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);
83 /* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */
85 /* Sum of RTT samples across an epoch in ticks. */
86 int64_t sum_rtt_ticks;
87 /* cwnd at the most recent congestion event. */
88 unsigned long max_cwnd;
89 /* cwnd at the previous congestion event. */
90 unsigned long prev_max_cwnd;
91 /* Number of congestion events. */
92 uint32_t num_cong_events;
93 /* Minimum observed rtt in ticks. */
95 /* Mean observed rtt between congestion epochs. */
97 /* ACKs since last congestion event. */
99 /* Time of last congestion event in ticks. */
103 static MALLOC_DEFINE(M_CUBIC, "cubic data",
104 "Per connection data required for the CUBIC congestion control algorithm");
106 struct cc_algo cubic_cc_algo = {
108 .ack_received = cubic_ack_received,
109 .cb_destroy = cubic_cb_destroy,
110 .cb_init = cubic_cb_init,
111 .cong_signal = cubic_cong_signal,
112 .conn_init = cubic_conn_init,
113 .mod_init = cubic_mod_init,
114 .post_recovery = cubic_post_recovery,
118 cubic_ack_received(struct cc_var *ccv, uint16_t type)
120 struct cubic *cubic_data;
121 unsigned long w_tf, w_cubic_next;
122 int ticks_since_cong;
124 cubic_data = ccv->cc_data;
125 cubic_record_rtt(ccv);
128 * Regular ACK and we're not in cong/fast recovery and we're cwnd
129 * limited and we're either not doing ABC or are slow starting or are
130 * doing ABC and we've sent a cwnd's worth of bytes.
132 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
133 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
134 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
135 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
136 /* Use the logic in NewReno ack_received() for slow start. */
137 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
138 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)
139 newreno_cc_algo.ack_received(ccv, type);
141 ticks_since_cong = ticks - cubic_data->t_last_cong;
144 * The mean RTT is used to best reflect the equations in
145 * the I-D. Using min_rtt in the tf_cwnd calculation
146 * causes w_tf to grow much faster than it should if the
147 * RTT is dominated by network buffering rather than
150 w_tf = tf_cwnd(ticks_since_cong,
151 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
154 w_cubic_next = cubic_cwnd(ticks_since_cong +
155 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
156 CCV(ccv, t_maxseg), cubic_data->K);
158 ccv->flags &= ~CCF_ABC_SENTAWND;
160 if (w_cubic_next < w_tf)
162 * TCP-friendly region, follow tf
165 CCV(ccv, snd_cwnd) = w_tf;
167 else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
169 * Concave or convex region, follow CUBIC
172 if (V_tcp_do_rfc3465)
173 CCV(ccv, snd_cwnd) = w_cubic_next;
175 CCV(ccv, snd_cwnd) += ((w_cubic_next -
176 CCV(ccv, snd_cwnd)) *
177 CCV(ccv, t_maxseg)) /
182 * If we're not in slow start and we're probing for a
183 * new cwnd limit at the start of a connection
184 * (happens when hostcache has a relevant entry),
185 * keep updating our current estimate of the
188 if (cubic_data->num_cong_events == 0 &&
189 cubic_data->max_cwnd < CCV(ccv, snd_cwnd))
190 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
196 cubic_cb_destroy(struct cc_var *ccv)
199 if (ccv->cc_data != NULL)
200 free(ccv->cc_data, M_CUBIC);
204 cubic_cb_init(struct cc_var *ccv)
206 struct cubic *cubic_data;
208 cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
210 if (cubic_data == NULL)
213 /* Init some key variables with sensible defaults. */
214 cubic_data->t_last_cong = ticks;
215 cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
216 cubic_data->mean_rtt_ticks = 1;
218 ccv->cc_data = cubic_data;
224 * Perform any necessary tasks before we enter congestion recovery.
