blob: 27ead0dd16bc7e444e96781ff01b10c444678396 [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>
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. */
Florian Westphal55d86942014-09-26 22:37:32 +020077void tcp_assign_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
Florian Westphal55d86942014-09-26 22:37:32 +020082 rcu_read_lock();
83 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
84 if (likely(try_module_get(ca->owner))) {
85 icsk->icsk_ca_ops = ca;
86 goto out;
Stephen Hemminger317a76f2005-06-23 12:19:55 -070087 }
Florian Westphal55d86942014-09-26 22:37:32 +020088 /* Fallback to next available. The last really
89 * guaranteed fallback is Reno from this list.
90 */
Stephen Hemminger317a76f2005-06-23 12:19:55 -070091 }
Florian Westphal55d86942014-09-26 22:37:32 +020092out:
93 rcu_read_unlock();
94
95 /* Clear out private data before diag gets it and
96 * the ca has not been initialized.
97 */
98 if (ca->get_info)
99 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
100}
101
102void tcp_init_congestion_control(struct sock *sk)
103{
104 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700105
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300106 if (icsk->icsk_ca_ops->init)
107 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700108}
109
110/* Manage refcounts on socket close. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300111void tcp_cleanup_congestion_control(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700112{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300113 struct inet_connection_sock *icsk = inet_csk(sk);
114
115 if (icsk->icsk_ca_ops->release)
116 icsk->icsk_ca_ops->release(sk);
117 module_put(icsk->icsk_ca_ops->owner);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700118}
119
120/* Used by sysctl to change default congestion control */
121int tcp_set_default_congestion_control(const char *name)
122{
123 struct tcp_congestion_ops *ca;
124 int ret = -ENOENT;
125
126 spin_lock(&tcp_cong_list_lock);
127 ca = tcp_ca_find(name);
Johannes Berg95a5afc2008-10-16 15:24:51 -0700128#ifdef CONFIG_MODULES
Eric Parisa8f80e82009-08-13 09:44:51 -0400129 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700130 spin_unlock(&tcp_cong_list_lock);
131
132 request_module("tcp_%s", name);
133 spin_lock(&tcp_cong_list_lock);
134 ca = tcp_ca_find(name);
135 }
136#endif
137
138 if (ca) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700139 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700140 list_move(&ca->list, &tcp_cong_list);
141 ret = 0;
142 }
143 spin_unlock(&tcp_cong_list_lock);
144
145 return ret;
146}
147
Stephen Hemmingerb1736a72006-10-31 17:31:33 -0800148/* Set default value from kernel configuration at bootup */
149static int __init tcp_congestion_default(void)
150{
151 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
152}
153late_initcall(tcp_congestion_default);
154
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800155/* Build string with list of available congestion control values */
156void tcp_get_available_congestion_control(char *buf, size_t maxlen)
157{
158 struct tcp_congestion_ops *ca;
159 size_t offs = 0;
160
161 rcu_read_lock();
162 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
163 offs += snprintf(buf + offs, maxlen - offs,
164 "%s%s",
165 offs == 0 ? "" : " ", ca->name);
Stephen Hemminger3ff825b2006-11-09 16:32:06 -0800166 }
167 rcu_read_unlock();
168}
169
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700170/* Get current default congestion control */
171void tcp_get_default_congestion_control(char *name)
172{
173 struct tcp_congestion_ops *ca;
174 /* We will always have reno... */
175 BUG_ON(list_empty(&tcp_cong_list));
176
177 rcu_read_lock();
178 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
179 strncpy(name, ca->name, TCP_CA_NAME_MAX);
180 rcu_read_unlock();
181}
182
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800183/* Built list of non-restricted congestion control values */
184void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
185{
186 struct tcp_congestion_ops *ca;
187 size_t offs = 0;
188
189 *buf = '\0';
190 rcu_read_lock();
191 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700192 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800193 continue;
194 offs += snprintf(buf + offs, maxlen - offs,
195 "%s%s",
196 offs == 0 ? "" : " ", ca->name);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800197 }
198 rcu_read_unlock();
199}
200
201/* Change list of non-restricted congestion control */
202int tcp_set_allowed_congestion_control(char *val)
203{
204 struct tcp_congestion_ops *ca;
Julia Lawallc34186e2010-08-27 19:31:56 -0700205 char *saved_clone, *clone, *name;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800206 int ret = 0;
207
Julia Lawallc34186e2010-08-27 19:31:56 -0700208 saved_clone = clone = kstrdup(val, GFP_USER);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800209 if (!clone)
210 return -ENOMEM;
211
212 spin_lock(&tcp_cong_list_lock);
213 /* pass 1 check for bad entries */
214 while ((name = strsep(&clone, " ")) && *name) {
215 ca = tcp_ca_find(name);
216 if (!ca) {
217 ret = -ENOENT;
218 goto out;
219 }
220 }
221
Stephen Hemminger164891a2007-04-23 22:26:16 -0700222 /* pass 2 clear old values */
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800223 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700224 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800225
226 /* pass 3 mark as allowed */
227 while ((name = strsep(&val, " ")) && *name) {
228 ca = tcp_ca_find(name);
229 WARN_ON(!ca);
230 if (ca)
Stephen Hemminger164891a2007-04-23 22:26:16 -0700231 ca->flags |= TCP_CONG_NON_RESTRICTED;
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800232 }
233out:
234 spin_unlock(&tcp_cong_list_lock);
Julia Lawallc34186e2010-08-27 19:31:56 -0700235 kfree(saved_clone);
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800236
237 return ret;
238}
