Commit | Line | Data |
---|---|---|
0f8782ea NC |
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 | * | |
9b9375b5 NC |
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 | * | |
0f8782ea NC |
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 | * | |
218af599 ED |
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. | |
0f8782ea NC |
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 */ | |
9b9375b5 | 83 | BBR_PROBE_RTT, /* cut inflight to min to probe min_rtt */ |
0f8782ea NC |
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 | struct skb_mstamp 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_cnt:3, /* number of rounds without large bw gains */ | |
114 | cycle_idx:3, /* current index in pacing_gain cycle array */ | |
115 | unused_b:6; | |
116 | u32 prior_cwnd; /* prior cwnd upon entering loss recovery */ | |
117 | u32 full_bw; /* recent bw, to estimate if pipe is full */ | |
118 | }; | |
119 | ||
120 | #define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */ | |
121 | ||
122 | /* Window length of bw filter (in rounds): */ | |
123 | static const int bbr_bw_rtts = CYCLE_LEN + 2; | |
124 | /* Window length of min_rtt filter (in sec): */ | |
125 | static const u32 bbr_min_rtt_win_sec = 10; | |
126 | /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */ | |
127 | static const u32 bbr_probe_rtt_mode_ms = 200; | |
128 | /* Skip TSO below the following bandwidth (bits/sec): */ | |
129 | static const int bbr_min_tso_rate = 1200000; | |
130 | ||
131 | /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain | |
132 | * that will allow a smoothly increasing pacing rate that will double each RTT | |
133 | * and send the same number of packets per RTT that an un-paced, slow-starting | |
134 | * Reno or CUBIC flow would: | |
135 | */ | |
136 | static const int bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; | |
137 | /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain | |
138 | * the queue created in BBR_STARTUP in a single round: | |
139 | */ | |
140 | static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885; | |
141 | /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */ | |
142 | static const int bbr_cwnd_gain = BBR_UNIT * 2; | |
143 | /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */ | |
144 | static const int bbr_pacing_gain[] = { | |
145 | BBR_UNIT * 5 / 4, /* probe for more available bw */ | |
146 | BBR_UNIT * 3 / 4, /* drain queue and/or yield bw to other flows */ | |
147 | BBR_UNIT, BBR_UNIT, BBR_UNIT, /* cruise at 1.0*bw to utilize pipe, */ | |
148 | BBR_UNIT, BBR_UNIT, BBR_UNIT /* without creating excess queue... */ | |
149 | }; | |
150 | /* Randomize the starting gain cycling phase over N phases: */ | |
151 | static const u32 bbr_cycle_rand = 7; | |
152 | ||
153 | /* Try to keep at least this many packets in flight, if things go smoothly. For | |
154 | * smooth functioning, a sliding window protocol ACKing every other packet | |
155 | * needs at least 4 packets in flight: | |
156 | */ | |
157 | static const u32 bbr_cwnd_min_target = 4; | |
158 | ||
159 | /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */ | |
160 | /* If bw has increased significantly (1.25x), there may be more bw available: */ | |
161 | static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4; | |
162 | /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */ | |
163 | static const u32 bbr_full_bw_cnt = 3; | |
164 | ||
165 | /* "long-term" ("LT") bandwidth estimator parameters... */ | |
166 | /* The minimum number of rounds in an LT bw sampling interval: */ | |
167 | static const u32 bbr_lt_intvl_min_rtts = 4; | |
168 | /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */ | |
169 | static const u32 bbr_lt_loss_thresh = 50; | |
170 | /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */ | |
171 | static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8; | |
172 | /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */ | |
173 | static const u32 bbr_lt_bw_diff = 4000 / 8; | |
174 | /* If we estimate we're policed, use lt_bw for this many round trips: */ | |
175 | static const u32 bbr_lt_bw_max_rtts = 48; | |
176 | ||
177 | /* Do we estimate that STARTUP filled the pipe? */ | |
178 | static bool bbr_full_bw_reached(const struct sock *sk) | |
179 | { | |
180 | const struct bbr *bbr = inet_csk_ca(sk); | |
181 | ||
182 | return bbr->full_bw_cnt >= bbr_full_bw_cnt; | |
183 | } | |
184 | ||
