blob: 0ea66c2c93444198fa21629dac617160f78a5cee [file] [log] [blame]
Neal Cardwell0f8782e2016-09-19 23:39:23 -04001/* 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 * BBR is described in detail in:
18 * "BBR: Congestion-Based Congestion Control",
19 * Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh,
20 * Van Jacobson. ACM Queue, Vol. 14 No. 5, September-October 2016.
21 *
22 * There is a public e-mail list for discussing BBR development and testing:
23 * https://groups.google.com/forum/#!forum/bbr-dev
24 *
25 * NOTE: BBR *must* be used with the fq qdisc ("man tc-fq") with pacing enabled,
26 * since pacing is integral to the BBR design and implementation.
27 * BBR without pacing would not function properly, and may incur unnecessary
28 * high packet loss rates.
29 */
30#include <linux/module.h>
31#include <net/tcp.h>
32#include <linux/inet_diag.h>
33#include <linux/inet.h>
34#include <linux/random.h>
35#include <linux/win_minmax.h>
36
37/* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth
38 * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps.
39 * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32.
40 * Since the minimum window is >=4 packets, the lower bound isn't
41 * an issue. The upper bound isn't an issue with existing technologies.
42 */
43#define BW_SCALE 24
44#define BW_UNIT (1 << BW_SCALE)
45
46#define BBR_SCALE 8 /* scaling factor for fractions in BBR (e.g. gains) */
47#define BBR_UNIT (1 << BBR_SCALE)
48
49/* BBR has the following modes for deciding how fast to send: */
50enum bbr_mode {
51 BBR_STARTUP, /* ramp up sending rate rapidly to fill pipe */
52 BBR_DRAIN, /* drain any queue created during startup */
53 BBR_PROBE_BW, /* discover, share bw: pace around estimated bw */
54 BBR_PROBE_RTT, /* cut cwnd to min to probe min_rtt */
55};
56
57/* BBR congestion control block */
58struct bbr {
59 u32 min_rtt_us; /* min RTT in min_rtt_win_sec window */
60 u32 min_rtt_stamp; /* timestamp of min_rtt_us */
61 u32 probe_rtt_done_stamp; /* end time for BBR_PROBE_RTT mode */
62 struct minmax bw; /* Max recent delivery rate in pkts/uS << 24 */
63 u32 rtt_cnt; /* count of packet-timed rounds elapsed */
64 u32 next_rtt_delivered; /* scb->tx.delivered at end of round */
65 struct skb_mstamp cycle_mstamp; /* time of this cycle phase start */
66 u32 mode:3, /* current bbr_mode in state machine */
67 prev_ca_state:3, /* CA state on previous ACK */
68 packet_conservation:1, /* use packet conservation? */
69 restore_cwnd:1, /* decided to revert cwnd to old value */
70 round_start:1, /* start of packet-timed tx->ack round? */
71 tso_segs_goal:7, /* segments we want in each skb we send */
72 idle_restart:1, /* restarting after idle? */
73 probe_rtt_round_done:1, /* a BBR_PROBE_RTT round at 4 pkts? */
74 unused:5,
75 lt_is_sampling:1, /* taking long-term ("LT") samples now? */
76 lt_rtt_cnt:7, /* round trips in long-term interval */
77 lt_use_bw:1; /* use lt_bw as our bw estimate? */
78 u32 lt_bw; /* LT est delivery rate in pkts/uS << 24 */
79 u32 lt_last_delivered; /* LT intvl start: tp->delivered */
80 u32 lt_last_stamp; /* LT intvl start: tp->delivered_mstamp */
81 u32 lt_last_lost; /* LT intvl start: tp->lost */
82 u32 pacing_gain:10, /* current gain for setting pacing rate */
83 cwnd_gain:10, /* current gain for setting cwnd */
84 full_bw_cnt:3, /* number of rounds without large bw gains */
85 cycle_idx:3, /* current index in pacing_gain cycle array */
86 unused_b:6;
87 u32 prior_cwnd; /* prior cwnd upon entering loss recovery */
88 u32 full_bw; /* recent bw, to estimate if pipe is full */
89};
90
91#define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */
92
93/* Window length of bw filter (in rounds): */
94static const int bbr_bw_rtts = CYCLE_LEN + 2;
95/* Window length of min_rtt filter (in sec): */
96static const u32 bbr_min_rtt_win_sec = 10;
97/* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
98static const u32 bbr_probe_rtt_mode_ms = 200;
99/* Skip TSO below the following bandwidth (bits/sec): */
