Line data Source code
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 1 : struct tcp_congestion_ops *tcp_ca_find(const char *name)
25 : {
26 1 : struct tcp_congestion_ops *e;
27 :
28 2 : list_for_each_entry_rcu(e, &tcp_cong_list, list) {
29 2 : if (strcmp(e->name, name) == 0)
30 1 : return e;
31 : }
32 :
33 : return NULL;
34 : }
35 :
36 : /* Must be called with rcu lock held */
37 1 : static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
38 : const char *name)
39 : {
40 2 : 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 0 : return ca;
51 : }
52 :
53 : /* Simple linear search, not much in here. */
54 2 : struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
55 : {
56 2 : struct tcp_congestion_ops *e;
57 :
58 3 : list_for_each_entry_rcu(e, &tcp_cong_list, list) {
59 1 : if (e->key == key)
60 0 : 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 2 : int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
71 : {
72 2 : int ret = 0;
73 :
74 : /* all algorithms must implement these */
75 2 : if (!ca->ssthresh || !ca->undo_cwnd ||
76 2 : !(ca->cong_avoid || ca->cong_control)) {
77 0 : pr_err("%s does not implement required ops\n", ca->name);
78 0 : return -EINVAL;
79 : }
80 :
81 2 : ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
82 :
83 2 : spin_lock(&tcp_cong_list_lock);
84 4 : if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
85 0 : pr_notice("%s already registered or non-unique key\n",
86 : ca->name);
87 0 : ret = -EEXIST;
88 : } else {
89 2 : list_add_tail_rcu(&ca->list, &tcp_cong_list);
90 2 : pr_debug("%s registered\n", ca->name);
91 : }
92 2 : spin_unlock(&tcp_cong_list_lock);
93 :
94 2 : 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 0 : void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
105 : {
106 0 : spin_lock(&tcp_cong_list_lock);
107 0 : list_del_rcu(&ca->list);
108 0 : 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 0 : synchronize_rcu();
118 0 : }
119 : EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120 :
121 0 : u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
122 : {
123 0 : const struct tcp_congestion_ops *ca;
124 0 : u32 key = TCP_CA_UNSPEC;
125 :
126 0 : might_sleep();
127 :
128 0 : rcu_read_lock();
129 0 : ca = tcp_ca_find_autoload(net, name);
130 0 : if (ca) {
131 0 : key = ca->key;
132 0 : *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
133 : }
134 0 : rcu_read_unlock();
135 :
136 0 : return key;
137 : }
138 : EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139 :
140 0 : char *tcp_ca_get_name_by_key(u32 key, char *buffer)
141 : {
142 0 : const struct tcp_congestion_ops *ca;
143 0 : char *ret = NULL;
144 :
145 0 : rcu_read_lock();
146 0 : ca = tcp_ca_find_key(key);
147 0 : if (ca)
148 0 : ret = strncpy(buffer, ca->name,
149 : TCP_CA_NAME_MAX);
150 0 : rcu_read_unlock();
151 :
152 0 : return ret;
153 : }
154 : EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155 :
156 : /* Assign choice of congestion control. */
157 7 : void tcp_assign_congestion_control(struct sock *sk)
158 : {
159 7 : struct net *net = sock_net(sk);
160 7 : struct inet_connection_sock *icsk = inet_csk(sk);
161 7 : const struct tcp_congestion_ops *ca;
162 :
163 7 : rcu_read_lock();
164 7 : ca = rcu_dereference(net->ipv4.tcp_congestion_control);
165 7 : if (unlikely(!bpf_try_module_get(ca, ca->owner)))
166 : ca = &tcp_reno;
167 7 : icsk->icsk_ca_ops = ca;
168 7 : rcu_read_unlock();
169 :
170 7 : memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
171 7 : if (ca->flags & TCP_CONG_NEEDS_ECN)
172 0 : INET_ECN_xmit(sk);
173 : else
174 7 : INET_ECN_dontxmit(sk);
175 7 : }
176 :
177 4 : void tcp_init_congestion_control(struct sock *sk)
178 : {
179 4 : struct inet_connection_sock *icsk = inet_csk(sk);
180 :
181 4 : tcp_sk(sk)->prior_ssthresh = 0;
182 4 : if (icsk->icsk_ca_ops->init)
183 4 : icsk->icsk_ca_ops->init(sk);
184 4 : if (tcp_ca_needs_ecn(sk))
185 0 : INET_ECN_xmit(sk);
186 : else
187 4 : INET_ECN_dontxmit(sk);
188 4 : icsk->icsk_ca_initialized = 1;
189 4 : }
190 :
191 0 : static void tcp_reinit_congestion_control(struct sock *sk,
192 : const struct tcp_congestion_ops *ca)
193 : {
194 0 : struct inet_connection_sock *icsk = inet_csk(sk);
195 :
196 0 : tcp_cleanup_congestion_control(sk);
197 0 : icsk->icsk_ca_ops = ca;
198 0 : icsk->icsk_ca_setsockopt = 1;
199 0 : memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
200 :
201 0 : if (ca->flags & TCP_CONG_NEEDS_ECN)
