xref: /linux/net/ipv4/tcp_bbr.c (revision 856e7c4b619af622d56b3b454f7bec32a170ac99)
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: */
79 enum 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 */
87 struct 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 		restore_cwnd:1,	     /* decided to revert cwnd to old value */
99 		round_start:1,	     /* start of packet-timed tx->ack round? */
100 		idle_restart:1,	     /* restarting after idle? */
101 		probe_rtt_round_done:1,  /* a BBR_PROBE_RTT round at 4 pkts? */
102 		unused:12,
103 		lt_is_sampling:1,    /* taking long-term ("LT") samples now? */
104 		lt_rtt_cnt:7,	     /* round trips in long-term interval */
105 		lt_use_bw:1;	     /* use lt_bw as our bw estimate? */
106 	u32	lt_bw;		     /* LT est delivery rate in pkts/uS << 24 */
107 	u32	lt_last_delivered;   /* LT intvl start: tp->delivered */
108 	u32	lt_last_stamp;	     /* LT intvl start: tp->delivered_mstamp */
109 	u32	lt_last_lost;	     /* LT intvl start: tp->lost */
110 	u32	pacing_gain:10,	/* current gain for setting pacing rate */
111 		cwnd_gain:10,	/* current gain for setting cwnd */
112 		full_bw_reached:1,   /* reached full bw in Startup? */
113 		full_bw_cnt:2,	/* number of rounds without large bw gains */
114 		cycle_idx:3,	/* current index in pacing_gain cycle array */
115 		has_seen_rtt:1, /* have we seen an RTT sample yet? */
116 		unused_b:5;
117 	u32	prior_cwnd;	/* prior cwnd upon entering loss recovery */
118 	u32	full_bw;	/* recent bw, to estimate if pipe is full */
119 };
120 
121 #define CYCLE_LEN	8	/* number of phases in a pacing gain cycle */
122 
123 /* Window length of bw filter (in rounds): */
124 static const int bbr_bw_rtts = CYCLE_LEN + 2;
125 /* Window length of min_rtt filter (in sec): */
126 static const u32 bbr_min_rtt_win_sec = 10;
127 /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
128 static const u32 bbr_probe_rtt_mode_ms = 200;
129 /* Skip TSO below the following bandwidth (bits/sec): */
130 static const int bbr_min_tso_rate = 1200000;
131 
132 /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
133  * that will allow a smoothly increasing pacing rate that will double each RTT
134  * and send the same number of packets per RTT that an un-paced, slow-starting
135  * Reno or CUBIC flow would:
136  */
137 static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
138 /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
139  * the queue created in BBR_STARTUP in a single round:
140  */
141 static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
142 /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
143 static const int bbr_cwnd_gain  = BBR_UNIT * 2;
144 /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
145 static const int bbr_pacing_gain[] = {
146 	BBR_UNIT * 5 / 4,	/* probe for more available bw */
147 	BBR_UNIT * 3 / 4,	/* drain queue and/or yield bw to other flows */
148 	BBR_UNIT, BBR_UNIT, BBR_UNIT,	/* cruise at 1.0*bw to utilize pipe, */
149 	BBR_UNIT, BBR_UNIT, BBR_UNIT	/* without creating excess queue... */
150 };
151 /* Randomize the starting gain cycling phase over N phases: */
152 static const u32 bbr_cycle_rand = 7;
153 
154 /* Try to keep at least this many packets in flight, if things go smoothly. For
155  * smooth functioning, a sliding window protocol ACKing every other packet
156  * needs at least 4 packets in flight:
157  */
158 static const u32 bbr_cwnd_min_target = 4;
159 
160 /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
161 /* If bw has increased significantly (1.25x), there may be more bw available: */
162 static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
163 /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
164 static const u32 bbr_full_bw_cnt = 3;
165 
166 /* "long-term" ("LT") bandwidth estimator parameters... */
167 /* The minimum number of rounds in an LT bw sampling interval: */
168 static const u32 bbr_lt_intvl_min_rtts = 4;
169 /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
170 static const u32 bbr_lt_loss_thresh = 50;
171 /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
172 static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
173 /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
174 static const u32 bbr_lt_bw_diff = 4000 / 8;
175 /* If we estimate we're policed, use lt_bw for this many round trips: */
176 static const u32 bbr_lt_bw_max_rtts = 48;
177 
178 /* Do we estimate that STARTUP filled the pipe? */
179 static bool bbr_full_bw_reached(const struct sock *sk)
180 {
181 	const struct bbr *bbr = inet_csk_ca(sk);
182 
183 	return bbr->full_bw_reached;
184 }
185 
186 /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
187 static u32 bbr_max_bw(const struct sock *sk)
188 {
189 	struct bbr *bbr = inet_csk_ca(sk);
190 
191 	return minmax_get(&bbr->bw);
192 }
193 
194 /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
195 static u32 bbr_bw(const struct sock *sk)
196 {
197 	struct bbr *bbr = inet_csk_ca(sk);
198 
199 	return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
200 }
201 
202 /* Return rate in bytes per second, optionally with a gain.
