Linux Audio

Check our new training course

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