100static const int bbr_min_tso_rate = 1200000;
101
102/* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
103 * that will allow a smoothly increasing pacing rate that will double each RTT
104 * and send the same number of packets per RTT that an un-paced, slow-starting
105 * Reno or CUBIC flow would:
106 */
107static const int bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
108/* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
109 * the queue created in BBR_STARTUP in a single round:
110 */
111static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
112/* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
113static const int bbr_cwnd_gain = BBR_UNIT * 2;
114/* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
115static const int bbr_pacing_gain[] = {
116 BBR_UNIT * 5 / 4, /* probe for more available bw */
117 BBR_UNIT * 3 / 4, /* drain queue and/or yield bw to other flows */
118 BBR_UNIT, BBR_UNIT, BBR_UNIT, /* cruise at 1.0*bw to utilize pipe, */
119 BBR_UNIT, BBR_UNIT, BBR_UNIT /* without creating excess queue... */
120};
121/* Randomize the starting gain cycling phase over N phases: */
122static const u32 bbr_cycle_rand = 7;
123
124/* Try to keep at least this many packets in flight, if things go smoothly. For
125 * smooth functioning, a sliding window protocol ACKing every other packet
126 * needs at least 4 packets in flight:
127 */
128static const u32 bbr_cwnd_min_target = 4;
129
130/* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
131/* If bw has increased significantly (1.25x), there may be more bw available: */
132static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
133/* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
134static const u32 bbr_full_bw_cnt = 3;
135
136/* "long-term" ("LT") bandwidth estimator parameters... */
137/* The minimum number of rounds in an LT bw sampling interval: */
138static const u32 bbr_lt_intvl_min_rtts = 4;
139/* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
140static const u32 bbr_lt_loss_thresh = 50;
141/* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
142static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
143/* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
144static const u32 bbr_lt_bw_diff = 4000 / 8;
145/* If we estimate we're policed, use lt_bw for this many round trips: */
146static const u32 bbr_lt_bw_max_rtts = 48;
147
148/* Do we estimate that STARTUP filled the pipe? */
149static bool bbr_full_bw_reached(const struct sock *sk)
150{
151 const struct bbr *bbr = inet_csk_ca(sk);
152
153 return bbr->full_bw_cnt >= bbr_full_bw_cnt;
154}
155
156/* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
157static u32 bbr_max_bw(const struct sock *sk)
158{
159 struct bbr *bbr = inet_csk_ca(sk);
160
161 return minmax_get(&bbr->bw);
162}
163
164/* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
165static u32 bbr_bw(const struct sock *sk)
166{
167 struct bbr *bbr = inet_csk_ca(sk);
168
169 return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
170}
171
172/* Return rate in bytes per second, optionally with a gain.
173 * The order here is chosen carefully to avoid overflow of u64. This should
174 * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
175 */
176static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
177{
178 rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache);
179 rate *= gain;
180 rate >>= BBR_SCALE;
181 rate *= USEC_PER_SEC;
182 return rate >> BW_SCALE;
183}
184
185/* Pace using current bw estimate and a gain factor. In order to help drive the
186 * network toward lower queues while maintaining high utilization and low
187 * latency, the average pacing rate aims to be slightly (~1%) lower than the
188 * estimated bandwidth. This is an important aspect of the design. In this
189 * implementation this slightly lower pacing rate is achieved implicitly by not
190 * including link-layer headers in the packet size used for the pacing rate.
