Linux Audio

Check our new training course

Loading...
v5.14.15
  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
 
 
 20static DEFINE_SPINLOCK(tcp_cong_list_lock);
 21static LIST_HEAD(tcp_cong_list);
 22
 23/* Simple linear search, don't expect many entries! */
 24struct tcp_congestion_ops *tcp_ca_find(const char *name)
 25{
 26	struct tcp_congestion_ops *e;
 27
 28	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
 29		if (strcmp(e->name, name) == 0)
 30			return e;
 31	}
 32
 33	return NULL;
 34}
 35
 36/* Must be called with rcu lock held */
 37static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
 38						       const char *name)
 39{
 40	struct tcp_congestion_ops *ca = tcp_ca_find(name);
 41
 42#ifdef CONFIG_MODULES
 43	if (!ca && capable(CAP_NET_ADMIN)) {
 44		rcu_read_unlock();
 45		request_module("tcp_%s", name);
 46		rcu_read_lock();
 47		ca = tcp_ca_find(name);
 48	}
 49#endif
 50	return ca;
 51}
 52
 53/* Simple linear search, not much in here. */
 54struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
 55{
 56	struct tcp_congestion_ops *e;
 57
 58	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
 59		if (e->key == key)
 60			return e;
 61	}
 62
 63	return NULL;
 64}
 65
 66/*
 67 * Attach new congestion control algorithm to the list
 68 * of available options.
 69 */
 70int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
 71{
 72	int ret = 0;
 73
 74	/* all algorithms must implement these */
 75	if (!ca->ssthresh || !ca->undo_cwnd ||
 76	    !(ca->cong_avoid || ca->cong_control)) {
 77		pr_err("%s does not implement required ops\n", ca->name);
 78		return -EINVAL;
 79	}
 80
 81	ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
 82
 83	spin_lock(&tcp_cong_list_lock);
 84	if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
 85		pr_notice("%s already registered or non-unique key\n",
 86			  ca->name);
 87		ret = -EEXIST;
 88	} else {
 89		list_add_tail_rcu(&ca->list, &tcp_cong_list);
 90		pr_debug("%s registered\n", ca->name);
 91	}
 92	spin_unlock(&tcp_cong_list_lock);
 93
 94	return ret;
 95}
 96EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
 97
 98/*
 99 * Remove congestion control algorithm, called from
100 * the module's remove function.  Module ref counts are used
101 * to ensure that this can't be done till all sockets using
102 * that method are closed.
103 */
104void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
105{
106	spin_lock(&tcp_cong_list_lock);
107	list_del_rcu(&ca->list);
108	spin_unlock(&tcp_cong_list_lock);
109
110	/* Wait for outstanding readers to complete before the
111	 * module gets removed entirely.
112	 *
113	 * A try_module_get() should fail by now as our module is
114	 * in "going" state since no refs are held anymore and
115	 * module_exit() handler being called.
116	 */
117	synchronize_rcu();
118}
119EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120
121u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
122{
123	const struct tcp_congestion_ops *ca;
124	u32 key = TCP_CA_UNSPEC;
125
126	might_sleep();
127
128	rcu_read_lock();
129	ca = tcp_ca_find_autoload(net, name);
130	if (ca) {
131		key = ca->key;
132		*ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
133	}
134	rcu_read_unlock();
135
136	return key;
137}
138EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139
140char *tcp_ca_get_name_by_key(u32 key, char *buffer)
141{
142	const struct tcp_congestion_ops *ca;
143	char *ret = NULL;
144
145	rcu_read_lock();
146	ca = tcp_ca_find_key(key);
147	if (ca)
148		ret = strncpy(buffer, ca->name,
149			      TCP_CA_NAME_MAX);
150	rcu_read_unlock();
151
152	return ret;
153}
154EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155
156/* Assign choice of congestion control. */
157void tcp_assign_congestion_control(struct sock *sk)
158{
159	struct net *net = sock_net(sk);
160	struct inet_connection_sock *icsk = inet_csk(sk);
161	const struct tcp_congestion_ops *ca;
162
163	rcu_read_lock();
164	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
165	if (unlikely(!bpf_try_module_get(ca, ca->owner)))
166		ca = &tcp_reno;
167	icsk->icsk_ca_ops = ca;
168	rcu_read_unlock();
169
170	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
171	if (ca->flags & TCP_CONG_NEEDS_ECN)
172		INET_ECN_xmit(sk);
173	else
174		INET_ECN_dontxmit(sk);
175}
 
