On Fri, Sep 27, 2019 at 10:22:51AM -0700, Jim Mattson wrote:
> On Fri, Sep 27, 2019 at 9:32 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> >
> > On 27/09/19 18:10, Jim Mattson wrote:
> > > On Fri, Sep 27, 2019 at 9:06 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> > >>
> > >> On 27/09/19 17:58, Xiaoyao Li wrote:
> > >>> Indeed, "KVM_GET_MSR_INDEX_LIST" returns the guest msrs that KVM supports and
> > >>> they are free from different guest configuration since they're initialized when
> > >>> kvm module is loaded.
> > >>>
> > >>> Even though some MSRs are not exposed to guest by clear their related cpuid
> > >>> bits, they are still saved/restored by QEMU in the same fashion.
> > >>>
> > >>> I wonder should we change "KVM_GET_MSR_INDEX_LIST" per VM?
> > >>
> > >> We can add a per-VM version too, yes.
> >
> > There is one problem with that: KVM_SET_CPUID2 is a vCPU ioctl, not a VM
> > ioctl.
> >
> > > Should the system-wide version continue to list *some* supported MSRs
> > > and *some* unsupported MSRs, with no rhyme or reason? Or should we
> > > codify what that list contains?
> >
> > The optimal thing would be for it to list only MSRs that are
> > unconditionally supported by all VMs and are part of the runtime state.
> > MSRs that are not part of the runtime state, such as the VMX
> > capabilities, should be returned by KVM_GET_MSR_FEATURE_INDEX_LIST.
> >
> > This also means that my own commit 95c5c7c77c06 ("KVM: nVMX: list VMX
> > MSRs in KVM_GET_MSR_INDEX_LIST", 2019-07-02) was incorrect.
> > Unfortunately, that commit was done because userspace (QEMU) has a
> > genuine need to detect whether KVM is new enough to support the
> > IA32_VMX_VMFUNC MSR.
> >
> > Perhaps we can make all MSRs supported unconditionally if
> > host_initiated. For unsupported performance counters it's easy to make
> > them return 0, and allow setting them to 0, if host_initiated (BTW, how
> > did you pick 32? is there any risk of conflicts with other MSRs?).
>
> 32 comes from INTEL_PMC_MAX_GENERIC. There are definitely conflicts.
> (Sorry; this should have occurred to me earlier.) 32 event selectors
> would occupy indices [0x186, 0x1a6). But on the architectural MSR
> list, only indices up through 0x197 are "reserved" (presumably for
> future event selectors). 32 GP counters would occupy indices [0xc1,
> 0xe1). But on the architectural MSR list, only indices up through 0xc8
> are defined for GP counters. None are marked "reserved" for future
> expansion, but none in the range (0xc8, 0xe1) are defined either.
>
> Perhaps INTEL_MAX_PMC_GENERIC should be reduced to 18. If we removed
> event selectors and counters above 18, would my original approach
> work?
Heh, VMX is technically available on P4 processors, which don't support
the architectural PMU. Generating the list based on hardware CPUID seems
both safer and easier.
