Re: [PATCH v2 4/7] x86/kvm/hyper-v: Introduce KVM_GET_SUPPORTED_HV_CPUID

[Roman Kagan <rkagan@virtuozzo.com>]

On Tue, Dec 11, 2018 at 04:04:01PM +0100, Vitaly Kuznetsov wrote:
> Roman Kagan <rkagan@virtuozzo.com> writes:
> 
> > On Tue, Dec 11, 2018 at 02:28:14PM +0100, Vitaly Kuznetsov wrote:
> >> Roman Kagan <rkagan@virtuozzo.com> writes:
> >> 
> >> > On Mon, Dec 10, 2018 at 06:21:56PM +0100, Vitaly Kuznetsov wrote:
> >> 
> >> >> +
> >> >> +Currently, the following list of CPUID leaves are returned:
> >> >> + HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS
> >> >> + HYPERV_CPUID_INTERFACE
> >> >> + HYPERV_CPUID_VERSION
> >> >> + HYPERV_CPUID_FEATURES
> >> >> + HYPERV_CPUID_ENLIGHTMENT_INFO
> >> >> + HYPERV_CPUID_IMPLEMENT_LIMITS
> >> >> + HYPERV_CPUID_NESTED_FEATURES
> >> >> +
> >> >> +HYPERV_CPUID_NESTED_FEATURES leaf is only exposed when Enlightened VMCS was
> >> >> +enabled on the corresponding vCPU (KVM_CAP_HYPERV_ENLIGHTENED_VMCS).
> >> >
> >> > IOW the output of ioctl(KVM_GET_SUPPORTED_HV_CPUID) depends on
> >> > whether ioctl(KVM_ENABLE_CAP, KVM_CAP_HYPERV_ENLIGHTENED_VMCS) has
> >> > already been called on that vcpu?  I wonder if this fits the intended
> >> > usage?
> >> 
> >> I added HYPERV_CPUID_NESTED_FEATURES in the list (and made the new ioctl
> >> per-cpu and not per-vm) for consistency. *In theory*
> >> KVM_CAP_HYPERV_ENLIGHTENED_VMCS is also enabled per-vcpu so some
> >> hypothetical userspace can later check enabled eVMCS versions (which can
> >> differ across vCPUs!) with KVM_GET_SUPPORTED_HV_CPUID. We will also have
> >> direct tlb flush and other nested features there so to avoid addning new
> >> KVM_CAP_* for them we need the CPUID.
> >
> > This is different from how KVM_GET_SUPPORTED_CPUID is used: QEMU assumes
> > that its output doesn't change between calls, and even caches the result
> > calling the ioctl only once.
> >
> 
> Yes, I'm not sure if we have to have full consistency between
> KVM_GET_SUPPORTED_CPUID and KVM_GET_SUPPORTED_HV_CPUID.

Neither do I.  I just noticed the difference and thought it might
matter.

> >> Another thing I'm thinking about is something like 'hv_all' cpu flag for
> >> Qemu which would enable everything by setting guest CPUIDs to what
> >> KVM_GET_SUPPORTED_HV_CPUID returns. In that case it would also be
> >> convenient to have HYPERV_CPUID_NESTED_FEATURES properly filled (or not
> >> filled when eVMCS was not enabled).
> >
> > I think this is orthogonal to the way you obtain capability info from
> > the kernel.
> 
> Not necessarily. If very dumb userspace does 'host passthrough' for
> Hyper-V features without doing anything (e.g. not enabling Enlightened
> VMCS) it will just put the result of KVM_GET_SUPPORTED_HV_CPUID in guest
> facing CPUIDs and it will all work. In case eVMCS was previously enabled
> it again just copies everything and this still works.
> 
> We don't probably need this for Qemu though. If you think it would be
> better to have HYPERV_CPUID_NESTED_FEATURES returned regardless of eVMCS
> enablement I'm ready to budge)

I have no opinion on this.  I hope Paolo, who requested the feature, can
explain the desired semantics :)

Roman.