 
176
177void tcp_init_congestion_control(struct sock *sk)
178{
179	struct inet_connection_sock *icsk = inet_csk(sk);
 
180
181	tcp_sk(sk)->prior_ssthresh = 0;
182	if (icsk->icsk_ca_ops->init)
183		icsk->icsk_ca_ops->init(sk);
184	if (tcp_ca_needs_ecn(sk))
185		INET_ECN_xmit(sk);
186	else
187		INET_ECN_dontxmit(sk);
188	icsk->icsk_ca_initialized = 1;
189}
190
191static void tcp_reinit_congestion_control(struct sock *sk,
192					  const struct tcp_congestion_ops *ca)
193{
194	struct inet_connection_sock *icsk = inet_csk(sk);
195
196	tcp_cleanup_congestion_control(sk);
197	icsk->icsk_ca_ops = ca;
198	icsk->icsk_ca_setsockopt = 1;
199	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
200
201	if (ca->flags & TCP_CONG_NEEDS_ECN)
202		INET_ECN_xmit(sk);
203	else
204		INET_ECN_dontxmit(sk);
205
206	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
207		tcp_init_congestion_control(sk);
208}
209
210/* Manage refcounts on socket close. */
211void tcp_cleanup_congestion_control(struct sock *sk)
212{
213	struct inet_connection_sock *icsk = inet_csk(sk);
214
215	if (icsk->icsk_ca_ops->release)
216		icsk->icsk_ca_ops->release(sk);
217	bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
218}
219
220/* Used by sysctl to change default congestion control */
221int tcp_set_default_congestion_control(struct net *net, const char *name)
222{
223	struct tcp_congestion_ops *ca;
224	const struct tcp_congestion_ops *prev;
225	int ret;
226
227	rcu_read_lock();
228	ca = tcp_ca_find_autoload(net, name);
229	if (!ca) {
230		ret = -ENOENT;
231	} else if (!bpf_try_module_get(ca, ca->owner)) {
232		ret = -EBUSY;
233	} else if (!net_eq(net, &init_net) &&
234			!(ca->flags & TCP_CONG_NON_RESTRICTED)) {
235		/* Only init netns can set default to a restricted algorithm */
236		ret = -EPERM;
237	} else {
238		prev = xchg(&net->ipv4.tcp_congestion_control, ca);
239		if (prev)
240			bpf_module_put(prev, prev->owner);
241
242		ca->flags |= TCP_CONG_NON_RESTRICTED;
 
 
 
 
 
 
 
 
243		ret = 0;
244	}
245	rcu_read_unlock();
246
247	return ret;
248}
249
250/* Set default value from kernel configuration at bootup */
251static int __init tcp_congestion_default(void)
252{
253	return tcp_set_default_congestion_control(&init_net,
254						  CONFIG_DEFAULT_TCP_CONG);
255}
256late_initcall(tcp_congestion_default);
257
 
258/* Build string with list of available congestion control values */
259void tcp_get_available_congestion_control(char *buf, size_t maxlen)
260{
261	struct tcp_congestion_ops *ca;
262	size_t offs = 0;
263
264	rcu_read_lock();
265	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
266		offs += snprintf(buf + offs, maxlen - offs,
267				 "%s%s",
268				 offs == 0 ? "" : " ", ca->name);
269
270		if (WARN_ON_ONCE(offs >= maxlen))
271			break;
272	}
273	rcu_read_unlock();
274}
275
276/* Get current default congestion control */
277void tcp_get_default_congestion_control(struct net *net, char *name)
278{
279	const struct tcp_congestion_ops *ca;
 
