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 these */
72 if (!ca->ssthresh || !ca->undo_cwnd ||
73 !(ca->cong_avoid || ca->cong_control)) {
74 pr_err("%s does not implement required ops\n", ca->name);
78 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
80 spin_lock(&tcp_cong_list_lock);
81 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
82 pr_notice("%s already registered or non-unique key\n",
86 list_add_tail_rcu(&ca->list, &tcp_cong_list);
87 pr_debug("%s registered\n", ca->name);
89 spin_unlock(&tcp_cong_list_lock);
93 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
96 * Remove congestion control algorithm, called from
97 * the module's remove function. Module ref counts are used
98 * to ensure that this can't be done till all sockets using
99 * that method are closed.
101 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
103 spin_lock(&tcp_cong_list_lock);
104 list_del_rcu(&ca->list);
105 spin_unlock(&tcp_cong_list_lock);
107 /* Wait for outstanding readers to complete before the
108 * module gets removed entirely.
110 * A try_module_get() should fail by now as our module is
111 * in "going" state since no refs are held anymore and
112 * module_exit() handler being called.
116 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
118 u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
120 const struct tcp_congestion_ops *ca;
121 u32 key = TCP_CA_UNSPEC;
126 ca = __tcp_ca_find_autoload(name);
129 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
135 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
137 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
139 const struct tcp_congestion_ops *ca;
143 ca = tcp_ca_find_key(key);
145 ret = strncpy(buffer, ca->name,
151 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
153 /* Assign choice of congestion control. */
154 void tcp_assign_congestion_control(struct sock *sk)
156 struct inet_connection_sock *icsk = inet_csk(sk);
157 struct tcp_congestion_ops *ca;
160 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
161 if (likely(try_module_get(ca->owner))) {
162 icsk->icsk_ca_ops = ca;
165 /* Fallback to next available. The last really
166 * guaranteed fallback is Reno from this list.
171 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
173 if (ca->flags & TCP_CONG_NEEDS_ECN)
176 INET_ECN_dontxmit(sk);
179 void tcp_init_congestion_control(struct sock *sk)
181 const struct inet_connection_sock *icsk = inet_csk(sk);
183 tcp_sk(sk)->prior_ssthresh = 0;
184 if (icsk->icsk_ca_ops->init)
185 icsk->icsk_ca_ops->init(sk);
186 if (tcp_ca_needs_ecn(sk))
189 INET_ECN_dontxmit(sk);
192 static void tcp_reinit_congestion_control(struct sock *sk,
193 const struct tcp_congestion_ops *ca)
195 struct inet_connection_sock *icsk = inet_csk(sk);
197 tcp_cleanup_congestion_control(sk);
198 icsk->icsk_ca_ops = ca;
199 icsk->icsk_ca_setsockopt = 1;
200 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
202 if (ca->flags & TCP_CONG_NEEDS_ECN)
205 INET_ECN_dontxmit(sk);
207 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
208 tcp_init_congestion_control(sk);
211 /* Manage refcounts on socket close. */
212 void tcp_cleanup_congestion_control(struct sock *sk)
214 struct inet_connection_sock *icsk = inet_csk(sk);
216 if (icsk->icsk_ca_ops->release)
217 icsk->icsk_ca_ops->release(sk);
218 module_put(icsk->icsk_ca_ops->owner);
221 /* Used by sysctl to change default congestion control */
222 int tcp_set_default_congestion_control(const char *name)
224 struct tcp_congestion_ops *ca;
227 spin_lock(&tcp_cong_list_lock);
228 ca = tcp_ca_find(name);
229 #ifdef CONFIG_MODULES
230 if (!ca && capable(CAP_NET_ADMIN)) {
231 spin_unlock(&tcp_cong_list_lock);
233 request_module("tcp_%s", name);
234 spin_lock(&tcp_cong_list_lock);
235 ca = tcp_ca_find(name);
240 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
241 list_move(&ca->list, &tcp_cong_list);
244 spin_unlock(&tcp_cong_list_lock);
249 /* Set default value from kernel configuration at bootup */
250 static int __init tcp_congestion_default(void)
252 return tcp_set_default_congestion_control(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(char *name)
274 struct tcp_congestion_ops *ca;
275 /* We will always have reno... */
276 BUG_ON(list_empty(&tcp_cong_list));
279 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
280 strncpy(name, ca->name, TCP_CA_NAME_MAX);
284 /* Built list of non-restricted congestion control values */
285 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
287 struct tcp_congestion_ops *ca;
292 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
293 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
295 offs += snprintf(buf + offs, maxlen - offs,
297 offs == 0 ? "" : " ", ca->name);
302 /* Change list of non-restricted congestion control */
303 int tcp_set_allowed_congestion_control(char *val)
305 struct tcp_congestion_ops *ca;
306 char *saved_clone, *clone, *name;
309 saved_clone = clone = kstrdup(val, GFP_USER);
313 spin_lock(&tcp_cong_list_lock);
314 /* pass 1 check for bad entries */
315 while ((name = strsep(&clone, " ")) && *name) {
316 ca = tcp_ca_find(name);
323 /* pass 2 clear old values */
324 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
325 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
327 /* pass 3 mark as allowed */
328 while ((name = strsep(&val, " ")) && *name) {
329 ca = tcp_ca_find(name);
332 ca->flags |= TCP_CONG_NON_RESTRICTED;
335 spin_unlock(&tcp_cong_list_lock);
341 /* Change congestion control for socket. If load is false, then it is the
342 * responsibility of the caller to call tcp_init_congestion_control or
343 * tcp_reinit_congestion_control (if the current congestion control was
344 * already initialized.
346 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
347 bool reinit, bool cap_net_admin)
349 struct inet_connection_sock *icsk = inet_csk(sk);
350 const struct tcp_congestion_ops *ca;
353 if (icsk->icsk_ca_dst_locked)
358 ca = tcp_ca_find(name);
360 ca = __tcp_ca_find_autoload(name);
361 /* No change asking for existing value */
362 if (ca == icsk->icsk_ca_ops) {
363 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,