tcp_cong.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Pluggable TCP congestion control support and newReno
  4 * congestion control.
  5 * Based on ideas from I/O scheduler support and Web100.
  6 *
  7 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
  8 */
  9
 10#define pr_fmt(fmt) "TCP: " fmt
 11
 12#include <linux/module.h>
 13#include <linux/mm.h>
 14#include <linux/types.h>
 15#include <linux/list.h>
 16#include <linux/gfp.h>
 17#include <linux/jhash.h>
 18#include <net/tcp.h>
 19#include <net/tcp_ecn.h>
 20#include <trace/events/tcp.h>
 21
 22static DEFINE_SPINLOCK(tcp_cong_list_lock);
 23static LIST_HEAD(tcp_cong_list);
 24
 25/* Simple linear search, don't expect many entries! */
 26struct tcp_congestion_ops *tcp_ca_find(const char *name)
 27{
 28	struct tcp_congestion_ops *e;
 29
 30	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
 31		if (strcmp(e->name, name) == 0)
 32			return e;
 33	}
 34
 35	return NULL;
 36}
 37
 38void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
 39{
 40	struct inet_connection_sock *icsk = inet_csk(sk);
 41
 42	trace_tcp_cong_state_set(sk, ca_state);
 43
 44	if (icsk->icsk_ca_ops->set_state)
 45		icsk->icsk_ca_ops->set_state(sk, ca_state);
 46	icsk->icsk_ca_state = ca_state;
 47}
 48
 49/* Must be called with rcu lock held */
 50static struct tcp_congestion_ops *tcp_ca_find_autoload(const char *name)
 51{
 52	struct tcp_congestion_ops *ca = tcp_ca_find(name);
 53
 54#ifdef CONFIG_MODULES
 55	if (!ca && capable(CAP_NET_ADMIN)) {
 56		rcu_read_unlock();
 57		request_module("tcp_%s", name);
 58		rcu_read_lock();
 59		ca = tcp_ca_find(name);
 60	}
 61#endif
 62	return ca;
 63}
 64
 65/* Simple linear search, not much in here. */
 66struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
 67{
 68	struct tcp_congestion_ops *e;
 69
 70	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
 71		if (e->key == key)
 72			return e;
 73	}
 74
 75	return NULL;
 76}
 77
 78int tcp_validate_congestion_control(struct tcp_congestion_ops *ca)
 79{
 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);
 84		return -EINVAL;
 85	}
 86
 87	return 0;
 88}
 89
 90/* Attach new congestion control algorithm to the list
 91 * of available options.
 92 */
 93int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
 94{
 95	int ret;
 96
 97	ret = tcp_validate_congestion_control(ca);
 98	if (ret)
 99		return ret;
100
101	ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
102
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",
106			  ca->name);
107		ret = -EEXIST;
108	} else {
109		list_add_tail_rcu(&ca->list, &tcp_cong_list);
110		pr_debug("%s registered\n", ca->name);
111	}
112	spin_unlock(&tcp_cong_list_lock);
113
114	return ret;
115}
116EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
117
118/*
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.
123 */
124void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
125{
126	spin_lock(&tcp_cong_list_lock);
127	list_del_rcu(&ca->list);
128	spin_unlock(&tcp_cong_list_lock);
129
130	/* Wait for outstanding readers to complete before the
131	 * module gets removed entirely.
132	 *
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.
136	 */
137	synchronize_rcu();
138}
139EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
140
141/* Replace a registered old ca with a new one.
142 *
143 * The new ca must have the same name as the old one, that has been
144 * registered.