185 | /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */ | |
186 | static u32 bbr_max_bw(const struct sock *sk) | |
187 | { | |
188 | struct bbr *bbr = inet_csk_ca(sk); | |
189 | ||
190 | return minmax_get(&bbr->bw); | |
191 | } | |
192 | ||
193 | /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */ | |
194 | static u32 bbr_bw(const struct sock *sk) | |
195 | { | |
196 | struct bbr *bbr = inet_csk_ca(sk); | |
197 | ||
198 | return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk); | |
199 | } | |
200 | ||
201 | /* Return rate in bytes per second, optionally with a gain. | |
202 | * The order here is chosen carefully to avoid overflow of u64. This should | |
203 | * work for input rates of up to 2.9Tbit/sec and gain of 2.89x. | |
204 | */ | |
205 | static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain) | |
206 | { | |
207 | rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache); | |
208 | rate *= gain; | |
209 | rate >>= BBR_SCALE; | |
210 | rate *= USEC_PER_SEC; | |
211 | return rate >> BW_SCALE; | |
212 | } | |
213 | ||
214 | /* Pace using current bw estimate and a gain factor. In order to help drive the | |
215 | * network toward lower queues while maintaining high utilization and low | |
216 | * latency, the average pacing rate aims to be slightly (~1%) lower than the | |
217 | * estimated bandwidth. This is an important aspect of the design. In this | |
218 | * implementation this slightly lower pacing rate is achieved implicitly by not | |
219 | * including link-layer headers in the packet size used for the pacing rate. | |
220 | */ | |
221 | static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain) | |
222 | { | |
223 | struct bbr *bbr = inet_csk_ca(sk); | |
224 | u64 rate = bw; | |
225 | ||
226 | rate = bbr_rate_bytes_per_sec(sk, rate, gain); | |
227 | rate = min_t(u64, rate, sk->sk_max_pacing_rate); | |
228 | if (bbr->mode != BBR_STARTUP || rate > sk->sk_pacing_rate) | |
229 | sk->sk_pacing_rate = rate; | |
230 | } | |
231 | ||
232 | /* Return count of segments we want in the skbs we send, or 0 for default. */ | |
233 | static u32 bbr_tso_segs_goal(struct sock *sk) | |
234 | { | |
235 | struct bbr *bbr = inet_csk_ca(sk); | |
236 | ||
237 | return bbr->tso_segs_goal; | |
238 | } | |
239 | ||
240 | static void bbr_set_tso_segs_goal(struct sock *sk) | |
241 | { | |
242 | struct tcp_sock *tp = tcp_sk(sk); | |
243 | struct bbr *bbr = inet_csk_ca(sk); | |
244 | u32 min_segs; | |
245 | ||
246 | min_segs = sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2; | |
247 | bbr->tso_segs_goal = min(tcp_tso_autosize(sk, tp->mss_cache, min_segs), | |
248 | 0x7FU); | |
249 | } | |
250 | ||
251 | /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ | |
252 | static void bbr_save_cwnd(struct sock *sk) | |
253 | { | |
254 | struct tcp_sock *tp = tcp_sk(sk); | |
255 | struct bbr *bbr = inet_csk_ca(sk); | |
256 | ||
257 | if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT) | |
258 | bbr->prior_cwnd = tp->snd_cwnd; /* this cwnd is good enough */ | |
259 | else /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */ | |
260 | bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd); | |
261 | } | |
262 | ||
263 | static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event) | |
264 | { | |
265 | struct tcp_sock *tp = tcp_sk(sk); | |
266 | struct bbr *bbr = inet_csk_ca(sk); | |
267 | ||
268 | if (event == CA_EVENT_TX_START && tp->app_limited) { | |
269 | bbr->idle_restart = 1; | |
270 | /* Avoid pointless buffer overflows: pace at est. bw if we don't | |
271 | * need more speed (we're restarting from idle and app-limited). | |
272 | */ | |
273 | if (bbr->mode == BBR_PROBE_BW) | |
274 | bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT); | |
275 | } | |
276 | } | |
277 | ||
278 | /* Find target cwnd. Right-size the cwnd based on min RTT and the | |
279 | * estimated bottleneck bandwidth: | |
280 | * | |
281 | * cwnd = bw * min_rtt * gain = BDP * gain | |
282 | * | |
283 | * The key factor, gain, controls the amount of queue. While a small gain | |
284 | * builds a smaller queue, it becomes more vulnerable to noise in RTT | |
285 | * measurements (e.g., delayed ACKs or other ACK compression effects). This | |
286 | * noise may cause BBR to under-estimate the rate. | |
287 | * | |
288 | * To achieve full performance in high-speed paths, we budget enough cwnd to | |
289 | * fit full-sized skbs in-flight on both end hosts to fully utilize the path: | |
