blob: 850c737e08e2a4a9185c64c8776711d83b8c7b08 [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
Stephen Hemminger317a76f2005-06-23 12:19:55 -07009#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/types.h>
12#include <linux/list.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090013#include <linux/gfp.h>
Stephen Hemminger317a76f2005-06-23 12:19:55 -070014#include <net/tcp.h>
15
John Heffner886236c2007-03-25 19:21:45 -070016int sysctl_tcp_max_ssthresh = 0;
17
Stephen Hemminger317a76f2005-06-23 12:19:55 -070018static DEFINE_SPINLOCK(tcp_cong_list_lock);
19static LIST_HEAD(tcp_cong_list);
20
21/* Simple linear search, don't expect many entries! */
22static struct tcp_congestion_ops *tcp_ca_find(const char *name)
23{
24 struct tcp_congestion_ops *e;
25
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070026 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -070027 if (strcmp(e->name, name) == 0)
28 return e;
29 }
30
31 return NULL;
32}
33
34/*
Robert P. J. Dayd08df602007-02-17 19:07:33 +010035 * Attach new congestion control algorithm to the list
Stephen Hemminger317a76f2005-06-23 12:19:55 -070036 * of available options.
37 */
38int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
39{
40 int ret = 0;
41
42 /* all algorithms must implement ssthresh and cong_avoid ops */
Stephen Hemminger72dc5b92006-06-05 17:30:08 -070043 if (!ca->ssthresh || !ca->cong_avoid) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -070044 printk(KERN_ERR "TCP %s does not implement required ops\n",
45 ca->name);
46 return -EINVAL;
47 }
48
49 spin_lock(&tcp_cong_list_lock);
50 if (tcp_ca_find(ca->name)) {
51 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
52 ret = -EEXIST;
53 } else {
Stephen Hemminger3d2573f2006-09-24 20:11:58 -070054 list_add_tail_rcu(&ca->list, &tcp_cong_list);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070055 printk(KERN_INFO "TCP %s registered\n", ca->name);
56 }
57 spin_unlock(&tcp_cong_list_lock);
58
59 return ret;
60}
61EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
62
63/*
64 * Remove congestion control algorithm, called from
65 * the module's remove function. Module ref counts are used
66 * to ensure that this can't be done till all sockets using
67 * that method are closed.
68 */
69void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
70{
71 spin_lock(&tcp_cong_list_lock);
72 list_del_rcu(&ca->list);
73 spin_unlock(&tcp_cong_list_lock);
74}
75EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
76
77/* Assign choice of congestion control. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030078void tcp_init_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -070079{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030080 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070081 struct tcp_congestion_ops *ca;
82
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070083 /* if no choice made yet assign the current value set as default */
84 if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
85 rcu_read_lock();
86 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
87 if (try_module_get(ca->owner)) {
88 icsk->icsk_ca_ops = ca;
89 break;
90 }
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070091
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070092 /* fallback to next available */
Stephen Hemminger317a76f2005-06-23 12:19:55 -070093 }
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070094 rcu_read_unlock();
Stephen Hemminger317a76f2005-06-23 12:19:55 -070095 }
Stephen Hemminger317a76f2005-06-23 12:19:55 -070096
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030097 if (icsk->icsk_ca_ops->init)
98 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070099}
100
101/* Manage refcounts on socket close. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300102void tcp_cleanup_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700103{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300104 struct inet_connection_sock *icsk = inet_csk(sk);
105
106 if (icsk->icsk_ca_ops->release)
107 icsk->icsk_ca_ops->release(sk);
108 module_put(icsk->icsk_ca_ops->owner);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700109}
110
111/* Used by sysctl to change default congestion control */
112int tcp_set_default_congestion_control(const char *name)
113{
114 struct tcp_congestion_ops *ca;
115 int ret = -ENOENT;
116
117 spin_lock(&tcp_cong_list_lock);
118 ca = tcp_ca_find(name);
Johannes Berg95a5afc2008-10-16 15:24:51 -0700119#ifdef CONFIG_MODULES
Eric Parisa8f80e82009-08-13 09:44:51 -0400120 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700121 spin_unlock(&tcp_cong_list_lock);
122
123 request_module("tcp_%s", name);
124 spin_lock(&tcp_cong_list_lock);
125 ca = tcp_ca_find(name);
126 }
127#endif
128
129 if (ca) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700130 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700131 list_move(&ca->list, &tcp_cong_list);
132 ret = 0;
133 }
134 spin_unlock(&tcp_cong_list_lock);
135
136 return ret;
137}
138
Stephen Hemmingerb1736a72006-10-31 17:31:33 -0800139/* Set default value from kernel configuration at bootup */
140static int __init tcp_congestion_default(void)
141{
142 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
143}
144late_initcall(tcp_congestion_default);
145
146
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800147/* Build string with list of available congestion control values */
148void tcp_get_available_congestion_control(char *buf, size_t maxlen)
149{
150 struct tcp_congestion_ops *ca;
151 size_t offs = 0;
152
153 rcu_read_lock();
154 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
155 offs += snprintf(buf + offs, maxlen - offs,