191 */
192static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
193{
194 struct bbr *bbr = inet_csk_ca(sk);
195 u64 rate = bw;
196
197 rate = bbr_rate_bytes_per_sec(sk, rate, gain);
198 rate = min_t(u64, rate, sk->sk_max_pacing_rate);
199 if (bbr->mode != BBR_STARTUP || rate > sk->sk_pacing_rate)
200 sk->sk_pacing_rate = rate;
201}
202
203/* Return count of segments we want in the skbs we send, or 0 for default. */
204static u32 bbr_tso_segs_goal(struct sock *sk)
205{
206 struct bbr *bbr = inet_csk_ca(sk);
207
208 return bbr->tso_segs_goal;
209}
210
211static void bbr_set_tso_segs_goal(struct sock *sk)
212{
213 struct tcp_sock *tp = tcp_sk(sk);
214 struct bbr *bbr = inet_csk_ca(sk);
215 u32 min_segs;
216
217 min_segs = sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
218 bbr->tso_segs_goal = min(tcp_tso_autosize(sk, tp->mss_cache, min_segs),
219 0x7FU);
220}
221
222/* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
223static void bbr_save_cwnd(struct sock *sk)
224{
225 struct tcp_sock *tp = tcp_sk(sk);
226 struct bbr *bbr = inet_csk_ca(sk);
227
228 if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
229 bbr->prior_cwnd = tp->snd_cwnd; /* this cwnd is good enough */
230 else /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
231 bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
232}
233
234static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
235{
236 struct tcp_sock *tp = tcp_sk(sk);
237 struct bbr *bbr = inet_csk_ca(sk);
238
239 if (event == CA_EVENT_TX_START && tp->app_limited) {
240 bbr->idle_restart = 1;
241 /* Avoid pointless buffer overflows: pace at est. bw if we don't
242 * need more speed (we're restarting from idle and app-limited).
243 */
244 if (bbr->mode == BBR_PROBE_BW)
245 bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
246 }
247}
248
249/* Find target cwnd. Right-size the cwnd based on min RTT and the
250 * estimated bottleneck bandwidth:
251 *
252 * cwnd = bw * min_rtt * gain = BDP * gain
253 *
254 * The key factor, gain, controls the amount of queue. While a small gain
255 * builds a smaller queue, it becomes more vulnerable to noise in RTT
256 * measurements (e.g., delayed ACKs or other ACK compression effects). This
257 * noise may cause BBR to under-estimate the rate.
258 *
259 * To achieve full performance in high-speed paths, we budget enough cwnd to
260 * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
261 * - one skb in sending host Qdisc,
262 * - one skb in sending host TSO/GSO engine
263 * - one skb being received by receiver host LRO/GRO/delayed-ACK engine
264 * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
265 * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
266 * which allows 2 outstanding 2-packet sequences, to try to keep pipe
267 * full even with ACK-every-other-packet delayed ACKs.
268 */
269static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain)
270{
271 struct bbr *bbr = inet_csk_ca(sk);
272 u32 cwnd;
273 u64 w;
274
275 /* If we've never had a valid RTT sample, cap cwnd at the initial
276 * default. This should only happen when the connection is not using TCP
277 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
278 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
279 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
280 */
281 if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */
282 return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/
283
284 w = (u64)bw * bbr->min_rtt_us;
285
286 /* Apply a gain to the given value, then remove the BW_SCALE shift. */
287 cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
288
289 /* Allow enough full-sized skbs in flight to utilize end systems. */
290 cwnd += 3 * bbr->tso_segs_goal;
291
292 /* Reduce delayed ACKs by rounding up cwnd to the next even number. */
293 cwnd = (cwnd + 1) & ~1U;
294
295 return cwnd;
296}
297
298/* An optimization in BBR to reduce losses: On the first round of recovery, we
299 * follow the packet conservation principle: send P packets per P packets acked.
300 * After that, we slow-start and send at most 2*P packets per P packets acked.
301 * After recovery finishes, or upon undo, we restore the cwnd we had when
302 * recovery started (capped by the target cwnd based on estimated BDP).
303 *
304 * TODO(ycheng/ncardwell): implement a rate-based approach.
305 */
306static bool bbr_set_cwnd_to_recover_or_restore(
307 struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
308{
309 struct tcp_sock *tp = tcp_sk(sk);
310 struct bbr *bbr = inet_csk_ca(sk);
311 u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
312 u32 cwnd = tp->snd_cwnd;
313
314 /* An ACK for P pkts should release at most 2*P packets. We do this
315 * in two steps. First, here we deduct the number of lost packets.
