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 algorithm 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 no choice made yet assign the current value set as default */
81 if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
83 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
84 if (try_module_get(ca->owner)) {
85 icsk->icsk_ca_ops = ca;
89 /* fallback to next available */
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);
117 if (!ca && capable(CAP_SYS_MODULE)) {
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 ca->non_restricted = 1; /* default is always allowed */
128 list_move(&ca->list, &tcp_cong_list);
131 spin_unlock(&tcp_cong_list_lock);
136 /* Set default value from kernel configuration at bootup */
137 static int __init tcp_congestion_default(void)
139 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
141 late_initcall(tcp_congestion_default);
144 /* Build string with list of available congestion control values */
145 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
147 struct tcp_congestion_ops *ca;
151 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
152 offs += snprintf(buf + offs, maxlen - offs,
154 offs == 0 ? "" : " ", ca->name);
160 /* Get current default congestion control */
161 void tcp_get_default_congestion_control(char *name)
163 struct tcp_congestion_ops *ca;
164 /* We will always have reno... */
165 BUG_ON(list_empty(&tcp_cong_list));
168 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
169 strncpy(name, ca->name, TCP_CA_NAME_MAX);
173 /* Built list of non-restricted congestion control values */
174 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
176 struct tcp_congestion_ops *ca;
181 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
182 if (!ca->non_restricted)
184 offs += snprintf(buf + offs, maxlen - offs,
186 offs == 0 ? "" : " ", ca->name);
192 /* Change list of non-restricted congestion control */
193 int tcp_set_allowed_congestion_control(char *val)
195 struct tcp_congestion_ops *ca;
199 clone = kstrdup(val, GFP_USER);
203 spin_lock(&tcp_cong_list_lock);
204 /* pass 1 check for bad entries */
205 while ((name = strsep(&clone, " ")) && *name) {
206 ca = tcp_ca_find(name);
214 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
215 ca->non_restricted = 0;
217 /* pass 3 mark as allowed */
218 while ((name = strsep(&val, " ")) && *name) {
219 ca = tcp_ca_find(name);
222 ca->non_restricted = 1;
225 spin_unlock(&tcp_cong_list_lock);
231 /* Change congestion control for socket */
232 int tcp_set_congestion_control(struct sock *sk, const char *name)
234 struct inet_connection_sock *icsk = inet_csk(sk);
235 struct tcp_congestion_ops *ca;
239 ca = tcp_ca_find(name);
241 /* no change asking for existing value */
242 if (ca == icsk->icsk_ca_ops)
246 /* not found attempt to autoload module */
247 if (!ca && capable(CAP_SYS_MODULE)) {
249 request_module("tcp_%s", name);
251 ca = tcp_ca_find(name);
257 else if (!(ca->non_restricted || capable(CAP_NET_ADMIN)))
260 else if (!try_module_get(ca->owner))
264 tcp_cleanup_congestion_control(sk);
265 icsk->icsk_ca_ops = ca;
267 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
268 icsk->icsk_ca_ops->init(sk);
277 * Linear increase during slow start
279 void tcp_slow_start(struct tcp_sock *tp)
281 if (sysctl_tcp_abc) {
282 /* RFC3465: Slow Start
283 * TCP sender SHOULD increase cwnd by the number of
284 * previously unacknowledged bytes ACKed by each incoming
285 * acknowledgment, provided the increase is not more than L
287 if (tp->bytes_acked < tp->mss_cache)
290 /* We MAY increase by 2 if discovered delayed ack */
291 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache) {
292 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
298 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
301 EXPORT_SYMBOL_GPL(tcp_slow_start);
304 * TCP Reno congestion control
305 * This is special case used for fallback as well.
307 /* This is Jacobson's slow start and congestion avoidance.
308 * SIGCOMM '88, p. 328.
310 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
313 struct tcp_sock *tp = tcp_sk(sk);
315 if (!tcp_is_cwnd_limited(sk, in_flight))
318 /* In "safe" area, increase. */
319 if (tp->snd_cwnd <= tp->snd_ssthresh)
322 /* In dangerous area, increase slowly. */
323 else if (sysctl_tcp_abc) {
324 /* RFC3465: Appropriate Byte Count
325 * increase once for each full cwnd acked
327 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
328 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
329 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
333 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
334 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
335 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
337 tp->snd_cwnd_cnt = 0;
342 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
344 /* Slow start threshold is half the congestion window (min 2) */
345 u32 tcp_reno_ssthresh(struct sock *sk)
347 const struct tcp_sock *tp = tcp_sk(sk);
348 return max(tp->snd_cwnd >> 1U, 2U);
350 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
352 /* Lower bound on congestion window with halving. */
353 u32 tcp_reno_min_cwnd(const struct sock *sk)
355 const struct tcp_sock *tp = tcp_sk(sk);
356 return tp->snd_ssthresh/2;
358 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
360 struct tcp_congestion_ops tcp_reno = {
363 .owner = THIS_MODULE,
364 .ssthresh = tcp_reno_ssthresh,
365 .cong_avoid = tcp_reno_cong_avoid,
366 .min_cwnd = tcp_reno_min_cwnd,
369 /* Initial congestion control used (until SYN)
370 * really reno under another name so we can tell difference
371 * during tcp_set_default_congestion_control
373 struct tcp_congestion_ops tcp_init_congestion_ops = {
375 .owner = THIS_MODULE,
376 .ssthresh = tcp_reno_ssthresh,
377 .cong_avoid = tcp_reno_cong_avoid,
378 .min_cwnd = tcp_reno_min_cwnd,
380 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);