xref: /openbmc/linux/net/ipv4/tcp_bbr.c (revision f87deada)
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 		tso_segs_goal:7,     /* segments we want in each skb we send */
101 		idle_restart:1,	     /* restarting after idle? */
102 		probe_rtt_round_done:1,  /* a BBR_PROBE_RTT round at 4 pkts? */
103 		unused:5,
104 		lt_is_sampling:1,    /* taking long-term ("LT") samples now? */
105 		lt_rtt_cnt:7,	     /* round trips in long-term interval */
106 		lt_use_bw:1;	     /* use lt_bw as our bw estimate? */
107 	u32	lt_bw;		     /* LT est delivery rate in pkts/uS << 24 */
108 	u32	lt_last_delivered;   /* LT intvl start: tp->delivered */
109 	u32	lt_last_stamp;	     /* LT intvl start: tp->delivered_mstamp */
110 	u32	lt_last_lost;	     /* LT intvl start: tp->lost */
111 	u32	pacing_gain:10,	/* current gain for setting pacing rate */
112 		cwnd_gain:10,	/* current gain for setting cwnd */
113 		full_bw_reached:1,   /* reached full bw in Startup? */
114 		full_bw_cnt:2,	/* number of rounds without large bw gains */
115 		cycle_idx:3,	/* current index in pacing_gain cycle array */
116 		has_seen_rtt:1, /* have we seen an RTT sample yet? */
117 		unused_b:5;
118 	u32	prior_cwnd;	/* prior cwnd upon entering loss recovery */
119 	u32	full_bw;	/* recent bw, to estimate if pipe is full */
120 };
121 
122 #define CYCLE_LEN	8	/* number of phases in a pacing gain cycle */
123 
124 /* Window length of bw filter (in rounds): */
125 static const int bbr_bw_rtts = CYCLE_LEN + 2;
126 /* Window length of min_rtt filter (in sec): */
127 static const u32 bbr_min_rtt_win_sec = 10;
128 /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
129 static const u32 bbr_probe_rtt_mode_ms = 200;
130 /* Skip TSO below the following bandwidth (bits/sec): */
131 static const int bbr_min_tso_rate = 1200000;
132 
133 /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
134  * that will allow a smoothly increasing pacing rate that will double each RTT
135  * and send the same number of packets per RTT that an un-paced, slow-starting
136  * Reno or CUBIC flow would:
137  */
138 static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
139 /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
140  * the queue created in BBR_STARTUP in a single round:
141  */
142 static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
143 /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
144 static const int bbr_cwnd_gain  = BBR_UNIT * 2;
145 /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
146 static const int bbr_pacing_gain[] = {
147 	BBR_UNIT * 5 / 4,	/* probe for more available bw */
148 	BBR_UNIT * 3 / 4,	/* drain queue and/or yield bw to other flows */
149 	BBR_UNIT, BBR_UNIT, BBR_UNIT,	/* cruise at 1.0*bw to utilize pipe, */
150 	BBR_UNIT, BBR_UNIT, BBR_UNIT	/* without creating excess queue... */
151 };
152 /* Randomize the starting gain cycling phase over N phases: */
153 static const u32 bbr_cycle_rand = 7;
154 
155 /* Try to keep at least this many packets in flight, if things go smoothly. For
156  * smooth functioning, a sliding window protocol ACKing every other packet
157  * needs at least 4 packets in flight:
158  */
159 static const u32 bbr_cwnd_min_target = 4;
160 
161 /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
162 /* If bw has increased significantly (1.25x), there may be more bw available: */
163 static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
164 /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
165 static const u32 bbr_full_bw_cnt = 3;
166 
167 /* "long-term" ("LT") bandwidth estimator parameters... */
168 /* The minimum number of rounds in an LT bw sampling interval: */
169 static const u32 bbr_lt_intvl_min_rtts = 4;
170 /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
171 static const u32 bbr_lt_loss_thresh = 50;
172 /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
173 static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
174 /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
175 static const u32 bbr_lt_bw_diff = 4000 / 8;
176 /* If we estimate we're policed, use lt_bw for this many round trips: */
177 static const u32 bbr_lt_bw_max_rtts = 48;
178 
179 /* Do we estimate that STARTUP filled the pipe? */
180 static bool bbr_full_bw_reached(const struct sock *sk)
181 {
182 	const struct bbr *bbr = inet_csk_ca(sk);
183 
184 	return bbr->full_bw_reached;
185 }
186 
187 /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
188 static u32 bbr_max_bw(const struct sock *sk)
189 {
190 	struct bbr *bbr = inet_csk_ca(sk);
191 
192 	return minmax_get(&bbr->bw);
193 }
194 
195 /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
196 static u32 bbr_bw(const struct sock *sk)
197 {
198 	struct bbr *bbr = inet_csk_ca(sk);
199 
200 	return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
201 }
202 
203 /* Return rate in bytes per second, optionally with a gain.
