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
21static DEFINE_SPINLOCK(tcp_cong_list_lock);
22static LIST_HEAD(tcp_cong_list);
23
24/* Simple linear search, don't expect many entries! */
25struct 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
37void 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 */
49static struct tcp_congestion_ops *tcp_ca_find_autoload(const char *name)
50{
51 struct tcp_congestion_ops *ca = tcp_ca_find(name);
52
53#ifdef CONFIG_MODULES
54 if (!ca && capable(CAP_NET_ADMIN)) {
55 rcu_read_unlock();
56 request_module("tcp_%s", name);
57 rcu_read_lock();
58 ca = tcp_ca_find(name);
59 }
60#endif
61 return ca;
62}
63
64/* Simple linear search, not much in here. */
65struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
66{
67 struct tcp_congestion_ops *e;
68
69 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
70 if (e->key == key)
71 return e;
72 }
73
74 return NULL;
75}
76
77int tcp_validate_congestion_control(struct tcp_congestion_ops *ca)
78{
79 /* all algorithms must implement these */
80 if (!ca->ssthresh || !ca->undo_cwnd ||
81 !(ca->cong_avoid || ca->cong_control)) {
82 pr_err("%s does not implement required ops\n", ca->name);
83 return -EINVAL;
84 }
85
86 return 0;
87}
88
89/* Attach new congestion control algorithm to the list
90 * of available options.
91 */
92int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
93{
94 int ret;
95
96 ret = tcp_validate_congestion_control(ca);
97 if (ret)
98 return ret;
99
100 ca->key = jhash(key: ca->name, length: sizeof(ca->name), strlen(ca->name));
101
102 spin_lock(lock: &tcp_cong_list_lock);
103 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(key: ca->key)) {
104 pr_notice("%s already registered or non-unique key\n",
105 ca->name);
106 ret = -EEXIST;
107 } else {
108 list_add_tail_rcu(new: &ca->list, head: &tcp_cong_list);
109 pr_debug("%s registered\n", ca->name);
110 }
111 spin_unlock(lock: &tcp_cong_list_lock);
112
113 return ret;
114}
115EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
116
117/*
118 * Remove congestion control algorithm, called from
119 * the module's remove function. Module ref counts are used
120 * to ensure that this can't be done till all sockets using
121 * that method are closed.
122 */
123void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
124{
125 spin_lock(lock: &tcp_cong_list_lock);
126 list_del_rcu(entry: &ca->list);
127 spin_unlock(lock: &tcp_cong_list_lock);
128
129 /* Wait for outstanding readers to complete before the
130 * module gets removed entirely.
131 *
132 * A try_module_get() should fail by now as our module is
133 * in "going" state since no refs are held anymore and
134 * module_exit() handler being called.
135 */
136 synchronize_rcu();
137}
138EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
139
140/* Replace a registered old ca with a new one.
141 *
142 * The new ca must have the same name as the old one, that has been
143 * registered.
144 */
145int tcp_update_congestion_control(struct tcp_congestion_ops *ca, struct tcp_congestion_ops *old_ca)
146{
147 struct tcp_congestion_ops *existing;
148 int ret = 0;
149
150 ca->key = jhash(key: ca->name, length: sizeof(ca->name), strlen(ca->name));
151
152 spin_lock(lock: &tcp_cong_list_lock);
153 existing = tcp_ca_find_key(key: old_ca->key);
154 if (ca->key == TCP_CA_UNSPEC || !existing || strcmp(existing->name, ca->name)) {
155 pr_notice("%s not registered or non-unique key\n",
156 ca->name);
157 ret = -EINVAL;
158 } else if (existing != old_ca) {
159 pr_notice("invalid old congestion control algorithm to replace\n");
160 ret = -EINVAL;
161 } else {
162 /* Add the new one before removing the old one to keep
163 * one implementation available all the time.
164 */
165 list_add_tail_rcu(new: &ca->list, head: &tcp_cong_list);
166 list_del_rcu(entry: &existing->list);
167 pr_debug("%s updated\n", ca->name);
168 }
169 spin_unlock(lock: &tcp_cong_list_lock);
170
171 /* Wait for outstanding readers to complete before the
172 * module or struct_ops gets removed entirely.
