Re: [PATCH] mm: vmscan: count only dirty pages as congested

From: Andrew Morton
Date: Wed Oct 15 2014 - 19:16:30 EST


On Wed, 15 Oct 2014 16:07:42 -0700 Jamie Liu <jamieliu@xxxxxxxxxx> wrote:

> On Wed, Oct 15, 2014 at 1:05 PM, Andrew Morton
> <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > On Wed, 15 Oct 2014 12:58:35 -0700 Jamie Liu <jamieliu@xxxxxxxxxx> wrote:
> >
> >> shrink_page_list() counts all pages with a mapping, including clean
> >> pages, toward nr_congested if they're on a write-congested BDI.
> >> shrink_inactive_list() then sets ZONE_CONGESTED if nr_dirty ==
> >> nr_congested. Fix this apples-to-oranges comparison by only counting
> >> pages for nr_congested if they count for nr_dirty.
> >>
> >> ...
> >>
> >> --- a/mm/vmscan.c
> >> +++ b/mm/vmscan.c
> >> @@ -875,7 +875,8 @@ static unsigned long shrink_page_list(struct list_head *page_list,
> >> * end of the LRU a second time.
> >> */
> >> mapping = page_mapping(page);
> >> - if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
> >> + if (((dirty || writeback) && mapping &&
> >> + bdi_write_congested(mapping->backing_dev_info)) ||
> >> (writeback && PageReclaim(page)))
> >> nr_congested++;
> >
> > What are the observed runtime effects of this change?
>
> wait_iff_congested() only waits if ZONE_CONGESTED is set (and at least
> one BDI is still congested). Modulo concurrent changes to BDI
> congestion status:
>
> After this change, the probability that a given shrink_inactive_list()
> sets ZONE_CONGESTED increases monotonically with the fraction of dirty
> pages on the LRU, to 100% if all dirty pages are backed by a
> write-congested BDI. This is in line with what appears to intended,
> judging by the comment:
>
> /*
> * Tag a zone as congested if all the dirty pages scanned were
> * backed by a congested BDI and wait_iff_congested will stall.
> */
> if (nr_dirty && nr_dirty == nr_congested)
> set_bit(ZONE_CONGESTED, &zone->flags);
>
> Before this change, the probability that a given
> shrink_inactive_list() sets ZONE_CONGESTED varies erratically. Because
> the ZONE_CONGESTED condition is nr_dirty && nr_dirty == nr_congested,
> the probability peaks when the fraction of dirty pages is equal to the
> fraction of file pages backed by congested BDIs. So under some
> circumstances, an increase in the fraction of dirty pages or in the
> fraction of congested pages can actually result in an *decreased*
> probability that reclaim will stall for writeback congestion, and vice
> versa; which is both counterintuitive and counterproductive.

(top-posting repaired. Please don't do that!)

OK, I buy all that. But has any runtime testing been performed to
confirm this and to quantify the effects?

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/