xref: /freebsd/sys/netinet/cc/cc_cubic.c (revision adc56f5a383771f594829b7db9c263b6f0dcf1bd)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
5  * Copyright (c) 2010 The FreeBSD Foundation
6  * All rights reserved.
7  *
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.
12  *
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.
16  *
17  * Redistribution and use in source and binary forms, with or without
18  * modification, are permitted provided that the following conditions
19  * are met:
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.
25  *
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
36  * SUCH DAMAGE.
37  */
38 
39 /*
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/
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 #include <sys/param.h>
54 #include <sys/kernel.h>
55 #include <sys/limits.h>
56 #include <sys/malloc.h>
57 #include <sys/module.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
60 #include <sys/sysctl.h>
61 #include <sys/systm.h>
62 
63 #include <net/vnet.h>
64 
65 #include <netinet/tcp.h>
66 #include <netinet/tcp_seq.h>
67 #include <netinet/tcp_timer.h>
68 #include <netinet/tcp_var.h>
69 #include <netinet/cc/cc.h>
70 #include <netinet/cc/cc_cubic.h>
71 #include <netinet/cc/cc_module.h>
72 
73 static void	cubic_ack_received(struct cc_var *ccv, uint16_t type);
74 static void	cubic_cb_destroy(struct cc_var *ccv);
75 static int	cubic_cb_init(struct cc_var *ccv);
76 static void	cubic_cong_signal(struct cc_var *ccv, uint32_t type);
77 static void	cubic_conn_init(struct cc_var *ccv);
78 static int	cubic_mod_init(void);
79 static void	cubic_post_recovery(struct cc_var *ccv);
80 static void	cubic_record_rtt(struct cc_var *ccv);
81 static void	cubic_ssthresh_update(struct cc_var *ccv);
82 static void	cubic_after_idle(struct cc_var *ccv);
83 
84 struct cubic {
85 	/* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */
86 	int64_t		K;
87 	/* Sum of RTT samples across an epoch in ticks. */
88 	int64_t		sum_rtt_ticks;
89 	/* cwnd at the most recent congestion event. */
90 	unsigned long	max_cwnd;
91 	/* cwnd at the previous congestion event. */
92 	unsigned long	prev_max_cwnd;
93 	/* Number of congestion events. */
94 	uint32_t	num_cong_events;
95 	/* Minimum observed rtt in ticks. */
96 	int		min_rtt_ticks;
97 	/* Mean observed rtt between congestion epochs. */
98 	int		mean_rtt_ticks;
99 	/* ACKs since last congestion event. */
100 	int		epoch_ack_count;
101 	/* Time of last congestion event in ticks. */
102 	int		t_last_cong;
103 };
104 
105 static MALLOC_DEFINE(M_CUBIC, "cubic data",
106     "Per connection data required for the CUBIC congestion control algorithm");
107 
108 struct cc_algo cubic_cc_algo = {
109 	.name = "cubic",
110 	.ack_received = cubic_ack_received,
111 	.cb_destroy = cubic_cb_destroy,
112 	.cb_init = cubic_cb_init,
113 	.cong_signal = cubic_cong_signal,
114 	.conn_init = cubic_conn_init,
115 	.mod_init = cubic_mod_init,
116 	.post_recovery = cubic_post_recovery,
117 	.after_idle = cubic_after_idle,
118 };
119 
120 static void
121 cubic_ack_received(struct cc_var *ccv, uint16_t type)
122 {
123 	struct cubic *cubic_data;
124 	unsigned long w_tf, w_cubic_next;
125 	int ticks_since_cong;
126 
127 	cubic_data = ccv->cc_data;
128 	cubic_record_rtt(ccv);
129 
130 	/*
131 	 * Regular ACK and we're not in cong/fast recovery and we're cwnd
132 	 * limited and we're either not doing ABC or are slow starting or are
133 	 * doing ABC and we've sent a cwnd's worth of bytes.
134 	 */
135 	if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
136 	    (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
137 	    CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
138 	    (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
139 		 /* Use the logic in NewReno ack_received() for slow start. */
140 		if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
141 		    cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)
142 			newreno_cc_algo.ack_received(ccv, type);
143 		else {
144 			if ((ticks_since_cong =
145 			    ticks - cubic_data->t_last_cong) < 0) {
146 				/*
147 				 * dragging t_last_cong along
148 				 */
149 				ticks_since_cong = INT_MAX;
150 				cubic_data->t_last_cong = ticks - INT_MAX;
151 			}
152 
153 			/*
154 			 * The mean RTT is used to best reflect the equations in
155 			 * the I-D. Using min_rtt in the tf_cwnd calculation
156 			 * causes w_tf to grow much faster than it should if the
157 			 * RTT is dominated by network buffering rather than
158 			 * propagation delay.