204  * The order here is chosen carefully to avoid overflow of u64. This should
205  * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
206  */
207 static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
208 {
209 	rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache);
210 	rate *= gain;
211 	rate >>= BBR_SCALE;
212 	rate *= USEC_PER_SEC;
213 	return rate >> BW_SCALE;
214 }
215 
216 /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
217 static u32 bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
218 {
219 	u64 rate = bw;
220 
221 	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
222 	rate = min_t(u64, rate, sk->sk_max_pacing_rate);
223 	return rate;
224 }
225 
226 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
227 static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
228 {
229 	struct tcp_sock *tp = tcp_sk(sk);
230 	struct bbr *bbr = inet_csk_ca(sk);
231 	u64 bw;
232 	u32 rtt_us;
233 
234 	if (tp->srtt_us) {		/* any RTT sample yet? */
235 		rtt_us = max(tp->srtt_us >> 3, 1U);
236 		bbr->has_seen_rtt = 1;
237 	} else {			 /* no RTT sample yet */
238 		rtt_us = USEC_PER_MSEC;	 /* use nominal default RTT */
239 	}
240 	bw = (u64)tp->snd_cwnd * BW_UNIT;
241 	do_div(bw, rtt_us);
242 	sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain);
243 }
244 
245 /* Pace using current bw estimate and a gain factor. In order to help drive the
246  * network toward lower queues while maintaining high utilization and low
247  * latency, the average pacing rate aims to be slightly (~1%) lower than the
248  * estimated bandwidth. This is an important aspect of the design. In this
249  * implementation this slightly lower pacing rate is achieved implicitly by not
250  * including link-layer headers in the packet size used for the pacing rate.
251  */
252 static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
253 {
254 	struct tcp_sock *tp = tcp_sk(sk);
255 	struct bbr *bbr = inet_csk_ca(sk);
256 	u32 rate = bbr_bw_to_pacing_rate(sk, bw, gain);
257 
258 	if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
259 		bbr_init_pacing_rate_from_rtt(sk);
260 	if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate)
261 		sk->sk_pacing_rate = rate;
262 }
263 
264 /* Return count of segments we want in the skbs we send, or 0 for default. */
265 static u32 bbr_tso_segs_goal(struct sock *sk)
266 {
267 	struct bbr *bbr = inet_csk_ca(sk);
268 
269 	return bbr->tso_segs_goal;
270 }
271 
272 static void bbr_set_tso_segs_goal(struct sock *sk)
273 {
274 	struct tcp_sock *tp = tcp_sk(sk);
275 	struct bbr *bbr = inet_csk_ca(sk);
276 	u32 min_segs;
277 
278 	min_segs = sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
279 	bbr->tso_segs_goal = min(tcp_tso_autosize(sk, tp->mss_cache, min_segs),
280 				 0x7FU);
281 }
282 
283 /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
284 static void bbr_save_cwnd(struct sock *sk)
285 {
286 	struct tcp_sock *tp = tcp_sk(sk);
287 	struct bbr *bbr = inet_csk_ca(sk);
288 
289 	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
290 		bbr->prior_cwnd = tp->snd_cwnd;  /* this cwnd is good enough */
291 	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
292 		bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
293 }
294 
295 static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
296 {
297 	struct tcp_sock *tp = tcp_sk(sk);
298 	struct bbr *bbr = inet_csk_ca(sk);
299 
300 	if (event == CA_EVENT_TX_START && tp->app_limited) {
301 		bbr->idle_restart = 1;
302 		/* Avoid pointless buffer overflows: pace at est. bw if we don't
303 		 * need more speed (we're restarting from idle and app-limited).
