]>
Commit | Line | Data |
---|---|---|
317a76f9 | 1 | /* |
b92022f3 | 2 | * Pluggable TCP congestion control support and newReno |
317a76f9 | 3 | * congestion control. |
02582e9b | 4 | * Based on ideas from I/O scheduler support and Web100. |
317a76f9 SH |
5 | * |
6 | * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org> | |
7 | */ | |
8 | ||
afd46503 JP |
9 | #define pr_fmt(fmt) "TCP: " fmt |
10 | ||
317a76f9 SH |
11 | #include <linux/module.h> |
12 | #include <linux/mm.h> | |
13 | #include <linux/types.h> | |
14 | #include <linux/list.h> | |
5a0e3ad6 | 15 | #include <linux/gfp.h> |
c5c6a8ab | 16 | #include <linux/jhash.h> |
317a76f9 SH |
17 | #include <net/tcp.h> |
18 | ||
19 | static DEFINE_SPINLOCK(tcp_cong_list_lock); | |
20 | static LIST_HEAD(tcp_cong_list); | |
21 | ||
22 | /* Simple linear search, don't expect many entries! */ | |
23 | static struct tcp_congestion_ops *tcp_ca_find(const char *name) | |
24 | { | |
25 | struct tcp_congestion_ops *e; | |
26 | ||
5f8ef48d | 27 | list_for_each_entry_rcu(e, &tcp_cong_list, list) { |
317a76f9 SH |
28 | if (strcmp(e->name, name) == 0) |
29 | return e; | |
30 | } | |
31 | ||
32 | return NULL; | |
33 | } | |
34 | ||
c5c6a8ab DB |
35 | /* Must be called with rcu lock held */ |
36 | static const struct tcp_congestion_ops *__tcp_ca_find_autoload(const char *name) | |
37 | { | |
38 | const struct tcp_congestion_ops *ca = tcp_ca_find(name); | |
39 | #ifdef CONFIG_MODULES | |
40 | if (!ca && capable(CAP_NET_ADMIN)) { | |
41 | rcu_read_unlock(); | |
42 | request_module("tcp_%s", name); | |
43 | rcu_read_lock(); | |
44 | ca = tcp_ca_find(name); | |
45 | } | |
46 | #endif | |
47 | return ca; | |
48 | } | |
49 | ||
50 | /* Simple linear search, not much in here. */ | |
51 | struct tcp_congestion_ops *tcp_ca_find_key(u32 key) | |
52 | { | |
53 | struct tcp_congestion_ops *e; | |
54 | ||
55 | list_for_each_entry_rcu(e, &tcp_cong_list, list) { | |
56 | if (e->key == key) | |
57 | return e; | |
58 | } | |
59 | ||
60 | return NULL; | |
61 | } | |
62 | ||
317a76f9 | 63 | /* |
d08df601 | 64 | * Attach new congestion control algorithm to the list |
317a76f9 SH |
65 | * of available options. |
66 | */ | |
67 | int tcp_register_congestion_control(struct tcp_congestion_ops *ca) | |
68 | { | |
69 | int ret = 0; | |
70 | ||
e9799183 FW |
71 | /* all algorithms must implement these */ |
72 | if (!ca->ssthresh || !ca->undo_cwnd || | |
73 | !(ca->cong_avoid || ca->cong_control)) { | |
afd46503 | 74 | pr_err("%s does not implement required ops\n", ca->name); |
317a76f9 SH |
75 | return -EINVAL; |
76 | } | |
77 | ||
c5c6a8ab DB |
78 | ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name)); |
79 | ||
317a76f9 | 80 | spin_lock(&tcp_cong_list_lock); |
c5c6a8ab DB |
81 | if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) { |
82 | pr_notice("%s already registered or non-unique key\n", | |
83 | ca->name); | |
317a76f9 SH |
84 | ret = -EEXIST; |
85 | } else { | |
3d2573f7 | 86 | list_add_tail_rcu(&ca->list, &tcp_cong_list); |
db2855ae | 87 | pr_debug("%s registered\n", ca->name); |
317a76f9 SH |
88 | } |
89 | spin_unlock(&tcp_cong_list_lock); | |
90 | ||
91 | return ret; | |
92 | } | |
93 | EXPORT_SYMBOL_GPL(tcp_register_congestion_control); | |
94 | ||
95 | /* | |
96 | * Remove congestion control algorithm, called from | |
97 | * the module's remove function. Module ref counts are used | |
98 | * to ensure that this can't be done till all sockets using | |
99 | * that method are closed. | |
100 | */ | |
101 | void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca) | |
