Re: [PATCH v2 0/6] KVM: x86: Apply NX mitigation more precisely

[Mingwei Zhang]

On Thu, Jul 28, 2022 at 2:20 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Thu, Jul 28, 2022, Paolo Bonzini wrote:
> > On 7/23/22 03:23, Sean Christopherson wrote:
> > > Patch 6 from Mingwei is the end goal of the series.  KVM incorrectly
> > > assumes that the NX huge page mitigation is the only scenario where KVM
> > > will create a non-leaf page instead of a huge page.   Precisely track
> > > (via kvm_mmu_page) if a non-huge page is being forced and use that info
> > > to avoid unnecessarily forcing smaller page sizes in
> > > disallowed_hugepage_adjust().
> > >
> > > v2: Rebase, tweak a changelog accordingly.
> > >
> > > v1:https://lore.kernel.org/all/20220409003847.819686-1-seanjc@xxxxxxxxxx
> > >
> > > Mingwei Zhang (1):
> > >    KVM: x86/mmu: explicitly check nx_hugepage in
> > >      disallowed_hugepage_adjust()
> > >
> > > Sean Christopherson (5):
> > >    KVM: x86/mmu: Tag disallowed NX huge pages even if they're not tracked
> > >    KVM: x86/mmu: Properly account NX huge page workaround for nonpaging
> > >      MMUs
> > >    KVM: x86/mmu: Set disallowed_nx_huge_page in TDP MMU before setting
> > >      SPTE
> > >    KVM: x86/mmu: Track the number of TDP MMU pages, but not the actual
> > >      pages
> > >    KVM: x86/mmu: Add helper to convert SPTE value to its shadow page
> >
> > Some of the benefits are cool, such as not having to track the pages for the
> > TDP MMU, and patch 2 is a borderline bugfix, but there's quite a lot of new
> > non-obvious complexity here.
>
> 100% agree on the complexity.
>
> > So the obligatory question is: is it worth a hundred lines of new code?
>
> Assuming I understanding the bug Mingwei's patch fixes, yes.  Though after
> re-reading that changelog, it should more explicitly call out the scenario we
> actually care about.
>
> Anyways, the bug we really care about is that by not precisely checking if a
> huge page is disallowed, KVM would refuse to create huge page after disabling
> dirty logging, which is a very noticeable performance issue for large VMs if
> a migration is canceled.  That particular bug has since been unintentionally
> fixed in the TDP MMU by zapping the non-leaf SPTE, but there are other paths
> that could similarly be affected, e.g. I believe zapping leaf SPTEs in response
> to a host page migration (mmu_notifier invalidation) to create a huge page would
> yield a similar result; KVM would see the shadow-present non-leaf SPTE and assume
> a huge page is disallowed.

Just a quick update: the kernel crash has been resolved. It turns out
to be a bug introduced when I rebase the patch.

I see the patch set is working now.