Re: [PATCH 1/5] mm, vmscan: Do not account skipped pages as scanned

From: Minchan Kim
Date: Wed Jul 27 2016 - 21:37:56 EST


On Mon, Jul 25, 2016 at 10:20:14AM +0100, Mel Gorman wrote:
> On Mon, Jul 25, 2016 at 05:04:56PM +0900, Minchan Kim wrote:
> > > @@ -1429,6 +1429,9 @@ static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
> > > continue;
> > > }
> > >
> > > + /* Pages skipped do not contribute to scan */
> > > + scan++;
> > > +
> >
> > As I mentioned in previous version, under irq-disabled-spin-lock, such
> > unbounded operation would make the latency spike worse if there are
> > lot of pages we should skip.
> >
> > Don't we take care it?
>
> It's not unbounded, it's bound by the size of the LRU list and it's not
> going to be enough to trigger a warning. While the lock hold time may be
> undesirable, unlocking it every SWAP_CLUSTER_MAX pages may increase overall
> contention. There also is the question of whether skipped pages should be
> temporarily putback before unlocking the LRU to avoid isolated pages being
> unavailable for too long. It also cannot easily just return early without
> prematurely triggering OOM due to a lack of progress. I didn't feel the
> complexity was justified.

I measured the lock holding time and it took max 96ms during 360M
scanning with hackbench. It was very easy to reproduce with node-lru
because it should skip too many pages.

Given that my box is much faster than usual mobile CPU, it would
take more time in embedded system. I think irq disable during 96ms would
be worth to be fixed.

Anyway, I'm done by that which I measured time by hand so it's up to you
that whether you want to fix or leave as it is until someone reports it with
more real workload.

>
> --
> Mel Gorman
> SUSE Labs
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>