tree ed34df6a3fe3d67d749594f4e9de0fe9c720ab17
parent 8b8b08cbfb9021af4b54b4175fc4c51d655aac8c
author Filipe Manana <fdmanana@suse.com> 1465910307 +0100
committer Filipe Manana <fdmanana@suse.com> 1470032473 +0100

Btrfs: add missing check for writeback errors on fsync

When we start an fsync we start ordered extents for all delalloc ranges.
However before attempting to log the inode, we only wait for those ordered
extents if we are not doing a full sync (bit BTRFS_INODE_NEEDS_FULL_SYNC
is set in the inode's flags). This means that if an ordered extent
completes with an IO error before we check if we can skip logging the
inode, we will not catch and report the IO error to user space. This is
because on an IO error, when the ordered extent completes we do not
update the inode, so if the inode was not previously updated by the
current transaction we end up not logging it through calls to fsync and
therefore not check its mapping flags for the presence of IO errors.

Fix this by checking for errors in the flags of the inode's mapping when
we notice we can skip logging the inode.

This caused sporadic failures in the test generic/331 (which explicitly
tests for IO errors during an fsync call).

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>