xref: /openbmc/linux/net/ipv4/tcp_bbr.c (revision 9be08a27)
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 	unsigned int mss = tcp_sk(sk)->mss_cache;
209 
210 	if (!tcp_needs_internal_pacing(sk))
211 		mss = tcp_mss_to_mtu(sk, mss);
212 	rate *= mss;
213 	rate *= gain;
214 	rate >>= BBR_SCALE;
215 	rate *= USEC_PER_SEC;
216 	return rate >> BW_SCALE;
217 }
218 
219 /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
220 static u32 bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
221 {
222 	u64 rate = bw;
223 
224 	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
225 	rate = min_t(u64, rate, sk->sk_max_pacing_rate);
226 	return rate;
227 }
228 
229 /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
230 static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
231 {
232 	struct tcp_sock *tp = tcp_sk(sk);
233 	struct bbr *bbr = inet_csk_ca(sk);
234 	u64 bw;
235 	u32 rtt_us;
236 
237 	if (tp->srtt_us) {		/* any RTT sample yet? */
238 		rtt_us = max(tp->srtt_us >> 3, 1U);
239 		bbr->has_seen_rtt = 1;
240 	} else {			 /* no RTT sample yet */
241 		rtt_us = USEC_PER_MSEC;	 /* use nominal default RTT */
242 	}
243 	bw = (u64)tp->snd_cwnd * BW_UNIT;
244 	do_div(bw, rtt_us);
245 	sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain);
246 }
247 
248 /* Pace using current bw estimate and a gain factor. In order to help drive the
249  * network toward lower queues while maintaining high utilization and low
250  * latency, the average pacing rate aims to be slightly (~1%) lower than the
251  * estimated bandwidth. This is an important aspect of the design. In this
252  * implementation this slightly lower pacing rate is achieved implicitly by not
253  * including link-layer headers in the packet size used for the pacing rate.
254  */
255 static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
256 {
257 	struct tcp_sock *tp = tcp_sk(sk);
258 	struct bbr *bbr = inet_csk_ca(sk);
259 	u32 rate = bbr_bw_to_pacing_rate(sk, bw, gain);
260 
261 	if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
262 		bbr_init_pacing_rate_from_rtt(sk);
263 	if (bbr_full_bw_reached(sk) || rate > sk->sk_pacing_rate)
264 		sk->sk_pacing_rate = rate;
265 }
266 
267 /* override sysctl_tcp_min_tso_segs */
268 static u32 bbr_min_tso_segs(struct sock *sk)
269 {
270 	return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
271 }
272 
273 static u32 bbr_tso_segs_goal(struct sock *sk)
274 {
275 	struct tcp_sock *tp = tcp_sk(sk);
276 	u32 segs, bytes;
277 
278 	/* Sort of tcp_tso_autosize() but ignoring
279 	 * driver provided sk_gso_max_size.
280 	 */
281 	bytes = min_t(u32, sk->sk_pacing_rate >> sk->sk_pacing_shift,
282 		      GSO_MAX_SIZE - 1 - MAX_TCP_HEADER);
283 	segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk));
284 
285 	return min(segs, 0x7FU);
286 }
287 
288 /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
289 static void bbr_save_cwnd(struct sock *sk)
290 {
291 	struct tcp_sock *tp = tcp_sk(sk);
292 	struct bbr *bbr = inet_csk_ca(sk);
293 
294 	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
295 		bbr->prior_cwnd = tp->snd_cwnd;  /* this cwnd is good enough */
296 	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
297 		bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
298 }
299 
300 static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
301 {
302 	struct tcp_sock *tp = tcp_sk(sk);
303 	struct bbr *bbr = inet_csk_ca(sk);
304 
305 	if (event == CA_EVENT_TX_START && tp->app_limited) {
306 		bbr->idle_restart = 1;
307 		/* Avoid pointless buffer overflows: pace at est. bw if we don't
308 		 * need more speed (we're restarting from idle and app-limited).