304 		 */
305 		if (bbr->mode == BBR_PROBE_BW)
306 			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
307 	}
308 }
309 
310 /* Find target cwnd. Right-size the cwnd based on min RTT and the
311  * estimated bottleneck bandwidth:
312  *
313  * cwnd = bw * min_rtt * gain = BDP * gain
314  *
315  * The key factor, gain, controls the amount of queue. While a small gain
316  * builds a smaller queue, it becomes more vulnerable to noise in RTT
317  * measurements (e.g., delayed ACKs or other ACK compression effects). This
318  * noise may cause BBR to under-estimate the rate.
319  *
320  * To achieve full performance in high-speed paths, we budget enough cwnd to
321  * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
322  *   - one skb in sending host Qdisc,
323  *   - one skb in sending host TSO/GSO engine
324  *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
325  * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
326  * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
327  * which allows 2 outstanding 2-packet sequences, to try to keep pipe
328  * full even with ACK-every-other-packet delayed ACKs.
329  */
330 static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain)
331 {
332 	struct bbr *bbr = inet_csk_ca(sk);
333 	u32 cwnd;
334 	u64 w;
335 
336 	/* If we've never had a valid RTT sample, cap cwnd at the initial
337 	 * default. This should only happen when the connection is not using TCP
338 	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
339 	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
340 	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
341 	 */
342 	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
343 		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
344 
345 	w = (u64)bw * bbr->min_rtt_us;
346 
347 	/* Apply a gain to the given value, then remove the BW_SCALE shift. */
348 	cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
349 
350 	/* Allow enough full-sized skbs in flight to utilize end systems. */
351 	cwnd += 3 * bbr->tso_segs_goal;
352 
353 	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
354 	cwnd = (cwnd + 1) & ~1U;
355 
356 	return cwnd;
357 }
358 
359 /* An optimization in BBR to reduce losses: On the first round of recovery, we
360  * follow the packet conservation principle: send P packets per P packets acked.
361  * After that, we slow-start and send at most 2*P packets per P packets acked.
362  * After recovery finishes, or upon undo, we restore the cwnd we had when
363  * recovery started (capped by the target cwnd based on estimated BDP).
364  *
365  * TODO(ycheng/ncardwell): implement a rate-based approach.
366  */
367 static bool bbr_set_cwnd_to_recover_or_restore(
368 	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
369 {
370 	struct tcp_sock *tp = tcp_sk(sk);
371 	struct bbr *bbr = inet_csk_ca(sk);
372 	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
373 	u32 cwnd = tp->snd_cwnd;
374 
375 	/* An ACK for P pkts should release at most 2*P packets. We do this
376 	 * in two steps. First, here we deduct the number of lost packets.
377 	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
378 	 */
379 	if (rs->losses > 0)
380 		cwnd = max_t(s32, cwnd - rs->losses, 1);
381 
382 	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
383 		/* Starting 1st round of Recovery, so do packet conservation. */
384 		bbr->packet_conservation = 1;
385 		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
386 		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
387 		cwnd = tcp_packets_in_flight(tp) + acked;
388 	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
389 		/* Exiting loss recovery; restore cwnd saved before recovery. */
390 		bbr->restore_cwnd = 1;
391 		bbr->packet_conservation = 0;
392 	}
393 	bbr->prev_ca_state = state;
394 
395 	if (bbr->restore_cwnd) {
396 		/* Restore cwnd after exiting loss recovery or PROBE_RTT. */
397 		cwnd = max(cwnd, bbr->prior_cwnd);
398 		bbr->restore_cwnd = 0;
399 	}
400 
401 	if (bbr->packet_conservation) {
402 		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
403 		return true;	/* yes, using packet conservation */
404 	}
405 	*new_cwnd = cwnd;
406 	return false;
407 }
408 
409 /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
410  * has drawn us down below target), or snap down to target if we're above it.
