Commit | Line | Data |
---|---|---|
317a76f9 SH |
1 | /* |
2 | * Plugable TCP congestion control support and newReno | |
3 | * congestion control. | |
4 | * Based on ideas from I/O scheduler suport and Web100. | |
5 | * | |
6 | * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org> | |
7 | */ | |
8 | ||
317a76f9 SH |
9 | #include <linux/module.h> |
10 | #include <linux/mm.h> | |
11 | #include <linux/types.h> | |
12 | #include <linux/list.h> | |
13 | #include <net/tcp.h> | |
14 | ||
886236c1 JH |
15 | int sysctl_tcp_max_ssthresh = 0; |
16 | ||
317a76f9 SH |
17 | static DEFINE_SPINLOCK(tcp_cong_list_lock); |
18 | static LIST_HEAD(tcp_cong_list); | |
19 | ||
20 | /* Simple linear search, don't expect many entries! */ | |
21 | static struct tcp_congestion_ops *tcp_ca_find(const char *name) | |
22 | { | |
23 | struct tcp_congestion_ops *e; | |
24 | ||
5f8ef48d | 25 | list_for_each_entry_rcu(e, &tcp_cong_list, list) { |
317a76f9 SH |
26 | if (strcmp(e->name, name) == 0) |
27 | return e; | |
28 | } | |
29 | ||
30 | return NULL; | |
31 | } | |
32 | ||
33 | /* | |
d08df601 | 34 | * Attach new congestion control algorithm to the list |
317a76f9 SH |
35 | * of available options. |
36 | */ | |
37 | int tcp_register_congestion_control(struct tcp_congestion_ops *ca) | |
38 | { | |
39 | int ret = 0; | |
40 | ||
41 | /* all algorithms must implement ssthresh and cong_avoid ops */ | |
72dc5b92 | 42 | if (!ca->ssthresh || !ca->cong_avoid) { |
317a76f9 SH |
43 | printk(KERN_ERR "TCP %s does not implement required ops\n", |
44 | ca->name); | |
45 | return -EINVAL; | |
46 | } | |
47 | ||
48 | spin_lock(&tcp_cong_list_lock); | |
49 | if (tcp_ca_find(ca->name)) { | |
50 | printk(KERN_NOTICE "TCP %s already registered\n", ca->name); | |
51 | ret = -EEXIST; | |
52 | } else { | |
3d2573f7 | 53 | list_add_tail_rcu(&ca->list, &tcp_cong_list); |
317a76f9 SH |
54 | printk(KERN_INFO "TCP %s registered\n", ca->name); |
55 | } | |
56 | spin_unlock(&tcp_cong_list_lock); | |
57 | ||
58 | return ret; | |
59 | } | |
60 | EXPORT_SYMBOL_GPL(tcp_register_congestion_control); | |
61 | ||
62 | /* | |
63 | * Remove congestion control algorithm, called from | |
64 | * the module's remove function. Module ref counts are used | |
65 | * to ensure that this can't be done till all sockets using | |
66 | * that method are closed. | |
67 | */ | |
68 | void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca) | |
69 | { | |
70 | spin_lock(&tcp_cong_list_lock); | |
71 | list_del_rcu(&ca->list); | |
72 | spin_unlock(&tcp_cong_list_lock); | |
73 | } | |
74 | EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control); | |
75 | ||
76 | /* Assign choice of congestion control. */ | |
6687e988 | 77 | void tcp_init_congestion_control(struct sock *sk) |
317a76f9 | 78 | { |
6687e988 | 79 | struct inet_connection_sock *icsk = inet_csk(sk); |
317a76f9 SH |
80 | struct tcp_congestion_ops *ca; |
81 | ||
4d4d3d1e SH |
82 | /* if no choice made yet assign the current value set as default */ |
83 | if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) { | |
84 | rcu_read_lock(); | |
85 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
86 | if (try_module_get(ca->owner)) { | |
87 | icsk->icsk_ca_ops = ca; | |
88 | break; | |
89 | } | |
5f8ef48d | 90 | |
4d4d3d1e | 91 | /* fallback to next available */ |
317a76f9 | 92 | } |
4d4d3d1e | 93 | rcu_read_unlock(); |
317a76f9 | 94 | } |
317a76f9 | 95 | |
6687e988 ACM |
96 | if (icsk->icsk_ca_ops->init) |