227 cubic_cong_signal(struct cc_var *ccv, uint32_t type)
229 struct cubic *cubic_data;
231 cubic_data = ccv->cc_data;
235 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
236 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
237 cubic_ssthresh_update(ccv);
238 cubic_data->num_cong_events++;
239 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
240 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
242 ENTER_RECOVERY(CCV(ccv, t_flags));
247 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
248 cubic_ssthresh_update(ccv);
249 cubic_data->num_cong_events++;
250 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
251 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
252 cubic_data->t_last_cong = ticks;
253 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
254 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
260 * Grab the current time and record it so we know when the
261 * most recent congestion event was. Only record it when the
262 * timeout has fired more than once, as there is a reasonable
263 * chance the first one is a false alarm and may not indicate
266 if (CCV(ccv, t_rxtshift) >= 2) {
267 cubic_data->num_cong_events++;
268 cubic_data->t_last_cong = ticks;
275 cubic_conn_init(struct cc_var *ccv)
277 struct cubic *cubic_data;
279 cubic_data = ccv->cc_data;
282 * Ensure we have a sane initial value for max_cwnd recorded. Without
283 * this here bad things happen when entries from the TCP hostcache
286 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
293 cubic_cc_algo.after_idle = newreno_cc_algo.after_idle;
299 * Perform any necessary tasks before we exit congestion recovery.
302 cubic_post_recovery(struct cc_var *ccv)
304 struct cubic *cubic_data;
307 cubic_data = ccv->cc_data;
310 /* Fast convergence heuristic. */
311 if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
312 cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
315 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
317 * If inflight data is less than ssthresh, set cwnd
318 * conservatively to avoid a burst of data, as suggested in
319 * the NewReno RFC. Otherwise, use the CUBIC method.
321 * XXXLAS: Find a way to do this without needing curack
323 if (V_tcp_do_rfc6675_pipe)
324 pipe = tcp_compute_pipe(ccv->ccvc.tcp);
326 pipe = CCV(ccv, snd_max) - ccv->curack;
328 if (pipe < CCV(ccv, snd_ssthresh))
329 CCV(ccv, snd_cwnd) = pipe + CCV(ccv, t_maxseg);
331 /* Update cwnd based on beta and adjusted max_cwnd. */
332 CCV(ccv, snd_cwnd) = max(1, ((CUBIC_BETA *
333 cubic_data->max_cwnd) >> CUBIC_SHIFT));
335 cubic_data->t_last_cong = ticks;
337 /* Calculate the average RTT between congestion epochs. */
338 if (cubic_data->epoch_ack_count > 0 &&
339 cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) {
340 cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
341 cubic_data->epoch_ack_count);
344 cubic_data->epoch_ack_count = 0;
345 cubic_data->sum_rtt_ticks = 0;
346 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
350 * Record the min RTT and sum samples for the epoch average RTT calculation.
353 cubic_record_rtt(struct cc_var *ccv)
355 struct cubic *cubic_data;
358 /* Ignore srtt until a min number of samples have been taken. */
359 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
360 cubic_data = ccv->cc_data;
361 t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
364 * Record the current SRTT as our minrtt if it's the smallest
365 * we've seen or minrtt is currently equal to its initialised
368 * XXXLAS: Should there be some hysteresis for minrtt?
370 if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
371 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) {
372 cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
375 * If the connection is within its first congestion
376 * epoch, ensure we prime mean_rtt_ticks with a
377 * reasonable value until the epoch average RTT is
378 * calculated in cubic_post_recovery().
380 if (cubic_data->min_rtt_ticks >
381 cubic_data->mean_rtt_ticks)
382 cubic_data->mean_rtt_ticks =
383 cubic_data->min_rtt_ticks;
386 /* Sum samples for epoch average RTT calculation. */
387 cubic_data->sum_rtt_ticks += t_srtt_ticks;
388 cubic_data->epoch_ack_count++;
393 * Update the ssthresh in the event of congestion.
396 cubic_ssthresh_update(struct cc_var *ccv)
398 struct cubic *cubic_data;
400 cubic_data = ccv->cc_data;
403 * On the first congestion event, set ssthresh to cwnd * 0.5, on
404 * subsequent congestion events, set it to cwnd * beta.
406 if (cubic_data->num_cong_events == 0)
407 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd) >> 1;
409 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) *
410 CUBIC_BETA) >> CUBIC_SHIFT;
414 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);