Re: [PATCH] mm: mglru: clear the reference counter for rejected folios

[Baolin Wang]

On 9/8/26 2:59 PM, Baoquan He wrote:
> On 09/08/26 at 12:01pm, Baolin Wang wrote:
> > On 9/8/26 11:03 AM, Baolin Wang wrote:
> > > On 9/8/26 10:34 AM, Barry Song wrote:
> > > > On Tue, Sep 8, 2026 at 10:30 AM Baoquan He <baoquan.he@xxxxxxxxx> wrote:
> > > > > Hi Baolin,
> > > > >
> > > > > On 09/07/26 at 11:25am, Baolin Wang wrote:
> > > > > ......snip...
> > > > > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > > > > index 40d3f1b48a74..42c0a09938ab 100644
> > > > > > --- a/mm/vmscan.c
> > > > > > +++ b/mm/vmscan.c
> > > > >
> > > > > Well, this seems to be based on Andrew's mm-new branch. I usually track
> > > > > mm-unstable branch. Maybe the subject should be marked as below?
> > > > > [PATCH mm-new] mm: mglru: clear the reference counter for rejected
> > >
> > > ACK.
> > >
> > > > > > @@ -5021,10 +5021,11 @@ static int evict_folios(unsigned
> > > > > > long nr_to_scan, struct lruvec *lruvec,
> > > > > >                 }
> > > > > >
> > > > > >                 /* don't add rejected folios to the oldest generation */
> > > > > > -             if (lru_gen_folio_seq(lruvec, folio, false) ==
> > > > > > min_seq[type]) {
> > > > > > -                     folio_set_lru_refs(folio, 0);
> > > > > > +             if (lru_gen_folio_seq(lruvec, folio, false) ==
> > > > > > min_seq[type])
> > > > > >                         folio_set_active(folio);
> > > > > > -             }
> > > > > > +
> > > > > > +             /* See the comments on LRU_REFS_FLAGS */
> > > > > > +             folio_set_lru_refs(folio, 0);
> > > > >
> > > > > This looks like a great catch, while the code change could bring issue.
> > > > >
> > > > > Because move_folios_to_lru() relies on folios' flags to decide their new
> > > > > generation. You just cleared it before move_folios_to_lru(). This is no
> > > > > problem for rejected folios that are determined to be put into the
> > > > > oldest generation. But for those rejected folios that are determined to
> > > > > be promoted, this could be wrong. E.g currently gen window is 4, and a
> > > > > folio is referenced, lru_gen_folio_seq() decides its new gen as 1, which
> > > > > is the 2nd oldest generation. While folio_set_lru_refs(folio, 0) clear
> > > > > referenced bit, this causes it being put into the oldest generation in
> > > > > move_folios_to_lru(), this is not expected.
> > >
> > > Yes. As I discussed with Barry earlier, lru_gen_folio_seq() also needs
> > > to be reconsidered regarding whether it should rely on PG_referenced
> > > [1].
> > >
> > > For commit 6cbdd9726fb5, we didn't discuss the impact on rejected folios
> > > either. Before commit 6cbdd9726fb5, if rejected folios did not have
> > > PG_active set by shrink_folio_list(), evict_folios() would set PG_active
> > > on these rejected folios.
> > >
> > > [1] https://lore.kernel.org/linux-mm/20260901220430.79810-1-
> > > baohua@xxxxxxxxxx/
> > >
> > > > The original code looks quite weird. It even prioritizes folios
> > > > that won't be promoted by `PG_active`. Do we need to change all
> > > > the cases just to call `PG_active`?
>
> I would agree if we can. While I have one concern. A rejected folio that
> should have gone into the oldest generation is now being promoted to the
> 2nd newest generation. In the original code, referenced folio is only
> being promoted to the next gen. I even think this is not a bug, but Yu

That's not quite true. Before commit 6cbdd9726fb5, a rejected referenced 
folio was also put back to the 2nd youngest gen.

> Zhao intentionally did it: the coldest folio is moved to 2nd newest gen,
> referenced folio (hot folio) is moved to new gen but carries the referenced
> bit. Both of them seems to be treated somewhat equally.
>
> To me, I would rather move both of them to the next gen, while keep their
> refs untouched.

IMHO, I strongly recommend not doing this, and that's exactly the 
motivation behind my patch. Because this is also being treated as a 
promotion, it should behave like folio_inc_gen() or folio_update_gen() 
and clear the refs after promotion. I think this is a fundamental 
principle of promotion. Otherwise, ref-based promotion is already 
completely broken.

Next, I also plan to clean up refs in lru_gen_set_refs() as discussed 
with Barry.

> > > Yes. Regarding this concern, I plan to change back to the original
> > > behavior:
> > >
> > >       /* See the comments on LRU_REFS_FLAGS */
> > >       folio_set_lru_refs(folio, 0);
> > >
> > >       /* don't add rejected folios to the oldest generation */
> > >       if (lru_gen_folio_seq(lruvec, folio, false) == min_seq[type])
> > >           folio_set_active(folio);
> > >
> > > What do you think?
> >
> > Just FYI, after above changes, the performance improvement on zram is no
> > longer obvious either. I think this also answers Kairui's earlier question
> > about why I saw a performance improvement (which seems related to commit
> > 6cbdd9726fb5). Also, there is no obvious performance regression either.
>
> Hi Baolin,
>
> Not sure if it's convenient to do a little more testing in your side.
> E.g rejected folios are moved to next gen, but not clearing their flags.

I'm not sure what you mean by "next gen" here. If you mean the 2nd 
oldest gen, I actually tested that too, that is, clearing refs after 
move_folios_to_lru(), and there wasn't any noticeable performance impact 
either (but the code was a bit hacky, so I didn't go with this approach).

> Then we can at least know which one is more efficient on yout testing
> platform.
>
> coldest folio      referenced folio
> ---------------------------------------------
> max_gen -1         next gen + referenced kept
> max_gen -1         max_gen -1
> next_gen           next gen + referenced kept
>
> Thanks
> Baoquan