Date:	Sun, 25 Nov 2012 23:10:41 -0500
From:	Johannes Weiner <hannes@...xchg.org>
To:	Rik van Riel <riel@...hat.com>
Cc:	Johannes Hirte <johannes.hirte@....tu-ilmenau.de>,
	akpm@...ux-foundation.org, mgorman@...e.de,
	Valdis.Kletnieks@...edu, jirislaby@...il.com, jslaby@...e.cz,
	zkabelac@...hat.com, mm-commits@...r.kernel.org,
	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	torvalds@...ux-foundation.org
Subject: Re: [PATCH] mm,vmscan: only loop back if compaction would fail in
 all zones

On Sun, Nov 25, 2012 at 10:15:18PM -0500, Johannes Weiner wrote:
> On Sun, Nov 25, 2012 at 07:16:45PM -0500, Rik van Riel wrote:
> > On Sun, 25 Nov 2012 17:44:33 -0500
> > Johannes Weiner <hannes@...xchg.org> wrote:
> > > On Sun, Nov 25, 2012 at 01:29:50PM -0500, Rik van Riel wrote:
> > 
> > > > Could you try this patch?
> > > 
> > > It's not quite enough because it's not reaching the conditions you
> > > changed, see analysis in https://lkml.org/lkml/2012/11/20/567
> > 
> > Johannes,
> > 
> > does the patch below fix your problem?
> 
> I can not reproduce the problem anymore with my smoke test.
> 
> > I suspect it would, because kswapd should only ever run into this
> > particular problem when we have a tiny memory zone in a pgdat,
> > and in that case we will also have a larger zone nearby, where
> > compaction would just succeed.
> 
> What if there is a higher order GFP_DMA allocation when the other
> zones in the system meet the high watermark for this order?
> 
> There is something else that worries me: if the preliminary zone scan
> finds the high watermark of all zones alright, end_zone is at its
> initialization value, 0.  The final compaction loop at `if (order)'
> goes through all zones up to and including end_zone, which was never
> really set to anything meaningful(?) and the only zone considered is
> the DMA zone again.  Very unlikely, granted, but if you'd ever hit
> that race and kswapd gets stuck, this will be fun to debug...

I actually liked your first idea better: force reclaim until the
compaction watermark is met.  The only problem was that still not
every check in there agreed when the zone was considered balanced and
so no actual reclaim happened.

So how about making everybody agree?  If the high watermark is met but
not the compaction one, keep doing reclaim AND don't consider the zone
balanced, AND don't make it contribute to balanced_pages etc.?  This
makes sure reclaim really does not bail and that the node is never
considered alright when it's actually not according to compaction.
This patch fixes the problem too (at least for the smoke test so far)
and IMO makes the code a bit more understandable.

We may be able to drop some of the relooping conditions.  We may also
be able to reduce the pressure from the DMA zone by passing the right
classzone_idx in there.  Needs more thought.

---
From: Johannes Weiner <hannes@...xchg.org>
Subject: [patch] mm: vmscan: fix endless loop in kswapd balancing

Kswapd does not in all places have the same criteria for when it
considers a zone balanced.  This leads to zones being not reclaimed
because they are considered just fine and the compaction checks to
loop over the zonelist again because they are considered unbalanced,
causing kswapd to run forever.

Add a function, zone_balanced(), that checks the watermark and if
compaction has enough free memory to do its job.  Then use it
uniformly for when kswapd needs to check if a zone is balanced.

Signed-off-by: Johannes Weiner <hannes@...xchg.org>
---
 mm/vmscan.c | 27 ++++++++++++++++++---------
 1 file changed, 18 insertions(+), 9 deletions(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 48550c6..3b0aef4 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2397,6 +2397,19 @@ static void age_active_anon(struct zone *zone, struct scan_control *sc)
 	} while (memcg);
 }
 
+static bool zone_balanced(struct zone *zone, int order,
+			  unsigned long balance_gap, int classzone_idx)
+{
+	if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
+				    balance_gap, classzone_idx, 0))
+		return false;
+
+	if (COMPACTION_BUILD && order && !compaction_suitable(zone, order))
+		return false;
+
+	return true;
+}
+
 /*
  * pgdat_balanced is used when checking if a node is balanced for high-order
  * allocations. Only zones that meet watermarks and are in a zone allowed
@@ -2475,8 +2488,7 @@ static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
 			continue;
 		}
 
-		if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
-							i, 0))
+		if (!zone_balanced(zone, order, 0, i))
 			all_zones_ok = false;
 		else
 			balanced += zone->present_pages;
@@ -2585,8 +2597,7 @@ static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
 				break;
 			}
 
-			if (!zone_watermark_ok_safe(zone, order,
-					high_wmark_pages(zone), 0, 0)) {
+			if (!zone_balanced(zone, order, 0, 0)) {
 				end_zone = i;
 				break;
 			} else {
@@ -2662,9 +2673,8 @@ static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
 				testorder = 0;
 
 			if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
-				    !zone_watermark_ok_safe(zone, testorder,
-					high_wmark_pages(zone) + balance_gap,
-					end_zone, 0)) {
+			    !zone_balanced(zone, testorder,
+					   balance_gap, end_zone)) {
 				shrink_zone(zone, &sc);
 
 				reclaim_state->reclaimed_slab = 0;
@@ -2691,8 +2701,7 @@ static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
 				continue;
 			}
 
-			if (!zone_watermark_ok_safe(zone, testorder,
-					high_wmark_pages(zone), end_zone, 0)) {
+			if (!zone_balanced(zone, testorder, 0, end_zone)) {
 				all_zones_ok = 0;
 				/*
 				 * We are still under min water mark.  This
-- 
1.7.11.7


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

Powered by blists - more mailing lists