Re: [PATCH v2 1/6] KVM: SVM: Use maxphyaddr in emulator RAX check for VMRUN/VMLOAD/VMSAVE

[Sean Christopherson <seanjc@google.com>]

On Wed, Mar 11, 2026, Yosry Ahmed wrote:
> > > Hold up, we're getting ahead of ourselves.
> > >
> > > The only legitimate reason the emulator is at all aware of VMSAVE, VMLOAD, and
> > > VMRUN is to deal with #GP due to the RAX check, because hardware checks the GPA
> > > against the host's physical address space.  See commit 82a11e9c6fa2 ("KVM: SVM:
> > > Add emulation support for #GP triggered by SVM instructions").
> > >
> > > The emulator "support" was originally added by commit 01de8b09e606 ("KVM: SVM:
> > > Add intercept checks for SVM instructions"), but AFAICT, for all intents and
> > > purposes that was dead code when it was added, because the emulator doesn't
> > > actually _emulate_ the instructions.  I assume if they aren't intercepted, and
> > > KVM is full on emulating instead of just decoding, they end up at EMULATION_FAILED
> > > and get a #UD or something.
> > >
> > > Outside of forced emulation or code stream rewriting, KVM should _never_ fully
> > > emulate any of the SVM instructions except VMMCALL (and that is a super special
> > > case).  KVM does need to _decode_ the instruction, and it needs to get the
> > > pre-intercept exception checks correct so that KVM correctly injects e.g. #GP
> > > instead of synthesizing a #VMEXIT for the CPL check, but KVM doesn't need to do
> > > *all* of the checks.
> > >
> > > Note, for L2, the SVME check is meaningless, as EFER.SVME has to be set for L2
> > > to be active, i.e. it's L1's responsibility to handle that check.
> > >
> > > Back to the physical address thing, KVM _already_ handles that check in the #GP
> > > path,
> >
> > I guess if KVM is not intercepting #GP, then the hardware injects the
> > #GP and the emulator still doesn't have to worry about it -- because
> > we don't support the case where RAX can be legal from the host's
> > perspective but not the guest's. Makes sense.
> >
> > > it's just wrong too:
> > >
> > >                 /* All SVM instructions expect page aligned RAX */
> > >                 if (svm->vmcb->save.rax & ~PAGE_MASK)
> > >                         goto reinject;
> > >
> > > So I think what we want is to
> > >
> > >   (a) fix the RAX check in gp_interception()
> > >   (b) drop the RAX check in the emulator
> > >   (c) add a CPL check in the emulator (because the intercepted #GP could have
> > >       been due to L2 executing at CPL>0, not due to a bad-but-good RAX).
> 
> Actually, I don't think (c) is needed. In the path where KVM
> intercepts #GP, it doesn't go through the emulation path which ends up
> calling check_svme(), it only uses the emulator to decode the
> instruction.

Oh, dagnabbit, that's stupidly obvious in hindsight.

> AFAICT, we can end up in the emulator only when the CPU does not
> produce a #GP, e.g. when we get a #NPF on the address in RAX. In this
> case, the CPU will have already checked the CPL for us, and the
> validity of the address. The emulator checking EFER.SVME check is
> probably also useless, because with Kevin's patches we should always
> be intercepting VMLOAD/VMSAVE when EFER.SVME is disabled by the guest
> and checking EFER.SVME there anyway.
> 
> Anyway, I want to touch the emulator as little as possible tbh, so I
> will still do (b) because it unblocks this series (removes the wrong
> GPA check that injects #GP), but will defer any further cleanups.

Yeah, works for me.