102 | { | |
103 | spin_lock(&tcp_cong_list_lock); | |
104 | list_del_rcu(&ca->list); | |
105 | spin_unlock(&tcp_cong_list_lock); | |
c5c6a8ab DB |
106 | |
107 | /* Wait for outstanding readers to complete before the | |
108 | * module gets removed entirely. | |
109 | * | |
110 | * A try_module_get() should fail by now as our module is | |
111 | * in "going" state since no refs are held anymore and | |
112 | * module_exit() handler being called. | |
113 | */ | |
114 | synchronize_rcu(); | |
317a76f9 SH |
115 | } |
116 | EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control); | |
117 | ||
c3a8d947 | 118 | u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca) |
c5c6a8ab DB |
119 | { |
120 | const struct tcp_congestion_ops *ca; | |
c3a8d947 | 121 | u32 key = TCP_CA_UNSPEC; |
c5c6a8ab DB |
122 | |
123 | might_sleep(); | |
124 | ||
125 | rcu_read_lock(); | |
126 | ca = __tcp_ca_find_autoload(name); | |
c3a8d947 DB |
127 | if (ca) { |
128 | key = ca->key; | |
129 | *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN; | |
130 | } | |
c5c6a8ab DB |
131 | rcu_read_unlock(); |
132 | ||
133 | return key; | |
134 | } | |
135 | EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name); | |
136 | ||
137 | char *tcp_ca_get_name_by_key(u32 key, char *buffer) | |
138 | { | |
139 | const struct tcp_congestion_ops *ca; | |
140 | char *ret = NULL; | |
141 | ||
142 | rcu_read_lock(); | |
143 | ca = tcp_ca_find_key(key); | |
144 | if (ca) | |
145 | ret = strncpy(buffer, ca->name, | |
146 | TCP_CA_NAME_MAX); | |
147 | rcu_read_unlock(); | |
148 | ||
149 | return ret; | |
150 | } | |
151 | EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key); | |
152 | ||
317a76f9 | 153 | /* Assign choice of congestion control. */ |
55d8694f | 154 | void tcp_assign_congestion_control(struct sock *sk) |
317a76f9 | 155 | { |
6687e988 | 156 | struct inet_connection_sock *icsk = inet_csk(sk); |
317a76f9 SH |
157 | struct tcp_congestion_ops *ca; |
158 | ||
55d8694f FW |
159 | rcu_read_lock(); |
160 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
161 | if (likely(try_module_get(ca->owner))) { | |
162 | icsk->icsk_ca_ops = ca; | |
163 | goto out; | |
317a76f9 | 164 | } |
55d8694f FW |
165 | /* Fallback to next available. The last really |
166 | * guaranteed fallback is Reno from this list. | |
167 | */ | |
317a76f9 | 168 | } |
55d8694f FW |
169 | out: |
170 | rcu_read_unlock(); | |
c1201444 | 171 | memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); |
55d8694f | 172 | |
6ac705b1 ED |
173 | if (ca->flags & TCP_CONG_NEEDS_ECN) |
174 | INET_ECN_xmit(sk); | |
175 | else | |
176 | INET_ECN_dontxmit(sk); | |
55d8694f FW |
177 | } |
178 | ||
179 | void tcp_init_congestion_control(struct sock *sk) | |
180 | { | |
181 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
317a76f9 | 182 | |
44abafc4 | 183 | tcp_sk(sk)->prior_ssthresh = 0; |
6687e988 ACM |
184 | if (icsk->icsk_ca_ops->init) |
185 | icsk->icsk_ca_ops->init(sk); | |
6ac705b1 ED |
186 | if (tcp_ca_needs_ecn(sk)) |
187 | INET_ECN_xmit(sk); | |
188 | else | |
189 | INET_ECN_dontxmit(sk); | |
317a76f9 SH |
190 | } |
191 | ||
ebfa00c5 SD |
192 | static void tcp_reinit_congestion_control(struct sock *sk, |
193 | const struct tcp_congestion_ops *ca) | |
29ba4fff DB |
194 | { |
195 | struct inet_connection_sock *icsk = inet_csk(sk); | |
196 | ||
197 | tcp_cleanup_congestion_control(sk); | |
198 | icsk->icsk_ca_ops = ca; | |
9f950415 | 199 | icsk->icsk_ca_setsockopt = 1; |
c1201444 | 200 | memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); |