97 | icsk->icsk_ca_ops->init(sk); | |
317a76f9 SH |
98 | } |
99 | ||
100 | /* Manage refcounts on socket close. */ | |
6687e988 | 101 | void tcp_cleanup_congestion_control(struct sock *sk) |
317a76f9 | 102 | { |
6687e988 ACM |
103 | struct inet_connection_sock *icsk = inet_csk(sk); |
104 | ||
105 | if (icsk->icsk_ca_ops->release) | |
106 | icsk->icsk_ca_ops->release(sk); | |
107 | module_put(icsk->icsk_ca_ops->owner); | |
317a76f9 SH |
108 | } |
109 | ||
110 | /* Used by sysctl to change default congestion control */ | |
111 | int tcp_set_default_congestion_control(const char *name) | |
112 | { | |
113 | struct tcp_congestion_ops *ca; | |
114 | int ret = -ENOENT; | |
115 | ||
116 | spin_lock(&tcp_cong_list_lock); | |
117 | ca = tcp_ca_find(name); | |
118 | #ifdef CONFIG_KMOD | |
35bfbc94 | 119 | if (!ca && capable(CAP_SYS_MODULE)) { |
317a76f9 SH |
120 | spin_unlock(&tcp_cong_list_lock); |
121 | ||
122 | request_module("tcp_%s", name); | |
123 | spin_lock(&tcp_cong_list_lock); | |
124 | ca = tcp_ca_find(name); | |
125 | } | |
126 | #endif | |
127 | ||
128 | if (ca) { | |
ce7bc3bf | 129 | ca->non_restricted = 1; /* default is always allowed */ |
317a76f9 SH |
130 | list_move(&ca->list, &tcp_cong_list); |
131 | ret = 0; | |
132 | } | |
133 | spin_unlock(&tcp_cong_list_lock); | |
134 | ||
135 | return ret; | |
136 | } | |
137 | ||
b1736a71 SH |
138 | /* Set default value from kernel configuration at bootup */ |
139 | static int __init tcp_congestion_default(void) | |
140 | { | |
141 | return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG); | |
142 | } | |
143 | late_initcall(tcp_congestion_default); | |
144 | ||
145 | ||
3ff825b2 SH |
146 | /* Build string with list of available congestion control values */ |
147 | void tcp_get_available_congestion_control(char *buf, size_t maxlen) | |
148 | { | |
149 | struct tcp_congestion_ops *ca; | |
150 | size_t offs = 0; | |
151 | ||
152 | rcu_read_lock(); | |
153 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
154 | offs += snprintf(buf + offs, maxlen - offs, | |
155 | "%s%s", | |
156 | offs == 0 ? "" : " ", ca->name); | |
157 | ||
158 | } | |
159 | rcu_read_unlock(); | |
160 | } | |
161 | ||
317a76f9 SH |
162 | /* Get current default congestion control */ |
163 | void tcp_get_default_congestion_control(char *name) | |
164 | { | |
165 | struct tcp_congestion_ops *ca; | |
166 | /* We will always have reno... */ | |
167 | BUG_ON(list_empty(&tcp_cong_list)); | |
168 | ||
169 | rcu_read_lock(); | |
170 | ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list); | |
171 | strncpy(name, ca->name, TCP_CA_NAME_MAX); | |
172 | rcu_read_unlock(); | |
173 | } | |
174 | ||
ce7bc3bf SH |
175 | /* Built list of non-restricted congestion control values */ |
176 | void tcp_get_allowed_congestion_control(char *buf, size_t maxlen) | |
177 | { | |
178 | struct tcp_congestion_ops *ca; | |
179 | size_t offs = 0; | |
180 | ||
181 | *buf = '\0'; | |
182 | rcu_read_lock(); | |
183 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) { | |
184 | if (!ca->non_restricted) | |
185 | continue; | |
186 | offs += snprintf(buf + offs, maxlen - offs, | |
187 | "%s%s", | |
188 | offs == 0 ? "" : " ", ca->name); | |
189 | ||
190 | } | |
191 | rcu_read_unlock(); | |
192 | } | |
193 | ||
194 | /* Change list of non-restricted congestion control */ | |
195 | int tcp_set_allowed_congestion_control(char *val) | |
196 | { | |
197 | struct tcp_congestion_ops *ca; | |
198 | char *clone, *name; | |
