: rfs: record flow in TCP receiving and sending pathes

rfs: record flow in TCP receiving and sending pathes

call sock_rps_record_flow() in function tcp_splice_read(), tcp_sendpage() and tcp_sendmsg().

----

 net/ipv4/tcp.c |    3 +++
 1 file changed, 3 insertions(+)

index 4e6ddfb..47537d8 100644
@@ -608,6 +608,7 @@ ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 	ssize_t spliced;
 	int ret;
 
+	sock_rps_record_flow(sk);
 	/*
 	 * We can't seek on a socket input
 	 */

@@ -862,6 +863,7 @@ ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 	ssize_t res;
 	struct sock *sk = sock->sk;
 
+	sock_rps_record_flow(sk);
 	if (!(sk->sk_route_caps & NETIF_F_SG) ||
 	    !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
 		return sock_no_sendpage(sock, page, offset, size, flags);

@@ -909,6 +911,7 @@ int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 	int sg, err, copied;
 	long timeo;
 
+	sock_rps_record_flow(sk);
 	lock_sock(sk);
 	TCP_CHECK_TIMER(sk);
 

From: Changli Gao <[email protected]>
Date: Fri, 9 Jul 2010 15:28:18 +0800

I don't think it's wise to pepper these calls all over the place if it is not necessary.

The only reason we vector straight to the TCP implementations of these I/O routines is to avoid the port autobinding made by the inet_*() functions.

But now that avoids also the RPS calls.

So it makes sense to just add a boolean state bit flag of some sort to "struct proto" which says to avoid the autobind calls, then make TCP vector through the inet_*() functions just like the other inet protocols do.

Then these extra send_rps_record_flow() annotations will not be necessary.