Re: [PATCH v3 2/8] KVM: x86: nSVM: Cache and validate vmcb12 g_pat

[Yosry Ahmed]

On Fri, Feb 06, 2026 at 11:12:36AM -0800, Sean Christopherson wrote:
> On Fri, Feb 06, 2026, Jim Mattson wrote:
> > On Fri, Feb 6, 2026 at 10:23 AM Yosry Ahmed <yosry.ahmed@xxxxxxxxx> wrote:
> > >
> > > February 6, 2026 at 10:19 AM, "Sean Christopherson" <seanjc@xxxxxxxxxx> wrote:
> > >
> > >
> > > >
> > > > On Thu, Feb 05, 2026, Jim Mattson wrote:
> > > >
> > > > >
> > > > > Cache g_pat from vmcb12 in svm->nested.gpat to avoid TOCTTOU issues, and
> > > > >  add a validity check so that when nested paging is enabled for vmcb12, an
> > > > >  invalid g_pat causes an immediate VMEXIT with exit code VMEXIT_INVALID, as
> > > > >  specified in the APM, volume 2: "Nested Paging and VMRUN/VMEXIT."
> > > > >
> > > > >  Fixes: 3d6368ef580a ("KVM: SVM: Add VMRUN handler")
> > > > >  Signed-off-by: Jim Mattson <jmattson@xxxxxxxxxx>
> > > > >  ---
> > > > >  arch/x86/kvm/svm/nested.c | 4 +++-
> > > > >  arch/x86/kvm/svm/svm.h | 3 +++
> > > > >  2 files changed, 6 insertions(+), 1 deletion(-)
> > > > >
> > > > >  diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> > > > >  index f72dbd10dcad..1d4ff6408b34 100644
> > > > >  --- a/arch/x86/kvm/svm/nested.c
> > > > >  +++ b/arch/x86/kvm/svm/nested.c
> > > > >  @@ -1027,9 +1027,11 @@ int nested_svm_vmrun(struct kvm_vcpu *vcpu)
> > > > >
> > > > >  nested_copy_vmcb_control_to_cache(svm, &vmcb12->control);
> > > > >  nested_copy_vmcb_save_to_cache(svm, &vmcb12->save);
> > > > >  + svm->nested.gpat = vmcb12->save.g_pat;
> > > > >
> > > > >  if (!nested_vmcb_check_save(vcpu) ||
> > > > >  - !nested_vmcb_check_controls(vcpu)) {
> > > > >  + !nested_vmcb_check_controls(vcpu) ||
> > > > >  + (nested_npt_enabled(svm) && !kvm_pat_valid(svm->nested.gpat))) {
> > > > >  vmcb12->control.exit_code = SVM_EXIT_ERR;
> > > > >  vmcb12->control.exit_info_1 = 0;
> > > > >  vmcb12->control.exit_info_2 = 0;
> > > > >  diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
> > > > >  index 986d90f2d4ca..42a4bf83b3aa 100644
> > > > >  --- a/arch/x86/kvm/svm/svm.h
> > > > >  +++ b/arch/x86/kvm/svm/svm.h
> > > > >  @@ -208,6 +208,9 @@ struct svm_nested_state {
> > > > >  */
> > > > >  struct vmcb_save_area_cached save;
> > > > >
> > > > >  + /* Cached guest PAT from vmcb12.save.g_pat */
> > > > >  + u64 gpat;
> > > > >
> > > > Shouldn't this go in vmcb_save_area_cached?
> > >
> > > I believe Jim changed it after this discussion on v2: https://lore.kernel.org/kvm/20260115232154.3021475-4-jmattson@xxxxxxxxxx/.
> 
> LOL, oh the irony:
> 
>   I'm going to cache it on its own to avoid confusion.
> 
> > Right. The two issues with putting it in vmcb_save_area_cached were:
> > 
> > 1. Checking all of vmcb_save_area_cached requires access to the
> > corresponding control area (or at least the boolean, "NTP enabled.")
> 
> Checking the control area seems like the right answer (I went down that path
> before reading this).
> 
> > 2. In the nested state serialization payload, everything else in the
> > vmcb_save_area_cached comes from L1 (host state to be restored at
> > emulated #VMEXIT.)
> 
> Hmm, right, but *because* it's ignored, that gives us carte blanche to clobber it.
> More below.
> 
> > The first issue was a little messy, but not that distasteful.
> 
> I actually find it the opposite of distasteful.  KVM definitely _should_ be
> checking the controls, not the vCPU state.  If it weren't for needing to get at
> MAXPHYADDR in CPUID, I'd push to drop @vcpu entirely.
> 
> > The second issue was really a mess.
> 
> I'd rather have the mess contained and document though.  Caching g_pat outside
> of vmcb_save_area_cached bleeds the mess into all of the relevant nSVM code, and
> doesn't leave any breadcrumbs in the code/comments to explain that it "needs" to
> be kept separate.
> 
> AFAICT, the only "problem" is that g_pat in the serialization payload will be
> garbage when restoring state from an older KVM.  But that's totally fine, precisely
> because L1's PAT isn't restored from vmcb01 on nested #VMEXIT, it's always resident
> in vcpu->arch.pat.  So can't we just do this to avoid a spurious -EINVAL?
> 
> 	/*
> 	 * Validate host state saved from before VMRUN (see
> 	 * nested_svm_check_permissions).
> 	 */
> 	__nested_copy_vmcb_save_to_cache(&save_cached, save);
> 
> 	/*
> 	 * Stuff gPAT in L1's save state, as older KVM may not have saved L1's
> 	 * gPAT.  L1's PAT, i.e. hPAT for the vCPU, is *always* tracked in
> 	 * vcpu->arch.pat, i.e. gPAT is a reflection of vcpu->arch.pat, not the
> 	 * other way around.
> 	 */
> 	save_cached.g_pat = vcpu->arch.pat;
> 
> 	if (!(save->cr0 & X86_CR0_PG) ||
> 	    !(save->cr0 & X86_CR0_PE) ||
> 	    (save->rflags & X86_EFLAGS_VM) ||
> 	    !nested_vmcb_check_save(vcpu, &ctl_cached, &save_cached))
> 		goto out_free;
> 
> Oh, and if we do plumb in @ctrl to __nested_vmcb_check_save(), I vote to
> opportunistically drop the useless single-use wrappers (probably in a standalone
> patch to plumb in @ctrl).  E.g. (completely untested)

They are dropped by
https://lore.kernel.org/kvm/20260206190851.860662-9-yosry.ahmed@xxxxxxxxx/.