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.
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(const char *name)
51 struct tcp_congestion_ops *ca = tcp_ca_find(name);
54 if (!ca && capable(CAP_NET_ADMIN)) {
56 request_module("tcp_%s", name);
58 ca = tcp_ca_find(name);
64 /* Simple linear search, not much in here. */
65 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
67 struct tcp_congestion_ops *e;
69 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
77 int tcp_validate_congestion_control(struct tcp_congestion_ops *ca)
79 /* all algorithms must implement these */
80 if (!ca->ssthresh || !ca->undo_cwnd ||
81 !(ca->cong_avoid || ca->cong_control)) {
82 pr_err("%s does not implement required ops\n", ca->name);
89 /* Attach new congestion control algorithm to the list
90 * of available options.
92 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
96 ret = tcp_validate_congestion_control(ca);
100 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
102 spin_lock(&tcp_cong_list_lock);
103 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
104 pr_notice("%s already registered or non-unique key\n",
108 list_add_tail_rcu(&ca->list, &tcp_cong_list);
109 pr_debug("%s registered\n", ca->name);
111 spin_unlock(&tcp_cong_list_lock);
115 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
118 * Remove congestion control algorithm, called from
119 * the module's remove function. Module ref counts are used
120 * to ensure that this can't be done till all sockets using
121 * that method are closed.
123 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
125 spin_lock(&tcp_cong_list_lock);
126 list_del_rcu(&ca->list);
127 spin_unlock(&tcp_cong_list_lock);
129 /* Wait for outstanding readers to complete before the
130 * module gets removed entirely.
132 * A try_module_get() should fail by now as our module is
133 * in "going" state since no refs are held anymore and
134 * module_exit() handler being called.
138 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
140 /* Replace a registered old ca with a new one.
142 * The new ca must have the same name as the old one, that has been
145 int tcp_update_congestion_control(struct tcp_congestion_ops *ca, struct tcp_congestion_ops *old_ca)
147 struct tcp_congestion_ops *existing;
150 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
152 spin_lock(&tcp_cong_list_lock);
153 existing = tcp_ca_find_key(old_ca->key);
154 if (ca->key == TCP_CA_UNSPEC || !existing || strcmp(existing->name, ca->name)) {
155 pr_notice("%s not registered or non-unique key\n",
158 } else if (existing != old_ca) {
159 pr_notice("invalid old congestion control algorithm to replace\n");
162 /* Add the new one before removing the old one to keep
163 * one implementation available all the time.
165 list_add_tail_rcu(&ca->list, &tcp_cong_list);
166 list_del_rcu(&existing->list);
167 pr_debug("%s updated\n", ca->name);
169 spin_unlock(&tcp_cong_list_lock);
171 /* Wait for outstanding readers to complete before the
172 * module or struct_ops gets removed entirely.
180 u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
182 const struct tcp_congestion_ops *ca;
183 u32 key = TCP_CA_UNSPEC;
188 ca = tcp_ca_find_autoload(name);
191 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
198 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
200 const struct tcp_congestion_ops *ca;
204 ca = tcp_ca_find_key(key);
206 strscpy(buffer, ca->name, TCP_CA_NAME_MAX);
214 /* Assign choice of congestion control. */
215 void tcp_assign_congestion_control(struct sock *sk)
217 struct net *net = sock_net(sk);
218 struct inet_connection_sock *icsk = inet_csk(sk);
219 const struct tcp_congestion_ops *ca;
222 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
223 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
225 icsk->icsk_ca_ops = ca;
228 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
229 if (ca->flags & TCP_CONG_NEEDS_ECN)
232 INET_ECN_dontxmit(sk);
235 void tcp_init_congestion_control(struct sock *sk)
237 struct inet_connection_sock *icsk = inet_csk(sk);
239 tcp_sk(sk)->prior_ssthresh = 0;
240 if (icsk->icsk_ca_ops->init)
241 icsk->icsk_ca_ops->init(sk);
242 if (tcp_ca_needs_ecn(sk))
245 INET_ECN_dontxmit(sk);
246 icsk->icsk_ca_initialized = 1;
249 static void tcp_reinit_congestion_control(struct sock *sk,
250 const struct tcp_congestion_ops *ca)
252 struct inet_connection_sock *icsk = inet_csk(sk);
254 tcp_cleanup_congestion_control(sk);
255 icsk->icsk_ca_ops = ca;
256 icsk->icsk_ca_setsockopt = 1;
257 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
259 if (ca->flags & TCP_CONG_NEEDS_ECN)
262 INET_ECN_dontxmit(sk);
264 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
265 tcp_init_congestion_control(sk);
268 /* Manage refcounts on socket close. */
269 void tcp_cleanup_congestion_control(struct sock *sk)
271 struct inet_connection_sock *icsk = inet_csk(sk);
273 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
274 icsk->icsk_ca_ops->release(sk);
275 icsk->icsk_ca_initialized = 0;
276 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
279 /* Used by sysctl to change default congestion control */
280 int tcp_set_default_congestion_control(struct net *net, const char *name)
282 struct tcp_congestion_ops *ca;
283 const struct tcp_congestion_ops *prev;
287 ca = tcp_ca_find_autoload(name);
290 } else if (!bpf_try_module_get(ca, ca->owner)) {
292 } else if (!net_eq(net, &init_net) &&
293 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
294 /* Only init netns can set default to a restricted algorithm */
297 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