145 */
146int tcp_update_congestion_control(struct tcp_congestion_ops *ca, struct tcp_congestion_ops *old_ca)
147{
148	struct tcp_congestion_ops *existing;
149	int ret = 0;
150
151	ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
152
153	spin_lock(&tcp_cong_list_lock);
154	existing = tcp_ca_find_key(old_ca->key);
155	if (ca->key == TCP_CA_UNSPEC || !existing || strcmp(existing->name, ca->name)) {
156		pr_notice("%s not registered or non-unique key\n",
157			  ca->name);
158		ret = -EINVAL;
159	} else if (existing != old_ca) {
160		pr_notice("invalid old congestion control algorithm to replace\n");
161		ret = -EINVAL;
162	} else {
163		/* Add the new one before removing the old one to keep
164		 * one implementation available all the time.
165		 */
166		list_add_tail_rcu(&ca->list, &tcp_cong_list);
167		list_del_rcu(&existing->list);
168		pr_debug("%s updated\n", ca->name);
169	}
170	spin_unlock(&tcp_cong_list_lock);
171
172	/* Wait for outstanding readers to complete before the
173	 * module or struct_ops gets removed entirely.
174	 */
175	if (!ret)
176		synchronize_rcu();
177
178	return ret;
179}
180
181u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
182{
183	const struct tcp_congestion_ops *ca;
184	u32 key = TCP_CA_UNSPEC;
185
186	might_sleep();
187
188	rcu_read_lock();
189	ca = tcp_ca_find_autoload(name);
190	if (ca) {
191		key = ca->key;
192		*ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
193	}
194	rcu_read_unlock();
195
196	return key;
197}
198
199char *tcp_ca_get_name_by_key(u32 key, char *buffer)
200{
201	const struct tcp_congestion_ops *ca;
202	char *ret = NULL;
203
204	rcu_read_lock();
205	ca = tcp_ca_find_key(key);
206	if (ca) {
207		strscpy(buffer, ca->name, TCP_CA_NAME_MAX);
208		ret = buffer;
209	}
210	rcu_read_unlock();
211
212	return ret;
213}
214
215/* Assign choice of congestion control. */
216void tcp_assign_congestion_control(struct sock *sk)
217{
218	struct net *net = sock_net(sk);
219	struct inet_connection_sock *icsk = inet_csk(sk);
220	const struct tcp_congestion_ops *ca;
221
222	rcu_read_lock();
223	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
224	if (unlikely(!bpf_try_module_get(ca, ca->owner)))
225		ca = &tcp_reno;
226	icsk->icsk_ca_ops = ca;
227	rcu_read_unlock();
228
229	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
230	if (ca->flags & TCP_CONG_NEEDS_ECN)
231		INET_ECN_xmit_ect_1_negotiation(sk);
232	else
233		INET_ECN_dontxmit(sk);
234}
235
236void tcp_init_congestion_control(struct sock *sk)
237{
238	struct inet_connection_sock *icsk = inet_csk(sk);
239
240	tcp_sk(sk)->prior_ssthresh = 0;
241	if (icsk->icsk_ca_ops->init)
242		icsk->icsk_ca_ops->init(sk);
243	if (tcp_ca_needs_ecn(sk))
244		INET_ECN_xmit(sk);
245	else
246		INET_ECN_dontxmit(sk);
247	icsk->icsk_ca_initialized = 1;
248}
249
250static void tcp_reinit_congestion_control(struct sock *sk,
251					  const struct tcp_congestion_ops *ca)
252{
253	struct inet_connection_sock *icsk = inet_csk(sk);
254
255	tcp_cleanup_congestion_control(sk);
256	icsk->icsk_ca_ops = ca;
257	icsk->icsk_ca_setsockopt = 1;
258	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
259
260	if (ca->flags & TCP_CONG_NEEDS_ECN)
261		INET_ECN_xmit_ect_1_negotiation(sk);
262	else
263		INET_ECN_dontxmit(sk);
264
265	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
266		tcp_init_congestion_control(sk);
267}
268
269/* Manage refcounts on socket close. */
270void tcp_cleanup_congestion_control(struct sock *sk)
271{
272	struct inet_connection_sock *icsk = inet_csk(sk);
273
274	if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
275		icsk->icsk_ca_ops->release(sk);
276	icsk->icsk_ca_initialized = 0;
277	bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
