Btrfs: fix file corruption after snapshotting due to mix of buffered/DIO writes

When we are mixing buffered writes with direct IO writes against the same
file and snapshotting is happening concurrently, we can end up with a
corrupt file content in the snapshot. Example:

1) Inode/file is empty.

2) Snapshotting starts.

2) Buffered write at offset 0 length 256Kb. This updates the i_size of the
   inode to 256Kb, disk_i_size remains zero. This happens after the task
   doing the snapshot flushes all existing delalloc.

3) DIO write at offset 256Kb length 768Kb. Once the ordered extent
   completes it sets the inode's disk_i_size to 1Mb (256Kb + 768Kb) and
   updates the inode item in the fs tree with a size of 1Mb (which is
   the value of disk_i_size).

4) The dealloc for the range [0, 256Kb[ did not start yet.

5) The transaction used in the DIO ordered extent completion, which updated
   the inode item, is committed by the snapshotting task.

6) Snapshot creation completes.

7) Dealloc for the range [0, 256Kb[ is flushed.

After that when reading the file from the snapshot we always get zeroes for
the range [0, 256Kb[, the file has a size of 1Mb and the data written by
the direct IO write is found. From an application's point of view this is
a corruption, since in the source subvolume it could never read a version
of the file that included the data from the direct IO write without the
data from the buffered write included as well. In the snapshot's tree,
file extent items are missing for the range [0, 256Kb[.

The issue, obviously, does not happen when using the -o flushoncommit
mount option.

Fix this by flushing delalloc for all the roots that are about to be
snapshotted when committing a transaction. This guarantees total ordering
when updating the disk_i_size of an inode since the flush for dealloc is
done when a transaction is in the TRANS_STATE_COMMIT_START state and wait
is done once no more external writers exist. This is similar to what we
do when using the flushoncommit mount option, but we do it only if the
transaction has snapshots to create and only for the roots of the
subvolumes to be snapshotted. The bulk of the dealloc is flushed in the
snapshot creation ioctl, so the flush work we do inside the transaction
is minimized.

This issue, involving buffered and direct IO writes with snapshotting, is
often triggered by fstest btrfs/078, and got reported by fsck when not
using the NO_HOLES features, for example:

  $ cat results/btrfs/078.full
  (...)
  _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent
  *** fsck.btrfs output ***
  [1/7] checking root items
  [2/7] checking extents
  [3/7] checking free space cache
  [4/7] checking fs roots
  root 258 inode 264 errors 100, file extent discount
  Found file extent holes:
        start: 524288, len: 65536
  ERROR: errors found in fs roots

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
1 file changed