Re: [DISCUSSION] Unconditionally lock folios when calling rmap_walk()

[Lokesh Gidra <lokeshgidra@google.com>]

On Tue, Sep 2, 2025 at 12:04 PM Lokesh Gidra <lokeshgidra@google.com> wrote:
>
> On Tue, Sep 2, 2025 at 12:01 PM David Hildenbrand <david@redhat.com> wrote:
> >
> > On 02.09.25 20:59, Lokesh Gidra wrote:
> > > On Fri, Aug 29, 2025 at 2:04 AM Lorenzo Stoakes
> > > <lorenzo.stoakes@oracle.com> wrote:
> > >>
> > >> On Fri, Aug 29, 2025 at 10:42:45AM +0200, David Hildenbrand wrote:
> > >>>>>
> > >>>>> I do wonder if we can identify this case and handle things differently.
> > >>>>>
> > >>>>> Perhaps even saying 'try and get the rmap lock, but if there's "too much"
> > >>>>> contention, grab the folio lock.
> > >>>>
> > >>>> Can you please elaborate what you mean? Where do you mean we can
> > >>>> possibly do something like this?
> > >>>>
> > >>>> UFFD move only works on PageAnonExclusive folios. So, would it help
> > >>>> (in terms of avoiding contention) if we were to change the condition:
> > >>>
> > >>> I think we shouldn't be using PAE here. Once could consider using
> > >>> folio_maybe_mapped_shared(), and assume contention on the folio lock if it
> > >>> is maybe mapped shared.
> > >>
> > >> Interesting!
> > >>
> > >>>
> > >>> But the real question is with whom we would be contending for the folio
> > >>> lock.
> > >>>
> > >>> Is it really other processes mapping that folio? I'm not so sure.
> > >>
> > >> Yeah, I might go off and do some research myself on this, actually. Nail down
> > >> wehre this might actually happen.
> > >>
> > >> Generally I'm softening on this and maybe we're good with the proposed change.
> > >>
> > >> But still want to be super careful here... :)
> > >>
> > > Anxiously waiting for your assessment. Fingers crossed :)
> >
> > I'd suggest you prepare an RFC patch where you neatly summarize all we
> > learned so far. :)
> >
Sent the patches:
https://lore.kernel.org/all/20250908044950.311548-1-lokeshgidra@google.com/

> Sounds good. Will do. Thanks.
> > --
> > Cheers
> >
> > David / dhildenb
> >