316 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
317 */
318 if (rs->losses > 0)
319 cwnd = max_t(s32, cwnd - rs->losses, 1);
320
321 if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
322 /* Starting 1st round of Recovery, so do packet conservation. */
323 bbr->packet_conservation = 1;
324 bbr->next_rtt_delivered = tp->delivered; /* start round now */
325 /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
326 cwnd = tcp_packets_in_flight(tp) + acked;
327 } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
328 /* Exiting loss recovery; restore cwnd saved before recovery. */
329 bbr->restore_cwnd = 1;
330 bbr->packet_conservation = 0;
331 }
332 bbr->prev_ca_state = state;
333
334 if (bbr->restore_cwnd) {
335 /* Restore cwnd after exiting loss recovery or PROBE_RTT. */
336 cwnd = max(cwnd, bbr->prior_cwnd);
337 bbr->restore_cwnd = 0;
338 }
339
340 if (bbr->packet_conservation) {
341 *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
342 return true; /* yes, using packet conservation */
343 }
344 *new_cwnd = cwnd;
345 return false;
346}
347
348/* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
349 * has drawn us down below target), or snap down to target if we're above it.
350 */
351static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
352 u32 acked, u32 bw, int gain)
353{
354 struct tcp_sock *tp = tcp_sk(sk);
355 struct bbr *bbr = inet_csk_ca(sk);
356 u32 cwnd = 0, target_cwnd = 0;
357
358 if (!acked)
359 return;
360
361 if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
362 goto done;
363
364 /* If we're below target cwnd, slow start cwnd toward target cwnd. */
365 target_cwnd = bbr_target_cwnd(sk, bw, gain);
366 if (bbr_full_bw_reached(sk)) /* only cut cwnd if we filled the pipe */
367 cwnd = min(cwnd + acked, target_cwnd);
368 else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
369 cwnd = cwnd + acked;
370 cwnd = max(cwnd, bbr_cwnd_min_target);
371
372done:
373 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); /* apply global cap */
374 if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */
375 tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
376}
377
378/* End cycle phase if it's time and/or we hit the phase's in-flight target. */
379static bool bbr_is_next_cycle_phase(struct sock *sk,
380 const struct rate_sample *rs)
381{
382 struct tcp_sock *tp = tcp_sk(sk);
383 struct bbr *bbr = inet_csk_ca(sk);
384 bool is_full_length =
385 skb_mstamp_us_delta(&tp->delivered_mstamp, &bbr->cycle_mstamp) >
386 bbr->min_rtt_us;
387 u32 inflight, bw;
388
389 /* The pacing_gain of 1.0 paces at the estimated bw to try to fully
390 * use the pipe without increasing the queue.
391 */
392 if (bbr->pacing_gain == BBR_UNIT)
393 return is_full_length; /* just use wall clock time */
394
395 inflight = rs->prior_in_flight; /* what was in-flight before ACK? */
396 bw = bbr_max_bw(sk);
397
398 /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
399 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
400 * small (e.g. on a LAN). We do not persist if packets are lost, since
401 * a path with small buffers may not hold that much.
402 */
403 if (bbr->pacing_gain > BBR_UNIT)
404 return is_full_length &&
405 (rs->losses || /* perhaps pacing_gain*BDP won't fit */
406 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain));
407
408 /* A pacing_gain < 1.0 tries to drain extra queue we added if bw
409 * probing didn't find more bw. If inflight falls to match BDP then we
410 * estimate queue is drained; persisting would underutilize the pipe.
