blob: 272a84593c855729d3d687ea3071ae9c800c444d [file] [log] [blame]
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001/*
2 * Plugable TCP congestion control support and newReno
3 * congestion control.
4 * Based on ideas from I/O scheduler suport and Web100.
5 *
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
7 */
8
Joe Perchesafd465032012-03-12 07:03:32 +00009#define pr_fmt(fmt) "TCP: " fmt
10
Stephen Hemminger317a76f2005-06-23 12:19:55 -070011#include <linux/module.h>
12#include <linux/mm.h>
13#include <linux/types.h>
14#include <linux/list.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090015#include <linux/gfp.h>
Stephen Hemminger317a76f2005-06-23 12:19:55 -070016#include <net/tcp.h>
17
John Heffner886236c2007-03-25 19:21:45 -070018int sysctl_tcp_max_ssthresh = 0;
19
Stephen Hemminger317a76f2005-06-23 12:19:55 -070020static DEFINE_SPINLOCK(tcp_cong_list_lock);
21static LIST_HEAD(tcp_cong_list);
22
23/* Simple linear search, don't expect many entries! */
24static struct tcp_congestion_ops *tcp_ca_find(const char *name)
25{
26 struct tcp_congestion_ops *e;
27
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070028 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -070029 if (strcmp(e->name, name) == 0)
30 return e;
31 }
32
33 return NULL;
34}
35
36/*
Robert P. J. Dayd08df602007-02-17 19:07:33 +010037 * Attach new congestion control algorithm to the list
Stephen Hemminger317a76f2005-06-23 12:19:55 -070038 * of available options.
39 */
40int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
41{
42 int ret = 0;
43
44 /* all algorithms must implement ssthresh and cong_avoid ops */
Stephen Hemminger72dc5b92006-06-05 17:30:08 -070045 if (!ca->ssthresh || !ca->cong_avoid) {
Joe Perchesafd465032012-03-12 07:03:32 +000046 pr_err("%s does not implement required ops\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070047 return -EINVAL;
48 }
49
50 spin_lock(&tcp_cong_list_lock);
51 if (tcp_ca_find(ca->name)) {
Joe Perchesafd465032012-03-12 07:03:32 +000052 pr_notice("%s already registered\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070053 ret = -EEXIST;
54 } else {
Stephen Hemminger3d2573f2006-09-24 20:11:58 -070055 list_add_tail_rcu(&ca->list, &tcp_cong_list);
Joe Perchesafd465032012-03-12 07:03:32 +000056 pr_info("%s registered\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070057 }
58 spin_unlock(&tcp_cong_list_lock);
59
60 return ret;
61}
62EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
63
64/*
65 * Remove congestion control algorithm, called from
66 * the module's remove function. Module ref counts are used
67 * to ensure that this can't be done till all sockets using
68 * that method are closed.
69 */
70void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
71{
72 spin_lock(&tcp_cong_list_lock);
73 list_del_rcu(&ca->list);
74 spin_unlock(&tcp_cong_list_lock);
75}
76EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
77
78/* Assign choice of congestion control. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030079void tcp_init_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -070080{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030081 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070082 struct tcp_congestion_ops *ca;
83
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070084 /* if no choice made yet assign the current value set as default */
85 if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
86 rcu_read_lock();
87 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
88 if (try_module_get(ca->owner)) {
89 icsk->icsk_ca_ops = ca;
90 break;
91 }
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070092
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070093 /* fallback to next available */
Stephen Hemminger317a76f2005-06-23 12:19:55 -070094 }
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070095 rcu_read_unlock();
Stephen Hemminger317a76f2005-06-23 12:19:55 -070096 }
Stephen Hemminger317a76f2005-06-23 12:19:55 -070097
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030098 if (icsk->icsk_ca_ops->init)
99 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700100}
101
102/* Manage refcounts on socket close. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300103void tcp_cleanup_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700104{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300105 struct inet_connection_sock *icsk = inet_csk(sk);
106
107 if (icsk->icsk_ca_ops->release)
108 icsk->icsk_ca_ops->release(sk);
109 module_put(icsk->icsk_ca_ops->owner);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700110}
111
112/* Used by sysctl to change default congestion control */
113int tcp_set_default_congestion_control(const char *name)
114{
115 struct tcp_congestion_ops *ca;
116 int ret = -ENOENT;
117
118 spin_lock(&tcp_cong_list_lock);
119 ca = tcp_ca_find(name);
Johannes Berg95a5afc2008-10-16 15:24:51 -0700120#ifdef CONFIG_MODULES
Eric Parisa8f80e82009-08-13 09:44:51 -0400121 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700122 spin_unlock(&tcp_cong_list_lock);
123
124 request_module("tcp_%s", name);
125 spin_lock(&tcp_cong_list_lock);
126 ca = tcp_ca_find(name);
127 }
128#endif
129
130 if (ca) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700131 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700132 list_move(&ca->list, &tcp_cong_list);
133 ret = 0;
134 }
135 spin_unlock(&tcp_cong_list_lock);
136
137 return ret;
138}
139
Stephen Hemmingerb1736a72006-10-31 17:31:33 -0800140/* Set default value from kernel configuration at bootup */
141static int __init tcp_congestion_default(void)
142{
143 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
144}
145late_initcall(tcp_congestion_default);
146
147
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800148/* Build string with list of available congestion control values */
