Re: [PATCH 2/2] x86: KVM: Emulate instruction when GPA can't be translated by EPT

[Sean Christopherson <seanjc@google.com>]

On Mon, Jan 08, 2024, Tao Su wrote:
> On Wed, Dec 20, 2023 at 05:42:56AM -0800, Jim Mattson wrote:
> > On Mon, Dec 18, 2023 at 6:08 AM Tao Su <tao1.su@linux.intel.com> wrote:
> > >
> > > With 4-level EPT, bits 51:48 of the guest physical address must all
> > > be zero; otherwise, an EPT violation always occurs, which is an unexpected
> > > VM exit in KVM currently.
> > >
> > > Even though KVM advertises the max physical bits to guest, guest may
> > > ignore MAXPHYADDR in CPUID and set a bigger physical bits to KVM.
> > > Rejecting invalid guest physical bits on KVM side is a choice, but it will
> > > break current KVM ABI, e.g., current QEMU ignores the physical bits
> > > advertised by KVM and uses host physical bits as guest physical bits by
> > > default when using '-cpu host', although we would like to send a patch to
> > > QEMU, it will still cause backward compatibility issues.
> > >
> > > For GPA that can't be translated by EPT but within host.MAXPHYADDR,
> > > emulation should be the best choice since KVM will inject #PF for the
> > > invalid GPA in guest's perspective and try to emulate the instructions
> > > which minimizes the impact on guests as much as possible.
> > >
> > > Signed-off-by: Tao Su <tao1.su@linux.intel.com>
> > > Tested-by: Yi Lai <yi1.lai@intel.com>
> > > ---
> > >  arch/x86/kvm/vmx/vmx.c | 7 +++++++
> > >  1 file changed, 7 insertions(+)
> > >
> > > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > > index be20a60047b1..a8aa2cfa2f5d 100644
> > > --- a/arch/x86/kvm/vmx/vmx.c
> > > +++ b/arch/x86/kvm/vmx/vmx.c
> > > @@ -5774,6 +5774,13 @@ static int handle_ept_violation(struct kvm_vcpu *vcpu)
> > >
> > >         vcpu->arch.exit_qualification = exit_qualification;
> > >
> > > +       /*
> > > +        * Emulate the instruction when accessing a GPA which is set any bits
> > > +        * beyond guest-physical bits that EPT can translate.
> > > +        */
> > > +       if (unlikely(gpa & rsvd_bits(kvm_mmu_tdp_maxphyaddr(), 63)))
> > > +               return kvm_emulate_instruction(vcpu, 0);
> > > +
> > 
> > This doesn't really work, since the KVM instruction emulator is
> > woefully incomplete. To make this work, first you have to teach the
> > KVM instruction emulator how to emulate *all* memory-accessing
> > instructions.
> 
> Please forget allow_smaller_maxphyaddr and #PF for a while.
> 
> I agree KVM instruction emulator is incomplete. However, hardware can’t
> execute instructions with GPA>48-bit and exits to KVM, KVM just repeatedly
> builds SPTE, I.e., current KVM is buggy.

Eh, hardware is just as much to blame as KVM.  Garbage in, garbage out.

> In this case, emulation may be a choice,

Yes, it can be a choice, but that choice needs to be made conciously by userspace,
not silently by KVM.