2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
5 * Copyright (c) 2010 The FreeBSD Foundation
8 * This software was developed by Lawrence Stewart while studying at the Centre
9 * for Advanced Internet Architectures, Swinburne University of Technology, made
10 * possible in part by a grant from the Cisco University Research Program Fund
11 * at Community Foundation Silicon Valley.
13 * Portions of this software were developed at the Centre for Advanced
14 * Internet Architectures, Swinburne University of Technology, Melbourne,
15 * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
26 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * An implementation of the CUBIC congestion control algorithm for FreeBSD,
41 * based on the Internet Draft "draft-rhee-tcpm-cubic-02" by Rhee, Xu and Ha.
42 * Originally released as part of the NewTCP research project at Swinburne
43 * University of Technology's Centre for Advanced Internet Architectures,
44 * Melbourne, Australia, which was made possible in part by a grant from the
45 * Cisco University Research Program Fund at Community Foundation Silicon
46 * Valley. More details are available at:
47 * http://caia.swin.edu.au/urp/newtcp/
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
53 #include <sys/param.h>
54 #include <sys/kernel.h>
55 #include <sys/malloc.h>
56 #include <sys/module.h>
57 #include <sys/socket.h>
58 #include <sys/socketvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/systm.h>
64 #include <netinet/tcp.h>
65 #include <netinet/tcp_seq.h>
66 #include <netinet/tcp_timer.h>
67 #include <netinet/tcp_var.h>
68 #include <netinet/cc/cc.h>
69 #include <netinet/cc/cc_cubic.h>
70 #include <netinet/cc/cc_module.h>
72 static void cubic_ack_received(struct cc_var *ccv, uint16_t type);
73 static void cubic_cb_destroy(struct cc_var *ccv);
74 static int cubic_cb_init(struct cc_var *ccv);
75 static void cubic_cong_signal(struct cc_var *ccv, uint32_t type);
76 static void cubic_conn_init(struct cc_var *ccv);
77 static int cubic_mod_init(void);
78 static void cubic_post_recovery(struct cc_var *ccv);
79 static void cubic_record_rtt(struct cc_var *ccv);
80 static void cubic_ssthresh_update(struct cc_var *ccv);
81 static void cubic_after_idle(struct cc_var *ccv);
84 /* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */
86 /* Sum of RTT samples across an epoch in ticks. */
87 int64_t sum_rtt_ticks;
88 /* cwnd at the most recent congestion event. */
89 unsigned long max_cwnd;
90 /* cwnd at the previous congestion event. */
91 unsigned long prev_max_cwnd;
94 #define CUBICFLAG_CONG_EVENT 0x00000001 /* congestion experienced */
95 #define CUBICFLAG_IN_SLOWSTART 0x00000002 /* in slow start */
96 #define CUBICFLAG_IN_APPLIMIT 0x00000004 /* application limited */
97 /* Minimum observed rtt in ticks. */
99 /* Mean observed rtt between congestion epochs. */
101 /* ACKs since last congestion event. */
103 /* Time of last congestion event in ticks. */
107 static MALLOC_DEFINE(M_CUBIC, "cubic data",
108 "Per connection data required for the CUBIC congestion control algorithm");
110 struct cc_algo cubic_cc_algo = {
112 .ack_received = cubic_ack_received,
113 .cb_destroy = cubic_cb_destroy,
114 .cb_init = cubic_cb_init,
115 .cong_signal = cubic_cong_signal,
116 .conn_init = cubic_conn_init,
117 .mod_init = cubic_mod_init,
118 .post_recovery = cubic_post_recovery,
119 .after_idle = cubic_after_idle,
123 cubic_ack_received(struct cc_var *ccv, uint16_t type)
125 struct cubic *cubic_data;
126 unsigned long w_tf, w_cubic_next;
127 int ticks_since_cong;
129 cubic_data = ccv->cc_data;
130 cubic_record_rtt(ccv);
133 * Regular ACK and we're not in cong/fast recovery and we're cwnd
134 * limited and we're either not doing ABC or are slow starting or are
135 * doing ABC and we've sent a cwnd's worth of bytes.
