2 * Pluggable TCP congestion control support and newReno
4 * Based on ideas from I/O scheduler support and Web100.
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
9 #define pr_fmt(fmt) "TCP: " fmt
11 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/list.h>
15 #include <linux/gfp.h>
16 #include <linux/jhash.h>
19 static DEFINE_SPINLOCK(tcp_cong_list_lock);
20 static LIST_HEAD(tcp_cong_list);
22 /* Simple linear search, don't expect many entries! */
23 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
25 struct tcp_congestion_ops *e;
27 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
28 if (strcmp(e->name, name) == 0)
35 /* Must be called with rcu lock held */
36 static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
39 struct tcp_congestion_ops *ca = tcp_ca_find(name);
42 if (!ca && capable(CAP_NET_ADMIN)) {
44 request_module("tcp_%s", name);
46 ca = tcp_ca_find(name);
52 /* Simple linear search, not much in here. */
53 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
55 struct tcp_congestion_ops *e;
57 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
66 * Attach new congestion control algorithm to the list
67 * of available options.
69 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
73 /* all algorithms must implement these */
74 if (!ca->ssthresh || !ca->undo_cwnd ||
75 !(ca->cong_avoid || ca->cong_control)) {
76 pr_err("%s does not implement required ops\n", ca->name);
80 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
82 spin_lock(&tcp_cong_list_lock);
83 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
84 pr_notice("%s already registered or non-unique key\n",
88 list_add_tail_rcu(&ca->list, &tcp_cong_list);
89 pr_debug("%s registered\n", ca->name);
91 spin_unlock(&tcp_cong_list_lock);
95 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
98 * Remove congestion control algorithm, called from
99 * the module's remove function. Module ref counts are used
100 * to ensure that this can't be done till all sockets using
101 * that method are closed.
103 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
105 spin_lock(&tcp_cong_list_lock);
106 list_del_rcu(&ca->list);
107 spin_unlock(&tcp_cong_list_lock);
109 /* Wait for outstanding readers to complete before the
110 * module gets removed entirely.
112 * A try_module_get() should fail by now as our module is
113 * in "going" state since no refs are held anymore and
114 * module_exit() handler being called.
118 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
122 const struct tcp_congestion_ops *ca;
123 u32 key = TCP_CA_UNSPEC;
128 ca = tcp_ca_find_autoload(net, name);
131 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
137 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
141 const struct tcp_congestion_ops *ca;
145 ca = tcp_ca_find_key(key);
147 ret = strncpy(buffer, ca->name,
153 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155 /* Assign choice of congestion control. */
156 void tcp_assign_congestion_control(struct sock *sk)
158 struct net *net = sock_net(sk);
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const struct tcp_congestion_ops *ca;
163 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
164 if (unlikely(!try_module_get(ca->owner)))
166 icsk->icsk_ca_ops = ca;
169 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
170 if (ca->flags & TCP_CONG_NEEDS_ECN)
173 INET_ECN_dontxmit(sk);
176 void tcp_init_congestion_control(struct sock *sk)
178 const struct inet_connection_sock *icsk = inet_csk(sk);
180 tcp_sk(sk)->prior_ssthresh = 0;
181 if (icsk->icsk_ca_ops->init)
182 icsk->icsk_ca_ops->init(sk);
183 if (tcp_ca_needs_ecn(sk))
186 INET_ECN_dontxmit(sk);
189 static void tcp_reinit_congestion_control(struct sock *sk,
190 const struct tcp_congestion_ops *ca)
192 struct inet_connection_sock *icsk = inet_csk(sk);
194 tcp_cleanup_congestion_control(sk);
195 icsk->icsk_ca_ops = ca;
196 icsk->icsk_ca_setsockopt = 1;
197 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
199 if (ca->flags & TCP_CONG_NEEDS_ECN)
202 INET_ECN_dontxmit(sk);
204 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
205 tcp_init_congestion_control(sk);
208 /* Manage refcounts on socket close. */
209 void tcp_cleanup_congestion_control(struct sock *sk)
211 struct inet_connection_sock *icsk = inet_csk(sk);
213 if (icsk->icsk_ca_ops->release)
214 icsk->icsk_ca_ops->release(sk);
215 module_put(icsk->icsk_ca_ops->owner);
218 /* Used by sysctl to change default congestion control */
219 int tcp_set_default_congestion_control(struct net *net, const char *name)
221 struct tcp_congestion_ops *ca;
222 const struct tcp_congestion_ops *prev;
226 ca = tcp_ca_find_autoload(net, name);
229 } else if (!try_module_get(ca->owner)) {
231 } else if (!net_eq(net, &init_net) &&
232 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
233 /* Only init netns can set default to a restricted algorithm */
236 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
238 module_put(prev->owner);
240 ca->flags |= TCP_CONG_NON_RESTRICTED;
248 /* Set default value from kernel configuration at bootup */
249 static int __init tcp_congestion_default(void)
251 return tcp_set_default_congestion_control(&init_net,
252 CONFIG_DEFAULT_TCP_CONG);
254 late_initcall(tcp_congestion_default);