202 0 : INET_ECN_xmit(sk);
203 : else
204 0 : INET_ECN_dontxmit(sk);
205 :
206 0 : if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
207 0 : tcp_init_congestion_control(sk);
208 0 : }
209 :
210 : /* Manage refcounts on socket close. */
211 4 : void tcp_cleanup_congestion_control(struct sock *sk)
212 : {
213 4 : struct inet_connection_sock *icsk = inet_csk(sk);
214 :
215 4 : if (icsk->icsk_ca_ops->release)
216 0 : icsk->icsk_ca_ops->release(sk);
217 0 : bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
218 4 : }
219 :
220 : /* Used by sysctl to change default congestion control */
221 1 : int tcp_set_default_congestion_control(struct net *net, const char *name)
222 : {
223 1 : struct tcp_congestion_ops *ca;
224 1 : const struct tcp_congestion_ops *prev;
225 1 : int ret;
226 :
227 1 : rcu_read_lock();
228 1 : ca = tcp_ca_find_autoload(net, name);
229 1 : if (!ca) {
230 : ret = -ENOENT;
231 1 : } else if (!bpf_try_module_get(ca, ca->owner)) {
232 : ret = -EBUSY;
233 : } else {
234 1 : prev = xchg(&net->ipv4.tcp_congestion_control, ca);
235 1 : if (prev)
236 1 : bpf_module_put(prev, prev->owner);
237 :
238 1 : ca->flags |= TCP_CONG_NON_RESTRICTED;
239 1 : ret = 0;
240 : }
241 1 : rcu_read_unlock();
242 :
243 1 : return ret;
244 : }
245 :
246 : /* Set default value from kernel configuration at bootup */
247 1 : static int __init tcp_congestion_default(void)
248 : {
249 1 : return tcp_set_default_congestion_control(&init_net,
250 : CONFIG_DEFAULT_TCP_CONG);
251 : }
252 : late_initcall(tcp_congestion_default);
253 :
254 : /* Build string with list of available congestion control values */
255 0 : void tcp_get_available_congestion_control(char *buf, size_t maxlen)
256 : {
257 0 : struct tcp_congestion_ops *ca;
258 0 : size_t offs = 0;
259 :
260 0 : rcu_read_lock();
261 0 : list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
262 0 : offs += snprintf(buf + offs, maxlen - offs,
263 : "%s%s",
264 0 : offs == 0 ? "" : " ", ca->name);
265 :
266 0 : if (WARN_ON_ONCE(offs >= maxlen))
267 : break;
268 : }
269 0 : rcu_read_unlock();
270 0 : }
271 :
272 : /* Get current default congestion control */
273 0 : void tcp_get_default_congestion_control(struct net *net, char *name)
274 : {
275 0 : const struct tcp_congestion_ops *ca;
276 :
277 0 : rcu_read_lock();
278 0 : ca = rcu_dereference(net->ipv4.tcp_congestion_control);
279 0 : strncpy(name, ca->name, TCP_CA_NAME_MAX);
280 0 : rcu_read_unlock();
281 0 : }
282 :
283 : /* Built list of non-restricted congestion control values */
284 0 : void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
285 : {
286 0 : struct tcp_congestion_ops *ca;
287 0 : size_t offs = 0;
288 :
289 0 : *buf = '\0';
290 0 : rcu_read_lock();
291 0 : list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
292 0 : if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
293 0 : continue;
294 0 : offs += snprintf(buf + offs, maxlen - offs,
295 : "%s%s",
296 0 : offs == 0 ? "" : " ", ca->name);
297 :
298 0 : if (WARN_ON_ONCE(offs >= maxlen))
299 : break;
300 : }
301 0 : rcu_read_unlock();
302 0 : }
303 :
304 : /* Change list of non-restricted congestion control */
305 0 : int tcp_set_allowed_congestion_control(char *val)
306 : {
307 0 : struct tcp_congestion_ops *ca;
308 0 : char *saved_clone, *clone, *name;
309 0 : int ret = 0;
310 :
311 0 : saved_clone = clone = kstrdup(val, GFP_USER);
312 0 : if (!clone)
313 : return -ENOMEM;
314 :
315 0 : spin_lock(&tcp_cong_list_lock);
316 : /* pass 1 check for bad entries */
317 0 : while ((name = strsep(&clone, " ")) && *name) {
318 0 : ca = tcp_ca_find(name);
319 0 : if (!ca) {
320 0 : ret = -ENOENT;
321 0 : goto out;
322 : }
323 : }
324 :
325 : /* pass 2 clear old values */
326 0 : list_for_each_entry_rcu(ca, &tcp_cong_list, list)
327 0 : ca->flags &= ~TCP_CONG_NON_RESTRICTED;
328 :
329 : /* pass 3 mark as allowed */
330 0 : while ((name = strsep(&val, " ")) && *name) {
331 0 : ca = tcp_ca_find(name);
332 0 : WARN_ON(!ca);
333 0 : if (ca)
334 0 : ca->flags |= TCP_CONG_NON_RESTRICTED;
335 : }
336 0 : out:
337 0 : spin_unlock(&tcp_cong_list_lock);
338 0 : kfree(saved_clone);
339 :
340 0 : return ret;
341 : }
342 :
343 : /* Change congestion control for socket. If load is false, then it is the
344 : * responsibility of the caller to call tcp_init_congestion_control or
345 : * tcp_reinit_congestion_control (if the current congestion control was
346 : * already initialized.