309 		 */
310 		if (bbr->mode == BBR_PROBE_BW)
311 			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
312 	}
313 }
314 
315 /* Find target cwnd. Right-size the cwnd based on min RTT and the
316  * estimated bottleneck bandwidth:
317  *
318  * cwnd = bw * min_rtt * gain = BDP * gain
319  *
320  * The key factor, gain, controls the amount of queue. While a small gain
321  * builds a smaller queue, it becomes more vulnerable to noise in RTT
322  * measurements (e.g., delayed ACKs or other ACK compression effects). This
323  * noise may cause BBR to under-estimate the rate.
324  *
325  * To achieve full performance in high-speed paths, we budget enough cwnd to
326  * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
327  *   - one skb in sending host Qdisc,
328  *   - one skb in sending host TSO/GSO engine
329  *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
330  * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
331  * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
332  * which allows 2 outstanding 2-packet sequences, to try to keep pipe
333  * full even with ACK-every-other-packet delayed ACKs.
334  */
335 static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain)
336 {
337 	struct bbr *bbr = inet_csk_ca(sk);
338 	u32 cwnd;
339 	u64 w;
340 
341 	/* If we've never had a valid RTT sample, cap cwnd at the initial
342 	 * default. This should only happen when the connection is not using TCP
343 	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
344 	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
345 	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
346 	 */
347 	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
348 		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
349 
350 	w = (u64)bw * bbr->min_rtt_us;
351 
352 	/* Apply a gain to the given value, then remove the BW_SCALE shift. */
353 	cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
354 
355 	/* Allow enough full-sized skbs in flight to utilize end systems. */
356 	cwnd += 3 * bbr_tso_segs_goal(sk);
357 
358 	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
359 	cwnd = (cwnd + 1) & ~1U;
360 
361 	/* Ensure gain cycling gets inflight above BDP even for small BDPs. */
362 	if (bbr->mode == BBR_PROBE_BW && gain > BBR_UNIT)
363 		cwnd += 2;
364 
365 	return cwnd;
366 }
367 
368 /* An optimization in BBR to reduce losses: On the first round of recovery, we
369  * follow the packet conservation principle: send P packets per P packets acked.
370  * After that, we slow-start and send at most 2*P packets per P packets acked.
371  * After recovery finishes, or upon undo, we restore the cwnd we had when
372  * recovery started (capped by the target cwnd based on estimated BDP).
373  *
374  * TODO(ycheng/ncardwell): implement a rate-based approach.
375  */
376 static bool bbr_set_cwnd_to_recover_or_restore(
377 	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
378 {
379 	struct tcp_sock *tp = tcp_sk(sk);
380 	struct bbr *bbr = inet_csk_ca(sk);
381 	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
382 	u32 cwnd = tp->snd_cwnd;
383 
384 	/* An ACK for P pkts should release at most 2*P packets. We do this
385 	 * in two steps. First, here we deduct the number of lost packets.
386 	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
387 	 */
388 	if (rs->losses > 0)
389 		cwnd = max_t(s32, cwnd - rs->losses, 1);
390 
391 	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
392 		/* Starting 1st round of Recovery, so do packet conservation. */
393 		bbr->packet_conservation = 1;
394 		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
395 		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
396 		cwnd = tcp_packets_in_flight(tp) + acked;
397 	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
398 		/* Exiting loss recovery; restore cwnd saved before recovery. */
399 		bbr->restore_cwnd = 1;
400 		bbr->packet_conservation = 0;
401 	}
402 	bbr->prev_ca_state = state;
403 
404 	if (bbr->restore_cwnd) {
405 		/* Restore cwnd after exiting loss recovery or PROBE_RTT. */
406 		cwnd = max(cwnd, bbr->prior_cwnd);
407 		bbr->restore_cwnd = 0;
408 	}
409 
410 	if (bbr->packet_conservation) {
411 		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
412 		return true;	/* yes, using packet conservation */
413 	}
414 	*new_cwnd = cwnd;
415 	return false;
416 }
417 
418 /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
419  * has drawn us down below target), or snap down to target if we're above it.
