blob: 63c29dba68a8f220d1b788d86ca11853590121ec [file] [log] [blame]
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001/*
Fabian Frederickb92022f2014-11-04 20:25:38 +01002 * Pluggable TCP congestion control support and newReno
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003 * congestion control.
Masanari Iida02582e92012-08-22 19:11:26 +09004 * Based on ideas from I/O scheduler support and Web100.
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005 *
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>
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +010016#include <linux/jhash.h>
Stephen Hemminger317a76f2005-06-23 12:19:55 -070017#include <net/tcp.h>
18
19static DEFINE_SPINLOCK(tcp_cong_list_lock);
20static LIST_HEAD(tcp_cong_list);
21
22/* Simple linear search, don't expect many entries! */
23static struct tcp_congestion_ops *tcp_ca_find(const char *name)
24{
25 struct tcp_congestion_ops *e;
26
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070027 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -070028 if (strcmp(e->name, name) == 0)
29 return e;
30 }
31
32 return NULL;
33}
34
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +010035/* Must be called with rcu lock held */
36static const struct tcp_congestion_ops *__tcp_ca_find_autoload(const char *name)
37{
38 const struct tcp_congestion_ops *ca = tcp_ca_find(name);
39#ifdef CONFIG_MODULES
40 if (!ca && capable(CAP_NET_ADMIN)) {
41 rcu_read_unlock();
42 request_module("tcp_%s", name);
43 rcu_read_lock();
44 ca = tcp_ca_find(name);
45 }
46#endif
47 return ca;
48}
49
50/* Simple linear search, not much in here. */
51struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
52{
53 struct tcp_congestion_ops *e;
54
55 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
56 if (e->key == key)
57 return e;
58 }
59
60 return NULL;
61}
62
Stephen Hemminger317a76f2005-06-23 12:19:55 -070063/*
Robert P. J. Dayd08df602007-02-17 19:07:33 +010064 * Attach new congestion control algorithm to the list
Stephen Hemminger317a76f2005-06-23 12:19:55 -070065 * of available options.
66 */
67int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
68{
69 int ret = 0;
70
71 /* all algorithms must implement ssthresh and cong_avoid ops */
Stephen Hemminger72dc5b92006-06-05 17:30:08 -070072 if (!ca->ssthresh || !ca->cong_avoid) {
Joe Perchesafd465032012-03-12 07:03:32 +000073 pr_err("%s does not implement required ops\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070074 return -EINVAL;
75 }
76
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +010077 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
78
Stephen Hemminger317a76f2005-06-23 12:19:55 -070079 spin_lock(&tcp_cong_list_lock);
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +010080 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
81 pr_notice("%s already registered or non-unique key\n",
82 ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070083 ret = -EEXIST;
84 } else {
Stephen Hemminger3d2573f2006-09-24 20:11:58 -070085 list_add_tail_rcu(&ca->list, &tcp_cong_list);
Joe Perchesafd465032012-03-12 07:03:32 +000086 pr_info("%s registered\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070087 }
88 spin_unlock(&tcp_cong_list_lock);
89
90 return ret;
91}
92EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
93
94/*
95 * Remove congestion control algorithm, called from
96 * the module's remove function. Module ref counts are used
97 * to ensure that this can't be done till all sockets using
98 * that method are closed.
99 */
100void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
101{
102 spin_lock(&tcp_cong_list_lock);
103 list_del_rcu(&ca->list);
104 spin_unlock(&tcp_cong_list_lock);
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +0100105
106 /* Wait for outstanding readers to complete before the
107 * module gets removed entirely.
108 *
109 * A try_module_get() should fail by now as our module is
110 * in "going" state since no refs are held anymore and
111 * module_exit() handler being called.
112 */
113 synchronize_rcu();
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700114}
115EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
116
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +0100117u32 tcp_ca_get_key_by_name(const char *name)
118{
119 const struct tcp_congestion_ops *ca;
120 u32 key;
121
122 might_sleep();
123
124 rcu_read_lock();
125 ca = __tcp_ca_find_autoload(name);
126 key = ca ? ca->key : TCP_CA_UNSPEC;
127 rcu_read_unlock();
128
129 return key;
130}
131EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
132
133char *tcp_ca_get_name_by_key(u32 key, char *buffer)
134{
135 const struct tcp_congestion_ops *ca;
136 char *ret = NULL;
137
138 rcu_read_lock();
139 ca = tcp_ca_find_key(key);
140 if (ca)
141 ret = strncpy(buffer, ca->name,
142 TCP_CA_NAME_MAX);
143 rcu_read_unlock();
144
145 return ret;
146}
147EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
148
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700149/* Assign choice of congestion control. */
Florian Westphal55d86942014-09-26 22:37:32 +0200150void tcp_assign_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700151{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300152 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700153 struct tcp_congestion_ops *ca;
154
Florian Westphal55d86942014-09-26 22:37:32 +0200155 rcu_read_lock();
156 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
157 if (likely(try_module_get(ca->owner))) {
158 icsk->icsk_ca_ops = ca;
159 goto out;
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700160 }
Florian Westphal55d86942014-09-26 22:37:32 +0200161 /* Fallback to next available. The last really
162 * guaranteed fallback is Reno from this list.
