net-gro: restore frag0 optimization
authorEric Dumazet <edumazet@google.com>
Sun, 30 Mar 2014 04:28:21 +0000 (21:28 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 31 Mar 2014 20:26:40 +0000 (16:26 -0400)
commita50e233c50dbc881abaa0e4070789064e8d12d70
tree5c4c25d28e0f08f258ff8304fd4582658c1a0780
parentbf39b4247b8799935ea91d90db250ab608a58e50
net-gro: restore frag0 optimization

Main difference between napi_frags_skb() and napi_gro_receive() is that
the later is called while ethernet header was already pulled by the NIC
driver (eth_type_trans() was called before napi_gro_receive())

Jerry Chu in commit 299603e8370a ("net-gro: Prepare GRO stack for the
upcoming tunneling support") tried to remove this difference by calling
eth_type_trans() from napi_frags_skb() instead of doing this later from
napi_frags_finish()

Goal was that napi_gro_complete() could call
ptype->callbacks.gro_complete(skb, 0)  (offset of first network header =
0)

Also, xxx_gro_receive() handlers all use off = skb_gro_offset(skb) to
point to their own header, for the current skb and ones held in gro_list

Problem is this cleanup work defeated the frag0 optimization:
It turns out the consecutive pskb_may_pull() calls are too expensive.

This patch brings back the frag0 stuff in napi_frags_skb().

As all skb have their mac header in skb head, we no longer need
skb_gro_mac_header()

Reported-by: Michal Schmidt <mschmidt@redhat.com>
Fixes: 299603e8370a ("net-gro: Prepare GRO stack for the upcoming tunneling support")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jerry Chu <hkchu@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/netdevice.h
net/core/dev.c