On 11/16/12 15:17, Ben Hutchings wrote:

The receive paths for skbs with inline and out-of-line VLAN tags (VLAN
RX accleration) were made largely consistent in 2.6.37, with tags
pulled out by software as necessary.  However GRO doesn't do this, so
it is not effective for VLAN-tagged packets received on devices
without VLAN RX acceleration.

napi_gro_frags() must not free the skb and does not advance the
skb->data pointer, so cannot use vlan_untag().  Extract the core of
vlan_untag() into a new function __vlan_untag() that allows the offset
to the VLAN tag to be specified and returns an error code.  Add
kernel-doc comments for both those functions.

Signed-off-by: Ben Hutchings <redacted>
---
Tested with sfc using both napi_gro_receive() and napi_gro_frags().  On
a Core i7 920 (Nehalem) system it increased TCP/IPv4 receive throughput
over a VLAN from ~8.0 to ~9.3 Gbit/s.

I verified similar results on myri10ge, using my recent GRO
patchset (minus the in-driver vtag removal) with just
napi_gro_frags().

I've no strong feeling as to whether or not this belongs in GRO or the
driver.  I'm just glad that it is being discussed.

Thank you,

Drew