tcp_cong.c revision 246eb2af060fc32650f07203c02bdc0456ad76c7 
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 <net/tcp.h>
14
15int sysctl_tcp_max_ssthresh = 0;
16
17static DEFINE_SPINLOCK(tcp_cong_list_lock);
18static LIST_HEAD(tcp_cong_list);
19
20/* Simple linear search, don't expect many entries! */
21static struct tcp_congestion_ops *tcp_ca_find(const char *name)
22{
23	struct tcp_congestion_ops *e;
24
25	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
26		if (strcmp(e->name, name) == 0)
27			return e;
28	}
29
30	return NULL;
31}
32
33/*
34 * Attach new congestion control algorithm to the list
35 * of available options.
36 */
37int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
38{
39	int ret = 0;
40
41	/* all algorithms must implement ssthresh and cong_avoid ops */
42	if (!ca->ssthresh || !ca->cong_avoid) {
43		printk(KERN_ERR "TCP %s does not implement required ops\n",
44		       ca->name);
45		return -EINVAL;
46	}
47
48	spin_lock(&tcp_cong_list_lock);
49	if (tcp_ca_find(ca->name)) {
50		printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
51		ret = -EEXIST;
52	} else {
53		list_add_tail_rcu(&ca->list, &tcp_cong_list);
54		printk(KERN_INFO "TCP %s registered\n", ca->name);
55	}
56	spin_unlock(&tcp_cong_list_lock);
57
58	return ret;
59}
60EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
61
62/*
63 * Remove congestion control algorithm, called from
64 * the module's remove function.  Module ref counts are used
65 * to ensure that this can't be done till all sockets using
66 * that method are closed.
67 */
68void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
69{
70	spin_lock(&tcp_cong_list_lock);
71	list_del_rcu(&ca->list);
72	spin_unlock(&tcp_cong_list_lock);
73}
74EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
75
76/* Assign choice of congestion control. */
77void tcp_init_congestion_control(struct sock *sk)
78{
79	struct inet_connection_sock *icsk = inet_csk(sk);
80	struct tcp_congestion_ops *ca;
81
82	/* if no choice made yet assign the current value set as default */
83	if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
84		rcu_read_lock();
85		list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
86			if (try_module_get(ca->owner)) {
87				icsk->icsk_ca_ops = ca;
88				break;
89			}
90
91			/* fallback to next available */
92		}
93		rcu_read_unlock();
94	}
95
96	if (icsk->icsk_ca_ops->init)
97		icsk->icsk_ca_ops->init(sk);
98}
99
100/* Manage refcounts on socket close. */
101void tcp_cleanup_congestion_control(struct sock *sk)
102{
103	struct inet_connection_sock *icsk = inet_csk(sk);
104
105	if (icsk->icsk_ca_ops->release)
106		icsk->icsk_ca_ops->release(sk);
107	module_put(icsk->icsk_ca_ops->owner);
108}
109
110/* Used by sysctl to change default congestion control */
111int tcp_set_default_congestion_control(const char *name)
112{
113	struct tcp_congestion_ops *ca;
114	int ret = -ENOENT;
115
116	spin_lock(&tcp_cong_list_lock);
117	ca = tcp_ca_find(name);
118#ifdef CONFIG_KMOD
119	if (!ca && capable(CAP_SYS_MODULE)) {
120		spin_unlock(&tcp_cong_list_lock);
121
122		request_module("tcp_%s", name);
123		spin_lock(&tcp_cong_list_lock);
124		ca = tcp_ca_find(name);
125	}
126#endif
127
128	if (ca) {
129		ca->flags |= TCP_CONG_NON_RESTRICTED;	/* default is always allowed */
130		list_move(&ca->list, &tcp_cong_list);
131		ret = 0;
132	}
133	spin_unlock(&tcp_cong_list_lock);
134
135	return ret;
136}
137
138/* Set default value from kernel configuration at bootup */
139static int __init tcp_congestion_default(void)
140{
141	return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
142}
143late_initcall(tcp_congestion_default);
144
145
146/* Build string with list of available congestion control values */
147void tcp_get_available_congestion_control(char *buf, size_t maxlen)
148{
