vmscan: kill hibernation specific reclaim logic and unify it
authorKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Tue, 15 Dec 2009 01:59:12 +0000 (17:59 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 15 Dec 2009 16:53:18 +0000 (08:53 -0800)
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement.  and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).

commit a06fe4d307 said

   Without the patch:
   Freed  14600 pages in  1749 jiffies = 32.61 MB/s (Anomolous!)
   Freed  88563 pages in 14719 jiffies = 23.50 MB/s
   Freed 205734 pages in 32389 jiffies = 24.81 MB/s

   With the patch:
   Freed  68252 pages in   496 jiffies = 537.52 MB/s
   Freed 116464 pages in   569 jiffies = 798.54 MB/s
   Freed 209699 pages in   705 jiffies = 1161.89 MB/s

At that time, their patch was pretty worth.  However, Modern Hardware
trend and recent VM improvement broke its worth.  From several reason, I
think we should remove shrink_all_zones() at all.

detail:

1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
  at no i/o congestion.
  but current shrink_zone() is sane, not slow.

2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
  fine on numa system.
  example)
    System has 4GB memory and each node have 2GB. and hibernate need 1GB.

    optimal)
       steal 500MB from each node.
    shrink_all_zones)
       steal 1GB from node-0.

  Oh, Cache balancing logic was broken. ;)
  Unfortunately, Desktop system moved ahead NUMA at nowadays.
  (Side note, if hibernate require 2GB, shrink_all_zones() never success
   on above machine)

3) if the node has several I/O flighting pages, shrink_all_zones() makes
  pretty bad result.

  schenario) hibernate need 1GB

  1) shrink_all_zones() try to reclaim 1GB from Node-0
  2) but it only reclaimed 990MB
  3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
  4) it reclaimed 990MB

  Oh, well. it reclaimed twice much than required.
  In the other hand, current shrink_zone() has sane baling out logic.
  then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.

4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
  shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.

Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.

side note: Reclaim logic unificication makes two good side effect.
 - Fix recursive reclaim bug on shrink_all_memory().
   it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
 - Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.

Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c

index 7b0d5c784c7e481720173239977860217d7d2f67..63bd521bb2297660a27c5abcedeaec59ba3f68c7 100644 (file)
@@ -58,6 +58,8 @@ struct scan_control {
        /* How many pages shrink_list() should reclaim */
        unsigned long nr_to_reclaim;
 
+       unsigned long hibernation_mode;
+
        /* This context's GFP mask */
        gfp_t gfp_mask;
 
@@ -1796,7 +1798,8 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
                }
 
                /* Take a nap, wait for some writeback to complete */
-               if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
+               if (!sc->hibernation_mode && sc->nr_scanned &&
+                   priority < DEF_PRIORITY - 2)
                        congestion_wait(BLK_RW_ASYNC, HZ/10);
        }
        /* top priority shrink_zones still had more to do? don't OOM, then */
@@ -2336,148 +2339,44 @@ unsigned long zone_reclaimable_pages(struct zone *zone)
 
 #ifdef CONFIG_HIBERNATION
 /*
- * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
- * from LRU lists system-wide, for given pass and priority.
- *
- * For pass > 3 we also try to shrink the LRU lists that contain a few pages
- */
-static void shrink_all_zones(unsigned long nr_pages, int prio,
-                                     int pass, struct scan_control *sc)
-{
-       struct zone *zone;
-       unsigned long nr_reclaimed = 0;
-       struct zone_reclaim_stat *reclaim_stat;
-
-       for_each_populated_zone(zone) {
-               enum lru_list l;
-
-               if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
-                       continue;
-
-               for_each_evictable_lru(l) {
-                       enum zone_stat_item ls = NR_LRU_BASE + l;
-                       unsigned long lru_pages = zone_page_state(zone, ls);
-
-                       /* For pass = 0, we don't shrink the active list */
-                       if (pass == 0 && (l == LRU_ACTIVE_ANON ||
-                                               l == LRU_ACTIVE_FILE))
-                               continue;
-
-                       reclaim_stat = get_reclaim_stat(zone, sc);
-                       reclaim_stat->nr_saved_scan[l] +=
-                                               (lru_pages >> prio) + 1;
-                       if (reclaim_stat->nr_saved_scan[l]
-                                               >= nr_pages || pass > 3) {
-                               unsigned long nr_to_scan;
-
-                               reclaim_stat->nr_saved_scan[l] = 0;
-                               nr_to_scan = min(nr_pages, lru_pages);
-                               nr_reclaimed += shrink_list(l, nr_to_scan, zone,
-                                                               sc, prio);
-                               if (nr_reclaimed >= nr_pages) {
-                                       sc->nr_reclaimed += nr_reclaimed;
-                                       return;
-                               }
-                       }
-               }
-       }
-       sc->nr_reclaimed += nr_reclaimed;
-}
-
-/*
- * Try to free `nr_pages' of memory, system-wide, and return the number of
+ * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
  * freed pages.
  *
  * Rather than trying to age LRUs the aim is to preserve the overall
  * LRU order by reclaiming preferentially
  * inactive > active > active referenced > active mapped
  */
-unsigned long shrink_all_memory(unsigned long nr_pages)
+unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
 {
-       unsigned long lru_pages, nr_slab;
-       int pass;
        struct reclaim_state reclaim_state;
        struct scan_control sc = {
-               .gfp_mask = GFP_KERNEL,
-               .may_unmap = 0,
+               .gfp_mask = GFP_HIGHUSER_MOVABLE,
+               .may_swap = 1,
+               .may_unmap = 1,
                .may_writepage = 1,
+               .swap_cluster_max = SWAP_CLUSTER_MAX,
+               .nr_to_reclaim = nr_to_reclaim,
+               .hibernation_mode = 1,
+               .swappiness = vm_swappiness,
+               .order = 0,
                .isolate_pages = isolate_pages_global,
-               .nr_reclaimed = 0,
        };
+       struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
+       struct task_struct *p = current;
+       unsigned long nr_reclaimed;
 
