Re: [Qemu-devel] [PATCH RFC v6 18/20] virtio: support revision-specific features

[Cornelia Huck <cornelia.huck@de.ibm.com>]

On Sun, 1 Feb 2015 23:29:20 +0200
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Fri, Jan 30, 2015 at 03:08:08PM +0100, Cornelia Huck wrote:
> > On Wed, 7 Jan 2015 21:10:07 +0200
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > On Wed, Jan 07, 2015 at 05:22:32PM +0100, Cornelia Huck wrote:
> > > > On Sun, 28 Dec 2014 10:32:06 +0200
> > > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > 
> > > > > On Thu, Dec 11, 2014 at 02:25:20PM +0100, Cornelia Huck wrote:
> > > > > > Devices may support different sets of feature bits depending on which
> > > > > > revision they're operating at. Let's give the transport a way to
> > > > > > re-query the device about its features when the revision has been
> > > > > > changed.
> > > > > > 
> > > > > > Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
> > > > > 
> > > > > So now we have both get_features and get_features_rev, and
> > > > > it's never clear which revision does host_features refer to.
> > > > > IMHO that's just too messy.
> > > > > Let's add get_legacy_features and host_legacy_features instead?
> > > > 
> > > > I wanted to avoid touching anything that does not support version 1.
> > > > And this interface might still work for later revisions, no?
> > > 
> > > We can add _modern_ then, or rename host_features to host_legacy_features
> > > everywhere as preparation.
> > > 
> > 
> > OK, I've ditched the "don't modify old stuff" goal and introduced
> > ->get_features_legacy(). For now, devices will add VERSION_1 in their
> > ->get_features() callback when they support it. (For many devices, this
> > will be the only difference between the two callbacks.)
> > 
> > Two sets of host_features don't make much sense to me.
> 
> It's the most natural implementation given that
> some features are only set for legacy and some (will be) only
> for modern interfaces.

We may be talking about different things: There's the host_features
field which is referenced by the different device types, and there are
the sets of features for legacy or modern devices. These are currently
set dynamically in the ->get_features() callbacks, although we could
possibly convert them to pre-set values like in the kernel. Then I
agree, two sets would be natural. But I'd rather stick with the current
method, if only to avoid churn.