Re: [PATCH 4/8] mm/swap: Use PG_dropbehind instead of PG_reclaim

[Yu Zhao <yuzhao@google.com>]

On Tue, Jan 14, 2025 at 9:28 PM Yu Zhao <yuzhao@google.com> wrote:
>
> On Tue, Jan 14, 2025 at 11:03 AM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > On Tue, Jan 14, 2025 at 12:12 AM Kirill A. Shutemov
> > <kirill.shutemov@linux.intel.com> wrote:
> > >
> > > On Mon, Jan 13, 2025 at 08:17:20AM -0800, Yosry Ahmed wrote:
> > > > On Mon, Jan 13, 2025 at 1:35 AM Kirill A. Shutemov
> > > > <kirill.shutemov@linux.intel.com> wrote:
> > > > >
> > > > > The recently introduced PG_dropbehind allows for freeing folios
> > > > > immediately after writeback. Unlike PG_reclaim, it does not need vmscan
> > > > > to be involved to get the folio freed.
> > > > >
> > > > > Instead of using folio_set_reclaim(), use folio_set_dropbehind() in
> > > > > lru_deactivate_file().
> > > > >
> > > > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > > > > ---
> > > > >  mm/swap.c | 8 +-------
> > > > >  1 file changed, 1 insertion(+), 7 deletions(-)
> > > > >
> > > > > diff --git a/mm/swap.c b/mm/swap.c
> > > > > index fc8281ef4241..4eb33b4804a8 100644
> > > > > --- a/mm/swap.c
> > > > > +++ b/mm/swap.c
> > > > > @@ -562,14 +562,8 @@ static void lru_deactivate_file(struct lruvec *lruvec, struct folio *folio)
> > > > >         folio_clear_referenced(folio);
> > > > >
> > > > >         if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
> > > > > -               /*
> > > > > -                * Setting the reclaim flag could race with
> > > > > -                * folio_end_writeback() and confuse readahead.  But the
> > > > > -                * race window is _really_ small and  it's not a critical
> > > > > -                * problem.
> > > > > -                */
> > > > >                 lruvec_add_folio(lruvec, folio);
> > > > > -               folio_set_reclaim(folio);
> > > > > +               folio_set_dropbehind(folio);
> > > > >         } else {
> > > > >                 /*
> > > > >                  * The folio's writeback ended while it was in the batch.
> > > >
> > > > Now there's a difference in behavior here depending on whether or not
> > > > the folio is under writeback (or will be written back soon). If it is,
> > > > we set PG_dropbehind to get it freed right after, but if writeback has
> > > > already ended we put it on the tail of the LRU to be freed later.
> > > >
> > > > It's a bit counterintuitive to me that folios with pending writeback
> > > > get freed faster than folios that completed their writeback already.
> > > > Am I missing something?
> > >
> > > Yeah, it is strange.
> > >
> > > I think we can drop the writeback/dirty check. Set PG_dropbehind and put
> > > the page on the tail of LRU unconditionally. The check was required to
> > > avoid confusion with PG_readahead.
> > >
> > > Comment above the function is not valid anymore.
> >
> > My read is that we don't put dirty/writeback folios at the tail of the
> > LRU because they cannot be freed immediately and we want to give them
> > time to be written back before reclaim reaches them. So I don't think
> > we want to change that and always put the pages at the tail.
> >
> > >
> > > But the folio that is still dirty under writeback will be freed faster as
> > > we get rid of the folio just after writeback is done while clean page can
> > > dangle on LRU for a while.
> >
> > Yeah if we reuse PG_dropbehind then we cannot avoid
> > folio_end_writeback() freeing the folio faster than clean ones.
> >
> > >
> > > I don't think we have any convenient place to free clean dropbehind page
> > > other than shrink_folio_list(). Or do we?
> >
> > Not sure tbh. FWIW I am not saying it's necessarily a bad thing to
> > free dirty/writeback folios before clean ones when deactivated, it's
> > just strange and a behavioral change from today that I wanted to point
> > out. Perhaps that's the best we can do for now.
> >
> > >
> > > Looking at shrink_folio_list(), I think we need to bypass page demotion
> > > for PG_dropbehind pages.
>
> I agree with Yosry. I don't think lru_deactivate_file() is still
> needed -- it was needed only because when truncation fails to free a
> dirty/writeback folio, page reclaim can do that quickly. For other
> conditions that mapping_evict_folio() returns 0, there isn't much page
> reclaim can do, and those conditions are not deactivate_file_folio()
> and lru_deactivate_file()'s intentions. So the following should be
> enough, and it's a lot cleaner :
>
> diff --git a/mm/truncate.c b/mm/truncate.c
> index e2e115adfbc5..12d2aa608517 100644
> --- a/mm/truncate.c
> +++ b/mm/truncate.c
> @@ -486,7 +486,7 @@ unsigned long mapping_try_invalidate(struct
> address_space *mapping,
>                          * of interest and try to speed up its reclaim.
>                          */
>                         if (!ret) {
> -                               deactivate_file_folio(folio);
> +                               folio_set_dropbehind(folio)
>                                 /* Likely in the lru cache of a remote CPU */
>                                 if (nr_failed)
>                                         (*nr_failed)++;
>
> Then we can drop deactivate_file_folio() and lru_deactivate_file().

And with the above and list_move_tail() removed, we can also remove
lruvec_add_folio_tail().