420  */
421 static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
422 			 u32 acked, u32 bw, int gain)
423 {
424 	struct tcp_sock *tp = tcp_sk(sk);
425 	struct bbr *bbr = inet_csk_ca(sk);
426 	u32 cwnd = 0, target_cwnd = 0;
427 
428 	if (!acked)
429 		return;
430 
431 	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
432 		goto done;
433 
434 	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
435 	target_cwnd = bbr_target_cwnd(sk, bw, gain);
436 	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
437 		cwnd = min(cwnd + acked, target_cwnd);
438 	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
439 		cwnd = cwnd + acked;
440 	cwnd = max(cwnd, bbr_cwnd_min_target);
441 
442 done:
443 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);	/* apply global cap */
444 	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
445 		tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
446 }
447 
448 /* End cycle phase if it's time and/or we hit the phase's in-flight target. */
449 static bool bbr_is_next_cycle_phase(struct sock *sk,
450 				    const struct rate_sample *rs)
451 {
452 	struct tcp_sock *tp = tcp_sk(sk);
453 	struct bbr *bbr = inet_csk_ca(sk);
454 	bool is_full_length =
455 		tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
456 		bbr->min_rtt_us;
457 	u32 inflight, bw;
458 
459 	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
460 	 * use the pipe without increasing the queue.
461 	 */
462 	if (bbr->pacing_gain == BBR_UNIT)
463 		return is_full_length;		/* just use wall clock time */
464 
465 	inflight = rs->prior_in_flight;  /* what was in-flight before ACK? */
466 	bw = bbr_max_bw(sk);
467 
468 	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
469 	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
470 	 * small (e.g. on a LAN). We do not persist if packets are lost, since
471 	 * a path with small buffers may not hold that much.
472 	 */
473 	if (bbr->pacing_gain > BBR_UNIT)
474 		return is_full_length &&
475 			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
476 			 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain));
477 
478 	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
479 	 * probing didn't find more bw. If inflight falls to match BDP then we
480 	 * estimate queue is drained; persisting would underutilize the pipe.
481 	 */
482 	return is_full_length ||
483 		inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT);
484 }
485 
486 static void bbr_advance_cycle_phase(struct sock *sk)
487 {
488 	struct tcp_sock *tp = tcp_sk(sk);
489 	struct bbr *bbr = inet_csk_ca(sk);
490 
491 	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
492 	bbr->cycle_mstamp = tp->delivered_mstamp;
493 	bbr->pacing_gain = bbr->lt_use_bw ? BBR_UNIT :
494 					    bbr_pacing_gain[bbr->cycle_idx];
495 }
496 
497 /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
498 static void bbr_update_cycle_phase(struct sock *sk,
499 				   const struct rate_sample *rs)
500 {
501 	struct bbr *bbr = inet_csk_ca(sk);
502 
503 	if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
504 		bbr_advance_cycle_phase(sk);
505 }
506 
507 static void bbr_reset_startup_mode(struct sock *sk)
508 {
509 	struct bbr *bbr = inet_csk_ca(sk);
510 
511 	bbr->mode = BBR_STARTUP;
512 	bbr->pacing_gain = bbr_high_gain;
513 	bbr->cwnd_gain	 = bbr_high_gain;
514 }
515 
516 static void bbr_reset_probe_bw_mode(struct sock *sk)
517 {
518 	struct bbr *bbr = inet_csk_ca(sk);
519 
520 	bbr->mode = BBR_PROBE_BW;
521 	bbr->pacing_gain = BBR_UNIT;
522 	bbr->cwnd_gain = bbr_cwnd_gain;
523 	bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
524 	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
525 }
526 
527 static void bbr_reset_mode(struct sock *sk)
528 {