278}
279
280/* Used by sysctl to change default congestion control */
281int tcp_set_default_congestion_control(struct net *net, const char *name)
282{
283	struct tcp_congestion_ops *ca;
284	const struct tcp_congestion_ops *prev;
285	int ret;
286
287	rcu_read_lock();
288	ca = tcp_ca_find_autoload(name);
289	if (!ca) {
290		ret = -ENOENT;
291	} else if (!bpf_try_module_get(ca, ca->owner)) {
292		ret = -EBUSY;
293	} else if (!net_eq(net, &init_net) &&
294			!(ca->flags & TCP_CONG_NON_RESTRICTED)) {
295		/* Only init netns can set default to a restricted algorithm */
296		ret = -EPERM;
297	} else {
298		prev = xchg(&net->ipv4.tcp_congestion_control, ca);
299		if (prev)
300			bpf_module_put(prev, prev->owner);
301
302		ca->flags |= TCP_CONG_NON_RESTRICTED;
303		ret = 0;
304	}
305	rcu_read_unlock();
306
307	return ret;
308}
309
310/* Set default value from kernel configuration at bootup */
311static int __init tcp_congestion_default(void)
312{
313	return tcp_set_default_congestion_control(&init_net,
314						  CONFIG_DEFAULT_TCP_CONG);
315}
316late_initcall(tcp_congestion_default);
317
318/* Build string with list of available congestion control values */
319void tcp_get_available_congestion_control(char *buf, size_t maxlen)
320{
321	struct tcp_congestion_ops *ca;
322	size_t offs = 0;
323
324	rcu_read_lock();
325	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
326		offs += snprintf(buf + offs, maxlen - offs,
327				 "%s%s",
328				 offs == 0 ? "" : " ", ca->name);
329
330		if (WARN_ON_ONCE(offs >= maxlen))
331			break;
332	}
333	rcu_read_unlock();
334}
335
336/* Get current default congestion control */
337void tcp_get_default_congestion_control(struct net *net, char *name)
338{
339	const struct tcp_congestion_ops *ca;
340
341	rcu_read_lock();
342	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
343	strscpy(name, ca->name, TCP_CA_NAME_MAX);
344	rcu_read_unlock();
345}
346
347/* Built list of non-restricted congestion control values */
348void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
349{
350	struct tcp_congestion_ops *ca;
351	size_t offs = 0;
352
353	*buf = '\0';
354	rcu_read_lock();
355	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
356		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
357			continue;
358		offs += snprintf(buf + offs, maxlen - offs,
359				 "%s%s",
360				 offs == 0 ? "" : " ", ca->name);
361
362		if (WARN_ON_ONCE(offs >= maxlen))
363			break;
364	}
365	rcu_read_unlock();
366}
367
368/* Change list of non-restricted congestion control */
369int tcp_set_allowed_congestion_control(char *val)
370{
371	struct tcp_congestion_ops *ca;
372	char *saved_clone, *clone, *name;
373	int ret = 0;
374
375	saved_clone = clone = kstrdup(val, GFP_USER);
376	if (!clone)
377		return -ENOMEM;
378
379	spin_lock(&tcp_cong_list_lock);
380	/* pass 1 check for bad entries */
381	while ((name = strsep(&clone, " ")) && *name) {
382		ca = tcp_ca_find(name);
383		if (!ca) {
384			ret = -ENOENT;
385			goto out;
386		}
387	}
388
389	/* pass 2 clear old values */
390	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
391		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
392
393	/* pass 3 mark as allowed */
394	while ((name = strsep(&val, " ")) && *name) {
395		ca = tcp_ca_find(name);
396		WARN_ON(!ca);
397		if (ca)
398			ca->flags |= TCP_CONG_NON_RESTRICTED;
399	}
400out:
401	spin_unlock(&tcp_cong_list_lock);
402	kfree(saved_clone);
403
404	return ret;
405}
406
407/* Change congestion control for socket. If load is false, then it is the
408 * responsibility of the caller to call tcp_init_congestion_control or
409 * tcp_reinit_congestion_control (if the current congestion control was
410 * already initialized.
