Re: [RFC PATCH v2 0/3] virtio-net: graceful drop of vhost for TAP

[Jason Wang]

在 2021/3/26 下午5:18, Jason Wang 写道:
> > > ?
> > I assume that  both src and dest started with vhost=on.
> >
> > As driver B supports both packed and split, you can switch from driver
> > A to driver B after migration
> > and driver B will work with split. Exactly as it does today.
> >
> > The key question is what is more important - vhost or features that
> > vhost does not support?
> > current code says: vhost is more important always
> > v1 patch says: features are more important always.
> > v2 patch says: vhost is more important at init time, features are more
> > important at migration time.
> > Because we are able to drop vhost but we can't drop features when we
> > have a running driver.
> > Do you agree?
>
>
> I think what came from cli is the most important. So if I understand 
> correclty:
>
> - vhost=on means "turn on vhost when possible" it implies that 
> fallback is allowed (we had already had fallback codes)
> - vhostforce=on means "turn on vhost unconditonally" it implies that 
> we can't do fallback
>
> So my understanding is that:
>
> - "vhost=on, packed=on", we can fallback to userspace but must keep 
> packed virtqueue works
> - "vhost=on,vhostforce=on,packed=on", we can't fallback and must keep 
> both vhost and packed virtqueue work, if we can't we need to fail
>
> Thanks


Daniel and Michael, am I right here?

We need some inputs to move forward to fix the migration compatibility 
issue.

Thanks