163 */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700164 }
Florian Westphal55d86942014-09-26 22:37:32 +0200165out:
166 rcu_read_unlock();
167
168 /* Clear out private data before diag gets it and
169 * the ca has not been initialized.
170 */
171 if (ca->get_info)
172 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
173}
174
175void tcp_init_congestion_control(struct sock *sk)
176{
177 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700178
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300179 if (icsk->icsk_ca_ops->init)
180 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700181}
182
Daniel Borkmann29ba4fff2015-01-05 23:57:45 +0100183static void tcp_reinit_congestion_control(struct sock *sk,
184 const struct tcp_congestion_ops *ca)
185{
186 struct inet_connection_sock *icsk = inet_csk(sk);
187
188 tcp_cleanup_congestion_control(sk);
189 icsk->icsk_ca_ops = ca;
190
191 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
192 icsk->icsk_ca_ops->init(sk);
193}
194
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700195/* Manage refcounts on socket close. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300196void tcp_cleanup_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700197{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300198 struct inet_connection_sock *icsk = inet_csk(sk);
199
200 if (icsk->icsk_ca_ops->release)
201 icsk->icsk_ca_ops->release(sk);
202 module_put(icsk->icsk_ca_ops->owner);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700203}
204
205/* Used by sysctl to change default congestion control */
206int tcp_set_default_congestion_control(const char *name)
207{
208 struct tcp_congestion_ops *ca;
209 int ret = -ENOENT;
210
211 spin_lock(&tcp_cong_list_lock);
212 ca = tcp_ca_find(name);
Johannes Berg95a5afc2008-10-16 15:24:51 -0700213#ifdef CONFIG_MODULES
Eric Parisa8f80e82009-08-13 09:44:51 -0400214 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700215 spin_unlock(&tcp_cong_list_lock);
216
217 request_module("tcp_%s", name);
218 spin_lock(&tcp_cong_list_lock);
219 ca = tcp_ca_find(name);
220 }
221#endif
222
223 if (ca) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700224 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700225 list_move(&ca->list, &tcp_cong_list);
226 ret = 0;
227 }
228 spin_unlock(&tcp_cong_list_lock);
229
230 return ret;
231}
232
Stephen Hemmingerb1736a72006-10-31 17:31:33 -0800233/* Set default value from kernel configuration at bootup */
234static int __init tcp_congestion_default(void)
235{
236 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
237}
238late_initcall(tcp_congestion_default);
239
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800240/* Build string with list of available congestion control values */
241void tcp_get_available_congestion_control(char *buf, size_t maxlen)
242{
243 struct tcp_congestion_ops *ca;
244 size_t offs = 0;
245
246 rcu_read_lock();
247 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
248 offs += snprintf(buf + offs, maxlen - offs,
249 "%s%s",
250 offs == 0 ? "" : " ", ca->name);
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800251 }
252 rcu_read_unlock();
253}
254
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700255/* Get current default congestion control */
256void tcp_get_default_congestion_control(char *name)
257{
258 struct tcp_congestion_ops *ca;
259 /* We will always have reno... */
260 BUG_ON(list_empty(&tcp_cong_list));
261
262 rcu_read_lock();
263 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
264 strncpy(name, ca->name, TCP_CA_NAME_MAX);
265 rcu_read_unlock();
266}
267
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800268/* Built list of non-restricted congestion control values */
269void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
270{
271 struct tcp_congestion_ops *ca;
272 size_t offs = 0;
273
274 *buf = '\0';
275 rcu_read_lock();
276 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700277 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800278 continue;
279 offs += snprintf(buf + offs, maxlen - offs,
280 "%s%s",
281 offs == 0 ? "" : " ", ca->name);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800282 }
283 rcu_read_unlock();
284}
285
286/* Change list of non-restricted congestion control */
287int tcp_set_allowed_congestion_control(char *val)
288{
289 struct tcp_congestion_ops *ca;
Julia Lawallc34186e2010-08-27 19:31:56 -0700290 char *saved_clone, *clone, *name;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800291 int ret = 0;
292
Julia Lawallc34186e2010-08-27 19:31:56 -0700293 saved_clone = clone = kstrdup(val, GFP_USER);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800294 if (!clone)
295 return -ENOMEM;
296
297 spin_lock(&tcp_cong_list_lock);
298 /* pass 1 check for bad entries */
299 while ((name = strsep(&clone, " ")) && *name) {
300 ca = tcp_ca_find(name);
301 if (!ca) {
302 ret = -ENOENT;
303 goto out;
304 }
305 }
306
Stephen Hemminger164891a2007-04-23 22:26:16 -0700307 /* pass 2 clear old values */
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800308 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700309 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800310
311 /* pass 3 mark as allowed */
312 while ((name = strsep(&val, " ")) && *name) {
313 ca = tcp_ca_find(name);
314 WARN_ON(!ca);
315 if (ca)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700316 ca->flags |= TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800317 }
318out:
319 spin_unlock(&tcp_cong_list_lock);