411 */
412int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
413			       bool cap_net_admin)
414{
415	struct inet_connection_sock *icsk = inet_csk(sk);
416	const struct tcp_congestion_ops *ca;
417	int err = 0;
418
419	if (icsk->icsk_ca_dst_locked)
420		return -EPERM;
421
422	rcu_read_lock();
423	if (!load)
424		ca = tcp_ca_find(name);
425	else
426		ca = tcp_ca_find_autoload(name);
427
428	/* No change asking for existing value */
429	if (ca == icsk->icsk_ca_ops) {
430		icsk->icsk_ca_setsockopt = 1;
431		goto out;
432	}
433
434	if (!ca)
435		err = -ENOENT;
436	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
437		err = -EPERM;
438	else if (!bpf_try_module_get(ca, ca->owner))
439		err = -EBUSY;
440	else
441		tcp_reinit_congestion_control(sk, ca);
442 out:
443	rcu_read_unlock();
444	return err;
445}
446
447/* Slow start is used when congestion window is no greater than the slow start
448 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
449 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
450 * something better;) a packet is only considered (s)acked in its entirety to
451 * defend the ACK attacks described in the RFC. Slow start processes a stretch
452 * ACK of degree N as if N acks of degree 1 are received back to back except
453 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
454 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
455 */
456__bpf_kfunc u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
457{
458	u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
459
460	acked -= cwnd - tcp_snd_cwnd(tp);
461	tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
462
463	return acked;
464}
465EXPORT_SYMBOL_GPL(tcp_slow_start);
466
467/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
468 * for every packet that was ACKed.
469 */
470__bpf_kfunc void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
471{
472	/* If credits accumulated at a higher w, apply them gently now. */
473	if (tp->snd_cwnd_cnt >= w) {
474		tp->snd_cwnd_cnt = 0;
475		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
476	}
477
478	tp->snd_cwnd_cnt += acked;
479	if (tp->snd_cwnd_cnt >= w) {
480		u32 delta = tp->snd_cwnd_cnt / w;
481
482		tp->snd_cwnd_cnt -= delta * w;
483		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
484	}
485	tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
486}
487EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
488
489/*
490 * TCP Reno congestion control
491 * This is special case used for fallback as well.
492 */
493/* This is Jacobson's slow start and congestion avoidance.
494 * SIGCOMM '88, p. 328.
495 */
496__bpf_kfunc void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
497{
498	struct tcp_sock *tp = tcp_sk(sk);
499
500	if (!tcp_is_cwnd_limited(sk))
501		return;
502
503	/* In "safe" area, increase. */
504	if (tcp_in_slow_start(tp)) {
505		acked = tcp_slow_start(tp, acked);
506		if (!acked)
507			return;
508	}
509	/* In dangerous area, increase slowly. */
510	tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
511}
512EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
513
514/* Slow start threshold is half the congestion window (min 2) */
515__bpf_kfunc u32 tcp_reno_ssthresh(struct sock *sk)
516{
517	const struct tcp_sock *tp = tcp_sk(sk);
518
519	return max(tcp_snd_cwnd(tp) >> 1U, 2U);
520}
521EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
522
523__bpf_kfunc u32 tcp_reno_undo_cwnd(struct sock *sk)
524{
525	const struct tcp_sock *tp = tcp_sk(sk);
526
527	return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
528}
529EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
530
531struct tcp_congestion_ops tcp_reno = {
532	.flags		= TCP_CONG_NON_RESTRICTED,
533	.name		= "reno",
534	.owner		= THIS_MODULE,
535	.ssthresh	= tcp_reno_ssthresh,
536	.cong_avoid	= tcp_reno_cong_avoid,
537	.undo_cwnd	= tcp_reno_undo_cwnd,
538};