29ba4fff | 201 | |
c1201444 | 202 | if (sk->sk_state != TCP_CLOSE) |
6ac705b1 | 203 | tcp_init_congestion_control(sk); |
29ba4fff DB |
204 | } |
205 | ||
317a76f9 | 206 | /* Manage refcounts on socket close. */ |
6687e988 | 207 | void tcp_cleanup_congestion_control(struct sock *sk) |
317a76f9 | 208 | { |
6687e988 ACM |
209 | struct inet_connection_sock *icsk = inet_csk(sk); |
210 | ||
211 | if (icsk->icsk_ca_ops->release) | |
212 | icsk->icsk_ca_ops->release(sk); | |
213 | module_put(icsk->icsk_ca_ops->owner); | |
317a76f9 SH |
214 | } |
215 | ||
216 | /* Used by sysctl to change default congestion control */ | |
217 | int tcp_set_default_congestion_control(const char *name) | |
218 | { | |
219 | struct tcp_congestion_ops *ca; | |
220 | int ret = -ENOENT; | |
221 | ||
222 | spin_lock(&tcp_cong_list_lock); | |
223 | ca = tcp_ca_find(name); | |
95a5afca | 224 | #ifdef CONFIG_MODULES |
a8f80e8f | 225 | if (!ca && capable(CAP_NET_ADMIN)) { |
317a76f9 SH |
226 | spin_unlock(&tcp_cong_list_lock); |
227 | ||
228 | request_module("tcp_%s", name); | |
229 | spin_lock(&tcp_cong_list_lock); | |
230 | ca = tcp_ca_find(name); | |
231 | } | |
232 | #endif | |
233 | ||
234 | if (ca) { | |
164891aa | 235 | ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */ |
317a76f9 SH |
236 | list_move(&ca->list, &tcp_cong_list); |
237 | ret = 0; | |
238 | } | |
239 | spin_unlock(&tcp_cong_list_lock); | |
240 | ||
241 | return ret; | |
242 | } | |
243 | ||
b1736a71 SH |
244 | /* Set default value from kernel configuration at bootup */ |
245 | static int __init tcp_congestion_default(void) | |
246 | { | |
247 | return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG); | |
248 | } | |
249 | late_initcall(tcp_congestion_default); | |
250 | ||
3ff825b2 SH |
251 | /* Build string with list of available congestion control values */ |
252 | void tcp_get_available_congestion_control(char *buf, size_t maxlen) | |
253 | { | |
254 | struct tcp_congestion_ops *ca; | |
255 | size_t offs = 0; | |
256 | ||
257 | rcu_read_lock(); | |
258 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
259 | offs += snprintf(buf + offs, maxlen - offs, | |
260 | "%s%s", | |
261 | offs == 0 ? "" : " ", ca->name); | |
3ff825b2 SH |
262 | } |
263 | rcu_read_unlock(); | |
264 | } | |
265 | ||
317a76f9 SH |
266 | /* Get current default congestion control */ |
267 | void tcp_get_default_congestion_control(char *name) | |
268 | { | |
269 | struct tcp_congestion_ops *ca; | |
270 | /* We will always have reno... */ | |
271 | BUG_ON(list_empty(&tcp_cong_list)); | |
272 | ||
273 | rcu_read_lock(); | |
274 | ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list); | |
275 | strncpy(name, ca->name, TCP_CA_NAME_MAX); | |
276 | rcu_read_unlock(); | |
277 | } | |
278 | ||
ce7bc3bf SH |
279 | /* Built list of non-restricted congestion control values */ |
280 | void tcp_get_allowed_congestion_control(char *buf, size_t maxlen) | |
281 | { | |
282 | struct tcp_congestion_ops *ca; | |
283 | size_t offs = 0; | |
284 | ||
285 | *buf = '\0'; | |
286 | rcu_read_lock(); | |
287 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
164891aa | 288 | if (!(ca->flags & TCP_CONG_NON_RESTRICTED)) |
ce7bc3bf SH |
289 | continue; |
290 | offs += snprintf(buf + offs, maxlen - offs, | |
291 | "%s%s", | |
292 | offs == 0 ? "" : " ", ca->name); | |
ce7bc3bf SH |
293 | } |
294 | rcu_read_unlock(); | |
295 | } | |
296 | ||
297 | /* Change list of non-restricted congestion control */ | |
298 | int tcp_set_allowed_congestion_control(char *val) | |
299 | { | |
