Re: [PATCH] mm: compaction: Abort compaction if too many pages areisolated and caller is asynchronous

[Mel Gorman]

On Sat, Jun 04, 2011 at 03:58:53PM +0900, Minchan Kim wrote:
> On Fri, Jun 03, 2011 at 03:49:41PM +0100, Mel Gorman wrote:
> > On Fri, Jun 03, 2011 at 03:09:20AM +0100, Mel Gorman wrote:
> > > On Thu, Jun 02, 2011 at 05:37:54PM +0200, Andrea Arcangeli wrote:
> > > > > There is an explanation in here somewhere because as I write this,
> > > > > the test machine has survived 14 hours under continual stress without
> > > > > the isolated counters going negative with over 128 million pages
> > > > > successfully migrated and a million pages failed to migrate due to
> > > > > direct compaction being called 80,000 times. It's possible it's a
> > > > > co-incidence but it's some co-incidence!
> > > > 
> > > > No idea...
> > > 
> > > I wasn't able to work on this most of the day but was looking at this
> > > closer this evening again and I think I might have thought of another
> > > theory that could cause this problem.
> > > 
> > > When THP is isolating pages, it accounts for the pages isolated against
> > > the zone of course. If it backs out, it finds the pages from the PTEs.
> > > On !SMP but PREEMPT, we may not have adequate protection against a new
> > > page from a different zone being inserted into the PTE causing us to
> > > decrement against the wrong zone. While the global counter is fine,
> > > the per-zone counters look corrupted. You'd still think it was the
> > > anon counter tht got screwed rather than the file one if it really was
> > > THP unfortunately so it's not the full picture. I'm going to start
> > > a test monitoring both zoneinfo and vmstat to see if vmstat looks
> > > fine while the per-zone counters that are negative are offset by a
> > > positive count on the other zones that when added together become 0.
> > > Hopefully it'll actually trigger overnight :/
> > > 
> > 
> > Right idea of the wrong zone being accounted for but wrong place. I
> > think the following patch should fix the problem;
> > 
> > ==== CUT HERE ===
> > mm: compaction: Ensure that the compaction free scanner does not move to the next zone
> > 
> > Compaction works with two scanners, a migration and a free
> > scanner. When the scanners crossover, migration within the zone is
> > complete. The location of the scanner is recorded on each cycle to
> > avoid excesive scanning.
> > 
> > When a zone is small and mostly reserved, it's very easy for the
> > migration scanner to be close to the end of the zone. Then the following
> > situation can occurs
> > 
> >   o migration scanner isolates some pages near the end of the zone
> >   o free scanner starts at the end of the zone but finds that the
> >     migration scanner is already there
> >   o free scanner gets reinitialised for the next cycle as
> >     cc->migrate_pfn + pageblock_nr_pages
> >     moving the free scanner into the next zone
> >   o migration scanner moves into the next zone but continues accounting
> >     against the old zone
> > 
> > When this happens, NR_ISOLATED accounting goes haywire because some
> > of the accounting happens against the wrong zone. One zones counter
> > remains positive while the other goes negative even though the overall
> > global count is accurate. This was reported on X86-32 with !SMP because
> > !SMP allows the negative counters to be visible. The fact that it is
> > difficult to reproduce on X86-64 is probably just a co-incidence as
> 
> I guess it's related to zone sizes.
> X86-64 has small DMA and large DMA32 zones for fallback of NORMAL while
> x86 has just a small DMA(16M) zone.
> 

Yep, this is a possibility as well as the use of lowmem reserves.

> I think DMA zone in x86 is easily full of non-LRU or non-movable pages.

Maybe not full, but it has more PageReserved pages than anywhere else
and few MIGRATE_MOVABLE blocks. MIGRATE_MOVABLE gets skipped during
async compaction we could easily reach the end of the DMA zone quickly.

> So isolate_migratepagse continues to scan for finding pages which are migratable
> and then it reaches near end of zone.
> 
> > the bug should theoritically be possible there.
>
> Finally, you found it. Congratulations on!
> > Signed-off-by: Mel Gorman <mgorman@xxxxxxx>
> Reviewed-by: Minchan Kim <minchan.kim@xxxxxxxxx>
> 
> When we are debugging this problem, we found a few of bugs and enhance points
> and submitted patches. It was a very good chance to fix Linux VM.
> 

Thanks.

-- 
Mel Gorman
SUSE Labs
--
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/