529 	if (!bbr_full_bw_reached(sk))
530 		bbr_reset_startup_mode(sk);
531 	else
532 		bbr_reset_probe_bw_mode(sk);
533 }
534 
535 /* Start a new long-term sampling interval. */
536 static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
537 {
538 	struct tcp_sock *tp = tcp_sk(sk);
539 	struct bbr *bbr = inet_csk_ca(sk);
540 
541 	bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
542 	bbr->lt_last_delivered = tp->delivered;
543 	bbr->lt_last_lost = tp->lost;
544 	bbr->lt_rtt_cnt = 0;
545 }
546 
547 /* Completely reset long-term bandwidth sampling. */
548 static void bbr_reset_lt_bw_sampling(struct sock *sk)
549 {
550 	struct bbr *bbr = inet_csk_ca(sk);
551 
552 	bbr->lt_bw = 0;
553 	bbr->lt_use_bw = 0;
554 	bbr->lt_is_sampling = false;
555 	bbr_reset_lt_bw_sampling_interval(sk);
556 }
557 
558 /* Long-term bw sampling interval is done. Estimate whether we're policed. */
559 static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
560 {
561 	struct bbr *bbr = inet_csk_ca(sk);
562 	u32 diff;
563 
564 	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
565 		/* Is new bw close to the lt_bw from the previous interval? */
566 		diff = abs(bw - bbr->lt_bw);
567 		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
568 		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
569 		     bbr_lt_bw_diff)) {
570 			/* All criteria are met; estimate we're policed. */
571 			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
572 			bbr->lt_use_bw = 1;
573 			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
574 			bbr->lt_rtt_cnt = 0;
575 			return;
576 		}
577 	}
578 	bbr->lt_bw = bw;
579 	bbr_reset_lt_bw_sampling_interval(sk);
580 }
581 
582 /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
583  * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
584  * explicitly models their policed rate, to reduce unnecessary losses. We
585  * estimate that we're policed if we see 2 consecutive sampling intervals with
586  * consistent throughput and high packet loss. If we think we're being policed,
587  * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
588  */
589 static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
590 {
591 	struct tcp_sock *tp = tcp_sk(sk);
592 	struct bbr *bbr = inet_csk_ca(sk);
593 	u32 lost, delivered;
594 	u64 bw;
595 	u32 t;
596 
597 	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
598 		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
599 		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
600 			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
601 			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
602 		}
603 		return;
604 	}
605 
606 	/* Wait for the first loss before sampling, to let the policer exhaust
607 	 * its tokens and estimate the steady-state rate allowed by the policer.
608 	 * Starting samples earlier includes bursts that over-estimate the bw.
609 	 */
610 	if (!bbr->lt_is_sampling) {
611 		if (!rs->losses)
612 			return;
613 		bbr_reset_lt_bw_sampling_interval(sk);
614 		bbr->lt_is_sampling = true;
615 	}
616 
617 	/* To avoid underestimates, reset sampling if we run out of data. */
618 	if (rs->is_app_limited) {
619 		bbr_reset_lt_bw_sampling(sk);
620 		return;
621 	}
622 
623 	if (bbr->round_start)
624 		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
625 	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
626 		return;		/* sampling interval needs to be longer */
627 	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
628 		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
629 		return;
630 	}
631 
632 	/* End sampling interval when a packet is lost, so we estimate the
633 	 * policer tokens were exhausted. Stopping the sampling before the
634 	 * tokens are exhausted under-estimates the policed rate.
