low just before the EDT transmit.
 436 */
 437static u32 bbr_packets_in_net_at_edt(struct sock *sk, u32 inflight_now)
 438{
 439	struct tcp_sock *tp = tcp_sk(sk);
 440	struct bbr *bbr = inet_csk_ca(sk);
 441	u64 now_ns, edt_ns, interval_us;
 442	u32 interval_delivered, inflight_at_edt;
 443
 444	now_ns = tp->tcp_clock_cache;
 445	edt_ns = max(tp->tcp_wstamp_ns, now_ns);
 446	interval_us = div_u64(edt_ns - now_ns, NSEC_PER_USEC);
 447	interval_delivered = (u64)bbr_bw(sk) * interval_us >> BW_SCALE;
 448	inflight_at_edt = inflight_now;
 449	if (bbr->pacing_gain > BBR_UNIT)              /* increasing inflight */
 450		inflight_at_edt += bbr_tso_segs_goal(sk);  /* include EDT skb */
 451	if (interval_delivered >= inflight_at_edt)
 452		return 0;
 453	return inflight_at_edt - interval_delivered;
 454}
 455
 456/* Find the cwnd increment based on estimate of ack aggregation */
 457static u32 bbr_ack_aggregation_cwnd(struct sock *sk)
 458{
 459	u32 max_aggr_cwnd, aggr_cwnd = 0;
 460
 461	if (bbr_extra_acked_gain && bbr_full_bw_reached(sk)) {
 462		max_aggr_cwnd = ((u64)bbr_bw(sk) * bbr_extra_acked_max_us)
 463				/ BW_UNIT;
 464		aggr_cwnd = (bbr_extra_acked_gain * bbr_extra_acked(sk))
 465			     >> BBR_SCALE;
 466		aggr_cwnd = min(aggr_cwnd, max_aggr_cwnd);
 467	}
 468
 469	return aggr_cwnd;
 470}
 471
 472/* An optimization in BBR to reduce losses: On the first round of recovery, we
 473 * follow the packet conservation principle: send P packets per P packets acked.
 474 * After that, we slow-start and send at most 2*P packets per P packets acked.
 475 * After recovery finishes, or upon undo, we restore the cwnd we had when
 476 * recovery started (capped by the target cwnd based on estimated BDP).
 477 *
 478 * TODO(ycheng/ncardwell): implement a rate-based approach.
 479 */
 480static bool bbr_set_cwnd_to_recover_or_restore(
 481	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
 482{
 483	struct tcp_sock *tp = tcp_sk(sk);
 484	struct bbr *bbr = inet_csk_ca(sk);
 485	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
 486	u32 cwnd = tcp_snd_cwnd(tp);
 487
 488	/* An ACK for P pkts should release at most 2*P packets. We do this
 489	 * in two steps. First, here we deduct the number of lost packets.
 490	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
 491	 */
 492	if (rs->losses > 0)
 493		cwnd = max_t(s32, cwnd - rs->losses, 1);
 494
 495	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
 496		/* Starting 1st round of Recovery, so do packet conservation. */
 497		bbr->packet_conservation = 1;
 498		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
 499		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
 500		cwnd = tcp_packets_in_flight(tp) + acked;
 501	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
 502		/* Exiting loss recovery; restore cwnd saved before recovery. */
 503		cwnd = max(cwnd, bbr->prior_cwnd);
 504		bbr->packet_conservation = 0;
 505	}
 506	bbr->prev_ca_state = state;
 507
 508	if (bbr->packet_conservation) {
 509		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
 510		return true;	/* yes, using packet conservation */
 511	}
 512	*new_cwnd = cwnd;
 513	return false;
 514}
 515
 516/* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
 517 * has drawn us down below target), or snap down to target if we're above it.
 518 */
 519static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
 520			 u32 acked, u32 bw, int gain)
 521{
 522	struct tcp_sock *tp = tcp_sk(sk);
 523	struct bbr *bbr = inet_csk_ca(sk);
 524	u32 cwnd = tcp_snd_cwnd(tp), target_cwnd = 0;
 525
 526	if (!acked)
 527		goto done;  /* no packet fully ACKed; just apply caps */
 528
 529	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
 530		goto done;
 531
 532	target_cwnd = bbr_bdp(sk, bw, gain);
 533
 534	/* Increment the cwnd to account for excess ACKed data that seems
 535	 * due to aggregation (of data and/or ACKs) visible in the ACK stream.