203  * The order here is chosen carefully to avoid overflow of u64. This should
204  * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
205  */
206 static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
207 {
208 	rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache);
209 	rate *= gain;
210 	rate >>= BBR_SCALE;
211 	rate *= USEC_PER_SEC;
212 	return rate >> BW_SCALE;
213 }
214 
215 /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
216 static u32 bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
217 {
218 	u64 rate = bw;
219 
220 	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
221 	rate = min_t(u64, rate, sk->sk_max_pacing_rate);
222 	return rate;
223 }
224 
225 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
226 static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
227 {
228 	struct tcp_sock *tp = tcp_sk(sk);
229 	struct bbr *bbr = inet_csk_ca(sk);
230 	u64 bw;
231 	u32 rtt_us;
232 
233 	if (tp->srtt_us) {		/* any RTT sample yet? */
234 		rtt_us = max(tp->srtt_us >> 3, 1U);
235 		bbr->has_seen_rtt = 1;
236 	} else {			 /* no RTT sample yet */
237 		rtt_us = USEC_PER_MSEC;	 /* use nominal default RTT */
238 	}
239 	bw = (u64)tp->snd_cwnd * BW_UNIT;
240 	do_div(bw, rtt_us);
241 	sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain);
242 }
243 
244 /* Pace using current bw estimate and a gain factor. In order to help drive the
245  * network toward lower queues while maintaining high utilization and low
246  * latency, the average pacing rate aims to be slightly (~1%) lower than the
247  * estimated bandwidth. This is an important aspect of the design. In this
248  * implementation this slightly lower pacing rate is achieved implicitly by not
249  * including link-layer headers in the packet size used for the pacing rate.
250  */
251 static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
252 {
253 	struct tcp_sock *tp = tcp_sk(sk);
254 	struct bbr *bbr = inet_csk_ca(sk);
255 	u32 rate = bbr_bw_to_pacing_rate(sk, bw, gain);
256 
257 	if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
258 		bbr_init_pacing_rate_from_rtt(sk);
259 	if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate)
260 		sk->sk_pacing_rate = rate;
261 }
262 
263 /* override sysctl_tcp_min_tso_segs */
264 static u32 bbr_min_tso_segs(struct sock *sk)
265 {
266 	return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
267 }
268 
269 static u32 bbr_tso_segs_goal(struct sock *sk)
270 {
271 	struct tcp_sock *tp = tcp_sk(sk);
272 	u32 segs, bytes;
273 
274 	/* Sort of tcp_tso_autosize() but ignoring
275 	 * driver provided sk_gso_max_size.