 
280
281	rcu_read_lock();
282	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
283	strncpy(name, ca->name, TCP_CA_NAME_MAX);
284	rcu_read_unlock();
285}
286
287/* Built list of non-restricted congestion control values */
288void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
289{
290	struct tcp_congestion_ops *ca;
291	size_t offs = 0;
292
293	*buf = '\0';
294	rcu_read_lock();
295	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
296		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
297			continue;
298		offs += snprintf(buf + offs, maxlen - offs,
299				 "%s%s",
300				 offs == 0 ? "" : " ", ca->name);
301
302		if (WARN_ON_ONCE(offs >= maxlen))
303			break;
304	}
305	rcu_read_unlock();
306}
307
308/* Change list of non-restricted congestion control */
309int tcp_set_allowed_congestion_control(char *val)
310{
311	struct tcp_congestion_ops *ca;
312	char *saved_clone, *clone, *name;
313	int ret = 0;
314
315	saved_clone = clone = kstrdup(val, GFP_USER);
316	if (!clone)
317		return -ENOMEM;
318
319	spin_lock(&tcp_cong_list_lock);
320	/* pass 1 check for bad entries */
321	while ((name = strsep(&clone, " ")) && *name) {
322		ca = tcp_ca_find(name);
323		if (!ca) {
324			ret = -ENOENT;
325			goto out;
326		}
327	}
328
329	/* pass 2 clear old values */
330	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
331		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
332
333	/* pass 3 mark as allowed */
334	while ((name = strsep(&val, " ")) && *name) {
335		ca = tcp_ca_find(name);
336		WARN_ON(!ca);
337		if (ca)
338			ca->flags |= TCP_CONG_NON_RESTRICTED;
339	}
340out:
341	spin_unlock(&tcp_cong_list_lock);
342	kfree(saved_clone);
343
344	return ret;
345}
346
347/* Change congestion control for socket. If load is false, then it is the
348 * responsibility of the caller to call tcp_init_congestion_control or
349 * tcp_reinit_congestion_control (if the current congestion control was
350 * already initialized.
351 */
352int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
353			       bool cap_net_admin)
354{
355	struct inet_connection_sock *icsk = inet_csk(sk);
356	const struct tcp_congestion_ops *ca;
357	int err = 0;
358
359	if (icsk->icsk_ca_dst_locked)
360		return -EPERM;
361
362	rcu_read_lock();
363	if (!load)
364		ca = tcp_ca_find(name);
365	else
366		ca = tcp_ca_find_autoload(sock_net(sk), name);
367
368	/* No change asking for existing value */
369	if (ca == icsk->icsk_ca_ops) {
370		icsk->icsk_ca_setsockopt = 1;
371		goto out;
372	}
373
 
 
 
 
 
 
 
 
 
374	if (!ca)
375		err = -ENOENT;
376	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
 
377		err = -EPERM;
378	else if (!bpf_try_module_get(ca, ca->owner))
 
379		err = -EBUSY;
380	else
381		tcp_reinit_congestion_control(sk, ca);
 
 
 
 
 
 
382 out:
383	rcu_read_unlock();
384	return err;
385}
386
387/* Slow start is used when congestion window is no greater than the slow start
388 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
389 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
390 * something better;) a packet is only considered (s)acked in its entirety to
391 * defend the ACK attacks described in the RFC. Slow start processes a stretch
392 * ACK of degree N as if N acks of degree 1 are received back to back except
393 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
394 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
395 */
396u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
397{
398	u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
 
399
400	acked -= cwnd - tp->snd_cwnd;
401	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
402
403	return acked;
 
 
 
 
 
 
404}
405EXPORT_SYMBOL_GPL(tcp_slow_start);
406
407/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
408 * for every packet that was ACKed.
 
