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 | ||
71 | /* all algorithms must implement ssthresh and cong_avoid ops */ | |
72dc5b92 | 72 | if (!ca->ssthresh || !ca->cong_avoid) { |
afd46503 | 73 | pr_err("%s does not implement required ops\n", ca->name); |
317a76f9 SH |
74 | return -EINVAL; |
75 | } | |
76 | ||
c5c6a8ab DB |
77 | ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name)); |
78 | ||
317a76f9 | 79 | spin_lock(&tcp_cong_list_lock); |
c5c6a8ab DB |
80 | if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) { |
81 | pr_notice("%s already registered or non-unique key\n", | |
82 | ca->name); | |
317a76f9 SH |
83 | ret = -EEXIST; |
84 | } else { | |
3d2573f7 | 85 | list_add_tail_rcu(&ca->list, &tcp_cong_list); |
db2855ae | 86 | pr_debug("%s registered\n", ca->name); |
317a76f9 SH |
87 | } |
88 | spin_unlock(&tcp_cong_list_lock); | |
89 | ||
90 | return ret; | |
91 | } | |
92 | EXPORT_SYMBOL_GPL(tcp_register_congestion_control); | |
93 | ||
94 | /* | |
95 | * Remove congestion control algorithm, called from | |
96 | * the module's remove function. Module ref counts are used | |
97 | * to ensure that this can't be done till all sockets using | |
98 | * that method are closed. | |
99 | */ | |
100 | void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca) | |
101 | { | |
102 | spin_lock(&tcp_cong_list_lock); | |
103 | list_del_rcu(&ca->list); | |
104 | spin_unlock(&tcp_cong_list_lock); | |
c5c6a8ab DB |
105 | |
106 | /* Wait for outstanding readers to complete before the | |
107 | * module gets removed entirely. | |
108 | * | |
109 | * A try_module_get() should fail by now as our module is | |
110 | * in "going" state since no refs are held anymore and | |
111 | * module_exit() handler being called. | |
112 | */ | |
113 | synchronize_rcu(); | |
317a76f9 SH |
114 | } |
115 | EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control); | |
116 | ||
c3a8d947 | 117 | u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca) |
c5c6a8ab DB |
118 | { |
119 | const struct tcp_congestion_ops *ca; | |
c3a8d947 | 120 | u32 key = TCP_CA_UNSPEC; |
c5c6a8ab DB |
121 | |
122 | might_sleep(); | |
123 | ||
124 | rcu_read_lock(); | |
125 | ca = __tcp_ca_find_autoload(name); | |
c3a8d947 DB |
126 | if (ca) { |
127 | key = ca->key; | |
128 | *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN; | |
129 | } | |
c5c6a8ab DB |
130 | rcu_read_unlock(); |
131 | ||
132 | return key; | |
133 | } | |
134 | EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name); | |
135 | ||
136 | char *tcp_ca_get_name_by_key(u32 key, char *buffer) | |
137 | { | |
138 | const struct tcp_congestion_ops *ca; | |
139 | char *ret = NULL; | |
140 | ||
141 | rcu_read_lock(); | |
142 | ca = tcp_ca_find_key(key); | |
143 | if (ca) | |
144 | ret = strncpy(buffer, ca->name, | |
145 | TCP_CA_NAME_MAX); | |
146 | rcu_read_unlock(); | |
147 | ||
148 | return ret; | |
149 | } | |
150 | EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key); | |
151 | ||
317a76f9 | 152 | /* Assign choice of congestion control. */ |
55d8694f | 153 | void tcp_assign_congestion_control(struct sock *sk) |
317a76f9 | 154 | { |
6687e988 | 155 | struct inet_connection_sock *icsk = inet_csk(sk); |
317a76f9 SH |
156 | struct tcp_congestion_ops *ca; |
157 | ||
55d8694f FW |
158 | rcu_read_lock(); |
159 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
160 | if (likely(try_module_get(ca->owner))) { | |
161 | icsk->icsk_ca_ops = ca; | |
162 | goto out; | |
317a76f9 | 163 | } |
55d8694f FW |
164 | /* Fallback to next available. The last really |
