Re: [Qemu-devel] [PATCH v2 1/2] i386/kvm: add support for Hyper-V reenlightenment MSRs

[Eduardo Habkost]

On Mon, Mar 19, 2018 at 06:29:08PM +0100, Vitaly Kuznetsov wrote:
> Roman Kagan <address@hidden> writes:
> 
> > On Fri, Mar 16, 2018 at 06:00:19PM +0100, Vitaly Kuznetsov wrote:
> >> KVM recently gained support for Hyper-V Reenlightenment MSRs which are
> >> required to make KVM-on-Hyper-V enable TSC page clocksource to its guests
> >> when INVTSC is not passed to it (and it is not passed by default in Qemu
> >> as it effectively blocks migration).
> >> 
> >> Signed-off-by: Vitaly Kuznetsov <address@hidden>
> >> ---
> >> Changes since v1:
> >> - add vmstate_msr_hyperv_reenlightenment subsection to vmstate_x86_cpu
> >>   [Paolo Bonzini]
> >> ---
[...]
> >> +
> >> +        if (has_msr_hv_reenlightenment) {
> >> +            env->features[FEAT_HYPERV_EAX] |=
> >> +                HV_ACCESS_REENLIGHTENMENTS_CONTROL;
> >> +        }
> >
> > Can you please add a matching comment to the definition of
> > feature_word_info[FEAT_HYPERV_EAX].feat_names[]?
> >
> 
> Sure, missed that.
> 
> > Also there appears to be no cpu property to turn this on/off, does it?
> > It's enabled based only on the support in the KVM it's running against.
> > So I guess we may have a problem migrating between the hosts with
> > different KVM versions, one supporting it and the other not.
> 
> Currently nested workloads don't migrate so I decided to take the
> opportunity and squeeze the new feature in without adding a new
> hv_reenlightenment cpu property (which would have to be added to libvirt
> at least).
> 
> > (This is also a problem with has_msr_hv_frequencies, and is in general a
> > long-standing issue of hv_* properties being done differently from the
> > rest of CPUID features.)
> 
> Suggestions? (To be honest I don't really like us adding new hv_*
> property for every new Hyper-V feature we support. I doubt anyone needs
> 'partial' Hyper-V emulation. It would be nice to have a single versioned
> 'hv' feature implying everything. We may then forbid migrations to older
> hv versions. But I don't really know the history of why we decided to go
> with a separate hv_* for every feature we add).

You will need "partial" emulation if you want to support
live-migration to/from a host where the KVM or QEMU don't support
all the features from the current host.

Is this something the current Hyper-V code already supports, or
it's something known to be broken?

-- 
Eduardo