635 	 */
636 	if (!rs->losses)
637 		return;
638 
639 	/* Calculate packets lost and delivered in sampling interval. */
640 	lost = tp->lost - bbr->lt_last_lost;
641 	delivered = tp->delivered - bbr->lt_last_delivered;
642 	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
643 	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
644 		return;
645 
646 	/* Find average delivery rate in this sampling interval. */
647 	t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
648 	if ((s32)t < 1)
649 		return;		/* interval is less than one ms, so wait */
650 	/* Check if can multiply without overflow */
651 	if (t >= ~0U / USEC_PER_MSEC) {
652 		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
653 		return;
654 	}
655 	t *= USEC_PER_MSEC;
656 	bw = (u64)delivered * BW_UNIT;
657 	do_div(bw, t);
658 	bbr_lt_bw_interval_done(sk, bw);
659 }
660 
661 /* Estimate the bandwidth based on how fast packets are delivered */
662 static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
663 {
664 	struct tcp_sock *tp = tcp_sk(sk);
665 	struct bbr *bbr = inet_csk_ca(sk);
666 	u64 bw;
667 
668 	bbr->round_start = 0;
669 	if (rs->delivered < 0 || rs->interval_us <= 0)
670 		return; /* Not a valid observation */
671 
672 	/* See if we've reached the next RTT */
673 	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
674 		bbr->next_rtt_delivered = tp->delivered;
675 		bbr->rtt_cnt++;
676 		bbr->round_start = 1;
677 		bbr->packet_conservation = 0;
678 	}
679 
680 	bbr_lt_bw_sampling(sk, rs);
681 
682 	/* Divide delivered by the interval to find a (lower bound) bottleneck
683 	 * bandwidth sample. Delivered is in packets and interval_us in uS and
684 	 * ratio will be <<1 for most connections. So delivered is first scaled.
685 	 */
686 	bw = (u64)rs->delivered * BW_UNIT;
687 	do_div(bw, rs->interval_us);
688 
689 	/* If this sample is application-limited, it is likely to have a very
690 	 * low delivered count that represents application behavior rather than
691 	 * the available network rate. Such a sample could drag down estimated
692 	 * bw, causing needless slow-down. Thus, to continue to send at the
693 	 * last measured network rate, we filter out app-limited samples unless
694 	 * they describe the path bw at least as well as our bw model.
695 	 *
696 	 * So the goal during app-limited phase is to proceed with the best
697 	 * network rate no matter how long. We automatically leave this
698 	 * phase when app writes faster than the network can deliver :)
699 	 */
700 	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
701 		/* Incorporate new sample into our max bw filter. */
702 		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
703 	}
704 }
705 
706 /* Estimate when the pipe is full, using the change in delivery rate: BBR
707  * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
708  * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
709  * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
710  * higher rwin, 3: we get higher delivery rate samples. Or transient
711  * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
712  * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
713  */
714 static void bbr_check_full_bw_reached(struct sock *sk,
715 				      const struct rate_sample *rs)
716 {
717 	struct bbr *bbr = inet_csk_ca(sk);
718 	u32 bw_thresh;
719 
720 	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
721 		return;
722 
723 	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
724 	if (bbr_max_bw(sk) >= bw_thresh) {
725 		bbr->full_bw = bbr_max_bw(sk);
726 		bbr->full_bw_cnt = 0;
727 		return;
728 	}
729 	++bbr->full_bw_cnt;
730 	bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
731 }
732 
733 /* If pipe is probably full, drain the queue and then enter steady-state. */
734 static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
735 {
736 	struct bbr *bbr = inet_csk_ca(sk);
737 
738 	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
739 		bbr->mode = BBR_DRAIN;	/* drain queue we created */
740 		bbr->pacing_gain = bbr_drain_gain;	/* pace slow to drain */
741 		bbr->cwnd_gain = bbr_high_gain;	/* maintain cwnd */
742 		tcp_sk(sk)->snd_ssthresh =
743 				bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT);
744 	}	/* fall through to check if in-flight is already small: */
745 	if (bbr->mode == BBR_DRAIN &&
746 	    tcp_packets_in_flight(tcp_sk(sk)) <=
747 	    bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT))
748 		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
749 }
750 
751 /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
752  * periodically drain the bottleneck queue, to converge to measure the true
753  * min_rtt (unloaded propagation delay). This allows the flows to keep queues
754  * small (reducing queuing delay and packet loss) and achieve fairness among
755  * BBR flows.