276 	 */
277 	bytes = min_t(u32, sk->sk_pacing_rate >> sk->sk_pacing_shift,
278 		      GSO_MAX_SIZE - 1 - MAX_TCP_HEADER);
279 	segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk));
280 
281 	return min(segs, 0x7FU);
282 }
283 
284 /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
285 static void bbr_save_cwnd(struct sock *sk)
286 {
287 	struct tcp_sock *tp = tcp_sk(sk);
288 	struct bbr *bbr = inet_csk_ca(sk);
289 
290 	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
291 		bbr->prior_cwnd = tp->snd_cwnd;  /* this cwnd is good enough */
292 	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
293 		bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
294 }
295 
296 static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
297 {
298 	struct tcp_sock *tp = tcp_sk(sk);
299 	struct bbr *bbr = inet_csk_ca(sk);
300 
301 	if (event == CA_EVENT_TX_START && tp->app_limited) {
302 		bbr->idle_restart = 1;
303 		/* Avoid pointless buffer overflows: pace at est. bw if we don't
304 		 * need more speed (we're restarting from idle and app-limited).
305 		 */
306 		if (bbr->mode == BBR_PROBE_BW)
307 			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
308 	}
309 }
310 
311 /* Find target cwnd. Right-size the cwnd based on min RTT and the
312  * estimated bottleneck bandwidth:
313  *
314  * cwnd = bw * min_rtt * gain = BDP * gain
315  *
316  * The key factor, gain, controls the amount of queue. While a small gain
317  * builds a smaller queue, it becomes more vulnerable to noise in RTT
318  * measurements (e.g., delayed ACKs or other ACK compression effects). This
319  * noise may cause BBR to under-estimate the rate.
320  *
321  * To achieve full performance in high-speed paths, we budget enough cwnd to
322  * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
323  *   - one skb in sending host Qdisc,
324  *   - one skb in sending host TSO/GSO engine
325  *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
326  * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
327  * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
328  * which allows 2 outstanding 2-packet sequences, to try to keep pipe
329  * full even with ACK-every-other-packet delayed ACKs.
330  */
331 static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain)
332 {
333 	struct bbr *bbr = inet_csk_ca(sk);
334 	u32 cwnd;
335 	u64 w;
336 
337 	/* If we've never had a valid RTT sample, cap cwnd at the initial
338 	 * default. This should only happen when the connection is not using TCP
339 	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
340 	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
341 	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
342 	 */
343 	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
344 		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
345 
346 	w = (u64)bw * bbr->min_rtt_us;
347 
348 	/* Apply a gain to the given value, then remove the BW_SCALE shift. */
349 	cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
350 
351 	/* Allow enough full-sized skbs in flight to utilize end systems. */
352 	cwnd += 3 * bbr_tso_segs_goal(sk);
353 
354 	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
355 	cwnd = (cwnd + 1) & ~1U;
356 
357 	return cwnd;
358 }
359 
360 /* An optimization in BBR to reduce losses: On the first round of recovery, we
361  * follow the packet conservation principle: send P packets per P packets acked.
362  * After that, we slow-start and send at most 2*P packets per P packets acked.
363  * After recovery finishes, or upon undo, we restore the cwnd we had when
364  * recovery started (capped by the target cwnd based on estimated BDP).
365  *
366  * TODO(ycheng/ncardwell): implement a rate-based approach.
367  */
368 static bool bbr_set_cwnd_to_recover_or_restore(
369 	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
370 {
371 	struct tcp_sock *tp = tcp_sk(sk);
372 	struct bbr *bbr = inet_csk_ca(sk);
373 	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
374 	u32 cwnd = tp->snd_cwnd;
375 
376 	/* An ACK for P pkts should release at most 2*P packets. We do this
377 	 * in two steps. First, here we deduct the number of lost packets.
