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 #include <trace/events/tcp.h>
20 
21 static DEFINE_SPINLOCK(tcp_cong_list_lock);
22 static LIST_HEAD(tcp_cong_list);
23 
24 /* Simple linear search, don't expect many entries! */
tcp_ca_find(const char * name)25 struct tcp_congestion_ops *tcp_ca_find(const char *name)
26 {
27 	struct tcp_congestion_ops *e;
28 
29 	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
30 		if (strcmp(e->name, name) == 0)
31 			return e;
32 	}
33 
34 	return NULL;
35 }
36 
tcp_set_ca_state(struct sock * sk,const u8 ca_state)37 void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
38 {
39 	struct inet_connection_sock *icsk = inet_csk(sk);
40 
41 	trace_tcp_cong_state_set(sk, ca_state);
42 
43 	if (icsk->icsk_ca_ops->set_state)
44 		icsk->icsk_ca_ops->set_state(sk, ca_state);
45 	icsk->icsk_ca_state = ca_state;
46 }
47 
48 /* Must be called with rcu lock held */
tcp_ca_find_autoload(struct net * net,const char * name)49 static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
50 						       const char *name)
51 {
52 	struct tcp_congestion_ops *ca = tcp_ca_find(name);
53 
54 #ifdef CONFIG_MODULES
55 	if (!ca && capable(CAP_NET_ADMIN)) {
56 		rcu_read_unlock();
57 		request_module("tcp_%s", name);
58 		rcu_read_lock();
59 		ca = tcp_ca_find(name);
60 	}
61 #endif
62 	return ca;
63 }
64 
65 /* Simple linear search, not much in here. */
tcp_ca_find_key(u32 key)66 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
67 {
68 	struct tcp_congestion_ops *e;
69 
70 	list_for_each_entry_rcu(e, &tcp_cong_list, list) {
71 		if (e->key == key)
72 			return e;
73 	}
74 
75 	return NULL;
76 }
77 
78 /*
79  * Attach new congestion control algorithm to the list
80  * of available options.
81  */
tcp_register_congestion_control(struct tcp_congestion_ops * ca)82 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
83 {
84 	int ret = 0;
85 
86 	/* all algorithms must implement these */
87 	if (!ca->ssthresh || !ca->undo_cwnd ||
88 	    !(ca->cong_avoid || ca->cong_control)) {
89 		pr_err("%s does not implement required ops\n", ca->name);
90 		return -EINVAL;
91 	}
92 
93 	ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
94 
95 	spin_lock(&tcp_cong_list_lock);
96 	if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
97 		pr_notice("%s already registered or non-unique key\n",
98 			  ca->name);
99 		ret = -EEXIST;
100 	} else {
101 		list_add_tail_rcu(&ca->list, &tcp_cong_list);
102 		pr_debug("%s registered\n", ca->name);
103 	}
104 	spin_unlock(&tcp_cong_list_lock);
105 
106 	return ret;
107 }
108 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
109 
110 /*
111  * Remove congestion control algorithm, called from
112  * the module's remove function.  Module ref counts are used
113  * to ensure that this can't be done till all sockets using
114  * that method are closed.
115  */
tcp_unregister_congestion_control(struct tcp_congestion_ops * ca)116 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
117 {
118 	spin_lock(&tcp_cong_list_lock);
119 	list_del_rcu(&ca->list);
120 	spin_unlock(&tcp_cong_list_lock);
121 
122 	/* Wait for outstanding readers to complete before the
123 	 * module gets removed entirely.
124 	 *
125 	 * A try_module_get() should fail by now as our module is
126 	 * in "going" state since no refs are held anymore and
127 	 * module_exit() handler being called.
