xref: /freebsd/sys/netinet/cc/cc_cubic.c (revision b4af4f93c682e445bf159f0d1ec90b636296c946)
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 	/* various flags */
94 	uint32_t	flags;
95 #define CUBICFLAG_CONG_EVENT	0x00000001	/* congestion experienced */
96 #define CUBICFLAG_IN_SLOWSTART	0x00000002	/* in slow start */
97 #define CUBICFLAG_IN_APPLIMIT	0x00000004	/* application limited */
98 	/* Minimum observed rtt in ticks. */
99 	int		min_rtt_ticks;
100 	/* Mean observed rtt between congestion epochs. */
101 	int		mean_rtt_ticks;
102 	/* ACKs since last congestion event. */
103 	int		epoch_ack_count;
104 	/* Time of last congestion event in ticks. */
105 	int		t_last_cong;
106 };
107 
108 static MALLOC_DEFINE(M_CUBIC, "cubic data",
109     "Per connection data required for the CUBIC congestion control algorithm");
110 
111 struct cc_algo cubic_cc_algo = {
112 	.name = "cubic",
113 	.ack_received = cubic_ack_received,
114 	.cb_destroy = cubic_cb_destroy,
115 	.cb_init = cubic_cb_init,
116 	.cong_signal = cubic_cong_signal,
117 	.conn_init = cubic_conn_init,
118 	.mod_init = cubic_mod_init,
119 	.post_recovery = cubic_post_recovery,
120 	.after_idle = cubic_after_idle,
121 };
122 
123 static void
124 cubic_ack_received(struct cc_var *ccv, uint16_t type)
125 {
126 	struct cubic *cubic_data;
127 	unsigned long w_tf, w_cubic_next;
128 	int ticks_since_cong;
129 
130 	cubic_data = ccv->cc_data;
131 	cubic_record_rtt(ccv);
132 
133 	/*
134 	 * Regular ACK and we're not in cong/fast recovery and we're cwnd
135 	 * limited and we're either not doing ABC or are slow starting or are
136 	 * doing ABC and we've sent a cwnd's worth of bytes.
137 	 */
138 	if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
139 	    (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 ||
140 	    CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
141 	    (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) {
142 		 /* Use the logic in NewReno ack_received() for slow start. */
143 		if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
144 		    cubic_data->min_rtt_ticks == TCPTV_SRTTBASE) {
145 			cubic_data->flags |= CUBICFLAG_IN_SLOWSTART;
146 			newreno_cc_algo.ack_received(ccv, type);
147 		} else {
148 			if ((ticks_since_cong =
149 			    ticks - cubic_data->t_last_cong) < 0) {
150 				/*
151 				 * dragging t_last_cong along
152 				 */
153 				ticks_since_cong = INT_MAX;
154 				cubic_data->t_last_cong = ticks - INT_MAX;
155 			}
156 
157 			if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART |
158 						 CUBICFLAG_IN_APPLIMIT)) {
159 				cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART |
160 						       CUBICFLAG_IN_APPLIMIT);
161 				cubic_data->t_last_cong = ticks;
162 				cubic_data->K = 0;
163 			}
164 			/*
165 			 * The mean RTT is used to best reflect the equations in
166 			 * the I-D. Using min_rtt in the tf_cwnd calculation
167 			 * causes w_tf to grow much faster than it should if the
168 			 * RTT is dominated by network buffering rather than
169 			 * propagation delay.
170 			 */
171 			w_tf = tf_cwnd(ticks_since_cong,
172 			    cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
173 			    CCV(ccv, t_maxseg));
174 
175 			w_cubic_next = cubic_cwnd(ticks_since_cong +
176 			    cubic_data->mean_rtt_ticks, cubic_data->max_cwnd,
177 			    CCV(ccv, t_maxseg), cubic_data->K);
178 
179 			ccv->flags &= ~CCF_ABC_SENTAWND;
180 
181 			if (w_cubic_next < w_tf) {
182 				/*
183 				 * TCP-friendly region, follow tf
184 				 * cwnd growth.
185 				 */
186 				if (CCV(ccv, snd_cwnd) < w_tf)
187 					CCV(ccv, snd_cwnd) = ulmin(w_tf, INT_MAX);
188 			} else if (CCV(ccv, snd_cwnd) < w_cubic_next) {
189 				/*
190 				 * Concave or convex region, follow CUBIC
191 				 * cwnd growth.
192 				 * Only update snd_cwnd, if it doesn't shrink.
