[TCP]: Sanitize initcwnd calculation, add new metrics.

1) Remove srtt etc. tests in tcp_init_cwnd, was buggy
   anyways.
2) Add RTAX_INITCWMD route metric, use this in tcp_init_cwnd()
   if non-zero.
3) Add RTAX_FEATURES (to enable/disable ECN/SACK/TIMESTAMPS
   on a per-route basis), currently unused.

include/linux/rtnetlink.h

@@ -284,9 +284,17 @@ enum
 #define RTAX_REORDERING RTAX_REORDERING
 	RTAX_HOPLIMIT,
 #define RTAX_HOPLIMIT RTAX_HOPLIMIT
+	RTAX_INITCWND,
+#define RTAX_INITCWND RTAX_INITCWND
+	RTAX_FEATURES,
+#define RTAX_FEATURES RTAX_FEATURES
 };
 
-#define RTAX_MAX RTAX_HOPLIMIT
+#define RTAX_MAX RTAX_FEATURES
+
+#define RTAX_FEATURE_ECN	0x00000001
+#define RTAX_FEATURE_SACK	0x00000002
+#define RTAX_FEATURE_TIMESTAMP	0x00000004
 
 struct rta_session
 {

include/net/tcp.h

@@ -1244,7 +1244,7 @@ static inline void tcp_enter_cwr(struct tcp_opt *tp)
 	}
 }
 
-extern __u32 tcp_init_cwnd(struct tcp_opt *tp);
+extern __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst);
 
 /* Slow start with delack produces 3 packets of burst, so that
  * it is safe "de facto".

net/ipv4/tcp_input.c

@@ -598,25 +598,17 @@ void tcp_update_metrics(struct sock *sk)
 	}
 }
 
-/* Increase initial CWND conservatively: if estimated
- * RTT is low enough (<20msec) or if we have some preset ssthresh.
- *
- * Numbers are taken from RFC2414.
- */
-__u32 tcp_init_cwnd(struct tcp_opt *tp)
+/* Numbers are taken from RFC2414.  */
+__u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst)
 {
-	__u32 cwnd;
-
-	if (tp->mss_cache > 1460)
-		return 2;
-
-	cwnd = (tp->mss_cache > 1095) ? 3 : 4;
-
-	if (!tp->srtt || (tp->snd_ssthresh >= 0xFFFF && tp->srtt > ((HZ/50)<<3)))
-		cwnd = 2;
-	else if (cwnd > tp->snd_ssthresh)
-		cwnd = tp->snd_ssthresh;
+	__u32 cwnd = dst_metric(dst, RTAX_INITCWND);
 
+	if (!cwnd) {
+		if (tp->mss_cache > 1460)
+			cwnd = 2;
+		else
+			cwnd = (tp->mss_cache > 1095) ? 3 : 4;
+	}
 	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 }
 
@@ -675,7 +667,7 @@ static void tcp_init_metrics(struct sock *sk)
 	tcp_bound_rto(tp);
 	if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
 		goto reset;
-	tp->snd_cwnd = tcp_init_cwnd(tp);
+	tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 	tp->snd_cwnd_stamp = tcp_time_stamp;
 	return;
 

net/ipv4/tcp_output.c

@@ -99,10 +99,10 @@ static __u16 tcp_advertise_mss(struct sock *sk)
 
 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  * This is the first part of cwnd validation mechanism. */
-static void tcp_cwnd_restart(struct tcp_opt *tp)
+static void tcp_cwnd_restart(struct tcp_opt *tp, struct dst_entry *dst)
 {
 	s32 delta = tcp_time_stamp - tp->lsndtime;
-	u32 restart_cwnd = tcp_init_cwnd(tp);
+	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
 	u32 cwnd = tp->snd_cwnd;
 
 	tp->snd_ssthresh = tcp_current_ssthresh(tp);
@@ -115,12 +115,12 @@ static void tcp_cwnd_restart(struct tcp_opt *tp)
 	tp->snd_cwnd_used = 0;
 }
 
-static __inline__ void tcp_event_data_sent(struct tcp_opt *tp, struct sk_buff *skb)
+static __inline__ void tcp_event_data_sent(struct tcp_opt *tp, struct sk_buff *skb, struct sock *sk)
 {
 	u32 now = tcp_time_stamp;
 
 	if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto)
-		tcp_cwnd_restart(tp);
+		tcp_cwnd_restart(tp, __sk_dst_get(sk));
 
 	tp->lsndtime = now;
 
@@ -272,7 +272,7 @@ int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
 			tcp_event_ack_sent(sk);
 
 		if (skb->len != tcp_header_size)
-			tcp_event_data_sent(tp, skb);
+			tcp_event_data_sent(tp, skb, sk);
 
 		TCP_INC_STATS(TcpOutSegs);