 536	 */
 537	target_cwnd += bbr_ack_aggregation_cwnd(sk);
 538	target_cwnd = bbr_quantization_budget(sk, target_cwnd);
 539
 540	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
 541	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
 542		cwnd = min(cwnd + acked, target_cwnd);
 543	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
 544		cwnd = cwnd + acked;
 545	cwnd = max(cwnd, bbr_cwnd_min_target);
 546
 547done:
 548	tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));	/* apply global cap */
 549	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
 550		tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), bbr_cwnd_min_target));
 551}
 552
 553/* End cycle phase if it's time and/or we hit the phase's in-flight target. */
 554static bool bbr_is_next_cycle_phase(struct sock *sk,
 555				    const struct rate_sample *rs)
 556{
 557	struct tcp_sock *tp = tcp_sk(sk);
 558	struct bbr *bbr = inet_csk_ca(sk);
 559	bool is_full_length =
 560		tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
 561		bbr->min_rtt_us;
 562	u32 inflight, bw;
 563
 564	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
 565	 * use the pipe without increasing the queue.
 566	 */
 567	if (bbr->pacing_gain == BBR_UNIT)
 568		return is_full_length;		/* just use wall clock time */
 569
 570	inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight);
 571	bw = bbr_max_bw(sk);
 572
 573	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
 574	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
 575	 * small (e.g. on a LAN). We do not persist if packets are lost, since
 576	 * a path with small buffers may not hold that much.
 577	 */
 578	if (bbr->pacing_gain > BBR_UNIT)
 579		return is_full_length &&
 580			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
 581			 inflight >= bbr_inflight(sk, bw, bbr->pacing_gain));
 582
 583	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
 584	 * probing didn't find more bw. If inflight falls to match BDP then we
 585	 * estimate queue is drained; persisting would underutilize the pipe.
 586	 */
 587	return is_full_length ||
 588		inflight <= bbr_inflight(sk, bw, BBR_UNIT);
 589}
 590
 591static void bbr_advance_cycle_phase(struct sock *sk)
 592{
 593	struct tcp_sock *tp = tcp_sk(sk);
 594	struct bbr *bbr = inet_csk_ca(sk);
 595
 596	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
 597	bbr->cycle_mstamp = tp->delivered_mstamp;
 598}
 599
 600/* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
 601static void bbr_update_cycle_phase(struct sock *sk,
 602				   const struct rate_sample *rs)
 603{
 604	struct bbr *bbr = inet_csk_ca(sk);
 605
 606	if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
 607		bbr_advance_cycle_phase(sk);
 608}
 609
 610static void bbr_reset_startup_mode(struct sock *sk)
 611{
 612	struct bbr *bbr = inet_csk_ca(sk);
 613
 614	bbr->mode = BBR_STARTUP;
 615}
 616
 617static void bbr_reset_probe_bw_mode(struct sock *sk)
 618{
 619	struct bbr *bbr = inet_csk_ca(sk);
 620
 621	bbr->mode = BBR_PROBE_BW;
 622	bbr->cycle_idx = CYCLE_LEN - 1 - get_random_u32_below(bbr_cycle_rand);
 623	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
 624}
 625
 626static void bbr_reset_mode(struct sock *sk)
 627{
 628	if (!bbr_full_bw_reached(sk))
 629		bbr_reset_startup_mode(sk);
 630	else
 631		bbr_reset_probe_bw_mode(sk);
 632}
 633
 634/* Start a new long-term sampling interval. */
 635static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
 636{
 637	struct tcp_sock *tp = tcp_sk(sk);
 638	struct bbr *bbr = inet_csk_ca(sk);
 639
 640	bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
 641	bbr->lt_last_delivered = tp->delivered;
 642	bbr->lt_last_lost = tp->lost;
 643	bbr->lt_rtt_cnt = 0;
 644}
 645
 646/* Completely reset long-term bandwidth sampling. */
 647static void bbr_reset_lt_bw_sampling(struct sock *sk)
 648{
 649	struct bbr *bbr = inet_csk_ca(sk);
 650
 651	bbr->lt_bw = 0;
 652	bbr->lt_use_bw = 0;
 653	bbr->lt_is_sampling = false;
 654	bbr_reset_lt_bw_sampling_interval(sk);
 655}
 656
 657/* Long-term bw sampling interval is done. Estimate whether we're policed. */
 658static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
 659{
 660	struct bbr *bbr = inet_csk_ca(sk);
 661	u32 diff;
 662
 663	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
 664		/* Is new bw close to the lt_bw from the previous interval? */