756  *
757  * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
758  * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
759  * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
760  * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
761  * re-enter the previous mode. BBR uses 200ms to approximately bound the
762  * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
763  *
764  * Note that flows need only pay 2% if they are busy sending over the last 10
765  * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
766  * natural silences or low-rate periods within 10 seconds where the rate is low
767  * enough for long enough to drain its queue in the bottleneck. We pick up
768  * these min RTT measurements opportunistically with our min_rtt filter. :-)
769  */
770 static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
771 {
772 	struct tcp_sock *tp = tcp_sk(sk);
773 	struct bbr *bbr = inet_csk_ca(sk);
774 	bool filter_expired;
775 
776 	/* Track min RTT seen in the min_rtt_win_sec filter window: */
777 	filter_expired = after(tcp_jiffies32,
778 			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
779 	if (rs->rtt_us >= 0 &&
780 	    (rs->rtt_us <= bbr->min_rtt_us ||
781 	     (filter_expired && !rs->is_ack_delayed))) {
782 		bbr->min_rtt_us = rs->rtt_us;
783 		bbr->min_rtt_stamp = tcp_jiffies32;
784 	}
785 
786 	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
787 	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
788 		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
789 		bbr->pacing_gain = BBR_UNIT;
790 		bbr->cwnd_gain = BBR_UNIT;
791 		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
792 		bbr->probe_rtt_done_stamp = 0;
793 	}
794 
795 	if (bbr->mode == BBR_PROBE_RTT) {
796 		/* Ignore low rate samples during this mode. */
797 		tp->app_limited =
798 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
799 		/* Maintain min packets in flight for max(200 ms, 1 round). */
800 		if (!bbr->probe_rtt_done_stamp &&
801 		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
802 			bbr->probe_rtt_done_stamp = tcp_jiffies32 +
803 				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
804 			bbr->probe_rtt_round_done = 0;
805 			bbr->next_rtt_delivered = tp->delivered;
806 		} else if (bbr->probe_rtt_done_stamp) {
807 			if (bbr->round_start)
808 				bbr->probe_rtt_round_done = 1;
809 			if (bbr->probe_rtt_round_done &&
810 			    after(tcp_jiffies32, bbr->probe_rtt_done_stamp)) {
811 				bbr->min_rtt_stamp = tcp_jiffies32;
812 				bbr->restore_cwnd = 1;  /* snap to prior_cwnd */
813 				bbr_reset_mode(sk);
814 			}
815 		}
816 	}
817 	/* Restart after idle ends only once we process a new S/ACK for data */
818 	if (rs->delivered > 0)
819 		bbr->idle_restart = 0;
820 }
821 
822 static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
823 {
824 	bbr_update_bw(sk, rs);
825 	bbr_update_cycle_phase(sk, rs);
826 	bbr_check_full_bw_reached(sk, rs);
827 	bbr_check_drain(sk, rs);
828 	bbr_update_min_rtt(sk, rs);
829 }
830 
831 static void bbr_main(struct sock *sk, const struct rate_sample *rs)
832 {
833 	struct bbr *bbr = inet_csk_ca(sk);
834 	u32 bw;
835 
836 	bbr_update_model(sk, rs);
837 
838 	bw = bbr_bw(sk);
839 	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
840 	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
841 }
842 
843 static void bbr_init(struct sock *sk)
844 {
845 	struct tcp_sock *tp = tcp_sk(sk);
846 	struct bbr *bbr = inet_csk_ca(sk);
847 
848 	bbr->prior_cwnd = 0;
849 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
850 	bbr->rtt_cnt = 0;
851 	bbr->next_rtt_delivered = 0;
852 	bbr->prev_ca_state = TCP_CA_Open;
853 	bbr->packet_conservation = 0;
854 
855 	bbr->probe_rtt_done_stamp = 0;
856 	bbr->probe_rtt_round_done = 0;
857 	bbr->min_rtt_us = tcp_min_rtt(tp);
858 	bbr->min_rtt_stamp = tcp_jiffies32;
859 
860 	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
861 
862 	bbr->has_seen_rtt = 0;
863 	bbr_init_pacing_rate_from_rtt(sk);
864 
865 	bbr->restore_cwnd = 0;
866 	bbr->round_start = 0;
867 	bbr->idle_restart = 0;
868 	bbr->full_bw_reached = 0;
869 	bbr->full_bw = 0;
870 	bbr->full_bw_cnt = 0;
871 	bbr->cycle_mstamp = 0;
872 	bbr->cycle_idx = 0;
873 	bbr_reset_lt_bw_sampling(sk);
874 	bbr_reset_startup_mode(sk);
875 
876 	cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
877 }
878 
879 static u32 bbr_sndbuf_expand(struct sock *sk)
880 {
881 	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
882 	return 3;
883 }
884 
885 /* In theory BBR does not need to undo the cwnd since it does not
886  * always reduce cwnd on losses (see bbr_main()). Keep it for now.
