Re: [PATCH 04/12] mm/swap: skip scanning cluster range if it's empty cluster

[Baoquan He <bhe@redhat.com>]

On 02/06/25 at 01:07am, Kairui Song wrote:
> On Wed, Feb 5, 2025 at 5:27 PM Baoquan He <bhe@redhat.com> wrote:
> >
> > Since ci->lock has been taken when isolating cluster from
> > si->free_clusters or taking si->percpu_cluster->next[order],
> > it's unnecessary to scan and check the cluster range availability
> > if i'ts empty cluster, and this can accelerate the huge page
> > swapping.
> >
> > Signed-off-by: Baoquan He <bhe@redhat.com>
> > ---
> >  mm/swapfile.c | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/mm/swapfile.c b/mm/swapfile.c
> > index 9c9a4ec6d4c6..61efde853eea 100644
> > --- a/mm/swapfile.c
> > +++ b/mm/swapfile.c
> > @@ -729,6 +729,9 @@ static bool cluster_scan_range(struct swap_info_struct *si,
> >         unsigned long offset, end = start + nr_pages;
> >         unsigned char *map = si->swap_map;
> >
> > +       if (cluster_is_empty(ci))
> > +               return true;
> > +
> 
> Hi Baoquan,
> 
> Thanks for the series.

Thanks for your reviewing.

> 
> Most commits are looking great, but this one is a bit questionable.
> cluster_scan_range is only called by alloc_swap_scan_cluster, and it
> already checks if the cluster has enough empty slots to use, so this
> might be redundant.

Hmm, maybe no. Assume we want to allocate 2M space on system with 4K
page size. Even if a empty cluster is taken into consideration,
cluster_scan_range() will loop 512 times to check if each slot is
available. That for sure is not necessary in the case, while the added
empty cluster checking is very cheap.

> 
> It is possible that cluster_scan_range sees an empty cluster if the
> cluster lock was dropped for reclaiming HAS_CACHE, but the chance
> should be extremely low, that this might be a negative optimization.

It may be not like that. If it's empty cluster, the added checking will
return directly. Then 'need_reclaim' is kept false, there's no chance to
drop cluster lock to do reclaiming for HAS_CACHE. Means for empty
cluster scanning, the ci->lock is kept held. Not sure if I missed
anything.

Thanks
Baoquan