137 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
138 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
139 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
140 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
141 /* Use the logic in NewReno ack_received() for slow start. */
142 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
143 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE) {
144 cubic_data->flags |= CUBICFLAG_IN_SLOWSTART;
145 newreno_cc_algo.ack_received(ccv, type);
147 ticks_since_cong = ticks - cubic_data->t_last_cong;
149 if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART |
150 CUBICFLAG_IN_APPLIMIT)) {
151 cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART |
152 CUBICFLAG_IN_APPLIMIT);
153 cubic_data->t_last_cong = ticks;
157 * The mean RTT is used to best reflect the equations in
158 * the I-D. Using min_rtt in the tf_cwnd calculation
159 * causes w_tf to grow much faster than it should if the
160 * RTT is dominated by network buffering rather than
163 w_tf = tf_cwnd(ticks_since_cong,
164 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
167 w_cubic_next = cubic_cwnd(ticks_since_cong +
168 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
169 CCV(ccv, t_maxseg), cubic_data->K);
171 ccv->flags &= ~CCF_ABC_SENTAWND;
173 if (w_cubic_next < w_tf)
175 * TCP-friendly region, follow tf
178 CCV(ccv, snd_cwnd) = w_tf;
180 else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
182 * Concave or convex region, follow CUBIC
185 if (V_tcp_do_rfc3465)
186 CCV(ccv, snd_cwnd) = w_cubic_next;
188 CCV(ccv, snd_cwnd) += ((w_cubic_next -
189 CCV(ccv, snd_cwnd)) *
190 CCV(ccv, t_maxseg)) /
195 * If we're not in slow start and we're probing for a
196 * new cwnd limit at the start of a connection
197 * (happens when hostcache has a relevant entry),
198 * keep updating our current estimate of the
201 if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) &&
202 cubic_data->max_cwnd < CCV(ccv, snd_cwnd)) {
203 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
204 cubic_data->K = cubic_k(cubic_data->max_cwnd /
208 } else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
209 !(ccv->flags & CCF_CWND_LIMITED)) {
210 cubic_data->flags |= CUBICFLAG_IN_APPLIMIT;
215 * This is a Cubic specific implementation of after_idle.
216 * - Reset cwnd by calling New Reno implementation of after_idle.
217 * - Reset t_last_cong.
220 cubic_after_idle(struct cc_var *ccv)
222 struct cubic *cubic_data;
224 cubic_data = ccv->cc_data;
226 cubic_data->max_cwnd = ulmax(cubic_data->max_cwnd, CCV(ccv, snd_cwnd));
227 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
229 newreno_cc_algo.after_idle(ccv);
230 cubic_data->t_last_cong = ticks;
235 cubic_cb_destroy(struct cc_var *ccv)
237 free(ccv->cc_data, M_CUBIC);
241 cubic_cb_init(struct cc_var *ccv)
243 struct cubic *cubic_data;
245 cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
247 if (cubic_data == NULL)
250 /* Init some key variables with sensible defaults. */
251 cubic_data->t_last_cong = ticks;
252 cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
253 cubic_data->mean_rtt_ticks = 1;
255 ccv->cc_data = cubic_data;
261 * Perform any necessary tasks before we enter congestion recovery.