290 | * - one skb in sending host Qdisc, | |
291 | * - one skb in sending host TSO/GSO engine | |
292 | * - one skb being received by receiver host LRO/GRO/delayed-ACK engine | |
293 | * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because | |
294 | * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets, | |
295 | * which allows 2 outstanding 2-packet sequences, to try to keep pipe | |
296 | * full even with ACK-every-other-packet delayed ACKs. | |
297 | */ | |
298 | static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain) | |
299 | { | |
300 | struct bbr *bbr = inet_csk_ca(sk); | |
301 | u32 cwnd; | |
302 | u64 w; | |
303 | ||
304 | /* If we've never had a valid RTT sample, cap cwnd at the initial | |
305 | * default. This should only happen when the connection is not using TCP | |
306 | * timestamps and has retransmitted all of the SYN/SYNACK/data packets | |
307 | * ACKed so far. In this case, an RTO can cut cwnd to 1, in which | |
308 | * case we need to slow-start up toward something safe: TCP_INIT_CWND. | |
309 | */ | |
310 | if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */ | |
311 | return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/ | |
312 | ||
313 | w = (u64)bw * bbr->min_rtt_us; | |
314 | ||
315 | /* Apply a gain to the given value, then remove the BW_SCALE shift. */ | |
316 | cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT; | |
317 | ||
318 | /* Allow enough full-sized skbs in flight to utilize end systems. */ | |
319 | cwnd += 3 * bbr->tso_segs_goal; | |
320 | ||
321 | /* Reduce delayed ACKs by rounding up cwnd to the next even number. */ | |
322 | cwnd = (cwnd + 1) & ~1U; | |
323 | ||
324 | return cwnd; | |
325 | } | |
326 | ||
327 | /* An optimization in BBR to reduce losses: On the first round of recovery, we | |
328 | * follow the packet conservation principle: send P packets per P packets acked. | |
329 | * After that, we slow-start and send at most 2*P packets per P packets acked. | |
330 | * After recovery finishes, or upon undo, we restore the cwnd we had when | |
331 | * recovery started (capped by the target cwnd based on estimated BDP). | |
332 | * | |
333 | * TODO(ycheng/ncardwell): implement a rate-based approach. | |
334 | */ | |
335 | static bool bbr_set_cwnd_to_recover_or_restore( | |
336 | struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd) | |
337 | { | |
338 | struct tcp_sock *tp = tcp_sk(sk); | |
339 | struct bbr *bbr = inet_csk_ca(sk); | |
340 | u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state; | |
341 | u32 cwnd = tp->snd_cwnd; | |
342 | ||
343 | /* An ACK for P pkts should release at most 2*P packets. We do this | |
344 | * in two steps. First, here we deduct the number of lost packets. | |
345 | * Then, in bbr_set_cwnd() we slow start up toward the target cwnd. | |
346 | */ | |
347 | if (rs->losses > 0) | |
348 | cwnd = max_t(s32, cwnd - rs->losses, 1); | |
349 | ||
350 | if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) { | |
351 | /* Starting 1st round of Recovery, so do packet conservation. */ | |
352 | bbr->packet_conservation = 1; | |
353 | bbr->next_rtt_delivered = tp->delivered; /* start round now */ | |
354 | /* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */ | |
355 | cwnd = tcp_packets_in_flight(tp) + acked; | |
356 | } else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) { | |
357 | /* Exiting loss recovery; restore cwnd saved before recovery. */ | |
358 | bbr->restore_cwnd = 1; | |
359 | bbr->packet_conservation = 0; | |
360 | } | |
361 | bbr->prev_ca_state = state; | |
362 | ||
363 | if (bbr->restore_cwnd) { | |
364 | /* Restore cwnd after exiting loss recovery or PROBE_RTT. */ | |
365 | cwnd = max(cwnd, bbr->prior_cwnd); | |
366 | bbr->restore_cwnd = 0; | |
367 | } | |
368 | ||
369 | if (bbr->packet_conservation) { | |
370 | *new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked); | |
371 | return true; /* yes, using packet conservation */ | |
372 | } | |
373 | *new_cwnd = cwnd; | |
374 | return false; | |
375 | } | |
376 | ||
377 | /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss | |
378 | * has drawn us down below target), or snap down to target if we're above it. | |
379 | */ | |
380 | static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs, | |
381 | u32 acked, u32 bw, int gain) | |
382 | { | |
383 | struct tcp_sock *tp = tcp_sk(sk); | |
384 | struct bbr *bbr = inet_csk_ca(sk); | |
385 | u32 cwnd = 0, target_cwnd = 0; | |
