On Thu, Nov 10, 2011 at 02:35:17PM -0600, Alex Elder wrote:
> Until the perag structures for the filesystem have been set up,
> initialization of the mount point only assigns computed values and
> has no other side-effects (such as allocating additional
> structures). So up to that point, returning a null pointer to
> signal an error is adequate. Once the perag initialization is done
> there needs to be some teardown in case of an error.
> Here the handling of errors is inconsistent. If early perag
> initialization fails, the code currently just exits. Then, if
> getting a reference to the root inode results in an error, a null
> pointer is returned but without first cleaning up the perag
> structures. Next, if rtmount_inodes() returns an error, the
> reference to the root inode (if any) is released, but again the
> perag structures are not cleaned up. Finally, when the perag data
> is read in, if an error occurs, a null pointer is returned but the
> root inode pointer reference is not released and the perag
> structures are not cleaned up.
> Remedy all of that by having each of these error cases jump to error
> handling code at the end of the function. That code needs to
> release the reference to the root inode and release all of the perag
> structures. This (plus a few other things that will be no-ops at
> this point in the mount process) is exactly what libxfs_umount(), so
> just use that function to implement this cleanup activity.
> Signed-off-by: Alex Elder <aelder@xxxxxxx>
Did you make sure libxfs_umount doesn't do too much work? I'm usually
more of a fan of uninding each operation individually.