fs/superblock: unregister sb shrinker before ->kill_sb()
authorDave Chinner <david@fromorbit.com>
Wed, 4 Jun 2014 23:10:46 +0000 (16:10 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 4 Jun 2014 23:54:11 +0000 (16:54 -0700)
commit28f2cd4f6da24a1aa06c226618ed5ad69e13df64
treeb8546f8ba98e3a8c9e546ea58046daeb44a4bb4e
parent850e9c69ca75f32aa9361a0edec6cad388a231b0
fs/superblock: unregister sb shrinker before ->kill_sb()

This series is aimed at regressions noticed during reclaim activity.  The
first two patches are shrinker patches that were posted ages ago but never
merged for reasons that are unclear to me.  I'm posting them again to see
if there was a reason they were dropped or if they just got lost.  Dave?
Time?  The last patch adjusts proportional reclaim.  Yuanhan Liu, can you
retest the vm scalability test cases on a larger machine?  Hugh, does this
work for you on the memcg test cases?

Based on ext4, I get the following results but unfortunately my larger
test machines are all unavailable so this is based on a relatively small
machine.

postmark
                                  3.15.0-rc5            3.15.0-rc5
                                     vanilla       proportion-v1r4
Ops/sec Transactions         21.00 (  0.00%)       25.00 ( 19.05%)
Ops/sec FilesCreate          39.00 (  0.00%)       45.00 ( 15.38%)
Ops/sec CreateTransact       10.00 (  0.00%)       12.00 ( 20.00%)
Ops/sec FilesDeleted       6202.00 (  0.00%)     6202.00 (  0.00%)
Ops/sec DeleteTransact       11.00 (  0.00%)       12.00 (  9.09%)
Ops/sec DataRead/MB          25.97 (  0.00%)       30.02 ( 15.59%)
Ops/sec DataWrite/MB         49.99 (  0.00%)       57.78 ( 15.58%)

ffsb (mail server simulator)
                                 3.15.0-rc5             3.15.0-rc5
                                    vanilla        proportion-v1r4
Ops/sec readall           9402.63 (  0.00%)      9805.74 (  4.29%)
Ops/sec create            4695.45 (  0.00%)      4781.39 (  1.83%)
Ops/sec delete             173.72 (  0.00%)       177.23 (  2.02%)
Ops/sec Transactions     14271.80 (  0.00%)     14764.37 (  3.45%)
Ops/sec Read                37.00 (  0.00%)        38.50 (  4.05%)
Ops/sec Write               18.20 (  0.00%)        18.50 (  1.65%)

dd of a large file
                                3.15.0-rc5            3.15.0-rc5
                                   vanilla       proportion-v1r4
WallTime DownloadTar       75.00 (  0.00%)       61.00 ( 18.67%)
WallTime DD               423.00 (  0.00%)      401.00 (  5.20%)
WallTime Delete             2.00 (  0.00%)        5.00 (-150.00%)

stutter (times mmap latency during large amounts of IO)

                            3.15.0-rc5            3.15.0-rc5
                               vanilla       proportion-v1r4
Unit >5ms Delays  80252.0000 (  0.00%)  81523.0000 ( -1.58%)
Unit Mmap min         8.2118 (  0.00%)      8.3206 ( -1.33%)
Unit Mmap mean       17.4614 (  0.00%)     17.2868 (  1.00%)
Unit Mmap stddev     24.9059 (  0.00%)     34.6771 (-39.23%)
Unit Mmap max      2811.6433 (  0.00%)   2645.1398 (  5.92%)
Unit Mmap 90%        20.5098 (  0.00%)     18.3105 ( 10.72%)
Unit Mmap 93%        22.9180 (  0.00%)     20.1751 ( 11.97%)
Unit Mmap 95%        25.2114 (  0.00%)     22.4988 ( 10.76%)
Unit Mmap 99%        46.1430 (  0.00%)     43.5952 (  5.52%)
Unit Ideal  Tput     85.2623 (  0.00%)     78.8906 (  7.47%)
Unit Tput min        44.0666 (  0.00%)     43.9609 (  0.24%)
Unit Tput mean       45.5646 (  0.00%)     45.2009 (  0.80%)
Unit Tput stddev      0.9318 (  0.00%)      1.1084 (-18.95%)
Unit Tput max        46.7375 (  0.00%)     46.7539 ( -0.04%)

This patch (of 3):

We will like to unregister the sb shrinker before ->kill_sb().  This will
allow cached objects to be counted without call to grab_super_passive() to
update ref count on sb.  We want to avoid locking during memory
reclamation especially when we are skipping the memory reclaim when we are
out of cached objects.

This is safe because grab_super_passive does a try-lock on the
sb->s_umount now, and so if we are in the unmount process, it won't ever
block.  That means what used to be a deadlock and races we were avoiding
by using grab_super_passive() is now:

        shrinker                        umount

        down_read(shrinker_rwsem)
                                        down_write(sb->s_umount)
                                        shrinker_unregister
                                          down_write(shrinker_rwsem)
                                            <blocks>
        grab_super_passive(sb)
          down_read_trylock(sb->s_umount)
            <fails>
        <shrinker aborts>
        ....
        <shrinkers finish running>
        up_read(shrinker_rwsem)
                                          <unblocks>
                                          <removes shrinker>
                                          up_write(shrinker_rwsem)
                                        ->kill_sb()
                                        ....

So it is safe to deregister the shrinker before ->kill_sb().

Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Tested-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Jan Kara <jack@suse.cz>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/super.c