On Fri, 2008-03-07 at 08:25 -0800, Stephen Hemminger wrote:

On Fri, 07 Mar 2008 14:09:57 +0000
Kieran Mansley [off-list ref] wrote:

quoted

We've seen a couple of problems when using a bridge or IP forwarding
combined with LRO packets generated by a network device driver.  As you
know, LRO packets can be either be page based (and passed up with
lro_receive_page()) or use the skb frag_list (and passed up with
lro_receive_skb()).  In both cases it is likely that the device driver
will have set CHECKSUM_UNNECESSARY to indicate that the packet has been
checksummed by the device, and gso_size to mark it as an LRO packet and
indicate the actual received MSS.

First off, no hardware should ever do LRO on non-local packets. If the
hardware isn't smart enough to do this, I guess the bridge code to have
an API to turn it off. IP should also turn it off if ip_forwarding
is enabled on that device.

If the only way to deal with this is to prevent LRO in those cases,
having an API to turn if off would clearly be helpful: working out in
the hardware or driver which packets can be forwarded or not is hard,
and would probably be a layering violation in itself.  

However, if there were a way in which it could be made to work, that
would in many ways be preferable.

quoted

If this skb goes directly to the network stack everything is fine.  The
problem comes when this packet instead goes into a bridge and is then
retransmitted on another device.  The skb seems to pass through the
bridge relatively unmodified and because it has gso_size set the
transmit path will attempt to segment it.  If page-based allocation has
been used, this is fine, but if the skb frag_list has been used the
transmit path BUGs in skb_gso_segment():

You can't do LRO with bridging, it is that simple, it is a protocol
layering violation.

Agreed.  LRO is a layering violation with or without bridging.  If the
consensus is that LRO with bridging is not viable, then that's
straightforward for us, but there are a lot of people who will want this
for performance.  Bridging and virtualisation are common partners, and
virtualisation and high performance networking are also rather popular
at the moment.  

Given that there seem to be a small number of places where this is
currently causing problems (LRO through a bridge to a stack is fine,
it's only if the packet is forwarded that things go wrong), and that an
LRO packet should in most cases be very easy to segment (it's already in
network-sized fragments) I wonder if making it work would be
easier/better than preventing it from happening.

Kieran