1 // SPDX-License-Identifier: GPL-2.0-only
3 * Pluggable TCP congestion control support and newReno
5 * Based on ideas from I/O scheduler support and Web100.
7 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
10 #define pr_fmt(fmt) "TCP: " fmt
12 #include <linux/module.h>
14 #include <linux/types.h>
15 #include <linux/list.h>
16 #include <linux/gfp.h>
17 #include <linux/jhash.h>
19 #include <trace/events/tcp.h>
21 static DEFINE_SPINLOCK(tcp_cong_list_lock);
22 static LIST_HEAD(tcp_cong_list);
24 /* Simple linear search, don't expect many entries! */
25 struct tcp_congestion_ops *tcp_ca_find(const char *name)
27 struct tcp_congestion_ops *e;
29 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
30 if (strcmp(e->name, name) == 0)
37 void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
39 struct inet_connection_sock *icsk = inet_csk(sk);
41 trace_tcp_cong_state_set(sk, ca_state);
43 if (icsk->icsk_ca_ops->set_state)
44 icsk->icsk_ca_ops->set_state(sk, ca_state);
45 icsk->icsk_ca_state = ca_state;
48 /* Must be called with rcu lock held */
49 static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
52 struct tcp_congestion_ops *ca = tcp_ca_find(name);
55 if (!ca && capable(CAP_NET_ADMIN)) {
57 request_module("tcp_%s", name);
59 ca = tcp_ca_find(name);
65 /* Simple linear search, not much in here. */
66 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
68 struct tcp_congestion_ops *e;
70 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
79 * Attach new congestion control algorithm to the list
80 * of available options.
82 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
86 /* all algorithms must implement these */
87 if (!ca->ssthresh || !ca->undo_cwnd ||
88 !(ca->cong_avoid || ca->cong_control)) {
89 pr_err("%s does not implement required ops\n", ca->name);
93 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
95 spin_lock(&tcp_cong_list_lock);
96 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
97 pr_notice("%s already registered or non-unique key\n",
101 list_add_tail_rcu(&ca->list, &tcp_cong_list);
102 pr_debug("%s registered\n", ca->name);
104 spin_unlock(&tcp_cong_list_lock);
108 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
111 * Remove congestion control algorithm, called from
112 * the module's remove function. Module ref counts are used
113 * to ensure that this can't be done till all sockets using
114 * that method are closed.
116 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
118 spin_lock(&tcp_cong_list_lock);
119 list_del_rcu(&ca->list);
120 spin_unlock(&tcp_cong_list_lock);
122 /* Wait for outstanding readers to complete before the
123 * module gets removed entirely.
125 * A try_module_get() should fail by now as our module is
126 * in "going" state since no refs are held anymore and
127 * module_exit() handler being called.
131 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
133 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
135 const struct tcp_congestion_ops *ca;
136 u32 key = TCP_CA_UNSPEC;
141 ca = tcp_ca_find_autoload(net, name);
144 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
151 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
153 const struct tcp_congestion_ops *ca;
157 ca = tcp_ca_find_key(key);
159 ret = strncpy(buffer, ca->name,
166 /* Assign choice of congestion control. */
167 void tcp_assign_congestion_control(struct sock *sk)
169 struct net *net = sock_net(sk);
170 struct inet_connection_sock *icsk = inet_csk(sk);
171 const struct tcp_congestion_ops *ca;
174 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
175 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
177 icsk->icsk_ca_ops = ca;
180 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
181 if (ca->flags & TCP_CONG_NEEDS_ECN)
184 INET_ECN_dontxmit(sk);
187 void tcp_init_congestion_control(struct sock *sk)
189 struct inet_connection_sock *icsk = inet_csk(sk);
191 tcp_sk(sk)->prior_ssthresh = 0;
192 if (icsk->icsk_ca_ops->init)
193 icsk->icsk_ca_ops->init(sk);
194 if (tcp_ca_needs_ecn(sk))
197 INET_ECN_dontxmit(sk);
198 icsk->icsk_ca_initialized = 1;
201 static void tcp_reinit_congestion_control(struct sock *sk,
202 const struct tcp_congestion_ops *ca)
204 struct inet_connection_sock *icsk = inet_csk(sk);
206 tcp_cleanup_congestion_control(sk);
207 icsk->icsk_ca_ops = ca;
208 icsk->icsk_ca_setsockopt = 1;
209 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
211 if (ca->flags & TCP_CONG_NEEDS_ECN)
214 INET_ECN_dontxmit(sk);
216 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
217 tcp_init_congestion_control(sk);
220 /* Manage refcounts on socket close. */
221 void tcp_cleanup_congestion_control(struct sock *sk)
223 struct inet_connection_sock *icsk = inet_csk(sk);
225 if (icsk->icsk_ca_ops->release)
226 icsk->icsk_ca_ops->release(sk);
227 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
230 /* Used by sysctl to change default congestion control */
231 int tcp_set_default_congestion_control(struct net *net, const char *name)
233 struct tcp_congestion_ops *ca;
234 const struct tcp_congestion_ops *prev;
238 ca = tcp_ca_find_autoload(net, name);
241 } else if (!bpf_try_module_get(ca, ca->owner)) {
243 } else if (!net_eq(net, &init_net) &&
244 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
245 /* Only init netns can set default to a restricted algorithm */
248 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
250 bpf_module_put(prev, prev->owner);
252 ca->flags |= TCP_CONG_NON_RESTRICTED;