Julia Lawallc34186e2010-08-27 19:31:56 -0700320 kfree(saved_clone);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800321
322 return ret;
323}
324
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700325/* Change congestion control for socket */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300326int tcp_set_congestion_control(struct sock *sk, const char *name)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700327{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300328 struct inet_connection_sock *icsk = inet_csk(sk);
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +0100329 const struct tcp_congestion_ops *ca;
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700330 int err = 0;
331
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +0100332 if (icsk->icsk_ca_dst_locked)
333 return -EPERM;
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700334
Daniel Borkmannc5c6a8a2015-01-05 23:57:46 +0100335 rcu_read_lock();
336 ca = __tcp_ca_find_autoload(name);
337 /* No change asking for existing value */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300338 if (ca == icsk->icsk_ca_ops)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700339 goto out;
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700340 if (!ca)
341 err = -ENOENT;
Eric W. Biederman52e804c2012-11-16 03:03:05 +0000342 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
343 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800344 err = -EPERM;
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700345 else if (!try_module_get(ca->owner))
346 err = -EBUSY;
Daniel Borkmann29ba4fff2015-01-05 23:57:45 +0100347 else
348 tcp_reinit_congestion_control(sk, ca);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700349 out:
350 rcu_read_unlock();
351 return err;
352}
353
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700354/* Slow start is used when congestion window is no greater than the slow start
355 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
356 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
357 * something better;) a packet is only considered (s)acked in its entirety to
358 * defend the ACK attacks described in the RFC. Slow start processes a stretch
359 * ACK of degree N as if N acks of degree 1 are received back to back except
360 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
361 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800362 */
Li RongQinga12a6012014-09-30 09:49:55 +0800363void tcp_slow_start(struct tcp_sock *tp, u32 acked)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800364{
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700365 u32 cwnd = tp->snd_cwnd + acked;
Eric Dumazet973ec442013-02-02 05:23:16 +0000366
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700367 if (cwnd > tp->snd_ssthresh)
368 cwnd = tp->snd_ssthresh + 1;
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700369 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800370}
371EXPORT_SYMBOL_GPL(tcp_slow_start);
372
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000373/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
374void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
375{
376 if (tp->snd_cwnd_cnt >= w) {
377 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
378 tp->snd_cwnd++;
379 tp->snd_cwnd_cnt = 0;
380 } else {
381 tp->snd_cwnd_cnt++;
382 }
383}
384EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
385
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700386/*
387 * TCP Reno congestion control
388 * This is special case used for fallback as well.
389 */
390/* This is Jacobson's slow start and congestion avoidance.
391 * SIGCOMM '88, p. 328.
392 */
Eric Dumazet24901552014-05-02 21:18:05 -0700393void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700394{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300395 struct tcp_sock *tp = tcp_sk(sk);
396
Eric Dumazet24901552014-05-02 21:18:05 -0700397 if (!tcp_is_cwnd_limited(sk))
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700398 return;
399
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800400 /* In "safe" area, increase. */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900401 if (tp->snd_cwnd <= tp->snd_ssthresh)
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700402 tcp_slow_start(tp, acked);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900403 /* In dangerous area, increase slowly. */
Stephen Hemmingerca2eb562013-02-05 07:25:17 +0000404 else
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000405 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700406}
407EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
408
409/* Slow start threshold is half the congestion window (min 2) */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300410u32 tcp_reno_ssthresh(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700411{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300412 const struct tcp_sock *tp = tcp_sk(sk);
stephen hemminger688d1942014-08-29 23:32:05 -0700413
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700414 return max(tp->snd_cwnd >> 1U, 2U);
415}
416EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
417
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700418struct tcp_congestion_ops tcp_reno = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700419 .flags = TCP_CONG_NON_RESTRICTED,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700420 .name = "reno",
421 .owner = THIS_MODULE,
422 .ssthresh = tcp_reno_ssthresh,
423 .cong_avoid = tcp_reno_cong_avoid,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700424};