2 * Plugable TCP congestion control support and newReno
4 * Based on ideas from I/O scheduler suport and Web100.
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
9 #include <linux/config.h>
10 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/list.h>
16 static DEFINE_SPINLOCK(tcp_cong_list_lock);
17 static LIST_HEAD(tcp_cong_list);
19 /* Simple linear search, don't expect many entries! */
20 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
22 struct tcp_congestion_ops *e;
24 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
25 if (strcmp(e->name, name) == 0)
33 * Attach new congestion control algorthim to the list
34 * of available options.
36 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
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",
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);
52 list_add_rcu(&ca->list, &tcp_cong_list);
53 printk(KERN_INFO "TCP %s registered\n", ca->name);
55 spin_unlock(&tcp_cong_list_lock);
59 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
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.
67 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
69 spin_lock(&tcp_cong_list_lock);
70 list_del_rcu(&ca->list);
71 spin_unlock(&tcp_cong_list_lock);
73 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
75 /* Assign choice of congestion control. */
76 void tcp_init_congestion_control(struct sock *sk)
78 struct inet_connection_sock *icsk = inet_csk(sk);
79 struct tcp_congestion_ops *ca;
81 if (icsk->icsk_ca_ops != &tcp_init_congestion_ops)
85 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
86 if (try_module_get(ca->owner)) {
87 icsk->icsk_ca_ops = ca;
94 if (icsk->icsk_ca_ops->init)
95 icsk->icsk_ca_ops->init(sk);
98 /* Manage refcounts on socket close. */
99 void tcp_cleanup_congestion_control(struct sock *sk)
101 struct inet_connection_sock *icsk = inet_csk(sk);
103 if (icsk->icsk_ca_ops->release)
104 icsk->icsk_ca_ops->release(sk);
105 module_put(icsk->icsk_ca_ops->owner);
108 /* Used by sysctl to change default congestion control */
109 int tcp_set_default_congestion_control(const char *name)
111 struct tcp_congestion_ops *ca;
114 spin_lock(&tcp_cong_list_lock);
115 ca = tcp_ca_find(name);
118 spin_unlock(&tcp_cong_list_lock);
120 request_module("tcp_%s", name);
121 spin_lock(&tcp_cong_list_lock);
122 ca = tcp_ca_find(name);
127 list_move(&ca->list, &tcp_cong_list);
130 spin_unlock(&tcp_cong_list_lock);
135 /* Get current default congestion control */
136 void tcp_get_default_congestion_control(char *name)
138 struct tcp_congestion_ops *ca;
139 /* We will always have reno... */
140 BUG_ON(list_empty(&tcp_cong_list));
143 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
144 strncpy(name, ca->name, TCP_CA_NAME_MAX);
148 /* Change congestion control for socket */
149 int tcp_set_congestion_control(struct sock *sk, const char *name)
151 struct inet_connection_sock *icsk = inet_csk(sk);
152 struct tcp_congestion_ops *ca;
156 ca = tcp_ca_find(name);
157 if (ca == icsk->icsk_ca_ops)
163 else if (!try_module_get(ca->owner))
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);
178 * TCP Reno congestion control
179 * This is special case used for fallback as well.
181 /* This is Jacobson's slow start and congestion avoidance.
182 * SIGCOMM '88, p. 328.
184 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
187 struct tcp_sock *tp = tcp_sk(sk);
189 if (in_flight < tp->snd_cwnd)
192 if (tp->snd_cwnd <= tp->snd_ssthresh) {
193 /* In "safe" area, increase. */
194 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
197 /* In dangerous area, increase slowly.
198 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
200 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
201 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
203 tp->snd_cwnd_cnt = 0;
208 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
210 /* Slow start threshold is half the congestion window (min 2) */
211 u32 tcp_reno_ssthresh(struct sock *sk)
213 const struct tcp_sock *tp = tcp_sk(sk);
214 return max(tp->snd_cwnd >> 1U, 2U);
216 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
218 /* Lower bound on congestion window. */
219 u32 tcp_reno_min_cwnd(struct sock *sk)
221 const struct tcp_sock *tp = tcp_sk(sk);
222 return tp->snd_ssthresh/2;
224 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
226 struct tcp_congestion_ops tcp_reno = {
228 .owner = THIS_MODULE,
229 .ssthresh = tcp_reno_ssthresh,
230 .cong_avoid = tcp_reno_cong_avoid,
231 .min_cwnd = tcp_reno_min_cwnd,
234 /* Initial congestion control used (until SYN)
235 * really reno under another name so we can tell difference
236 * during tcp_set_default_congestion_control
238 struct tcp_congestion_ops tcp_init_congestion_ops = {
240 .owner = THIS_MODULE,
241 .ssthresh = tcp_reno_ssthresh,
242 .cong_avoid = tcp_reno_cong_avoid,
243 .min_cwnd = tcp_reno_min_cwnd,
245 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);