347 : */
348 0 : int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
349 : bool cap_net_admin)
350 : {
351 0 : struct inet_connection_sock *icsk = inet_csk(sk);
352 0 : const struct tcp_congestion_ops *ca;
353 0 : int err = 0;
354 :
355 0 : if (icsk->icsk_ca_dst_locked)
356 : return -EPERM;
357 :
358 0 : rcu_read_lock();
359 0 : if (!load)
360 0 : ca = tcp_ca_find(name);
361 : else
362 0 : ca = tcp_ca_find_autoload(sock_net(sk), name);
363 :
364 : /* No change asking for existing value */
365 0 : if (ca == icsk->icsk_ca_ops) {
366 0 : icsk->icsk_ca_setsockopt = 1;
367 0 : goto out;
368 : }
369 :
370 0 : if (!ca)
371 : err = -ENOENT;
372 0 : else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
373 : err = -EPERM;
374 0 : else if (!bpf_try_module_get(ca, ca->owner))
375 : err = -EBUSY;
376 : else
377 0 : tcp_reinit_congestion_control(sk, ca);
378 0 : out:
379 0 : rcu_read_unlock();
380 0 : return err;
381 : }
382 :
383 : /* Slow start is used when congestion window is no greater than the slow start
384 : * threshold. We base on RFC2581 and also handle stretch ACKs properly.
385 : * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
386 : * something better;) a packet is only considered (s)acked in its entirety to
387 : * defend the ACK attacks described in the RFC. Slow start processes a stretch
388 : * ACK of degree N as if N acks of degree 1 are received back to back except
389 : * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
390 : * returns the leftover acks to adjust cwnd in congestion avoidance mode.
391 : */
392 0 : u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
393 : {
394 0 : u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
395 :
396 0 : acked -= cwnd - tp->snd_cwnd;
397 0 : tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
398 :
399 0 : return acked;
400 : }
401 : EXPORT_SYMBOL_GPL(tcp_slow_start);
402 :
403 : /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
404 : * for every packet that was ACKed.
405 : */
406 0 : void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
407 : {
408 : /* If credits accumulated at a higher w, apply them gently now. */
409 0 : if (tp->snd_cwnd_cnt >= w) {
410 0 : tp->snd_cwnd_cnt = 0;
411 0 : tp->snd_cwnd++;
412 : }
413 :
414 0 : tp->snd_cwnd_cnt += acked;
415 0 : if (tp->snd_cwnd_cnt >= w) {
416 0 : u32 delta = tp->snd_cwnd_cnt / w;
417 :
418 0 : tp->snd_cwnd_cnt -= delta * w;
419 0 : tp->snd_cwnd += delta;
420 : }
421 0 : tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
422 0 : }
423 : EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
424 :
425 : /*
426 : * TCP Reno congestion control
427 : * This is special case used for fallback as well.
428 : */
429 : /* This is Jacobson's slow start and congestion avoidance.
430 : * SIGCOMM '88, p. 328.
431 : */
432 0 : void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
433 : {
434 0 : struct tcp_sock *tp = tcp_sk(sk);
435 :
436 0 : if (!tcp_is_cwnd_limited(sk))
437 : return;
438 :
439 : /* In "safe" area, increase. */
440 0 : if (tcp_in_slow_start(tp)) {
441 0 : acked = tcp_slow_start(tp, acked);
442 0 : if (!acked)
443 : return;
444 : }
445 : /* In dangerous area, increase slowly. */
446 0 : tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
447 : }
448 : EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
449 :
450 : /* Slow start threshold is half the congestion window (min 2) */
451 0 : u32 tcp_reno_ssthresh(struct sock *sk)
452 : {
453 0 : const struct tcp_sock *tp = tcp_sk(sk);
454 :
455 0 : return max(tp->snd_cwnd >> 1U, 2U);
456 : }
457 : EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
458 :
459 0 : u32 tcp_reno_undo_cwnd(struct sock *sk)
460 : {
461 0 : const struct tcp_sock *tp = tcp_sk(sk);
462 :
463 0 : return max(tp->snd_cwnd, tp->prior_cwnd);
464 : }
465 : EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
466 :
467 : struct tcp_congestion_ops tcp_reno = {
468 : .flags = TCP_CONG_NON_RESTRICTED,
469 : .name = "reno",
470 : .owner = THIS_MODULE,
471 : .ssthresh = tcp_reno_ssthresh,
472 : .cong_avoid = tcp_reno_cong_avoid,
473 : .undo_cwnd = tcp_reno_undo_cwnd,
474 : };
|
