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) {
78 int tcp_validate_congestion_control(struct tcp_congestion_ops *ca)
80 /* all algorithms must implement these */
81 if (!ca->ssthresh || !ca->undo_cwnd ||
82 !(ca->cong_avoid || ca->cong_control)) {
83 pr_err("%s does not implement required ops\n", ca->name);
90 /* Attach new congestion control algorithm to the list
91 * of available options.
93 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
97 ret = tcp_validate_congestion_control(ca);
101 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
103 spin_lock(&tcp_cong_list_lock);
104 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
105 pr_notice("%s already registered or non-unique key\n",
109 list_add_tail_rcu(&ca->list, &tcp_cong_list);
110 pr_debug("%s registered\n", ca->name);
112 spin_unlock(&tcp_cong_list_lock);
116 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
119 * Remove congestion control algorithm, called from
120 * the module's remove function. Module ref counts are used
121 * to ensure that this can't be done till all sockets using
122 * that method are closed.
124 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
126 spin_lock(&tcp_cong_list_lock);
127 list_del_rcu(&ca->list);
128 spin_unlock(&tcp_cong_list_lock);
130 /* Wait for outstanding readers to complete before the
131 * module gets removed entirely.
133 * A try_module_get() should fail by now as our module is
134 * in "going" state since no refs are held anymore and
135 * module_exit() handler being called.
139 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
141 /* Replace a registered old ca with a new one.
143 * The new ca must have the same name as the old one, that has been
146 int tcp_update_congestion_control(struct tcp_congestion_ops *ca, struct tcp_congestion_ops *old_ca)
148 struct tcp_congestion_ops *existing;
151 ret = tcp_validate_congestion_control(ca);
155 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
157 spin_lock(&tcp_cong_list_lock);
158 existing = tcp_ca_find_key(old_ca->key);
159 if (ca->key == TCP_CA_UNSPEC || !existing || strcmp(existing->name, ca->name)) {
160 pr_notice("%s not registered or non-unique key\n",
163 } else if (existing != old_ca) {
164 pr_notice("invalid old congestion control algorithm to replace\n");
167 /* Add the new one before removing the old one to keep
168 * one implementation available all the time.
170 list_add_tail_rcu(&ca->list, &tcp_cong_list);
171 list_del_rcu(&existing->list);
172 pr_debug("%s updated\n", ca->name);
174 spin_unlock(&tcp_cong_list_lock);
176 /* Wait for outstanding readers to complete before the
177 * module or struct_ops gets removed entirely.
185 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
187 const struct tcp_congestion_ops *ca;
188 u32 key = TCP_CA_UNSPEC;
193 ca = tcp_ca_find_autoload(net, name);
196 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
203 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
205 const struct tcp_congestion_ops *ca;
209 ca = tcp_ca_find_key(key);
211 ret = strncpy(buffer, ca->name,
218 /* Assign choice of congestion control. */
219 void tcp_assign_congestion_control(struct sock *sk)
221 struct net *net = sock_net(sk);
222 struct inet_connection_sock *icsk = inet_csk(sk);
223 const struct tcp_congestion_ops *ca;
226 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
227 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
229 icsk->icsk_ca_ops = ca;
232 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
233 if (ca->flags & TCP_CONG_NEEDS_ECN)
236 INET_ECN_dontxmit(sk);
239 void tcp_init_congestion_control(struct sock *sk)
241 struct inet_connection_sock *icsk = inet_csk(sk);
243 tcp_sk(sk)->prior_ssthresh = 0;
244 if (icsk->icsk_ca_ops->init)
245 icsk->icsk_ca_ops->init(sk);
246 if (tcp_ca_needs_ecn(sk))
249 INET_ECN_dontxmit(sk);
250 icsk->icsk_ca_initialized = 1;
253 static void tcp_reinit_congestion_control(struct sock *sk,
254 const struct tcp_congestion_ops *ca)
256 struct inet_connection_sock *icsk = inet_csk(sk);
258 tcp_cleanup_congestion_control(sk);
259 icsk->icsk_ca_ops = ca;
260 icsk->icsk_ca_setsockopt = 1;
261 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
263 if (ca->flags & TCP_CONG_NEEDS_ECN)
266 INET_ECN_dontxmit(sk);
268 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
269 tcp_init_congestion_control(sk);
272 /* Manage refcounts on socket close. */
273 void tcp_cleanup_congestion_control(struct sock *sk)
275 struct inet_connection_sock *icsk = inet_csk(sk);
277 if (icsk->icsk_ca_ops->release)
278 icsk->icsk_ca_ops->release(sk);
279 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
282 /* Used by sysctl to change default congestion control */
283 int tcp_set_default_congestion_control(struct net *net, const char *name)
285 struct tcp_congestion_ops *ca;
286 const struct tcp_congestion_ops *prev;
290 ca = tcp_ca_find_autoload(net, name);
293 } else if (!bpf_try_module_get(ca, ca->owner)) {
295 } else if (!net_eq(net, &init_net) &&
296 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
297 /* Only init netns can set default to a restricted algorithm */
300 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
302 bpf_module_put(prev, prev->owner);
304 ca->flags |= TCP_CONG_NON_RESTRICTED;
312 /* Set default value from kernel configuration at bootup */
313 static int __init tcp_congestion_default(void)
315 return tcp_set_default_congestion_control(&init_net,
