When we have a buffered write that starts at an offset greater than or
equals to the file's size happening concurrently with a full ranged
fiemap, we can end up leaking an extent state structure.

Suppose we have a file with a size of 1Mb, and before the buffered write
and fiemap are performed, it has a single extent state in its io tree
representing the range from 0 to 1Mb, with the EXTENT_DELALLOC bit set.

The following sequence diagram shows how the memory leak happens if a
fiemap a buffered write, starting at offset 1Mb and with a length of
4Kb, are performed concurrently.

          CPU 1                                                  CPU 2

  extent_fiemap()
    --> it's a full ranged fiemap
        range from 0 to LLONG_MAX - 1
        (9223372036854775807)

    --> locks range in the inode's
        io tree
      --> after this we have 2 extent
          states in the io tree:
          --> 1 for range [0, 1Mb[ with
              the bits EXTENT_LOCKED and
              EXTENT_DELALLOC_BITS set
          --> 1 for the range
              [1Mb, LLONG_MAX[ with
              the EXTENT_LOCKED bit set

                                                  --> start buffered write at offset
                                                      1Mb with a length of 4Kb

                                                  btrfs_file_write_iter()

                                                    btrfs_buffered_write()
                                                      --> cached_state is NULL

                                                      lock_and_cleanup_extent_if_need()
                                                        --> returns 0 and does not lock
                                                            range because it starts
                                                            at current i_size / eof

                                                      --> cached_state remains NULL

                                                      btrfs_dirty_pages()
                                                        btrfs_set_extent_delalloc()
                                                          (...)
                                                          __set_extent_bit()

                                                            --> splits extent state for range
                                                                [1Mb, LLONG_MAX[ and now we
                                                                have 2 extent states:

                                                                --> one for the range
                                                                    [1Mb, 1Mb + 4Kb[ with
                                                                    EXTENT_LOCKED set
                                                                --> another one for the range
                                                                    [1Mb + 4Kb, LLONG_MAX[ with
                                                                    EXTENT_LOCKED set as well

                                                            --> sets EXTENT_DELALLOC on the
                                                                extent state for the range
                                                                [1Mb, 1Mb + 4Kb[
                                                            --> caches extent state
                                                                [1Mb, 1Mb + 4Kb[ into
                                                                @cached_state because it has
                                                                the bit EXTENT_LOCKED set

                                                    --> btrfs_buffered_write() ends up
                                                        with a non-NULL cached_state and
                                                        never calls anything to release its
                                                        reference on it, resulting in a
                                                        memory leak

Fix this by calling free_extent_state() on cached_state if the range was
not locked by lock_and_cleanup_extent_if_need().

The same issue can happen if anything else other than fiemap locks a range
that covers eof and beyond.

This could be triggered, sporadically, by test case generic/561 from the
fstests suite, which makes duperemove run concurrently with fsstress, and
duperemove does plenty of calls to fiemap. When CONFIG_BTRFS_DEBUG is set
the leak is reported in dmesg/syslog when removing the btrfs module with
a message like the following:

  [77100.039461] BTRFS: state leak: start 6574080 end 6582271 state 16402 in tree 0 refs 1

Otherwise (CONFIG_BTRFS_DEBUG not set) detectable with kmemleak.

CC: stable@vger.kernel.org # 4.16+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>