411  */
412 static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
413 			 u32 acked, u32 bw, int gain)
414 {
415 	struct tcp_sock *tp = tcp_sk(sk);
416 	struct bbr *bbr = inet_csk_ca(sk);
417 	u32 cwnd = 0, target_cwnd = 0;
418 
419 	if (!acked)
420 		return;
421 
422 	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
423 		goto done;
424 
425 	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
426 	target_cwnd = bbr_target_cwnd(sk, bw, gain);
427 	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
428 		cwnd = min(cwnd + acked, target_cwnd);
429 	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
430 		cwnd = cwnd + acked;
431 	cwnd = max(cwnd, bbr_cwnd_min_target);
432 
433 done:
434 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);	/* apply global cap */
435 	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
436 		tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
437 }
438 
439 /* End cycle phase if it's time and/or we hit the phase's in-flight target. */
440 static bool bbr_is_next_cycle_phase(struct sock *sk,
441 				    const struct rate_sample *rs)
442 {
443 	struct tcp_sock *tp = tcp_sk(sk);
444 	struct bbr *bbr = inet_csk_ca(sk);
445 	bool is_full_length =
446 		tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
447 		bbr->min_rtt_us;
448 	u32 inflight, bw;
449 
450 	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
451 	 * use the pipe without increasing the queue.
452 	 */
453 	if (bbr->pacing_gain == BBR_UNIT)
454 		return is_full_length;		/* just use wall clock time */
455 
456 	inflight = rs->prior_in_flight;  /* what was in-flight before ACK? */
457 	bw = bbr_max_bw(sk);
458 
459 	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
460 	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
461 	 * small (e.g. on a LAN). We do not persist if packets are lost, since
462 	 * a path with small buffers may not hold that much.
463 	 */
464 	if (bbr->pacing_gain > BBR_UNIT)
465 		return is_full_length &&
466 			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
467 			 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain));
468 
469 	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
470 	 * probing didn't find more bw. If inflight falls to match BDP then we
471 	 * estimate queue is drained; persisting would underutilize the pipe.
472 	 */
473 	return is_full_length ||
474 		inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT);
475 }
476 
477 static void bbr_advance_cycle_phase(struct sock *sk)
478 {
479 	struct tcp_sock *tp = tcp_sk(sk);
480 	struct bbr *bbr = inet_csk_ca(sk);
481 
482 	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
483 	bbr->cycle_mstamp = tp->delivered_mstamp;
484 	bbr->pacing_gain = bbr->lt_use_bw ? BBR_UNIT :
485 					    bbr_pacing_gain[bbr->cycle_idx];
486 }
487 
488 /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
489 static void bbr_update_cycle_phase(struct sock *sk,
490 				   const struct rate_sample *rs)
491 {
492 	struct bbr *bbr = inet_csk_ca(sk);
493 
494 	if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
495 		bbr_advance_cycle_phase(sk);
496 }
497 
498 static void bbr_reset_startup_mode(struct sock *sk)
499 {
500 	struct bbr *bbr = inet_csk_ca(sk);
501 
502 	bbr->mode = BBR_STARTUP;
503 	bbr->pacing_gain = bbr_high_gain;
504 	bbr->cwnd_gain	 = bbr_high_gain;
505 }
506 
507 static void bbr_reset_probe_bw_mode(struct sock *sk)
508 {
509 	struct bbr *bbr = inet_csk_ca(sk);
510 
511 	bbr->mode = BBR_PROBE_BW;
512 	bbr->pacing_gain = BBR_UNIT;
513 	bbr->cwnd_gain = bbr_cwnd_gain;
514 	bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
515 	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
516 }
517 
518 static void bbr_reset_mode(struct sock *sk)
519 {
