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 const struct tcp_congestion_ops *__tcp_ca_find_autoload(const char *name)
38 const struct tcp_congestion_ops *ca = tcp_ca_find(name);
40 if (!ca && capable(CAP_NET_ADMIN)) {
42 request_module("tcp_%s", name);
44 ca = tcp_ca_find(name);
50 /* Simple linear search, not much in here. */
51 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
53 struct tcp_congestion_ops *e;
55 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
64 * Attach new congestion control algorithm to the list
65 * of available options.
67 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
71 /* all algorithms must implement ssthresh and cong_avoid ops */
72 if (!ca->ssthresh || !ca->cong_avoid) {
73 pr_err("%s does not implement required ops\n", ca->name);
77 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
79 spin_lock(&tcp_cong_list_lock);
80 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
81 pr_notice("%s already registered or non-unique key\n",
85 list_add_tail_rcu(&ca->list, &tcp_cong_list);
86 pr_debug("%s registered\n", ca->name);
88 spin_unlock(&tcp_cong_list_lock);
92 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
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.
100 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
102 spin_lock(&tcp_cong_list_lock);
103 list_del_rcu(&ca->list);
104 spin_unlock(&tcp_cong_list_lock);
106 /* Wait for outstanding readers to complete before the
107 * module gets removed entirely.
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.
115 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
117 u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
119 const struct tcp_congestion_ops *ca;
120 u32 key = TCP_CA_UNSPEC;
125 ca = __tcp_ca_find_autoload(name);
128 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
134 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
136 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
138 const struct tcp_congestion_ops *ca;
142 ca = tcp_ca_find_key(key);
144 ret = strncpy(buffer, ca->name,
150 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
152 /* Assign choice of congestion control. */
153 void tcp_assign_congestion_control(struct sock *sk)
155 struct inet_connection_sock *icsk = inet_csk(sk);
156 struct tcp_congestion_ops *ca;
159 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
160 if (likely(try_module_get(ca->owner))) {
161 icsk->icsk_ca_ops = ca;
164 /* Fallback to next available. The last really
165 * guaranteed fallback is Reno from this list.
171 /* Clear out private data before diag gets it and
172 * the ca has not been initialized.
175 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
178 void tcp_init_congestion_control(struct sock *sk)
180 const struct inet_connection_sock *icsk = inet_csk(sk);
182 if (icsk->icsk_ca_ops->init)
183 icsk->icsk_ca_ops->init(sk);
186 static void tcp_reinit_congestion_control(struct sock *sk,
187 const struct tcp_congestion_ops *ca)
189 struct inet_connection_sock *icsk = inet_csk(sk);
191 tcp_cleanup_congestion_control(sk);
192 icsk->icsk_ca_ops = ca;
193 icsk->icsk_ca_setsockopt = 1;
195 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
196 icsk->icsk_ca_ops->init(sk);
199 /* Manage refcounts on socket close. */
200 void tcp_cleanup_congestion_control(struct sock *sk)
202 struct inet_connection_sock *icsk = inet_csk(sk);
204 if (icsk->icsk_ca_ops->release)
205 icsk->icsk_ca_ops->release(sk);
206 module_put(icsk->icsk_ca_ops->owner);
209 /* Used by sysctl to change default congestion control */
210 int tcp_set_default_congestion_control(const char *name)
212 struct tcp_congestion_ops *ca;
215 spin_lock(&tcp_cong_list_lock);
216 ca = tcp_ca_find(name);
217 #ifdef CONFIG_MODULES
218 if (!ca && capable(CAP_NET_ADMIN)) {
219 spin_unlock(&tcp_cong_list_lock);
221 request_module("tcp_%s", name);
222 spin_lock(&tcp_cong_list_lock);
223 ca = tcp_ca_find(name);
228 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
229 list_move(&ca->list, &tcp_cong_list);
232 spin_unlock(&tcp_cong_list_lock);
237 /* Set default value from kernel configuration at bootup */
238 static int __init tcp_congestion_default(void)
240 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
242 late_initcall(tcp_congestion_default);
244 /* Build string with list of available congestion control values */
245 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
247 struct tcp_congestion_ops *ca;
251 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