149void tcp_get_available_congestion_control(char *buf, size_t maxlen)
150{
151 struct tcp_congestion_ops *ca;
152 size_t offs = 0;
153
154 rcu_read_lock();
155 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
156 offs += snprintf(buf + offs, maxlen - offs,
157 "%s%s",
158 offs == 0 ? "" : " ", ca->name);
159
160 }
161 rcu_read_unlock();
162}
163
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700164/* Get current default congestion control */
165void tcp_get_default_congestion_control(char *name)
166{
167 struct tcp_congestion_ops *ca;
168 /* We will always have reno... */
169 BUG_ON(list_empty(&tcp_cong_list));
170
171 rcu_read_lock();
172 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
173 strncpy(name, ca->name, TCP_CA_NAME_MAX);
174 rcu_read_unlock();
175}
176
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800177/* Built list of non-restricted congestion control values */
178void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
179{
180 struct tcp_congestion_ops *ca;
181 size_t offs = 0;
182
183 *buf = '\0';
184 rcu_read_lock();
185 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700186 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800187 continue;
188 offs += snprintf(buf + offs, maxlen - offs,
189 "%s%s",
190 offs == 0 ? "" : " ", ca->name);
191
192 }
193 rcu_read_unlock();
194}
195
196/* Change list of non-restricted congestion control */
197int tcp_set_allowed_congestion_control(char *val)
198{
199 struct tcp_congestion_ops *ca;
Julia Lawallc34186e2010-08-27 19:31:56 -0700200 char *saved_clone, *clone, *name;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800201 int ret = 0;
202
Julia Lawallc34186e2010-08-27 19:31:56 -0700203 saved_clone = clone = kstrdup(val, GFP_USER);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800204 if (!clone)
205 return -ENOMEM;
206
207 spin_lock(&tcp_cong_list_lock);
208 /* pass 1 check for bad entries */
209 while ((name = strsep(&clone, " ")) && *name) {
210 ca = tcp_ca_find(name);
211 if (!ca) {
212 ret = -ENOENT;
213 goto out;
214 }
215 }
216
Stephen Hemminger164891a2007-04-23 22:26:16 -0700217 /* pass 2 clear old values */
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800218 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700219 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800220
221 /* pass 3 mark as allowed */
222 while ((name = strsep(&val, " ")) && *name) {
223 ca = tcp_ca_find(name);
224 WARN_ON(!ca);
225 if (ca)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700226 ca->flags |= TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800227 }
228out:
229 spin_unlock(&tcp_cong_list_lock);
Julia Lawallc34186e2010-08-27 19:31:56 -0700230 kfree(saved_clone);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800231
232 return ret;
233}
234
235
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700236/* Change congestion control for socket */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300237int tcp_set_congestion_control(struct sock *sk, const char *name)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700238{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300239 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700240 struct tcp_congestion_ops *ca;
241 int err = 0;
242
243 rcu_read_lock();
244 ca = tcp_ca_find(name);
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700245
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800246 /* no change asking for existing value */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300247 if (ca == icsk->icsk_ca_ops)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700248 goto out;
249
Johannes Berg95a5afc2008-10-16 15:24:51 -0700250#ifdef CONFIG_MODULES
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800251 /* not found attempt to autoload module */
Eric Parisa8f80e82009-08-13 09:44:51 -0400252 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800253 rcu_read_unlock();
254 request_module("tcp_%s", name);
255 rcu_read_lock();
256 ca = tcp_ca_find(name);
257 }
258#endif
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700259 if (!ca)
260 err = -ENOENT;
261
Stephen Hemminger164891a2007-04-23 22:26:16 -0700262 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800263 err = -EPERM;
264
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700265 else if (!try_module_get(ca->owner))
266 err = -EBUSY;
267
268 else {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300269 tcp_cleanup_congestion_control(sk);
270 icsk->icsk_ca_ops = ca;
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700271
272 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300273 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700274 }
275 out:
276 rcu_read_unlock();
277 return err;
278}
279
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800280/* RFC2861 Check whether we are limited by application or congestion window
281 * This is the inverse of cwnd check in tcp_tso_should_defer
282 */
283int tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
284{
285 const struct tcp_sock *tp = tcp_sk(sk);
286 u32 left;
287
288 if (in_flight >= tp->snd_cwnd)
289 return 1;
290
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800291 left = tp->snd_cwnd - in_flight;
John Heffnerce447eb2008-04-29 03:13:02 -0700292 if (sk_can_gso(sk) &&
John Heffner246eb2a2008-04-29 03:13:52 -0700293 left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
294 left * tp->mss_cache < sk->sk_gso_max_size)
John Heffnerce447eb2008-04-29 03:13:02 -0700295 return 1;
Neal Cardwell6b5a5c02011-11-21 17:15:14 +0000296 return left <= tcp_max_tso_deferred_mss(tp);
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800297}
298EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800299
300/*
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700301 * Slow start is used when congestion window is less than slow start
