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));
176 if (ca->flags & TCP_CONG_NEEDS_ECN)
179 INET_ECN_dontxmit(sk);
182 void tcp_init_congestion_control(struct sock *sk)
184 const struct inet_connection_sock *icsk = inet_csk(sk);
186 tcp_sk(sk)->prior_ssthresh = 0;
187 if (icsk->icsk_ca_ops->init)
188 icsk->icsk_ca_ops->init(sk);
189 if (tcp_ca_needs_ecn(sk))
192 INET_ECN_dontxmit(sk);
195 static void tcp_reinit_congestion_control(struct sock *sk,
196 const struct tcp_congestion_ops *ca)
198 struct inet_connection_sock *icsk = inet_csk(sk);
200 tcp_cleanup_congestion_control(sk);
201 icsk->icsk_ca_ops = ca;
202 icsk->icsk_ca_setsockopt = 1;
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(const char *name)
221 struct tcp_congestion_ops *ca;
224 spin_lock(&tcp_cong_list_lock);
225 ca = tcp_ca_find(name);
226 #ifdef CONFIG_MODULES
227 if (!ca && capable(CAP_NET_ADMIN)) {
228 spin_unlock(&tcp_cong_list_lock);
230 request_module("tcp_%s", name);
231 spin_lock(&tcp_cong_list_lock);
232 ca = tcp_ca_find(name);
237 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
238 list_move(&ca->list, &tcp_cong_list);
241 spin_unlock(&tcp_cong_list_lock);
246 /* Set default value from kernel configuration at bootup */
247 static int __init tcp_congestion_default(void)
249 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
251 late_initcall(tcp_congestion_default);
253 /* Build string with list of available congestion control values */
254 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
256 struct tcp_congestion_ops *ca;
260 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
261 offs += snprintf(buf + offs, maxlen - offs,
263 offs == 0 ? "" : " ", ca->name);
268 /* Get current default congestion control */
269 void tcp_get_default_congestion_control(char *name)
271 struct tcp_congestion_ops *ca;
272 /* We will always have reno... */
273 BUG_ON(list_empty(&tcp_cong_list));
276 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
277 strncpy(name, ca->name, TCP_CA_NAME_MAX);
281 /* Built list of non-restricted congestion control values */
282 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
284 struct tcp_congestion_ops *ca;
289 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
290 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
292 offs += snprintf(buf + offs, maxlen - offs,
294 offs == 0 ? "" : " ", ca->name);
299 /* Change list of non-restricted congestion control */
300 int tcp_set_allowed_congestion_control(char *val)
302 struct tcp_congestion_ops *ca;
303 char *saved_clone, *clone, *name;
306 saved_clone = clone = kstrdup(val, GFP_USER);
310 spin_lock(&tcp_cong_list_lock);
311 /* pass 1 check for bad entries */
312 while ((name = strsep(&clone, " ")) && *name) {
313 ca = tcp_ca_find(name);
320 /* pass 2 clear old values */
321 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
322 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
324 /* pass 3 mark as allowed */
325 while ((name = strsep(&val, " ")) && *name) {
326 ca = tcp_ca_find(name);
329 ca->flags |= TCP_CONG_NON_RESTRICTED;
332 spin_unlock(&tcp_cong_list_lock);
338 /* Change congestion control for socket */
339 int tcp_set_congestion_control(struct sock *sk, const char *name)
341 struct inet_connection_sock *icsk = inet_csk(sk);
342 const struct tcp_congestion_ops *ca;
345 if (icsk->icsk_ca_dst_locked)
349 ca = __tcp_ca_find_autoload(name);
350 /* No change asking for existing value */
351 if (ca == icsk->icsk_ca_ops) {
352 icsk->icsk_ca_setsockopt = 1;
357 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
358 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
360 else if (!try_module_get(ca->owner))
363 tcp_reinit_congestion_control(sk, ca);
369 /* Slow start is used when congestion window is no greater than the slow start
370 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
371 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
372 * something better;) a packet is only considered (s)acked in its entirety to
373 * defend the ACK attacks described in the RFC. Slow start processes a stretch
374 * ACK of degree N as if N acks of degree 1 are received back to back except
375 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
376 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
378 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
380 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
382 acked -= cwnd - tp->snd_cwnd;
383 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
387 EXPORT_SYMBOL_GPL(tcp_slow_start);
389 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
390 * for every packet that was ACKed.
392 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
394 /* If credits accumulated at a higher w, apply them gently now. */
395 if (tp->snd_cwnd_cnt >= w) {
396 tp->snd_cwnd_cnt = 0;
400 tp->snd_cwnd_cnt += acked;
401 if (tp->snd_cwnd_cnt >= w) {
402 u32 delta = tp->snd_cwnd_cnt / w;
404 tp->snd_cwnd_cnt -= delta * w;
405 tp->snd_cwnd += delta;
407 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
409 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
412 * TCP Reno congestion control
413 * This is special case used for fallback as well.
415 /* This is Jacobson's slow start and congestion avoidance.
416 * SIGCOMM '88, p. 328.
418 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
420 struct tcp_sock *tp = tcp_sk(sk);
422 if (!tcp_is_cwnd_limited(sk))
425 /* In "safe" area, increase. */
426 if (tcp_in_slow_start(tp)) {
427 acked = tcp_slow_start(tp, acked);
431 /* In dangerous area, increase slowly. */
432 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
434 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
436 /* Slow start threshold is half the congestion window (min 2) */
437 u32 tcp_reno_ssthresh(struct sock *sk)
439 const struct tcp_sock *tp = tcp_sk(sk);
441 return max(tp->snd_cwnd >> 1U, 2U);
443 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
445 struct tcp_congestion_ops tcp_reno = {
446 .flags = TCP_CONG_NON_RESTRICTED,
448 .owner = THIS_MODULE,
449 .ssthresh = tcp_reno_ssthresh,
450 .cong_avoid = tcp_reno_cong_avoid,