2 * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
3 * Copyright (c) 2010 The FreeBSD Foundation
6 * This software was developed by Lawrence Stewart while studying at the Centre
7 * for Advanced Internet Architectures, Swinburne University of Technology, made
8 * possible in part by a grant from the Cisco University Research Program Fund
9 * at Community Foundation Silicon Valley.
11 * Portions of this software were developed at the Centre for Advanced
12 * Internet Architectures, Swinburne University of Technology, Melbourne,
13 * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
24 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * An implementation of the CUBIC congestion control algorithm for FreeBSD,
39 * based on the Internet Draft "draft-rhee-tcpm-cubic-02" by Rhee, Xu and Ha.
40 * Originally released as part of the NewTCP research project at Swinburne
41 * University of Technology's Centre for Advanced Internet Architectures,
42 * Melbourne, Australia, which was made possible in part by a grant from the
43 * Cisco University Research Program Fund at Community Foundation Silicon
44 * Valley. More details are available at:
45 * http://caia.swin.edu.au/urp/newtcp/
48 #include <sys/cdefs.h>
49 __FBSDID("$FreeBSD$");
51 #include <sys/param.h>
52 #include <sys/kernel.h>
53 #include <sys/malloc.h>
54 #include <sys/module.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
62 #include <netinet/cc.h>
63 #include <netinet/tcp_seq.h>
64 #include <netinet/tcp_timer.h>
65 #include <netinet/tcp_var.h>
67 #include <netinet/cc/cc_cubic.h>
68 #include <netinet/cc/cc_module.h>
70 static void cubic_ack_received(struct cc_var *ccv, uint16_t type);
71 static void cubic_cb_destroy(struct cc_var *ccv);
72 static int cubic_cb_init(struct cc_var *ccv);
73 static void cubic_cong_signal(struct cc_var *ccv, uint32_t type);
74 static void cubic_conn_init(struct cc_var *ccv);
75 static int cubic_mod_init(void);
76 static void cubic_post_recovery(struct cc_var *ccv);
77 static void cubic_record_rtt(struct cc_var *ccv);
78 static void cubic_ssthresh_update(struct cc_var *ccv);
81 /* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */
83 /* Sum of RTT samples across an epoch in ticks. */
84 int64_t sum_rtt_ticks;
85 /* cwnd at the most recent congestion event. */
86 unsigned long max_cwnd;
87 /* cwnd at the previous congestion event. */
88 unsigned long prev_max_cwnd;
89 /* Number of congestion events. */
90 uint32_t num_cong_events;
91 /* Minimum observed rtt in ticks. */
93 /* Mean observed rtt between congestion epochs. */
95 /* ACKs since last congestion event. */
97 /* Time of last congestion event in ticks. */
101 static MALLOC_DEFINE(M_CUBIC, "cubic data",
102 "Per connection data required for the CUBIC congestion control algorithm");
104 struct cc_algo cubic_cc_algo = {
106 .ack_received = cubic_ack_received,
107 .cb_destroy = cubic_cb_destroy,
108 .cb_init = cubic_cb_init,
109 .cong_signal = cubic_cong_signal,
110 .conn_init = cubic_conn_init,
111 .mod_init = cubic_mod_init,
112 .post_recovery = cubic_post_recovery,
116 cubic_ack_received(struct cc_var *ccv, uint16_t type)
118 struct cubic *cubic_data;
119 unsigned long w_tf, w_cubic_next;
120 int ticks_since_cong;
122 cubic_data = ccv->cc_data;
123 cubic_record_rtt(ccv);
126 * Regular ACK and we're not in cong/fast recovery and we're cwnd
127 * limited and we're either not doing ABC or are slow starting or are
128 * doing ABC and we've sent a cwnd's worth of bytes.
130 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
131 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
132 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
133 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
134 /* Use the logic in NewReno ack_received() for slow start. */
135 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
136 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)
137 newreno_cc_algo.ack_received(ccv, type);
139 ticks_since_cong = ticks - cubic_data->t_last_cong;
142 * The mean RTT is used to best reflect the equations in
143 * the I-D. Using min_rtt in the tf_cwnd calculation
144 * causes w_tf to grow much faster than it should if the
145 * RTT is dominated by network buffering rather than
148 w_tf = tf_cwnd(ticks_since_cong,
149 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
152 w_cubic_next = cubic_cwnd(ticks_since_cong +
153 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
154 CCV(ccv, t_maxseg), cubic_data->K);
156 ccv->flags &= ~CCF_ABC_SENTAWND;
158 if (w_cubic_next < w_tf)
160 * TCP-friendly region, follow tf
163 CCV(ccv, snd_cwnd) = w_tf;
165 else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
167 * Concave or convex region, follow CUBIC
170 if (V_tcp_do_rfc3465)
171 CCV(ccv, snd_cwnd) = w_cubic_next;
173 CCV(ccv, snd_cwnd) += ((w_cubic_next -
174 CCV(ccv, snd_cwnd)) *
175 CCV(ccv, t_maxseg)) /
180 * If we're not in slow start and we're probing for a
181 * new cwnd limit at the start of a connection
182 * (happens when hostcache has a relevant entry),
183 * keep updating our current estimate of the
186 if (cubic_data->num_cong_events == 0 &&
187 cubic_data->max_cwnd < CCV(ccv, snd_cwnd))
188 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
194 cubic_cb_destroy(struct cc_var *ccv)
197 if (ccv->cc_data != NULL)
198 free(ccv->cc_data, M_CUBIC);
202 cubic_cb_init(struct cc_var *ccv)
204 struct cubic *cubic_data;
206 cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
208 if (cubic_data == NULL)
211 /* Init some key variables with sensible defaults. */
212 cubic_data->t_last_cong = ticks;
213 cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
214 cubic_data->mean_rtt_ticks = 1;
216 ccv->cc_data = cubic_data;
222 * Perform any necessary tasks before we enter congestion recovery.
