Re: [PATCH] mm/compaction: do not skip the free scanner's periodic lock drop

From: Kiryl Shutsemau (Meta)

Date: Tue Nov 17 2026 - 05:30:07 EST


On Mon, Nov 16, 2026 at 05:37:02PM -0800, SJ Park wrote:
> On Mon, 16 Nov 2026 12:57:45 +0000 "Kiryl Shutsemau (Meta)" <kas@xxxxxxxxxx> wrote:
>
> > isolate_freepages_block() drops zone->lock, checks for a fatal signal
> > and calls cond_resched() at every pfn that is a multiple of
> > COMPACT_CLUSTER_MAX.
> >
> > With a stride above one, the scan steps over a compound page, or over a
> > high-order free page it has just isolated, to a pfn that is not a
> > multiple of the stride. From there it never reaches a multiple of
> > COMPACT_CLUSTER_MAX, so the rest of the pageblock is scanned without the
> > check, possibly with zone->lock held and interrupts off.
> >
> > Count probes instead, and check every COMPACT_CLUSTER_MAX of them.
>
> Makes sense to me.
>
> >
> > Fixes: 4fca9730c51d ("mm, compaction: sample pageblocks for free pages")
>
> I understand you intentionally not Cc-ing stable@, buecase the user impact is
> not that bad. I think that makes sense. I just wanted to call out this for a
> case that it was not intentional.

Right, I have not see the actual user effect of the bug.

> > Signed-off-by: Kiryl Shutsemau (Meta) <kas@xxxxxxxxxx>
> > ---
> > mm/compaction.c | 12 +++++++-----
> > 1 file changed, 7 insertions(+), 5 deletions(-)
> >
> > diff --git a/mm/compaction.c b/mm/compaction.c
> > index f1b2060eb201..4465c1551406 100644
> > --- a/mm/compaction.c
> > +++ b/mm/compaction.c
> > @@ -579,7 +579,7 @@ static unsigned long isolate_freepages_block(struct compact_control *cc,
> > unsigned int stride,
> > bool strict)
> > {
> > - int nr_scanned = 0, total_isolated = 0;
> > + int nr_scanned = 0, nr_probes = 0, total_isolated = 0;
> > struct page *page;
> > unsigned long flags = 0;
> > bool locked = false;
> > @@ -599,13 +599,15 @@ static unsigned long isolate_freepages_block(struct compact_control *cc,
> > /*
> > * Periodically drop the lock (if held) regardless of its
> > * contention, to give chance to IRQs. Abort if fatal signal
> > - * pending.
> > + * pending. Count probes rather than pfns, as the scan does not
> > + * always land on a multiple of COMPACT_CLUSTER_MAX.
> > */
> > - if (!(blockpfn % COMPACT_CLUSTER_MAX)
> > - && compact_unlock_should_abort(&cc->zone->lock, flags,
> > - &locked, cc))
> > + if (!(nr_probes % COMPACT_CLUSTER_MAX) &&
> > + compact_unlock_should_abort(&cc->zone->lock, flags,
> > + &locked, cc))
>
> Why don't you let the later two lines unchanged? No strong opinion, no
> blocker.

&& on the new line is odd to me.


> Reviewed-by: SJ Park <sj@xxxxxxxxxx>

Thanks!

--
Kiryl Shutsemau / Kirill A. Shutemov