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