ipv4: set transport header earlier

[ Upstream commit 21d1196a ]

commit 45f00f99 ("ipv4: tcp: clean up tcp_v4_early_demux()") added a
performance regression for non GRO traffic, basically disabling
IP early demux.

IPv6 stack resets transport header in ip6_rcv() before calling
IP early demux in ip6_rcv_finish(), while IPv4 does this only in
ip_local_deliver_finish(), _after_ IP early demux.

GRO traffic happened to enable IP early demux because transport header
is also set in inet_gro_receive()

Instead of reverting the faulty commit, we can make IPv4/IPv6 behave the
same : transport_header should be set in ip_rcv() instead of
ip_local_deliver_finish()

ip_local_deliver_finish() can also use skb_network_header_len() which is
faster than ip_hdrlen()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

net/ipv4/ip_input.c

@@ -190,10 +190,7 @@ static int ip_local_deliver_finish(struct sk_buff *skb)
 {
 	struct net *net = dev_net(skb->dev);
 
-	__skb_pull(skb, ip_hdrlen(skb));
-
-	/* Point into the IP datagram, just past the header. */
-	skb_reset_transport_header(skb);
+	__skb_pull(skb, skb_network_header_len(skb));
 
 	rcu_read_lock();
 	{
@@ -437,6 +434,8 @@ int ip_rcv(struct sk_buff *skb, struct net_device *dev, struct packet_type *pt,
 		goto drop;
 	}
 
+	skb->transport_header = skb->network_header + iph->ihl*4;
+
 	/* Remove any debris in the socket control block */
 	memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));