releases/2.6.38.8/mm-vmscan-correctly-check-if-reclaimer-should-schedule.patch

   1 From f06590bd718ed950c98828e30ef93204028f3210 Mon Sep 17 00:00:00 2001
   2 From: Minchan Kim <minchan.kim@gmail.com>
   3 Date: Tue, 24 May 2011 17:11:11 -0700
   4 Subject: mm: vmscan: correctly check if reclaimer should schedule
   5  during shrink_slab
   6
   7 From: Minchan Kim <minchan.kim@gmail.com>
   8
   9 commit f06590bd718ed950c98828e30ef93204028f3210 upstream.
  10
  11 It has been reported on some laptops that kswapd is consuming large
  12 amounts of CPU and not being scheduled when SLUB is enabled during large
  13 amounts of file copying.  It is expected that this is due to kswapd
  14 missing every cond_resched() point because;
  15
  16 shrink_page_list() calls cond_resched() if inactive pages were isolated
  17         which in turn may not happen if all_unreclaimable is set in
  18         shrink_zones(). If for whatver reason, all_unreclaimable is
  19         set on all zones, we can miss calling cond_resched().
  20
  21 balance_pgdat() only calls cond_resched if the zones are not
  22         balanced. For a high-order allocation that is balanced, it
  23         checks order-0 again. During that window, order-0 might have
  24         become unbalanced so it loops again for order-0 and returns
  25         that it was reclaiming for order-0 to kswapd(). It can then
  26         find that a caller has rewoken kswapd for a high-order and
  27         re-enters balance_pgdat() without ever calling cond_resched().
  28
  29 shrink_slab only calls cond_resched() if we are reclaiming slab
  30         pages. If there are a large number of direct reclaimers, the
  31         shrinker_rwsem can be contended and prevent kswapd calling
  32         cond_resched().
  33
  34 This patch modifies the shrink_slab() case.  If the semaphore is
  35 contended, the caller will still check cond_resched().  After each
  36 successful call into a shrinker, the check for cond_resched() remains in
  37 case one shrinker is particularly slow.
  38
  39 [mgorman@suse.de: preserve call to cond_resched after each call into shrinker]
  40 Signed-off-by: Mel Gorman <mgorman@suse.de>
  41 Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
  42 Cc: Rik van Riel <riel@redhat.com>
  43 Cc: Johannes Weiner <hannes@cmpxchg.org>
  44 Cc: Wu Fengguang <fengguang.wu@intel.com>
  45 Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
  46 Tested-by: Colin King <colin.king@canonical.com>
  47 Cc: Raghavendra D Prabhu <raghu.prabhu13@gmail.com>
  48 Cc: Jan Kara <jack@suse.cz>
  49 Cc: Chris Mason <chris.mason@oracle.com>
  50 Cc: Christoph Lameter <cl@linux.com>
  51 Cc: Pekka Enberg <penberg@kernel.org>
  52 Cc: Rik van Riel <riel@redhat.com>
  53 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  54 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
  55 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
  56
  57 ---
  58  mm/vmscan.c |    9 +++++++--
  59  1 file changed, 7 insertions(+), 2 deletions(-)
  60
  61 --- a/mm/vmscan.c
  62 +++ b/mm/vmscan.c
  63 @@ -230,8 +230,11 @@ unsigned long shrink_slab(unsigned long
  64         if (scanned == 0)
  65                 scanned = SWAP_CLUSTER_MAX;
  66
  67 -       if (!down_read_trylock(&shrinker_rwsem))
  68 -               return 1;       /* Assume we'll be able to shrink next time */
  69 +       if (!down_read_trylock(&shrinker_rwsem)) {
  70 +               /* Assume we'll be able to shrink next time */
  71 +               ret = 1;
  72 +               goto out;
  73 +       }
  74
  75         list_for_each_entry(shrinker, &shrinker_list, list) {
  76                 unsigned long long delta;
  77 @@ -282,6 +285,8 @@ unsigned long shrink_slab(unsigned long
  78                 shrinker->nr += total_scan;
  79         }
  80         up_read(&shrinker_rwsem);
  81 +out:
  82 +       cond_resched();
  83         return ret;
  84  }
  85