520 	if (!bbr_full_bw_reached(sk))
521 		bbr_reset_startup_mode(sk);
522 	else
523 		bbr_reset_probe_bw_mode(sk);
524 }
525 
526 /* Start a new long-term sampling interval. */
527 static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
528 {
529 	struct tcp_sock *tp = tcp_sk(sk);
530 	struct bbr *bbr = inet_csk_ca(sk);
531 
532 	bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
533 	bbr->lt_last_delivered = tp->delivered;
534 	bbr->lt_last_lost = tp->lost;
535 	bbr->lt_rtt_cnt = 0;
536 }
537 
538 /* Completely reset long-term bandwidth sampling. */
539 static void bbr_reset_lt_bw_sampling(struct sock *sk)
540 {
541 	struct bbr *bbr = inet_csk_ca(sk);
542 
543 	bbr->lt_bw = 0;
544 	bbr->lt_use_bw = 0;
545 	bbr->lt_is_sampling = false;
546 	bbr_reset_lt_bw_sampling_interval(sk);
547 }
548 
549 /* Long-term bw sampling interval is done. Estimate whether we're policed. */
550 static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
551 {
552 	struct bbr *bbr = inet_csk_ca(sk);
553 	u32 diff;
554 
555 	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
556 		/* Is new bw close to the lt_bw from the previous interval? */
557 		diff = abs(bw - bbr->lt_bw);
558 		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
559 		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
560 		     bbr_lt_bw_diff)) {
561 			/* All criteria are met; estimate we're policed. */
562 			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
563 			bbr->lt_use_bw = 1;
564 			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
565 			bbr->lt_rtt_cnt = 0;
566 			return;
567 		}
568 	}
569 	bbr->lt_bw = bw;
570 	bbr_reset_lt_bw_sampling_interval(sk);
571 }
572 
573 /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
574  * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
575  * explicitly models their policed rate, to reduce unnecessary losses. We
576  * estimate that we're policed if we see 2 consecutive sampling intervals with
577  * consistent throughput and high packet loss. If we think we're being policed,
578  * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
579  */
580 static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
581 {
582 	struct tcp_sock *tp = tcp_sk(sk);
583 	struct bbr *bbr = inet_csk_ca(sk);
584 	u32 lost, delivered;
585 	u64 bw;
586 	u32 t;
587 
588 	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
589 		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
590 		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
591 			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
592 			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
593 		}
594 		return;
595 	}
596 
597 	/* Wait for the first loss before sampling, to let the policer exhaust
598 	 * its tokens and estimate the steady-state rate allowed by the policer.
599 	 * Starting samples earlier includes bursts that over-estimate the bw.
600 	 */
601 	if (!bbr->lt_is_sampling) {
602 		if (!rs->losses)
603 			return;
604 		bbr_reset_lt_bw_sampling_interval(sk);
605 		bbr->lt_is_sampling = true;
606 	}
607 
608 	/* To avoid underestimates, reset sampling if we run out of data. */
609 	if (rs->is_app_limited) {
610 		bbr_reset_lt_bw_sampling(sk);
611 		return;
612 	}
613 
614 	if (bbr->round_start)
615 		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
616 	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
617 		return;		/* sampling interval needs to be longer */
618 	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
619 		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
620 		return;
621 	}
622 
623 	/* End sampling interval when a packet is lost, so we estimate the
624 	 * policer tokens were exhausted. Stopping the sampling before the
625 	 * tokens are exhausted under-estimates the policed rate.