256 /* Build string with list of available congestion control values */
257 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
259 struct tcp_congestion_ops *ca;
263 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
264 offs += snprintf(buf + offs, maxlen - offs,
266 offs == 0 ? "" : " ", ca->name);
271 /* Get current default congestion control */
272 void tcp_get_default_congestion_control(struct net *net, char *name)
274 const struct tcp_congestion_ops *ca;
277 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
278 strncpy(name, ca->name, TCP_CA_NAME_MAX);
282 /* Built list of non-restricted congestion control values */
283 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
285 struct tcp_congestion_ops *ca;
290 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
291 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
293 offs += snprintf(buf + offs, maxlen - offs,
295 offs == 0 ? "" : " ", ca->name);
300 /* Change list of non-restricted congestion control */
301 int tcp_set_allowed_congestion_control(char *val)
303 struct tcp_congestion_ops *ca;
304 char *saved_clone, *clone, *name;
307 saved_clone = clone = kstrdup(val, GFP_USER);
311 spin_lock(&tcp_cong_list_lock);
312 /* pass 1 check for bad entries */
313 while ((name = strsep(&clone, " ")) && *name) {
314 ca = tcp_ca_find(name);
321 /* pass 2 clear old values */
322 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
323 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
325 /* pass 3 mark as allowed */
326 while ((name = strsep(&val, " ")) && *name) {
327 ca = tcp_ca_find(name);
330 ca->flags |= TCP_CONG_NON_RESTRICTED;
333 spin_unlock(&tcp_cong_list_lock);
339 /* Change congestion control for socket. If load is false, then it is the
340 * responsibility of the caller to call tcp_init_congestion_control or
341 * tcp_reinit_congestion_control (if the current congestion control was
342 * already initialized.
344 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
345 bool reinit, bool cap_net_admin)
347 struct inet_connection_sock *icsk = inet_csk(sk);
348 const struct tcp_congestion_ops *ca;
351 if (icsk->icsk_ca_dst_locked)
356 ca = tcp_ca_find(name);
358 ca = tcp_ca_find_autoload(sock_net(sk), name);
360 /* No change asking for existing value */
361 if (ca == icsk->icsk_ca_ops) {
362 icsk->icsk_ca_setsockopt = 1;
369 const struct tcp_congestion_ops *old_ca = icsk->icsk_ca_ops;
371 if (try_module_get(ca->owner)) {
373 tcp_reinit_congestion_control(sk, ca);
375 icsk->icsk_ca_ops = ca;
376 module_put(old_ca->owner);
381 } else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin)) {
383 } else if (!try_module_get(ca->owner)) {
386 tcp_reinit_congestion_control(sk, ca);
393 /* Slow start is used when congestion window is no greater than the slow start
394 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
395 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
396 * something better;) a packet is only considered (s)acked in its entirety to
397 * defend the ACK attacks described in the RFC. Slow start processes a stretch
398 * ACK of degree N as if N acks of degree 1 are received back to back except
399 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
400 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
402 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
404 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
406 acked -= cwnd - tp->snd_cwnd;
407 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
411 EXPORT_SYMBOL_GPL(tcp_slow_start);
413 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
414 * for every packet that was ACKed.
416 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
418 /* If credits accumulated at a higher w, apply them gently now. */
419 if (tp->snd_cwnd_cnt >= w) {
420 tp->snd_cwnd_cnt = 0;
424 tp->snd_cwnd_cnt += acked;
425 if (tp->snd_cwnd_cnt >= w) {
426 u32 delta = tp->snd_cwnd_cnt / w;
428 tp->snd_cwnd_cnt -= delta * w;
429 tp->snd_cwnd += delta;
431 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
433 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
436 * TCP Reno congestion control
437 * This is special case used for fallback as well.
439 /* This is Jacobson's slow start and congestion avoidance.
440 * SIGCOMM '88, p. 328.
442 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
444 struct tcp_sock *tp = tcp_sk(sk);
446 if (!tcp_is_cwnd_limited(sk))
449 /* In "safe" area, increase. */
450 if (tcp_in_slow_start(tp)) {
451 acked = tcp_slow_start(tp, acked);
455 /* In dangerous area, increase slowly. */
456 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
458 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
460 /* Slow start threshold is half the congestion window (min 2) */
461 u32 tcp_reno_ssthresh(struct sock *sk)
463 const struct tcp_sock *tp = tcp_sk(sk);
465 return max(tp->snd_cwnd >> 1U, 2U);
467 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
469 u32 tcp_reno_undo_cwnd(struct sock *sk)
471 const struct tcp_sock *tp = tcp_sk(sk);
473 return max(tp->snd_cwnd, tp->prior_cwnd);
475 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
477 struct tcp_congestion_ops tcp_reno = {
478 .flags = TCP_CONG_NON_RESTRICTED,
480 .owner = THIS_MODULE,
481 .ssthresh = tcp_reno_ssthresh,
482 .cong_avoid = tcp_reno_cong_avoid,
483 .undo_cwnd = tcp_reno_undo_cwnd,