 665		diff = abs(bw - bbr->lt_bw);
 666		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
 667		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
 668		     bbr_lt_bw_diff)) {
 669			/* All criteria are met; estimate we're policed. */
 670			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
 671			bbr->lt_use_bw = 1;
 672			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
 673			bbr->lt_rtt_cnt = 0;
 674			return;
 675		}
 676	}
 677	bbr->lt_bw = bw;
 678	bbr_reset_lt_bw_sampling_interval(sk);
 679}
 680
 681/* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
 682 * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
 683 * explicitly models their policed rate, to reduce unnecessary losses. We
 684 * estimate that we're policed if we see 2 consecutive sampling intervals with
 685 * consistent throughput and high packet loss. If we think we're being policed,
 686 * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
 687 */
 688static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
 689{
 690	struct tcp_sock *tp = tcp_sk(sk);
 691	struct bbr *bbr = inet_csk_ca(sk);
 692	u32 lost, delivered;
 693	u64 bw;
 694	u32 t;
 695
 696	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
 697		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
 698		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
 699			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
 700			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
 701		}
 702		return;
 703	}
 704
 705	/* Wait for the first loss before sampling, to let the policer exhaust
 706	 * its tokens and estimate the steady-state rate allowed by the policer.
 707	 * Starting samples earlier includes bursts that over-estimate the bw.
 708	 */
 709	if (!bbr->lt_is_sampling) {
 710		if (!rs->losses)
 711			return;
 712		bbr_reset_lt_bw_sampling_interval(sk);
 713		bbr->lt_is_sampling = true;
 714	}
 715
 716	/* To avoid underestimates, reset sampling if we run out of data. */
 717	if (rs->is_app_limited) {
 718		bbr_reset_lt_bw_sampling(sk);
 719		return;
 720	}
 721
 722	if (bbr->round_start)
 723		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
 724	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
 725		return;		/* sampling interval needs to be longer */
 726	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
 727		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
 728		return;
 729	}
 730
 731	/* End sampling interval when a packet is lost, so we estimate the
 732	 * policer tokens were exhausted. Stopping the sampling before the
 733	 * tokens are exhausted under-estimates the policed rate.
 734	 */
 735	if (!rs->losses)
 736		return;
 737
 738	/* Calculate packets lost and delivered in sampling interval. */
 739	lost = tp->lost - bbr->lt_last_lost;
 740	delivered = tp->delivered - bbr->lt_last_delivered;
 741	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
 742	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
 743		return;
 744
 745	/* Find average delivery rate in this sampling interval. */
 746	t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
 747	if ((s32)t < 1)
 748		return;		/* interval is less than one ms, so wait */
 749	/* Check if can multiply without overflow */
 750	if (t >= ~0U / USEC_PER_MSEC) {
 751		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
 752		return;
 753	}
 754	t *= USEC_PER_MSEC;
 755	bw = (u64)delivered * BW_UNIT;
 756	do_div(bw, t);
 757	bbr_lt_bw_interval_done(sk, bw);
 758}
 759
 760/* Estimate the bandwidth based on how fast packets are delivered */
 761static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
 762{
 763	struct tcp_sock *tp = tcp_sk(sk);
 764	struct bbr *bbr = inet_csk_ca(sk);
 765	u64 bw;
 766
 767	bbr->round_start = 0;
 768	if (rs->delivered < 0 || rs->interval_us <= 0)
 769		return; /* Not a valid observation */
 770
 771	/* See if we've reached the next RTT */
 772	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
 773		bbr->next_rtt_delivered = tp->delivered;
 774		bbr->rtt_cnt++;
 775		bbr->round_start = 1;
 776		bbr->packet_conservation = 0;
 777	}
 778
 779	bbr_lt_bw_sampling(sk, rs);
 780
 781	/* Divide delivered by the interval to find a (lower bound) bottleneck
 782	 * bandwidth sample. Delivered is in packets and interval_us in uS and
 783	 * ratio will be <<1 for most connections. So delivered is first scaled.
