mm: page_alloc: exempt GFP_THISNODE allocations from zone fairness

Jan Stancek reports manual page migration encountering allocation
failures after some pages when there is still plenty of memory free, and
bisected the problem down to commit 81c0a2bb515f ("mm: page_alloc: fair
zone allocator policy").

The problem is that GFP_THISNODE obeys the zone fairness allocation
batches on one hand, but doesn't reset them and wake kswapd on the other
hand.  After a few of those allocations, the batches are exhausted and
the allocations fail.

Fixing this means either having GFP_THISNODE wake up kswapd, or
GFP_THISNODE not participating in zone fairness at all.  The latter
seems safer as an acute bugfix, we can clean up later.

Reported-by: Jan Stancek <jstancek@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: <stable@kernel.org>		[3.12+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 22 insertions, 4 deletions

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 3d1bf889465a..3bac76ae4b30 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1238,6 +1238,15 @@ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
         }
         local_irq_restore(flags);
 }
+static bool gfp_thisnode_allocation(gfp_t gfp_mask)
+{
+        return (gfp_mask & GFP_THISNODE) == GFP_THISNODE;
+}
+#else
+static bool gfp_thisnode_allocation(gfp_t gfp_mask)
+{
+        return false;
+}
 #endif
 
 /*
@@ -1574,7 +1583,13 @@ again:
                                           get_pageblock_migratetype(page));
         }
 
-        __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
+        /*
+         * NOTE: GFP_THISNODE allocations do not partake in the kswapd
+         * aging protocol, so they can't be fair.
+         */
+        if (!gfp_thisnode_allocation(gfp_flags))
+                __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
+
         __count_zone_vm_events(PGALLOC, zone, 1 << order);
         zone_statistics(preferred_zone, zone, gfp_flags);
         local_irq_restore(flags);
@@ -1946,8 +1961,12 @@ zonelist_scan:
                  * ultimately fall back to remote zones that do not
                  * partake in the fairness round-robin cycle of this
                  * zonelist.
+                 *
+                 * NOTE: GFP_THISNODE allocations do not partake in
+                 * the kswapd aging protocol, so they can't be fair.
                  */
-                if (alloc_flags & ALLOC_WMARK_LOW) {
+                if ((alloc_flags & ALLOC_WMARK_LOW) &&
+                    !gfp_thisnode_allocation(gfp_mask)) {
                         if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
                                 continue;
                         if (!zone_local(preferred_zone, zone))
@@ -2503,8 +2522,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
          * allowed per node queues are empty and that nodes are
          * over allocated.
          */
-        if (IS_ENABLED(CONFIG_NUMA) &&
-                        (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
+        if (gfp_thisnode_allocation(gfp_mask))
                 goto nopage;
 
 restart: