2 * Plugable TCP congestion control support and newReno
4 * Based on ideas from I/O scheduler suport and Web100.
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);
116 if (!ca && capable(CAP_SYS_MODULE)) {
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 ca->non_restricted = 1; /* default is always allowed */
127 list_move(&ca->list, &tcp_cong_list);
130 spin_unlock(&tcp_cong_list_lock);
135 /* Set default value from kernel configuration at bootup */
136 static int __init tcp_congestion_default(void)
138 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
140 late_initcall(tcp_congestion_default);
143 /* Build string with list of available congestion control values */
144 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
146 struct tcp_congestion_ops *ca;
150 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
151 offs += snprintf(buf + offs, maxlen - offs,
153 offs == 0 ? "" : " ", ca->name);
159 /* Get current default congestion control */
160 void tcp_get_default_congestion_control(char *name)
162 struct tcp_congestion_ops *ca;
163 /* We will always have reno... */
164 BUG_ON(list_empty(&tcp_cong_list));
167 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
168 strncpy(name, ca->name, TCP_CA_NAME_MAX);
172 /* Built list of non-restricted congestion control values */
173 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
175 struct tcp_congestion_ops *ca;
180 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
181 if (!ca->non_restricted)
183 offs += snprintf(buf + offs, maxlen - offs,
185 offs == 0 ? "" : " ", ca->name);
191 /* Change list of non-restricted congestion control */
192 int tcp_set_allowed_congestion_control(char *val)
194 struct tcp_congestion_ops *ca;
198 clone = kstrdup(val, GFP_USER);
202 spin_lock(&tcp_cong_list_lock);
203 /* pass 1 check for bad entries */
204 while ((name = strsep(&clone, " ")) && *name) {
205 ca = tcp_ca_find(name);
213 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
214 ca->non_restricted = 0;
216 /* pass 3 mark as allowed */
217 while ((name = strsep(&val, " ")) && *name) {
218 ca = tcp_ca_find(name);
221 ca->non_restricted = 1;
224 spin_unlock(&tcp_cong_list_lock);
230 /* Change congestion control for socket */
231 int tcp_set_congestion_control(struct sock *sk, const char *name)
233 struct inet_connection_sock *icsk = inet_csk(sk);
234 struct tcp_congestion_ops *ca;
238 ca = tcp_ca_find(name);
239 /* no change asking for existing value */
240 if (ca == icsk->icsk_ca_ops)
244 /* not found attempt to autoload module */
245 if (!ca && capable(CAP_SYS_MODULE)) {
247 request_module("tcp_%s", name);
249 ca = tcp_ca_find(name);
255 else if (!(ca->non_restricted || capable(CAP_NET_ADMIN)))
258 else if (!try_module_get(ca->owner))
262 tcp_cleanup_congestion_control(sk);
263 icsk->icsk_ca_ops = ca;
264 if (icsk->icsk_ca_ops->init)
265 icsk->icsk_ca_ops->init(sk);
274 * Linear increase during slow start
276 void tcp_slow_start(struct tcp_sock *tp)
278 if (sysctl_tcp_abc) {
279 /* RFC3465: Slow Start
280 * TCP sender SHOULD increase cwnd by the number of
281 * previously unacknowledged bytes ACKed by each incoming
282 * acknowledgment, provided the increase is not more than L
284 if (tp->bytes_acked < tp->mss_cache)
287 /* We MAY increase by 2 if discovered delayed ack */
288 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache) {
289 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
295 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
298 EXPORT_SYMBOL_GPL(tcp_slow_start);
301 * TCP Reno congestion control
302 * This is special case used for fallback as well.
304 /* This is Jacobson's slow start and congestion avoidance.
305 * SIGCOMM '88, p. 328.
307 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
310 struct tcp_sock *tp = tcp_sk(sk);
312 if (!tcp_is_cwnd_limited(sk, in_flight))
315 /* In "safe" area, increase. */
316 if (tp->snd_cwnd <= tp->snd_ssthresh)
319 /* In dangerous area, increase slowly. */
320 else if (sysctl_tcp_abc) {
321 /* RFC3465: Appropriate Byte Count
322 * increase once for each full cwnd acked
324 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
325 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
326 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
330 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
331 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
332 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
334 tp->snd_cwnd_cnt = 0;
339 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
341 /* Slow start threshold is half the congestion window (min 2) */
342 u32 tcp_reno_ssthresh(struct sock *sk)
344 const struct tcp_sock *tp = tcp_sk(sk);
345 return max(tp->snd_cwnd >> 1U, 2U);
347 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
349 /* Lower bound on congestion window with halving. */
350 u32 tcp_reno_min_cwnd(const struct sock *sk)
352 const struct tcp_sock *tp = tcp_sk(sk);
353 return tp->snd_ssthresh/2;
355 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
357 struct tcp_congestion_ops tcp_reno = {
360 .owner = THIS_MODULE,
361 .ssthresh = tcp_reno_ssthresh,
362 .cong_avoid = tcp_reno_cong_avoid,
363 .min_cwnd = tcp_reno_min_cwnd,
366 /* Initial congestion control used (until SYN)
367 * really reno under another name so we can tell difference
368 * during tcp_set_default_congestion_control
370 struct tcp_congestion_ops tcp_init_congestion_ops = {
372 .owner = THIS_MODULE,
373 .ssthresh = tcp_reno_ssthresh,
374 .cong_avoid = tcp_reno_cong_avoid,
375 .min_cwnd = tcp_reno_min_cwnd,
377 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);