386 | ||
387 | if (!acked) | |
388 | return; | |
389 | ||
390 | if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd)) | |
391 | goto done; | |
392 | ||
393 | /* If we're below target cwnd, slow start cwnd toward target cwnd. */ | |
394 | target_cwnd = bbr_target_cwnd(sk, bw, gain); | |
395 | if (bbr_full_bw_reached(sk)) /* only cut cwnd if we filled the pipe */ | |
396 | cwnd = min(cwnd + acked, target_cwnd); | |
397 | else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND) | |
398 | cwnd = cwnd + acked; | |
399 | cwnd = max(cwnd, bbr_cwnd_min_target); | |
400 | ||
401 | done: | |
402 | tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); /* apply global cap */ | |
403 | if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */ | |
404 | tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target); | |
405 | } | |
406 | ||
407 | /* End cycle phase if it's time and/or we hit the phase's in-flight target. */ | |
408 | static bool bbr_is_next_cycle_phase(struct sock *sk, | |
409 | const struct rate_sample *rs) | |
410 | { | |
411 | struct tcp_sock *tp = tcp_sk(sk); | |
412 | struct bbr *bbr = inet_csk_ca(sk); | |
413 | bool is_full_length = | |
414 | skb_mstamp_us_delta(&tp->delivered_mstamp, &bbr->cycle_mstamp) > | |
415 | bbr->min_rtt_us; | |
416 | u32 inflight, bw; | |
417 | ||
418 | /* The pacing_gain of 1.0 paces at the estimated bw to try to fully | |
419 | * use the pipe without increasing the queue. | |
420 | */ | |
421 | if (bbr->pacing_gain == BBR_UNIT) | |
422 | return is_full_length; /* just use wall clock time */ | |
423 | ||
424 | inflight = rs->prior_in_flight; /* what was in-flight before ACK? */ | |
425 | bw = bbr_max_bw(sk); | |
426 | ||
427 | /* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at | |
428 | * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is | |
429 | * small (e.g. on a LAN). We do not persist if packets are lost, since | |
430 | * a path with small buffers may not hold that much. | |
431 | */ | |
432 | if (bbr->pacing_gain > BBR_UNIT) | |
433 | return is_full_length && | |
434 | (rs->losses || /* perhaps pacing_gain*BDP won't fit */ | |
435 | inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain)); | |
436 | ||
437 | /* A pacing_gain < 1.0 tries to drain extra queue we added if bw | |
438 | * probing didn't find more bw. If inflight falls to match BDP then we | |
439 | * estimate queue is drained; persisting would underutilize the pipe. | |
440 | */ | |
441 | return is_full_length || | |
442 | inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT); | |
443 | } | |
444 | ||
445 | static void bbr_advance_cycle_phase(struct sock *sk) | |
446 | { | |
447 | struct tcp_sock *tp = tcp_sk(sk); | |
448 | struct bbr *bbr = inet_csk_ca(sk); | |
449 | ||
450 | bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1); | |
451 | bbr->cycle_mstamp = tp->delivered_mstamp; | |
452 | bbr->pacing_gain = bbr_pacing_gain[bbr->cycle_idx]; | |
453 | } | |
454 | ||
455 | /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */ | |
456 | static void bbr_update_cycle_phase(struct sock *sk, | |
457 | const struct rate_sample *rs) | |
458 | { | |
459 | struct bbr *bbr = inet_csk_ca(sk); | |
460 | ||
461 | if ((bbr->mode == BBR_PROBE_BW) && !bbr->lt_use_bw && | |
462 | bbr_is_next_cycle_phase(sk, rs)) | |
463 | bbr_advance_cycle_phase(sk); | |
464 | } | |
465 | ||
466 | static void bbr_reset_startup_mode(struct sock *sk) | |
467 | { | |
468 | struct bbr *bbr = inet_csk_ca(sk); | |
469 | ||
470 | bbr->mode = BBR_STARTUP; | |
471 | bbr->pacing_gain = bbr_high_gain; | |
472 | bbr->cwnd_gain = bbr_high_gain; | |
473 | } | |
474 | ||
475 | static void bbr_reset_probe_bw_mode(struct sock *sk) | |
476 | { | |
477 | struct bbr *bbr = inet_csk_ca(sk); | |
478 | ||
479 | bbr->mode = BBR_PROBE_BW; | |
480 | bbr->pacing_gain = BBR_UNIT; | |
481 | bbr->cwnd_gain = bbr_cwnd_gain; | |
482 | bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand); | |
483 | bbr_advance_cycle_phase(sk); /* flip to next phase of gain cycle */ | |
484 | } | |
485 | ||
486 | static void bbr_reset_mode(struct sock *sk) | |
487 | { | |
488 | if (!bbr_full_bw_reached(sk)) | |
489 | bbr_reset_startup_mode(sk); | |
490 | else | |
491 | bbr_reset_probe_bw_mode(sk); | |
492 | } | |
493 | ||
494 | /* Start a new long-term sampling interval. */ | |
495 | static void bbr_reset_lt_bw_sampling_interval(struct sock *sk) | |
496 | { | |
497 | struct tcp_sock *tp = tcp_sk(sk); | |