378 	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
379 	 */
380 	if (rs->losses > 0)
381 		cwnd = max_t(s32, cwnd - rs->losses, 1);
382 
383 	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
384 		/* Starting 1st round of Recovery, so do packet conservation. */
385 		bbr->packet_conservation = 1;
386 		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
387 		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
388 		cwnd = tcp_packets_in_flight(tp) + acked;
389 	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
390 		/* Exiting loss recovery; restore cwnd saved before recovery. */
391 		bbr->restore_cwnd = 1;
392 		bbr->packet_conservation = 0;
393 	}
394 	bbr->prev_ca_state = state;
395 
396 	if (bbr->restore_cwnd) {
397 		/* Restore cwnd after exiting loss recovery or PROBE_RTT. */
398 		cwnd = max(cwnd, bbr->prior_cwnd);
399 		bbr->restore_cwnd = 0;
400 	}
401 
402 	if (bbr->packet_conservation) {
403 		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
404 		return true;	/* yes, using packet conservation */
405 	}
406 	*new_cwnd = cwnd;
407 	return false;
408 }
409 
410 /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
411  * has drawn us down below target), or snap down to target if we're above it.
412  */
413 static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
414 			 u32 acked, u32 bw, int gain)
415 {
416 	struct tcp_sock *tp = tcp_sk(sk);
417 	struct bbr *bbr = inet_csk_ca(sk);
418 	u32 cwnd = 0, target_cwnd = 0;
419 
420 	if (!acked)
421 		return;
422 
423 	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
424 		goto done;
425 
426 	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
427 	target_cwnd = bbr_target_cwnd(sk, bw, gain);
428 	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
429 		cwnd = min(cwnd + acked, target_cwnd);
430 	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
431 		cwnd = cwnd + acked;
432 	cwnd = max(cwnd, bbr_cwnd_min_target);
433 
434 done:
435 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);	/* apply global cap */
436 	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
437 		tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
438 }
439 
440 /* End cycle phase if it's time and/or we hit the phase's in-flight target. */
441 static bool bbr_is_next_cycle_phase(struct sock *sk,
442 				    const struct rate_sample *rs)
443 {
444 	struct tcp_sock *tp = tcp_sk(sk);
445 	struct bbr *bbr = inet_csk_ca(sk);
446 	bool is_full_length =
447 		tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
448 		bbr->min_rtt_us;
449 	u32 inflight, bw;
450 
451 	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
452 	 * use the pipe without increasing the queue.
453 	 */
454 	if (bbr->pacing_gain == BBR_UNIT)
455 		return is_full_length;		/* just use wall clock time */
456 
457 	inflight = rs->prior_in_flight;  /* what was in-flight before ACK? */
458 	bw = bbr_max_bw(sk);
459 
460 	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
461 	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
462 	 * small (e.g. on a LAN). We do not persist if packets are lost, since
463 	 * a path with small buffers may not hold that much.
464 	 */
465 	if (bbr->pacing_gain > BBR_UNIT)
466 		return is_full_length &&
467 			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
468 			 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain));
469 
470 	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
471 	 * probing didn't find more bw. If inflight falls to match BDP then we
472 	 * estimate queue is drained; persisting would underutilize the pipe.
473 	 */
474 	return is_full_length ||
475 		inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT);
476 }
477 
478 static void bbr_advance_cycle_phase(struct sock *sk)
479 {
480 	struct tcp_sock *tp = tcp_sk(sk);
481 	struct bbr *bbr = inet_csk_ca(sk);
482 
483 	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
484 	bbr->cycle_mstamp = tp->delivered_mstamp;
485 	bbr->pacing_gain = bbr->lt_use_bw ? BBR_UNIT :
486 					    bbr_pacing_gain[bbr->cycle_idx];
487 }
488 
489 /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
490 static void bbr_update_cycle_phase(struct sock *sk,
491 				   const struct rate_sample *rs)
492 {
493 	struct bbr *bbr = inet_csk_ca(sk);
494 
495 	if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
496 		bbr_advance_cycle_phase(sk);
497 }
498 