411 */
412 return is_full_length ||
413 inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT);
414}
415
416static void bbr_advance_cycle_phase(struct sock *sk)
417{
418 struct tcp_sock *tp = tcp_sk(sk);
419 struct bbr *bbr = inet_csk_ca(sk);
420
421 bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
422 bbr->cycle_mstamp = tp->delivered_mstamp;
423 bbr->pacing_gain = bbr_pacing_gain[bbr->cycle_idx];
424}
425
426/* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
427static void bbr_update_cycle_phase(struct sock *sk,
428 const struct rate_sample *rs)
429{
430 struct bbr *bbr = inet_csk_ca(sk);
431
432 if ((bbr->mode == BBR_PROBE_BW) && !bbr->lt_use_bw &&
433 bbr_is_next_cycle_phase(sk, rs))
434 bbr_advance_cycle_phase(sk);
435}
436
437static void bbr_reset_startup_mode(struct sock *sk)
438{
439 struct bbr *bbr = inet_csk_ca(sk);
440
441 bbr->mode = BBR_STARTUP;
442 bbr->pacing_gain = bbr_high_gain;
443 bbr->cwnd_gain = bbr_high_gain;
444}
445
446static void bbr_reset_probe_bw_mode(struct sock *sk)
447{
448 struct bbr *bbr = inet_csk_ca(sk);
449
450 bbr->mode = BBR_PROBE_BW;
451 bbr->pacing_gain = BBR_UNIT;
452 bbr->cwnd_gain = bbr_cwnd_gain;
453 bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
454 bbr_advance_cycle_phase(sk); /* flip to next phase of gain cycle */
455}
456
457static void bbr_reset_mode(struct sock *sk)
458{
459 if (!bbr_full_bw_reached(sk))
460 bbr_reset_startup_mode(sk);
461 else
462 bbr_reset_probe_bw_mode(sk);
463}
464
465/* Start a new long-term sampling interval. */
466static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
467{
468 struct tcp_sock *tp = tcp_sk(sk);
469 struct bbr *bbr = inet_csk_ca(sk);
470
471 bbr->lt_last_stamp = tp->delivered_mstamp.stamp_jiffies;
472 bbr->lt_last_delivered = tp->delivered;
473 bbr->lt_last_lost = tp->lost;
474 bbr->lt_rtt_cnt = 0;
475}
476
477/* Completely reset long-term bandwidth sampling. */
478static void bbr_reset_lt_bw_sampling(struct sock *sk)
479{
480 struct bbr *bbr = inet_csk_ca(sk);
481
482 bbr->lt_bw = 0;
483 bbr->lt_use_bw = 0;
484 bbr->lt_is_sampling = false;
485 bbr_reset_lt_bw_sampling_interval(sk);
486}
487
488/* Long-term bw sampling interval is done. Estimate whether we're policed. */
489static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
490{
491 struct bbr *bbr = inet_csk_ca(sk);
492 u32 diff;
493
494 if (bbr->lt_bw) { /* do we have bw from a previous interval? */
495 /* Is new bw close to the lt_bw from the previous interval? */
496 diff = abs(bw - bbr->lt_bw);
497 if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
498 (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
499 bbr_lt_bw_diff)) {
500 /* All criteria are met; estimate we're policed. */
501 bbr->lt_bw = (bw + bbr->lt_bw) >> 1; /* avg 2 intvls */
502 bbr->lt_use_bw = 1;
503 bbr->pacing_gain = BBR_UNIT; /* try to avoid drops */
504 bbr->lt_rtt_cnt = 0;
505 return;
506 }
507 }
508 bbr->lt_bw = bw;
509 bbr_reset_lt_bw_sampling_interval(sk);
510}
511
512/* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
513 * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
514 * explicitly models their policed rate, to reduce unnecessary losses. We
515 * estimate that we're policed if we see 2 consecutive sampling intervals with
516 * consistent throughput and high packet loss. If we think we're being policed,
517 * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
518 */
519static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
520{
521 struct tcp_sock *tp = tcp_sk(sk);
522 struct bbr *bbr = inet_csk_ca(sk);
523 u32 lost, delivered;
524 u64 bw;
525 s32 t;
526
527 if (bbr->lt_use_bw) { /* already using long-term rate, lt_bw? */
528 if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
529 ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
530 bbr_reset_lt_bw_sampling(sk); /* stop using lt_bw */
531 bbr_reset_probe_bw_mode(sk); /* restart gain cycling */
532 }
533 return;
534 }
535
536 /* Wait for the first loss before sampling, to let the policer exhaust
537 * its tokens and estimate the steady-state rate allowed by the policer.
538 * Starting samples earlier includes bursts that over-estimate the bw.