239
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700240/* Change congestion control for socket */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300241int tcp_set_congestion_control(struct sock *sk, const char *name)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700242{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300243 struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700244 struct tcp_congestion_ops *ca;
245 int err = 0;
246
247 rcu_read_lock();
248 ca = tcp_ca_find(name);
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700249
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800250 /* no change asking for existing value */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300251 if (ca == icsk->icsk_ca_ops)
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700252 goto out;
253
Johannes Berg95a5afc2008-10-16 15:24:51 -0700254#ifdef CONFIG_MODULES
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800255 /* not found attempt to autoload module */
Eric Parisa8f80e82009-08-13 09:44:51 -0400256 if (!ca && capable(CAP_NET_ADMIN)) {
Stephen Hemminger35bfbc92006-11-09 16:36:36 -0800257 rcu_read_unlock();
258 request_module("tcp_%s", name);
259 rcu_read_lock();
260 ca = tcp_ca_find(name);
261 }
262#endif
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700263 if (!ca)
264 err = -ENOENT;
265
Eric W. Biederman52e804c2012-11-16 03:03:05 +0000266 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
267 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
Stephen Hemmingerce7bc3b2006-11-09 16:35:15 -0800268 err = -EPERM;
269
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700270 else if (!try_module_get(ca->owner))
271 err = -EBUSY;
272
273 else {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300274 tcp_cleanup_congestion_control(sk);
275 icsk->icsk_ca_ops = ca;
Stephen Hemminger4d4d3d12007-04-23 22:32:11 -0700276
277 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300278 icsk->icsk_ca_ops->init(sk);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -0700279 }
280 out:
281 rcu_read_unlock();
282 return err;
283}
284
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700285/* Slow start is used when congestion window is no greater than the slow start
286 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
287 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
288 * something better;) a packet is only considered (s)acked in its entirety to
289 * defend the ACK attacks described in the RFC. Slow start processes a stretch
290 * ACK of degree N as if N acks of degree 1 are received back to back except
291 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
292 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800293 */
Li RongQinga12a6012014-09-30 09:49:55 +0800294void tcp_slow_start(struct tcp_sock *tp, u32 acked)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800295{
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700296 u32 cwnd = tp->snd_cwnd + acked;
Eric Dumazet973ec442013-02-02 05:23:16 +0000297
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700298 if (cwnd > tp->snd_ssthresh)
299 cwnd = tp->snd_ssthresh + 1;
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700300 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800301}
302EXPORT_SYMBOL_GPL(tcp_slow_start);
303
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000304/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
305void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
306{
307 if (tp->snd_cwnd_cnt >= w) {
308 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
309 tp->snd_cwnd++;
310 tp->snd_cwnd_cnt = 0;
311 } else {
312 tp->snd_cwnd_cnt++;
313 }
314}
315EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
316
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700317/*
318 * TCP Reno congestion control
319 * This is special case used for fallback as well.
320 */
321/* This is Jacobson's slow start and congestion avoidance.
322 * SIGCOMM '88, p. 328.
323 */
Eric Dumazet24901552014-05-02 21:18:05 -0700324void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700325{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300326 struct tcp_sock *tp = tcp_sk(sk);
327
Eric Dumazet24901552014-05-02 21:18:05 -0700328 if (!tcp_is_cwnd_limited(sk))
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700329 return;
330
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800331 /* In "safe" area, increase. */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900332 if (tp->snd_cwnd <= tp->snd_ssthresh)
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700333 tcp_slow_start(tp, acked);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900334 /* In dangerous area, increase slowly. */
Stephen Hemmingerca2eb562013-02-05 07:25:17 +0000335 else
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000336 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700337}
338EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
339
340/* Slow start threshold is half the congestion window (min 2) */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300341u32 tcp_reno_ssthresh(struct sock *sk)
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700342{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300343 const struct tcp_sock *tp = tcp_sk(sk);
stephen hemminger688d1942014-08-29 23:32:05 -0700344
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700345 return max(tp->snd_cwnd >> 1U, 2U);
346}
347EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
348
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700349struct tcp_congestion_ops tcp_reno = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700350 .flags = TCP_CONG_NON_RESTRICTED,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700351 .name = "reno",
352 .owner = THIS_MODULE,
353 .ssthresh = tcp_reno_ssthresh,
354 .cong_avoid = tcp_reno_cong_avoid,
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700355};