2 * Plugable TCP congestion control support and newReno
4 * Based on ideas from I/O scheduler support and Web100.
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
9 #define pr_fmt(fmt) "TCP: " fmt
11 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/list.h>
15 #include <linux/gfp.h>
18 static DEFINE_SPINLOCK(tcp_cong_list_lock);
19 static LIST_HEAD(tcp_cong_list);
21 /* Simple linear search, don't expect many entries! */
22 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
24 struct tcp_congestion_ops *e;
26 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
27 if (strcmp(e->name, name) == 0)
35 * Attach new congestion control algorithm to the list
36 * of available options.
38 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
42 /* all algorithms must implement ssthresh and cong_avoid ops */
43 if (!ca->ssthresh || !ca->cong_avoid) {
44 pr_err("%s does not implement required ops\n", ca->name);
48 spin_lock(&tcp_cong_list_lock);
49 if (tcp_ca_find(ca->name)) {
50 pr_notice("%s already registered\n", ca->name);
53 list_add_tail_rcu(&ca->list, &tcp_cong_list);
54 pr_info("%s registered\n", ca->name);
56 spin_unlock(&tcp_cong_list_lock);
60 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
63 * Remove congestion control algorithm, called from
64 * the module's remove function. Module ref counts are used
65 * to ensure that this can't be done till all sockets using
66 * that method are closed.
68 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
70 spin_lock(&tcp_cong_list_lock);
71 list_del_rcu(&ca->list);
72 spin_unlock(&tcp_cong_list_lock);
74 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
76 /* Assign choice of congestion control. */
77 void tcp_assign_congestion_control(struct sock *sk)
79 struct inet_connection_sock *icsk = inet_csk(sk);
80 struct tcp_congestion_ops *ca;
83 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
84 if (likely(try_module_get(ca->owner))) {
85 icsk->icsk_ca_ops = ca;
88 /* Fallback to next available. The last really
89 * guaranteed fallback is Reno from this list.
95 /* Clear out private data before diag gets it and
96 * the ca has not been initialized.
99 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
102 void tcp_init_congestion_control(struct sock *sk)
104 const struct inet_connection_sock *icsk = inet_csk(sk);
106 if (icsk->icsk_ca_ops->init)
107 icsk->icsk_ca_ops->init(sk);
110 /* Manage refcounts on socket close. */
111 void tcp_cleanup_congestion_control(struct sock *sk)
113 struct inet_connection_sock *icsk = inet_csk(sk);
115 if (icsk->icsk_ca_ops->release)
116 icsk->icsk_ca_ops->release(sk);
117 module_put(icsk->icsk_ca_ops->owner);
120 /* Used by sysctl to change default congestion control */
121 int tcp_set_default_congestion_control(const char *name)
123 struct tcp_congestion_ops *ca;
126 spin_lock(&tcp_cong_list_lock);
127 ca = tcp_ca_find(name);
128 #ifdef CONFIG_MODULES
129 if (!ca && capable(CAP_NET_ADMIN)) {
130 spin_unlock(&tcp_cong_list_lock);
132 request_module("tcp_%s", name);
133 spin_lock(&tcp_cong_list_lock);
134 ca = tcp_ca_find(name);
139 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
140 list_move(&ca->list, &tcp_cong_list);
143 spin_unlock(&tcp_cong_list_lock);
148 /* Set default value from kernel configuration at bootup */
149 static int __init tcp_congestion_default(void)
151 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
153 late_initcall(tcp_congestion_default);
155 /* Build string with list of available congestion control values */
156 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
158 struct tcp_congestion_ops *ca;
162 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
163 offs += snprintf(buf + offs, maxlen - offs,
165 offs == 0 ? "" : " ", ca->name);
170 /* Get current default congestion control */
171 void tcp_get_default_congestion_control(char *name)
173 struct tcp_congestion_ops *ca;
174 /* We will always have reno... */
175 BUG_ON(list_empty(&tcp_cong_list));
178 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
