Btrfs: fix unexpected EEXIST from btrfs_get_extent
authorLiu Bo <bo.li.liu@oracle.com>
Fri, 5 Jan 2018 19:51:09 +0000 (12:51 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 26 Apr 2018 09:02:09 +0000 (11:02 +0200)
commitebe064401f078ae68f2532833ad8f212b6ee2be2
treed935b7678d72f06ca9291f603f097fd992d8cf1b
parentc231cec825a9d72642b712b6b62bb9c99b7407fd
Btrfs: fix unexpected EEXIST from btrfs_get_extent

[ Upstream commit 18e83ac75bfe67009c4ddcdd581bba8eb16f4030 ]

This fixes a corner case that is caused by a race of dio write vs dio
read/write.

Here is how the race could happen.

Suppose that no extent map has been loaded into memory yet.
There is a file extent [0, 32K), two jobs are running concurrently
against it, t1 is doing dio write to [8K, 32K) and t2 is doing dio
read from [0, 4K) or [4K, 8K).

t1 goes ahead of t2 and splits em [0, 32K) to em [0K, 8K) and [8K 32K).

------------------------------------------------------
             t1                                t2
      btrfs_get_blocks_direct()         btrfs_get_blocks_direct()
       -> btrfs_get_extent()              -> btrfs_get_extent()
           -> lookup_extent_mapping()
           -> add_extent_mapping()            -> lookup_extent_mapping()
              # load [0, 32K)
       -> btrfs_new_extent_direct()
           -> btrfs_drop_extent_cache()
              # split [0, 32K) and
      # drop [8K, 32K)
           -> add_extent_mapping()
              # add [8K, 32K)
                                              -> add_extent_mapping()
                                                 # handle -EEXIST when adding
                                                 # [0, 32K)
------------------------------------------------------
About how t2(dio read/write) runs into -EEXIST:

a) add_extent_mapping() gets -EEXIST for adding em [0, 32k),

b) search_extent_mapping() then returns [0, 8k) as the existing em,
   even though start == existing->start, em is [0, 32k) so that
   extent_map_end(em) > extent_map_end(existing), i.e. 32k > 8k,

c) then it goes thru merge_extent_mapping() which tries to add a [8k, 8k)
   (with a length 0) and returns -EEXIST as [8k, 32k) is already in tree,

d) so btrfs_get_extent() ends up returning -EEXIST to dio read/write,
   which is confusing applications.

Here I conclude all the possible situations,
1) start < existing->start

            +-----------+em+-----------+
+--prev---+ |     +-------------+      |
|         | |     |             |      |
+---------+ +     +---+existing++      ++
                +
                |
                +
             start

2) start == existing->start

      +------------em------------+
      |     +-------------+      |
      |     |             |      |
      +     +----existing-+      +
            |
            |
            +
         start

3) start > existing->start && start < (existing->start + existing->len)

      +------------em------------+
      |     +-------------+      |
      |     |             |      |
      +     +----existing-+      +
               |
               |
               +
             start

4) start >= (existing->start + existing->len)

+-----------+em+-----------+
|     +-------------+      | +--next---+
|     |             |      | |         |
+     +---+existing++      + +---------+
                      +
                      |
                      +
                   start

As we can see, it turns out that if start is within existing em (front
inclusive), then the existing em should be returned as is, otherwise,
we try our best to merge candidate em with sibling ems to form a
larger em (in order to reduce the total number of em).

Reported-by: David Vallender <david.vallender@landmark.co.uk>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/btrfs/inode.c