199 | int ret = 0; | |
200 | ||
201 | clone = kstrdup(val, GFP_USER); | |
202 | if (!clone) | |
203 | return -ENOMEM; | |
204 | ||
205 | spin_lock(&tcp_cong_list_lock); | |
206 | /* pass 1 check for bad entries */ | |
207 | while ((name = strsep(&clone, " ")) && *name) { | |
208 | ca = tcp_ca_find(name); | |
209 | if (!ca) { | |
210 | ret = -ENOENT; | |
211 | goto out; | |
212 | } | |
213 | } | |
214 | ||
215 | /* pass 2 clear */ | |
216 | list_for_each_entry_rcu(ca, &tcp_cong_list, list) | |
217 | ca->non_restricted = 0; | |
218 | ||
219 | /* pass 3 mark as allowed */ | |
220 | while ((name = strsep(&val, " ")) && *name) { | |
221 | ca = tcp_ca_find(name); | |
222 | WARN_ON(!ca); | |
223 | if (ca) | |
224 | ca->non_restricted = 1; | |
225 | } | |
226 | out: | |
227 | spin_unlock(&tcp_cong_list_lock); | |
228 | ||
229 | return ret; | |
230 | } | |
231 | ||
232 | ||
5f8ef48d | 233 | /* Change congestion control for socket */ |
6687e988 | 234 | int tcp_set_congestion_control(struct sock *sk, const char *name) |
5f8ef48d | 235 | { |
6687e988 | 236 | struct inet_connection_sock *icsk = inet_csk(sk); |
5f8ef48d SH |
237 | struct tcp_congestion_ops *ca; |
238 | int err = 0; | |
239 | ||
240 | rcu_read_lock(); | |
241 | ca = tcp_ca_find(name); | |
4d4d3d1e | 242 | |
35bfbc94 | 243 | /* no change asking for existing value */ |
6687e988 | 244 | if (ca == icsk->icsk_ca_ops) |
5f8ef48d SH |
245 | goto out; |
246 | ||
35bfbc94 SH |
247 | #ifdef CONFIG_KMOD |
248 | /* not found attempt to autoload module */ | |
249 | if (!ca && capable(CAP_SYS_MODULE)) { | |
250 | rcu_read_unlock(); | |
251 | request_module("tcp_%s", name); | |
252 | rcu_read_lock(); | |
253 | ca = tcp_ca_find(name); | |
254 | } | |
255 | #endif | |
5f8ef48d SH |
256 | if (!ca) |
257 | err = -ENOENT; | |
258 | ||
ce7bc3bf SH |
259 | else if (!(ca->non_restricted || capable(CAP_NET_ADMIN))) |
260 | err = -EPERM; | |
261 | ||
5f8ef48d SH |
262 | else if (!try_module_get(ca->owner)) |
263 | err = -EBUSY; | |
264 | ||
265 | else { | |
6687e988 ACM |
266 | tcp_cleanup_congestion_control(sk); |
267 | icsk->icsk_ca_ops = ca; | |
4d4d3d1e SH |
268 | |
269 | if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init) | |
6687e988 | 270 | icsk->icsk_ca_ops->init(sk); |
5f8ef48d SH |
271 | } |
272 | out: | |
273 | rcu_read_unlock(); | |
274 | return err; | |
275 | } | |
276 | ||
40efc6fa SH |
277 | |
278 | /* | |
886236c1 JH |
279 | * Slow start (exponential increase) with |
280 | * RFC3742 Limited Slow Start (fast linear increase) support. | |
40efc6fa SH |
281 | */ |
282 | void tcp_slow_start(struct tcp_sock *tp) | |
283 | { | |
886236c1 JH |
284 | int cnt = 0; |
285 | ||
40efc6fa SH |
286 | if (sysctl_tcp_abc) { |
287 | /* RFC3465: Slow Start | |
288 | * TCP sender SHOULD increase cwnd by the number of | |
289 | * previously unacknowledged bytes ACKed by each incoming | |
290 | * acknowledgment, provided the increase is not more than L | |
291 | */ | |
292 | if (tp->bytes_acked < tp->mss_cache) | |
293 | return; | |
40efc6fa | 294 | } |
886236c1 JH |
295 | |
296 | if (sysctl_tcp_max_ssthresh > 0 && | |
297 | tp->snd_cwnd > sysctl_tcp_max_ssthresh) | |
298 | cnt += sysctl_tcp_max_ssthresh>>1; | |
299 | else | |
300 | cnt += tp->snd_cwnd; | |
301 | ||
302 | /* RFC3465: We MAY increase by 2 if discovered delayed ack */ | |
303 | if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache) | |
304 | cnt <<= 1; | |
40efc6fa SH |
305 | tp->bytes_acked = 0; |
