Re: [Qemu-devel] [PATCH v2 3/8] virtio: Add support for guest setting of queue size

[Michael S. Tsirkin]

On Thu, Jul 25, 2013 at 01:30:15PM +0100, Peter Maydell wrote:
> On 25 July 2013 10:03, Michael S. Tsirkin <address@hidden> wrote:
> > On Thu, Jul 25, 2013 at 09:50:21AM +0100, Peter Maydell wrote:
> >> On 25 July 2013 06:38, Michael S. Tsirkin <address@hidden> wrote:
> >> > Probably needs to go back to default value on reset?
> >>
> >> Tricky, since the default value is "whatever was passed to
> >> virtio_add_queue()" and we don't save that anywhere.
> >>
> >> For virtio-mmio it is a guest bug to fail to write to the
> >> QueueNum register, so the current behaviour is not out of
> >> specification (and not harmful either AFAICT).
> >
> > Best not to leak info across reboots.
> > Also if guest sets num = 0 it will cause all kind of
> > harm, no?
> >
> >> I guess we could add a vring.defaultnum, which would be
> >> set by virtio_add_queue/virtio_del_queue, and have reset
> >> copy defaultnum into num. No migration needed for defaultnum
> >> because it's always the same for a particular qemu config.
> 
> So I had a look at implementing this, and I noticed that
> we already have some odd behaviour on reset. Specifically,
> virtio backends like net can create virtio queues based on
> guest behaviour (ie setting feature bits). But on reset,
> these queues aren't deleted, so a post-reset QEMU will look
> different from a from-scratch restarted QEMU...

That's a bug. Thanks for the report. :)

> This in turn makes 'save defaultnum and have reset copy it
> into num' awkward, because defaultnum now needs to be
> migrated (otherwise it would do the wrong thing on a reset
> after a VM migration).
> 
> -- PMM

Looks like we'll have to fix the bug first :(