626 	 */
627 	if (!rs->losses)
628 		return;
629 
630 	/* Calculate packets lost and delivered in sampling interval. */
631 	lost = tp->lost - bbr->lt_last_lost;
632 	delivered = tp->delivered - bbr->lt_last_delivered;
633 	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
634 	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
635 		return;
636 
637 	/* Find average delivery rate in this sampling interval. */
638 	t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
639 	if ((s32)t < 1)
640 		return;		/* interval is less than one ms, so wait */
641 	/* Check if can multiply without overflow */
642 	if (t >= ~0U / USEC_PER_MSEC) {
643 		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
644 		return;
645 	}
646 	t *= USEC_PER_MSEC;
647 	bw = (u64)delivered * BW_UNIT;
648 	do_div(bw, t);
649 	bbr_lt_bw_interval_done(sk, bw);
650 }
651 
652 /* Estimate the bandwidth based on how fast packets are delivered */
653 static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
654 {
655 	struct tcp_sock *tp = tcp_sk(sk);
656 	struct bbr *bbr = inet_csk_ca(sk);
657 	u64 bw;
658 
659 	bbr->round_start = 0;
660 	if (rs->delivered < 0 || rs->interval_us <= 0)
661 		return; /* Not a valid observation */
662 
663 	/* See if we've reached the next RTT */
664 	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
665 		bbr->next_rtt_delivered = tp->delivered;
666 		bbr->rtt_cnt++;
667 		bbr->round_start = 1;
668 		bbr->packet_conservation = 0;
669 	}
670 
671 	bbr_lt_bw_sampling(sk, rs);
672 
673 	/* Divide delivered by the interval to find a (lower bound) bottleneck
674 	 * bandwidth sample. Delivered is in packets and interval_us in uS and
675 	 * ratio will be <<1 for most connections. So delivered is first scaled.
676 	 */
677 	bw = (u64)rs->delivered * BW_UNIT;
678 	do_div(bw, rs->interval_us);
679 
680 	/* If this sample is application-limited, it is likely to have a very
681 	 * low delivered count that represents application behavior rather than
682 	 * the available network rate. Such a sample could drag down estimated
683 	 * bw, causing needless slow-down. Thus, to continue to send at the
684 	 * last measured network rate, we filter out app-limited samples unless
685 	 * they describe the path bw at least as well as our bw model.
686 	 *
687 	 * So the goal during app-limited phase is to proceed with the best
688 	 * network rate no matter how long. We automatically leave this
689 	 * phase when app writes faster than the network can deliver :)
690 	 */
691 	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
692 		/* Incorporate new sample into our max bw filter. */
693 		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
694 	}
695 }
696 
697 /* Estimate when the pipe is full, using the change in delivery rate: BBR
698  * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
699  * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
700  * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
701  * higher rwin, 3: we get higher delivery rate samples. Or transient
702  * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
703  * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
704  */
705 static void bbr_check_full_bw_reached(struct sock *sk,
706 				      const struct rate_sample *rs)
707 {
708 	struct bbr *bbr = inet_csk_ca(sk);
709 	u32 bw_thresh;
710 
711 	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
712 		return;
713 
714 	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
715 	if (bbr_max_bw(sk) >= bw_thresh) {
716 		bbr->full_bw = bbr_max_bw(sk);
717 		bbr->full_bw_cnt = 0;
718 		return;
719 	}
720 	++bbr->full_bw_cnt;
721 	bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
722 }
723 
724 /* If pipe is probably full, drain the queue and then enter steady-state. */
725 static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
726 {
727 	struct bbr *bbr = inet_csk_ca(sk);
728 
729 	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
730 		bbr->mode = BBR_DRAIN;	/* drain queue we created */
731 		bbr->pacing_gain = bbr_drain_gain;	/* pace slow to drain */
732 		bbr->cwnd_gain = bbr_high_gain;	/* maintain cwnd */
733 	}	/* fall through to check if in-flight is already small: */
734 	if (bbr->mode == BBR_DRAIN &&
735 	    tcp_packets_in_flight(tcp_sk(sk)) <=
736 	    bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT))
737 		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
738 }
739 
740 /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
741  * periodically drain the bottleneck queue, to converge to measure the true
742  * min_rtt (unloaded propagation delay). This allows the flows to keep queues
743  * small (reducing queuing delay and packet loss) and achieve fairness among
744  * BBR flows.