499 static void bbr_reset_startup_mode(struct sock *sk)
500 {
501 	struct bbr *bbr = inet_csk_ca(sk);
502 
503 	bbr->mode = BBR_STARTUP;
504 	bbr->pacing_gain = bbr_high_gain;
505 	bbr->cwnd_gain	 = bbr_high_gain;
506 }
507 
508 static void bbr_reset_probe_bw_mode(struct sock *sk)
509 {
510 	struct bbr *bbr = inet_csk_ca(sk);
511 
512 	bbr->mode = BBR_PROBE_BW;
513 	bbr->pacing_gain = BBR_UNIT;
514 	bbr->cwnd_gain = bbr_cwnd_gain;
515 	bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
516 	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
517 }
518 
519 static void bbr_reset_mode(struct sock *sk)
520 {
521 	if (!bbr_full_bw_reached(sk))
522 		bbr_reset_startup_mode(sk);
523 	else
524 		bbr_reset_probe_bw_mode(sk);
525 }
526 
527 /* Start a new long-term sampling interval. */
528 static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
529 {
530 	struct tcp_sock *tp = tcp_sk(sk);
531 	struct bbr *bbr = inet_csk_ca(sk);
532 
533 	bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
534 	bbr->lt_last_delivered = tp->delivered;
535 	bbr->lt_last_lost = tp->lost;
536 	bbr->lt_rtt_cnt = 0;
537 }
538 
539 /* Completely reset long-term bandwidth sampling. */
540 static void bbr_reset_lt_bw_sampling(struct sock *sk)
541 {
542 	struct bbr *bbr = inet_csk_ca(sk);
543 
544 	bbr->lt_bw = 0;
545 	bbr->lt_use_bw = 0;
546 	bbr->lt_is_sampling = false;
547 	bbr_reset_lt_bw_sampling_interval(sk);
548 }
549 
550 /* Long-term bw sampling interval is done. Estimate whether we're policed. */
551 static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
552 {
553 	struct bbr *bbr = inet_csk_ca(sk);
554 	u32 diff;
555 
556 	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
557 		/* Is new bw close to the lt_bw from the previous interval? */
558 		diff = abs(bw - bbr->lt_bw);
559 		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
560 		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
561 		     bbr_lt_bw_diff)) {
562 			/* All criteria are met; estimate we're policed. */
563 			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
564 			bbr->lt_use_bw = 1;
565 			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
566 			bbr->lt_rtt_cnt = 0;
567 			return;
568 		}
569 	}
570 	bbr->lt_bw = bw;
571 	bbr_reset_lt_bw_sampling_interval(sk);
572 }
573 
574 /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
575  * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
576  * explicitly models their policed rate, to reduce unnecessary losses. We
577  * estimate that we're policed if we see 2 consecutive sampling intervals with
578  * consistent throughput and high packet loss. If we think we're being policed,
579  * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
580  */
581 static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
582 {
583 	struct tcp_sock *tp = tcp_sk(sk);
584 	struct bbr *bbr = inet_csk_ca(sk);
585 	u32 lost, delivered;
586 	u64 bw;
587 	u32 t;
588 
589 	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
590 		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
591 		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
592 			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
593 			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
594 		}
595 		return;
596 	}
597 
598 	/* Wait for the first loss before sampling, to let the policer exhaust
599 	 * its tokens and estimate the steady-state rate allowed by the policer.
600 	 * Starting samples earlier includes bursts that over-estimate the bw.
601 	 */
602 	if (!bbr->lt_is_sampling) {
603 		if (!rs->losses)
604 			return;
605 		bbr_reset_lt_bw_sampling_interval(sk);
606 		bbr->lt_is_sampling = true;
607 	}
608 
609 	/* To avoid underestimates, reset sampling if we run out of data. */
610 	if (rs->is_app_limited) {
611 		bbr_reset_lt_bw_sampling(sk);
612 		return;
613 	}
614 
615 	if (bbr->round_start)
616 		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
617 	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
618 		return;		/* sampling interval needs to be longer */
619 	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
620 		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
621 		return;
622 	}
623 
624 	/* End sampling interval when a packet is lost, so we estimate the
625 	 * policer tokens were exhausted. Stopping the sampling before the
626 	 * tokens are exhausted under-estimates the policed rate.