498 | struct bbr *bbr = inet_csk_ca(sk); | |
499 | ||
500 | bbr->lt_last_stamp = tp->delivered_mstamp.stamp_jiffies; | |
501 | bbr->lt_last_delivered = tp->delivered; | |
502 | bbr->lt_last_lost = tp->lost; | |
503 | bbr->lt_rtt_cnt = 0; | |
504 | } | |
505 | ||
506 | /* Completely reset long-term bandwidth sampling. */ | |
507 | static void bbr_reset_lt_bw_sampling(struct sock *sk) | |
508 | { | |
509 | struct bbr *bbr = inet_csk_ca(sk); | |
510 | ||
511 | bbr->lt_bw = 0; | |
512 | bbr->lt_use_bw = 0; | |
513 | bbr->lt_is_sampling = false; | |
514 | bbr_reset_lt_bw_sampling_interval(sk); | |
515 | } | |
516 | ||
517 | /* Long-term bw sampling interval is done. Estimate whether we're policed. */ | |
518 | static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw) | |
519 | { | |
520 | struct bbr *bbr = inet_csk_ca(sk); | |
521 | u32 diff; | |
522 | ||
523 | if (bbr->lt_bw) { /* do we have bw from a previous interval? */ | |
524 | /* Is new bw close to the lt_bw from the previous interval? */ | |
525 | diff = abs(bw - bbr->lt_bw); | |
526 | if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) || | |
527 | (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <= | |
528 | bbr_lt_bw_diff)) { | |
529 | /* All criteria are met; estimate we're policed. */ | |
530 | bbr->lt_bw = (bw + bbr->lt_bw) >> 1; /* avg 2 intvls */ | |
531 | bbr->lt_use_bw = 1; | |
532 | bbr->pacing_gain = BBR_UNIT; /* try to avoid drops */ | |
533 | bbr->lt_rtt_cnt = 0; | |
534 | return; | |
535 | } | |
536 | } | |
537 | bbr->lt_bw = bw; | |
538 | bbr_reset_lt_bw_sampling_interval(sk); | |
539 | } | |
540 | ||
541 | /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of | |
542 | * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and | |
543 | * explicitly models their policed rate, to reduce unnecessary losses. We | |
544 | * estimate that we're policed if we see 2 consecutive sampling intervals with | |
545 | * consistent throughput and high packet loss. If we think we're being policed, | |
546 | * set lt_bw to the "long-term" average delivery rate from those 2 intervals. | |
547 | */ | |
548 | static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs) | |
549 | { | |
550 | struct tcp_sock *tp = tcp_sk(sk); | |
551 | struct bbr *bbr = inet_csk_ca(sk); | |
552 | u32 lost, delivered; | |
553 | u64 bw; | |
554 | s32 t; | |
555 | ||
556 | if (bbr->lt_use_bw) { /* already using long-term rate, lt_bw? */ | |
557 | if (bbr->mode == BBR_PROBE_BW && bbr->round_start && | |
558 | ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) { | |
559 | bbr_reset_lt_bw_sampling(sk); /* stop using lt_bw */ | |
560 | bbr_reset_probe_bw_mode(sk); /* restart gain cycling */ | |
561 | } | |
562 | return; | |
563 | } | |
564 | ||
565 | /* Wait for the first loss before sampling, to let the policer exhaust | |
566 | * its tokens and estimate the steady-state rate allowed by the policer. | |
567 | * Starting samples earlier includes bursts that over-estimate the bw. | |
568 | */ | |
569 | if (!bbr->lt_is_sampling) { | |
570 | if (!rs->losses) | |
571 | return; | |
572 | bbr_reset_lt_bw_sampling_interval(sk); | |
573 | bbr->lt_is_sampling = true; | |
574 | } | |
575 | ||
576 | /* To avoid underestimates, reset sampling if we run out of data. */ | |
577 | if (rs->is_app_limited) { | |
578 | bbr_reset_lt_bw_sampling(sk); | |
579 | return; | |
580 | } | |
581 | ||
582 | if (bbr->round_start) | |
583 | bbr->lt_rtt_cnt++; /* count round trips in this interval */ | |
584 | if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts) | |
585 | return; /* sampling interval needs to be longer */ | |
586 | if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) { | |
587 | bbr_reset_lt_bw_sampling(sk); /* interval is too long */ | |
588 | return; | |
589 | } | |
590 | ||
591 | /* End sampling interval when a packet is lost, so we estimate the | |
592 | * policer tokens were exhausted. Stopping the sampling before the | |
593 | * tokens are exhausted under-estimates the policed rate. | |
594 | */ | |
595 | if (!rs->losses) | |
596 | return; | |
597 | ||
598 | /* Calculate packets lost and delivered in sampling interval. */ | |
599 | lost = tp->lost - bbr->lt_last_lost; | |
600 | delivered = tp->delivered - bbr->lt_last_delivered; | |
601 | /* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */ | |
602 | if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered) | |
603 | return; | |
604 | ||