 784	 */
 785	bw = div64_long((u64)rs->delivered * BW_UNIT, rs->interval_us);
 786
 787	/* If this sample is application-limited, it is likely to have a very
 788	 * low delivered count that represents application behavior rather than
 789	 * the available network rate. Such a sample could drag down estimated
 790	 * bw, causing needless slow-down. Thus, to continue to send at the
 791	 * last measured network rate, we filter out app-limited samples unless
 792	 * they describe the path bw at least as well as our bw model.
 793	 *
 794	 * So the goal during app-limited phase is to proceed with the best
 795	 * network rate no matter how long. We automatically leave this
 796	 * phase when app writes faster than the network can deliver :)
 797	 */
 798	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
 799		/* Incorporate new sample into our max bw filter. */
 800		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
 801	}
 802}
 803
 804/* Estimates the windowed max degree of ack aggregation.
 805 * This is used to provision extra in-flight data to keep sending during
 806 * inter-ACK silences.
 807 *
 808 * Degree of ack aggregation is estimated as extra data acked beyond expected.
 809 *
 810 * max_extra_acked = "maximum recent excess data ACKed beyond max_bw * interval"
 811 * cwnd += max_extra_acked
 812 *
 813 * Max extra_acked is clamped by cwnd and bw * bbr_extra_acked_max_us (100 ms).
 814 * Max filter is an approximate sliding window of 5-10 (packet timed) round
 815 * trips.
 816 */
 817static void bbr_update_ack_aggregation(struct sock *sk,
 818				       const struct rate_sample *rs)
 819{
 820	u32 epoch_us, expected_acked, extra_acked;
 821	struct bbr *bbr = inet_csk_ca(sk);
 822	struct tcp_sock *tp = tcp_sk(sk);
 823
 824	if (!bbr_extra_acked_gain || rs->acked_sacked <= 0 ||
 825	    rs->delivered < 0 || rs->interval_us <= 0)
 826		return;
 827
 828	if (bbr->round_start) {
 829		bbr->extra_acked_win_rtts = min(0x1F,
 830						bbr->extra_acked_win_rtts + 1);
 831		if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
 832			bbr->extra_acked_win_rtts = 0;
 833			bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
 834						   0 : 1;
 835			bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
 836		}
 837	}
 838
 839	/* Compute how many packets we expected to be delivered over epoch. */
 840	epoch_us = tcp_stamp_us_delta(tp->delivered_mstamp,
 841				      bbr->ack_epoch_mstamp);
 842	expected_acked = ((u64)bbr_bw(sk) * epoch_us) / BW_UNIT;
 843
 844	/* Reset the aggregation epoch if ACK rate is below expected rate or
 845	 * significantly large no. of ack received since epoch (potentially
 846	 * quite old epoch).