264 cubic_cong_signal(struct cc_var *ccv, uint32_t type)
266 struct cubic *cubic_data;
268 cubic_data = ccv->cc_data;
272 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
273 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
274 cubic_ssthresh_update(ccv);
275 cubic_data->flags |= CUBICFLAG_CONG_EVENT;
276 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
277 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
278 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
280 ENTER_RECOVERY(CCV(ccv, t_flags));
285 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
286 cubic_ssthresh_update(ccv);
287 cubic_data->flags |= CUBICFLAG_CONG_EVENT;
288 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
289 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
290 cubic_data->t_last_cong = ticks;
291 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
292 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
293 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
299 * Grab the current time and record it so we know when the
300 * most recent congestion event was. Only record it when the
301 * timeout has fired more than once, as there is a reasonable
302 * chance the first one is a false alarm and may not indicate
305 if (CCV(ccv, t_rxtshift) >= 2) {
306 cubic_data->flags |= CUBICFLAG_CONG_EVENT;
307 cubic_data->t_last_cong = ticks;
314 cubic_conn_init(struct cc_var *ccv)
316 struct cubic *cubic_data;
318 cubic_data = ccv->cc_data;
321 * Ensure we have a sane initial value for max_cwnd recorded. Without
322 * this here bad things happen when entries from the TCP hostcache
325 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
335 * Perform any necessary tasks before we exit congestion recovery.
338 cubic_post_recovery(struct cc_var *ccv)
340 struct cubic *cubic_data;
343 cubic_data = ccv->cc_data;
346 /* Fast convergence heuristic. */
347 if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
348 cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
351 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
353 * If inflight data is less than ssthresh, set cwnd
354 * conservatively to avoid a burst of data, as suggested in
355 * the NewReno RFC. Otherwise, use the CUBIC method.
357 * XXXLAS: Find a way to do this without needing curack
359 if (V_tcp_do_rfc6675_pipe)
360 pipe = tcp_compute_pipe(ccv->ccvc.tcp);
362 pipe = CCV(ccv, snd_max) - ccv->curack;
364 if (pipe < CCV(ccv, snd_ssthresh))
366 * Ensure that cwnd does not collapse to 1 MSS under
367 * adverse conditions. Implements RFC6582
369 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) +
372 /* Update cwnd based on beta and adjusted max_cwnd. */
373 CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->max_cwnd *
374 CUBIC_BETA) >> CUBIC_SHIFT,
375 2 * CCV(ccv, t_maxseg));
377 cubic_data->t_last_cong = ticks;
379 /* Calculate the average RTT between congestion epochs. */
380 if (cubic_data->epoch_ack_count > 0 &&
381 cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) {
382 cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
383 cubic_data->epoch_ack_count);
386 cubic_data->epoch_ack_count = 0;
387 cubic_data->sum_rtt_ticks = 0;
388 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
392 * Record the min RTT and sum samples for the epoch average RTT calculation.
395 cubic_record_rtt(struct cc_var *ccv)
397 struct cubic *cubic_data;
400 /* Ignore srtt until a min number of samples have been taken. */
401 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
402 cubic_data = ccv->cc_data;
403 t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
406 * Record the current SRTT as our minrtt if it's the smallest
407 * we've seen or minrtt is currently equal to its initialised
410 * XXXLAS: Should there be some hysteresis for minrtt?
412 if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
413 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) {
414 cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
417 * If the connection is within its first congestion
418 * epoch, ensure we prime mean_rtt_ticks with a
419 * reasonable value until the epoch average RTT is
420 * calculated in cubic_post_recovery().
422 if (cubic_data->min_rtt_ticks >
423 cubic_data->mean_rtt_ticks)
424 cubic_data->mean_rtt_ticks =
425 cubic_data->min_rtt_ticks;
428 /* Sum samples for epoch average RTT calculation. */
429 cubic_data->sum_rtt_ticks += t_srtt_ticks;
430 cubic_data->epoch_ack_count++;
435 * Update the ssthresh in the event of congestion.
438 cubic_ssthresh_update(struct cc_var *ccv)
440 struct cubic *cubic_data;
443 cubic_data = ccv->cc_data;
446 * On the first congestion event, set ssthresh to cwnd * 0.5, on
447 * subsequent congestion events, set it to cwnd * beta.
449 if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0)
450 ssthresh = CCV(ccv, snd_cwnd) >> 1;
452 ssthresh = ((uint64_t)CCV(ccv, snd_cwnd) *
453 CUBIC_BETA) >> CUBIC_SHIFT;
454 CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * CCV(ccv, t_maxseg));
458 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);