 
 
 
409 */
410void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
411{
412	/* If credits accumulated at a higher w, apply them gently now. */
413	if (tp->snd_cwnd_cnt >= w) {
414		tp->snd_cwnd_cnt = 0;
415		tp->snd_cwnd++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
416	}
 
 
417
418	tp->snd_cwnd_cnt += acked;
 
 
419	if (tp->snd_cwnd_cnt >= w) {
420		u32 delta = tp->snd_cwnd_cnt / w;
421
422		tp->snd_cwnd_cnt -= delta * w;
423		tp->snd_cwnd += delta;
 
424	}
425	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
426}
427EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
428
429/*
430 * TCP Reno congestion control
431 * This is special case used for fallback as well.
432 */
433/* This is Jacobson's slow start and congestion avoidance.
434 * SIGCOMM '88, p. 328.
435 */
436void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
437{
438	struct tcp_sock *tp = tcp_sk(sk);
439
440	if (!tcp_is_cwnd_limited(sk))
441		return;
442
443	/* In "safe" area, increase. */
444	if (tcp_in_slow_start(tp)) {
445		acked = tcp_slow_start(tp, acked);
446		if (!acked)
447			return;
448	}
449	/* In dangerous area, increase slowly. */
450	tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
 
 
 
 
 
 
 
 
 
 
 
451}
452EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
453
454/* Slow start threshold is half the congestion window (min 2) */
455u32 tcp_reno_ssthresh(struct sock *sk)
456{
457	const struct tcp_sock *tp = tcp_sk(sk);
458
459	return max(tp->snd_cwnd >> 1U, 2U);
460}
461EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
462
463u32 tcp_reno_undo_cwnd(struct sock *sk)
 
464{
465	const struct tcp_sock *tp = tcp_sk(sk);
466
467	return max(tp->snd_cwnd, tp->prior_cwnd);
468}
469EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
470
471struct tcp_congestion_ops tcp_reno = {
472	.flags		= TCP_CONG_NON_RESTRICTED,
473	.name		= "reno",
474	.owner		= THIS_MODULE,
475	.ssthresh	= tcp_reno_ssthresh,
476	.cong_avoid	= tcp_reno_cong_avoid,
477	.undo_cwnd	= tcp_reno_undo_cwnd,
 
 
 
 
 
 
 
 
 
 
 
 
478};
v3.5.6
 
  1/*
  2 * Plugable TCP congestion control support and newReno
  3 * congestion control.
  4 * Based on ideas from I/O scheduler suport and Web100.
  5 *
  6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
  7 */
  8
  9#define pr_fmt(fmt) "TCP: " fmt
 10
 11#include <linux/module.h>
 12#include <linux/mm.h>
 13#include <linux/types.h>
 14#include <linux/list.h>
 15#include <linux/gfp.h>
 
 16#include <net/tcp.h>
 17
 18int sysctl_tcp_max_ssthresh = 0;
 19
 20static DEFINE_SPINLOCK(tcp_cong_list_lock);
 21static LIST_HEAD(tcp_cong_list);
 22
 23/* Simple linear search, don't expect many entries! */
 24static struct tcp_congestion_ops *tcp_ca_find(const char *name)
 25{
 26	struct tcp_congestion_ops *e;
 27
 28	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
 29		if (strcmp(e->name, name) == 0)
 30			return e;
 31	}
 32
 33	return NULL;
 34}
 35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 36/*
 37 * Attach new congestion control algorithm to the list
 38 * of available options.
 39 */
 40int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
 41{
 42	int ret = 0;
 43
 44	/* all algorithms must implement ssthresh and cong_avoid ops */
 45	if (!ca->ssthresh || !ca->cong_avoid) {
 
 46		pr_err("%s does not implement required ops\n", ca->name);
 47		return -EINVAL;
 48	}
 49
 
