[Qemu-devel] [PATCH 00/19 v2] Add virtio-net/tap support for partial csums and GSO

[Mark McLoughlin <markmc@redhat.com>]

Hey,
        Over a year ago we added some code to qemu-kvm.git which takes
advantage of the recent tun/tap IFF_VNET_HDR feature in order to allow
virtio-net to send and receive packets with partial checksums and
segmentation offloaded:

  http://article.gmane.org/gmane.comp.emulators.kvm.devel/20440

  "This allows us to pass larger packets and packets with
  partial checkums between the guest and the host, greatly
  increasing the achievable bandwidth."

        Unfortunately, that implementation was quite hacky as it
made some assumptions that would break if e.g. you added another
network client to a vlan where the feature had enabled.

        Now that we have the -netdev parameter, we can more safely
pair the NIC and backend, allowing us to negatiate features like
this.

        What follows is a somewhat cleaned up version of the code
from qemu-kvm.git. Further cleanups are probably possible, but I
think this much is mergeable. Some points of discussion:

  - If you start a guest on a host which supports IFF_VNET_HDR and
    migrate to a host which doesn't support it, we can't just tell
    the guest to stop using GSO.

    We need prevent this by rejecting the migration on the destination
    if IFF_VNET_HDR is required but not available. However, in my
    testing if a load() function returns an error, the migration is
    seen to have completed on the source but failed on the destination.
    Clearly that's wrong, but I haven't investigated further yet.

    Also, we provide a vnet_hdr= arg so that people can prevent the
    feature being used on the source or require that is available on
    the destination.

  - I've slightly abused QemuOpts in 05/19. Comments welcome.

  - The whole mess around adding a 'raw' packet flag for the gratuitous
    ARP packet is gratuitously messy.

  - I'm always enabling IFF_VNET_HDR if it's available. I don't think
    that's a problem, especially compared to what would be required to
    only selectively enable it.

        Changes since v1:

  - add has_vnet_hdr and has_ufo fields for ease of vmstate conversion

  - use qemu_error() in virtio_net_load()


Cheers,
Mark.