156 "%s%s",
157 offs == 0 ? "" : " ", ca->name);
158
159 }
160 rcu_read_unlock();
161}
162
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700163/* Get current default congestion control */
164void tcp_get_default_congestion_control(char *name)
165{
166 struct tcp_congestion_ops *ca;
167 /* We will always have reno... */
168 BUG_ON(list_empty(&tcp_cong_list));
169
170 rcu_read_lock();
171 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
172 strncpy(name, ca->name, TCP_CA_NAME_MAX);
173 rcu_read_unlock();
174}
175
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800176/* Built list of non-restricted congestion control values */
177void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
178{
179 struct tcp_congestion_ops *ca;
180 size_t offs = 0;
181
182 *buf = '\0';
183 rcu_read_lock();
184 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700185 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800186 continue;
187 offs += snprintf(buf + offs, maxlen - offs,
188 "%s%s",
189 offs == 0 ? "" : " ", ca->name);
190
191 }
192 rcu_read_unlock();
193}
194
195/* Change list of non-restricted congestion control */
196int tcp_set_allowed_congestion_control(char *val)
197{
198 struct tcp_congestion_ops *ca;
Julia Lawallc34186e2010-08-27 19:31:56 -0700199 char *saved_clone, *clone, *name;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800200 int ret = 0;
201
Julia Lawallc34186e2010-08-27 19:31:56 -0700202 saved_clone = clone = kstrdup(val, GFP_USER);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800203 if (!clone)
204 return -ENOMEM;
205
206 spin_lock(&tcp_cong_list_lock);
207 /* pass 1 check for bad entries */
208 while ((name = strsep(&clone, " ")) && *name) {
209 ca = tcp_ca_find(name);
210 if (!ca) {
211 ret = -ENOENT;
212 goto out;
213 }
214 }
215
Stephen Hemminger164891a2007-04-23 22:26:16 -0700216 /* pass 2 clear old values */
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800217 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700218 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800219
220 /* pass 3 mark as allowed */
221 while ((name = strsep(&val, " ")) && *name) {
222 ca = tcp_ca_find(name);
223 WARN_ON(!ca);
224 if (ca)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700225 ca->flags |= TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800226 }
227out:
228 spin_unlock(&tcp_cong_list_lock);
Julia Lawallc34186e2010-08-27 19:31:56 -0700229 kfree(saved_clone);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800230
231 return ret;
232}
233
234
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700235/* Change congestion control for socket */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300236int tcp_set_congestion_control(struct sock *sk, const char *name)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700237{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300238 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700239 struct tcp_congestion_ops *ca;
240 int err = 0;
241
242 rcu_read_lock();
243 ca = tcp_ca_find(name);
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700244
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800245 /* no change asking for existing value */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300246 if (ca == icsk->icsk_ca_ops)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700247 goto out;
248
Johannes Berg95a5afc2008-10-16 15:24:51 -0700249#ifdef CONFIG_MODULES
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800250 /* not found attempt to autoload module */
Eric Parisa8f80e82009-08-13 09:44:51 -0400251 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800252 rcu_read_unlock();
253 request_module("tcp_%s", name);
254 rcu_read_lock();
255 ca = tcp_ca_find(name);
256 }
257#endif
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700258 if (!ca)
259 err = -ENOENT;
260
Stephen Hemminger164891a2007-04-23 22:26:16 -0700261 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800262 err = -EPERM;
263
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700264 else if (!try_module_get(ca->owner))
265 err = -EBUSY;
266
267 else {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300268 tcp_cleanup_congestion_control(sk);
269 icsk->icsk_ca_ops = ca;
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700270
271 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300272 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700273 }
274 out:
275 rcu_read_unlock();
276 return err;
277}
278
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800279/* RFC2861 Check whether we are limited by application or congestion window
280 * This is the inverse of cwnd check in tcp_tso_should_defer
281 */
282int tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
283{
284 const struct tcp_sock *tp = tcp_sk(sk);
285 u32 left;
286
287 if (in_flight >= tp->snd_cwnd)
288 return 1;
289
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800290 left = tp->snd_cwnd - in_flight;
John Heffnerce447eb2008-04-29 03:13:02 -0700291 if (sk_can_gso(sk) &&
John Heffner246eb2a2008-04-29 03:13:52 -0700292 left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
293 left * tp->mss_cache < sk->sk_gso_max_size)
John Heffnerce447eb2008-04-29 03:13:02 -0700294 return 1;
295 return left <= tcp_max_burst(tp);
Ilpo Järvinencea14e02008-01-12 03:19:12 -0800296}
297EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800298
299/*
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700300 * Slow start is used when congestion window is less than slow start
301 * threshold. This version implements the basic RFC2581 version
302 * and optionally supports:
303 * RFC3742 Limited Slow Start - growth limited to max_ssthresh