 
 50	spin_lock(&tcp_cong_list_lock);
 51	if (tcp_ca_find(ca->name)) {
 52		pr_notice("%s already registered\n", ca->name);
 
 53		ret = -EEXIST;
 54	} else {
 55		list_add_tail_rcu(&ca->list, &tcp_cong_list);
 56		pr_info("%s registered\n", ca->name);
 57	}
 58	spin_unlock(&tcp_cong_list_lock);
 59
 60	return ret;
 61}
 62EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
 63
 64/*
 65 * Remove congestion control algorithm, called from
 66 * the module's remove function.  Module ref counts are used
 67 * to ensure that this can't be done till all sockets using
 68 * that method are closed.
 69 */
 70void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
 71{
 72	spin_lock(&tcp_cong_list_lock);
 73	list_del_rcu(&ca->list);
 74	spin_unlock(&tcp_cong_list_lock);
 
 
 
 
 
 
 
 
 
 75}
 76EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
 77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 78/* Assign choice of congestion control. */
 79void tcp_init_congestion_control(struct sock *sk)
 80{
 
 81	struct inet_connection_sock *icsk = inet_csk(sk);
 82	struct tcp_congestion_ops *ca;
 
 
 
 
 
 
 
 83
 84	/* if no choice made yet assign the current value set as default */
 85	if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
 86		rcu_read_lock();
 87		list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 88			if (try_module_get(ca->owner)) {
 89				icsk->icsk_ca_ops = ca;
 90				break;
 91			}
 92
 93			/* fallback to next available */
 94		}
 95		rcu_read_unlock();
 96	}
 97
 
 98	if (icsk->icsk_ca_ops->init)
 99		icsk->icsk_ca_ops->init(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100}
101
102/* Manage refcounts on socket close. */
103void tcp_cleanup_congestion_control(struct sock *sk)
104{
105	struct inet_connection_sock *icsk = inet_csk(sk);
106
107	if (icsk->icsk_ca_ops->release)
108		icsk->icsk_ca_ops->release(sk);
109	module_put(icsk->icsk_ca_ops->owner);
110}
111
112/* Used by sysctl to change default congestion control */
113int tcp_set_default_congestion_control(const char *name)
114{
115	struct tcp_congestion_ops *ca;
116	int ret = -ENOENT;
 
117
118	spin_lock(&tcp_cong_list_lock);
119	ca = tcp_ca_find(name);
120#ifdef CONFIG_MODULES
121	if (!ca && capable(CAP_NET_ADMIN)) {
122		spin_unlock(&tcp_cong_list_lock);
 
 
 
 
 
 
 
 
 
123
124		request_module("tcp_%s", name);
125		spin_lock(&tcp_cong_list_lock);
126		ca = tcp_ca_find(name);
127	}
128#endif
129
130	if (ca) {
131		ca->flags |= TCP_CONG_NON_RESTRICTED;	/* default is always allowed */
132		list_move(&ca->list, &tcp_cong_list);
133		ret = 0;
134	}
135	spin_unlock(&tcp_cong_list_lock);
136
137	return ret;
138}
139
140/* Set default value from kernel configuration at bootup */
141static int __init tcp_congestion_default(void)
142{
143	return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 
144}
145late_initcall(tcp_congestion_default);
146
147
148/* Build string with list of available congestion control values */
149void tcp_get_available_congestion_control(char *buf, size_t maxlen)
150{
151	struct tcp_congestion_ops *ca;
152	size_t offs = 0;
153
154	rcu_read_lock();
155	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
156		offs += snprintf(buf + offs, maxlen - offs,
157				 "%s%s",
158				 offs == 0 ? "" : " ", ca->name);
159
 