225 cubic_cong_signal(struct cc_var *ccv, uint32_t type)
227 struct cubic *cubic_data;
229 cubic_data = ccv->cc_data;
233 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
234 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
235 cubic_ssthresh_update(ccv);
236 cubic_data->num_cong_events++;
237 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
238 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
240 ENTER_RECOVERY(CCV(ccv, t_flags));
245 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
246 cubic_ssthresh_update(ccv);
247 cubic_data->num_cong_events++;
248 cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
249 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
250 cubic_data->t_last_cong = ticks;
251 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
252 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
258 * Grab the current time and record it so we know when the
259 * most recent congestion event was. Only record it when the
260 * timeout has fired more than once, as there is a reasonable
261 * chance the first one is a false alarm and may not indicate
264 if (CCV(ccv, t_rxtshift) >= 2)
265 cubic_data->num_cong_events++;
266 cubic_data->t_last_cong = ticks;
272 cubic_conn_init(struct cc_var *ccv)
274 struct cubic *cubic_data;
276 cubic_data = ccv->cc_data;
279 * Ensure we have a sane initial value for max_cwnd recorded. Without
280 * this here bad things happen when entries from the TCP hostcache
283 cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
290 cubic_cc_algo.after_idle = newreno_cc_algo.after_idle;
296 * Perform any necessary tasks before we exit congestion recovery.
299 cubic_post_recovery(struct cc_var *ccv)
301 struct cubic *cubic_data;
303 cubic_data = ccv->cc_data;
305 /* Fast convergence heuristic. */
306 if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
307 cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
310 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
312 * If inflight data is less than ssthresh, set cwnd
313 * conservatively to avoid a burst of data, as suggested in
314 * the NewReno RFC. Otherwise, use the CUBIC method.
316 * XXXLAS: Find a way to do this without needing curack
318 if (SEQ_GT(ccv->curack + CCV(ccv, snd_ssthresh),
320 CCV(ccv, snd_cwnd) = CCV(ccv, snd_max) - ccv->curack +
323 /* Update cwnd based on beta and adjusted max_cwnd. */
324 CCV(ccv, snd_cwnd) = max(1, ((CUBIC_BETA *
325 cubic_data->max_cwnd) >> CUBIC_SHIFT));
327 cubic_data->t_last_cong = ticks;
329 /* Calculate the average RTT between congestion epochs. */
330 if (cubic_data->epoch_ack_count > 0 &&
331 cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) {
332 cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
333 cubic_data->epoch_ack_count);
336 cubic_data->epoch_ack_count = 0;
337 cubic_data->sum_rtt_ticks = 0;
338 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
342 * Record the min RTT and sum samples for the epoch average RTT calculation.
345 cubic_record_rtt(struct cc_var *ccv)
347 struct cubic *cubic_data;
350 /* Ignore srtt until a min number of samples have been taken. */
351 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
352 cubic_data = ccv->cc_data;
353 t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
356 * Record the current SRTT as our minrtt if it's the smallest
357 * we've seen or minrtt is currently equal to its initialised
360 * XXXLAS: Should there be some hysteresis for minrtt?
362 if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
363 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) {
364 cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
367 * If the connection is within its first congestion
368 * epoch, ensure we prime mean_rtt_ticks with a
369 * reasonable value until the epoch average RTT is
370 * calculated in cubic_post_recovery().
372 if (cubic_data->min_rtt_ticks >
373 cubic_data->mean_rtt_ticks)
374 cubic_data->mean_rtt_ticks =
375 cubic_data->min_rtt_ticks;
378 /* Sum samples for epoch average RTT calculation. */
379 cubic_data->sum_rtt_ticks += t_srtt_ticks;
380 cubic_data->epoch_ack_count++;
385 * Update the ssthresh in the event of congestion.
388 cubic_ssthresh_update(struct cc_var *ccv)
390 struct cubic *cubic_data;
392 cubic_data = ccv->cc_data;
395 * On the first congestion event, set ssthresh to cwnd * 0.5, on
396 * subsequent congestion events, set it to cwnd * beta.
398 if (cubic_data->num_cong_events == 0)
399 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd) >> 1;
401 CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_cwnd) * CUBIC_BETA)
406 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);