mm: vmscan: move direct reclaim wait_iff_congested into shrink_list
authorMel Gorman <mgorman@suse.de>
Wed, 3 Jul 2013 22:02:02 +0000 (15:02 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 3 Jul 2013 23:07:29 +0000 (16:07 -0700)
shrink_inactive_list makes decisions on whether to stall based on the
number of dirty pages encountered.  The wait_iff_congested() call in
shrink_page_list does no such thing and it's arbitrary.

This patch moves the decision on whether to set ZONE_CONGESTED and the
wait_iff_congested call into shrink_page_list.  This keeps all the
decisions on whether to stall or not in the one place.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c

index 5f80d018bffaf162c3f74a933501161e8430e72f..4898daf074cfa88e7047d5a301d65d6db7367258 100644 (file)
@@ -695,7 +695,9 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                                      struct zone *zone,
                                      struct scan_control *sc,
                                      enum ttu_flags ttu_flags,
+                                     unsigned long *ret_nr_dirty,
                                      unsigned long *ret_nr_unqueued_dirty,
+                                     unsigned long *ret_nr_congested,
                                      unsigned long *ret_nr_writeback,
                                      unsigned long *ret_nr_immediate,
                                      bool force_reclaim)
@@ -1017,20 +1019,13 @@ keep:
                VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
        }
 
-       /*
-        * Tag a zone as congested if all the dirty pages encountered were
-        * backed by a congested BDI. In this case, reclaimers should just
-        * back off and wait for congestion to clear because further reclaim
-        * will encounter the same problem
-        */
-       if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
-               zone_set_flag(zone, ZONE_CONGESTED);
-
        free_hot_cold_page_list(&free_pages, 1);
 
        list_splice(&ret_pages, page_list);
        count_vm_events(PGACTIVATE, pgactivate);
        mem_cgroup_uncharge_end();
+       *ret_nr_dirty += nr_dirty;
+       *ret_nr_congested += nr_congested;
        *ret_nr_unqueued_dirty += nr_unqueued_dirty;
        *ret_nr_writeback += nr_writeback;
        *ret_nr_immediate += nr_immediate;
@@ -1045,7 +1040,7 @@ unsigned long reclaim_clean_pages_from_list(struct zone *zone,
                .priority = DEF_PRIORITY,
                .may_unmap = 1,
        };
-       unsigned long ret, dummy1, dummy2, dummy3;
+       unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
        struct page *page, *next;
        LIST_HEAD(clean_pages);
 
@@ -1057,8 +1052,8 @@ unsigned long reclaim_clean_pages_from_list(struct zone *zone,
        }
 
        ret = shrink_page_list(&clean_pages, zone, &sc,
-                               TTU_UNMAP|TTU_IGNORE_ACCESS,
-                               &dummy1, &dummy2, &dummy3, true);
+                       TTU_UNMAP|TTU_IGNORE_ACCESS,
+                       &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
        list_splice(&clean_pages, page_list);
        __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
        return ret;
@@ -1352,6 +1347,8 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
        unsigned long nr_scanned;
        unsigned long nr_reclaimed = 0;
        unsigned long nr_taken;
+       unsigned long nr_dirty = 0;
+       unsigned long nr_congested = 0;
        unsigned long nr_unqueued_dirty = 0;
        unsigned long nr_writeback = 0;
        unsigned long nr_immediate = 0;
@@ -1396,8 +1393,9 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
                return 0;
 
        nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
-                       &nr_unqueued_dirty, &nr_writeback, &nr_immediate,
-                       false);
+                               &nr_dirty, &nr_unqueued_dirty, &nr_congested,
+                               &nr_writeback, &nr_immediate,
+                               false);
 
        spin_lock_irq(&zone->lru_lock);
 
@@ -1431,7 +1429,7 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
         * same way balance_dirty_pages() manages.
         *
         * This scales the number of dirty pages that must be under writeback
-        * before throttling depending on priority. It is a simple backoff
+        * before a zone gets flagged ZONE_WRITEBACK. It is a simple backoff
         * function that has the most effect in the range DEF_PRIORITY to
         * DEF_PRIORITY-2 which is the priority reclaim is considered to be
         * in trouble and reclaim is considered to be in trouble.
@@ -1442,18 +1440,27 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
         * ...
         * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
         *                     isolated page is PageWriteback
+        *
+        * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
+        * of pages under pages flagged for immediate reclaim and stall if any
+        * are encountered in the nr_immediate check below.
         */
        if (nr_writeback && nr_writeback >=
-                       (nr_taken >> (DEF_PRIORITY - sc->priority))) {
+                       (nr_taken >> (DEF_PRIORITY - sc->priority)))
                zone_set_flag(zone, ZONE_WRITEBACK);
-               wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
-       }
 
        /*
         * memcg will stall in page writeback so only consider forcibly
         * stalling for global reclaim
         */
        if (global_reclaim(sc)) {
+               /*
+                * Tag a zone as congested if all the dirty pages scanned were
+                * backed by a congested BDI and wait_iff_congested will stall.
+                */
+               if (nr_dirty && nr_dirty == nr_congested)
+                       zone_set_flag(zone, ZONE_CONGESTED);
+
                /*
                 * If dirty pages are scanned that are not queued for IO, it
                 * implies that flushers are not keeping up. In this case, flag
@@ -1474,6 +1481,14 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
                        congestion_wait(BLK_RW_ASYNC, HZ/10);
        }
 
+       /*
+        * Stall direct reclaim for IO completions if underlying BDIs or zone
+        * is congested. Allow kswapd to continue until it starts encountering
+        * unqueued dirty pages or cycling through the LRU too quickly.
+        */
+       if (!sc->hibernation_mode && !current_is_kswapd())
+               wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
+
        trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
                zone_idx(zone),
                nr_scanned, nr_reclaimed,
@@ -2374,17 +2389,6 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
                                                WB_REASON_TRY_TO_FREE_PAGES);
                        sc->may_writepage = 1;
                }
-
-               /* Take a nap, wait for some writeback to complete */
-               if (!sc->hibernation_mode && sc->nr_scanned &&
-                   sc->priority < DEF_PRIORITY - 2) {
-                       struct zone *preferred_zone;
-
-                       first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
-                                               &cpuset_current_mems_allowed,
-                                               &preferred_zone);
-                       wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
-               }
        } while (--sc->priority >= 0);
 
 out: