Re: UDF madness

[Christoph Hellwig <hch@lst.de>]

> Yes, me too.  generic_shutdown_super() takes lock_super().  And udf uses
> lock_super for protecting its block allocation data strutures.  Trivial
> deadlock on unmount.
> 
> Filesystems really shouldn't be using lock_super() for internal purposes,
> and the main filesystems have been taught to not do that any more, but UDF
> is a holdout.
> 
> It seems that this deadlock was introduced on Jan 5 by the "udf: fix
> reservation discarding" patch which added the udf_discard_prealloc() call
> into udf_clear_inode().  The below dopey patch prevents the deadlock, but
> perhaps we can think of something more appealing.  Ideally, use a new lock
> altogether?

Yes, the lock_super usage in UDF only protects it's block allocation and
doesn't try to cross-protect anything with the VFS usage of it.

I'll cook up a patch to add a balloc_mutex to the UDF superblock and
give it some testing.