blob: 7b09d8b49fa51271cc0fa99a3fcda9f017c0f469 [file] [log] [blame]
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001/*
2 * Plugable TCP congestion control support and newReno
3 * 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>
Stephen Hemminger317a76f2005-06-23 12:19:55 -070016#include <net/tcp.h>
17
18static 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) {
Joe Perchesafd465032012-03-12 07:03:32 +000044 pr_err("%s does not implement required ops\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070045 return -EINVAL;
46 }
47
48 spin_lock(&tcp_cong_list_lock);
49 if (tcp_ca_find(ca->name)) {
Joe Perchesafd465032012-03-12 07:03:32 +000050 pr_notice("%s already registered\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070051 ret = -EEXIST;
52 } else {
Stephen Hemminger3d2573f2006-09-24 20:11:58 -070053 list_add_tail_rcu(&ca->list, &tcp_cong_list);
Joe Perchesafd465032012-03-12 07:03:32 +000054 pr_info("%s registered\n", ca->name);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070055 }
56 spin_unlock(&tcp_cong_list_lock);
57
58 return ret;
59}
60EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
61
62/*
63 * Remove congestion control algorithm, called from
64 * the module's remove function. Module ref counts are used
65 * to ensure that this can't be done till all sockets using
66 * that method are closed.
67 */
68void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
69{
70 spin_lock(&tcp_cong_list_lock);
71 list_del_rcu(&ca->list);
72 spin_unlock(&tcp_cong_list_lock);
73}
74EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
75
76/* Assign choice of congestion control. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030077void tcp_init_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -070078{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030079 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070080 struct tcp_congestion_ops *ca;
81
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070082 /* if no choice made yet assign the current value set as default */
83 if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
84 rcu_read_lock();
85 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
86 if (try_module_get(ca->owner)) {
87 icsk->icsk_ca_ops = ca;
88 break;
89 }
Stephen Hemminger5f8ef482005-06-23 20:37:36 -070090
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070091 /* fallback to next available */
Stephen Hemminger317a76f2005-06-23 12:19:55 -070092 }
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -070093 rcu_read_unlock();
Stephen Hemminger317a76f2005-06-23 12:19:55 -070094 }
Stephen Hemminger317a76f2005-06-23 12:19:55 -070095
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030096 if (icsk->icsk_ca_ops->init)
97 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -070098}
99
100/* Manage refcounts on socket close. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300101void tcp_cleanup_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700102{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300103 struct inet_connection_sock *icsk = inet_csk(sk);
104
105 if (icsk->icsk_ca_ops->release)
106 icsk->icsk_ca_ops->release(sk);
107 module_put(icsk->icsk_ca_ops->owner);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700108}
109
110/* Used by sysctl to change default congestion control */
111int tcp_set_default_congestion_control(const char *name)
112{
113 struct tcp_congestion_ops *ca;
114 int ret = -ENOENT;
115
116 spin_lock(&tcp_cong_list_lock);
117 ca = tcp_ca_find(name);
Johannes Berg95a5afc2008-10-16 15:24:51 -0700118#ifdef CONFIG_MODULES
Eric Parisa8f80e82009-08-13 09:44:51 -0400119 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700120 spin_unlock(&tcp_cong_list_lock);
121
122 request_module("tcp_%s", name);
123 spin_lock(&tcp_cong_list_lock);
124 ca = tcp_ca_find(name);
125 }
126#endif
127
128 if (ca) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700129 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700130 list_move(&ca->list, &tcp_cong_list);
131 ret = 0;
132 }
133 spin_unlock(&tcp_cong_list_lock);
134
135 return ret;
136}
137
Stephen Hemmingerb1736a72006-10-31 17:31:33 -0800138/* Set default value from kernel configuration at bootup */
139static int __init tcp_congestion_default(void)
140{
141 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
142}
143late_initcall(tcp_congestion_default);
144
145
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800146/* Build string with list of available congestion control values */
147void tcp_get_available_congestion_control(char *buf, size_t maxlen)
148{
149 struct tcp_congestion_ops *ca;
150 size_t offs = 0;
151
152 rcu_read_lock();
153 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
154 offs += snprintf(buf + offs, maxlen - offs,
155 "%s%s",
156 offs == 0 ? "" : " ", ca->name);
157
158 }
159 rcu_read_unlock();
160}
161
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700162/* Get current default congestion control */
163void tcp_get_default_congestion_control(char *name)
164{
165 struct tcp_congestion_ops *ca;
166 /* We will always have reno... */
167 BUG_ON(list_empty(&tcp_cong_list));
168
169 rcu_read_lock();
170 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
171 strncpy(name, ca->name, TCP_CA_NAME_MAX);
172 rcu_read_unlock();
173}
174
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800175/* Built list of non-restricted congestion control values */
176void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
177{
178 struct tcp_congestion_ops *ca;
179 size_t offs = 0;
180
181 *buf = '\0';
182 rcu_read_lock();
183 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700184 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800185 continue;
186 offs += snprintf(buf + offs, maxlen - offs,
187 "%s%s",
188 offs == 0 ? "" : " ", ca->name);
189
190 }
191 rcu_read_unlock();
192}
193
194/* Change list of non-restricted congestion control */
195int tcp_set_allowed_congestion_control(char *val)
196{
197 struct tcp_congestion_ops *ca;
Julia Lawallc34186e2010-08-27 19:31:56 -0700198 char *saved_clone, *clone, *name;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800199 int ret = 0;
200
Julia Lawallc34186e2010-08-27 19:31:56 -0700201 saved_clone = clone = kstrdup(val, GFP_USER);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800202 if (!clone)
203 return -ENOMEM;
204
205 spin_lock(&tcp_cong_list_lock);
206 /* pass 1 check for bad entries */
207 while ((name = strsep(&clone, " ")) && *name) {
208 ca = tcp_ca_find(name);
209 if (!ca) {
210 ret = -ENOENT;
211 goto out;
