Btrfs: fix memory leak due to concurrent append writes with fiemap
authorFilipe Manana <fdmanana@suse.com>
Mon, 30 Sep 2019 09:20:25 +0000 (10:20 +0100)
committerDavid Sterba <dsterba@suse.com>
Tue, 1 Oct 2019 16:40:58 +0000 (18:40 +0200)
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>
fs/btrfs/file.c

index 8fe4eb7..27e5b26 100644 (file)
@@ -1591,7 +1591,6 @@ static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
        struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
        struct btrfs_root *root = BTRFS_I(inode)->root;
        struct page **pages = NULL;
-       struct extent_state *cached_state = NULL;
        struct extent_changeset *data_reserved = NULL;
        u64 release_bytes = 0;
        u64 lockstart;
@@ -1611,6 +1610,7 @@ static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
                return -ENOMEM;
 
        while (iov_iter_count(i) > 0) {
+               struct extent_state *cached_state = NULL;
                size_t offset = offset_in_page(pos);
                size_t sector_offset;
                size_t write_bytes = min(iov_iter_count(i),
@@ -1758,9 +1758,20 @@ again:
                if (copied > 0)
                        ret = btrfs_dirty_pages(inode, pages, dirty_pages,
                                                pos, copied, &cached_state);
+
+               /*
+                * If we have not locked the extent range, because the range's
+                * start offset is >= i_size, we might still have a non-NULL
+                * cached extent state, acquired while marking the extent range
+                * as delalloc through btrfs_dirty_pages(). Therefore free any
+                * possible cached extent state to avoid a memory leak.
+                */
                if (extents_locked)
                        unlock_extent_cached(&BTRFS_I(inode)->io_tree,
                                             lockstart, lockend, &cached_state);
+               else
+                       free_extent_state(cached_state);
+
                btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
                                               true);
                if (ret) {