627 	 */
628 	if (!rs->losses)
629 		return;
630 
631 	/* Calculate packets lost and delivered in sampling interval. */
632 	lost = tp->lost - bbr->lt_last_lost;
633 	delivered = tp->delivered - bbr->lt_last_delivered;
634 	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
635 	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
636 		return;
637 
638 	/* Find average delivery rate in this sampling interval. */
639 	t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
640 	if ((s32)t < 1)
641 		return;		/* interval is less than one ms, so wait */
642 	/* Check if can multiply without overflow */
643 	if (t >= ~0U / USEC_PER_MSEC) {
644 		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
645 		return;
646 	}
647 	t *= USEC_PER_MSEC;
648 	bw = (u64)delivered * BW_UNIT;
649 	do_div(bw, t);
650 	bbr_lt_bw_interval_done(sk, bw);
651 }
652 
653 /* Estimate the bandwidth based on how fast packets are delivered */
654 static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
655 {
656 	struct tcp_sock *tp = tcp_sk(sk);
657 	struct bbr *bbr = inet_csk_ca(sk);
658 	u64 bw;
659 
660 	bbr->round_start = 0;
661 	if (rs->delivered < 0 || rs->interval_us <= 0)
662 		return; /* Not a valid observation */
663 
664 	/* See if we've reached the next RTT */
665 	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
666 		bbr->next_rtt_delivered = tp->delivered;
667 		bbr->rtt_cnt++;
668 		bbr->round_start = 1;
669 		bbr->packet_conservation = 0;
670 	}
671 
672 	bbr_lt_bw_sampling(sk, rs);
673 
674 	/* Divide delivered by the interval to find a (lower bound) bottleneck
675 	 * bandwidth sample. Delivered is in packets and interval_us in uS and
676 	 * ratio will be <<1 for most connections. So delivered is first scaled.
677 	 */
678 	bw = (u64)rs->delivered * BW_UNIT;
679 	do_div(bw, rs->interval_us);
680 
681 	/* If this sample is application-limited, it is likely to have a very
682 	 * low delivered count that represents application behavior rather than
683 	 * the available network rate. Such a sample could drag down estimated
684 	 * bw, causing needless slow-down. Thus, to continue to send at the
685 	 * last measured network rate, we filter out app-limited samples unless
686 	 * they describe the path bw at least as well as our bw model.
687 	 *
688 	 * So the goal during app-limited phase is to proceed with the best
689 	 * network rate no matter how long. We automatically leave this
690 	 * phase when app writes faster than the network can deliver :)
691 	 */
692 	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
693 		/* Incorporate new sample into our max bw filter. */
694 		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
695 	}
696 }
697 
698 /* Estimate when the pipe is full, using the change in delivery rate: BBR
699  * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
700  * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
701  * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
702  * higher rwin, 3: we get higher delivery rate samples. Or transient
703  * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
704  * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
705  */
706 static void bbr_check_full_bw_reached(struct sock *sk,
707 				      const struct rate_sample *rs)
708 {
709 	struct bbr *bbr = inet_csk_ca(sk);
710 	u32 bw_thresh;
711 
712 	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
713 		return;
714 
715 	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
716 	if (bbr_max_bw(sk) >= bw_thresh) {
717 		bbr->full_bw = bbr_max_bw(sk);
718 		bbr->full_bw_cnt = 0;
719 		return;
720 	}
721 	++bbr->full_bw_cnt;
722 	bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
723 }
724 
725 /* If pipe is probably full, drain the queue and then enter steady-state. */
726 static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
727 {
728 	struct bbr *bbr = inet_csk_ca(sk);
729 
730 	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
731 		bbr->mode = BBR_DRAIN;	/* drain queue we created */
732 		bbr->pacing_gain = bbr_drain_gain;	/* pace slow to drain */
733 		bbr->cwnd_gain = bbr_high_gain;	/* maintain cwnd */
734 		tcp_sk(sk)->snd_ssthresh =
735 				bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT);
736 	}	/* fall through to check if in-flight is already small: */
737 	if (bbr->mode == BBR_DRAIN &&
738 	    tcp_packets_in_flight(tcp_sk(sk)) <=
739 	    bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT))
740 		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
741 }
742 
743 /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
744  * periodically drain the bottleneck queue, to converge to measure the true
745  * min_rtt (unloaded propagation delay). This allows the flows to keep queues
746  * small (reducing queuing delay and packet loss) and achieve fairness among
747  * BBR flows.