173 */
174 if (!ret)
175 synchronize_rcu();
176
177 return ret;
178}
179
180u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
181{
182 const struct tcp_congestion_ops *ca;
183 u32 key = TCP_CA_UNSPEC;
184
185 might_sleep();
186
187 rcu_read_lock();
188 ca = tcp_ca_find_autoload(name);
189 if (ca) {
190 key = ca->key;
191 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
192 }
193 rcu_read_unlock();
194
195 return key;
196}
197
198char *tcp_ca_get_name_by_key(u32 key, char *buffer)
199{
200 const struct tcp_congestion_ops *ca;
201 char *ret = NULL;
202
203 rcu_read_lock();
204 ca = tcp_ca_find_key(key);
205 if (ca) {
206 strscpy(buffer, ca->name, TCP_CA_NAME_MAX);
207 ret = buffer;
208 }
209 rcu_read_unlock();
210
211 return ret;
212}
213
214/* Assign choice of congestion control. */
215void tcp_assign_congestion_control(struct sock *sk)
216{
217 struct net *net = sock_net(sk);
218 struct inet_connection_sock *icsk = inet_csk(sk);
219 const struct tcp_congestion_ops *ca;
220
221 rcu_read_lock();
222 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
223 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
224 ca = &tcp_reno;
225 icsk->icsk_ca_ops = ca;
226 rcu_read_unlock();
227
228 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
229 if (ca->flags & TCP_CONG_NEEDS_ECN)
230 INET_ECN_xmit(sk);
231 else
232 INET_ECN_dontxmit(sk);
233}
234
235void tcp_init_congestion_control(struct sock *sk)
236{
237 struct inet_connection_sock *icsk = inet_csk(sk);
238
239 tcp_sk(sk)->prior_ssthresh = 0;
240 if (icsk->icsk_ca_ops->init)
241 icsk->icsk_ca_ops->init(sk);
242 if (tcp_ca_needs_ecn(sk))
243 INET_ECN_xmit(sk);
244 else
245 INET_ECN_dontxmit(sk);
246 icsk->icsk_ca_initialized = 1;
247}
248
249static void tcp_reinit_congestion_control(struct sock *sk,
250 const struct tcp_congestion_ops *ca)
251{
252 struct inet_connection_sock *icsk = inet_csk(sk);
253
254 tcp_cleanup_congestion_control(sk);
255 icsk->icsk_ca_ops = ca;
256 icsk->icsk_ca_setsockopt = 1;
257 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
258
259 if (ca->flags & TCP_CONG_NEEDS_ECN)
260 INET_ECN_xmit(sk);
261 else
262 INET_ECN_dontxmit(sk);
263
264 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
265 tcp_init_congestion_control(sk);
266}
267
268/* Manage refcounts on socket close. */
269void tcp_cleanup_congestion_control(struct sock *sk)
270{
271 struct inet_connection_sock *icsk = inet_csk(sk);
272
273 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
274 icsk->icsk_ca_ops->release(sk);
275 icsk->icsk_ca_initialized = 0;
276 bpf_module_put(data: icsk->icsk_ca_ops, owner: icsk->icsk_ca_ops->owner);
277}
278
279/* Used by sysctl to change default congestion control */
280int tcp_set_default_congestion_control(struct net *net, const char *name)
281{
282 struct tcp_congestion_ops *ca;
283 const struct tcp_congestion_ops *prev;
284 int ret;
285
286 rcu_read_lock();
287 ca = tcp_ca_find_autoload(name);
288 if (!ca) {
289 ret = -ENOENT;
290 } else if (!bpf_try_module_get(data: ca, owner: ca->owner)) {
291 ret = -EBUSY;
292 } else if (!net_eq(net1: net, net2: &init_net) &&
293 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
294 /* Only init netns can set default to a restricted algorithm */
295 ret = -EPERM;
296 } else {
297 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
298 if (prev)
299 bpf_module_put(data: prev, owner: prev->owner);
300
301 ca->flags |= TCP_CONG_NON_RESTRICTED;
302 ret = 0;
303 }
304 rcu_read_unlock();
305
306 return ret;
307}
308
309/* Set default value from kernel configuration at bootup */
310static int __init tcp_congestion_default(void)
311{
312 return tcp_set_default_congestion_control(net: &init_net,
313 CONFIG_DEFAULT_TCP_CONG);
314}
315late_initcall(tcp_congestion_default);
316
317/* Build string with list of available congestion control values */
318void tcp_get_available_congestion_control(char *buf, size_t maxlen)
319{
320 struct tcp_congestion_ops *ca;
321 size_t offs = 0;
322
323 rcu_read_lock();
324 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
325 offs += snprintf(buf: buf + offs, size: maxlen - offs,
326 fmt: "%s%s",
327 offs == 0 ? "" : " ", ca->name);
328
329 if (WARN_ON_ONCE(offs >= maxlen))
330 break;
331 }
332 rcu_read_unlock();
333}
334
335/* Get current default congestion control */
336void tcp_get_default_congestion_control(struct net *net, char *name)
337{
338 const struct tcp_congestion_ops *ca;
339
340 rcu_read_lock();
341 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
342 strscpy(name, ca->name, TCP_CA_NAME_MAX);
343 rcu_read_unlock();
344}
345
346/* Built list of non-restricted congestion control values */
347void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
348{
349 struct tcp_congestion_ops *ca;
350 size_t offs = 0;
351
352 *buf = '\0';
353 rcu_read_lock();
354 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
355 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
356 continue;
357 offs += snprintf(buf: buf + offs, size: maxlen - offs,
358 fmt: "%s%s",
359 offs == 0 ? "" : " ", ca->name);
360
361 if (WARN_ON_ONCE(offs >= maxlen))
362 break;
363 }
364 rcu_read_unlock();
365}
366
367/* Change list of non-restricted congestion control */
368int tcp_set_allowed_congestion_control(char *val)
369{
370 struct tcp_congestion_ops *ca;
371 char *saved_clone, *clone, *name;
372 int ret = 0;
373
374 saved_clone = clone = kstrdup(s: val, GFP_USER);
375 if (!clone)
376 return -ENOMEM;
377
378 spin_lock(lock: &tcp_cong_list_lock);
379 /* pass 1 check for bad entries */
380 while ((name = strsep(&clone, " ")) && *name) {
381 ca = tcp_ca_find(name);
382 if (!ca) {
383 ret = -ENOENT;
384 goto out;
385 }
386 }
387
388 /* pass 2 clear old values */
389 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
390 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
391
392 /* pass 3 mark as allowed */
393 while ((name = strsep(&val, " ")) && *name) {
394 ca = tcp_ca_find(name);
395 WARN_ON(!ca);
396 if (ca)
397 ca->flags |= TCP_CONG_NON_RESTRICTED;
398 }
399out:
400 spin_unlock(lock: &tcp_cong_list_lock);
401 kfree(objp: saved_clone);
402
403 return ret;
404}
405
406/* Change congestion control for socket. If load is false, then it is the
407 * responsibility of the caller to call tcp_init_congestion_control or
408 * tcp_reinit_congestion_control (if the current congestion control was
409 * already initialized.
