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/module.h>
11 #include <linux/types.h>
12 #include <linux/list.h>
15 static DEFINE_SPINLOCK(tcp_cong_list_lock);
16 static LIST_HEAD(tcp_cong_list);
18 /* Simple linear search, don't expect many entries! */
19 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
21 struct tcp_congestion_ops *e;
23 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
24 if (strcmp(e->name, name) == 0)
32 * Attach new congestion control algorthim to the list
33 * of available options.
35 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
39 /* all algorithms must implement ssthresh and cong_avoid ops */
40 if (!ca->ssthresh || !ca->cong_avoid) {
41 printk(KERN_ERR "TCP %s does not implement required ops\n",
46 spin_lock(&tcp_cong_list_lock);
47 if (tcp_ca_find(ca->name)) {
48 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
51 list_add_tail_rcu(&ca->list, &tcp_cong_list);
52 printk(KERN_INFO "TCP %s registered\n", ca->name);
54 spin_unlock(&tcp_cong_list_lock);
58 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
61 * Remove congestion control algorithm, called from
62 * the module's remove function. Module ref counts are used
63 * to ensure that this can't be done till all sockets using
64 * that method are closed.
66 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
68 spin_lock(&tcp_cong_list_lock);
69 list_del_rcu(&ca->list);
70 spin_unlock(&tcp_cong_list_lock);
72 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
74 /* Assign choice of congestion control. */
75 void tcp_init_congestion_control(struct sock *sk)
77 struct inet_connection_sock *icsk = inet_csk(sk);
78 struct tcp_congestion_ops *ca;
80 if (icsk->icsk_ca_ops != &tcp_init_congestion_ops)
84 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
85 if (try_module_get(ca->owner)) {
86 icsk->icsk_ca_ops = ca;
93 if (icsk->icsk_ca_ops->init)
94 icsk->icsk_ca_ops->init(sk);
97 /* Manage refcounts on socket close. */
98 void tcp_cleanup_congestion_control(struct sock *sk)
100 struct inet_connection_sock *icsk = inet_csk(sk);
102 if (icsk->icsk_ca_ops->release)
103 icsk->icsk_ca_ops->release(sk);
104 module_put(icsk->icsk_ca_ops->owner);
107 /* Used by sysctl to change default congestion control */
108 int tcp_set_default_congestion_control(const char *name)
110 struct tcp_congestion_ops *ca;
113 spin_lock(&tcp_cong_list_lock);
114 ca = tcp_ca_find(name);
117 spin_unlock(&tcp_cong_list_lock);
119 request_module("tcp_%s", name);
120 spin_lock(&tcp_cong_list_lock);
121 ca = tcp_ca_find(name);
126 list_move(&ca->list, &tcp_cong_list);
129 spin_unlock(&tcp_cong_list_lock);
134 /* Set default value from kernel configuration at bootup */
135 static int __init tcp_congestion_default(void)
137 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
139 late_initcall(tcp_congestion_default);
142 /* Build string with list of available congestion control values */
143 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
145 struct tcp_congestion_ops *ca;
149 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
150 offs += snprintf(buf + offs, maxlen - offs,
152 offs == 0 ? "" : " ", ca->name);
158 /* Get current default congestion control */
159 void tcp_get_default_congestion_control(char *name)
161 struct tcp_congestion_ops *ca;
162 /* We will always have reno... */
163 BUG_ON(list_empty(&tcp_cong_list));
166 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
167 strncpy(name, ca->name, TCP_CA_NAME_MAX);
171 /* Change congestion control for socket */
172 int tcp_set_congestion_control(struct sock *sk, const char *name)
174 struct inet_connection_sock *icsk = inet_csk(sk);
175 struct tcp_congestion_ops *ca;
179 ca = tcp_ca_find(name);
180 if (ca == icsk->icsk_ca_ops)
186 else if (!try_module_get(ca->owner))
190 tcp_cleanup_congestion_control(sk);
191 icsk->icsk_ca_ops = ca;
192 if (icsk->icsk_ca_ops->init)
193 icsk->icsk_ca_ops->init(sk);
202 * Linear increase during slow start
204 void tcp_slow_start(struct tcp_sock *tp)
206 if (sysctl_tcp_abc) {
207 /* RFC3465: Slow Start
208 * TCP sender SHOULD increase cwnd by the number of
209 * previously unacknowledged bytes ACKed by each incoming
210 * acknowledgment, provided the increase is not more than L
212 if (tp->bytes_acked < tp->mss_cache)
215 /* We MAY increase by 2 if discovered delayed ack */
216 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache) {
217 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
223 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
226 EXPORT_SYMBOL_GPL(tcp_slow_start);
229 * TCP Reno congestion control
230 * This is special case used for fallback as well.
232 /* This is Jacobson's slow start and congestion avoidance.
233 * SIGCOMM '88, p. 328.
235 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
238 struct tcp_sock *tp = tcp_sk(sk);
240 if (!tcp_is_cwnd_limited(sk, in_flight))
243 /* In "safe" area, increase. */
244 if (tp->snd_cwnd <= tp->snd_ssthresh)
247 /* In dangerous area, increase slowly. */
248 else if (sysctl_tcp_abc) {
249 /* RFC3465: Appropriate Byte Count
250 * increase once for each full cwnd acked
252 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
253 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
254 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
258 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
259 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
260 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
262 tp->snd_cwnd_cnt = 0;
267 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
269 /* Slow start threshold is half the congestion window (min 2) */
270 u32 tcp_reno_ssthresh(struct sock *sk)
272 const struct tcp_sock *tp = tcp_sk(sk);
273 return max(tp->snd_cwnd >> 1U, 2U);
275 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
277 /* Lower bound on congestion window with halving. */
278 u32 tcp_reno_min_cwnd(const struct sock *sk)
280 const struct tcp_sock *tp = tcp_sk(sk);
281 return tp->snd_ssthresh/2;
283 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
285 struct tcp_congestion_ops tcp_reno = {
287 .owner = THIS_MODULE,
288 .ssthresh = tcp_reno_ssthresh,
289 .cong_avoid = tcp_reno_cong_avoid,
290 .min_cwnd = tcp_reno_min_cwnd,
293 /* Initial congestion control used (until SYN)
294 * really reno under another name so we can tell difference
295 * during tcp_set_default_congestion_control
297 struct tcp_congestion_ops tcp_init_congestion_ops = {
299 .owner = THIS_MODULE,
300 .ssthresh = tcp_reno_ssthresh,
301 .cong_avoid = tcp_reno_cong_avoid,
302 .min_cwnd = tcp_reno_min_cwnd,
304 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);