748  *
749  * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
750  * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
751  * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
752  * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
753  * re-enter the previous mode. BBR uses 200ms to approximately bound the
754  * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
755  *
756  * Note that flows need only pay 2% if they are busy sending over the last 10
757  * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
758  * natural silences or low-rate periods within 10 seconds where the rate is low
759  * enough for long enough to drain its queue in the bottleneck. We pick up
760  * these min RTT measurements opportunistically with our min_rtt filter. :-)
761  */
762 static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
763 {
764 	struct tcp_sock *tp = tcp_sk(sk);
765 	struct bbr *bbr = inet_csk_ca(sk);
766 	bool filter_expired;
767 
768 	/* Track min RTT seen in the min_rtt_win_sec filter window: */
769 	filter_expired = after(tcp_jiffies32,
770 			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
771 	if (rs->rtt_us >= 0 &&
772 	    (rs->rtt_us <= bbr->min_rtt_us ||
773 	     (filter_expired && !rs->is_ack_delayed))) {
774 		bbr->min_rtt_us = rs->rtt_us;
775 		bbr->min_rtt_stamp = tcp_jiffies32;
776 	}
777 
778 	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
779 	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
780 		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
781 		bbr->pacing_gain = BBR_UNIT;
782 		bbr->cwnd_gain = BBR_UNIT;
783 		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
784 		bbr->probe_rtt_done_stamp = 0;
785 	}
786 
787 	if (bbr->mode == BBR_PROBE_RTT) {
788 		/* Ignore low rate samples during this mode. */
789 		tp->app_limited =
790 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
791 		/* Maintain min packets in flight for max(200 ms, 1 round). */
792 		if (!bbr->probe_rtt_done_stamp &&
793 		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
794 			bbr->probe_rtt_done_stamp = tcp_jiffies32 +
795 				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
796 			bbr->probe_rtt_round_done = 0;
797 			bbr->next_rtt_delivered = tp->delivered;
798 		} else if (bbr->probe_rtt_done_stamp) {
799 			if (bbr->round_start)
800 				bbr->probe_rtt_round_done = 1;
801 			if (bbr->probe_rtt_round_done &&
802 			    after(tcp_jiffies32, bbr->probe_rtt_done_stamp)) {
803 				bbr->min_rtt_stamp = tcp_jiffies32;
804 				bbr->restore_cwnd = 1;  /* snap to prior_cwnd */
805 				bbr_reset_mode(sk);
806 			}
807 		}
808 	}
809 	/* Restart after idle ends only once we process a new S/ACK for data */
810 	if (rs->delivered > 0)
811 		bbr->idle_restart = 0;
812 }
813 
814 static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
815 {
816 	bbr_update_bw(sk, rs);
817 	bbr_update_cycle_phase(sk, rs);
818 	bbr_check_full_bw_reached(sk, rs);
819 	bbr_check_drain(sk, rs);
820 	bbr_update_min_rtt(sk, rs);
821 }
822 
823 static void bbr_main(struct sock *sk, const struct rate_sample *rs)
824 {
825 	struct bbr *bbr = inet_csk_ca(sk);
826 	u32 bw;
827 
828 	bbr_update_model(sk, rs);
829 
830 	bw = bbr_bw(sk);
831 	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
832 	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
833 }
834 
835 static void bbr_init(struct sock *sk)
836 {
837 	struct tcp_sock *tp = tcp_sk(sk);
838 	struct bbr *bbr = inet_csk_ca(sk);
839 
840 	bbr->prior_cwnd = 0;
841 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
842 	bbr->rtt_cnt = 0;
843 	bbr->next_rtt_delivered = 0;
844 	bbr->prev_ca_state = TCP_CA_Open;
845 	bbr->packet_conservation = 0;
846 
847 	bbr->probe_rtt_done_stamp = 0;
848 	bbr->probe_rtt_round_done = 0;
849 	bbr->min_rtt_us = tcp_min_rtt(tp);
850 	bbr->min_rtt_stamp = tcp_jiffies32;
851 
852 	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
853 
854 	bbr->has_seen_rtt = 0;
855 	bbr_init_pacing_rate_from_rtt(sk);
856 
857 	bbr->restore_cwnd = 0;
858 	bbr->round_start = 0;
859 	bbr->idle_restart = 0;
860 	bbr->full_bw_reached = 0;
861 	bbr->full_bw = 0;
862 	bbr->full_bw_cnt = 0;
863 	bbr->cycle_mstamp = 0;
864 	bbr->cycle_idx = 0;
865 	bbr_reset_lt_bw_sampling(sk);
866 	bbr_reset_startup_mode(sk);
867 
868 	cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
869 }
870 
871 static u32 bbr_sndbuf_expand(struct sock *sk)
872 {
873 	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
874 	return 3;
875 }
876 
877 /* In theory BBR does not need to undo the cwnd since it does not
878  * always reduce cwnd on losses (see bbr_main()). Keep it for now.