299 bpf_module_put(prev, prev->owner);
301 ca->flags |= TCP_CONG_NON_RESTRICTED;
309 /* Set default value from kernel configuration at bootup */
310 static int __init tcp_congestion_default(void)
312 return tcp_set_default_congestion_control(&init_net,
313 CONFIG_DEFAULT_TCP_CONG);
315 late_initcall(tcp_congestion_default);
317 /* Build string with list of available congestion control values */
318 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
320 struct tcp_congestion_ops *ca;
324 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
325 offs += snprintf(buf + offs, maxlen - offs,
327 offs == 0 ? "" : " ", ca->name);
329 if (WARN_ON_ONCE(offs >= maxlen))
335 /* Get current default congestion control */
336 void tcp_get_default_congestion_control(struct net *net, char *name)
338 const struct tcp_congestion_ops *ca;
341 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
342 strscpy(name, ca->name, TCP_CA_NAME_MAX);
346 /* Built list of non-restricted congestion control values */
347 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
349 struct tcp_congestion_ops *ca;
354 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
355 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
357 offs += snprintf(buf + offs, maxlen - offs,
359 offs == 0 ? "" : " ", ca->name);
361 if (WARN_ON_ONCE(offs >= maxlen))
367 /* Change list of non-restricted congestion control */
368 int tcp_set_allowed_congestion_control(char *val)
370 struct tcp_congestion_ops *ca;
371 char *saved_clone, *clone, *name;
374 saved_clone = clone = kstrdup(val, GFP_USER);
378 spin_lock(&tcp_cong_list_lock);
379 /* pass 1 check for bad entries */
380 while ((name = strsep(&clone, " ")) && *name) {
381 ca = tcp_ca_find(name);
388 /* pass 2 clear old values */
389 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
390 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
392 /* pass 3 mark as allowed */
393 while ((name = strsep(&val, " ")) && *name) {
394 ca = tcp_ca_find(name);
397 ca->flags |= TCP_CONG_NON_RESTRICTED;
400 spin_unlock(&tcp_cong_list_lock);
406 /* Change congestion control for socket. If load is false, then it is the
407 * responsibility of the caller to call tcp_init_congestion_control or
408 * tcp_reinit_congestion_control (if the current congestion control was
409 * already initialized.
411 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 const struct tcp_congestion_ops *ca;
418 if (icsk->icsk_ca_dst_locked)
423 ca = tcp_ca_find(name);
425 ca = tcp_ca_find_autoload(name);
427 /* No change asking for existing value */
428 if (ca == icsk->icsk_ca_ops) {
429 icsk->icsk_ca_setsockopt = 1;
435 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
437 else if (!bpf_try_module_get(ca, ca->owner))
440 tcp_reinit_congestion_control(sk, ca);
446 /* Slow start is used when congestion window is no greater than the slow start
447 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
448 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
449 * something better;) a packet is only considered (s)acked in its entirety to
450 * defend the ACK attacks described in the RFC. Slow start processes a stretch
451 * ACK of degree N as if N acks of degree 1 are received back to back except
452 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
453 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
455 __bpf_kfunc u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
457 u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
459 acked -= cwnd - tcp_snd_cwnd(tp);
460 tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
464 EXPORT_SYMBOL_GPL(tcp_slow_start);
466 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
467 * for every packet that was ACKed.
469 __bpf_kfunc void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
471 /* If credits accumulated at a higher w, apply them gently now. */
472 if (tp->snd_cwnd_cnt >= w) {
473 tp->snd_cwnd_cnt = 0;
474 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
477 tp->snd_cwnd_cnt += acked;
478 if (tp->snd_cwnd_cnt >= w) {
479 u32 delta = tp->snd_cwnd_cnt / w;
481 tp->snd_cwnd_cnt -= delta * w;
482 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
484 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
486 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
489 * TCP Reno congestion control
490 * This is special case used for fallback as well.
492 /* This is Jacobson's slow start and congestion avoidance.
493 * SIGCOMM '88, p. 328.
495 __bpf_kfunc void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
497 struct tcp_sock *tp = tcp_sk(sk);
499 if (!tcp_is_cwnd_limited(sk))
502 /* In "safe" area, increase. */
503 if (tcp_in_slow_start(tp)) {
504 acked = tcp_slow_start(tp, acked);
508 /* In dangerous area, increase slowly. */
509 tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
511 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
513 /* Slow start threshold is half the congestion window (min 2) */
514 __bpf_kfunc u32 tcp_reno_ssthresh(struct sock *sk)
516 const struct tcp_sock *tp = tcp_sk(sk);
518 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
520 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
522 __bpf_kfunc u32 tcp_reno_undo_cwnd(struct sock *sk)
524 const struct tcp_sock *tp = tcp_sk(sk);
526 return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
528 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
530 struct tcp_congestion_ops tcp_reno = {
531 .flags = TCP_CONG_NON_RESTRICTED,
533 .owner = THIS_MODULE,
534 .ssthresh = tcp_reno_ssthresh,
535 .cong_avoid = tcp_reno_cong_avoid,
536 .undo_cwnd = tcp_reno_undo_cwnd,