[libvirt] [RFC] Question regarding the validity of QEMU capabilities

[Jiri Denemark]

On Fri, Oct 20, 2017 at 13:12:59 +0100, Daniel P. Berrange wrote:
> On Fri, Oct 20, 2017 at 01:10:14PM +0200, Pavel Hrdina wrote:
> > On Fri, Oct 20, 2017 at 12:59:51PM +0200, Marc Hartmayer wrote:
> > > Hi all,
> > > 
> > > the actual capabilities of QEMU are depending on the host. This
> > > includes dependencies like which kernel modules are loaded or which
> > > kernel parameters are used (e.g. kvm.nested). Therefore, after a
> > > restart we cannot be sure that the QEMU capabilities remain the same.
> > > 
> > > How can we solve this problem?
> > 
> > Hi,
> > 
> > thanks for bringing up this issue, we kind of already know about it
> > but it's good idea to discuss it publicly.
> > 
> > > 
> > > I have come up with two ways:
> > >  - reprobe the capabilities with every host reboot
> > 
> > This is the solution that was agreed on but nobody was motivated enough
> > to write the code :).
> 
> Reprobing QEMU is quite heavy.  I wonder if we should change slightly.
> 
> Make our capabilities cache *only* contain stuff reported by the QEMU
> binary. For other stuff we detect from the host, never cache it.

This is what we've always been doing. But some stuff we get from QEMU
actually depends on the host. In other words, it's QEMU probing the host
and reporting the results to us.

Jirka