On Mon, Jan 30, 2012 at 02:06:50PM +0100, Michal Nazarewicz wrote:

quoted

quoted

@@ -1017,11 +1049,14 @@ __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)

			rmv_page_order(page);

			/* Take ownership for orders >= pageblock_order */
-			if (current_order >= pageblock_order)
+			if (current_order >= pageblock_order &&
+			    !is_pageblock_cma(page))
				change_pageblock_range(page, current_order,
							start_migratetype);

-			expand(zone, page, order, current_order, area, migratetype);
+			expand(zone, page, order, current_order, area,
+			       is_migrate_cma(start_migratetype)
+			     ? start_migratetype : migratetype);

What is this check meant to be doing?

start_migratetype is determined by allocflags_to_migratetype() and
that never will be MIGRATE_CMA so is_migrate_cma(start_migratetype)
should always be false.

Right, thanks!  This should be the other way around, ie.:

+			expand(zone, page, order, current_order, area,
+			       is_migrate_cma(migratetype)
+			     ? migratetype : start_migratetype);

I'll fix this and the calls to is_pageblock_cma().

That makes a lot more sense. Thanks.

I have a vague recollection that there was a problem with finding
unmovable pages in MIGRATE_CMA regions. This might have been part of
the problem.

-- 
Mel Gorman
SUSE Labs