Re: [RFC PATCH 0/4] KVM: x86/mmu: Rework marking folios dirty/accessed

[Sean Christopherson <seanjc@google.com>]

On Fri, Apr 05, 2024, David Hildenbrand wrote:
> On 05.04.24 11:37, Paolo Bonzini wrote:
> > On Fri, Apr 5, 2024 at 8:53 AM David Hildenbrand <david@redhat.com> wrote:
> > > >        mmu_notifier_invalidate_range_start(&range);
> > > >        tlb_start_vma(&tlb, vma);
> > > >        walk_page_range(vma->vm_mm, range.start, range.end,
> > > >                        &madvise_free_walk_ops, &tlb);
> > > >        tlb_end_vma(&tlb, vma);
> > > >        mmu_notifier_invalidate_range_end(&range);
> > > > 
> > > 
> > > Indeed, we do setup the MMU notifier invalidation. We do the start/end
> > > ... I was looking for PTE notifications.
> > > 
> > > I spotted the set_pte_at(), not a set_pte_at_notify() like we do in
> > > other code. Maybe that's not required here (digging through
> > > documentation I'm still left clueless). [...]
> > > Absolutely unclear to me when we *must* to use it, or if it is. Likely
> > > its a pure optimization when we're *changing* a PTE.
> > 
> > Yes, .change_pte() is just an optimization. The original point of it
> > was for KSM, so that KVM could flip the sPTE to a new location without
> > first zapping it. At the time there was also an .invalidate_page()
> > callback, and both of them were *not* bracketed by calls to
> > mmu_notifier_invalidate_range_{start,end}()
> > 
> > Later on, both callbacks were changed to occur within an
> > invalidate_range_start/end() block.
> > 
> > Commit 6bdb913f0a70 ("mm: wrap calls to set_pte_at_notify with
> > invalidate_range_start and invalidate_range_end", 2012-10-09) did so
> > for .change_pte(). The reason to do so was a bit roundabout, namely to
> > allow sleepable .invalidate_page() hooks (because .change_pte() is not
> > sleepable and at the time .invalidate_page() was used as a fallback
> > for .change_pte()).
> > 
> > This however made KVM's usage of the .change_pte() callback completely
> > moot, because KVM unmaps the sPTEs during .invalidate_range_start()
> > and therefore .change_pte() has no hope of succeeding.
> > 
> > (Commit 369ea8242c0f ("mm/rmap: update to new mmu_notifier semantic
> > v2", 2017-08-31) is where the other set of non-bracketed calls to MMU
> > notifier callbacks was changed; calls to
> > mmu_notifier_invalidate_page() were replaced by calls to
> > mmu_notifier_invalidate_range(), bracketed by calls to
> > mmu_notifier_invalidate_range_{start,end}()).
> > 
> > Since KVM is the only user of .change_pte(), we can remove
> > .change_pte() and set_pte_at_notify() completely.
> 
> Nice, thanks for all that information!

Ya, from commit c13fda237f08 ("KVM: Assert that notifier count is elevated in
.change_pte()"):

    x86 and MIPS are clearcut nops if the old SPTE is not-present, and that
    is guaranteed due to the prior invalidation.  PPC simply unmaps the SPTE,
    which again should be a nop due to the invalidation.  arm64 is a bit
    murky, but it's also likely a nop because kvm_pgtable_stage2_map() is
    called without a cache pointer, which means it will map an entry if and
    only if an existing PTE was found.

I'm 100% in favor of removing .change_pte().  As I've said multiple times, the
only reason I haven't sent a patch is because I didn't want it to prompt someone
into resurrecting the original behavior. :-)