410 */
411int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
412 bool cap_net_admin)
413{
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 const struct tcp_congestion_ops *ca;
416 int err = 0;
417
418 if (icsk->icsk_ca_dst_locked)
419 return -EPERM;
420
421 rcu_read_lock();
422 if (!load)
423 ca = tcp_ca_find(name);
424 else
425 ca = tcp_ca_find_autoload(name);
426
427 /* No change asking for existing value */
428 if (ca == icsk->icsk_ca_ops) {
429 icsk->icsk_ca_setsockopt = 1;
430 goto out;
431 }
432
433 if (!ca)
434 err = -ENOENT;
435 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
436 err = -EPERM;
437 else if (!bpf_try_module_get(data: ca, owner: ca->owner))
438 err = -EBUSY;
439 else
440 tcp_reinit_congestion_control(sk, ca);
441 out:
442 rcu_read_unlock();
443 return err;
444}
445
446/* Slow start is used when congestion window is no greater than the slow start
447 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
448 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
449 * something better;) a packet is only considered (s)acked in its entirety to
450 * defend the ACK attacks described in the RFC. Slow start processes a stretch
451 * ACK of degree N as if N acks of degree 1 are received back to back except
452 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
453 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
454 */
455__bpf_kfunc u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
456{
457 u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
458
459 acked -= cwnd - tcp_snd_cwnd(tp);
460 tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
461
462 return acked;
463}
464EXPORT_SYMBOL_GPL(tcp_slow_start);
465
466/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
467 * for every packet that was ACKed.
468 */
469__bpf_kfunc void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
470{
471 /* If credits accumulated at a higher w, apply them gently now. */
472 if (tp->snd_cwnd_cnt >= w) {
473 tp->snd_cwnd_cnt = 0;
474 tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) + 1);
475 }
476
477 tp->snd_cwnd_cnt += acked;
478 if (tp->snd_cwnd_cnt >= w) {
479 u32 delta = tp->snd_cwnd_cnt / w;
480
481 tp->snd_cwnd_cnt -= delta * w;
482 tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) + delta);
483 }
484 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
485}
486EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
487
488/*
489 * TCP Reno congestion control
490 * This is special case used for fallback as well.
491 */
492/* This is Jacobson's slow start and congestion avoidance.
493 * SIGCOMM '88, p. 328.
494 */
495__bpf_kfunc void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
496{
497 struct tcp_sock *tp = tcp_sk(sk);
498
499 if (!tcp_is_cwnd_limited(sk))
500 return;
501
502 /* In "safe" area, increase. */
503 if (tcp_in_slow_start(tp)) {
504 acked = tcp_slow_start(tp, acked);
505 if (!acked)
506 return;
507 }
508 /* In dangerous area, increase slowly. */
509 tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
510}
511EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
512
513/* Slow start threshold is half the congestion window (min 2) */
514__bpf_kfunc u32 tcp_reno_ssthresh(struct sock *sk)
515{
516 const struct tcp_sock *tp = tcp_sk(sk);
517
518 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
519}
520EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
521
522__bpf_kfunc u32 tcp_reno_undo_cwnd(struct sock *sk)
523{
524 const struct tcp_sock *tp = tcp_sk(sk);
525
526 return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
527}
528EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
529
530struct tcp_congestion_ops tcp_reno = {
531 .flags = TCP_CONG_NON_RESTRICTED,
532 .name = "reno",
533 .owner = THIS_MODULE,
534 .ssthresh = tcp_reno_ssthresh,
535 .cong_avoid = tcp_reno_cong_avoid,
536 .undo_cwnd = tcp_reno_undo_cwnd,
537};
538

source code of linux/net/ipv4/tcp_cong.c