-       current->reclaim_state = &reclaim_state;
-
-       lru_pages = global_reclaimable_pages();
-       nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
-       /* If slab caches are huge, it's better to hit them first */
-       while (nr_slab >= lru_pages) {
-               reclaim_state.reclaimed_slab = 0;
-               shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
-               if (!reclaim_state.reclaimed_slab)
-                       break;
-
-               sc.nr_reclaimed += reclaim_state.reclaimed_slab;
-               if (sc.nr_reclaimed >= nr_pages)
-                       goto out;
-
-               nr_slab -= reclaim_state.reclaimed_slab;
-       }
-
-       /*
-        * We try to shrink LRUs in 5 passes:
-        * 0 = Reclaim from inactive_list only
-        * 1 = Reclaim from active list but don't reclaim mapped
-        * 2 = 2nd pass of type 1
-        * 3 = Reclaim mapped (normal reclaim)
-        * 4 = 2nd pass of type 3
-        */
-       for (pass = 0; pass < 5; pass++) {
-               int prio;
-
-               /* Force reclaiming mapped pages in the passes #3 and #4 */
-               if (pass > 2)
-                       sc.may_unmap = 1;
-
-               for (prio = DEF_PRIORITY; prio >= 0; prio--) {
-                       unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
-
-                       sc.nr_scanned = 0;
-                       sc.swap_cluster_max = nr_to_scan;
-                       shrink_all_zones(nr_to_scan, prio, pass, &sc);
-                       if (sc.nr_reclaimed >= nr_pages)
-                               goto out;
-
-                       reclaim_state.reclaimed_slab = 0;
-                       shrink_slab(sc.nr_scanned, sc.gfp_mask,
-                                   global_reclaimable_pages());
-                       sc.nr_reclaimed += reclaim_state.reclaimed_slab;
-                       if (sc.nr_reclaimed >= nr_pages)
-                               goto out;
-
-                       if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
-                               congestion_wait(BLK_RW_ASYNC, HZ / 10);
-               }
-       }
-
-       /*
-        * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
-        * something in slab caches
-        */
-       if (!sc.nr_reclaimed) {
-               do {
-                       reclaim_state.reclaimed_slab = 0;
-                       shrink_slab(nr_pages, sc.gfp_mask,
-                                   global_reclaimable_pages());
-                       sc.nr_reclaimed += reclaim_state.reclaimed_slab;
-               } while (sc.nr_reclaimed < nr_pages &&
-                               reclaim_state.reclaimed_slab > 0);
-       }
+       p->flags |= PF_MEMALLOC;
+       lockdep_set_current_reclaim_state(sc.gfp_mask);
+       reclaim_state.reclaimed_slab = 0;
+       p->reclaim_state = &reclaim_state;
 
+       nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
 
-out:
-       current->reclaim_state = NULL;
+       p->reclaim_state = NULL;
+       lockdep_clear_current_reclaim_state();
+       p->flags &= ~PF_MEMALLOC;
 
-       return sc.nr_reclaimed;
+       return nr_reclaimed;
 }
 #endif /* CONFIG_HIBERNATION */