128 	 */
129 	synchronize_rcu();
130 }
131 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
132 
tcp_ca_get_key_by_name(struct net * net,const char * name,bool * ecn_ca)133 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
134 {
135 	const struct tcp_congestion_ops *ca;
136 	u32 key = TCP_CA_UNSPEC;
137 
138 	might_sleep();
139 
140 	rcu_read_lock();
141 	ca = tcp_ca_find_autoload(net, name);
142 	if (ca) {
143 		key = ca->key;
144 		*ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
145 	}
146 	rcu_read_unlock();
147 
148 	return key;
149 }
150 
tcp_ca_get_name_by_key(u32 key,char * buffer)151 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
152 {
153 	const struct tcp_congestion_ops *ca;
154 	char *ret = NULL;
155 
156 	rcu_read_lock();
157 	ca = tcp_ca_find_key(key);
158 	if (ca)
159 		ret = strncpy(buffer, ca->name,
160 			      TCP_CA_NAME_MAX);
161 	rcu_read_unlock();
162 
163 	return ret;
164 }
165 
166 /* Assign choice of congestion control. */
tcp_assign_congestion_control(struct sock * sk)167 void tcp_assign_congestion_control(struct sock *sk)
168 {
169 	struct net *net = sock_net(sk);
170 	struct inet_connection_sock *icsk = inet_csk(sk);
171 	const struct tcp_congestion_ops *ca;
172 
173 	rcu_read_lock();
174 	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
175 	if (unlikely(!bpf_try_module_get(ca, ca->owner)))
176 		ca = &tcp_reno;
177 	icsk->icsk_ca_ops = ca;
178 	rcu_read_unlock();
179 
180 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
181 	if (ca->flags & TCP_CONG_NEEDS_ECN)
182 		INET_ECN_xmit(sk);
183 	else
184 		INET_ECN_dontxmit(sk);
185 }
186 
tcp_init_congestion_control(struct sock * sk)187 void tcp_init_congestion_control(struct sock *sk)
188 {
189 	struct inet_connection_sock *icsk = inet_csk(sk);
190 
191 	tcp_sk(sk)->prior_ssthresh = 0;
192 	if (icsk->icsk_ca_ops->init)
193 		icsk->icsk_ca_ops->init(sk);
194 	if (tcp_ca_needs_ecn(sk))
195 		INET_ECN_xmit(sk);
196 	else
197 		INET_ECN_dontxmit(sk);
198 	icsk->icsk_ca_initialized = 1;
199 }
200 
tcp_reinit_congestion_control(struct sock * sk,const struct tcp_congestion_ops * ca)201 static void tcp_reinit_congestion_control(struct sock *sk,
202 					  const struct tcp_congestion_ops *ca)
203 {
204 	struct inet_connection_sock *icsk = inet_csk(sk);
205 
206 	tcp_cleanup_congestion_control(sk);
207 	icsk->icsk_ca_ops = ca;
208 	icsk->icsk_ca_setsockopt = 1;
209 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
210 
211 	if (ca->flags & TCP_CONG_NEEDS_ECN)
212 		INET_ECN_xmit(sk);
213 	else
214 		INET_ECN_dontxmit(sk);
215 
216 	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
217 		tcp_init_congestion_control(sk);
218 }
219 
220 /* Manage refcounts on socket close. */
tcp_cleanup_congestion_control(struct sock * sk)221 void tcp_cleanup_congestion_control(struct sock *sk)
222 {
223 	struct inet_connection_sock *icsk = inet_csk(sk);
224 
225 	if (icsk->icsk_ca_ops->release)
226 		icsk->icsk_ca_ops->release(sk);
227 	bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
228 }
229 
230 /* Used by sysctl to change default congestion control */
tcp_set_default_congestion_control(struct net * net,const char * name)231 int tcp_set_default_congestion_control(struct net *net, const char *name)
232 {
233 	struct tcp_congestion_ops *ca;
234 	const struct tcp_congestion_ops *prev;
235 	int ret;
236 
237 	rcu_read_lock();
238 	ca = tcp_ca_find_autoload(net, name);
239 	if (!ca) {
240 		ret = -ENOENT;
241 	} else if (!bpf_try_module_get(ca, ca->owner)) {
242 		ret = -EBUSY;
243 	} else if (!net_eq(net, &init_net) &&