539 */
540 if (!bbr->lt_is_sampling) {
541 if (!rs->losses)
542 return;
543 bbr_reset_lt_bw_sampling_interval(sk);
544 bbr->lt_is_sampling = true;
545 }
546
547 /* To avoid underestimates, reset sampling if we run out of data. */
548 if (rs->is_app_limited) {
549 bbr_reset_lt_bw_sampling(sk);
550 return;
551 }
552
553 if (bbr->round_start)
554 bbr->lt_rtt_cnt++; /* count round trips in this interval */
555 if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
556 return; /* sampling interval needs to be longer */
557 if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
558 bbr_reset_lt_bw_sampling(sk); /* interval is too long */
559 return;
560 }
561
562 /* End sampling interval when a packet is lost, so we estimate the
563 * policer tokens were exhausted. Stopping the sampling before the
564 * tokens are exhausted under-estimates the policed rate.
565 */
566 if (!rs->losses)
567 return;
568
569 /* Calculate packets lost and delivered in sampling interval. */
570 lost = tp->lost - bbr->lt_last_lost;
571 delivered = tp->delivered - bbr->lt_last_delivered;
572 /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
573 if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
574 return;
575
576 /* Find average delivery rate in this sampling interval. */
577 t = (s32)(tp->delivered_mstamp.stamp_jiffies - bbr->lt_last_stamp);
578 if (t < 1)
579 return; /* interval is less than one jiffy, so wait */
580 t = jiffies_to_usecs(t);
581 /* Interval long enough for jiffies_to_usecs() to return a bogus 0? */
582 if (t < 1) {
583 bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */
584 return;
585 }
586 bw = (u64)delivered * BW_UNIT;
587 do_div(bw, t);
588 bbr_lt_bw_interval_done(sk, bw);
589}
590
591/* Estimate the bandwidth based on how fast packets are delivered */
592static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
593{
594 struct tcp_sock *tp = tcp_sk(sk);
595 struct bbr *bbr = inet_csk_ca(sk);
596 u64 bw;
597
598 bbr->round_start = 0;
599 if (rs->delivered < 0 || rs->interval_us <= 0)
600 return; /* Not a valid observation */
601
602 /* See if we've reached the next RTT */
603 if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
604 bbr->next_rtt_delivered = tp->delivered;
605 bbr->rtt_cnt++;
606 bbr->round_start = 1;
607 bbr->packet_conservation = 0;
608 }
609
610 bbr_lt_bw_sampling(sk, rs);
611
612 /* Divide delivered by the interval to find a (lower bound) bottleneck
613 * bandwidth sample. Delivered is in packets and interval_us in uS and
614 * ratio will be <<1 for most connections. So delivered is first scaled.
615 */
616 bw = (u64)rs->delivered * BW_UNIT;
617 do_div(bw, rs->interval_us);
618
619 /* If this sample is application-limited, it is likely to have a very
620 * low delivered count that represents application behavior rather than
621 * the available network rate. Such a sample could drag down estimated
622 * bw, causing needless slow-down. Thus, to continue to send at the
623 * last measured network rate, we filter out app-limited samples unless
624 * they describe the path bw at least as well as our bw model.
625 *
626 * So the goal during app-limited phase is to proceed with the best
627 * network rate no matter how long. We automatically leave this
628 * phase when app writes faster than the network can deliver :)
629 */
630 if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
631 /* Incorporate new sample into our max bw filter. */
632 minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
633 }
634}
635
636/* Estimate when the pipe is full, using the change in delivery rate: BBR
637 * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
638 * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
639 * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
640 * higher rwin, 3: we get higher delivery rate samples. Or transient
641 * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
642 * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
643 */
644static void bbr_check_full_bw_reached(struct sock *sk,
645 const struct rate_sample *rs)
646{
647 struct bbr *bbr = inet_csk_ca(sk);
648 u32 bw_thresh;
649
650 if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
651 return;
652
653 bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
654 if (bbr_max_bw(sk) >= bw_thresh) {
655 bbr->full_bw = bbr_max_bw(sk);
656 bbr->full_bw_cnt = 0;
657 return;
658 }
659 ++bbr->full_bw_cnt;
660}
661
662/* If pipe is probably full, drain the queue and then enter steady-state. */
663static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
664{
665 struct bbr *bbr = inet_csk_ca(sk);
666
667 if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
668 bbr->mode = BBR_DRAIN; /* drain queue we created */
669 bbr->pacing_gain = bbr_drain_gain; /* pace slow to drain */
670 bbr->cwnd_gain = bbr_high_gain; /* maintain cwnd */
671 } /* fall through to check if in-flight is already small: */
672 if (bbr->mode == BBR_DRAIN &&
673 tcp_packets_in_flight(tcp_sk(sk)) <=
674 bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT))
675 bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */
676}
677
678/* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
679 * periodically drain the bottleneck queue, to converge to measure the true
680 * min_rtt (unloaded propagation delay). This allows the flows to keep queues
681 * small (reducing queuing delay and packet loss) and achieve fairness among
682 * BBR flows.