 847	 */
 848	if (bbr->ack_epoch_acked <= expected_acked ||
 849	    (bbr->ack_epoch_acked + rs->acked_sacked >=
 850	     bbr_ack_epoch_acked_reset_thresh)) {
 851		bbr->ack_epoch_acked = 0;
 852		bbr->ack_epoch_mstamp = tp->delivered_mstamp;
 853		expected_acked = 0;
 854	}
 855
 856	/* Compute excess data delivered, beyond what was expected. */
 857	bbr->ack_epoch_acked = min_t(u32, 0xFFFFF,
 858				     bbr->ack_epoch_acked + rs->acked_sacked);
 859	extra_acked = bbr->ack_epoch_acked - expected_acked;
 860	extra_acked = min(extra_acked, tcp_snd_cwnd(tp));
 861	if (extra_acked > bbr->extra_acked[bbr->extra_acked_win_idx])
 862		bbr->extra_acked[bbr->extra_acked_win_idx] = extra_acked;
 863}
 864
 865/* Estimate when the pipe is full, using the change in delivery rate: BBR
 866 * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
 867 * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
 868 * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
 869 * higher rwin, 3: we get higher delivery rate samples. Or transient
 870 * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
 871 * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
 872 */
 873static void bbr_check_full_bw_reached(struct sock *sk,
 874				      const struct rate_sample *rs)
 875{
 876	struct bbr *bbr = inet_csk_ca(sk);
 877	u32 bw_thresh;
 878
 879	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
 880		return;
 881
 882	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
 883	if (bbr_max_bw(sk) >= bw_thresh) {
 884		bbr->full_bw = bbr_max_bw(sk);
 885		bbr->full_bw_cnt = 0;
 886		return;
 887	}
 888	++bbr->full_bw_cnt;
 889	bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
 890}
 891
 892/* If pipe is probably full, drain the queue and then enter steady-state. */
 893static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
 894{
 895	struct bbr *bbr = inet_csk_ca(sk);
 896
 897	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
 898		bbr->mode = BBR_DRAIN;	/* drain queue we created */
 899		tcp_sk(sk)->snd_ssthresh =
 900				bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT);
 901	}	/* fall through to check if in-flight is already small: */
 902	if (bbr->mode == BBR_DRAIN &&
 903	    bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <=
 904	    bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT))
 905		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
 906}
 907
 908static void bbr_check_probe_rtt_done(struct sock *sk)
 909{
 910	struct tcp_sock *tp = tcp_sk(sk);
 911	struct bbr *bbr = inet_csk_ca(sk);
 912
 913	if (!(bbr->probe_rtt_done_stamp &&
 914	      after(tcp_jiffies32, bbr->probe_rtt_done_stamp)))
 915		return;
 916
 917	bbr->min_rtt_stamp = tcp_jiffies32;  /* wait a while until PROBE_RTT */
 918	tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp), bbr->prior_cwnd));
 919	bbr_reset_mode(sk);
 920}
 921
 922/* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
 923 * periodically drain the bottleneck queue, to converge to measure the true
 924 * min_rtt (unloaded propagation delay). This allows the flows to keep queues
 925 * small (reducing queuing delay and packet loss) and achieve fairness among
 926 * BBR flows.
 927 *
 928 * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
 929 * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
 930 * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
 931 * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
 932 * re-enter the previous mode. BBR uses 200ms to approximately bound the
 933 * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
 934 *
 935 * Note that flows need only pay 2% if they are busy sending over the last 10
 936 * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
 937 * natural silences or low-rate periods within 10 seconds where the rate is low
 938 * enough for long enough to drain its queue in the bottleneck. We pick up
 939 * these min RTT measurements opportunistically with our min_rtt filter. :-)
 940 */
 941static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
 942{
 943	struct tcp_sock *tp = tcp_sk(sk);
 944	struct bbr *bbr = inet_csk_ca(sk);
 945	bool filter_expired;
 946
 947	/* Track min RTT seen in the min_rtt_win_sec filter window: */
 948	filter_expired = after(tcp_jiffies32,
 949			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
 950	if (rs->rtt_us >= 0 &&
 951	    (rs->rtt_us < bbr->min_rtt_us ||
 952	     (filter_expired && !rs->is_ack_delayed))) {
 953		bbr->min_rtt_us = rs->rtt_us;
 954		bbr->min_rtt_stamp = tcp_jiffies32;
 955	}
 956
 957	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
 958	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
 959		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
 960		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
 961		bbr->probe_rtt_done_stamp = 0;
 962	}
 963
 964	if (bbr->mode == BBR_PROBE_RTT) {
 965		/* Ignore low rate samples during this mode. */
 966		tp->app_limited =
 967			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
 968		/* Maintain min packets in flight for max(200 ms, 1 round). */
 969		if (!bbr->probe_rtt_done_stamp &&
 970		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
 971			bbr->probe_rtt_done_stamp = tcp_jiffies32 +
 972				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
 973			bbr->probe_rtt_round_done = 0;
 974			bbr->next_rtt_delivered = tp->delivered;
 975		} else if (bbr->probe_rtt_done_stamp) {
 976			if (bbr->round_start)
 977				bbr->probe_rtt_round_done = 1;
 978			if (bbr->probe_rtt_round_done)
 979				bbr_check_probe_rtt_done(sk);
 980		}
 981	}
 982	/* Restart after idle ends only once we process a new S/ACK for data */
 983	if (rs->delivered > 0)
 984		bbr->idle_restart = 0;
 985}
 986
 987static void bbr_update_gains(struct sock *sk)
 988{
 989	struct bbr *bbr = inet_csk_ca(sk);
 990
 991	switch (bbr->mode) {
 992	case BBR_STARTUP:
 993		bbr->pacing_gain = bbr_high_gain;
 994		bbr->cwnd_gain	 = bbr_high_gain;
 995		break;
 996	case BBR_DRAIN:
 997		bbr->pacing_gain = bbr_drain_gain;	/* slow, to drain */
 998		bbr->cwnd_gain	 = bbr_high_gain;	/* keep cwnd */
 999		break;
1000	case BBR_PROBE_BW:
1001		bbr->pacing_gain = (bbr->lt_use_bw ?
