On Thu, 10 Oct 2013 16:23:50 +0300
"Michael S. Tsirkin" <> wrote:
> On Thu, Oct 10, 2013 at 02:40:28PM +0200, Cornelia Huck wrote:
> > On Thu, 10 Oct 2013 14:15:33 +0300
> > "Michael S. Tsirkin" <> wrote:
> >
> > > On Thu, Oct 10, 2013 at 01:02:57PM +0200, Cornelia Huck wrote:
> > > > On Thu, 10 Oct 2013 11:34:30 +0300
> > > > "Michael S. Tsirkin" <> wrote:
> > > >
> > > > > On Tue, Oct 08, 2013 at 05:19:05PM +0200, Cornelia Huck wrote:
> > > > > > Provide a new ccw that allows devices and drivers to operate on selected
> > > > > > revision levels.
> > > > > >
> > > > > > VIRTIO-42
> > > > > >
> > > > > > Signed-off-by: Cornelia Huck <>
> > > > >
> > > > > So basically, CCW_CMD_SET_VIRTIO_REV re-implements the VIRTIO_1 feature
> > > > > bit in a device specific manner.
> > > > > Now that we have FEATURES_OK, this does not seem to be needed?
> > > >
> > > > The idea was to implement something like a 'revision id', which we
> > > > didn't have for ccw. It also allows for negotiating (the driver can try
> > > > revisions until it finds one that the device accepts) and extra
> > > > configuration specifics (the currently unused data field). As it needs
> > > > to be done before any other channel commands, it also allows for early
> > > > fencing.
> > >
> > > So does not the new initialization sequence and the new
> > > FEATURES_OK flag handle all that?
> > > I just did a commit - could you take a look please?
> >
> > I don't think the features stuff would allow for extra configuration
> > data?
>
> Yes but what is this extra data?
This might be special queue-configuration limits or so. Just a
placeholder so that we don't need to come up with a new mechanism later.
>
>
> > >
> > > If not and we need an extra hand-shake - shouldn't it be
> > > generic and not transport specific?
> > >
> >
> > I'm trying to introduce a revision mechanism here, which pci/mmio seem
> > to have natively but ccw hasn't.
> > If we'll never want to use the
> > exisiting revision mechanism for anything virtio specific, we can
> > probably ditch this. But if we do, all transports should have something
> > like that.
>
> So far revisions didn't work well for PCI.
> Basically the issue is that it's a vague kind of thing.
> Imagine that we had the proposed mechanism in place.
> Which revision does a transitional driver advertise?
> I guess it could start with some value and decrement until success,
> but you don't actually seem to say in spec that it should.
That's actually what I had in mind: Try a revision; if it doesn't work,
try another one. If your driver does not want to support legacy
devices, don't try revision 0.
Was my description too vague?
>
