On Fri, Apr 15, 2016 at 05:02:45PM -0700, Daniele Di Proietto wrote:
> The system userspace testsuite uses the userspace datapath with
> netdev-linux devices, connected to veth pairs with the AF_PACKET socket:
>
> (veth pair) (AF_PACKET)
> TCP stack -> p0 ---> ovs-p0 -------------> netdev-linux (userspace OVS)
>
> Unfortunately this configuration has some problems with offloads: a
> packet generated by the TCP stack maybe sent to p0 without being
> checksummed or segmented. The AF_PACKET socket, by default, ignores the
> offloads and just transmits the data of the packets to userspace, but:
>
> 1. The packet may need GSO, so the data will be too big to be received
> by the userspace datapath
> 2. The packet might have incomplete checksums, so it will likely be
> discarded by the receiver.
>
> Problem 1 causes TCP connections to see a congestion window smaller than
> the MTU, which hurts performance but doesn't prevent communication.
>
> Problem 2 was hidden in the testsuite by a Linux kernel bug, fixed by
> commit ce8c839b74e3("veth: don’t modify ip_summed; doing so treats
> packets with bad checksums as good"). In the kernels that include the
> fix, the userspace datapath is able to process pings, but not tcp or udp
> data.
>
> Unfortunately I couldn't find a way to ask the AF_PACKET to perform
> offloads in kernel. A possible fix would be to use the PACKET_VNET_HDR
> sockopt and perform the offloads in userspace.
>
> Until a proper fix is worked out for netdev-linux, this commit disables
> offloads on the non-OVS side of the veth pair, as a workaround.
>
> Signed-off-by: Daniele Di Proietto <[email protected]>
> ---
I understand why you need this but I don't think it should be part of
this patchset.
We should start noting somewhere things that we need to revisit in
the future.
Acked-by: Flavio Leitner <[email protected]>
_______________________________________________
dev mailing list
[email protected]
http://openvswitch.org/mailman/listinfo/dev
