On Tue, Mar 13, 2018 at 05:28:07PM -0700, Samudrala, Sridhar wrote:
> > I am not sure if it's a good idea to leave the
> > virtio_bypass around if running into failure: the guest is not
> > migratable as the VF doesn't have a backup path,
> Are you talking about a failure when registering backup netdev? This should
> happen, but i guess we can improve error handling in such scenario.
A nice way to do this would be to clear the backup feature bit.
> > And perhaps the worse
> > part is that, it now has two interfaces with identical MAC address but
> > one of them is invalid (user cannot use the virtio interface as it has
> > a dampened datapath). IMHO the virtio_bypass and its lower netdev
> > should be destroyed at all when it fails to bind the VF, and
> > technically, there should be some way to propogate the failure status
> > to the hypervisor/backend, indicating that the VM is not migratable
> > because of guest software errors (maybe by clearing out the backup
> > feature from the guest virtio driver so host can see/learn it).
> In BACKUP mode, user can only use the upper virtio_bypass netdev and that will
> always be there. Any failure to enslave VF netdev is not fatal, but i will see
> if we can improve the error handling of failure to enslave backup netdev.
> Also, i don't think the BACKUP feature bit is negotiable with the host.
All bits are negotiable. It's up to the host whether to support
a device with this bit clear, or to fail negotiation.