On Fri, Sep 22, 2023, Jim Mattson wrote:
> On Fri, Sep 22, 2023 at 12:40 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> >
> > On Fri, Sep 22, 2023, Jim Mattson wrote:
> > > Okay. What about the IA32_MISC_ENABLE bits above?
> >
> > One of the exceptions where I don't see a better option, and hopefully something
> > that Intel won't repeat in the future. Though I'm not exactly brimming with
> > confidence that Intel won't retroactively add more "gotcha! unsupported!" bits
> > in the future when they realize they forgot add a useful CPUID feature bit.
>
> I don't understand the difference here. Why not make userspace
> responsible for setting these bits as well?
That probably would have been the ideal approach. I'm not entirely sure it would
have actually been feasible though, as I suspect enumerting X86_FEATURE_DS without
any kind of guard would break userspace that reflects KVM_GET_SUPPORTED_CPUID
back into KVM_SET_CPUID(2).
Even better would have been to never merge PEBS support in KVM in its current
form. The whole thing is a house of cards, e.g. if counters are "cross-mapped"
then the guest counters simply stop working. And those warts aside, the entire
enabling was a chaotic mess. See commit 9fc222967a39 ("KVM: x86: Give host
userspace full control of MSR_IA32_MISC_ENABLES").
In other words, setting the UNAVAILABLE bits was the least awful way to salvage
the mess.