605 | /* Find average delivery rate in this sampling interval. */ | |
606 | t = (s32)(tp->delivered_mstamp.stamp_jiffies - bbr->lt_last_stamp); | |
607 | if (t < 1) | |
608 | return; /* interval is less than one jiffy, so wait */ | |
609 | t = jiffies_to_usecs(t); | |
610 | /* Interval long enough for jiffies_to_usecs() to return a bogus 0? */ | |
611 | if (t < 1) { | |
612 | bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */ | |
613 | return; | |
614 | } | |
615 | bw = (u64)delivered * BW_UNIT; | |
616 | do_div(bw, t); | |
617 | bbr_lt_bw_interval_done(sk, bw); | |
618 | } | |
619 | ||
620 | /* Estimate the bandwidth based on how fast packets are delivered */ | |
621 | static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs) | |
622 | { | |
623 | struct tcp_sock *tp = tcp_sk(sk); | |
624 | struct bbr *bbr = inet_csk_ca(sk); | |
625 | u64 bw; | |
626 | ||
627 | bbr->round_start = 0; | |
628 | if (rs->delivered < 0 || rs->interval_us <= 0) | |
629 | return; /* Not a valid observation */ | |
630 | ||
631 | /* See if we've reached the next RTT */ | |
632 | if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) { | |
633 | bbr->next_rtt_delivered = tp->delivered; | |
634 | bbr->rtt_cnt++; | |
635 | bbr->round_start = 1; | |
636 | bbr->packet_conservation = 0; | |
637 | } | |
638 | ||
639 | bbr_lt_bw_sampling(sk, rs); | |
640 | ||
641 | /* Divide delivered by the interval to find a (lower bound) bottleneck | |
642 | * bandwidth sample. Delivered is in packets and interval_us in uS and | |
643 | * ratio will be <<1 for most connections. So delivered is first scaled. | |
644 | */ | |
645 | bw = (u64)rs->delivered * BW_UNIT; | |
646 | do_div(bw, rs->interval_us); | |
647 | ||
648 | /* If this sample is application-limited, it is likely to have a very | |
649 | * low delivered count that represents application behavior rather than | |
650 | * the available network rate. Such a sample could drag down estimated | |
651 | * bw, causing needless slow-down. Thus, to continue to send at the | |
652 | * last measured network rate, we filter out app-limited samples unless | |
653 | * they describe the path bw at least as well as our bw model. | |
654 | * | |
655 | * So the goal during app-limited phase is to proceed with the best | |
656 | * network rate no matter how long. We automatically leave this | |
657 | * phase when app writes faster than the network can deliver :) | |
658 | */ | |
659 | if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) { | |
660 | /* Incorporate new sample into our max bw filter. */ | |
661 | minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw); | |
662 | } | |
663 | } | |
664 | ||
665 | /* Estimate when the pipe is full, using the change in delivery rate: BBR | |
666 | * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by | |
667 | * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited | |
668 | * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the | |
669 | * higher rwin, 3: we get higher delivery rate samples. Or transient | |
670 | * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar | |
671 | * design goal, but uses delay and inter-ACK spacing instead of bandwidth. | |
672 | */ | |
673 | static void bbr_check_full_bw_reached(struct sock *sk, | |
674 | const struct rate_sample *rs) | |
675 | { | |
676 | struct bbr *bbr = inet_csk_ca(sk); | |
677 | u32 bw_thresh; | |
678 | ||
679 | if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited) | |
680 | return; | |
681 | ||
682 | bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE; | |
683 | if (bbr_max_bw(sk) >= bw_thresh) { | |
684 | bbr->full_bw = bbr_max_bw(sk); | |
685 | bbr->full_bw_cnt = 0; | |
686 | return; | |
687 | } | |
688 | ++bbr->full_bw_cnt; | |
689 | } | |
690 | ||
691 | /* If pipe is probably full, drain the queue and then enter steady-state. */ | |
692 | static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs) | |
693 | { | |
694 | struct bbr *bbr = inet_csk_ca(sk); | |
695 | ||
696 | if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) { | |
697 | bbr->mode = BBR_DRAIN; /* drain queue we created */ | |
698 | bbr->pacing_gain = bbr_drain_gain; /* pace slow to drain */ | |
699 | bbr->cwnd_gain = bbr_high_gain; /* maintain cwnd */ | |
700 | } /* fall through to check if in-flight is already small: */ | |
701 | if (bbr->mode == BBR_DRAIN && | |
702 | tcp_packets_in_flight(tcp_sk(sk)) <= | |
703 | bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT)) | |
