[IPV4]: fib_trie initialzation fix
[linux-2.6.git] / net / ipv4 / tcp_cong.c
index 665394a..e688c68 100644 (file)
@@ -21,7 +21,7 @@ static struct tcp_congestion_ops *tcp_ca_find(const char *name)
 {
        struct tcp_congestion_ops *e;
 
-       list_for_each_entry(e, &tcp_cong_list, list) {
+       list_for_each_entry_rcu(e, &tcp_cong_list, list) {
                if (strcmp(e->name, name) == 0)
                        return e;
        }
@@ -73,30 +73,36 @@ void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
 
 /* Assign choice of congestion control. */
-void tcp_init_congestion_control(struct tcp_sock *tp)
+void tcp_init_congestion_control(struct sock *sk)
 {
+       struct inet_connection_sock *icsk = inet_csk(sk);
        struct tcp_congestion_ops *ca;
 
+       if (icsk->icsk_ca_ops != &tcp_init_congestion_ops)
+               return;
+
        rcu_read_lock();
        list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
                if (try_module_get(ca->owner)) {
-                       tp->ca_ops = ca;
+                       icsk->icsk_ca_ops = ca;
                        break;
                }
 
        }
        rcu_read_unlock();
 
-       if (tp->ca_ops->init)
-               tp->ca_ops->init(tp);
+       if (icsk->icsk_ca_ops->init)
+               icsk->icsk_ca_ops->init(sk);
 }
 
 /* Manage refcounts on socket close. */
-void tcp_cleanup_congestion_control(struct tcp_sock *tp)
+void tcp_cleanup_congestion_control(struct sock *sk)
 {
-       if (tp->ca_ops->release)
-               tp->ca_ops->release(tp);
-       module_put(tp->ca_ops->owner);
+       struct inet_connection_sock *icsk = inet_csk(sk);
+
+       if (icsk->icsk_ca_ops->release)
+               icsk->icsk_ca_ops->release(sk);
+       module_put(icsk->icsk_ca_ops->owner);
 }
 
 /* Used by sysctl to change default congestion control */
@@ -139,6 +145,63 @@ void tcp_get_default_congestion_control(char *name)
        rcu_read_unlock();
 }
 
+/* Change congestion control for socket */
+int tcp_set_congestion_control(struct sock *sk, const char *name)
+{
+       struct inet_connection_sock *icsk = inet_csk(sk);
+       struct tcp_congestion_ops *ca;
+       int err = 0;
+
+       rcu_read_lock();
+       ca = tcp_ca_find(name);
+       if (ca == icsk->icsk_ca_ops)
+               goto out;
+
+       if (!ca)
+               err = -ENOENT;
+
+       else if (!try_module_get(ca->owner))
+               err = -EBUSY;
+
+       else {
+               tcp_cleanup_congestion_control(sk);
+               icsk->icsk_ca_ops = ca;
+               if (icsk->icsk_ca_ops->init)
+                       icsk->icsk_ca_ops->init(sk);
+       }
+ out:
+       rcu_read_unlock();
+       return err;
+}
+
+
+/*
+ * Linear increase during slow start
+ */
+void tcp_slow_start(struct tcp_sock *tp)
+{
+       if (sysctl_tcp_abc) {
+               /* RFC3465: Slow Start
+                * TCP sender SHOULD increase cwnd by the number of
+                * previously unacknowledged bytes ACKed by each incoming
+                * acknowledgment, provided the increase is not more than L
+                */
+               if (tp->bytes_acked < tp->mss_cache)
+                       return;
+
+               /* We MAY increase by 2 if discovered delayed ack */
+               if (sysctl_tcp_abc > 1 && tp->bytes_acked > 2*tp->mss_cache) {
+                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                               tp->snd_cwnd++;
+               }
+       }
+       tp->bytes_acked = 0;
+
+       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+               tp->snd_cwnd++;
+}
+EXPORT_SYMBOL_GPL(tcp_slow_start);
+
 /*
  * TCP Reno congestion control
  * This is special case used for fallback as well.
@@ -146,40 +209,52 @@ void tcp_get_default_congestion_control(char *name)
 /* This is Jacobson's slow start and congestion avoidance.
  * SIGCOMM '88, p. 328.
  */
-void tcp_reno_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt, u32 in_flight,
+void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
                         int flag)
 {
-       if (in_flight < tp->snd_cwnd)
+       struct tcp_sock *tp = tcp_sk(sk);
+
+       if (!tcp_is_cwnd_limited(sk, in_flight))
                return;
 
-        if (tp->snd_cwnd <= tp->snd_ssthresh) {
-                /* In "safe" area, increase. */
-               if (tp->snd_cwnd < tp->snd_cwnd_clamp)
-                       tp->snd_cwnd++;
-       } else {
-                /* In dangerous area, increase slowly.
-                * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
-                */
-               if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
-                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
-                               tp->snd_cwnd++;
-                       tp->snd_cwnd_cnt = 0;
-               } else
-                       tp->snd_cwnd_cnt++;
-       }
+       /* In "safe" area, increase. */
+        if (tp->snd_cwnd <= tp->snd_ssthresh)
+               tcp_slow_start(tp);
+
+       /* In dangerous area, increase slowly. */
+       else if (sysctl_tcp_abc) {
+               /* RFC3465: Apppriate Byte Count
+                * increase once for each full cwnd acked
+                */
+               if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
+                       tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
+                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                               tp->snd_cwnd++;
+               }
+       } else {
+               /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
+               if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
+                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                               tp->snd_cwnd++;
+                       tp->snd_cwnd_cnt = 0;
+               } else
+                       tp->snd_cwnd_cnt++;
+       }
 }
 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 
 /* Slow start threshold is half the congestion window (min 2) */
-u32 tcp_reno_ssthresh(struct tcp_sock *tp)
+u32 tcp_reno_ssthresh(struct sock *sk)
 {
+       const struct tcp_sock *tp = tcp_sk(sk);
        return max(tp->snd_cwnd >> 1U, 2U);
 }
 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 
 /* Lower bound on congestion window. */
-u32 tcp_reno_min_cwnd(struct tcp_sock *tp)
+u32 tcp_reno_min_cwnd(struct sock *sk)
 {
+       const struct tcp_sock *tp = tcp_sk(sk);
        return tp->snd_ssthresh/2;
 }
 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
@@ -192,4 +267,15 @@ struct tcp_congestion_ops tcp_reno = {
        .min_cwnd       = tcp_reno_min_cwnd,
 };
 
-EXPORT_SYMBOL_GPL(tcp_reno);
+/* Initial congestion control used (until SYN)
+ * really reno under another name so we can tell difference
+ * during tcp_set_default_congestion_control
+ */
+struct tcp_congestion_ops tcp_init_congestion_ops  = {
+       .name           = "",
+       .owner          = THIS_MODULE,
+       .ssthresh       = tcp_reno_ssthresh,
+       .cong_avoid     = tcp_reno_cong_avoid,
+       .min_cwnd       = tcp_reno_min_cwnd,
+};
+EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);