879  */
880 static u32 bbr_undo_cwnd(struct sock *sk)
881 {
882 	struct bbr *bbr = inet_csk_ca(sk);
883 
884 	bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
885 	bbr->full_bw_cnt = 0;
886 	bbr_reset_lt_bw_sampling(sk);
887 	return tcp_sk(sk)->snd_cwnd;
888 }
889 
890 /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
891 static u32 bbr_ssthresh(struct sock *sk)
892 {
893 	bbr_save_cwnd(sk);
894 	return tcp_sk(sk)->snd_ssthresh;
895 }
896 
897 static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
898 			   union tcp_cc_info *info)
899 {
900 	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
901 	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
902 		struct tcp_sock *tp = tcp_sk(sk);
903 		struct bbr *bbr = inet_csk_ca(sk);
904 		u64 bw = bbr_bw(sk);
905 
906 		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
907 		memset(&info->bbr, 0, sizeof(info->bbr));
908 		info->bbr.bbr_bw_lo		= (u32)bw;
909 		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
910 		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
911 		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
912 		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
913 		*attr = INET_DIAG_BBRINFO;
914 		return sizeof(info->bbr);
915 	}
916 	return 0;
917 }
918 
919 static void bbr_set_state(struct sock *sk, u8 new_state)
920 {
921 	struct bbr *bbr = inet_csk_ca(sk);
922 
923 	if (new_state == TCP_CA_Loss) {
924 		struct rate_sample rs = { .losses = 1 };
925 
926 		bbr->prev_ca_state = TCP_CA_Loss;
927 		bbr->full_bw = 0;
928 		bbr->round_start = 1;	/* treat RTO like end of a round */
929 		bbr_lt_bw_sampling(sk, &rs);
930 	}
931 }
932 
933 static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
934 	.flags		= TCP_CONG_NON_RESTRICTED,
935 	.name		= "bbr",
936 	.owner		= THIS_MODULE,
937 	.init		= bbr_init,
938 	.cong_control	= bbr_main,
939 	.sndbuf_expand	= bbr_sndbuf_expand,
940 	.undo_cwnd	= bbr_undo_cwnd,
941 	.cwnd_event	= bbr_cwnd_event,
942 	.ssthresh	= bbr_ssthresh,
943 	.min_tso_segs	= bbr_min_tso_segs,
944 	.get_info	= bbr_get_info,
945 	.set_state	= bbr_set_state,
946 };
947 
948 static int __init bbr_register(void)
949 {
950 	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
951 	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
952 }
953 
954 static void __exit bbr_unregister(void)
955 {
956 	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
957 }
958 
959 module_init(bbr_register);
960 module_exit(bbr_unregister);
961 
962 MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
963 MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
964 MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
965 MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
966 MODULE_LICENSE("Dual BSD/GPL");
967 MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");
968