300 | struct tcp_congestion_ops *ca; | |
c34186ed | 301 | char *saved_clone, *clone, *name; |
ce7bc3bf SH |
302 | int ret = 0; |
303 | ||
c34186ed | 304 | saved_clone = clone = kstrdup(val, GFP_USER); |
ce7bc3bf SH |
305 | if (!clone) |
306 | return -ENOMEM; | |
307 | ||
308 | spin_lock(&tcp_cong_list_lock); | |
309 | /* pass 1 check for bad entries */ | |
310 | while ((name = strsep(&clone, " ")) && *name) { | |
311 | ca = tcp_ca_find(name); | |
312 | if (!ca) { | |
313 | ret = -ENOENT; | |
314 | goto out; | |
315 | } | |
316 | } | |
317 | ||
164891aa | 318 | /* pass 2 clear old values */ |
ce7bc3bf | 319 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) |
164891aa | 320 | ca->flags &= ~TCP_CONG_NON_RESTRICTED; |
ce7bc3bf SH |
321 | |
322 | /* pass 3 mark as allowed */ | |
323 | while ((name = strsep(&val, " ")) && *name) { | |
324 | ca = tcp_ca_find(name); | |
325 | WARN_ON(!ca); | |
326 | if (ca) | |
164891aa | 327 | ca->flags |= TCP_CONG_NON_RESTRICTED; |
ce7bc3bf SH |
328 | } |
329 | out: | |
330 | spin_unlock(&tcp_cong_list_lock); | |
c34186ed | 331 | kfree(saved_clone); |
ce7bc3bf SH |
332 | |
333 | return ret; | |
334 | } | |
335 | ||
91b5b21c LB |
336 | /* Change congestion control for socket. If load is false, then it is the |
337 | * responsibility of the caller to call tcp_init_congestion_control or | |
338 | * tcp_reinit_congestion_control (if the current congestion control was | |
339 | * already initialized. | |
340 | */ | |
ebfa00c5 | 341 | int tcp_set_congestion_control(struct sock *sk, const char *name, bool load, bool reinit) |
5f8ef48d | 342 | { |
6687e988 | 343 | struct inet_connection_sock *icsk = inet_csk(sk); |
c5c6a8ab | 344 | const struct tcp_congestion_ops *ca; |
5f8ef48d SH |
345 | int err = 0; |
346 | ||
c5c6a8ab DB |
347 | if (icsk->icsk_ca_dst_locked) |
348 | return -EPERM; | |
4d4d3d1e | 349 | |
c5c6a8ab | 350 | rcu_read_lock(); |
91b5b21c LB |
351 | if (!load) |
352 | ca = tcp_ca_find(name); | |
353 | else | |
354 | ca = __tcp_ca_find_autoload(name); | |
c5c6a8ab | 355 | /* No change asking for existing value */ |
9f950415 NC |
356 | if (ca == icsk->icsk_ca_ops) { |
357 | icsk->icsk_ca_setsockopt = 1; | |
5f8ef48d | 358 | goto out; |
9f950415 | 359 | } |
91b5b21c | 360 | if (!ca) { |
5f8ef48d | 361 | err = -ENOENT; |
91b5b21c | 362 | } else if (!load) { |
ebfa00c5 SD |
363 | const struct tcp_congestion_ops *old_ca = icsk->icsk_ca_ops; |
364 | ||
365 | if (try_module_get(ca->owner)) { | |
366 | if (reinit) { | |
367 | tcp_reinit_congestion_control(sk, ca); | |
368 | } else { | |
369 | icsk->icsk_ca_ops = ca; | |
370 | module_put(old_ca->owner); | |
371 | } | |
372 | } else { | |
91b5b21c | 373 | err = -EBUSY; |
ebfa00c5 | 374 | } |
91b5b21c LB |
375 | } else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || |
376 | ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN))) { | |
ce7bc3bf | 377 | err = -EPERM; |
91b5b21c | 378 | } else if (!try_module_get(ca->owner)) { |
5f8ef48d | 379 | err = -EBUSY; |
91b5b21c | 380 | } else { |
29ba4fff | 381 | tcp_reinit_congestion_control(sk, ca); |
91b5b21c | 382 | } |
5f8ef48d SH |
383 | out: |
384 | rcu_read_unlock(); | |
385 | return err; | |
386 | } | |
387 | ||
9f9843a7 YC |
388 | /* Slow start is used when congestion window is no greater than the slow start |
389 | * threshold. We base on RFC2581 and also handle stretch ACKs properly. | |
390 | * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but | |
391 | * something better;) a packet is only considered (s)acked in its entirety to | |