165 | * guaranteed fallback is Reno from this list. | |
166 | */ | |
317a76f9 | 167 | } |
55d8694f FW |
168 | out: |
169 | rcu_read_unlock(); | |
170 | ||
171 | /* Clear out private data before diag gets it and | |
172 | * the ca has not been initialized. | |
173 | */ | |
174 | if (ca->get_info) | |
175 | memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); | |
6ac705b1 ED |
176 | if (ca->flags & TCP_CONG_NEEDS_ECN) |
177 | INET_ECN_xmit(sk); | |
178 | else | |
179 | INET_ECN_dontxmit(sk); | |
55d8694f FW |
180 | } |
181 | ||
182 | void tcp_init_congestion_control(struct sock *sk) | |
183 | { | |
184 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
317a76f9 | 185 | |
6687e988 ACM |
186 | if (icsk->icsk_ca_ops->init) |
187 | icsk->icsk_ca_ops->init(sk); | |
6ac705b1 ED |
188 | if (tcp_ca_needs_ecn(sk)) |
189 | INET_ECN_xmit(sk); | |
190 | else | |
191 | INET_ECN_dontxmit(sk); | |
317a76f9 SH |
192 | } |
193 | ||
29ba4fff DB |
194 | static void tcp_reinit_congestion_control(struct sock *sk, |
195 | const struct tcp_congestion_ops *ca) | |
196 | { | |
197 | struct inet_connection_sock *icsk = inet_csk(sk); | |
198 | ||
199 | tcp_cleanup_congestion_control(sk); | |
200 | icsk->icsk_ca_ops = ca; | |
9f950415 | 201 | icsk->icsk_ca_setsockopt = 1; |
29ba4fff | 202 | |
6ac705b1 ED |
203 | if (sk->sk_state != TCP_CLOSE) |
204 | tcp_init_congestion_control(sk); | |
29ba4fff DB |
205 | } |
206 | ||
317a76f9 | 207 | /* Manage refcounts on socket close. */ |
6687e988 | 208 | void tcp_cleanup_congestion_control(struct sock *sk) |
317a76f9 | 209 | { |
6687e988 ACM |
210 | struct inet_connection_sock *icsk = inet_csk(sk); |
211 | ||
212 | if (icsk->icsk_ca_ops->release) | |
213 | icsk->icsk_ca_ops->release(sk); | |
214 | module_put(icsk->icsk_ca_ops->owner); | |
317a76f9 SH |
215 | } |
216 | ||
217 | /* Used by sysctl to change default congestion control */ | |
218 | int tcp_set_default_congestion_control(const char *name) | |
219 | { | |
220 | struct tcp_congestion_ops *ca; | |
221 | int ret = -ENOENT; | |
222 | ||
223 | spin_lock(&tcp_cong_list_lock); | |
224 | ca = tcp_ca_find(name); | |
95a5afca | 225 | #ifdef CONFIG_MODULES |
a8f80e8f | 226 | if (!ca && capable(CAP_NET_ADMIN)) { |
317a76f9 SH |
227 | spin_unlock(&tcp_cong_list_lock); |
228 | ||
229 | request_module("tcp_%s", name); | |
230 | spin_lock(&tcp_cong_list_lock); | |
231 | ca = tcp_ca_find(name); | |
232 | } | |
233 | #endif | |
234 | ||
235 | if (ca) { | |
164891aa | 236 | ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */ |
317a76f9 SH |
237 | list_move(&ca->list, &tcp_cong_list); |
238 | ret = 0; | |
239 | } | |
240 | spin_unlock(&tcp_cong_list_lock); | |
241 | ||
242 | return ret; | |
243 | } | |
244 | ||
b1736a71 SH |
245 | /* Set default value from kernel configuration at bootup */ |
246 | static int __init tcp_congestion_default(void) | |
247 | { | |
248 | return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG); | |
249 | } | |
250 | late_initcall(tcp_congestion_default); | |
251 | ||
3ff825b2 SH |
252 | /* Build string with list of available congestion control values */ |
253 | void tcp_get_available_congestion_control(char *buf, size_t maxlen) | |
254 | { | |
255 | struct tcp_congestion_ops *ca; | |
256 | size_t offs = 0; | |
257 | ||
258 | rcu_read_lock(); | |
259 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
260 | offs += snprintf(buf + offs, maxlen - offs, | |
261 | "%s%s", | |
262 | offs == 0 ? "" : " ", ca->name); | |
3ff825b2 SH |
263 | } |
264 | rcu_read_unlock(); | |
265 | } | |
266 | ||
317a76f9 SH |
267 | /* Get current default congestion control */ |