159 			 */
160 			w_tf = tf_cwnd(ticks_since_cong,
161 			    cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
162 			    CCV(ccv, t_maxseg));
163 
164 			w_cubic_next = cubic_cwnd(ticks_since_cong +
165 			    cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
166 			    CCV(ccv, t_maxseg), cubic_data->K);
167 
168 			ccv->flags &= ~CCF_ABC_SENTAWND;
169 
170 			if (w_cubic_next < w_tf) {
171 				/*
172 				 * TCP-friendly region, follow tf
173 				 * cwnd growth.
174 				 */
175 				if (CCV(ccv, snd_cwnd) < w_tf)
176 					CCV(ccv, snd_cwnd) = ulmin(w_tf, INT_MAX);
177 			}
178 
179 			else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
180 				/*
181 				 * Concave or convex region, follow CUBIC
182 				 * cwnd growth.
183 				 */
184 				if (V_tcp_do_rfc3465)
185 					CCV(ccv, snd_cwnd) = ulmin(w_cubic_next,
186 					    INT_MAX);
187 				else
188 					CCV(ccv, snd_cwnd) += ulmax(1,
189 					    ((ulmin(w_cubic_next, INT_MAX) -
190 					    CCV(ccv, snd_cwnd)) *
191 					    CCV(ccv, t_maxseg)) /
192 					    CCV(ccv, snd_cwnd));
193 			}
194 
195 			/*
196 			 * If we're not in slow start and we're probing for a
197 			 * new cwnd limit at the start of a connection
198 			 * (happens when hostcache has a relevant entry),
199 			 * keep updating our current estimate of the
200 			 * max_cwnd.
201 			 */
202 			if (cubic_data->num_cong_events == 0 &&
203 			    cubic_data->max_cwnd < CCV(ccv, snd_cwnd))
204 				cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
205 				cubic_data->K = cubic_k(cubic_data->max_cwnd /
206 				    CCV(ccv, t_maxseg));
207 		}
208 	}
209 }
210 
211 /*
212  * This is a Cubic specific implementation of after_idle.
213  *   - Reset cwnd by calling New Reno implementation of after_idle.
214  *   - Reset t_last_cong.
215  */
216 static void
217 cubic_after_idle(struct cc_var *ccv)
218 {
219 	struct cubic *cubic_data;
220 
221 	cubic_data = ccv->cc_data;
222 
223 	cubic_data->max_cwnd = ulmax(cubic_data->max_cwnd, CCV(ccv, snd_cwnd));
224 	cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
225 
226 	newreno_cc_algo.after_idle(ccv);
227 	cubic_data->t_last_cong = ticks;
228 }
229 
230 
231 static void
232 cubic_cb_destroy(struct cc_var *ccv)
233 {
234 	free(ccv->cc_data, M_CUBIC);
235 }
236 
237 static int
238 cubic_cb_init(struct cc_var *ccv)
239 {
240 	struct cubic *cubic_data;
241 
242 	cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
243 
244 	if (cubic_data == NULL)
245 		return (ENOMEM);
246 
247 	/* Init some key variables with sensible defaults. */
248 	cubic_data->t_last_cong = ticks;
249 	cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
250 	cubic_data->mean_rtt_ticks = 1;
251 
252 	ccv->cc_data = cubic_data;
253 
254 	return (0);
255 }
256 
257 /*
258  * Perform any necessary tasks before we enter congestion recovery.
259  */
260 static void
261 cubic_cong_signal(struct cc_var *ccv, uint32_t type)
262 {
263 	struct cubic *cubic_data;
264 
265 	cubic_data = ccv->cc_data;
266 
267 	switch (type) {
268 	case CC_NDUPACK:
269 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
270 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
271 				cubic_ssthresh_update(ccv);
272 				cubic_data->num_cong_events++;
273 				cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
274 				cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
275 			}
276 			ENTER_RECOVERY(CCV(ccv, t_flags));
277 		}
278 		break;
279 
280 	case CC_ECN:
281 		if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
282 			cubic_ssthresh_update(ccv);
283 			cubic_data->num_cong_events++;
284 			cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
285 			cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
286 			cubic_data->t_last_cong = ticks;
287 			CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
288 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
289 		}
290 		break;
291 
292 	case CC_RTO:
293 		/*
294 		 * Grab the current time and record it so we know when the
295 		 * most recent congestion event was. Only record it when the
296 		 * timeout has fired more than once, as there is a reasonable
297 		 * chance the first one is a false alarm and may not indicate
298 		 * congestion.
