xref: /openbmc/linux/net/ipv4/tcp_cong.c (revision b2765275)
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 	const 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 }
189 
190 static void tcp_reinit_congestion_control(struct sock *sk,
191 					  const struct tcp_congestion_ops *ca)
192 {
193 	struct inet_connection_sock *icsk = inet_csk(sk);
194 
195 	tcp_cleanup_congestion_control(sk);
196 	icsk->icsk_ca_ops = ca;
197 	icsk->icsk_ca_setsockopt = 1;
198 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
199 
200 	if (sk->sk_state != TCP_CLOSE)
201 		tcp_init_congestion_control(sk);
202 }
203 
204 /* Manage refcounts on socket close. */
205 void tcp_cleanup_congestion_control(struct sock *sk)
206 {
207 	struct inet_connection_sock *icsk = inet_csk(sk);
208 
209 	if (icsk->icsk_ca_ops->release)
210 		icsk->icsk_ca_ops->release(sk);
211 	bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
212 }
213 
214 /* Used by sysctl to change default congestion control */
215 int tcp_set_default_congestion_control(struct net *net, const char *name)
216 {
217 	struct tcp_congestion_ops *ca;
218 	const struct tcp_congestion_ops *prev;
219 	int ret;
220 
221 	rcu_read_lock();
222 	ca = tcp_ca_find_autoload(net, name);
223 	if (!ca) {
224 		ret = -ENOENT;
225 	} else if (!bpf_try_module_get(ca, ca->owner)) {
226 		ret = -EBUSY;
227 	} else {
228 		prev = xchg(&net->ipv4.tcp_congestion_control, ca);
229 		if (prev)
230 			bpf_module_put(prev, prev->owner);
231 
232 		ca->flags |= TCP_CONG_NON_RESTRICTED;
233 		ret = 0;
234 	}
235 	rcu_read_unlock();
236 
237 	return ret;
238 }
239 
240 /* Set default value from kernel configuration at bootup */
241 static int __init tcp_congestion_default(void)
242 {
243 	return tcp_set_default_congestion_control(&init_net,
244 						  CONFIG_DEFAULT_TCP_CONG);
245 }
246 late_initcall(tcp_congestion_default);
247 
248 /* Build string with list of available congestion control values */
249 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
250 {
251 	struct tcp_congestion_ops *ca;
252 	size_t offs = 0;
253 
254 	rcu_read_lock();
255 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
256 		offs += snprintf(buf + offs, maxlen - offs,
257 				 "%s%s",
258 				 offs == 0 ? "" : " ", ca->name);
259 
260 		if (WARN_ON_ONCE(offs >= maxlen))
261 			break;
262 	}
263 	rcu_read_unlock();
264 }
265 
266 /* Get current default congestion control */
267 void tcp_get_default_congestion_control(struct net *net, char *name)
268 {
269 	const struct tcp_congestion_ops *ca;
270 
271 	rcu_read_lock();
272 	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
273 	strncpy(name, ca->name, TCP_CA_NAME_MAX);
274 	rcu_read_unlock();
275 }
276 
277 /* Built list of non-restricted congestion control values */
278 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
279 {
280 	struct tcp_congestion_ops *ca;
281 	size_t offs = 0;
282 
283 	*buf = '\0';
284 	rcu_read_lock();
285 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
286 		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
287 			continue;
288 		offs += snprintf(buf + offs, maxlen - offs,
289 				 "%s%s",
290 				 offs == 0 ? "" : " ", ca->name);
291 
292 		if (WARN_ON_ONCE(offs >= maxlen))
293 			break;
294 	}
295 	rcu_read_unlock();
296 }
297 
298 /* Change list of non-restricted congestion control */
299 int tcp_set_allowed_congestion_control(char *val)
300 {
301 	struct tcp_congestion_ops *ca;
302 	char *saved_clone, *clone, *name;
303 	int ret = 0;
304 
305 	saved_clone = clone = kstrdup(val, GFP_USER);
306 	if (!clone)
307 		return -ENOMEM;
308 
309 	spin_lock(&tcp_cong_list_lock);
310 	/* pass 1 check for bad entries */
311 	while ((name = strsep(&clone, " ")) && *name) {
312 		ca = tcp_ca_find(name);
313 		if (!ca) {
314 			ret = -ENOENT;
315 			goto out;
316 		}
317 	}
318 
319 	/* pass 2 clear old values */
320 	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
321 		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
322 
323 	/* pass 3 mark as allowed */
324 	while ((name = strsep(&val, " ")) && *name) {
325 		ca = tcp_ca_find(name);
326 		WARN_ON(!ca);
327 		if (ca)
328 			ca->flags |= TCP_CONG_NON_RESTRICTED;
329 	}
330 out:
331 	spin_unlock(&tcp_cong_list_lock);
332 	kfree(saved_clone);
333 
334 	return ret;
335 }
336 
337 /* Change congestion control for socket. If load is false, then it is the
338  * responsibility of the caller to call tcp_init_congestion_control or
339  * tcp_reinit_congestion_control (if the current congestion control was
340  * already initialized.