304 * RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800305 */
306void tcp_slow_start(struct tcp_sock *tp)
307{
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700308 int cnt; /* increase in packets */
John Heffner886236c2007-03-25 19:21:45 -0700309
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700310 /* RFC3465: ABC Slow start
311 * Increase only after a full MSS of bytes is acked
312 *
313 * TCP sender SHOULD increase cwnd by the number of
314 * previously unacknowledged bytes ACKed by each incoming
315 * acknowledgment, provided the increase is not more than L
316 */
317 if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
318 return;
John Heffner886236c2007-03-25 19:21:45 -0700319
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700320 if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
321 cnt = sysctl_tcp_max_ssthresh >> 1; /* limited slow start */
John Heffner886236c2007-03-25 19:21:45 -0700322 else
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700323 cnt = tp->snd_cwnd; /* exponential increase */
John Heffner886236c2007-03-25 19:21:45 -0700324
Stephen Hemmingera02ba042007-05-17 00:04:18 -0700325 /* RFC3465: ABC
326 * We MAY increase by 2 if discovered delayed ack
327 */
John Heffner886236c2007-03-25 19:21:45 -0700328 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
329 cnt <<= 1;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800330 tp->bytes_acked = 0;
331
John Heffner886236c2007-03-25 19:21:45 -0700332 tp->snd_cwnd_cnt += cnt;
333 while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
334 tp->snd_cwnd_cnt -= tp->snd_cwnd;
335 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
336 tp->snd_cwnd++;
337 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800338}
339EXPORT_SYMBOL_GPL(tcp_slow_start);
340
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000341/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
342void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
343{
344 if (tp->snd_cwnd_cnt >= w) {
345 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
346 tp->snd_cwnd++;
347 tp->snd_cwnd_cnt = 0;
348 } else {
349 tp->snd_cwnd_cnt++;
350 }
351}
352EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
353
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700354/*
355 * TCP Reno congestion control
356 * This is special case used for fallback as well.
357 */
358/* This is Jacobson's slow start and congestion avoidance.
359 * SIGCOMM '88, p. 328.
360 */
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200361void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700362{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300363 struct tcp_sock *tp = tcp_sk(sk);
364
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -0800365 if (!tcp_is_cwnd_limited(sk, in_flight))
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700366 return;
367
Stephen Hemminger7faffa1c2005-11-10 17:07:24 -0800368 /* In "safe" area, increase. */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900369 if (tp->snd_cwnd <= tp->snd_ssthresh)
Stephen Hemminger7faffa1c2005-11-10 17:07:24 -0800370 tcp_slow_start(tp);
Stephen Hemminger9772efb2005-11-10 17:09:53 -0800371
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900372 /* In dangerous area, increase slowly. */
Stephen Hemminger9772efb2005-11-10 17:09:53 -0800373 else if (sysctl_tcp_abc) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900374 /* RFC3465: Appropriate Byte Count
375 * increase once for each full cwnd acked
376 */
377 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
378 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
379 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
380 tp->snd_cwnd++;
381 }
382 } else {
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000383 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900384 }
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700385}
386EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
387
388/* Slow start threshold is half the congestion window (min 2) */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300389u32 tcp_reno_ssthresh(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700390{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300391 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700392 return max(tp->snd_cwnd >> 1U, 2U);
393}
394EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
395
Stephen Hemminger72dc5b92006-06-05 17:30:08 -0700396/* Lower bound on congestion window with halving. */
397u32 tcp_reno_min_cwnd(const struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700398{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300399 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700400 return tp->snd_ssthresh/2;
401}
402EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
403
404struct tcp_congestion_ops tcp_reno = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700405 .flags = TCP_CONG_NON_RESTRICTED,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700406 .name = "reno",
407 .owner = THIS_MODULE,
408 .ssthresh = tcp_reno_ssthresh,
409 .cong_avoid = tcp_reno_cong_avoid,
410 .min_cwnd = tcp_reno_min_cwnd,
411};
412
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700413/* Initial congestion control used (until SYN)
414 * really reno under another name so we can tell difference
415 * during tcp_set_default_congestion_control
416 */
417struct tcp_congestion_ops tcp_init_congestion_ops = {
418 .name = "",
419 .owner = THIS_MODULE,
420 .ssthresh = tcp_reno_ssthresh,
421 .cong_avoid = tcp_reno_cong_avoid,
422 .min_cwnd = tcp_reno_min_cwnd,
423};
424EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);