268 | void tcp_get_default_congestion_control(char *name) | |
269 | { | |
270 | struct tcp_congestion_ops *ca; | |
271 | /* We will always have reno... */ | |
272 | BUG_ON(list_empty(&tcp_cong_list)); | |
273 | ||
274 | rcu_read_lock(); | |
275 | ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list); | |
276 | strncpy(name, ca->name, TCP_CA_NAME_MAX); | |
277 | rcu_read_unlock(); | |
278 | } | |
279 | ||
ce7bc3bf SH |
280 | /* Built list of non-restricted congestion control values */ |
281 | void tcp_get_allowed_congestion_control(char *buf, size_t maxlen) | |
282 | { | |
283 | struct tcp_congestion_ops *ca; | |
284 | size_t offs = 0; | |
285 | ||
286 | *buf = '\0'; | |
287 | rcu_read_lock(); | |
288 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
164891aa | 289 | if (!(ca->flags & TCP_CONG_NON_RESTRICTED)) |
ce7bc3bf SH |
290 | continue; |
291 | offs += snprintf(buf + offs, maxlen - offs, | |
292 | "%s%s", | |
293 | offs == 0 ? "" : " ", ca->name); | |
ce7bc3bf SH |
294 | } |
295 | rcu_read_unlock(); | |
296 | } | |
297 | ||
298 | /* Change list of non-restricted congestion control */ | |
299 | int tcp_set_allowed_congestion_control(char *val) | |
300 | { | |
301 | struct tcp_congestion_ops *ca; | |
c34186ed | 302 | char *saved_clone, *clone, *name; |
ce7bc3bf SH |
303 | int ret = 0; |
304 | ||
c34186ed | 305 | saved_clone = clone = kstrdup(val, GFP_USER); |
ce7bc3bf SH |
306 | if (!clone) |
307 | return -ENOMEM; | |
308 | ||
309 | spin_lock(&tcp_cong_list_lock); | |
310 | /* pass 1 check for bad entries */ | |
311 | while ((name = strsep(&clone, " ")) && *name) { | |
312 | ca = tcp_ca_find(name); | |
313 | if (!ca) { | |
314 | ret = -ENOENT; | |
315 | goto out; | |
316 | } | |
317 | } | |
318 | ||
164891aa | 319 | /* pass 2 clear old values */ |
ce7bc3bf | 320 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) |
164891aa | 321 | ca->flags &= ~TCP_CONG_NON_RESTRICTED; |
ce7bc3bf SH |
322 | |
323 | /* pass 3 mark as allowed */ | |
324 | while ((name = strsep(&val, " ")) && *name) { | |
325 | ca = tcp_ca_find(name); | |
326 | WARN_ON(!ca); | |
327 | if (ca) | |
164891aa | 328 | ca->flags |= TCP_CONG_NON_RESTRICTED; |
ce7bc3bf SH |
329 | } |
330 | out: | |
331 | spin_unlock(&tcp_cong_list_lock); | |
c34186ed | 332 | kfree(saved_clone); |
ce7bc3bf SH |
333 | |
334 | return ret; | |
335 | } | |
336 | ||
5f8ef48d | 337 | /* Change congestion control for socket */ |
6687e988 | 338 | int tcp_set_congestion_control(struct sock *sk, const char *name) |
5f8ef48d | 339 | { |
6687e988 | 340 | struct inet_connection_sock *icsk = inet_csk(sk); |
c5c6a8ab | 341 | const struct tcp_congestion_ops *ca; |
5f8ef48d SH |
342 | int err = 0; |
343 | ||
c5c6a8ab DB |
344 | if (icsk->icsk_ca_dst_locked) |
345 | return -EPERM; | |
4d4d3d1e | 346 | |
c5c6a8ab DB |
347 | rcu_read_lock(); |
348 | ca = __tcp_ca_find_autoload(name); | |
349 | /* No change asking for existing value */ | |
9f950415 NC |
350 | if (ca == icsk->icsk_ca_ops) { |
351 | icsk->icsk_ca_setsockopt = 1; | |
5f8ef48d | 352 | goto out; |
9f950415 | 353 | } |
5f8ef48d SH |
354 | if (!ca) |
355 | err = -ENOENT; | |
52e804c6 EB |
356 | else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || |
357 | ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN))) | |
ce7bc3bf | 358 | err = -EPERM; |
5f8ef48d SH |
359 | else if (!try_module_get(ca->owner)) |
360 | err = -EBUSY; | |
29ba4fff DB |
361 | else |
362 | tcp_reinit_congestion_control(sk, ca); | |
5f8ef48d SH |
363 | out: |
364 | rcu_read_unlock(); | |
365 | return err; | |
366 | } | |
367 | ||
9f9843a7 YC |
