On Wed, Jul 27, 2022 at 07:04:35PM +0000, Sean Christopherson wrote:
> On Wed, Jul 27, 2022, Yan Zhao wrote:
> > On Sat, Jul 23, 2022 at 01:23:23AM +0000, Sean Christopherson wrote:
> >
> > <snip>
> >
> > > @@ -386,16 +385,18 @@ static void handle_changed_spte_dirty_log(struct kvm *kvm, int as_id, gfn_t gfn,
> > > static void tdp_mmu_unlink_sp(struct kvm *kvm, struct kvm_mmu_page *sp,
> > > bool shared)
> > > {
> > > + atomic64_dec(&kvm->arch.tdp_mmu_pages);
> > > +
> > > + if (!sp->nx_huge_page_disallowed)
> > > + return;
> > > +
> > Does this read of sp->nx_huge_page_disallowed also need to be protected by
> > tdp_mmu_pages_lock in shared path?
>
>
> No, because only one CPU can call tdp_mmu_unlink_sp() for a shadow page. E.g. in
> a shared walk, the SPTE is zapped atomically and only the CPU that "wins" gets to
> unlink the s[. The extra lock is needed to prevent list corruption, but the
> sp itself is thread safe.
>
> FWIW, even if that guarantee didn't hold, checking the flag outside of tdp_mmu_pages_lock
> is safe because false positives are ok. untrack_possible_nx_huge_page() checks that
> the shadow page is actually on the list, i.e. it's a nop if a different task unlinks
> the page first.
>
> False negatives need to be avoided, but nx_huge_page_disallowed is cleared only
> when untrack_possible_nx_huge_page() is guaranteed to be called, i.e. true false
> negatives can't occur.
>
> Hmm, but I think there's a missing smp_rmb(), which is needed to ensure
> nx_huge_page_disallowed is read after observing the shadow-present SPTE (that's
> being unlinked). I'll add that in the next version.
It makes sense. Thanks for such detailed explanation!
Thanks
Yan
