xref: /openbmc/linux/net/ipv4/tcp_cong.c (revision e368cd72) 
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Pluggable TCP congestion control support and newReno
4  * congestion control.
5  * Based on ideas from I/O scheduler support and Web100.
6  *
7  * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
8  */
9 
10 #define pr_fmt(fmt) "TCP: " fmt
11 
12 #include <linux/module.h>
13 #include <linux/mm.h>
14 #include <linux/types.h>
15 #include <linux/list.h>
16 #include <linux/gfp.h>
17 #include <linux/jhash.h>
18 #include <net/tcp.h>
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 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 /* Must be called with rcu lock held */
37 static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
38 						       const char *name)
39 {
40 	struct tcp_congestion_ops *ca = tcp_ca_find(name);
41 
42 #ifdef CONFIG_MODULES
43 	if (!ca && capable(CAP_NET_ADMIN)) {
44 		rcu_read_unlock();
45 		request_module("tcp_%s", name);
46 		rcu_read_lock();
47 		ca = tcp_ca_find(name);
48 	}
49 #endif
50 	return ca;
51 }
52 
53 /* Simple linear search, not much in here. */
54 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
55 {
56 	struct tcp_congestion_ops *e;
57 
58 	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
59 		if (e->key == key)
60 			return e;
61 	}
62 
63 	return NULL;
64 }
65 
66 /*
67  * Attach new congestion control algorithm to the list
68  * of available options.
69  */
70 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
71 {
72 	int ret = 0;
73 
74 	/* all algorithms must implement these */
75 	if (!ca->ssthresh || !ca->undo_cwnd ||
76 	    !(ca->cong_avoid || ca->cong_control)) {
77 		pr_err("%s does not implement required ops\n", ca->name);
78 		return -EINVAL;
79 	}
80 
81 	ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
82 
83 	spin_lock(&tcp_cong_list_lock);
84 	if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
85 		pr_notice("%s already registered or non-unique key\n",
86 			  ca->name);
87 		ret = -EEXIST;
88 	} else {
89 		list_add_tail_rcu(&ca->list, &tcp_cong_list);
90 		pr_debug("%s registered\n", ca->name);
91 	}
92 	spin_unlock(&tcp_cong_list_lock);
93 
94 	return ret;
95 }
96 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
97 
98 /*
99  * Remove congestion control algorithm, called from
100  * the module's remove function.  Module ref counts are used
101  * to ensure that this can't be done till all sockets using
102  * that method are closed.
103  */
104 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
105 {
106 	spin_lock(&tcp_cong_list_lock);
107 	list_del_rcu(&ca->list);
108 	spin_unlock(&tcp_cong_list_lock);
109 
110 	/* Wait for outstanding readers to complete before the
111 	 * module gets removed entirely.
112 	 *
113 	 * A try_module_get() should fail by now as our module is
114 	 * in "going" state since no refs are held anymore and
115 	 * module_exit() handler being called.
116 	 */
117 	synchronize_rcu();
118 }
119 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120 
121 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
122 {
123 	const struct tcp_congestion_ops *ca;
124 	u32 key = TCP_CA_UNSPEC;
125 
126 	might_sleep();
127 
128 	rcu_read_lock();
129 	ca = tcp_ca_find_autoload(net, name);
130 	if (ca) {
131 		key = ca->key;
132 		*ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
133 	}
134 	rcu_read_unlock();
135 
136 	return key;
137 }
138 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139 
140 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
141 {
142 	const struct tcp_congestion_ops *ca;
143 	char *ret = NULL;
144 
145 	rcu_read_lock();
146 	ca = tcp_ca_find_key(key);
147 	if (ca)
148 		ret = strncpy(buffer, ca->name,
149 			      TCP_CA_NAME_MAX);
150 	rcu_read_unlock();
151 
152 	return ret;
153 }
154 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155 
156 /* Assign choice of congestion control. */
157 void tcp_assign_congestion_control(struct sock *sk)
158 {
159 	struct net *net = sock_net(sk);
160 	struct inet_connection_sock *icsk = inet_csk(sk);
161 	const struct tcp_congestion_ops *ca;
162 
163 	rcu_read_lock();
164 	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
165 	if (unlikely(!bpf_try_module_get(ca, ca->owner)))
166 		ca = &tcp_reno;
167 	icsk->icsk_ca_ops = ca;
168 	rcu_read_unlock();
169 
170 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
171 	if (ca->flags & TCP_CONG_NEEDS_ECN)
172 		INET_ECN_xmit(sk);
173 	else
174 		INET_ECN_dontxmit(sk);
175 }
176 
177 void tcp_init_congestion_control(struct sock *sk)
178 {
179 	struct inet_connection_sock *icsk = inet_csk(sk);
180 
181 	tcp_sk(sk)->prior_ssthresh = 0;
182 	if (icsk->icsk_ca_ops->init)
183 		icsk->icsk_ca_ops->init(sk);
184 	if (tcp_ca_needs_ecn(sk))