683 *
684 * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
685 * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
686 * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
687 * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
688 * re-enter the previous mode. BBR uses 200ms to approximately bound the
689 * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
690 *
691 * Note that flows need only pay 2% if they are busy sending over the last 10
692 * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
693 * natural silences or low-rate periods within 10 seconds where the rate is low
694 * enough for long enough to drain its queue in the bottleneck. We pick up
695 * these min RTT measurements opportunistically with our min_rtt filter. :-)
696 */
697static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
698{
699 struct tcp_sock *tp = tcp_sk(sk);
700 struct bbr *bbr = inet_csk_ca(sk);
701 bool filter_expired;
702
703 /* Track min RTT seen in the min_rtt_win_sec filter window: */
704 filter_expired = after(tcp_time_stamp,
705 bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
706 if (rs->rtt_us >= 0 &&
707 (rs->rtt_us <= bbr->min_rtt_us || filter_expired)) {
708 bbr->min_rtt_us = rs->rtt_us;
709 bbr->min_rtt_stamp = tcp_time_stamp;
710 }
711
712 if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
713 !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
714 bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */
715 bbr->pacing_gain = BBR_UNIT;
716 bbr->cwnd_gain = BBR_UNIT;
717 bbr_save_cwnd(sk); /* note cwnd so we can restore it */
718 bbr->probe_rtt_done_stamp = 0;
719 }
720
721 if (bbr->mode == BBR_PROBE_RTT) {
722 /* Ignore low rate samples during this mode. */
723 tp->app_limited =
724 (tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
725 /* Maintain min packets in flight for max(200 ms, 1 round). */
726 if (!bbr->probe_rtt_done_stamp &&
727 tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
728 bbr->probe_rtt_done_stamp = tcp_time_stamp +
729 msecs_to_jiffies(bbr_probe_rtt_mode_ms);
730 bbr->probe_rtt_round_done = 0;
731 bbr->next_rtt_delivered = tp->delivered;
732 } else if (bbr->probe_rtt_done_stamp) {
733 if (bbr->round_start)
734 bbr->probe_rtt_round_done = 1;
735 if (bbr->probe_rtt_round_done &&
736 after(tcp_time_stamp, bbr->probe_rtt_done_stamp)) {
737 bbr->min_rtt_stamp = tcp_time_stamp;
738 bbr->restore_cwnd = 1; /* snap to prior_cwnd */
739 bbr_reset_mode(sk);
740 }
741 }
742 }
743 bbr->idle_restart = 0;
744}
745
746static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
747{
748 bbr_update_bw(sk, rs);
749 bbr_update_cycle_phase(sk, rs);
750 bbr_check_full_bw_reached(sk, rs);
751 bbr_check_drain(sk, rs);
752 bbr_update_min_rtt(sk, rs);
753}
754
755static void bbr_main(struct sock *sk, const struct rate_sample *rs)
756{
757 struct bbr *bbr = inet_csk_ca(sk);
758 u32 bw;
759
760 bbr_update_model(sk, rs);
761
762 bw = bbr_bw(sk);
763 bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
764 bbr_set_tso_segs_goal(sk);
765 bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
766}
767
768static void bbr_init(struct sock *sk)
769{
770 struct tcp_sock *tp = tcp_sk(sk);
771 struct bbr *bbr = inet_csk_ca(sk);
772 u64 bw;
773
774 bbr->prior_cwnd = 0;
775 bbr->tso_segs_goal = 0; /* default segs per skb until first ACK */
776 bbr->rtt_cnt = 0;
777 bbr->next_rtt_delivered = 0;
778 bbr->prev_ca_state = TCP_CA_Open;
779 bbr->packet_conservation = 0;
780
781 bbr->probe_rtt_done_stamp = 0;
782 bbr->probe_rtt_round_done = 0;
