mm, compaction: khugepaged should not give up due to need_resched()

Async compaction aborts when it detects zone lock contention or
need_resched() is true.  David Rientjes has reported that in practice,
most direct async compactions for THP allocation abort due to
need_resched().  This means that a second direct compaction is never
attempted, which might be OK for a page fault, but khugepaged is intended
to attempt a sync compaction in such case and in these cases it won't.

This patch replaces "bool contended" in compact_control with an int that
distinguishes between aborting due to need_resched() and aborting due to
lock contention.  This allows propagating the abort through all compaction
functions as before, but passing the abort reason up to
__alloc_pages_slowpath() which decides when to continue with direct
reclaim and another compaction attempt.

Another problem is that try_to_compact_pages() did not act upon the
reported contention (both need_resched() or lock contention) immediately
and would proceed with another zone from the zonelist.  When
need_resched() is true, that means initializing another zone compaction,
only to check again need_resched() in isolate_migratepages() and aborting.
 For zone lock contention, the unintended consequence is that the lock
contended status reported back to the allocator is detrmined from the last
zone where compaction was attempted, which is rather arbitrary.

This patch fixes the problem in the following way:
- async compaction of a zone aborting due to need_resched() or fatal signal
  pending means that further zones should not be tried. We report
  COMPACT_CONTENDED_SCHED to the allocator.
- aborting zone compaction due to lock contention means we can still try
  another zone, since it has different set of locks. We report back
  COMPACT_CONTENDED_LOCK only if *all* zones where compaction was attempted,
  it was aborted due to lock contention.

As a result of these fixes, khugepaged will proceed with second sync
compaction as intended, when the preceding async compaction aborted due to
need_resched().  Page fault compactions aborting due to need_resched()
will spare some cycles previously wasted by initializing another zone
compaction only to abort again.  Lock contention will be reported only
when compaction in all zones aborted due to lock contention, and therefore
it's not a good idea to try again after reclaim.

In stress-highalloc from mmtests configured to use __GFP_NO_KSWAPD, this
has improved number of THP collapse allocations by 10%, which shows
positive effect on khugepaged.  The benchmark's success rates are
unchanged as it is not recognized as khugepaged.  Numbers of compact_stall
and compact_fail events have however decreased by 20%, with
compact_success still a bit improved, which is good.  With benchmark
configured not to use __GFP_NO_KSWAPD, there is 6% improvement in THP
collapse allocations, and only slight improvement in stalls and failures.

[akpm@linux-foundation.org: fix warnings]
Reported-by: David Rientjes <rientjes@google.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Minchan Kim <minchan@kernel.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 33 insertions, 12 deletions

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index dfbf54b51649..313338d74095 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2297,7 +2297,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
 	struct zonelist *zonelist, enum zone_type high_zoneidx,
 	nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
 	int classzone_idx, int migratetype, enum migrate_mode mode,
-	bool *contended_compaction, bool *deferred_compaction)
+	int *contended_compaction, bool *deferred_compaction)
 {
 	struct zone *last_compact_zone = NULL;
 	unsigned long compact_result;
@@ -2371,7 +2371,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
 	struct zonelist *zonelist, enum zone_type high_zoneidx,
 	nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
 	int classzone_idx, int migratetype, enum migrate_mode mode,
-	bool *contended_compaction, bool *deferred_compaction)
+	int *contended_compaction, bool *deferred_compaction)
 {
 	return NULL;
 }
@@ -2547,7 +2547,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	unsigned long did_some_progress;
 	enum migrate_mode migration_mode = MIGRATE_ASYNC;
 	bool deferred_compaction = false;
-	bool contended_compaction = false;
+	int contended_compaction = COMPACT_CONTENDED_NONE;
 
 	/*
 	 * In the slowpath, we sanity check order to avoid ever trying to
@@ -2651,15 +2651,36 @@ rebalance:
 	if (page)
 		goto got_pg;
 
-	/*
-	 * If compaction is deferred for high-order allocations, it is because
-	 * sync compaction recently failed. In this is the case and the caller
-	 * requested a movable allocation that does not heavily disrupt the
-	 * system then fail the allocation instead of entering direct reclaim.
-	 */
-	if ((deferred_compaction || contended_compaction) &&
-						(gfp_mask & __GFP_NO_KSWAPD))
-		goto nopage;
+	/* Checks for THP-specific high-order allocations */
+	if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
+		/*
+		 * If compaction is deferred for high-order allocations, it is
+		 * because sync compaction recently failed. If this is the case
+		 * and the caller requested a THP allocation, we do not want
+		 * to heavily disrupt the system, so we fail the allocation
+		 * instead of entering direct reclaim.
+		 */
+		if (deferred_compaction)
+			goto nopage;
+
+		/*
+		 * In all zones where compaction was attempted (and not
+		 * deferred or skipped), lock contention has been detected.
+		 * For THP allocation we do not want to disrupt the others
+		 * so we fallback to base pages instead.
+		 */
+		if (contended_compaction == COMPACT_CONTENDED_LOCK)
+			goto nopage;
+
+		/*
+		 * If compaction was aborted due to need_resched(), we do not
+		 * want to further increase allocation latency, unless it is
+		 * khugepaged trying to collapse.
+		 */
+		if (contended_compaction == COMPACT_CONTENDED_SCHED
+			&& !(current->flags & PF_KTHREAD))
+			goto nopage;
+	}
 
 	/*
 	 * It can become very expensive to allocate transparent hugepages at