Merge master.kernel.org:/pub/scm/linux/kernel/git/wim/linux-2.6-watchdog
[linux-2.6] / net / ipv4 / tcp_cong.c
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
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <linux/mm.h>
12 #include <linux/types.h>
13 #include <linux/list.h>
14 #include <net/tcp.h>
15
16 static DEFINE_SPINLOCK(tcp_cong_list_lock);
17 static LIST_HEAD(tcp_cong_list);
18
19 /* Simple linear search, don't expect many entries! */
20 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
21 {
22         struct tcp_congestion_ops *e;
23
24         list_for_each_entry_rcu(e, &tcp_cong_list, list) {
25                 if (strcmp(e->name, name) == 0)
26                         return e;
27         }
28
29         return NULL;
30 }
31
32 /*
33  * Attach new congestion control algorthim to the list
34  * of available options.
35  */
36 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
37 {
38         int ret = 0;
39
40         /* all algorithms must implement ssthresh and cong_avoid ops */
41         if (!ca->ssthresh || !ca->cong_avoid || !ca->min_cwnd) {
42                 printk(KERN_ERR "TCP %s does not implement required ops\n",
43                        ca->name);
44                 return -EINVAL;
45         }
46
47         spin_lock(&tcp_cong_list_lock);
48         if (tcp_ca_find(ca->name)) {
49                 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
50                 ret = -EEXIST;
51         } else {
52                 list_add_rcu(&ca->list, &tcp_cong_list);
53                 printk(KERN_INFO "TCP %s registered\n", ca->name);
54         }
55         spin_unlock(&tcp_cong_list_lock);
56
57         return ret;
58 }
59 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
60
61 /*
62  * Remove congestion control algorithm, called from
63  * the module's remove function.  Module ref counts are used
64  * to ensure that this can't be done till all sockets using
65  * that method are closed.
66  */
67 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
68 {
69         spin_lock(&tcp_cong_list_lock);
70         list_del_rcu(&ca->list);
71         spin_unlock(&tcp_cong_list_lock);
72 }
73 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
74
75 /* Assign choice of congestion control. */
76 void tcp_init_congestion_control(struct sock *sk)
77 {
78         struct inet_connection_sock *icsk = inet_csk(sk);
79         struct tcp_congestion_ops *ca;
80
81         if (icsk->icsk_ca_ops != &tcp_init_congestion_ops)
82                 return;
83
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                 }
90
91         }
92         rcu_read_unlock();
93
94         if (icsk->icsk_ca_ops->init)
95                 icsk->icsk_ca_ops->init(sk);
96 }
97
98 /* Manage refcounts on socket close. */
99 void tcp_cleanup_congestion_control(struct sock *sk)
100 {
101         struct inet_connection_sock *icsk = inet_csk(sk);
102
103         if (icsk->icsk_ca_ops->release)
104                 icsk->icsk_ca_ops->release(sk);
105         module_put(icsk->icsk_ca_ops->owner);
106 }
107
108 /* Used by sysctl to change default congestion control */
109 int tcp_set_default_congestion_control(const char *name)
110 {
111         struct tcp_congestion_ops *ca;
112         int ret = -ENOENT;
113
114         spin_lock(&tcp_cong_list_lock);
115         ca = tcp_ca_find(name);
116 #ifdef CONFIG_KMOD
117         if (!ca) {
118                 spin_unlock(&tcp_cong_list_lock);
119
120                 request_module("tcp_%s", name);
121                 spin_lock(&tcp_cong_list_lock);
122                 ca = tcp_ca_find(name);
123         }
124 #endif
125
126         if (ca) {
127                 list_move(&ca->list, &tcp_cong_list);
128                 ret = 0;
129         }
130         spin_unlock(&tcp_cong_list_lock);
131
132         return ret;
133 }
134
135 /* Get current default congestion control */
136 void tcp_get_default_congestion_control(char *name)
137 {
138         struct tcp_congestion_ops *ca;
139         /* We will always have reno... */
140         BUG_ON(list_empty(&tcp_cong_list));
141
142         rcu_read_lock();
143         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
144         strncpy(name, ca->name, TCP_CA_NAME_MAX);
145         rcu_read_unlock();
146 }
147
148 /* Change congestion control for socket */
149 int tcp_set_congestion_control(struct sock *sk, const char *name)
150 {
151         struct inet_connection_sock *icsk = inet_csk(sk);
152         struct tcp_congestion_ops *ca;
153         int err = 0;
154
155         rcu_read_lock();
