[libvirt] [RFC] Host CPU passthrough

[Daniel P. Berrange]

On Mon, Aug 08, 2011 at 01:48:59PM +0200, Jiri Denemark wrote:
> Hi,
> 
> AFAIK this topic is not new but I think we still do not have a good solution
> for it. Libvirt already supports specifying what CPU and its features a guest
> should see but imagine one wants to run a guest on the best possible CPU. The
> current way is to copy the <cpu> element from capabilities XML into domain
> XML. This approach is fine since it provides reproducible environment and such
> domain can even be migrated to a different host. But the CPU shown provided to
> a guest is not the same as the real host CPU. It's based on a model which
> doesn't reflect all aspects of real CPUs. Ideally, we would model everything
> but that's quite complicated and may need updating anytime a new CPU is
> introduced.
> 
> I think we should add a possibility to passthrough host CPU into a domain no
> matter what the CPU looks like. Domains configured in such a way will not be
> allowed to be migrated since we cannot guarantee the same CPU on destination
> host but they would be able to see the real CPU. Another benefit is that a
> single XML configuration would work on any CPU without modifications. Since
> possible issues with such domains will be harder to debug we can mark them as
> tainted just like we do if someone calls qemu monitor commands directly.
> 
> I'm not sure what XML representation makes the most sense so I'll mention
> several options which I'm thinking about:
> 
> - "host" ("native", "passthrough", whatever) special model; same name space as
>   real CPU models, I don't like this too much:
> 
>   <cpu match='exact'>
>     <model>host</model>
>   </cpu>

This is not nice because the end mgmt app doesn't see what the guest is
actually running with. So you get the result that the app has to have
two sets of logic, one looking at the guest XML CPU model, and other
case where the app has to look at the host capabilities CPU model.

> - "host" match policy; has nice first-look semantics that the guest CPU
>   matches host's but doesn't play very nice with the defined semantics of
>   "match" attribute:
> 
>   <cpu match='host'/>

Yeah I don't much like that.

> - new "mode" attribute; IMHO the best one, similar to what we have for SMBIOS:
> 
>   <cpu mode='host'/>
> 
>   vs.
> 
>   <cpu mode='custom' match='...'>
>     <model>...</model.
>     ...
>   </cpu>

This option is almost right with one change.

When the guest is inactive:

   <cpu mode='host'/>

But when the guest is actually launched, it should get filled
in with the real data. ie

   <cpu mode='host' match='...'>
     <model>...</model>
     ...
   </cpu>

and then cleared out again when the guest stops. This is what we do with
the SELinux dynamic security labels. During save/restore/ migration, of
course the contents should remain present.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|


-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|