185 		INET_ECN_xmit(sk);
186 	else
187 		INET_ECN_dontxmit(sk);
188 	icsk->icsk_ca_initialized = 1;
189 }
190 
191 static void tcp_reinit_congestion_control(struct sock *sk,
192 					  const struct tcp_congestion_ops *ca)
193 {
194 	struct inet_connection_sock *icsk = inet_csk(sk);
195 
196 	tcp_cleanup_congestion_control(sk);
197 	icsk->icsk_ca_ops = ca;
198 	icsk->icsk_ca_setsockopt = 1;
199 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
200 
201 	if (ca->flags & TCP_CONG_NEEDS_ECN)
202 		INET_ECN_xmit(sk);
203 	else
204 		INET_ECN_dontxmit(sk);
205 
206 	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
207 		tcp_init_congestion_control(sk);
208 }
209 
210 /* Manage refcounts on socket close. */
211 void tcp_cleanup_congestion_control(struct sock *sk)
212 {
213 	struct inet_connection_sock *icsk = inet_csk(sk);
214 
215 	if (icsk->icsk_ca_ops->release)
216 		icsk->icsk_ca_ops->release(sk);
217 	bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
218 }
219 
220 /* Used by sysctl to change default congestion control */
221 int tcp_set_default_congestion_control(struct net *net, const char *name)
222 {
223 	struct tcp_congestion_ops *ca;
224 	const struct tcp_congestion_ops *prev;
225 	int ret;
226 
227 	rcu_read_lock();
228 	ca = tcp_ca_find_autoload(net, name);
229 	if (!ca) {
230 		ret = -ENOENT;
231 	} else if (!bpf_try_module_get(ca, ca->owner)) {
232 		ret = -EBUSY;
233 	} else if (!net_eq(net, &init_net) &&
234 			!(ca->flags & TCP_CONG_NON_RESTRICTED)) {
235 		/* Only init netns can set default to a restricted algorithm */
236 		ret = -EPERM;
237 	} else {
238 		prev = xchg(&net->ipv4.tcp_congestion_control, ca);
239 		if (prev)
240 			bpf_module_put(prev, prev->owner);
241 
242 		ca->flags |= TCP_CONG_NON_RESTRICTED;
243 		ret = 0;
244 	}
245 	rcu_read_unlock();
246 
247 	return ret;
248 }
249 
250 /* Set default value from kernel configuration at bootup */
251 static int __init tcp_congestion_default(void)
252 {
253 	return tcp_set_default_congestion_control(&init_net,
254 						  CONFIG_DEFAULT_TCP_CONG);
255 }
256 late_initcall(tcp_congestion_default);
257 
258 /* Build string with list of available congestion control values */
259 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
260 {
261 	struct tcp_congestion_ops *ca;
262 	size_t offs = 0;
263 
264 	rcu_read_lock();
265 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
266 		offs += snprintf(buf + offs, maxlen - offs,
267 				 "%s%s",
268 				 offs == 0 ? "" : " ", ca->name);
269 
270 		if (WARN_ON_ONCE(offs >= maxlen))
271 			break;
272 	}
273 	rcu_read_unlock();
274 }
275 
276 /* Get current default congestion control */
277 void tcp_get_default_congestion_control(struct net *net, char *name)
278 {
279 	const struct tcp_congestion_ops *ca;
280 
281 	rcu_read_lock();
282 	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
283 	strncpy(name, ca->name, TCP_CA_NAME_MAX);
284 	rcu_read_unlock();
285 }
286 
287 /* Built list of non-restricted congestion control values */
288 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
289 {
290 	struct tcp_congestion_ops *ca;
291 	size_t offs = 0;
292 
293 	*buf = '\0';
294 	rcu_read_lock();
295 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
296 		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
297 			continue;
298 		offs += snprintf(buf + offs, maxlen - offs,
299 				 "%s%s",
300 				 offs == 0 ? "" : " ", ca->name);
301 
302 		if (WARN_ON_ONCE(offs >= maxlen))
303 			break;
304 	}
305 	rcu_read_unlock();
306 }
307 
308 /* Change list of non-restricted congestion control */
309 int tcp_set_allowed_congestion_control(char *val)
310 {
311 	struct tcp_congestion_ops *ca;
312 	char *saved_clone, *clone, *name;
313 	int ret = 0;
314 
315 	saved_clone = clone = kstrdup(val, GFP_USER);
316 	if (!clone)
317 		return -ENOMEM;
318 
319 	spin_lock(&tcp_cong_list_lock);
320 	/* pass 1 check for bad entries */
321 	while ((name = strsep(&clone, " ")) && *name) {
322 		ca = tcp_ca_find(name);
323 		if (!ca) {
324 			ret = -ENOENT;
325 			goto out;
326 		}
327 	}
328 
329 	/* pass 2 clear old values */
330 	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
331 		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
332 
333 	/* pass 3 mark as allowed */
334 	while ((name = strsep(&val, " ")) && *name) {
335 		ca = tcp_ca_find(name);
336 		WARN_ON(!ca);
337 		if (ca)
338 			ca->flags |= TCP_CONG_NON_RESTRICTED;
339 	}
340 out:
341 	spin_unlock(&tcp_cong_list_lock);
342 	kfree(saved_clone);
343 
344 	return ret;
345 }
346 
347 /* Change congestion control for socket. If load is false, then it is the
348  * responsibility of the caller to call tcp_init_congestion_control or
349  * tcp_reinit_congestion_control (if the current congestion control was
350  * already initialized.
