On Tue, 21 Jun 2016, Vlastimil Babka wrote:
> > diff --git a/mm/compaction.c b/mm/compaction.c
> > --- a/mm/compaction.c
> > +++ b/mm/compaction.c
> > @@ -494,24 +494,22 @@ static unsigned long isolate_freepages_block(struct
> > compact_control *cc,
> >
> > /* Found a free page, will break it into order-0 pages */
> > order = page_order(page);
> > - isolated = __isolate_free_page(page, page_order(page));
> > + isolated = __isolate_free_page(page, order);
> > + if (!isolated)
> > + break;
>
> This seems to fix as a side-effect a bug in Joonsoo's mmotm patch
> mm-compaction-split-freepages-without-holding-the-zone-lock.patch, that
> Minchan found: http://marc.info/?l=linux-mm&m=146607176528495&w=2
>
> So it should be noted somewhere so they are merged together. Or Joonsoo posts
> an isolated fix and this patch has to rebase.
>
Indeed, I hadn't noticed the differences between Linus's tree and -mm.
Thanks very much for pointing it out.
My interest is to eventually backport this to a much older kernel where we
suffer from the same issue: it seems that we have always not terminated
the freeing scanner when splitting the free page fails and we feel it
because some of our systems have 128GB zones and migrate_pages() can call
compaction_alloc() several times if it keeps getting -EAGAIN. It's very
expensive.
I'm not sure we should label it as a -fix for
mm-compaction-split-freepages-without-holding-the-zone-lock.patch since
the problem this patch is addressing has seemingly existed for years.
Perhaps it would be better to have two patches, one as a -fix and then the
abort on page split failure on top. I'll send out a two patch series in
this form.
> > set_page_private(page, order);
> > total_isolated += isolated;
> > list_add_tail(&page->lru, freelist);
> >
> > - /* If a page was split, advance to the end of it */
> > - if (isolated) {
> > - cc->nr_freepages += isolated;
> > - if (!strict &&
> > - cc->nr_migratepages <= cc->nr_freepages) {
> > - blockpfn += isolated;
> > - break;
> > - }
> > -
> > - blockpfn += isolated - 1;
> > - cursor += isolated - 1;
> > - continue;
> > + /* Advance to the end of split page */
> > + cc->nr_freepages += isolated;
> > + if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
> > + blockpfn += isolated;
> > + break;
> > }
> > + blockpfn += isolated - 1;
> > + cursor += isolated - 1;
> > + continue;
> >
> > isolate_fail:
> > if (strict)
> > @@ -521,6 +519,9 @@ isolate_fail:
> >
> > }
> >
> > + if (locked)
> > + spin_unlock_irqrestore(&cc->zone->lock, flags);
> > +
> > /*
> > * There is a tiny chance that we have read bogus compound_order(),
> > * so be careful to not go outside of the pageblock.
> > @@ -542,9 +543,6 @@ isolate_fail:
> > if (strict && blockpfn < end_pfn)
> > total_isolated = 0;
> >
> > - if (locked)
> > - spin_unlock_irqrestore(&cc->zone->lock, flags);
> > -
> > /* Update the pageblock-skip if the whole pageblock was scanned */
> > if (blockpfn == end_pfn)
> > update_pageblock_skip(cc, valid_page, total_isolated, false);
> > @@ -622,7 +620,7 @@ isolate_freepages_range(struct compact_control *cc,
> > */
> > }
> >
> > - /* split_free_page does not map the pages */
> > + /* __isolate_free_page() does not map the pages */
> > map_pages(&freelist);
> >
> > if (pfn < end_pfn) {
> > @@ -1071,6 +1069,7 @@ static void isolate_freepages(struct compact_control
> > *cc)
> > block_end_pfn = block_start_pfn,
> > block_start_pfn -= pageblock_nr_pages,
> > isolate_start_pfn = block_start_pfn) {
> > + unsigned long isolated;
> >
> > /*
> > * This can iterate a massively long zone without finding any
> > @@ -1095,8 +1094,12 @@ static void isolate_freepages(struct compact_control
> > *cc)
> > continue;
> >
> > /* Found a block suitable for isolating free pages from. */
> > - isolate_freepages_block(cc, &isolate_start_pfn,
> > - block_end_pfn, freelist, false);
> > + isolated = isolate_freepages_block(cc, &isolate_start_pfn,
> > + block_end_pfn, freelist,
> > false);
> > + /* If free page split failed, do not continue needlessly */
>
> More accurately, free page isolation failed?
>
Eek, maybe. The condition should only work if we terminated early because
- need_resched() or zone->lock contention for MIGRATE_ASYNC, or
- __isolate_free_page() fails.
And the latter can only fail because of this (somewhat arbitrary) split
watermark check. I'll rename it because it includes both, but I thought
the next immediate condition check for cc->contended and its comment was
explanatory enough.
> > + if (!isolated && isolate_start_pfn < block_end_pfn &&
> > + cc->nr_freepages <= cc->nr_migratepages)
> > + break;
> >
> > /*
> > * If we isolated enough freepages, or aborted due to async
> > @@ -1124,7 +1127,7 @@ static void isolate_freepages(struct compact_control
> > *cc)
> > }
> > }
> >
> > - /* split_free_page does not map the pages */
> > + /* __isolate_free_page() does not map the pages */
> > map_pages(freelist);
> >
> > /*
> > @@ -1703,6 +1706,12 @@ enum compact_result try_to_compact_pages(gfp_t
> > gfp_mask, unsigned int order,
> > continue;
> > }
> >
> > + /* Don't attempt compaction if splitting free page will fail
> > */
> > + if (!zone_watermark_ok(zone, 0,
> > + low_wmark_pages(zone) + (1 << order),
> > + 0, 0))
> > + continue;
> > +
>
> Please don't add this, compact_zone already checks this via
> compaction_suitable() (and the usual 2 << order gap), so this is adding yet
> another watermark check with a different kind of gap.
>
Good point, thanks.
