Re: [PATCH v4 3/9] KVM: SVM: Properly check RAX on #GP intercept of SVM instructions

[Sean Christopherson]

On Fri, Apr 03, 2026, Yosry Ahmed wrote:
> On Fri, Apr 3, 2026 at 12:00 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > On Fri, Apr 03, 2026, Sean Christopherson wrote:
> > > >     svm_exit_code = svm_instr_exit_code(vcpu);
> > > >     if (svm_exit_code) {
> > > > -           /* All SVM instructions expect page aligned RAX */
> > > > -           if (svm->vmcb->save.rax & ~PAGE_MASK)
> > > > +           unsigned long rax = kvm_register_read(vcpu, VCPU_REGS_RAX);
> > > > +
> > > > +           if (!page_address_valid(vcpu, rax))
> > >
> > > Eh, let it poke out, i.e.
> > >
> > >               if (!page_address_valid(vcpu, kvm_register_read(vcpu, VCPU_REGS_RAX)))
> >
> > Argh, looking at the rest of this series, and at KVM's existing code, having to
> > use kvm_register_read() is awful.  This really should be able to use kvm_rax_read(),
> > but that won't handle the truncation.
> >
> > There are only a handful of likely-benign goofs due to this mess, but there is a
> > pile of manual truncation and casting going on.  In addition to _raw() variants,
> > and mode-aware defaults, add "e" versions would be helpful, as many of the
> > explicit truncation flows are cases where e.g. EAX, ECX, and EDX are architecturally
> > accessed.
> >
> > I'll put together patches, and think more on how to handle this series (the
> > dependencies aren't terrible, but they certainly are annoying).  I'm tempted
> > to squeeze this into 7.1 to make future life easier...
> 
> Just to make sure I understand this correctly, you'll keep this series
> using kvm_register_read() and send patches on top to make
> kvm_rax_read() a viable alternative and switch it, right?

Yep, exactly!