mm: page_alloc: fix zone allocation fairness on UP
authorJohannes Weiner <hannes@cmpxchg.org>
Thu, 2 Oct 2014 23:21:10 +0000 (16:21 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 2 Oct 2014 23:28:44 +0000 (16:28 -0700)
The zone allocation batches can easily underflow due to higher-order
allocations or spills to remote nodes.  On SMP that's fine, because
underflows are expected from concurrency and dealt with by returning 0.
But on UP, zone_page_state will just return a wrapped unsigned long,
which will get past the <= 0 check and then consider the zone eligible
until its watermarks are hit.

Commit 3a025760fc15 ("mm: page_alloc: spill to remote nodes before
waking kswapd") already made the counter-resetting use
atomic_long_read() to accomodate underflows from remote spills, but it
didn't go all the way with it.

Make it clear that these batches are expected to go negative regardless
of concurrency, and use atomic_long_read() everywhere.

Fixes: 81c0a2bb515f ("mm: page_alloc: fair zone allocator policy")
Reported-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Leon Romanovsky <leon@leon.nu>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org> [3.12+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/page_alloc.c

index 18cee0d4c8a20705a4b3e7dd73e7d1d8a8b8d595..eee961958021e8c31e02ac6845b290a21d5a1690 100644 (file)
@@ -1612,7 +1612,7 @@ again:
        }
 
        __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
-       if (zone_page_state(zone, NR_ALLOC_BATCH) == 0 &&
+       if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
            !zone_is_fair_depleted(zone))
                zone_set_flag(zone, ZONE_FAIR_DEPLETED);
 
@@ -5701,9 +5701,8 @@ static void __setup_per_zone_wmarks(void)
                zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
 
                __mod_zone_page_state(zone, NR_ALLOC_BATCH,
-                                     high_wmark_pages(zone) -
-                                     low_wmark_pages(zone) -
-                                     zone_page_state(zone, NR_ALLOC_BATCH));
+                       high_wmark_pages(zone) - low_wmark_pages(zone) -
+                       atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
 
                setup_zone_migrate_reserve(zone);
                spin_unlock_irqrestore(&zone->lock, flags);