Re: [virtio] Re: [virtio-comment] *-over-virtio

[Stefan Hajnoczi <stefanha@redhat.com>]

On Tue, Aug 06, 2013 at 08:55:27AM -0500, Anthony Liguori wrote:
> Stefan Hajnoczi <stefanha@redhat.com> writes:
> 
> > On Mon, Aug 05, 2013 at 12:22:06PM -0400, Sasha Levin wrote:
> >> We seem to have a family of several devices that are rather simple on the virtio level. Some
> >> identifying features include:
> >> 
> >>  * No real usage of the configuration space
> >>  * No device features
> >>  * No control vqs
> >>  * Usually one, but possibly 2 vqs used for straightforward message passing outside the virtio
> >> scope.
> >> 
> >> The use case is simple: use virtio as a simple transport for protocol specific messages, this
> >> can pretty much go over any other transport exactly the same way.
> >> 
> >> Would it make sense to move their actual definition out of the spec and just reserve device
> >> IDs for them there?
> >
> > Why don't they use virtio-serial or the new vsock from the start?  Both
> > transports are built to support arbitrary protocols for tasks that don't
> > need virtio device-level features.  Seems like the right fit instead of
> > reserving unique virtio device IDs.
> 
> IMHO, anything that may be of interest across multiple guests and
> multiple hypervisors should be considered for formal specification.
> 
> I don't think the answer should be to tunnel everything through
> virtio-serial or vsock.  There's nothing wrong with simple virtio
> devices.
> 
> A reasonable analogy is the HID specification for USB.  It supports a
> very large number of often simple devices.  It's pretty useful to have
> standard drivers for such devices because then there is no need for
> vendor specific drivers.

If the devices are simple and don't re-invent virtio-serial or vsock,
then I agree.

But I actually misunderstood Sasha, I thought he was referring to new,
custom devices rather than things like 9p or rng.

Stefan