341  */
342 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
343 			       bool reinit, bool cap_net_admin)
344 {
345 	struct inet_connection_sock *icsk = inet_csk(sk);
346 	const struct tcp_congestion_ops *ca;
347 	int err = 0;
348 
349 	if (icsk->icsk_ca_dst_locked)
350 		return -EPERM;
351 
352 	rcu_read_lock();
353 	if (!load)
354 		ca = tcp_ca_find(name);
355 	else
356 		ca = tcp_ca_find_autoload(sock_net(sk), name);
357 
358 	/* No change asking for existing value */
359 	if (ca == icsk->icsk_ca_ops) {
360 		icsk->icsk_ca_setsockopt = 1;
361 		goto out;
362 	}
363 
364 	if (!ca) {
365 		err = -ENOENT;
366 	} else if (!load) {
367 		const struct tcp_congestion_ops *old_ca = icsk->icsk_ca_ops;
368 
369 		if (bpf_try_module_get(ca, ca->owner)) {
370 			if (reinit) {
371 				tcp_reinit_congestion_control(sk, ca);
372 			} else {
373 				icsk->icsk_ca_ops = ca;
374 				bpf_module_put(old_ca, old_ca->owner);
375 			}
376 		} else {
377 			err = -EBUSY;
378 		}
379 	} else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin)) {
380 		err = -EPERM;
381 	} else if (!bpf_try_module_get(ca, ca->owner)) {
382 		err = -EBUSY;
383 	} else {
384 		tcp_reinit_congestion_control(sk, ca);
385 	}
386  out:
387 	rcu_read_unlock();
388 	return err;
389 }
390 
391 /* Slow start is used when congestion window is no greater than the slow start
392  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
393  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
394  * something better;) a packet is only considered (s)acked in its entirety to
395  * defend the ACK attacks described in the RFC. Slow start processes a stretch
396  * ACK of degree N as if N acks of degree 1 are received back to back except
397  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
398  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
399  */
400 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
401 {
402 	u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
403 
404 	acked -= cwnd - tp->snd_cwnd;
405 	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
406 
407 	return acked;
408 }
409 EXPORT_SYMBOL_GPL(tcp_slow_start);
410 
411 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
412  * for every packet that was ACKed.
413  */
414 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
415 {
416 	/* If credits accumulated at a higher w, apply them gently now. */
417 	if (tp->snd_cwnd_cnt >= w) {
418 		tp->snd_cwnd_cnt = 0;
419 		tp->snd_cwnd++;
420 	}
421 
422 	tp->snd_cwnd_cnt += acked;
423 	if (tp->snd_cwnd_cnt >= w) {
424 		u32 delta = tp->snd_cwnd_cnt / w;
425 
426 		tp->snd_cwnd_cnt -= delta * w;
427 		tp->snd_cwnd += delta;
428 	}
429 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
430 }
431 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
432 
433 /*
434  * TCP Reno congestion control
435  * This is special case used for fallback as well.
436  */
437 /* This is Jacobson's slow start and congestion avoidance.
438  * SIGCOMM '88, p. 328.
439  */
440 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
441 {
442 	struct tcp_sock *tp = tcp_sk(sk);
443 
444 	if (!tcp_is_cwnd_limited(sk))
445 		return;
446 
447 	/* In "safe" area, increase. */
448 	if (tcp_in_slow_start(tp)) {
449 		acked = tcp_slow_start(tp, acked);
450 		if (!acked)
451 			return;
452 	}
453 	/* In dangerous area, increase slowly. */
454 	tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
455 }
456 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
457 
458 /* Slow start threshold is half the congestion window (min 2) */
459 u32 tcp_reno_ssthresh(struct sock *sk)
460 {
461 	const struct tcp_sock *tp = tcp_sk(sk);
462 
463 	return max(tp->snd_cwnd >> 1U, 2U);
464 }
465 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
466 
467 u32 tcp_reno_undo_cwnd(struct sock *sk)
468 {
469 	const struct tcp_sock *tp = tcp_sk(sk);
470 
471 	return max(tp->snd_cwnd, tp->prior_cwnd);
472 }
473 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
474 
475 struct tcp_congestion_ops tcp_reno = {
476 	.flags		= TCP_CONG_NON_RESTRICTED,
477 	.name		= "reno",
478 	.owner		= THIS_MODULE,
479 	.ssthresh	= tcp_reno_ssthresh,
480 	.cong_avoid	= tcp_reno_cong_avoid,
481 	.undo_cwnd	= tcp_reno_undo_cwnd,
482 };
483