783 bbr->min_rtt_us = tcp_min_rtt(tp);
784 bbr->min_rtt_stamp = tcp_time_stamp;
785
786 minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */
787
788 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
789 bw = (u64)tp->snd_cwnd * BW_UNIT;
790 do_div(bw, (tp->srtt_us >> 3) ? : USEC_PER_MSEC);
791 sk->sk_pacing_rate = 0; /* force an update of sk_pacing_rate */
792 bbr_set_pacing_rate(sk, bw, bbr_high_gain);
793
794 bbr->restore_cwnd = 0;
795 bbr->round_start = 0;
796 bbr->idle_restart = 0;
797 bbr->full_bw = 0;
798 bbr->full_bw_cnt = 0;
799 bbr->cycle_mstamp.v64 = 0;
800 bbr->cycle_idx = 0;
801 bbr_reset_lt_bw_sampling(sk);
802 bbr_reset_startup_mode(sk);
803}
804
805static u32 bbr_sndbuf_expand(struct sock *sk)
806{
807 /* Provision 3 * cwnd since BBR may slow-start even during recovery. */
808 return 3;
809}
810
811/* In theory BBR does not need to undo the cwnd since it does not
812 * always reduce cwnd on losses (see bbr_main()). Keep it for now.
813 */
814static u32 bbr_undo_cwnd(struct sock *sk)
815{
816 return tcp_sk(sk)->snd_cwnd;
817}
818
819/* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
820static u32 bbr_ssthresh(struct sock *sk)
821{
822 bbr_save_cwnd(sk);
823 return TCP_INFINITE_SSTHRESH; /* BBR does not use ssthresh */
824}
825
826static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
827 union tcp_cc_info *info)
828{
829 if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
830 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
831 struct tcp_sock *tp = tcp_sk(sk);
832 struct bbr *bbr = inet_csk_ca(sk);
833 u64 bw = bbr_bw(sk);
834
835 bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
836 memset(&info->bbr, 0, sizeof(info->bbr));
837 info->bbr.bbr_bw_lo = (u32)bw;
838 info->bbr.bbr_bw_hi = (u32)(bw >> 32);
839 info->bbr.bbr_min_rtt = bbr->min_rtt_us;
840 info->bbr.bbr_pacing_gain = bbr->pacing_gain;
841 info->bbr.bbr_cwnd_gain = bbr->cwnd_gain;
842 *attr = INET_DIAG_BBRINFO;
843 return sizeof(info->bbr);
844 }
845 return 0;
846}
847
848static void bbr_set_state(struct sock *sk, u8 new_state)
849{
850 struct bbr *bbr = inet_csk_ca(sk);
851
852 if (new_state == TCP_CA_Loss) {
853 struct rate_sample rs = { .losses = 1 };
854
855 bbr->prev_ca_state = TCP_CA_Loss;
856 bbr->full_bw = 0;
857 bbr->round_start = 1; /* treat RTO like end of a round */
858 bbr_lt_bw_sampling(sk, &rs);
859 }
860}
861
862static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
863 .flags = TCP_CONG_NON_RESTRICTED,
864 .name = "bbr",
865 .owner = THIS_MODULE,
866 .init = bbr_init,
867 .cong_control = bbr_main,
868 .sndbuf_expand = bbr_sndbuf_expand,
869 .undo_cwnd = bbr_undo_cwnd,
870 .cwnd_event = bbr_cwnd_event,
871 .ssthresh = bbr_ssthresh,
872 .tso_segs_goal = bbr_tso_segs_goal,
873 .get_info = bbr_get_info,
874 .set_state = bbr_set_state,
875};
876
877static int __init bbr_register(void)
878{
879 BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
880 return tcp_register_congestion_control(&tcp_bbr_cong_ops);
881}
882
883static void __exit bbr_unregister(void)
884{
885 tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
886}
887
888module_init(bbr_register);
889module_exit(bbr_unregister);
890
891MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
892MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
893MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
894MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
895MODULE_LICENSE("Dual BSD/GPL");
896MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");