[TCP]: Allow choosing TCP congestion control via sockopt.
[linux-2.6.git] / 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 tcp_sock *tp)
77 {
78         struct tcp_congestion_ops *ca;
79
80         if (tp->ca_ops != &tcp_init_congestion_ops)
81                 return;
82
83         rcu_read_lock();
84         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
85                 if (try_module_get(ca->owner)) {
86                         tp->ca_ops = ca;
87                         break;
88                 }
89
90         }
91         rcu_read_unlock();
92
93         if (tp->ca_ops->init)
94                 tp->ca_ops->init(tp);
95 }
96
97 /* Manage refcounts on socket close. */
98 void tcp_cleanup_congestion_control(struct tcp_sock *tp)
99 {
100         if (tp->ca_ops->release)
101                 tp->ca_ops->release(tp);
102         module_put(tp->ca_ops->owner);
103 }
104
105 /* Used by sysctl to change default congestion control */
106 int tcp_set_default_congestion_control(const char *name)
107 {
108         struct tcp_congestion_ops *ca;
109         int ret = -ENOENT;
110
111         spin_lock(&tcp_cong_list_lock);
112         ca = tcp_ca_find(name);
113 #ifdef CONFIG_KMOD
114         if (!ca) {
115                 spin_unlock(&tcp_cong_list_lock);
116
117                 request_module("tcp_%s", name);
118                 spin_lock(&tcp_cong_list_lock);
119                 ca = tcp_ca_find(name);
120         }
121 #endif
122
123         if (ca) {
124                 list_move(&ca->list, &tcp_cong_list);
125                 ret = 0;
126         }
127         spin_unlock(&tcp_cong_list_lock);
128
129         return ret;
130 }
131
132 /* Get current default congestion control */
133 void tcp_get_default_congestion_control(char *name)
134 {
135         struct tcp_congestion_ops *ca;
136         /* We will always have reno... */
137         BUG_ON(list_empty(&tcp_cong_list));
138
139         rcu_read_lock();
140         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
141         strncpy(name, ca->name, TCP_CA_NAME_MAX);
142         rcu_read_unlock();
143 }
144
145 /* Change congestion control for socket */
146 int tcp_set_congestion_control(struct tcp_sock *tp, const char *name)
147 {
148         struct tcp_congestion_ops *ca;
149         int err = 0;
150
151         rcu_read_lock();
152         ca = tcp_ca_find(name);
153         if (ca == tp->ca_ops)
154                 goto out;
155
156         if (!ca)
157                 err = -ENOENT;
158
159         else if (!try_module_get(ca->owner))
160                 err = -EBUSY;
161
162         else {
163                 tcp_cleanup_congestion_control(tp);
164                 tp->ca_ops = ca;
165                 if (tp->ca_ops->init)
166                         tp->ca_ops->init(tp);
167         }
168  out:
169         rcu_read_unlock();
170         return err;
171 }
172
173 /*
174  * TCP Reno congestion control
175  * This is special case used for fallback as well.
176  */
177 /* This is Jacobson's slow start and congestion avoidance.
178  * SIGCOMM '88, p. 328.
179  */
180 void tcp_reno_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt, u32 in_flight,
181                          int flag)
182 {
183         if (in_flight < tp->snd_cwnd)
184                 return;
185
186         if (tp->snd_cwnd <= tp->snd_ssthresh) {
187                 /* In "safe" area, increase. */
188                 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
189                         tp->snd_cwnd++;
190         } else {
191                 /* In dangerous area, increase slowly.
192                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
193                  */
194                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
195                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
196                                 tp->snd_cwnd++;
197                         tp->snd_cwnd_cnt = 0;
198                 } else
199                         tp->snd_cwnd_cnt++;
200         }
201 }
202 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
203
204 /* Slow start threshold is half the congestion window (min 2) */
205 u32 tcp_reno_ssthresh(struct tcp_sock *tp)
206 {
207         return max(tp->snd_cwnd >> 1U, 2U);
208 }
209 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
210
211 /* Lower bound on congestion window. */
212 u32 tcp_reno_min_cwnd(struct tcp_sock *tp)
213 {
214         return tp->snd_ssthresh/2;
215 }
216 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
217
218 struct tcp_congestion_ops tcp_reno = {
219         .name           = "reno",
220         .owner          = THIS_MODULE,
221         .ssthresh       = tcp_reno_ssthresh,
222         .cong_avoid     = tcp_reno_cong_avoid,
223         .min_cwnd       = tcp_reno_min_cwnd,
224 };
225
226 /* Initial congestion control used (until SYN)
227  * really reno under another name so we can tell difference
228  * during tcp_set_default_congestion_control
229  */
230 struct tcp_congestion_ops tcp_init_congestion_ops  = {
231         .name           = "",
232         .owner          = THIS_MODULE,
233         .ssthresh       = tcp_reno_ssthresh,
234         .cong_avoid     = tcp_reno_cong_avoid,
235         .min_cwnd       = tcp_reno_min_cwnd,
236 };
237 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);