Re: [PATCH v22 1/4] mm: add MAP_DROPPABLE for designating always lazily freeable mappings

[Yu Zhao <yuzhao@google.com>]

On Thu, Jul 11, 2024 at 1:49 PM Yu Zhao <yuzhao@google.com> wrote:
>
> On Thu, Jul 11, 2024 at 1:20 PM David Hildenbrand <david@redhat.com> wrote:
> >
> > On 11.07.24 21:18, David Hildenbrand wrote:
> > > On 11.07.24 20:56, David Hildenbrand wrote:
> > >> On 11.07.24 20:54, Jason A. Donenfeld wrote:
> > >>> On Thu, Jul 11, 2024 at 08:24:07PM +0200, David Hildenbrand wrote:
> > >>>>> And PG_large_rmappable seems to only be used for hugetlb branches.
> > >>>>
> > >>>> It should be set for THP/large folios.
> > >>>
> > >>> And it's tested too, apparently.
> > >>>
> > >>> Okay, well, how disappointing is this below? Because I'm running out of
> > >>> tricks for flag reuse.
> > >>>
> > >>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> > >>> index b9e914e1face..c1ea49a7f198 100644
> > >>> --- a/include/linux/page-flags.h
> > >>> +++ b/include/linux/page-flags.h
> > >>> @@ -110,6 +110,7 @@ enum pageflags {
> > >>>             PG_workingset,
> > >>>             PG_error,
> > >>>             PG_owner_priv_1,        /* Owner use. If pagecache, fs may use*/
> > >>> +   PG_owner_priv_2,
> > >>
> > >> Oh no, no new page flags please :)
> > >>
> > >> Maybe just follow what Linux suggested: pass vma to pte_dirty() and
> > >> always return false for these special VMAs.
> > >
> > > ... or look into removing that one case that gives us headake.
> > >
> > > No idea what would happen if we do the following:
> > >
> > > CCing Yu Zhao.
> > >
> > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > index 0761f91b407f..d1dfbd4fd38d 100644
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -4280,14 +4280,9 @@ static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_c
> > >                   return true;
> > >           }
> > >
> > > -       /* dirty lazyfree */
> > > -       if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
> > > -               success = lru_gen_del_folio(lruvec, folio, true);
> > > -               VM_WARN_ON_ONCE_FOLIO(!success, folio);
> > > -               folio_set_swapbacked(folio);
> > > -               lruvec_add_folio_tail(lruvec, folio);
> > > -               return true;
> > > -       }
> > > +       /* lazyfree: we may not be allowed to set swapbacked: MAP_DROPPABLE */
> > > +       if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio))
> > > +               return false;
>
> This is an optimization to avoid an unnecessary trip to
> shrink_folio_list(), so it's safe to delete the entire 'if' block, and
> that would be preferable than leaving a dangling 'if'.
>
> > Note that something is unclear to me: are we maybe running into that
> > code also if folio_set_swapbacked() is already set and we are not in the
> > lazyfree path (in contrast to what is documented)?
>
> Not sure what you mean: either rmap sees pte_dirty() and does
> folio_mark_dirty() and then folio_set_swapbacked(); or MGLRU does the
> same sequence, with the first two steps in walk_pte_range() and the
> last one here.

Rationale: rmap is expensive (cache unfriendly) and MGLRU tries to
avoid using it.