244 			!(ca->flags & TCP_CONG_NON_RESTRICTED)) {
245 		/* Only init netns can set default to a restricted algorithm */
246 		ret = -EPERM;
247 	} else {
248 		prev = xchg(&net->ipv4.tcp_congestion_control, ca);
249 		if (prev)
250 			bpf_module_put(prev, prev->owner);
251 
252 		ca->flags |= TCP_CONG_NON_RESTRICTED;
253 		ret = 0;
254 	}
255 	rcu_read_unlock();
256 
257 	return ret;
258 }
259 
260 /* Set default value from kernel configuration at bootup */
tcp_congestion_default(void)261 static int __init tcp_congestion_default(void)
262 {
263 	return tcp_set_default_congestion_control(&init_net,
264 						  CONFIG_DEFAULT_TCP_CONG);
265 }
266 late_initcall(tcp_congestion_default);
267 
268 /* Build string with list of available congestion control values */
tcp_get_available_congestion_control(char * buf,size_t maxlen)269 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
270 {
271 	struct tcp_congestion_ops *ca;
272 	size_t offs = 0;
273 
274 	rcu_read_lock();
275 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
276 		offs += snprintf(buf + offs, maxlen - offs,
277 				 "%s%s",
278 				 offs == 0 ? "" : " ", ca->name);
279 
280 		if (WARN_ON_ONCE(offs >= maxlen))
281 			break;
282 	}
283 	rcu_read_unlock();
284 }
285 
286 /* Get current default congestion control */
tcp_get_default_congestion_control(struct net * net,char * name)287 void tcp_get_default_congestion_control(struct net *net, char *name)
288 {
289 	const struct tcp_congestion_ops *ca;
290 
291 	rcu_read_lock();
292 	ca = rcu_dereference(net->ipv4.tcp_congestion_control);
293 	strncpy(name, ca->name, TCP_CA_NAME_MAX);
294 	rcu_read_unlock();
295 }
296 
297 /* Built list of non-restricted congestion control values */
tcp_get_allowed_congestion_control(char * buf,size_t maxlen)298 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
299 {
300 	struct tcp_congestion_ops *ca;
301 	size_t offs = 0;
302 
303 	*buf = '\0';
304 	rcu_read_lock();
305 	list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
306 		if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
307 			continue;
308 		offs += snprintf(buf + offs, maxlen - offs,
309 				 "%s%s",
310 				 offs == 0 ? "" : " ", ca->name);
311 
312 		if (WARN_ON_ONCE(offs >= maxlen))
313 			break;
314 	}
315 	rcu_read_unlock();
316 }
317 
318 /* Change list of non-restricted congestion control */
tcp_set_allowed_congestion_control(char * val)319 int tcp_set_allowed_congestion_control(char *val)
320 {
321 	struct tcp_congestion_ops *ca;
322 	char *saved_clone, *clone, *name;
323 	int ret = 0;
324 
325 	saved_clone = clone = kstrdup(val, GFP_USER);
326 	if (!clone)
327 		return -ENOMEM;
328 
329 	spin_lock(&tcp_cong_list_lock);
330 	/* pass 1 check for bad entries */
331 	while ((name = strsep(&clone, " ")) && *name) {
332 		ca = tcp_ca_find(name);
333 		if (!ca) {
334 			ret = -ENOENT;
335 			goto out;
336 		}
337 	}
338 
339 	/* pass 2 clear old values */
340 	list_for_each_entry_rcu(ca, &tcp_cong_list, list)
341 		ca->flags &= ~TCP_CONG_NON_RESTRICTED;
342 
343 	/* pass 3 mark as allowed */
344 	while ((name = strsep(&val, " ")) && *name) {
345 		ca = tcp_ca_find(name);
346 		WARN_ON(!ca);
347 		if (ca)
348 			ca->flags |= TCP_CONG_NON_RESTRICTED;
349 	}
350 out:
351 	spin_unlock(&tcp_cong_list_lock);
352 	kfree(saved_clone);
353 
354 	return ret;
355 }
356 
357 /* Change congestion control for socket. If load is false, then it is the
358  * responsibility of the caller to call tcp_init_congestion_control or
359  * tcp_reinit_congestion_control (if the current congestion control was
360  * already initialized.
