mm: compaction: make isolate_lru_page() filter-aware again
authorMel Gorman <mgorman@suse.de>
Fri, 13 Jan 2012 01:19:38 +0000 (17:19 -0800)
committerBen Hutchings <ben@decadent.org.uk>
Thu, 2 Aug 2012 13:37:31 +0000 (14:37 +0100)
commit c82449352854ff09e43062246af86bdeb628f0c3 upstream.

Stable note: Not tracked in Bugzilla. A fix aimed at preserving page aging
information by reducing LRU list churning had the side-effect of
reducing THP allocation success rates. This was part of a series
to restore the success rates while preserving the reclaim fix.

Commit 39deaf85 ("mm: compaction: make isolate_lru_page() filter-aware")
noted that compaction does not migrate dirty or writeback pages and that
is was meaningless to pick the page and re-add it to the LRU list.  This
had to be partially reverted because some dirty pages can be migrated by
compaction without blocking.

This patch updates "mm: compaction: make isolate_lru_page" by skipping
over pages that migration has no possibility of migrating to minimise LRU
disruption.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel<riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Andy Isaacson <adi@hexapodia.org>
Cc: Nai Xia <nai.xia@gmail.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
include/linux/mmzone.h
mm/compaction.c
mm/vmscan.c

index 905b1e1..25842b6 100644 (file)
@@ -173,6 +173,8 @@ static inline int is_unevictable_lru(enum lru_list l)
 #define ISOLATE_CLEAN          ((__force isolate_mode_t)0x4)
 /* Isolate unmapped file */
 #define ISOLATE_UNMAPPED       ((__force isolate_mode_t)0x8)
+/* Isolate for asynchronous migration */
+#define ISOLATE_ASYNC_MIGRATE  ((__force isolate_mode_t)0x10)
 
 /* LRU Isolation modes. */
 typedef unsigned __bitwise__ isolate_mode_t;
index b81625d..979a919 100644 (file)
@@ -371,6 +371,9 @@ static isolate_migrate_t isolate_migratepages(struct zone *zone,
                        continue;
                }
 
+               if (!cc->sync)
+                       mode |= ISOLATE_ASYNC_MIGRATE;
+
                /* Try isolate the page */
                if (__isolate_lru_page(page, mode, 0) != 0)
                        continue;
index 8342119..1b95e4c 100644 (file)
@@ -1061,8 +1061,39 @@ int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
 
        ret = -EBUSY;
 
-       if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
-               return ret;
+       /*
+        * To minimise LRU disruption, the caller can indicate that it only
+        * wants to isolate pages it will be able to operate on without
+        * blocking - clean pages for the most part.
+        *
+        * ISOLATE_CLEAN means that only clean pages should be isolated. This
+        * is used by reclaim when it is cannot write to backing storage
+        *
+        * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
+        * that it is possible to migrate without blocking
+        */
+       if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
+               /* All the caller can do on PageWriteback is block */
+               if (PageWriteback(page))
+                       return ret;
+
+               if (PageDirty(page)) {
+                       struct address_space *mapping;
+
+                       /* ISOLATE_CLEAN means only clean pages */
+                       if (mode & ISOLATE_CLEAN)
+                               return ret;
+
+                       /*
+                        * Only pages without mappings or that have a
+                        * ->migratepage callback are possible to migrate
+                        * without blocking
+                        */
+                       mapping = page_mapping(page);
+                       if (mapping && !mapping->a_ops->migratepage)
+                               return ret;
+               }
+       }
 
        if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
                return ret;