156         ca = tcp_ca_find(name);
157         if (ca == icsk->icsk_ca_ops)
158                 goto out;
159
160         if (!ca)
161                 err = -ENOENT;
162
163         else if (!try_module_get(ca->owner))
164                 err = -EBUSY;
165
166         else {
167                 tcp_cleanup_congestion_control(sk);
168                 icsk->icsk_ca_ops = ca;
169                 if (icsk->icsk_ca_ops->init)
170                         icsk->icsk_ca_ops->init(sk);
171         }
172  out:
173         rcu_read_unlock();
174         return err;
175 }
176
177
178 /*
179  * Linear increase during slow start
180  */
181 void tcp_slow_start(struct tcp_sock *tp)
182 {
183         if (sysctl_tcp_abc) {
184                 /* RFC3465: Slow Start
185                  * TCP sender SHOULD increase cwnd by the number of
186                  * previously unacknowledged bytes ACKed by each incoming
187                  * acknowledgment, provided the increase is not more than L
188                  */
189                 if (tp->bytes_acked < tp->mss_cache)
190                         return;
191
192                 /* We MAY increase by 2 if discovered delayed ack */
193                 if (sysctl_tcp_abc > 1 && tp->bytes_acked > 2*tp->mss_cache) {
194                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
195                                 tp->snd_cwnd++;
196                 }
197         }
198         tp->bytes_acked = 0;
199
200         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
201                 tp->snd_cwnd++;
202 }
203 EXPORT_SYMBOL_GPL(tcp_slow_start);
204
205 /*
206  * TCP Reno congestion control
207  * This is special case used for fallback as well.
208  */
209 /* This is Jacobson's slow start and congestion avoidance.
210  * SIGCOMM '88, p. 328.
211  */
212 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
213                          int flag)
214 {
215         struct tcp_sock *tp = tcp_sk(sk);
216
217         if (!tcp_is_cwnd_limited(sk, in_flight))
218                 return;
219
220         /* In "safe" area, increase. */
221         if (tp->snd_cwnd <= tp->snd_ssthresh)
222                 tcp_slow_start(tp);
223
224         /* In dangerous area, increase slowly. */
225         else if (sysctl_tcp_abc) {
226                 /* RFC3465: Appropriate Byte Count
227                  * increase once for each full cwnd acked
228                  */
229                 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
230                         tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
231                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
232                                 tp->snd_cwnd++;
233                 }
234         } else {
235                 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
236                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
237                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
238                                 tp->snd_cwnd++;
239                         tp->snd_cwnd_cnt = 0;
240                 } else
241                         tp->snd_cwnd_cnt++;
242         }
243 }
244 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
245
246 /* Slow start threshold is half the congestion window (min 2) */
247 u32 tcp_reno_ssthresh(struct sock *sk)
248 {
249         const struct tcp_sock *tp = tcp_sk(sk);
250         return max(tp->snd_cwnd >> 1U, 2U);
251 }
252 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
253
254 /* Lower bound on congestion window. */
255 u32 tcp_reno_min_cwnd(struct sock *sk)
256 {
257         const struct tcp_sock *tp = tcp_sk(sk);
258         return tp->snd_ssthresh/2;
259 }
260 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
261
262 struct tcp_congestion_ops tcp_reno = {
263         .name           = "reno",
264         .owner          = THIS_MODULE,
265         .ssthresh       = tcp_reno_ssthresh,
266         .cong_avoid     = tcp_reno_cong_avoid,
267         .min_cwnd       = tcp_reno_min_cwnd,
268 };
269
270 /* Initial congestion control used (until SYN)
271  * really reno under another name so we can tell difference
272  * during tcp_set_default_congestion_control
273  */
274 struct tcp_congestion_ops tcp_init_congestion_ops  = {
275         .name           = "",
276         .owner          = THIS_MODULE,
277         .ssthresh       = tcp_reno_ssthresh,
278         .cong_avoid     = tcp_reno_cong_avoid,
279         .min_cwnd       = tcp_reno_min_cwnd,
280 };
281 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);