xref: /openbmc/linux/net/ipv4/tcp_cong.c (revision 9cdb81c7)
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 
18 int sysctl_tcp_max_ssthresh = 0;
19 
20 static DEFINE_SPINLOCK(tcp_cong_list_lock);
21 static LIST_HEAD(tcp_cong_list);
22 
23 /* Simple linear search, don't expect many entries! */
24 static 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  */
40 int 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 }
62 EXPORT_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  */
70 void 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 }
76 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
77 
78 /* Assign choice of congestion control. */
79 void 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. */
103 void 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 */
113 int 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 */
141 static int __init tcp_congestion_default(void)
142 {
143 	return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
144 }
145 late_initcall(tcp_congestion_default);
146 
147 
148 /* Build string with list of available congestion control values */
149 void 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 */
165 void 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 */
178 void 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 */
197 int 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 	}
228 out:
229 	spin_unlock(&tcp_cong_list_lock);
230 	kfree(saved_clone);
231 
232 	return ret;
233 }
234 
235 
236 /* Change congestion control for socket */
237 int 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  */
283 int 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 1;
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 		return 1;
296 	return left <= tcp_max_tso_deferred_mss(tp);
297 }
298 EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
299 
300 /*
301  * Slow start is used when congestion window is less than slow start
302  * threshold. This version implements the basic RFC2581 version
303  * and optionally supports:
304  * 	RFC3742 Limited Slow Start  	  - growth limited to max_ssthresh
305  *	RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
306  */
307 void tcp_slow_start(struct tcp_sock *tp)
308 {
309 	int cnt; /* increase in packets */
310 
311 	/* RFC3465: ABC Slow start
312 	 * Increase only after a full MSS of bytes is acked
313 	 *
314 	 * TCP sender SHOULD increase cwnd by the number of
315 	 * previously unacknowledged bytes ACKed by each incoming
316 	 * acknowledgment, provided the increase is not more than L
317 	 */
318 	if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
319 		return;
320 
321 	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
322 		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
323 	else
324 		cnt = tp->snd_cwnd;			/* exponential increase */
325 
326 	/* RFC3465: ABC
327 	 * We MAY increase by 2 if discovered delayed ack
328 	 */
329 	if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
330 		cnt <<= 1;
331 	tp->bytes_acked = 0;
332 
333 	tp->snd_cwnd_cnt += cnt;
334 	while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
335 		tp->snd_cwnd_cnt -= tp->snd_cwnd;
336 		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
337 			tp->snd_cwnd++;
338 	}
339 }
340 EXPORT_SYMBOL_GPL(tcp_slow_start);
341 
342 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
343 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
344 {
345 	if (tp->snd_cwnd_cnt >= w) {
346 		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
347 			tp->snd_cwnd++;
348 		tp->snd_cwnd_cnt = 0;
349 	} else {
350 		tp->snd_cwnd_cnt++;
351 	}
352 }
353 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
354 
355 /*
356  * TCP Reno congestion control
357  * This is special case used for fallback as well.
358  */
359 /* This is Jacobson's slow start and congestion avoidance.
360  * SIGCOMM '88, p. 328.
361  */
362 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
363 {
364 	struct tcp_sock *tp = tcp_sk(sk);
365 
366 	if (!tcp_is_cwnd_limited(sk, in_flight))
367 		return;
368 
369 	/* In "safe" area, increase. */
370 	if (tp->snd_cwnd <= tp->snd_ssthresh)
371 		tcp_slow_start(tp);
372 
373 	/* In dangerous area, increase slowly. */
374 	else if (sysctl_tcp_abc) {
375 		/* RFC3465: Appropriate Byte Count
376 		 * increase once for each full cwnd acked
377 		 */
378 		if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
379 			tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
380 			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
381 				tp->snd_cwnd++;
382 		}
383 	} else {
384 		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
385 	}
386 }
387 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
388 
389 /* Slow start threshold is half the congestion window (min 2) */
390 u32 tcp_reno_ssthresh(struct sock *sk)
391 {
392 	const struct tcp_sock *tp = tcp_sk(sk);
393 	return max(tp->snd_cwnd >> 1U, 2U);
394 }
395 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
396 
397 /* Lower bound on congestion window with halving. */
398 u32 tcp_reno_min_cwnd(const struct sock *sk)
399 {
400 	const struct tcp_sock *tp = tcp_sk(sk);
401 	return tp->snd_ssthresh/2;
402 }
403 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
404 
405 struct tcp_congestion_ops tcp_reno = {
406 	.flags		= TCP_CONG_NON_RESTRICTED,
407 	.name		= "reno",
408 	.owner		= THIS_MODULE,
409 	.ssthresh	= tcp_reno_ssthresh,
410 	.cong_avoid	= tcp_reno_cong_avoid,
411 	.min_cwnd	= tcp_reno_min_cwnd,
412 };
413 
414 /* Initial congestion control used (until SYN)
415  * really reno under another name so we can tell difference
416  * during tcp_set_default_congestion_control
417  */
418 struct tcp_congestion_ops tcp_init_congestion_ops  = {
419 	.name		= "",
420 	.owner		= THIS_MODULE,
421 	.ssthresh	= tcp_reno_ssthresh,
422 	.cong_avoid	= tcp_reno_cong_avoid,
423 	.min_cwnd	= tcp_reno_min_cwnd,
424 };
425 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
426