316 CONFIG_DEFAULT_TCP_CONG);
318 late_initcall(tcp_congestion_default);
320 /* Build string with list of available congestion control values */
321 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
323 struct tcp_congestion_ops *ca;
327 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
328 offs += snprintf(buf + offs, maxlen - offs,
330 offs == 0 ? "" : " ", ca->name);
332 if (WARN_ON_ONCE(offs >= maxlen))
338 /* Get current default congestion control */
339 void tcp_get_default_congestion_control(struct net *net, char *name)
341 const struct tcp_congestion_ops *ca;
344 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
345 strncpy(name, ca->name, TCP_CA_NAME_MAX);
349 /* Built list of non-restricted congestion control values */
350 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
352 struct tcp_congestion_ops *ca;
357 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
358 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
360 offs += snprintf(buf + offs, maxlen - offs,
362 offs == 0 ? "" : " ", ca->name);
364 if (WARN_ON_ONCE(offs >= maxlen))
370 /* Change list of non-restricted congestion control */
371 int tcp_set_allowed_congestion_control(char *val)
373 struct tcp_congestion_ops *ca;
374 char *saved_clone, *clone, *name;
377 saved_clone = clone = kstrdup(val, GFP_USER);
381 spin_lock(&tcp_cong_list_lock);
382 /* pass 1 check for bad entries */
383 while ((name = strsep(&clone, " ")) && *name) {
384 ca = tcp_ca_find(name);
391 /* pass 2 clear old values */
392 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
393 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
395 /* pass 3 mark as allowed */
396 while ((name = strsep(&val, " ")) && *name) {
397 ca = tcp_ca_find(name);
400 ca->flags |= TCP_CONG_NON_RESTRICTED;
403 spin_unlock(&tcp_cong_list_lock);
409 /* Change congestion control for socket. If load is false, then it is the
410 * responsibility of the caller to call tcp_init_congestion_control or
411 * tcp_reinit_congestion_control (if the current congestion control was
412 * already initialized.
414 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
417 struct inet_connection_sock *icsk = inet_csk(sk);
418 const struct tcp_congestion_ops *ca;
421 if (icsk->icsk_ca_dst_locked)
426 ca = tcp_ca_find(name);
428 ca = tcp_ca_find_autoload(sock_net(sk), name);
430 /* No change asking for existing value */
431 if (ca == icsk->icsk_ca_ops) {
432 icsk->icsk_ca_setsockopt = 1;
438 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
440 else if (!bpf_try_module_get(ca, ca->owner))
443 tcp_reinit_congestion_control(sk, ca);
449 /* Slow start is used when congestion window is no greater than the slow start
450 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
451 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
452 * something better;) a packet is only considered (s)acked in its entirety to
453 * defend the ACK attacks described in the RFC. Slow start processes a stretch
454 * ACK of degree N as if N acks of degree 1 are received back to back except
455 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
456 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
458 __bpf_kfunc u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
460 u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
462 acked -= cwnd - tcp_snd_cwnd(tp);
463 tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
467 EXPORT_SYMBOL_GPL(tcp_slow_start);
469 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
470 * for every packet that was ACKed.
472 __bpf_kfunc void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
474 /* If credits accumulated at a higher w, apply them gently now. */
475 if (tp->snd_cwnd_cnt >= w) {
476 tp->snd_cwnd_cnt = 0;
477 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
480 tp->snd_cwnd_cnt += acked;
481 if (tp->snd_cwnd_cnt >= w) {
482 u32 delta = tp->snd_cwnd_cnt / w;
484 tp->snd_cwnd_cnt -= delta * w;
485 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
487 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
489 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
492 * TCP Reno congestion control
493 * This is special case used for fallback as well.
495 /* This is Jacobson's slow start and congestion avoidance.
496 * SIGCOMM '88, p. 328.
498 __bpf_kfunc void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
500 struct tcp_sock *tp = tcp_sk(sk);
502 if (!tcp_is_cwnd_limited(sk))
505 /* In "safe" area, increase. */
506 if (tcp_in_slow_start(tp)) {
507 acked = tcp_slow_start(tp, acked);
511 /* In dangerous area, increase slowly. */
512 tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
514 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
516 /* Slow start threshold is half the congestion window (min 2) */
517 __bpf_kfunc u32 tcp_reno_ssthresh(struct sock *sk)
519 const struct tcp_sock *tp = tcp_sk(sk);
521 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
523 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
525 __bpf_kfunc u32 tcp_reno_undo_cwnd(struct sock *sk)
527 const struct tcp_sock *tp = tcp_sk(sk);
529 return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
531 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
533 struct tcp_congestion_ops tcp_reno = {
534 .flags = TCP_CONG_NON_RESTRICTED,
536 .owner = THIS_MODULE,
537 .ssthresh = tcp_reno_ssthresh,
538 .cong_avoid = tcp_reno_cong_avoid,
539 .undo_cwnd = tcp_reno_undo_cwnd,