1002				    BBR_UNIT :
1003				    bbr_pacing_gain[bbr->cycle_idx]);
1004		bbr->cwnd_gain	 = bbr_cwnd_gain;
1005		break;
1006	case BBR_PROBE_RTT:
1007		bbr->pacing_gain = BBR_UNIT;
1008		bbr->cwnd_gain	 = BBR_UNIT;
1009		break;
1010	default:
1011		WARN_ONCE(1, "BBR bad mode: %u\n", bbr->mode);
1012		break;
1013	}
1014}
1015
1016static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
1017{
1018	bbr_update_bw(sk, rs);
1019	bbr_update_ack_aggregation(sk, rs);
1020	bbr_update_cycle_phase(sk, rs);
1021	bbr_check_full_bw_reached(sk, rs);
1022	bbr_check_drain(sk, rs);
1023	bbr_update_min_rtt(sk, rs);
1024	bbr_update_gains(sk);
1025}
1026
1027__bpf_kfunc static void bbr_main(struct sock *sk, u32 ack, int flag, const struct rate_sample *rs)
1028{
1029	struct bbr *bbr = inet_csk_ca(sk);
1030	u32 bw;
1031
1032	bbr_update_model(sk, rs);
1033
1034	bw = bbr_bw(sk);
1035	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
1036	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
1037}
1038
1039__bpf_kfunc static void bbr_init(struct sock *sk)
1040{
1041	struct tcp_sock *tp = tcp_sk(sk);
1042	struct bbr *bbr = inet_csk_ca(sk);
1043
1044	bbr->prior_cwnd = 0;
1045	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
1046	bbr->rtt_cnt = 0;
1047	bbr->next_rtt_delivered = tp->delivered;
1048	bbr->prev_ca_state = TCP_CA_Open;
1049	bbr->packet_conservation = 0;
1050
1051	bbr->probe_rtt_done_stamp = 0;
1052	bbr->probe_rtt_round_done = 0;
1053	bbr->min_rtt_us = tcp_min_rtt(tp);
1054	bbr->min_rtt_stamp = tcp_jiffies32;
1055
1056	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
1057
1058	bbr->has_seen_rtt = 0;
1059	bbr_init_pacing_rate_from_rtt(sk);
1060
1061	bbr->round_start = 0;
1062	bbr->idle_restart = 0;
1063	bbr->full_bw_reached = 0;
1064	bbr->full_bw = 0;
1065	bbr->full_bw_cnt = 0;
1066	bbr->cycle_mstamp = 0;
1067	bbr->cycle_idx = 0;
1068	bbr_reset_lt_bw_sampling(sk);
1069	bbr_reset_startup_mode(sk);
1070
1071	bbr->ack_epoch_mstamp = tp->tcp_mstamp;
1072	bbr->ack_epoch_acked = 0;
1073	bbr->extra_acked_win_rtts = 0;
1074	bbr->extra_acked_win_idx = 0;
1075	bbr->extra_acked[0] = 0;
1076	bbr->extra_acked[1] = 0;
1077
1078	cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
1079}
1080
1081__bpf_kfunc static u32 bbr_sndbuf_expand(struct sock *sk)
1082{
1083	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
1084	return 3;
1085}
1086
1087/* In theory BBR does not need to undo the cwnd since it does not
1088 * always reduce cwnd on losses (see bbr_main()). Keep it for now.