149	struct tcp_congestion_ops *ca;
150	size_t offs = 0;
151
152	rcu_read_lock();
153	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
154		offs += snprintf(buf + offs, maxlen - offs,
155				 "%s%s",
156				 offs == 0 ? "" : " ", ca->name);
157
158	}
159	rcu_read_unlock();
160}
161
162/* Get current default congestion control */
163void tcp_get_default_congestion_control(char *name)
164{
165	struct tcp_congestion_ops *ca;
166	/* We will always have reno... */
167	BUG_ON(list_empty(&tcp_cong_list));
168
169	rcu_read_lock();
170	ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
171	strncpy(name, ca->name, TCP_CA_NAME_MAX);
172	rcu_read_unlock();
173}
174
175/* Built list of non-restricted congestion control values */
176void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
177{
178	struct tcp_congestion_ops *ca;
179	size_t offs = 0;
180
181	*buf = '\0';
182	rcu_read_lock();
183	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
184		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
185			continue;
186		offs += snprintf(buf + offs, maxlen - offs,
187				 "%s%s",
188				 offs == 0 ? "" : " ", ca->name);
189
190	}
191	rcu_read_unlock();
192}
193
194/* Change list of non-restricted congestion control */
195int tcp_set_allowed_congestion_control(char *val)
196{
197	struct tcp_congestion_ops *ca;
198	char *clone, *name;
199	int ret = 0;
200
201	clone = kstrdup(val, GFP_USER);
202	if (!clone)
203		return -ENOMEM;
204
205	spin_lock(&tcp_cong_list_lock);
206	/* pass 1 check for bad entries */
207	while ((name = strsep(&clone, " ")) && *name) {
208		ca = tcp_ca_find(name);
209		if (!ca) {
210			ret = -ENOENT;
211			goto out;
212		}
213	}
214
215	/* pass 2 clear old values */
216	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
217		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
218
219	/* pass 3 mark as allowed */
220	while ((name = strsep(&val, " ")) && *name) {
221		ca = tcp_ca_find(name);
222		WARN_ON(!ca);
223		if (ca)
224			ca->flags |= TCP_CONG_NON_RESTRICTED;
225	}
226out:
227	spin_unlock(&tcp_cong_list_lock);
228
229	return ret;
230}
231
232
233/* Change congestion control for socket */
234int tcp_set_congestion_control(struct sock *sk, const char *name)
235{
236	struct inet_connection_sock *icsk = inet_csk(sk);
237	struct tcp_congestion_ops *ca;
238	int err = 0;
239
240	rcu_read_lock();
241	ca = tcp_ca_find(name);
242
243	/* no change asking for existing value */
244	if (ca == icsk->icsk_ca_ops)
245		goto out;
246
247#ifdef CONFIG_KMOD
248	/* not found attempt to autoload module */
249	if (!ca && capable(CAP_SYS_MODULE)) {
250		rcu_read_unlock();
251		request_module("tcp_%s", name);
252		rcu_read_lock();
253		ca = tcp_ca_find(name);
254	}
255#endif
256	if (!ca)
257		err = -ENOENT;
258
259	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
260		err = -EPERM;
261
262	else if (!try_module_get(ca->owner))
263		err = -EBUSY;
264
265	else {
266		tcp_cleanup_congestion_control(sk);
267		icsk->icsk_ca_ops = ca;
268
269		if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
270			icsk->icsk_ca_ops->init(sk);
271	}
272 out:
273	rcu_read_unlock();
274	return err;
275}
276
277/* RFC2861 Check whether we are limited by application or congestion window
278 * This is the inverse of cwnd check in tcp_tso_should_defer
279 */
280int tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
281{
282	const struct tcp_sock *tp = tcp_sk(sk);
283	u32 left;
284
285	if (in_flight >= tp->snd_cwnd)
286		return 1;
287
288	left = tp->snd_cwnd - in_flight;
289	if (sk_can_gso(sk) &&
290	    left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
291	    left * tp->mss_cache < sk->sk_gso_max_size)
292		return 1;
293	return left <= tcp_max_burst(tp);
294}
295EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
296
297/*
