xref: /openbmc/linux/net/ipv4/tcp_cong.c (revision 275876e2)
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 
18 static DEFINE_SPINLOCK(tcp_cong_list_lock);
19 static LIST_HEAD(tcp_cong_list);
20 
21 /* Simple linear search, don't expect many entries! */
22 static 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  */
38 int 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("%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("%s already registered\n", ca->name);
51 		ret = -EEXIST;
52 	} else {
53 		list_add_tail_rcu(&ca->list, &tcp_cong_list);
54 		pr_info("%s registered\n", ca->name);
55 	}
56 	spin_unlock(&tcp_cong_list_lock);
57 
58 	return ret;
59 }
60 EXPORT_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  */
68 void 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 }
74 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
75 
76 /* Assign choice of congestion control. */
77 void 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. */
101 void 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 */
111 int 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 */
139 static int __init tcp_congestion_default(void)
140 {
141 	return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
142 }
143 late_initcall(tcp_congestion_default);
144 
145 
146 /* Build string with list of available congestion control values */
147 void 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 */
163 void 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 */
176 void 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 */
195 int 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 	}
226 out:
227 	spin_unlock(&tcp_cong_list_lock);
228 	kfree(saved_clone);
229 
230 	return ret;
231 }
232 
233 
234 /* Change congestion control for socket */
235 int 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) ||
261 		   ns_capable(sock_net(sk)->user_ns, 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 /* Slow start is used when congestion window is no greater than the slow start
280  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
281  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
282  * something better;) a packet is only considered (s)acked in its entirety to
283  * defend the ACK attacks described in the RFC. Slow start processes a stretch
284  * ACK of degree N as if N acks of degree 1 are received back to back except
285  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
286  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
287  */
288 int tcp_slow_start(struct tcp_sock *tp, u32 acked)
289 {
290 	u32 cwnd = tp->snd_cwnd + acked;
291 
292 	if (cwnd > tp->snd_ssthresh)
293 		cwnd = tp->snd_ssthresh + 1;
294 	acked -= cwnd - tp->snd_cwnd;
295 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
296 	return acked;
297 }
298 EXPORT_SYMBOL_GPL(tcp_slow_start);
299 
300 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
301 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
302 {
303 	if (tp->snd_cwnd_cnt >= w) {
304 		if (tp->snd_cwnd < tp->snd_cwnd_clamp)
305 			tp->snd_cwnd++;
306 		tp->snd_cwnd_cnt = 0;
307 	} else {
308 		tp->snd_cwnd_cnt++;
309 	}
310 }
311 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
312 
313 /*
314  * TCP Reno congestion control
315  * This is special case used for fallback as well.
316  */
317 /* This is Jacobson's slow start and congestion avoidance.
318  * SIGCOMM '88, p. 328.
319  */
320 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
321 {
322 	struct tcp_sock *tp = tcp_sk(sk);
323 
324 	if (!tcp_is_cwnd_limited(sk))
325 		return;
326 
327 	/* In "safe" area, increase. */
328 	if (tp->snd_cwnd <= tp->snd_ssthresh)
329 		tcp_slow_start(tp, acked);
330 	/* In dangerous area, increase slowly. */
331 	else
332 		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
333 }
334 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
335 
336 /* Slow start threshold is half the congestion window (min 2) */
337 u32 tcp_reno_ssthresh(struct sock *sk)
338 {
339 	const struct tcp_sock *tp = tcp_sk(sk);
340 	return max(tp->snd_cwnd >> 1U, 2U);
341 }
342 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
343 
344 struct tcp_congestion_ops tcp_reno = {
345 	.flags		= TCP_CONG_NON_RESTRICTED,
346 	.name		= "reno",
347 	.owner		= THIS_MODULE,
348 	.ssthresh	= tcp_reno_ssthresh,
349 	.cong_avoid	= tcp_reno_cong_avoid,
350 };
351 
352 /* Initial congestion control used (until SYN)
353  * really reno under another name so we can tell difference
354  * during tcp_set_default_congestion_control
355  */
356 struct tcp_congestion_ops tcp_init_congestion_ops  = {
357 	.name		= "",
358 	.owner		= THIS_MODULE,
359 	.ssthresh	= tcp_reno_ssthresh,
360 	.cong_avoid	= tcp_reno_cong_avoid,
361 };
362 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
363