tcp_cong.c revision 02582e9bcc36ed503ffede46e104a885dea222fb 
1/*
2 * Plugable TCP congestion control support and newReno
3 * congestion control.
4 * Based on ideas from I/O scheduler support 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	unsigned int delta = 0;
312
313	/* RFC3465: ABC Slow start
314	 * Increase only after a full MSS of bytes is acked
315	 *
316	 * TCP sender SHOULD increase cwnd by the number of
317	 * previously unacknowledged bytes ACKed by each incoming
318	 * acknowledgment, provided the increase is not more than L
319	 */
320	if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
321		return;
322
323	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
324		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
325	else
326		cnt = tp->snd_cwnd;			/* exponential increase */
327
328	/* RFC3465: ABC
329	 * We MAY increase by 2 if discovered delayed ack
330	 */
331	if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
332		cnt <<= 1;
333	tp->bytes_acked = 0;
334
335	tp->snd_cwnd_cnt += cnt;
336	while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
337		tp->snd_cwnd_cnt -= tp->snd_cwnd;
338		delta++;
339	}
340	tp->snd_cwnd = min(tp->snd_cwnd + delta, tp->snd_cwnd_clamp);
341}
342EXPORT_SYMBOL_GPL(tcp_slow_start);
343
344/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
345void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
346{
347	if (tp->snd_cwnd_cnt >= w) {
348		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
349			tp->snd_cwnd++;
350		tp->snd_cwnd_cnt = 0;
351	} else {
352		tp->snd_cwnd_cnt++;
353	}
354}
355EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
356
357/*
358 * TCP Reno congestion control
359 * This is special case used for fallback as well.
360 */
361/* This is Jacobson's slow start and congestion avoidance.
362 * SIGCOMM '88, p. 328.
363 */
364void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
365{
366	struct tcp_sock *tp = tcp_sk(sk);
367
368	if (!tcp_is_cwnd_limited(sk, in_flight))
369		return;
370
371	/* In "safe" area, increase. */
372	if (tp->snd_cwnd <= tp->snd_ssthresh)
373		tcp_slow_start(tp);
374
375	/* In dangerous area, increase slowly. */
376	else if (sysctl_tcp_abc) {
377		/* RFC3465: Appropriate Byte Count
378		 * increase once for each full cwnd acked
379		 */
380		if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
381			tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
382			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
383				tp->snd_cwnd++;
384		}
385	} else {
386		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
387	}
388}
389EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
390
391/* Slow start threshold is half the congestion window (min 2) */
392u32 tcp_reno_ssthresh(struct sock *sk)
393{
394	const struct tcp_sock *tp = tcp_sk(sk);
395	return max(tp->snd_cwnd >> 1U, 2U);
396}
397EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
398
399/* Lower bound on congestion window with halving. */
400u32 tcp_reno_min_cwnd(const struct sock *sk)
401{
402	const struct tcp_sock *tp = tcp_sk(sk);
403	return tp->snd_ssthresh/2;
404}
405EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
406
407struct tcp_congestion_ops tcp_reno = {
408	.flags		= TCP_CONG_NON_RESTRICTED,
409	.name		= "reno",
410	.owner		= THIS_MODULE,
411	.ssthresh	= tcp_reno_ssthresh,
412	.cong_avoid	= tcp_reno_cong_avoid,
413	.min_cwnd	= tcp_reno_min_cwnd,
414};
415
416/* Initial congestion control used (until SYN)
417 * really reno under another name so we can tell difference
418 * during tcp_set_default_congestion_control
419 */
420struct tcp_congestion_ops tcp_init_congestion_ops  = {
421	.name		= "",
422	.owner		= THIS_MODULE,
423	.ssthresh	= tcp_reno_ssthresh,
424	.cong_avoid	= tcp_reno_cong_avoid,
425	.min_cwnd	= tcp_reno_min_cwnd,
426};
427EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
428