Re: [PATCH v4 2/4] hw/i386: Attach CPUs to machine

["Daniel P. Berrangé" <berrange@redhat.com>]

On Mon, Feb 07, 2022 at 10:36:42AM +0100, Peter Krempa wrote:
> On Mon, Feb 07, 2022 at 10:18:43 +0100, Igor Mammedov wrote:
> > On Mon, 7 Feb 2022 09:14:37 +0100
> > Igor Mammedov <imammedo@redhat.com> wrote:
> > 
> > > On Sat,  5 Feb 2022 13:45:24 +0100
> > > Philippe Mathieu-Daudé <f4bug@amsat.org> wrote:
> > > 
> > > > Previously CPUs were exposed in the QOM tree at a path
> > > > 
> > > >   /machine/unattached/device[nn]
> > > > 
> > > > where the 'nn' of the first CPU is usually zero, but can
> > > > vary depending on what devices were already created.
> > > > 
> > > > With this change the CPUs are now at
> > > > 
> > > >   /machine/cpu[nn]
> > > > 
> > > > where the 'nn' of the first CPU is always zero.  
> > > 
> > > Could you add to commit message the reason behind the change?
> > 
> > regardless, it looks like unwarranted movement to me
> > prompted by livirt accessing/expecting a QOM patch which is
> > not stable ABI. I'd rather get it fixed on libvirt side.
> > 
> > If libvirt needs for some reason access a CPU instance,
> > it should use @query-hotpluggable-cpus to get a list of CPUs
> > (which includes QOM path of already present CPUs) instead of
> > hard-codding some 'well-known' path as there is no any guarantee 
> > that it will stay stable whatsoever.
> 
> I don't disagree with you about the use of hardcoded path, but the way
> of using @query-hotpluggable-cpus is not really aligning well for how
> it's being used.
> 
> To shed a bit more light, libvirt uses the following hardcoded path
> 
> #define QOM_CPU_PATH  "/machine/unattached/device[0]"
> 
> in code which is used to query CPU flags. That code doesn't care at all
> which cpus are present but wants to get any of them. So yes, calling
> query-hotpluggable-cpus is possible but a bit pointless.
> 
> In general the code probing cpu flags via qom-get is very cumbersome as
> it ends up doing ~400 QMP calls at startup of a VM in cases when we deem
> it necessary to probe the cpu fully.

Yes, that's one QMP call per CPUID feature bit that QEMU knows
about. It is a massive performance bottleneck that we need a
much better solution for.  We really should have raised this
with QEMU right away when we found we had this need for 100's
of QMP commands.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|