704 | bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */ | |
705 | } | |
706 | ||
707 | /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and | |
708 | * periodically drain the bottleneck queue, to converge to measure the true | |
709 | * min_rtt (unloaded propagation delay). This allows the flows to keep queues | |
710 | * small (reducing queuing delay and packet loss) and achieve fairness among | |
711 | * BBR flows. | |
712 | * | |
713 | * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires, | |
714 | * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets. | |
715 | * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed | |
716 | * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and | |
717 | * re-enter the previous mode. BBR uses 200ms to approximately bound the | |
718 | * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s). | |
719 | * | |
720 | * Note that flows need only pay 2% if they are busy sending over the last 10 | |
721 | * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have | |
722 | * natural silences or low-rate periods within 10 seconds where the rate is low | |
723 | * enough for long enough to drain its queue in the bottleneck. We pick up | |
724 | * these min RTT measurements opportunistically with our min_rtt filter. :-) | |
725 | */ | |
726 | static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs) | |
727 | { | |
728 | struct tcp_sock *tp = tcp_sk(sk); | |
729 | struct bbr *bbr = inet_csk_ca(sk); | |
730 | bool filter_expired; | |
731 | ||
732 | /* Track min RTT seen in the min_rtt_win_sec filter window: */ | |
733 | filter_expired = after(tcp_time_stamp, | |
734 | bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ); | |
735 | if (rs->rtt_us >= 0 && | |
736 | (rs->rtt_us <= bbr->min_rtt_us || filter_expired)) { | |
737 | bbr->min_rtt_us = rs->rtt_us; | |
738 | bbr->min_rtt_stamp = tcp_time_stamp; | |
739 | } | |
740 | ||
741 | if (bbr_probe_rtt_mode_ms > 0 && filter_expired && | |
742 | !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) { | |
743 | bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */ | |
744 | bbr->pacing_gain = BBR_UNIT; | |
745 | bbr->cwnd_gain = BBR_UNIT; | |
746 | bbr_save_cwnd(sk); /* note cwnd so we can restore it */ | |
747 | bbr->probe_rtt_done_stamp = 0; | |
748 | } | |
749 | ||
750 | if (bbr->mode == BBR_PROBE_RTT) { | |
751 | /* Ignore low rate samples during this mode. */ | |
752 | tp->app_limited = | |
753 | (tp->delivered + tcp_packets_in_flight(tp)) ? : 1; | |
754 | /* Maintain min packets in flight for max(200 ms, 1 round). */ | |
755 | if (!bbr->probe_rtt_done_stamp && | |
756 | tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) { | |
757 | bbr->probe_rtt_done_stamp = tcp_time_stamp + | |
758 | msecs_to_jiffies(bbr_probe_rtt_mode_ms); | |
759 | bbr->probe_rtt_round_done = 0; | |
760 | bbr->next_rtt_delivered = tp->delivered; | |
761 | } else if (bbr->probe_rtt_done_stamp) { | |
762 | if (bbr->round_start) | |
763 | bbr->probe_rtt_round_done = 1; | |
764 | if (bbr->probe_rtt_round_done && | |
765 | after(tcp_time_stamp, bbr->probe_rtt_done_stamp)) { | |
766 | bbr->min_rtt_stamp = tcp_time_stamp; | |
767 | bbr->restore_cwnd = 1; /* snap to prior_cwnd */ | |
768 | bbr_reset_mode(sk); | |
769 | } | |
770 | } | |
771 | } | |
772 | bbr->idle_restart = 0; | |
773 | } | |
774 | ||
775 | static void bbr_update_model(struct sock *sk, const struct rate_sample *rs) | |
776 | { | |
777 | bbr_update_bw(sk, rs); | |
778 | bbr_update_cycle_phase(sk, rs); | |
779 | bbr_check_full_bw_reached(sk, rs); | |
780 | bbr_check_drain(sk, rs); | |
781 | bbr_update_min_rtt(sk, rs); | |
782 | } | |
783 | ||
784 | static void bbr_main(struct sock *sk, const struct rate_sample *rs) | |
785 | { | |
786 | struct bbr *bbr = inet_csk_ca(sk); | |
787 | u32 bw; | |
788 | ||
789 | bbr_update_model(sk, rs); | |
790 | ||
791 | bw = bbr_bw(sk); | |
792 | bbr_set_pacing_rate(sk, bw, bbr->pacing_gain); | |
793 | bbr_set_tso_segs_goal(sk); | |
794 | bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain); | |
795 | } | |
796 | ||
797 | static void bbr_init(struct sock *sk) | |
798 | { | |
799 | struct tcp_sock *tp = tcp_sk(sk); | |
800 | struct bbr *bbr = inet_csk_ca(sk); | |
801 | u64 bw; | |
802 | ||
803 | bbr->prior_cwnd = 0; | |
804 | bbr->tso_segs_goal = 0; /* default segs per skb until first ACK */ | |
805 | bbr->rtt_cnt = 0; | |
806 | bbr->next_rtt_delivered = 0; | |
807 | bbr->prev_ca_state = TCP_CA_Open; | |
