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(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 tcp_sock *tp)
78 struct tcp_congestion_ops *ca;
81 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
82 if (try_module_get(ca->owner)) {
94 /* Manage refcounts on socket close. */
95 void tcp_cleanup_congestion_control(struct tcp_sock *tp)
97 if (tp->ca_ops->release)
98 tp->ca_ops->release(tp);
99 module_put(tp->ca_ops->owner);
102 /* Used by sysctl to change default congestion control */
103 int tcp_set_default_congestion_control(const char *name)
105 struct tcp_congestion_ops *ca;
108 spin_lock(&tcp_cong_list_lock);
109 ca = tcp_ca_find(name);
112 spin_unlock(&tcp_cong_list_lock);
114 request_module("tcp_%s", name);
115 spin_lock(&tcp_cong_list_lock);
116 ca = tcp_ca_find(name);
121 list_move(&ca->list, &tcp_cong_list);
124 spin_unlock(&tcp_cong_list_lock);
129 /* Get current default congestion control */
130 void tcp_get_default_congestion_control(char *name)
132 struct tcp_congestion_ops *ca;
133 /* We will always have reno... */
134 BUG_ON(list_empty(&tcp_cong_list));
137 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
138 strncpy(name, ca->name, TCP_CA_NAME_MAX);
143 * TCP Reno congestion control
144 * This is special case used for fallback as well.
146 /* This is Jacobson's slow start and congestion avoidance.
147 * SIGCOMM '88, p. 328.
149 void tcp_reno_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt, u32 in_flight,
152 if (in_flight < tp->snd_cwnd)
155 if (tp->snd_cwnd <= tp->snd_ssthresh) {
156 /* In "safe" area, increase. */
157 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
160 /* In dangerous area, increase slowly.
161 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
163 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
164 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
166 tp->snd_cwnd_cnt = 0;
171 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
173 /* Slow start threshold is half the congestion window (min 2) */
174 u32 tcp_reno_ssthresh(struct tcp_sock *tp)
176 return max(tp->snd_cwnd >> 1U, 2U);
178 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
180 /* Lower bound on congestion window. */
181 u32 tcp_reno_min_cwnd(struct tcp_sock *tp)
183 return tp->snd_ssthresh/2;
185 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
187 struct tcp_congestion_ops tcp_reno = {
189 .owner = THIS_MODULE,
190 .ssthresh = tcp_reno_ssthresh,
191 .cong_avoid = tcp_reno_cong_avoid,
192 .min_cwnd = tcp_reno_min_cwnd,
195 EXPORT_SYMBOL_GPL(tcp_reno);