298 * Slow start is used when congestion window is less than slow start
299 * threshold. This version implements the basic RFC2581 version
300 * and optionally supports:
301 * 	RFC3742 Limited Slow Start  	  - growth limited to max_ssthresh
302 *	RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
303 */
304void tcp_slow_start(struct tcp_sock *tp)
305{
306	int cnt; /* increase in packets */
307
308	/* RFC3465: ABC Slow start
309	 * Increase only after a full MSS of bytes is acked
310	 *
311	 * TCP sender SHOULD increase cwnd by the number of
312	 * previously unacknowledged bytes ACKed by each incoming
313	 * acknowledgment, provided the increase is not more than L
314	 */
315	if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
316		return;
317
318	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
319		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
320	else
321		cnt = tp->snd_cwnd;			/* exponential increase */
322
323	/* RFC3465: ABC
324	 * We MAY increase by 2 if discovered delayed ack
325	 */
326	if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
327		cnt <<= 1;
328	tp->bytes_acked = 0;
329
330	tp->snd_cwnd_cnt += cnt;
331	while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
332		tp->snd_cwnd_cnt -= tp->snd_cwnd;
333		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
334			tp->snd_cwnd++;
335	}
336}
337EXPORT_SYMBOL_GPL(tcp_slow_start);
338
339/*
340 * TCP Reno congestion control
341 * This is special case used for fallback as well.
342 */
343/* This is Jacobson's slow start and congestion avoidance.
344 * SIGCOMM '88, p. 328.
345 */
346void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
347{
348	struct tcp_sock *tp = tcp_sk(sk);
349
350	if (!tcp_is_cwnd_limited(sk, in_flight))
351		return;
352
353	/* In "safe" area, increase. */
354	if (tp->snd_cwnd <= tp->snd_ssthresh)
355		tcp_slow_start(tp);
356
357	/* In dangerous area, increase slowly. */
358	else if (sysctl_tcp_abc) {
359		/* RFC3465: Appropriate Byte Count
360		 * increase once for each full cwnd acked
361		 */
362		if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
363			tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
364			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
365				tp->snd_cwnd++;
366		}
367	} else {
368		/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
369		if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
370			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
371				tp->snd_cwnd++;
372			tp->snd_cwnd_cnt = 0;
373		} else
374			tp->snd_cwnd_cnt++;
375	}
376}
377EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
378
379/* Slow start threshold is half the congestion window (min 2) */
380u32 tcp_reno_ssthresh(struct sock *sk)
381{
382	const struct tcp_sock *tp = tcp_sk(sk);
383	return max(tp->snd_cwnd >> 1U, 2U);
384}
385EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
386
387/* Lower bound on congestion window with halving. */
388u32 tcp_reno_min_cwnd(const struct sock *sk)
389{
390	const struct tcp_sock *tp = tcp_sk(sk);
391	return tp->snd_ssthresh/2;
392}
393EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
394
395struct tcp_congestion_ops tcp_reno = {
396	.flags		= TCP_CONG_NON_RESTRICTED,
397	.name		= "reno",
398	.owner		= THIS_MODULE,
399	.ssthresh	= tcp_reno_ssthresh,
400	.cong_avoid	= tcp_reno_cong_avoid,
401	.min_cwnd	= tcp_reno_min_cwnd,
402};
403
404/* Initial congestion control used (until SYN)
405 * really reno under another name so we can tell difference
406 * during tcp_set_default_congestion_control
407 */
408struct tcp_congestion_ops tcp_init_congestion_ops  = {
409	.name		= "",
410	.owner		= THIS_MODULE,
411	.ssthresh	= tcp_reno_ssthresh,
412	.cong_avoid	= tcp_reno_cong_avoid,
413	.min_cwnd	= tcp_reno_min_cwnd,
414};
415EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
416