361  */
tcp_set_congestion_control(struct sock * sk,const char * name,bool load,bool cap_net_admin)362 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
363 			       bool cap_net_admin)
364 {
365 	struct inet_connection_sock *icsk = inet_csk(sk);
366 	const struct tcp_congestion_ops *ca;
367 	int err = 0;
368 
369 	if (icsk->icsk_ca_dst_locked)
370 		return -EPERM;
371 
372 	rcu_read_lock();
373 	if (!load)
374 		ca = tcp_ca_find(name);
375 	else
376 		ca = tcp_ca_find_autoload(sock_net(sk), name);
377 
378 	/* No change asking for existing value */
379 	if (ca == icsk->icsk_ca_ops) {
380 		icsk->icsk_ca_setsockopt = 1;
381 		goto out;
382 	}
383 
384 	if (!ca)
385 		err = -ENOENT;
386 	else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
387 		err = -EPERM;
388 	else if (!bpf_try_module_get(ca, ca->owner))
389 		err = -EBUSY;
390 	else
391 		tcp_reinit_congestion_control(sk, ca);
392  out:
393 	rcu_read_unlock();
394 	return err;
395 }
396 
397 /* Slow start is used when congestion window is no greater than the slow start
398  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
399  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
400  * something better;) a packet is only considered (s)acked in its entirety to
401  * defend the ACK attacks described in the RFC. Slow start processes a stretch
402  * ACK of degree N as if N acks of degree 1 are received back to back except
403  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
404  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
405  */
tcp_slow_start(struct tcp_sock * tp,u32 acked)406 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
407 {
408 	u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
409 
410 	acked -= cwnd - tcp_snd_cwnd(tp);
411 	tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
412 
413 	return acked;
414 }
415 EXPORT_SYMBOL_GPL(tcp_slow_start);
416 
417 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
418  * for every packet that was ACKed.
419  */
tcp_cong_avoid_ai(struct tcp_sock * tp,u32 w,u32 acked)420 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
421 {
422 	/* If credits accumulated at a higher w, apply them gently now. */
423 	if (tp->snd_cwnd_cnt >= w) {
424 		tp->snd_cwnd_cnt = 0;
425 		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
426 	}
427 
428 	tp->snd_cwnd_cnt += acked;
429 	if (tp->snd_cwnd_cnt >= w) {
430 		u32 delta = tp->snd_cwnd_cnt / w;
431 
432 		tp->snd_cwnd_cnt -= delta * w;
433 		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
434 	}
435 	tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
436 }
437 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
438 
439 /*
440  * TCP Reno congestion control
441  * This is special case used for fallback as well.
442  */
443 /* This is Jacobson's slow start and congestion avoidance.
444  * SIGCOMM '88, p. 328.
445  */
tcp_reno_cong_avoid(struct sock * sk,u32 ack,u32 acked)446 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
447 {
448 	struct tcp_sock *tp = tcp_sk(sk);
449 
450 	if (!tcp_is_cwnd_limited(sk))
451 		return;
452 
453 	/* In "safe" area, increase. */
454 	if (tcp_in_slow_start(tp)) {
455 		acked = tcp_slow_start(tp, acked);
456 		if (!acked)
457 			return;
458 	}
459 	/* In dangerous area, increase slowly. */
460 	tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
461 }
462 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
463 
464 /* Slow start threshold is half the congestion window (min 2) */
tcp_reno_ssthresh(struct sock * sk)465 u32 tcp_reno_ssthresh(struct sock *sk)
466 {
467 	const struct tcp_sock *tp = tcp_sk(sk);
468 
469 	return max(tcp_snd_cwnd(tp) >> 1U, 2U);
470 }
471 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
472 
tcp_reno_undo_cwnd(struct sock * sk)473 u32 tcp_reno_undo_cwnd(struct sock *sk)
474 {
475 	const struct tcp_sock *tp = tcp_sk(sk);
476 
477 	return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
478 }
479 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
480 
481 struct tcp_congestion_ops tcp_reno = {
482 	.flags		= TCP_CONG_NON_RESTRICTED,
483 	.name		= "reno",
484 	.owner		= THIS_MODULE,
485 	.ssthresh	= tcp_reno_ssthresh,
486 	.cong_avoid	= tcp_reno_cong_avoid,
487 	.undo_cwnd	= tcp_reno_undo_cwnd,
488 };
489