351  */
352 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
353 			       bool cap_net_admin)
354 {
355 	struct inet_connection_sock *icsk = inet_csk(sk);
356 	const struct tcp_congestion_ops *ca;
357 	int err = 0;
358 
359 	if (icsk->icsk_ca_dst_locked)
360 		return -EPERM;
361 
362 	rcu_read_lock();
363 	if (!load)
364 		ca = tcp_ca_find(name);
365 	else
366 		ca = tcp_ca_find_autoload(sock_net(sk), name);
367 
368 	/* No change asking for existing value */
369 	if (ca == icsk->icsk_ca_ops) {
370 		icsk->icsk_ca_setsockopt = 1;
371 		goto out;
372 	}
373 
374 	if (!ca)
375 		err = -ENOENT;
376 	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
377 		err = -EPERM;
378 	else if (!bpf_try_module_get(ca, ca->owner))
379 		err = -EBUSY;
380 	else
381 		tcp_reinit_congestion_control(sk, ca);
382  out:
383 	rcu_read_unlock();
384 	return err;
385 }
386 
387 /* Slow start is used when congestion window is no greater than the slow start
388  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
389  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
390  * something better;) a packet is only considered (s)acked in its entirety to
391  * defend the ACK attacks described in the RFC. Slow start processes a stretch
392  * ACK of degree N as if N acks of degree 1 are received back to back except
393  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
394  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
395  */
396 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
397 {
398 	u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
399 
400 	acked -= cwnd - tp->snd_cwnd;
401 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
402 
403 	return acked;
404 }
405 EXPORT_SYMBOL_GPL(tcp_slow_start);
406 
407 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
408  * for every packet that was ACKed.
409  */
410 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
411 {
412 	/* If credits accumulated at a higher w, apply them gently now. */
413 	if (tp->snd_cwnd_cnt >= w) {
414 		tp->snd_cwnd_cnt = 0;
415 		tp->snd_cwnd++;
416 	}
417 
418 	tp->snd_cwnd_cnt += acked;
419 	if (tp->snd_cwnd_cnt >= w) {
420 		u32 delta = tp->snd_cwnd_cnt / w;
421 
422 		tp->snd_cwnd_cnt -= delta * w;
423 		tp->snd_cwnd += delta;
424 	}
425 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
426 }
427 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
428 
429 /*
430  * TCP Reno congestion control
431  * This is special case used for fallback as well.
432  */
433 /* This is Jacobson's slow start and congestion avoidance.
434  * SIGCOMM '88, p. 328.
435  */
436 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
437 {
438 	struct tcp_sock *tp = tcp_sk(sk);
439 
440 	if (!tcp_is_cwnd_limited(sk))
441 		return;
442 
443 	/* In "safe" area, increase. */
444 	if (tcp_in_slow_start(tp)) {
445 		acked = tcp_slow_start(tp, acked);
446 		if (!acked)
447 			return;
448 	}
449 	/* In dangerous area, increase slowly. */
450 	tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
451 }
452 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
453 
454 /* Slow start threshold is half the congestion window (min 2) */
455 u32 tcp_reno_ssthresh(struct sock *sk)
456 {
457 	const struct tcp_sock *tp = tcp_sk(sk);
458 
459 	return max(tp->snd_cwnd >> 1U, 2U);
460 }
461 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
462 
463 u32 tcp_reno_undo_cwnd(struct sock *sk)
464 {
465 	const struct tcp_sock *tp = tcp_sk(sk);
466 
467 	return max(tp->snd_cwnd, tp->prior_cwnd);
468 }
469 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
470 
471 struct tcp_congestion_ops tcp_reno = {
472 	.flags		= TCP_CONG_NON_RESTRICTED,
473 	.name		= "reno",
474 	.owner		= THIS_MODULE,
475 	.ssthresh	= tcp_reno_ssthresh,
476 	.cong_avoid	= tcp_reno_cong_avoid,
477 	.undo_cwnd	= tcp_reno_undo_cwnd,
478 };
479