Re: [PATCH 21/32] union-mount: Make lookup work for union-mounted file systems

From: Miklos Szeredi <miklos@szeredi.hu>
To: vaurora@redhat.com
Cc: miklos@szeredi.hu, jblunck@suse.de, linux-kernel@vger.kernel.org,
	linux-fsdevel@vger.kernel.org, viro@zeniv.linux.org.uk,
	bharata@in.ibm.com, dwmw2@infradead.org, mszeredi@suse.cz
Subject: Re: [PATCH 21/32] union-mount: Make lookup work for union-mounted file systems
Date: Wed, 20 May 2009 12:21:10 +0200	[thread overview]
Message-ID: <E1M6iv8-0003CB-4O@pomaz-ex.szeredi.hu> (raw)
In-Reply-To: <20090519173038.GC722@shell> (message from Valerie Aurora on Tue, 19 May 2009 13:30:38 -0400)

On Tue, 19 May 2009, Valerie Aurora wrote:
> On Tue, May 19, 2009 at 06:15:52PM +0200, Miklos Szeredi wrote:

> > Worse, it looks like there are still i_mutex lock ordering issues
> > (__hash_lookup_topmost()/__hash_lookup_build_union()).  What happens
> > if two separate unions of two filesystems are built where the order of
> > branches is reversed?
> 
> We have a similar problem in union_copyup_dir().  Hm, thinking about
> this, only one of the file systems can actually change while we are
> doing work.  That might help us get out of the lock ordering problems.
> Thoughts?

Right, we talked about this with Jan, and came basically to the same
conclusion.  The lookup on the lower branches needs to be separate
from the atomic lookup/create on the top branch.  Which means some
restructuring in the callers...

I'm not sure how all this could be simplified, I don't yet even
understand what all these different lookup functions are meant to do
(header comments might help).

Thanks,
Miklos