808 | bbr->packet_conservation = 0; | |
809 | ||
810 | bbr->probe_rtt_done_stamp = 0; | |
811 | bbr->probe_rtt_round_done = 0; | |
812 | bbr->min_rtt_us = tcp_min_rtt(tp); | |
813 | bbr->min_rtt_stamp = tcp_time_stamp; | |
814 | ||
815 | minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */ | |
816 | ||
817 | /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */ | |
818 | bw = (u64)tp->snd_cwnd * BW_UNIT; | |
819 | do_div(bw, (tp->srtt_us >> 3) ? : USEC_PER_MSEC); | |
820 | sk->sk_pacing_rate = 0; /* force an update of sk_pacing_rate */ | |
821 | bbr_set_pacing_rate(sk, bw, bbr_high_gain); | |
822 | ||
823 | bbr->restore_cwnd = 0; | |
824 | bbr->round_start = 0; | |
825 | bbr->idle_restart = 0; | |
826 | bbr->full_bw = 0; | |
827 | bbr->full_bw_cnt = 0; | |
828 | bbr->cycle_mstamp.v64 = 0; | |
829 | bbr->cycle_idx = 0; | |
830 | bbr_reset_lt_bw_sampling(sk); | |
831 | bbr_reset_startup_mode(sk); | |
218af599 ED |
832 | |
833 | cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED); | |
0f8782ea NC |
834 | } |
835 | ||
836 | static u32 bbr_sndbuf_expand(struct sock *sk) | |
837 | { | |
838 | /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ | |
839 | return 3; | |
840 | } | |
841 | ||
842 | /* In theory BBR does not need to undo the cwnd since it does not | |
843 | * always reduce cwnd on losses (see bbr_main()). Keep it for now. | |
844 | */ | |
845 | static u32 bbr_undo_cwnd(struct sock *sk) | |
846 | { | |
847 | return tcp_sk(sk)->snd_cwnd; | |
848 | } | |
849 | ||
850 | /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */ | |
851 | static u32 bbr_ssthresh(struct sock *sk) | |
852 | { | |
853 | bbr_save_cwnd(sk); | |
854 | return TCP_INFINITE_SSTHRESH; /* BBR does not use ssthresh */ | |
855 | } | |
856 | ||
857 | static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr, | |
858 | union tcp_cc_info *info) | |
859 | { | |
860 | if (ext & (1 << (INET_DIAG_BBRINFO - 1)) || | |
861 | ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | |
862 | struct tcp_sock *tp = tcp_sk(sk); | |
863 | struct bbr *bbr = inet_csk_ca(sk); | |
864 | u64 bw = bbr_bw(sk); | |
865 | ||
866 | bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE; | |
867 | memset(&info->bbr, 0, sizeof(info->bbr)); | |
868 | info->bbr.bbr_bw_lo = (u32)bw; | |
869 | info->bbr.bbr_bw_hi = (u32)(bw >> 32); | |
870 | info->bbr.bbr_min_rtt = bbr->min_rtt_us; | |
871 | info->bbr.bbr_pacing_gain = bbr->pacing_gain; | |
872 | info->bbr.bbr_cwnd_gain = bbr->cwnd_gain; | |
873 | *attr = INET_DIAG_BBRINFO; | |
874 | return sizeof(info->bbr); | |
875 | } | |
876 | return 0; | |
877 | } | |
878 | ||
879 | static void bbr_set_state(struct sock *sk, u8 new_state) | |
880 | { | |
881 | struct bbr *bbr = inet_csk_ca(sk); | |
882 | ||
883 | if (new_state == TCP_CA_Loss) { | |
884 | struct rate_sample rs = { .losses = 1 }; | |
885 | ||
886 | bbr->prev_ca_state = TCP_CA_Loss; | |
887 | bbr->full_bw = 0; | |
888 | bbr->round_start = 1; /* treat RTO like end of a round */ | |
889 | bbr_lt_bw_sampling(sk, &rs); | |
890 | } | |
891 | } | |
892 | ||
893 | static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { | |
894 | .flags = TCP_CONG_NON_RESTRICTED, | |
895 | .name = "bbr", | |
896 | .owner = THIS_MODULE, | |
897 | .init = bbr_init, | |
898 | .cong_control = bbr_main, | |
899 | .sndbuf_expand = bbr_sndbuf_expand, | |
900 | .undo_cwnd = bbr_undo_cwnd, | |
901 | .cwnd_event = bbr_cwnd_event, | |
902 | .ssthresh = bbr_ssthresh, | |
903 | .tso_segs_goal = bbr_tso_segs_goal, | |
904 | .get_info = bbr_get_info, | |
905 | .set_state = bbr_set_state, | |
906 | }; | |
907 | ||
908 | static int __init bbr_register(void) | |
909 | { | |
910 | BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE); | |
911 | return tcp_register_congestion_control(&tcp_bbr_cong_ops); | |
912 | } | |
913 | ||
914 | static void __exit bbr_unregister(void) | |
915 | { | |
916 | tcp_unregister_congestion_control(&tcp_bbr_cong_ops); | |
917 | } | |
918 | ||
919 | module_init(bbr_register); | |
920 | module_exit(bbr_unregister); | |
921 | ||
922 | MODULE_AUTHOR("Van Jacobson <vanj@google.com>"); | |
923 | MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>"); | |
924 | MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>"); | |
925 | MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>"); | |
926 | MODULE_LICENSE("Dual BSD/GPL"); | |
927 | MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)"); |