Re: [Qemu-devel] Re: [PATCH] qemu/virtio: move features to an inline function

[Anthony Liguori]

Michael S. Tsirkin wrote:
> > Let's say we supported virtio-vbus along with virtio-pci.  What does  
> > virtio_blk_get_features() do to mask out sg_indirect?  For all  
> > virtio-blk knows, it could be on top of virtio-vbus.
> >     
>
> So? VIRTIO_RING_F_INDIRECT_DESC applies to all transports.
> Just clear this bit.
>   

You can have many layers with virtio.  device + transport + ring

virtio-vbus would have a different transport and a different ring 
implementation.  So no, VIRTIO_RING_F_INDIRECT_DESC wouldn't apply to 
virtio-vbus.

> > >   This would break things like
> > > migrating between userspace and kernel virtio (something that I
> > > support).
> > >       
> > The PIT uses a common state structure and common code for save/restore.   
> > This makes migration compatible.
> >     
>
> Isn't device name put in the machine config, which presumably is
> send along as well?
>   

Good question.  I don't know the best way to resolve this.

Maybe migration between devices isn't such a good idea.  It's 
conceivable that vhost will require some state that isn't present in the 
userspace virtio-net.  I think this requires some thought.

> > In this case, it's two separate implementations of the same device.  I  
> > think it makes sense for them to be separate devices.
> >
> > Regards,
> >
> > Anthony Liguori
> >     
>
> Hmm, I see what you mean. But kernel virtio is harder. Unlike
> PIT/APIC, it is not a separate codepath.  It still needs
> all of userspace virtio to support live migration and non-MSI guests.
> Really, it's the same device that switches between kernel and userspace
> modes on the fly.
>
> This will become clearer from code when I implement migration for vhost,
> but basically you switch to userspace when you start migration, and
> back to kernel if migration fails. You also switch to kernel when MSI
> is enabled and back to userspace when it is disabled.
>   

Why bother switching to userspace for migration?  Can't you just have 
get/set ioctls for the state?

Regards,

Anthony Liguori