XFS supports and typically allows concurrent asynchronous direct I/O
submission to a single file. One exception to the rule is that file
extending dio writes that start beyond the current EOF (e.g.,
potentially create a hole at EOF) require exclusive I/O access to the
file. This is because such writes must zero any pre-existing blocks
beyond EOF that are exposed by virtue of now residing within EOF as a
result of the write about to be submitted.

Before EOF zeroing can occur, the current file i_size must be stabilized
to avoid data corruption. In this scenario, XFS upgrades the iolock to
exclude any further I/O submission, waits on in-flight I/O to complete
to ensure i_size is up to date (i_size is updated on dio write
completion) and restarts the various checks against the state of the
file. The problem is that this protection sequence is triggered only
when the iolock is currently held shared. While this is true for async
dio in most cases, the caller may upgrade the lock in advance based on
arbitrary circumstances with respect to EOF zeroing. For example, the
iolock is always acquired exclusively if the start offset is not block
aligned. This means that even though the iolock is already held
exclusive for such I/Os, pending I/O is not drained and thus EOF zeroing
can occur based on an unstable i_size.

This problem has been reproduced as guest data corruption in virtual
machines with file-backed qcow2 virtual disks hosted on an XFS
filesystem. The virtual disks must be configured with aio=native mode
and the must not be truncated out to the maximum file size (as some virt
managers will do).

Update xfs_file_aio_write_checks() to unconditionally drain in-flight
dio before EOF zeroing can occur. Rather than trigger the wait based on
iolock state, use a new flag and upgrade the iolock when necessary. Note
that this results in a full restart of the inode checks even when the
iolock was already held exclusive when technically it is only required
to recheck i_size. This should be a rare enough occurrence that it is
preferable to keep the code simple rather than create an alternate
restart jump target.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>