212 }
213 }
214
Stephen Hemminger164891a2007-04-23 22:26:16 -0700215 /* pass 2 clear old values */
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800216 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700217 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800218
219 /* pass 3 mark as allowed */
220 while ((name = strsep(&val, " ")) && *name) {
221 ca = tcp_ca_find(name);
222 WARN_ON(!ca);
223 if (ca)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700224 ca->flags |= TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800225 }
226out:
227 spin_unlock(&tcp_cong_list_lock);
Julia Lawallc34186e2010-08-27 19:31:56 -0700228 kfree(saved_clone);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800229
230 return ret;
231}
232
233
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700234/* Change congestion control for socket */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300235int tcp_set_congestion_control(struct sock *sk, const char *name)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700236{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300237 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700238 struct tcp_congestion_ops *ca;
239 int err = 0;
240
241 rcu_read_lock();
242 ca = tcp_ca_find(name);
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700243
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800244 /* no change asking for existing value */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300245 if (ca == icsk->icsk_ca_ops)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700246 goto out;
247
Johannes Berg95a5afc2008-10-16 15:24:51 -0700248#ifdef CONFIG_MODULES
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800249 /* not found attempt to autoload module */
Eric Parisa8f80e82009-08-13 09:44:51 -0400250 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800251 rcu_read_unlock();
252 request_module("tcp_%s", name);
253 rcu_read_lock();
254 ca = tcp_ca_find(name);
255 }
256#endif
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700257 if (!ca)
258 err = -ENOENT;
259
Eric W. Biederman52e804c2012-11-16 03:03:05 +0000260 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
261 ns_capable(sock_net(sk)->user_ns, 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
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700279/* Slow start is used when congestion window is no greater than the slow start
280 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
281 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
282 * something better;) a packet is only considered (s)acked in its entirety to
283 * defend the ACK attacks described in the RFC. Slow start processes a stretch
284 * ACK of degree N as if N acks of degree 1 are received back to back except
285 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
286 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800287 */
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700288int tcp_slow_start(struct tcp_sock *tp, u32 acked)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800289{
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700290 u32 cwnd = tp->snd_cwnd + acked;
Eric Dumazet973ec442013-02-02 05:23:16 +0000291
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700292 if (cwnd > tp->snd_ssthresh)
293 cwnd = tp->snd_ssthresh + 1;
294 acked -= cwnd - tp->snd_cwnd;
295 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
296 return acked;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800297}
298EXPORT_SYMBOL_GPL(tcp_slow_start);
299
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000300/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
301void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
302{
303 if (tp->snd_cwnd_cnt >= w) {
304 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
305 tp->snd_cwnd++;
306 tp->snd_cwnd_cnt = 0;
307 } else {
308 tp->snd_cwnd_cnt++;
309 }
310}
311EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
312
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700313/*
314 * TCP Reno congestion control
315 * This is special case used for fallback as well.
316 */
317/* This is Jacobson's slow start and congestion avoidance.
318 * SIGCOMM '88, p. 328.
319 */
Eric Dumazet24901552014-05-02 21:18:05 -0700320void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700321{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300322 struct tcp_sock *tp = tcp_sk(sk);
323
Eric Dumazet24901552014-05-02 21:18:05 -0700324 if (!tcp_is_cwnd_limited(sk))
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700325 return;
326
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800327 /* In "safe" area, increase. */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900328 if (tp->snd_cwnd <= tp->snd_ssthresh)
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700329 tcp_slow_start(tp, acked);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900330 /* In dangerous area, increase slowly. */
Stephen Hemmingerca2eb562013-02-05 07:25:17 +0000331 else
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000332 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700333}
334EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
335
336/* Slow start threshold is half the congestion window (min 2) */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300337u32 tcp_reno_ssthresh(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700338{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300339 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700340 return max(tp->snd_cwnd >> 1U, 2U);
341}
342EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
343
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700344struct tcp_congestion_ops tcp_reno = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700345 .flags = TCP_CONG_NON_RESTRICTED,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700346 .name = "reno",
347 .owner = THIS_MODULE,
348 .ssthresh = tcp_reno_ssthresh,
349 .cong_avoid = tcp_reno_cong_avoid,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700350};
351
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700352/* Initial congestion control used (until SYN)
353 * really reno under another name so we can tell difference
354 * during tcp_set_default_congestion_control
355 */
356struct tcp_congestion_ops tcp_init_congestion_ops = {
357 .name = "",
358 .owner = THIS_MODULE,
359 .ssthresh = tcp_reno_ssthresh,
360 .cong_avoid = tcp_reno_cong_avoid,
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700361};
362EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);