On Mon, Mar 11, 2024, Xu Yilun wrote:
> On Fri, Mar 08, 2024 at 03:28:08PM -0800, Sean Christopherson wrote:
> > On Fri, Mar 08, 2024, Xu Yilun wrote:
> > > On Tue, Feb 27, 2024 at 06:41:40PM -0800, Sean Christopherson wrote:
> > > > @@ -4410,6 +4405,16 @@ static int kvm_faultin_pfn(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault,
> > > > fault->mmu_seq = vcpu->kvm->mmu_invalidate_seq;
> > > > smp_rmb();
> > > >
> > > > + /*
> > > > + * Check for a private vs. shared mismatch *after* taking a snapshot of
> > > > + * mmu_invalidate_seq, as changes to gfn attributes are guarded by the
> > > > + * invalidation notifier.
> > >
> > > I didn't see how mmu_invalidate_seq influences gfn attribute judgement.
> > > And there is no synchronization between the below check and
> > > kvm_vm_set_mem_attributes(), the gfn attribute could still be changing
> > > after the snapshot.
> >
> > There is synchronization. If kvm_vm_set_mem_attributes() changes the attributes,
> > and thus bumps mmu_invalidate_seq, after kvm_faultin_pfn() takes its snapshot,
> > then is_page_fault_stale() will detect that an invalidation related to the gfn
> > occured and resume the guest *without* installing a mapping in KVM's page tables.
> >
> > I.e. KVM may read the old, stale gfn attributes, but it will never actually
> > expose the stale attirubtes to the guest.
>
> That makes sense! I was just thinking of the racing for below few lines,
>
> if (fault->is_private != kvm_mem_is_private(vcpu->kvm, fault->gfn)) {
> kvm_mmu_prepare_memory_fault_exit(vcpu, fault);
> return -EFAULT;
> }
>
> But the guarding is actually for the whole kvm_faultin_pfn(). It is the
> the same mechanism between getting old gfn attributes and getting old pfn.
>
> I wonder if we could instead add some general comments at
>
> fault->mmu_seq = vcpu->kvm->mmu_invalidate_seq;
>
> about the snapshot and is_page_fault_stale() thing.
Yeah, I'll add a comment. The only reason not to add a comment is that, ideally,
the comment/documentation would live in common KVM code, not x86. But this code
already has a few big comments about the mmu_notifier retry logic, one more
definitely won't hurt.