745  *
746  * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
747  * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
748  * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
749  * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
750  * re-enter the previous mode. BBR uses 200ms to approximately bound the
751  * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
752  *
753  * Note that flows need only pay 2% if they are busy sending over the last 10
754  * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
755  * natural silences or low-rate periods within 10 seconds where the rate is low
756  * enough for long enough to drain its queue in the bottleneck. We pick up
757  * these min RTT measurements opportunistically with our min_rtt filter. :-)
758  */
759 static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
760 {
761 	struct tcp_sock *tp = tcp_sk(sk);
762 	struct bbr *bbr = inet_csk_ca(sk);
763 	bool filter_expired;
764 
765 	/* Track min RTT seen in the min_rtt_win_sec filter window: */
766 	filter_expired = after(tcp_jiffies32,
767 			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
768 	if (rs->rtt_us >= 0 &&
769 	    (rs->rtt_us <= bbr->min_rtt_us ||
770 	     (filter_expired && !rs->is_ack_delayed))) {
771 		bbr->min_rtt_us = rs->rtt_us;
772 		bbr->min_rtt_stamp = tcp_jiffies32;
773 	}
774 
775 	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
776 	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
777 		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
778 		bbr->pacing_gain = BBR_UNIT;
779 		bbr->cwnd_gain = BBR_UNIT;
780 		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
781 		bbr->probe_rtt_done_stamp = 0;
782 	}
783 
784 	if (bbr->mode == BBR_PROBE_RTT) {
785 		/* Ignore low rate samples during this mode. */
786 		tp->app_limited =
787 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
788 		/* Maintain min packets in flight for max(200 ms, 1 round). */
789 		if (!bbr->probe_rtt_done_stamp &&
790 		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
791 			bbr->probe_rtt_done_stamp = tcp_jiffies32 +
792 				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
793 			bbr->probe_rtt_round_done = 0;
794 			bbr->next_rtt_delivered = tp->delivered;
795 		} else if (bbr->probe_rtt_done_stamp) {
796 			if (bbr->round_start)
797 				bbr->probe_rtt_round_done = 1;
798 			if (bbr->probe_rtt_round_done &&
799 			    after(tcp_jiffies32, bbr->probe_rtt_done_stamp)) {
800 				bbr->min_rtt_stamp = tcp_jiffies32;
801 				bbr->restore_cwnd = 1;  /* snap to prior_cwnd */
802 				bbr_reset_mode(sk);
803 			}
804 		}
805 	}
806 	bbr->idle_restart = 0;
807 }
808 
809 static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
810 {
811 	bbr_update_bw(sk, rs);
812 	bbr_update_cycle_phase(sk, rs);
813 	bbr_check_full_bw_reached(sk, rs);
814 	bbr_check_drain(sk, rs);
815 	bbr_update_min_rtt(sk, rs);
816 }
817 
818 static void bbr_main(struct sock *sk, const struct rate_sample *rs)
819 {
820 	struct bbr *bbr = inet_csk_ca(sk);
821 	u32 bw;
822 
823 	bbr_update_model(sk, rs);
824 
825 	bw = bbr_bw(sk);
826 	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
827 	bbr_set_tso_segs_goal(sk);
828 	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
829 }
830 
831 static void bbr_init(struct sock *sk)
832 {
833 	struct tcp_sock *tp = tcp_sk(sk);
834 	struct bbr *bbr = inet_csk_ca(sk);
835 
836 	bbr->prior_cwnd = 0;
837 	bbr->tso_segs_goal = 0;	 /* default segs per skb until first ACK */
838 	bbr->rtt_cnt = 0;
839 	bbr->next_rtt_delivered = 0;
840 	bbr->prev_ca_state = TCP_CA_Open;
841 	bbr->packet_conservation = 0;
842 
843 	bbr->probe_rtt_done_stamp = 0;
844 	bbr->probe_rtt_round_done = 0;
845 	bbr->min_rtt_us = tcp_min_rtt(tp);
846 	bbr->min_rtt_stamp = tcp_jiffies32;
847 
848 	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
849 
850 	bbr->has_seen_rtt = 0;
851 	bbr_init_pacing_rate_from_rtt(sk);
852 
853 	bbr->restore_cwnd = 0;
854 	bbr->round_start = 0;
855 	bbr->idle_restart = 0;
856 	bbr->full_bw_reached = 0;
857 	bbr->full_bw = 0;
858 	bbr->full_bw_cnt = 0;
859 	bbr->cycle_mstamp = 0;
860 	bbr->cycle_idx = 0;
861 	bbr_reset_lt_bw_sampling(sk);
862 	bbr_reset_startup_mode(sk);
863 
864 	cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
865 }
866 
867 static u32 bbr_sndbuf_expand(struct sock *sk)
868 {
869 	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
870 	return 3;
871 }
872 
873 /* In theory BBR does not need to undo the cwnd since it does not
874  * always reduce cwnd on losses (see bbr_main()). Keep it for now.
