1
2
3
4
5
6
7
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
20static DEFINE_SPINLOCK(tcp_cong_list_lock);
21static LIST_HEAD(tcp_cong_list);
22
23
24struct 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
37static 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
54struct 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
68
69
70int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
71{
72 int ret = 0;
73
74
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}
96EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
97
98
99
100
101
102
103
104void 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
111
112
113
114
115
116
117 synchronize_rcu();
118}
119EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120
121u32 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}
138EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139
140char *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}
154EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155
156
157void 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
177void 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
191static 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
211void 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
221int 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 {
234 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
235 if (prev)
236 bpf_module_put(prev, prev->owner);
237
238 ca->flags |= TCP_CONG_NON_RESTRICTED;
239 ret = 0;
240 }
241 rcu_read_unlock();
242
243 return ret;
244}
245
246
247static int __init tcp_congestion_default(void)
248{
249 return tcp_set_default_congestion_control(&init_net,
250 CONFIG_DEFAULT_TCP_CONG);
251}
252late_initcall(tcp_congestion_default);
253
254
255void tcp_get_available_congestion_control(char *buf, size_t maxlen)
256{
257 struct tcp_congestion_ops *ca;
258 size_t offs = 0;
259
260 rcu_read_lock();
261 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
262 offs += snprintf(buf + offs, maxlen - offs,
263 "%s%s",
264 offs == 0 ? "" : " ", ca->name);
265
266 if (WARN_ON_ONCE(offs >= maxlen))
267 break;
268 }
269 rcu_read_unlock();
270}
271
272
273void tcp_get_default_congestion_control(struct net *net, char *name)
274{
275 const struct tcp_congestion_ops *ca;
276
277 rcu_read_lock();
278 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
279 strncpy(name, ca->name, TCP_CA_NAME_MAX);
280 rcu_read_unlock();
281}
282
283
284void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
285{
286 struct tcp_congestion_ops *ca;
287 size_t offs = 0;
288
289 *buf = '\0';
290 rcu_read_lock();
291 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
292 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
293 continue;
294 offs += snprintf(buf + offs, maxlen - offs,
295 "%s%s",
296 offs == 0 ? "" : " ", ca->name);
297
298 if (WARN_ON_ONCE(offs >= maxlen))
299 break;
300 }
301 rcu_read_unlock();
302}
303
304
305int tcp_set_allowed_congestion_control(char *val)
306{
307 struct tcp_congestion_ops *ca;
308 char *saved_clone, *clone, *name;
309 int ret = 0;
310
311 saved_clone = clone = kstrdup(val, GFP_USER);
312 if (!clone)
313 return -ENOMEM;
314
315 spin_lock(&tcp_cong_list_lock);
316
317 while ((name = strsep(&clone, " ")) && *name) {
318 ca = tcp_ca_find(name);
319 if (!ca) {
320 ret = -ENOENT;
321 goto out;
322 }
323 }
324
325
326 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
327 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
328
329
330 while ((name = strsep(&val, " ")) && *name) {
331 ca = tcp_ca_find(name);
332 WARN_ON(!ca);
333 if (ca)
334 ca->flags |= TCP_CONG_NON_RESTRICTED;
335 }
336out:
337 spin_unlock(&tcp_cong_list_lock);
338 kfree(saved_clone);
339
340 return ret;
341}
342
343
344
345
346
347
348int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
349 bool cap_net_admin)
350{
351 struct inet_connection_sock *icsk = inet_csk(sk);
352 const struct tcp_congestion_ops *ca;
353 int err = 0;
354
355 if (icsk->icsk_ca_dst_locked)
356 return -EPERM;
357
358 rcu_read_lock();
359 if (!load)
360 ca = tcp_ca_find(name);
361 else
362 ca = tcp_ca_find_autoload(sock_net(sk), name);
363
364
365 if (ca == icsk->icsk_ca_ops) {
366 icsk->icsk_ca_setsockopt = 1;
367 goto out;
368 }
369
370 if (!ca)
371 err = -ENOENT;
372 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
373 err = -EPERM;
374 else if (!bpf_try_module_get(ca, ca->owner))
375 err = -EBUSY;
376 else
377 tcp_reinit_congestion_control(sk, ca);
378 out:
379 rcu_read_unlock();
380 return err;
381}
382
383
384
385
386
387
388
389
390
391
392u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
393{
394 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
395
396 acked -= cwnd - tp->snd_cwnd;
397 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
398
399 return acked;
400}
401EXPORT_SYMBOL_GPL(tcp_slow_start);
402
403
404
405
406void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
407{
408
409 if (tp->snd_cwnd_cnt >= w) {
410 tp->snd_cwnd_cnt = 0;
411 tp->snd_cwnd++;
412 }
413
414 tp->snd_cwnd_cnt += acked;
415 if (tp->snd_cwnd_cnt >= w) {
416 u32 delta = tp->snd_cwnd_cnt / w;
417
418 tp->snd_cwnd_cnt -= delta * w;
419 tp->snd_cwnd += delta;
420 }
421 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
422}
423EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
424
425
426
427
428
429
430
431
432void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
433{
434 struct tcp_sock *tp = tcp_sk(sk);
435
436 if (!tcp_is_cwnd_limited(sk))
437 return;
438
439
440 if (tcp_in_slow_start(tp)) {
441 acked = tcp_slow_start(tp, acked);
442 if (!acked)
443 return;
444 }
445
446 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
447}
448EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
449
450
451u32 tcp_reno_ssthresh(struct sock *sk)
452{
453 const struct tcp_sock *tp = tcp_sk(sk);
454
455 return max(tp->snd_cwnd >> 1U, 2U);
456}
457EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
458
459u32 tcp_reno_undo_cwnd(struct sock *sk)
460{
461 const struct tcp_sock *tp = tcp_sk(sk);
462
463 return max(tp->snd_cwnd, tp->prior_cwnd);
464}
465EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
466
467struct 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};
475