 
160	}
161	rcu_read_unlock();
162}
163
164/* Get current default congestion control */
165void tcp_get_default_congestion_control(char *name)
166{
167	struct tcp_congestion_ops *ca;
168	/* We will always have reno... */
169	BUG_ON(list_empty(&tcp_cong_list));
170
171	rcu_read_lock();
172	ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
173	strncpy(name, ca->name, TCP_CA_NAME_MAX);
174	rcu_read_unlock();
175}
176
177/* Built list of non-restricted congestion control values */
178void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
179{
180	struct tcp_congestion_ops *ca;
181	size_t offs = 0;
182
183	*buf = '\0';
184	rcu_read_lock();
185	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
186		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
187			continue;
188		offs += snprintf(buf + offs, maxlen - offs,
189				 "%s%s",
190				 offs == 0 ? "" : " ", ca->name);
191
 
 
192	}
193	rcu_read_unlock();
194}
195
196/* Change list of non-restricted congestion control */
197int tcp_set_allowed_congestion_control(char *val)
198{
199	struct tcp_congestion_ops *ca;
200	char *saved_clone, *clone, *name;
201	int ret = 0;
202
203	saved_clone = clone = kstrdup(val, GFP_USER);
204	if (!clone)
205		return -ENOMEM;
206
207	spin_lock(&tcp_cong_list_lock);
208	/* pass 1 check for bad entries */
209	while ((name = strsep(&clone, " ")) && *name) {
210		ca = tcp_ca_find(name);
211		if (!ca) {
212			ret = -ENOENT;
213			goto out;
214		}
215	}
216
217	/* pass 2 clear old values */
218	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
219		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
220
221	/* pass 3 mark as allowed */
222	while ((name = strsep(&val, " ")) && *name) {
223		ca = tcp_ca_find(name);
224		WARN_ON(!ca);
225		if (ca)
226			ca->flags |= TCP_CONG_NON_RESTRICTED;
227	}
228out:
229	spin_unlock(&tcp_cong_list_lock);
230	kfree(saved_clone);
231
232	return ret;
233}
234
235
236/* Change congestion control for socket */
237int tcp_set_congestion_control(struct sock *sk, const char *name)
 
 
 
 
238{
239	struct inet_connection_sock *icsk = inet_csk(sk);
240	struct tcp_congestion_ops *ca;
241	int err = 0;
242
 
 
 
243	rcu_read_lock();
244	ca = tcp_ca_find(name);
 
 
 
245
246	/* no change asking for existing value */
247	if (ca == icsk->icsk_ca_ops)
 
248		goto out;
 
249
250#ifdef CONFIG_MODULES
251	/* not found attempt to autoload module */
252	if (!ca && capable(CAP_NET_ADMIN)) {
253		rcu_read_unlock();
254		request_module("tcp_%s", name);
255		rcu_read_lock();
256		ca = tcp_ca_find(name);
257	}
258#endif
259	if (!ca)
260		err = -ENOENT;
261
262	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
263		err = -EPERM;
264
265	else if (!try_module_get(ca->owner))
266		err = -EBUSY;
267
268	else {
269		tcp_cleanup_congestion_control(sk);
270		icsk->icsk_ca_ops = ca;
271
272		if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
273			icsk->icsk_ca_ops->init(sk);
274	}
275 out:
276	rcu_read_unlock();
277	return err;
278}
279
280/* RFC2861 Check whether we are limited by application or congestion window
281 * This is the inverse of cwnd check in tcp_tso_should_defer
 
 
 
 
 