179 strncpy(name, ca->name, TCP_CA_NAME_MAX);
183 /* Built list of non-restricted congestion control values */
184 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
186 struct tcp_congestion_ops *ca;
191 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
192 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
194 offs += snprintf(buf + offs, maxlen - offs,
196 offs == 0 ? "" : " ", ca->name);
201 /* Change list of non-restricted congestion control */
202 int tcp_set_allowed_congestion_control(char *val)
204 struct tcp_congestion_ops *ca;
205 char *saved_clone, *clone, *name;
208 saved_clone = clone = kstrdup(val, GFP_USER);
212 spin_lock(&tcp_cong_list_lock);
213 /* pass 1 check for bad entries */
214 while ((name = strsep(&clone, " ")) && *name) {
215 ca = tcp_ca_find(name);
222 /* pass 2 clear old values */
223 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
224 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
226 /* pass 3 mark as allowed */
227 while ((name = strsep(&val, " ")) && *name) {
228 ca = tcp_ca_find(name);
231 ca->flags |= TCP_CONG_NON_RESTRICTED;
234 spin_unlock(&tcp_cong_list_lock);
240 /* Change congestion control for socket */
241 int tcp_set_congestion_control(struct sock *sk, const char *name)
243 struct inet_connection_sock *icsk = inet_csk(sk);
244 struct tcp_congestion_ops *ca;
248 ca = tcp_ca_find(name);
250 /* no change asking for existing value */
251 if (ca == icsk->icsk_ca_ops)
254 #ifdef CONFIG_MODULES
255 /* not found attempt to autoload module */
256 if (!ca && capable(CAP_NET_ADMIN)) {
258 request_module("tcp_%s", name);
260 ca = tcp_ca_find(name);
266 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
267 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
270 else if (!try_module_get(ca->owner))
274 tcp_cleanup_congestion_control(sk);
275 icsk->icsk_ca_ops = ca;
277 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
278 icsk->icsk_ca_ops->init(sk);
285 /* Slow start is used when congestion window is no greater than the slow start
286 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
287 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
288 * something better;) a packet is only considered (s)acked in its entirety to
289 * defend the ACK attacks described in the RFC. Slow start processes a stretch
290 * ACK of degree N as if N acks of degree 1 are received back to back except
291 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
292 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
294 void tcp_slow_start(struct tcp_sock *tp, u32 acked)
296 u32 cwnd = tp->snd_cwnd + acked;
298 if (cwnd > tp->snd_ssthresh)
299 cwnd = tp->snd_ssthresh + 1;
300 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
302 EXPORT_SYMBOL_GPL(tcp_slow_start);
304 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
305 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
307 if (tp->snd_cwnd_cnt >= w) {
308 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
310 tp->snd_cwnd_cnt = 0;
315 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
318 * TCP Reno congestion control
319 * This is special case used for fallback as well.
321 /* This is Jacobson's slow start and congestion avoidance.
322 * SIGCOMM '88, p. 328.
324 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
326 struct tcp_sock *tp = tcp_sk(sk);
328 if (!tcp_is_cwnd_limited(sk))
331 /* In "safe" area, increase. */
332 if (tp->snd_cwnd <= tp->snd_ssthresh)
333 tcp_slow_start(tp, acked);
334 /* In dangerous area, increase slowly. */
336 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
338 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
340 /* Slow start threshold is half the congestion window (min 2) */
341 u32 tcp_reno_ssthresh(struct sock *sk)
343 const struct tcp_sock *tp = tcp_sk(sk);
345 return max(tp->snd_cwnd >> 1U, 2U);
347 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
349 struct tcp_congestion_ops tcp_reno = {
350 .flags = TCP_CONG_NON_RESTRICTED,
352 .owner = THIS_MODULE,
353 .ssthresh = tcp_reno_ssthresh,
354 .cong_avoid = tcp_reno_cong_avoid,