tcp_cong.c revision 058bd4d2a4ff0aaa4a5381c67e776729d840c785 
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		pr_err("TCP %s does not implement required ops\n", ca->name);
45		return -EINVAL;
46	}
47
48	spin_lock(&tcp_cong_list_lock);
49	if (tcp_ca_find(ca->name)) {
50		pr_notice("TCP %s already registered\n", ca->name);
51		ret = -EEXIST;
52	} else {
53		list_add_tail_rcu(&ca->list, &tcp_cong_list);
54		pr_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_MODULES
119	if (!ca && capable(CAP_NET_ADMIN)) {
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 *saved_clone, *clone, *name;
199	int ret = 0;
200
201	saved_clone = 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	kfree(saved_clone);
229
230	return ret;
231}
232
233
234/* Change congestion control for socket */
235int tcp_set_congestion_control(struct sock *sk, const char *name)
236{
237	struct inet_connection_sock *icsk = inet_csk(sk);
238	struct tcp_congestion_ops *ca;
239	int err = 0;
240
241	rcu_read_lock();
242	ca = tcp_ca_find(name);
243
244	/* no change asking for existing value */
245	if (ca == icsk->icsk_ca_ops)
246		goto out;
247
248#ifdef CONFIG_MODULES
249	/* not found attempt to autoload module */
250	if (!ca && capable(CAP_NET_ADMIN)) {
251		rcu_read_unlock();
252		request_module("tcp_%s", name);
253		rcu_read_lock();
254		ca = tcp_ca_find(name);
255	}
256#endif
257	if (!ca)
258		err = -ENOENT;
259
260	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
261		err = -EPERM;
262
263	else if (!try_module_get(ca->owner))
264		err = -EBUSY;
265
266	else {
267		tcp_cleanup_congestion_control(sk);
268		icsk->icsk_ca_ops = ca;
269
270		if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
271			icsk->icsk_ca_ops->init(sk);
272	}
273 out:
274	rcu_read_unlock();
275	return err;
276}
277
278/* RFC2861 Check whether we are limited by application or congestion window
279 * This is the inverse of cwnd check in tcp_tso_should_defer
280 */
281int tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
282{
283	const struct tcp_sock *tp = tcp_sk(sk);
284	u32 left;
285
286	if (in_flight >= tp->snd_cwnd)
287		return 1;
288
289	left = tp->snd_cwnd - in_flight;
290	if (sk_can_gso(sk) &&
291	    left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
292	    left * tp->mss_cache < sk->sk_gso_max_size)
293		return 1;
294	return left <= tcp_max_tso_deferred_mss(tp);
295}
296EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
297
298/*
299 * Slow start is used when congestion window is less than slow start
300 * threshold. This version implements the basic RFC2581 version
301 * and optionally supports:
302 * 	RFC3742 Limited Slow Start  	  - growth limited to max_ssthresh
303 *	RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
304 */
305void tcp_slow_start(struct tcp_sock *tp)
306{
307	int cnt; /* increase in packets */
308
309	/* RFC3465: ABC Slow start
310	 * Increase only after a full MSS of bytes is acked
311	 *
312	 * TCP sender SHOULD increase cwnd by the number of
313	 * previously unacknowledged bytes ACKed by each incoming
314	 * acknowledgment, provided the increase is not more than L
315	 */
316	if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
317		return;
318
319	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
320		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
321	else
322		cnt = tp->snd_cwnd;			/* exponential increase */
323
324	/* RFC3465: ABC
325	 * We MAY increase by 2 if discovered delayed ack
326	 */
327	if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
328		cnt <<= 1;
329	tp->bytes_acked = 0;
330
331	tp->snd_cwnd_cnt += cnt;
332	while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
333		tp->snd_cwnd_cnt -= tp->snd_cwnd;
334		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
335			tp->snd_cwnd++;
336	}
337}
338EXPORT_SYMBOL_GPL(tcp_slow_start);
339
340/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
341void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
342{
343	if (tp->snd_cwnd_cnt >= w) {
344		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
345			tp->snd_cwnd++;
346		tp->snd_cwnd_cnt = 0;
347	} else {
348		tp->snd_cwnd_cnt++;
349	}
350}
351EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
352
353/*
354 * TCP Reno congestion control
355 * This is special case used for fallback as well.
356 */
357/* This is Jacobson's slow start and congestion avoidance.
358 * SIGCOMM '88, p. 328.
359 */
360void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
361{
362	struct tcp_sock *tp = tcp_sk(sk);
363
364	if (!tcp_is_cwnd_limited(sk, in_flight))
365		return;
366
367	/* In "safe" area, increase. */
368	if (tp->snd_cwnd <= tp->snd_ssthresh)
369		tcp_slow_start(tp);
370
371	/* In dangerous area, increase slowly. */
372	else if (sysctl_tcp_abc) {
373		/* RFC3465: Appropriate Byte Count
374		 * increase once for each full cwnd acked
375		 */
376		if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
377			tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
378			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
379				tp->snd_cwnd++;
380		}
381	} else {
382		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
383	}
384}
385EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
386
387/* Slow start threshold is half the congestion window (min 2) */
388u32 tcp_reno_ssthresh(struct sock *sk)
389{
390	const struct tcp_sock *tp = tcp_sk(sk);
391	return max(tp->snd_cwnd >> 1U, 2U);
392}
393EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
394
395/* Lower bound on congestion window with halving. */
396u32 tcp_reno_min_cwnd(const struct sock *sk)
397{
398	const struct tcp_sock *tp = tcp_sk(sk);
399	return tp->snd_ssthresh/2;
400}
401EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
402
403struct tcp_congestion_ops tcp_reno = {
404	.flags		= TCP_CONG_NON_RESTRICTED,
405	.name		= "reno",
406	.owner		= THIS_MODULE,
407	.ssthresh	= tcp_reno_ssthresh,
408	.cong_avoid	= tcp_reno_cong_avoid,
409	.min_cwnd	= tcp_reno_min_cwnd,
410};
411
412/* Initial congestion control used (until SYN)
413 * really reno under another name so we can tell difference
414 * during tcp_set_default_congestion_control
415 */
416struct tcp_congestion_ops tcp_init_congestion_ops  = {
417	.name		= "",
418	.owner		= THIS_MODULE,
419	.ssthresh	= tcp_reno_ssthresh,
420	.cong_avoid	= tcp_reno_cong_avoid,
421	.min_cwnd	= tcp_reno_min_cwnd,
422};
423EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
424