299 		 */
300 		if (CCV(ccv, t_rxtshift) >= 2) {
301 			cubic_data->num_cong_events++;
302 			cubic_data->t_last_cong = ticks;
303 		}
304 		break;
305 	}
306 }
307 
308 static void
309 cubic_conn_init(struct cc_var *ccv)
310 {
311 	struct cubic *cubic_data;
312 
313 	cubic_data = ccv->cc_data;
314 
315 	/*
316 	 * Ensure we have a sane initial value for max_cwnd recorded. Without
317 	 * this here bad things happen when entries from the TCP hostcache
318 	 * get used.
319 	 */
320 	cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
321 }
322 
323 static int
324 cubic_mod_init(void)
325 {
326 	return (0);
327 }
328 
329 /*
330  * Perform any necessary tasks before we exit congestion recovery.
331  */
332 static void
333 cubic_post_recovery(struct cc_var *ccv)
334 {
335 	struct cubic *cubic_data;
336 	int pipe;
337 
338 	cubic_data = ccv->cc_data;
339 	pipe = 0;
340 
341 	/* Fast convergence heuristic. */
342 	if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
343 		cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
344 		    >> CUBIC_SHIFT;
345 
346 	if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
347 		/*
348 		 * If inflight data is less than ssthresh, set cwnd
349 		 * conservatively to avoid a burst of data, as suggested in
350 		 * the NewReno RFC. Otherwise, use the CUBIC method.
351 		 *
352 		 * XXXLAS: Find a way to do this without needing curack
353 		 */
354 		if (V_tcp_do_rfc6675_pipe)
355 			pipe = tcp_compute_pipe(ccv->ccvc.tcp);
356 		else
357 			pipe = CCV(ccv, snd_max) - ccv->curack;
358 
359 		if (pipe < CCV(ccv, snd_ssthresh))
360 			/*
361 			 * Ensure that cwnd does not collapse to 1 MSS under
362 			 * adverse conditions. Implements RFC6582
363 			 */
364 			CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) +
365 			    CCV(ccv, t_maxseg);
366 		else
367 			/* Update cwnd based on beta and adjusted max_cwnd. */
368 			CCV(ccv, snd_cwnd) = max(1, ((CUBIC_BETA *
369 			    cubic_data->max_cwnd) >> CUBIC_SHIFT));
370 	}
371 	cubic_data->t_last_cong = ticks;
372 
373 	/* Calculate the average RTT between congestion epochs. */
374 	if (cubic_data->epoch_ack_count > 0 &&
375 	    cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) {
376 		cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
377 		    cubic_data->epoch_ack_count);
378 	}
379 
380 	cubic_data->epoch_ack_count = 0;
381 	cubic_data->sum_rtt_ticks = 0;
382 	cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
383 }
384 
385 /*
386  * Record the min RTT and sum samples for the epoch average RTT calculation.
387  */
388 static void
389 cubic_record_rtt(struct cc_var *ccv)
390 {
391 	struct cubic *cubic_data;
392 	int t_srtt_ticks;
393 
394 	/* Ignore srtt until a min number of samples have been taken. */
395 	if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
396 		cubic_data = ccv->cc_data;
397 		t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
398 
399 		/*
400 		 * Record the current SRTT as our minrtt if it's the smallest
401 		 * we've seen or minrtt is currently equal to its initialised
402 		 * value.
403 		 *
404 		 * XXXLAS: Should there be some hysteresis for minrtt?
405 		 */
406 		if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
407 		    cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) {
408 			cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
409 
410 			/*
411 			 * If the connection is within its first congestion
412 			 * epoch, ensure we prime mean_rtt_ticks with a
413 			 * reasonable value until the epoch average RTT is
414 			 * calculated in cubic_post_recovery().
415 			 */
416 			if (cubic_data->min_rtt_ticks >
417 			    cubic_data->mean_rtt_ticks)
418 				cubic_data->mean_rtt_ticks =
419 				    cubic_data->min_rtt_ticks;
420 		}
421 
422 		/* Sum samples for epoch average RTT calculation. */
423 		cubic_data->sum_rtt_ticks += t_srtt_ticks;
424 		cubic_data->epoch_ack_count++;
425 	}
426 }
427 
428 /*
429  * Update the ssthresh in the event of congestion.
430  */
431 static void
432 cubic_ssthresh_update(struct cc_var *ccv)
433 {
434 	struct cubic *cubic_data;
435 
436 	cubic_data = ccv->cc_data;
437 
438 	/*
439 	 * On the first congestion event, set ssthresh to cwnd * 0.5, on
440 	 * subsequent congestion events, set it to cwnd * beta.
441 	 */
442 	if (cubic_data->num_cong_events == 0)
443 		CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd) >> 1;
444 	else
445 		CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) *
446 		    CUBIC_BETA) >> CUBIC_SHIFT;
447 }
448 
449 
450 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);
451