193 				 */
194 				if (V_tcp_do_rfc3465)
195 					CCV(ccv, snd_cwnd) = ulmin(w_cubic_next,
196 					    INT_MAX);
197 				else
198 					CCV(ccv, snd_cwnd) += ulmax(1,
199 					    ((ulmin(w_cubic_next, INT_MAX) -
200 					    CCV(ccv, snd_cwnd)) *
201 					    CCV(ccv, t_maxseg)) /
202 					    CCV(ccv, snd_cwnd));
203 			}
204 
205 			/*
206 			 * If we're not in slow start and we're probing for a
207 			 * new cwnd limit at the start of a connection
208 			 * (happens when hostcache has a relevant entry),
209 			 * keep updating our current estimate of the
210 			 * max_cwnd.
211 			 */
212 			if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) &&
213 			    cubic_data->max_cwnd < CCV(ccv, snd_cwnd)) {
214 				cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
215 				cubic_data->K = cubic_k(cubic_data->max_cwnd /
216 				    CCV(ccv, t_maxseg));
217 			}
218 		}
219 	} else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
220 	    !(ccv->flags & CCF_CWND_LIMITED)) {
221 		cubic_data->flags |= CUBICFLAG_IN_APPLIMIT;
222 	}
223 }
224 
225 /*
226  * This is a Cubic specific implementation of after_idle.
227  *   - Reset cwnd by calling New Reno implementation of after_idle.
228  *   - Reset t_last_cong.
229  */
230 static void
231 cubic_after_idle(struct cc_var *ccv)
232 {
233 	struct cubic *cubic_data;
234 
235 	cubic_data = ccv->cc_data;
236 
237 	cubic_data->max_cwnd = ulmax(cubic_data->max_cwnd, CCV(ccv, snd_cwnd));
238 	cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
239 
240 	newreno_cc_algo.after_idle(ccv);
241 	cubic_data->t_last_cong = ticks;
242 }
243 
244 
245 static void
246 cubic_cb_destroy(struct cc_var *ccv)
247 {
248 	free(ccv->cc_data, M_CUBIC);
249 }
250 
251 static int
252 cubic_cb_init(struct cc_var *ccv)
253 {
254 	struct cubic *cubic_data;
255 
256 	cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO);
257 
258 	if (cubic_data == NULL)
259 		return (ENOMEM);
260 
261 	/* Init some key variables with sensible defaults. */
262 	cubic_data->t_last_cong = ticks;
263 	cubic_data->min_rtt_ticks = TCPTV_SRTTBASE;
264 	cubic_data->mean_rtt_ticks = 1;
265 
266 	ccv->cc_data = cubic_data;
267 
268 	return (0);
269 }
270 
271 /*
272  * Perform any necessary tasks before we enter congestion recovery.
273  */
274 static void
275 cubic_cong_signal(struct cc_var *ccv, uint32_t type)
276 {
277 	struct cubic *cubic_data;
278 
279 	cubic_data = ccv->cc_data;
280 
281 	switch (type) {
282 	case CC_NDUPACK:
283 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
284 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
285 				cubic_ssthresh_update(ccv);
286 				cubic_data->flags |= CUBICFLAG_CONG_EVENT;
287 				cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
288 				cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
289 				cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
290 			}
291 			ENTER_RECOVERY(CCV(ccv, t_flags));
292 		}
293 		break;
294 
295 	case CC_ECN:
296 		if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
297 			cubic_ssthresh_update(ccv);
298 			cubic_data->flags |= CUBICFLAG_CONG_EVENT;
299 			cubic_data->prev_max_cwnd = cubic_data->max_cwnd;
300 			cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
301 			cubic_data->t_last_cong = ticks;
302 			cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
303 			CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
304 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
305 		}
306 		break;
307 
308 	case CC_RTO:
309 		/*
310 		 * Grab the current time and record it so we know when the
311 		 * most recent congestion event was. Only record it when the
312 		 * timeout has fired more than once, as there is a reasonable
313 		 * chance the first one is a false alarm and may not indicate
314 		 * congestion.
315 		 * This will put Cubic firmly into the concave / TCP friendly
316 		 * region, for a slower ramp-up after two consecutive RTOs.
317 		 */
318 		if (CCV(ccv, t_rxtshift) >= 2) {
319 			cubic_data->flags |= CUBICFLAG_CONG_EVENT;
320 			cubic_data->t_last_cong = ticks;
321 			cubic_data->max_cwnd = CCV(ccv, snd_cwnd_prev);
322 			cubic_data->K = cubic_k(cubic_data->max_cwnd /
323 						CCV(ccv, t_maxseg));
324 		}
325 		break;
326 	}
327 }
328 
329 static void
330 cubic_conn_init(struct cc_var *ccv)
331 {
332 	struct cubic *cubic_data;
333 
334 	cubic_data = ccv->cc_data;
335 
336 	/*
337 	 * Ensure we have a sane initial value for max_cwnd recorded. Without
338 	 * this here bad things happen when entries from the TCP hostcache
339 	 * get used.