368 | /* Slow start is used when congestion window is no greater than the slow start |
369 | * threshold. We base on RFC2581 and also handle stretch ACKs properly. | |
370 | * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but | |
371 | * something better;) a packet is only considered (s)acked in its entirety to | |
372 | * defend the ACK attacks described in the RFC. Slow start processes a stretch | |
373 | * ACK of degree N as if N acks of degree 1 are received back to back except | |
374 | * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and | |
375 | * returns the leftover acks to adjust cwnd in congestion avoidance mode. | |
40efc6fa | 376 | */ |
e73ebb08 | 377 | u32 tcp_slow_start(struct tcp_sock *tp, u32 acked) |
40efc6fa | 378 | { |
76174004 | 379 | u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh); |
a02ba041 | 380 | |
e73ebb08 | 381 | acked -= cwnd - tp->snd_cwnd; |
9f9843a7 | 382 | tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp); |
e73ebb08 NC |
383 | |
384 | return acked; | |
40efc6fa SH |
385 | } |
386 | EXPORT_SYMBOL_GPL(tcp_slow_start); | |
387 | ||
814d488c NC |
388 | /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w), |
389 | * for every packet that was ACKed. | |
390 | */ | |
e73ebb08 | 391 | void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked) |
758ce5c8 | 392 | { |
9949afa4 NC |
393 | /* If credits accumulated at a higher w, apply them gently now. */ |
394 | if (tp->snd_cwnd_cnt >= w) { | |
395 | tp->snd_cwnd_cnt = 0; | |
396 | tp->snd_cwnd++; | |
397 | } | |
398 | ||
814d488c | 399 | tp->snd_cwnd_cnt += acked; |
758ce5c8 | 400 | if (tp->snd_cwnd_cnt >= w) { |
814d488c NC |
401 | u32 delta = tp->snd_cwnd_cnt / w; |
402 | ||
403 | tp->snd_cwnd_cnt -= delta * w; | |
404 | tp->snd_cwnd += delta; | |
758ce5c8 | 405 | } |
814d488c | 406 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp); |
758ce5c8 IJ |
407 | } |
408 | EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai); | |
409 | ||
317a76f9 SH |
410 | /* |
411 | * TCP Reno congestion control | |
412 | * This is special case used for fallback as well. | |
413 | */ | |
414 | /* This is Jacobson's slow start and congestion avoidance. | |
415 | * SIGCOMM '88, p. 328. | |
416 | */ | |
24901551 | 417 | void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
317a76f9 | 418 | { |
6687e988 ACM |
419 | struct tcp_sock *tp = tcp_sk(sk); |
420 | ||
24901551 | 421 | if (!tcp_is_cwnd_limited(sk)) |
317a76f9 SH |
422 | return; |
423 | ||
7faffa1c | 424 | /* In "safe" area, increase. */ |
071d5080 | 425 | if (tcp_in_slow_start(tp)) { |
c22bdca9 NC |
426 | acked = tcp_slow_start(tp, acked); |
427 | if (!acked) | |
428 | return; | |
429 | } | |
e905a9ed | 430 | /* In dangerous area, increase slowly. */ |
c22bdca9 | 431 | tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked); |
317a76f9 SH |
432 | } |
433 | EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid); | |
434 | ||
435 | /* Slow start threshold is half the congestion window (min 2) */ | |
6687e988 | 436 | u32 tcp_reno_ssthresh(struct sock *sk) |
317a76f9 | 437 | { |
6687e988 | 438 | const struct tcp_sock *tp = tcp_sk(sk); |
688d1945 | 439 | |
317a76f9 SH |
440 | return max(tp->snd_cwnd >> 1U, 2U); |
441 | } | |
442 | EXPORT_SYMBOL_GPL(tcp_reno_ssthresh); | |
443 | ||
317a76f9 | 444 | struct tcp_congestion_ops tcp_reno = { |
164891aa | 445 | .flags = TCP_CONG_NON_RESTRICTED, |
317a76f9 SH |
446 | .name = "reno", |
447 | .owner = THIS_MODULE, | |
448 | .ssthresh = tcp_reno_ssthresh, | |
449 | .cong_avoid = tcp_reno_cong_avoid, | |
317a76f9 | 450 | }; |