302 * threshold. This version implements the basic RFC2581 version
303 * and optionally supports:
304 * RFC3742 Limited Slow Start - growth limited to max_ssthresh
305 * RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800306 */
307void tcp_slow_start(struct tcp_sock *tp)
308{
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700309 int cnt; /* increase in packets */
John Heffner886236c2007-03-25 19:21:45 -0700310
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700311 /* RFC3465: ABC Slow start
312 * Increase only after a full MSS of bytes is acked
313 *
314 * TCP sender SHOULD increase cwnd by the number of
315 * previously unacknowledged bytes ACKed by each incoming
316 * acknowledgment, provided the increase is not more than L
317 */
318 if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
319 return;
John Heffner886236c2007-03-25 19:21:45 -0700320
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700321 if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
322 cnt = sysctl_tcp_max_ssthresh >> 1; /* limited slow start */
John Heffner886236c2007-03-25 19:21:45 -0700323 else
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700324 cnt = tp->snd_cwnd; /* exponential increase */
John Heffner886236c2007-03-25 19:21:45 -0700325
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700326 /* RFC3465: ABC
327 * We MAY increase by 2 if discovered delayed ack
328 */
John Heffner886236c2007-03-25 19:21:45 -0700329 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
330 cnt <<= 1;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800331 tp->bytes_acked = 0;
332
John Heffner886236c2007-03-25 19:21:45 -0700333 tp->snd_cwnd_cnt += cnt;
334 while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
335 tp->snd_cwnd_cnt -= tp->snd_cwnd;
336 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
337 tp->snd_cwnd++;
338 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800339}
340EXPORT_SYMBOL_GPL(tcp_slow_start);
341
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000342/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
343void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
344{
345 if (tp->snd_cwnd_cnt >= w) {
346 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
347 tp->snd_cwnd++;
348 tp->snd_cwnd_cnt = 0;
349 } else {
350 tp->snd_cwnd_cnt++;
351 }
352}
353EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
354
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700355/*
356 * TCP Reno congestion control
357 * This is special case used for fallback as well.
358 */
359/* This is Jacobson's slow start and congestion avoidance.
360 * SIGCOMM '88, p. 328.
361 */
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200362void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700363{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300364 struct tcp_sock *tp = tcp_sk(sk);
365
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -0800366 if (!tcp_is_cwnd_limited(sk, in_flight))
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700367 return;
368
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800369 /* In "safe" area, increase. */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900370 if (tp->snd_cwnd <= tp->snd_ssthresh)
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800371 tcp_slow_start(tp);
Stephen Hemminger9772efb2005-11-10 17:09:53 -0800372
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900373 /* In dangerous area, increase slowly. */
Stephen Hemminger9772efb2005-11-10 17:09:53 -0800374 else if (sysctl_tcp_abc) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900375 /* RFC3465: Appropriate Byte Count
376 * increase once for each full cwnd acked
377 */
378 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
379 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
380 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
381 tp->snd_cwnd++;
382 }
383 } else {
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000384 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900385 }
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700386}
387EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
388
389/* Slow start threshold is half the congestion window (min 2) */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300390u32 tcp_reno_ssthresh(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700391{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300392 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700393 return max(tp->snd_cwnd >> 1U, 2U);
394}
395EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
396
Stephen Hemminger72dc5b92006-06-05 17:30:08 -0700397/* Lower bound on congestion window with halving. */
398u32 tcp_reno_min_cwnd(const struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700399{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300400 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700401 return tp->snd_ssthresh/2;
402}
403EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
404
405struct tcp_congestion_ops tcp_reno = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700406 .flags = TCP_CONG_NON_RESTRICTED,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700407 .name = "reno",
408 .owner = THIS_MODULE,
409 .ssthresh = tcp_reno_ssthresh,
410 .cong_avoid = tcp_reno_cong_avoid,
411 .min_cwnd = tcp_reno_min_cwnd,
412};
413
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700414/* Initial congestion control used (until SYN)
415 * really reno under another name so we can tell difference
416 * during tcp_set_default_congestion_control
417 */
418struct tcp_congestion_ops tcp_init_congestion_ops = {
419 .name = "",
420 .owner = THIS_MODULE,
421 .ssthresh = tcp_reno_ssthresh,
422 .cong_avoid = tcp_reno_cong_avoid,
423 .min_cwnd = tcp_reno_min_cwnd,
424};
425EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);