[libvirt] [Qemu-devel] [uq/master PATCH 0/7] x86 CPU subclasses, take 7

Eduardo Habkost ehabkost at redhat.com
Fri Jan 31 19:25:59 UTC 2014

On Fri, Jan 31, 2014 at 08:18:39PM +0100, Igor Mammedov wrote:
> On Fri, 31 Jan 2014 11:56:18 -0700
> Eric Blake <eblake at redhat.com> wrote:
> > On 01/31/2014 11:51 AM, Eduardo Habkost wrote:
> > 
> > >> Allowing -device may be okay, since in the (very?) long term -device
> > >> can be replaced by -object.  But -object is definitive.
> > > 
> > > OK, one additional reason to try device_add first.
> > > 
> > > But then we have one additional problem:
> > > 
> > >  * We want to allow libvirt to probe for CPU model information when
> > >    running QEMU using "-machine none" (because libvirt already does
> > >    that, and we don't want to require libvirt to run QEMU multiple
> > >    times)
> > >  * "device_add driver=<model>-x86_64-cpu" requires an icc-bus to be present
> > >  * -machine none doesn't have any bus
> > >  * I don't see a way to create an icc-bus through QMP (is there a way?)
> > 
> > Is the icc-bus something that makes sense for all architectures, so that
> > libvirt could just blindly request a command line that uses '-machine
> > none' but also instantiates the icc-bus?  Even if icc-bus is
> > x86-specific, libvirt DOES have some notion of what architecture a qemu
> > executable will be targetting, and could modify the command line based
> > on what architecture it guesses the binary will support, if only for the
> > purpose of minimizing qemu invocations for its probe of supported cpus.
> Since -machine none, will not produce accurate CPUID info anyway and libvirt
> knows in advance that it's going to create x86 machine, it might probe with
> default machine type. It would be more accurate than -machine none,
> but still might be not accurate if another machine type will be eventually
> used for running guest.

AFAIK, libvirt logic about CPU model features doesn't even take into
account that it may change depending machine-type. I believe getting the
non-machine-type-specific data will be better than the current state
where libvirt relies on the data from cpu_map.xml, which duplicates (and
sometimes don't match) what's inside QEMU.

We will also need a way to figure out machine-type-specific information
about each CPU model without re-running QEMU, but I think we can do this
one step at a time.


More information about the libvir-list mailing list