rps: tcp: fix rps_sock_flow_table table updates

I believe a moderate SYN flood attack can corrupt RFS flow table
(rps_sock_flow_table), making RPS/RFS much less effective.

Even in a normal situation, server handling short lived sessions suffer
from bad steering for the first data packet of a session, if another SYN
packet is received for another session.

We do following action in tcp_v4_rcv() :

	sock_rps_save_rxhash(sk, skb->rxhash);

We should _not_ do this if sk is a LISTEN socket, as about each
packet received on a LISTEN socket has a different rxhash than
previous one.
 -> RPS_NO_CPU markers are spread all over rps_sock_flow_table.

Also, it makes sense to protect sk->rxhash field changes with socket
lock (We currently can change it even if user thread owns the lock
and might use rxhash)

This patch moves sock_rps_save_rxhash() to a sock locked section,
and only for non LISTEN sockets.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index 202cf09..fe193e5 100644 (file)
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -1555,6 +1555,7 @@ int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
 #endif
 
        if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
+               sock_rps_save_rxhash(sk, skb->rxhash);
                TCP_CHECK_TIMER(sk);
                if (tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len)) {
                        rsk = sk;
@@ -1579,7 +1580,9 @@ int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
                        }
                        return 0;
                }
-       }
+       } else
+               sock_rps_save_rxhash(sk, skb->rxhash);
+
 
        TCP_CHECK_TIMER(sk);
        if (tcp_rcv_state_process(sk, skb, tcp_hdr(skb), skb->len)) {
@@ -1672,8 +1675,6 @@ process:
 
        skb->dev = NULL;
 
-       sock_rps_save_rxhash(sk, skb->rxhash);
-
        bh_lock_sock_nested(sk);
        ret = 0;
        if (!sock_owned_by_user(sk)) {