 
282 */
283bool tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
284{
285	const struct tcp_sock *tp = tcp_sk(sk);
286	u32 left;
287
288	if (in_flight >= tp->snd_cwnd)
289		return true;
290
291	left = tp->snd_cwnd - in_flight;
292	if (sk_can_gso(sk) &&
293	    left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
294	    left * tp->mss_cache < sk->sk_gso_max_size &&
295	    left < sk->sk_gso_max_segs)
296		return true;
297	return left <= tcp_max_tso_deferred_mss(tp);
298}
299EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
300
301/*
302 * Slow start is used when congestion window is less than slow start
303 * threshold. This version implements the basic RFC2581 version
304 * and optionally supports:
305 * 	RFC3742 Limited Slow Start  	  - growth limited to max_ssthresh
306 *	RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
307 */
308void tcp_slow_start(struct tcp_sock *tp)
309{
310	int cnt; /* increase in packets */
311
312	/* RFC3465: ABC Slow start
313	 * Increase only after a full MSS of bytes is acked
314	 *
315	 * TCP sender SHOULD increase cwnd by the number of
316	 * previously unacknowledged bytes ACKed by each incoming
317	 * acknowledgment, provided the increase is not more than L
318	 */
319	if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
320		return;
321
322	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
323		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
324	else
325		cnt = tp->snd_cwnd;			/* exponential increase */
326
327	/* RFC3465: ABC
328	 * We MAY increase by 2 if discovered delayed ack
329	 */
330	if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
331		cnt <<= 1;
332	tp->bytes_acked = 0;
333
334	tp->snd_cwnd_cnt += cnt;
335	while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
336		tp->snd_cwnd_cnt -= tp->snd_cwnd;
337		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
338			tp->snd_cwnd++;
339	}
340}
341EXPORT_SYMBOL_GPL(tcp_slow_start);
342
343/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
344void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
345{
346	if (tp->snd_cwnd_cnt >= w) {
347		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
348			tp->snd_cwnd++;
349		tp->snd_cwnd_cnt = 0;
350	} else {
351		tp->snd_cwnd_cnt++;
352	}
 
353}
354EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
355
356/*
357 * TCP Reno congestion control
358 * This is special case used for fallback as well.
359 */
360/* This is Jacobson's slow start and congestion avoidance.
361 * SIGCOMM '88, p. 328.
362 */
363void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
364{
365	struct tcp_sock *tp = tcp_sk(sk);
366
367	if (!tcp_is_cwnd_limited(sk, in_flight))
368		return;
369
370	/* In "safe" area, increase. */
371	if (tp->snd_cwnd <= tp->snd_ssthresh)
372		tcp_slow_start(tp);
373
 
 
374	/* In dangerous area, increase slowly. */
375	else if (sysctl_tcp_abc) {
376		/* RFC3465: Appropriate Byte Count
377		 * increase once for each full cwnd acked
378		 */
379		if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
380			tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
381			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
382				tp->snd_cwnd++;
383		}
384	} else {
385		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
386	}
387}
388EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
389
390/* Slow start threshold is half the congestion window (min 2) */
391u32 tcp_reno_ssthresh(struct sock *sk)
392{
393	const struct tcp_sock *tp = tcp_sk(sk);
 
394	return max(tp->snd_cwnd >> 1U, 2U);
395}
396EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
397
398/* Lower bound on congestion window with halving. */
399u32 tcp_reno_min_cwnd(const struct sock *sk)
400{
401	const struct tcp_sock *tp = tcp_sk(sk);
402	return tp->snd_ssthresh/2;
 
403}
404EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
405
406struct tcp_congestion_ops tcp_reno = {
407	.flags		= TCP_CONG_NON_RESTRICTED,
408	.name		= "reno",
409	.owner		= THIS_MODULE,
410	.ssthresh	= tcp_reno_ssthresh,
411	.cong_avoid	= tcp_reno_cong_avoid,
412	.min_cwnd	= tcp_reno_min_cwnd,
413};
414
415/* Initial congestion control used (until SYN)
416 * really reno under another name so we can tell difference
417 * during tcp_set_default_congestion_control
418 */
419struct tcp_congestion_ops tcp_init_congestion_ops  = {
420	.name		= "",
421	.owner		= THIS_MODULE,
422	.ssthresh	= tcp_reno_ssthresh,
423	.cong_avoid	= tcp_reno_cong_avoid,
424	.min_cwnd	= tcp_reno_min_cwnd,
425};
426EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);