306 | ||
886236c1 JH |
307 | tp->snd_cwnd_cnt += cnt; |
308 | while (tp->snd_cwnd_cnt >= tp->snd_cwnd) { | |
309 | tp->snd_cwnd_cnt -= tp->snd_cwnd; | |
310 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | |
311 | tp->snd_cwnd++; | |
312 | } | |
40efc6fa SH |
313 | } |
314 | EXPORT_SYMBOL_GPL(tcp_slow_start); | |
315 | ||
317a76f9 SH |
316 | /* |
317 | * TCP Reno congestion control | |
318 | * This is special case used for fallback as well. | |
319 | */ | |
320 | /* This is Jacobson's slow start and congestion avoidance. | |
321 | * SIGCOMM '88, p. 328. | |
322 | */ | |
6687e988 | 323 | void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight, |
317a76f9 SH |
324 | int flag) |
325 | { | |
6687e988 ACM |
326 | struct tcp_sock *tp = tcp_sk(sk); |
327 | ||
f4805ede | 328 | if (!tcp_is_cwnd_limited(sk, in_flight)) |
317a76f9 SH |
329 | return; |
330 | ||
7faffa1c | 331 | /* In "safe" area, increase. */ |
e905a9ed | 332 | if (tp->snd_cwnd <= tp->snd_ssthresh) |
7faffa1c | 333 | tcp_slow_start(tp); |
9772efb9 | 334 | |
e905a9ed | 335 | /* In dangerous area, increase slowly. */ |
9772efb9 | 336 | else if (sysctl_tcp_abc) { |
e905a9ed YH |
337 | /* RFC3465: Appropriate Byte Count |
338 | * increase once for each full cwnd acked | |
339 | */ | |
340 | if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) { | |
341 | tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache; | |
342 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | |
343 | tp->snd_cwnd++; | |
344 | } | |
345 | } else { | |
346 | /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */ | |
347 | if (tp->snd_cwnd_cnt >= tp->snd_cwnd) { | |
348 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | |
349 | tp->snd_cwnd++; | |
350 | tp->snd_cwnd_cnt = 0; | |
351 | } else | |
352 | tp->snd_cwnd_cnt++; | |
353 | } | |
317a76f9 SH |
354 | } |
355 | EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid); | |
356 | ||
357 | /* Slow start threshold is half the congestion window (min 2) */ | |
6687e988 | 358 | u32 tcp_reno_ssthresh(struct sock *sk) |
317a76f9 | 359 | { |
6687e988 | 360 | const struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 SH |
361 | return max(tp->snd_cwnd >> 1U, 2U); |
362 | } | |
363 | EXPORT_SYMBOL_GPL(tcp_reno_ssthresh); | |
364 | ||
72dc5b92 SH |
365 | /* Lower bound on congestion window with halving. */ |
366 | u32 tcp_reno_min_cwnd(const struct sock *sk) | |
317a76f9 | 367 | { |
6687e988 | 368 | const struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 SH |
369 | return tp->snd_ssthresh/2; |
370 | } | |
371 | EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd); | |
372 | ||
373 | struct tcp_congestion_ops tcp_reno = { | |
374 | .name = "reno", | |
ce7bc3bf | 375 | .non_restricted = 1, |
317a76f9 SH |
376 | .owner = THIS_MODULE, |
377 | .ssthresh = tcp_reno_ssthresh, | |
378 | .cong_avoid = tcp_reno_cong_avoid, | |
379 | .min_cwnd = tcp_reno_min_cwnd, | |
380 | }; | |
381 | ||
5f8ef48d SH |
382 | /* Initial congestion control used (until SYN) |
383 | * really reno under another name so we can tell difference | |
384 | * during tcp_set_default_congestion_control | |
385 | */ | |
386 | struct tcp_congestion_ops tcp_init_congestion_ops = { | |
387 | .name = "", | |
388 | .owner = THIS_MODULE, | |
389 | .ssthresh = tcp_reno_ssthresh, | |
390 | .cong_avoid = tcp_reno_cong_avoid, | |
391 | .min_cwnd = tcp_reno_min_cwnd, | |
392 | }; | |
393 | EXPORT_SYMBOL_GPL(tcp_init_congestion_ops); |