Re: [PATCH 2/2] KVM: VMX: Ignore userspace MSR filters for x2APIC when APICV is enabled

[Peter Xu]

On Wed, Oct 07, 2020 at 04:01:59PM +0200, Alexander Graf wrote:
> 
> 
> On 05.10.20 21:55, Sean Christopherson wrote:
> > 
> > Rework the resetting of the MSR bitmap for x2APIC MSRs to ignore
> > userspace filtering when APICV is enabled.  Allowing userspace to
> > intercept reads to x2APIC MSRs when APICV is fully enabled for the guest
> > simply can't work.   The LAPIC and thus virtual APIC is in-kernel and
> > cannot be directly accessed by userspace.  If userspace wants to
> > intercept x2APIC MSRs, then it should first disable APICV.
> > 
> > Opportunistically change the behavior to reset the full range of MSRs if
> > and only if APICV is enabled for KVM.  The MSR bitmaps are initialized
> > to intercept all reads and writes by default, and enable_apicv cannot be
> > toggled after KVM is loaded.  I.e. if APICV is disabled, simply toggle
> > the TPR MSR accordingly.
> > 
> > Note, this still allows userspace to intercept reads and writes to TPR,
> > and writes to EOI and SELF_IPI.  It is at least plausible userspace
> > interception could work for those registers, though it is still silly.
> > 
> > Cc: Alexander Graf <graf@xxxxxxxxxx>
> > Cc: Aaron Lewis <aaronlewis@xxxxxxxxxx>
> > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> 
> I'm not opposed in general to leaving APICV handled registers out of the
> filtering logic. However, this really needs a note in the documentation
> then, no?

If we want to forbid apicv msrs, should we even fail KVM_X86_SET_MSR_FILTER
directly then?

I've no strong opinion on whether these msrs should be restricted. I'm not sure
whether my understanding is correct here - to me, kvm should always depend on
the userspace to do the right thing to make the vm work.  To me, as long as the
error is self-contained and it does not affect kvm as a whole or the host, then
it seems still fine.

However I do agree that I also worried about vmx_update_msr_bitmap_x2apic()
being slower.  Majorly I see calls from vmx_refresh_apicv_exec_ctrl() or
nested, so I'm not sure whether that could make sense for some workload.  Btw,
that seems to be another change corresponds to the idea to restrict msr
filitering on apicv regs.

Thanks,

-- 
Peter Xu