392 | * defend the ACK attacks described in the RFC. Slow start processes a stretch | |
393 | * ACK of degree N as if N acks of degree 1 are received back to back except | |
394 | * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and | |
395 | * returns the leftover acks to adjust cwnd in congestion avoidance mode. | |
40efc6fa | 396 | */ |
e73ebb08 | 397 | u32 tcp_slow_start(struct tcp_sock *tp, u32 acked) |
40efc6fa | 398 | { |
76174004 | 399 | u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh); |
a02ba041 | 400 | |
e73ebb08 | 401 | acked -= cwnd - tp->snd_cwnd; |
9f9843a7 | 402 | tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); |
e73ebb08 NC |
403 | |
404 | return acked; | |
40efc6fa SH |
405 | } |
406 | EXPORT_SYMBOL_GPL(tcp_slow_start); | |
407 | ||
814d488c NC |
408 | /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w), |
409 | * for every packet that was ACKed. | |
410 | */ | |
e73ebb08 | 411 | void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked) |
758ce5c8 | 412 | { |
9949afa4 NC |
413 | /* If credits accumulated at a higher w, apply them gently now. */ |
414 | if (tp->snd_cwnd_cnt >= w) { | |
415 | tp->snd_cwnd_cnt = 0; | |
416 | tp->snd_cwnd++; | |
417 | } | |
418 | ||
814d488c | 419 | tp->snd_cwnd_cnt += acked; |
758ce5c8 | 420 | if (tp->snd_cwnd_cnt >= w) { |
814d488c NC |
421 | u32 delta = tp->snd_cwnd_cnt / w; |
422 | ||
423 | tp->snd_cwnd_cnt -= delta * w; | |
424 | tp->snd_cwnd += delta; | |
758ce5c8 | 425 | } |
814d488c | 426 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp); |
758ce5c8 IJ |
427 | } |
428 | EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai); | |
429 | ||
317a76f9 SH |
430 | /* |
431 | * TCP Reno congestion control | |
432 | * This is special case used for fallback as well. | |
433 | */ | |
434 | /* This is Jacobson's slow start and congestion avoidance. | |
435 | * SIGCOMM '88, p. 328. | |
436 | */ | |
24901551 | 437 | void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
317a76f9 | 438 | { |
6687e988 ACM |
439 | struct tcp_sock *tp = tcp_sk(sk); |
440 | ||
24901551 | 441 | if (!tcp_is_cwnd_limited(sk)) |
317a76f9 SH |
442 | return; |
443 | ||
7faffa1c | 444 | /* In "safe" area, increase. */ |
071d5080 | 445 | if (tcp_in_slow_start(tp)) { |
c22bdca9 NC |
446 | acked = tcp_slow_start(tp, acked); |
447 | if (!acked) | |
448 | return; | |
449 | } | |
e905a9ed | 450 | /* In dangerous area, increase slowly. */ |
c22bdca9 | 451 | tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked); |
317a76f9 SH |
452 | } |
453 | EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid); | |
454 | ||
455 | /* Slow start threshold is half the congestion window (min 2) */ | |
6687e988 | 456 | u32 tcp_reno_ssthresh(struct sock *sk) |
317a76f9 | 457 | { |
6687e988 | 458 | const struct tcp_sock *tp = tcp_sk(sk); |
688d1945 | 459 | |
317a76f9 SH |
460 | return max(tp->snd_cwnd >> 1U, 2U); |
461 | } | |
462 | EXPORT_SYMBOL_GPL(tcp_reno_ssthresh); | |
463 | ||
e9799183 FW |
464 | u32 tcp_reno_undo_cwnd(struct sock *sk) |
465 | { | |
466 | const struct tcp_sock *tp = tcp_sk(sk); | |
467 | ||
468 | return max(tp->snd_cwnd, tp->snd_ssthresh << 1); | |
469 | } | |
470 | EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd); | |
471 | ||
317a76f9 | 472 | struct tcp_congestion_ops tcp_reno = { |
164891aa | 473 | .flags = TCP_CONG_NON_RESTRICTED, |
317a76f9 SH |
474 | .name = "reno", |
475 | .owner = THIS_MODULE, | |
476 | .ssthresh = tcp_reno_ssthresh, | |
477 | .cong_avoid = tcp_reno_cong_avoid, | |
e9799183 | 478 | .undo_cwnd = tcp_reno_undo_cwnd, |
317a76f9 | 479 | }; |