Re: [PATCH v3 1/7] KVM: SVM: Drop RAX check for SVM instructions from the emulator

[Paolo Bonzini]

On Tue, Mar 17, 2026 at 3:58 PM Jim Mattson <jmattson@xxxxxxxxxx> wrote:
> > Yeah, that part was clearly broken when physical address bits broke
> > AMD's 48-bit limit (and I don't think the details of what is
> > considered an invalid physical address are part of the architectural
> > description, for example some SMM ranges are carved out by the
> > VMRUN/VMLOAD/VMSAVE microcode).
>
> IIUC, VLS is broken because it applies the host's microarchitectural
> physical address restrictions to untranslated *guest* physical
> addresses. Is that right?

Yes, look at gp_interception() in svm.c. I'm not sure if AMD ever said
publicly whether it was an erratum, or intentional even with VLS
enabled (https://lkml.org/lkml/2021/1/13/513).

(Also I hope that, with VLS enabled, the check is performed *again*
after translating the address).

Paolo