260 /* Set default value from kernel configuration at bootup */
261 static int __init tcp_congestion_default(void)
263 return tcp_set_default_congestion_control(&init_net,
264 CONFIG_DEFAULT_TCP_CONG);
266 late_initcall(tcp_congestion_default);
268 /* Build string with list of available congestion control values */
269 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
271 struct tcp_congestion_ops *ca;
275 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
276 offs += snprintf(buf + offs, maxlen - offs,
278 offs == 0 ? "" : " ", ca->name);
280 if (WARN_ON_ONCE(offs >= maxlen))
286 /* Get current default congestion control */
287 void tcp_get_default_congestion_control(struct net *net, char *name)
289 const struct tcp_congestion_ops *ca;
292 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
293 strncpy(name, ca->name, TCP_CA_NAME_MAX);
297 /* Built list of non-restricted congestion control values */
298 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
300 struct tcp_congestion_ops *ca;
305 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
306 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
308 offs += snprintf(buf + offs, maxlen - offs,
310 offs == 0 ? "" : " ", ca->name);
312 if (WARN_ON_ONCE(offs >= maxlen))
318 /* Change list of non-restricted congestion control */
319 int tcp_set_allowed_congestion_control(char *val)
321 struct tcp_congestion_ops *ca;
322 char *saved_clone, *clone, *name;
325 saved_clone = clone = kstrdup(val, GFP_USER);
329 spin_lock(&tcp_cong_list_lock);
330 /* pass 1 check for bad entries */
331 while ((name = strsep(&clone, " ")) && *name) {
332 ca = tcp_ca_find(name);
339 /* pass 2 clear old values */
340 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
341 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
343 /* pass 3 mark as allowed */
344 while ((name = strsep(&val, " ")) && *name) {
345 ca = tcp_ca_find(name);
348 ca->flags |= TCP_CONG_NON_RESTRICTED;
351 spin_unlock(&tcp_cong_list_lock);
357 /* Change congestion control for socket. If load is false, then it is the
358 * responsibility of the caller to call tcp_init_congestion_control or
359 * tcp_reinit_congestion_control (if the current congestion control was
360 * already initialized.
362 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
365 struct inet_connection_sock *icsk = inet_csk(sk);
366 const struct tcp_congestion_ops *ca;
369 if (icsk->icsk_ca_dst_locked)
374 ca = tcp_ca_find(name);
376 ca = tcp_ca_find_autoload(sock_net(sk), name);
378 /* No change asking for existing value */
379 if (ca == icsk->icsk_ca_ops) {
380 icsk->icsk_ca_setsockopt = 1;
386 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
388 else if (!bpf_try_module_get(ca, ca->owner))
391 tcp_reinit_congestion_control(sk, ca);
397 /* Slow start is used when congestion window is no greater than the slow start
398 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
399 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
400 * something better;) a packet is only considered (s)acked in its entirety to
401 * defend the ACK attacks described in the RFC. Slow start processes a stretch
402 * ACK of degree N as if N acks of degree 1 are received back to back except
403 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
404 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
406 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
408 u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
410 acked -= cwnd - tcp_snd_cwnd(tp);
411 tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
415 EXPORT_SYMBOL_GPL(tcp_slow_start);
417 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
418 * for every packet that was ACKed.
420 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
422 /* If credits accumulated at a higher w, apply them gently now. */
423 if (tp->snd_cwnd_cnt >= w) {
424 tp->snd_cwnd_cnt = 0;
425 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
428 tp->snd_cwnd_cnt += acked;
429 if (tp->snd_cwnd_cnt >= w) {
430 u32 delta = tp->snd_cwnd_cnt / w;
432 tp->snd_cwnd_cnt -= delta * w;
433 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
435 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
437 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
440 * TCP Reno congestion control
441 * This is special case used for fallback as well.
443 /* This is Jacobson's slow start and congestion avoidance.
444 * SIGCOMM '88, p. 328.
446 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
448 struct tcp_sock *tp = tcp_sk(sk);
450 if (!tcp_is_cwnd_limited(sk))
453 /* In "safe" area, increase. */
454 if (tcp_in_slow_start(tp)) {
455 acked = tcp_slow_start(tp, acked);
459 /* In dangerous area, increase slowly. */
460 tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
462 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
464 /* Slow start threshold is half the congestion window (min 2) */
465 u32 tcp_reno_ssthresh(struct sock *sk)
467 const struct tcp_sock *tp = tcp_sk(sk);
469 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
471 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
473 u32 tcp_reno_undo_cwnd(struct sock *sk)
475 const struct tcp_sock *tp = tcp_sk(sk);
477 return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
479 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
481 struct tcp_congestion_ops tcp_reno = {
482 .flags = TCP_CONG_NON_RESTRICTED,
484 .owner = THIS_MODULE,
485 .ssthresh = tcp_reno_ssthresh,
486 .cong_avoid = tcp_reno_cong_avoid,
487 .undo_cwnd = tcp_reno_undo_cwnd,