1089 */
1090__bpf_kfunc static u32 bbr_undo_cwnd(struct sock *sk)
1091{
1092	struct bbr *bbr = inet_csk_ca(sk);
1093
1094	bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
1095	bbr->full_bw_cnt = 0;
1096	bbr_reset_lt_bw_sampling(sk);
1097	return tcp_snd_cwnd(tcp_sk(sk));
1098}
1099
1100/* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
1101__bpf_kfunc static u32 bbr_ssthresh(struct sock *sk)
1102{
1103	bbr_save_cwnd(sk);
1104	return tcp_sk(sk)->snd_ssthresh;
1105}
1106
1107static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
1108			   union tcp_cc_info *info)
1109{
1110	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
1111	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
1112		struct tcp_sock *tp = tcp_sk(sk);
1113		struct bbr *bbr = inet_csk_ca(sk);
1114		u64 bw = bbr_bw(sk);
1115
1116		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
1117		memset(&info->bbr, 0, sizeof(info->bbr));
1118		info->bbr.bbr_bw_lo		= (u32)bw;
1119		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
1120		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
1121		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
1122		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
1123		*attr = INET_DIAG_BBRINFO;
1124		return sizeof(info->bbr);
1125	}
1126	return 0;
1127}
1128
1129__bpf_kfunc static void bbr_set_state(struct sock *sk, u8 new_state)
1130{
1131	struct bbr *bbr = inet_csk_ca(sk);
1132
1133	if (new_state == TCP_CA_Loss) {
1134		struct rate_sample rs = { .losses = 1 };
1135
1136		bbr->prev_ca_state = TCP_CA_Loss;
1137		bbr->full_bw = 0;
1138		bbr->round_start = 1;	/* treat RTO like end of a round */
1139		bbr_lt_bw_sampling(sk, &rs);
1140	}
1141}
1142
1143static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
1144	.flags		= TCP_CONG_NON_RESTRICTED,
1145	.name		= "bbr",
1146	.owner		= THIS_MODULE,
1147	.init		= bbr_init,
1148	.cong_control	= bbr_main,
1149	.sndbuf_expand	= bbr_sndbuf_expand,
1150	.undo_cwnd	= bbr_undo_cwnd,
1151	.cwnd_event	= bbr_cwnd_event,
1152	.ssthresh	= bbr_ssthresh,
1153	.min_tso_segs	= bbr_min_tso_segs,
1154	.get_info	= bbr_get_info,
1155	.set_state	= bbr_set_state,
1156};
1157
1158BTF_KFUNCS_START(tcp_bbr_check_kfunc_ids)
 
 
1159BTF_ID_FLAGS(func, bbr_init)
1160BTF_ID_FLAGS(func, bbr_main)
1161BTF_ID_FLAGS(func, bbr_sndbuf_expand)
1162BTF_ID_FLAGS(func, bbr_undo_cwnd)
1163BTF_ID_FLAGS(func, bbr_cwnd_event)
1164BTF_ID_FLAGS(func, bbr_ssthresh)
1165BTF_ID_FLAGS(func, bbr_min_tso_segs)
1166BTF_ID_FLAGS(func, bbr_set_state)
1167BTF_KFUNCS_END(tcp_bbr_check_kfunc_ids)
 
 
1168
1169static const struct btf_kfunc_id_set tcp_bbr_kfunc_set = {
1170	.owner = THIS_MODULE,
1171	.set   = &tcp_bbr_check_kfunc_ids,
1172};
1173
1174static int __init bbr_register(void)
1175{
1176	int ret;
1177
1178	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
1179
1180	ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_bbr_kfunc_set);
1181	if (ret < 0)
1182		return ret;
1183	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
1184}
1185
1186static void __exit bbr_unregister(void)
1187{
1188	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
1189}
1190
1191module_init(bbr_register);
1192module_exit(bbr_unregister);
1193
1194MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
1195MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
1196MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
1197MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
1198MODULE_LICENSE("Dual BSD/GPL");
1199MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");