Re: [Qemu-devel] [PATCH v1] qemu: command: rework cpu feature argument support

[Eduardo Habkost <ehabkost@redhat.com>]

On Thu, Nov 17, 2016 at 04:04:33PM +0100, Markus Armbruster wrote:
> Eduardo Habkost <ehabkost@redhat.com> writes:
> 
> > CCing qemu-devel.
> >
> > CCing Markus, in case he has any insights about the interface
> > introspection.
> >
> > On Tue, Nov 15, 2016 at 08:42:12AM +0100, Jiri Denemark wrote:
> >> On Mon, Nov 14, 2016 at 18:02:29 -0200, Eduardo Habkost wrote:
> >> > On Mon, Nov 14, 2016 at 02:26:03PM -0500, Collin L. Walling wrote:
> >> > > cpu features are passed to the qemu command with feature=on/off
> >> > > instead of +/-feature.
> >> > > 
> >> > > Signed-off-by: Collin L. Walling <walling@linux.vnet.ibm.com>
> >> > 
> >> > If I'm not mistaken, the "feature=on|off" syntax was added on
> >> > QEMU 2.0.0. Does current libvirt support older QEMU versions?
> >> 
> >> Of course it does. I'd love to switch to feature=on|off, but how can we
> >> check if QEMU supports it? We can't really start using this syntax
> >> without it.
> >
> > Actually, I was wrong, this was added in v2.4.0. "feat=on|off"
> > needs two things to work (in x86):
> >
> > * Translation of all "foo=bar" options to QOM property setting.
> >   This was added in v2.0.0-rc0~162^2
> > * The actual QOM properties for feature names to be present. They
> >   were added in v2.4.0-rc0~101^2~1
> >
> > So you can be sure "feat=on" is supported by checking if the
> > feature flags are present in device-list-properties output for
> > the CPU model. But device-list-properties is also messy[1].
> >
> > Maybe we can use the availability of query-cpu-model-expansion to
> > check if we can safely use the new "feat=on|off" system? It's
> > easier than taking all the variables above into account.
> >
> > ---
> >
> > [1] * device-list-properties support for x86 CPU QOM classes
> >       will be in QEMU 2.8.
> >     * device-list-properties on x86 CPU QOM classes returns
> >       an error on QEMU 2.5-2.7.
> >     * device-list-properties on x86 CPU classes may crash QEMU in
> >       QEMU older than 2.5 (see commit 4c315c27).
> >   But:
> >     * query-cpu-definitions will probably return the CPU QOM
> >       typename in QEMU 2.9+ only.
> 
> In other words:
> 
> * feature=on|off works since 2.4, but what's the best way to probe for
>   it?
> 
> * device-list-properties can act as a witness, but there are two
>   problems: you need to map between CPU model and QOM typename to use
>   it, and it can fail or even crash before 2.8.
> 
> * 2.9 will take provide the information to map between CPU model and QOM
>   typename.
> 
> I guess the sane choice is to probe for the CPU model - QOM typename
> mapping information, and if it's there, assume device-list-properties
> and feature=on|off work.  Eduardo, what do you think?

That would work for x86, but not for s390x on QEMU <= 2.8 (where
only feat=on|off is supported). Using query-cpu-model-expansion
as a witness would work for both s390x and i386, even with
QEMU <= 2.8.

-- 
Eduardo