887  */
888 static u32 bbr_undo_cwnd(struct sock *sk)
889 {
890 	struct bbr *bbr = inet_csk_ca(sk);
891 
892 	bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
893 	bbr->full_bw_cnt = 0;
894 	bbr_reset_lt_bw_sampling(sk);
895 	return tcp_sk(sk)->snd_cwnd;
896 }
897 
898 /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
899 static u32 bbr_ssthresh(struct sock *sk)
900 {
901 	bbr_save_cwnd(sk);
902 	return tcp_sk(sk)->snd_ssthresh;
903 }
904 
905 static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
906 			   union tcp_cc_info *info)
907 {
908 	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
909 	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
910 		struct tcp_sock *tp = tcp_sk(sk);
911 		struct bbr *bbr = inet_csk_ca(sk);
912 		u64 bw = bbr_bw(sk);
913 
914 		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
915 		memset(&info->bbr, 0, sizeof(info->bbr));
916 		info->bbr.bbr_bw_lo		= (u32)bw;
917 		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
918 		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
919 		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
920 		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
921 		*attr = INET_DIAG_BBRINFO;
922 		return sizeof(info->bbr);
923 	}
924 	return 0;
925 }
926 
927 static void bbr_set_state(struct sock *sk, u8 new_state)
928 {
929 	struct bbr *bbr = inet_csk_ca(sk);
930 
931 	if (new_state == TCP_CA_Loss) {
932 		struct rate_sample rs = { .losses = 1 };
933 
934 		bbr->prev_ca_state = TCP_CA_Loss;
935 		bbr->full_bw = 0;
936 		bbr->round_start = 1;	/* treat RTO like end of a round */
937 		bbr_lt_bw_sampling(sk, &rs);
938 	}
939 }
940 
941 static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
942 	.flags		= TCP_CONG_NON_RESTRICTED,
943 	.name		= "bbr",
944 	.owner		= THIS_MODULE,
945 	.init		= bbr_init,
946 	.cong_control	= bbr_main,
947 	.sndbuf_expand	= bbr_sndbuf_expand,
948 	.undo_cwnd	= bbr_undo_cwnd,
949 	.cwnd_event	= bbr_cwnd_event,
950 	.ssthresh	= bbr_ssthresh,
951 	.min_tso_segs	= bbr_min_tso_segs,
952 	.get_info	= bbr_get_info,
953 	.set_state	= bbr_set_state,
954 };
955 
956 static int __init bbr_register(void)
957 {
958 	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
959 	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
960 }
961 
962 static void __exit bbr_unregister(void)
963 {
964 	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
965 }
966 
967 module_init(bbr_register);
968 module_exit(bbr_unregister);
969 
970 MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
971 MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
972 MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
973 MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
974 MODULE_LICENSE("Dual BSD/GPL");
975 MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");
976