252 offs += snprintf(buf + offs, maxlen - offs,
254 offs == 0 ? "" : " ", ca->name);
259 /* Get current default congestion control */
260 void tcp_get_default_congestion_control(char *name)
262 struct tcp_congestion_ops *ca;
263 /* We will always have reno... */
264 BUG_ON(list_empty(&tcp_cong_list));
267 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
268 strncpy(name, ca->name, TCP_CA_NAME_MAX);
272 /* Built list of non-restricted congestion control values */
273 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
275 struct tcp_congestion_ops *ca;
280 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
281 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
283 offs += snprintf(buf + offs, maxlen - offs,
285 offs == 0 ? "" : " ", ca->name);
290 /* Change list of non-restricted congestion control */
291 int tcp_set_allowed_congestion_control(char *val)
293 struct tcp_congestion_ops *ca;
294 char *saved_clone, *clone, *name;
297 saved_clone = clone = kstrdup(val, GFP_USER);
301 spin_lock(&tcp_cong_list_lock);
302 /* pass 1 check for bad entries */
303 while ((name = strsep(&clone, " ")) && *name) {
304 ca = tcp_ca_find(name);
311 /* pass 2 clear old values */
312 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
313 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
315 /* pass 3 mark as allowed */
316 while ((name = strsep(&val, " ")) && *name) {
317 ca = tcp_ca_find(name);
320 ca->flags |= TCP_CONG_NON_RESTRICTED;
323 spin_unlock(&tcp_cong_list_lock);
329 /* Change congestion control for socket */
330 int tcp_set_congestion_control(struct sock *sk, const char *name)
332 struct inet_connection_sock *icsk = inet_csk(sk);
333 const struct tcp_congestion_ops *ca;
336 if (icsk->icsk_ca_dst_locked)
340 ca = __tcp_ca_find_autoload(name);
341 /* No change asking for existing value */
342 if (ca == icsk->icsk_ca_ops) {
343 icsk->icsk_ca_setsockopt = 1;
348 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
349 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
351 else if (!try_module_get(ca->owner))
354 tcp_reinit_congestion_control(sk, ca);
360 /* Slow start is used when congestion window is no greater than the slow start
361 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
362 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
363 * something better;) a packet is only considered (s)acked in its entirety to
364 * defend the ACK attacks described in the RFC. Slow start processes a stretch
365 * ACK of degree N as if N acks of degree 1 are received back to back except
366 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
367 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
369 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
371 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
373 acked -= cwnd - tp->snd_cwnd;
374 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
378 EXPORT_SYMBOL_GPL(tcp_slow_start);
380 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
381 * for every packet that was ACKed.
383 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
385 /* If credits accumulated at a higher w, apply them gently now. */
386 if (tp->snd_cwnd_cnt >= w) {
387 tp->snd_cwnd_cnt = 0;
391 tp->snd_cwnd_cnt += acked;
392 if (tp->snd_cwnd_cnt >= w) {
393 u32 delta = tp->snd_cwnd_cnt / w;
395 tp->snd_cwnd_cnt -= delta * w;
396 tp->snd_cwnd += delta;
398 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
400 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
403 * TCP Reno congestion control
404 * This is special case used for fallback as well.
406 /* This is Jacobson's slow start and congestion avoidance.
407 * SIGCOMM '88, p. 328.
409 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
411 struct tcp_sock *tp = tcp_sk(sk);
413 if (!tcp_is_cwnd_limited(sk))
416 /* In "safe" area, increase. */
417 if (tcp_in_slow_start(tp)) {
418 acked = tcp_slow_start(tp, acked);
422 /* In dangerous area, increase slowly. */
423 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
425 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
427 /* Slow start threshold is half the congestion window (min 2) */
428 u32 tcp_reno_ssthresh(struct sock *sk)
430 const struct tcp_sock *tp = tcp_sk(sk);
432 return max(tp->snd_cwnd >> 1U, 2U);
434 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
436 struct tcp_congestion_ops tcp_reno = {
437 .flags = TCP_CONG_NON_RESTRICTED,
439 .owner = THIS_MODULE,
440 .ssthresh = tcp_reno_ssthresh,
441 .cong_avoid = tcp_reno_cong_avoid,