340 	 */
341 	cubic_data->max_cwnd = CCV(ccv, snd_cwnd);
342 }
343 
344 static int
345 cubic_mod_init(void)
346 {
347 	return (0);
348 }
349 
350 /*
351  * Perform any necessary tasks before we exit congestion recovery.
352  */
353 static void
354 cubic_post_recovery(struct cc_var *ccv)
355 {
356 	struct cubic *cubic_data;
357 	int pipe;
358 
359 	cubic_data = ccv->cc_data;
360 	pipe = 0;
361 
362 	/* Fast convergence heuristic. */
363 	if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd)
364 		cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR)
365 		    >> CUBIC_SHIFT;
366 
367 	if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
368 		/*
369 		 * If inflight data is less than ssthresh, set cwnd
370 		 * conservatively to avoid a burst of data, as suggested in
371 		 * the NewReno RFC. Otherwise, use the CUBIC method.
372 		 *
373 		 * XXXLAS: Find a way to do this without needing curack
374 		 */
375 		if (V_tcp_do_rfc6675_pipe)
376 			pipe = tcp_compute_pipe(ccv->ccvc.tcp);
377 		else
378 			pipe = CCV(ccv, snd_max) - ccv->curack;
379 
380 		if (pipe < CCV(ccv, snd_ssthresh))
381 			/*
382 			 * Ensure that cwnd does not collapse to 1 MSS under
383 			 * adverse conditions. Implements RFC6582
384 			 */
385 			CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) +
386 			    CCV(ccv, t_maxseg);
387 		else
388 			/* Update cwnd based on beta and adjusted max_cwnd. */
389 			CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->max_cwnd *
390 			    CUBIC_BETA) >> CUBIC_SHIFT,
391 			    2 * CCV(ccv, t_maxseg));
392 	}
393 	cubic_data->t_last_cong = ticks;
394 
395 	/* Calculate the average RTT between congestion epochs. */
396 	if (cubic_data->epoch_ack_count > 0 &&
397 	    cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) {
398 		cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks /
399 		    cubic_data->epoch_ack_count);
400 	}
401 
402 	cubic_data->epoch_ack_count = 0;
403 	cubic_data->sum_rtt_ticks = 0;
404 	cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg));
405 }
406 
407 /*
408  * Record the min RTT and sum samples for the epoch average RTT calculation.
409  */
410 static void
411 cubic_record_rtt(struct cc_var *ccv)
412 {
413 	struct cubic *cubic_data;
414 	int t_srtt_ticks;
415 
416 	/* Ignore srtt until a min number of samples have been taken. */
417 	if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
418 		cubic_data = ccv->cc_data;
419 		t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE;
420 
421 		/*
422 		 * Record the current SRTT as our minrtt if it's the smallest
423 		 * we've seen or minrtt is currently equal to its initialised
424 		 * value.
425 		 *
426 		 * XXXLAS: Should there be some hysteresis for minrtt?
427 		 */
428 		if ((t_srtt_ticks < cubic_data->min_rtt_ticks ||
429 		    cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) {
430 			cubic_data->min_rtt_ticks = max(1, t_srtt_ticks);
431 
432 			/*
433 			 * If the connection is within its first congestion
434 			 * epoch, ensure we prime mean_rtt_ticks with a
435 			 * reasonable value until the epoch average RTT is
436 			 * calculated in cubic_post_recovery().
437 			 */
438 			if (cubic_data->min_rtt_ticks >
439 			    cubic_data->mean_rtt_ticks)
440 				cubic_data->mean_rtt_ticks =
441 				    cubic_data->min_rtt_ticks;
442 		}
443 
444 		/* Sum samples for epoch average RTT calculation. */
445 		cubic_data->sum_rtt_ticks += t_srtt_ticks;
446 		cubic_data->epoch_ack_count++;
447 	}
448 }
449 
450 /*
451  * Update the ssthresh in the event of congestion.
452  */
453 static void
454 cubic_ssthresh_update(struct cc_var *ccv)
455 {
456 	struct cubic *cubic_data;
457 	uint32_t ssthresh;
458 
459 	cubic_data = ccv->cc_data;
460 
461 	/*
462 	 * On the first congestion event, set ssthresh to cwnd * 0.5, on
463 	 * subsequent congestion events, set it to cwnd * beta.
464 	 */
465 	if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0)
466 		ssthresh = CCV(ccv, snd_cwnd) >> 1;
467 	else
468 		ssthresh = ((uint64_t)CCV(ccv, snd_cwnd) *
469 		    CUBIC_BETA) >> CUBIC_SHIFT;
470 	CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * CCV(ccv, t_maxseg));
471 }
472 
473 
474 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);
475 MODULE_VERSION(cubic, 1);
476