875  */
876 static u32 bbr_undo_cwnd(struct sock *sk)
877 {
878 	struct bbr *bbr = inet_csk_ca(sk);
879 
880 	bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
881 	bbr->full_bw_cnt = 0;
882 	bbr_reset_lt_bw_sampling(sk);
883 	return tcp_sk(sk)->snd_cwnd;
884 }
885 
886 /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
887 static u32 bbr_ssthresh(struct sock *sk)
888 {
889 	bbr_save_cwnd(sk);
890 	return TCP_INFINITE_SSTHRESH;	 /* BBR does not use ssthresh */
891 }
892 
893 static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
894 			   union tcp_cc_info *info)
895 {
896 	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
897 	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
898 		struct tcp_sock *tp = tcp_sk(sk);
899 		struct bbr *bbr = inet_csk_ca(sk);
900 		u64 bw = bbr_bw(sk);
901 
902 		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
903 		memset(&info->bbr, 0, sizeof(info->bbr));
904 		info->bbr.bbr_bw_lo		= (u32)bw;
905 		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
906 		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
907 		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
908 		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
909 		*attr = INET_DIAG_BBRINFO;
910 		return sizeof(info->bbr);
911 	}
912 	return 0;
913 }
914 
915 static void bbr_set_state(struct sock *sk, u8 new_state)
916 {
917 	struct bbr *bbr = inet_csk_ca(sk);
918 
919 	if (new_state == TCP_CA_Loss) {
920 		struct rate_sample rs = { .losses = 1 };
921 
922 		bbr->prev_ca_state = TCP_CA_Loss;
923 		bbr->full_bw = 0;
924 		bbr->round_start = 1;	/* treat RTO like end of a round */
925 		bbr_lt_bw_sampling(sk, &rs);
926 	}
927 }
928 
929 static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
930 	.flags		= TCP_CONG_NON_RESTRICTED,
931 	.name		= "bbr",
932 	.owner		= THIS_MODULE,
933 	.init		= bbr_init,
934 	.cong_control	= bbr_main,
935 	.sndbuf_expand	= bbr_sndbuf_expand,
936 	.undo_cwnd	= bbr_undo_cwnd,
937 	.cwnd_event	= bbr_cwnd_event,
938 	.ssthresh	= bbr_ssthresh,
939 	.tso_segs_goal	= bbr_tso_segs_goal,
940 	.get_info	= bbr_get_info,
941 	.set_state	= bbr_set_state,
942 };
943 
944 static int __init bbr_register(void)
945 {
946 	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
947 	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
948 }
949 
950 static void __exit bbr_unregister(void)
951 {
952 	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
953 }
954 
955 module_init(bbr_register);
956 module_exit(bbr_unregister);
957 
958 MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
959 MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
960 MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
961 MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
962 MODULE_LICENSE("Dual BSD/GPL");
963 MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");
964