Kernel 2.6.30.4 XFS(..?) regression (& with/2.6.31-rc6)

[Christoph Hellwig]

Ok, let's see where errors could happen then.

There are four theoretical possibilities:

 (a) XFS
 (b) loop driver
 (c) crypto loop code
 (d) block layer

Or combinations thereof.

I would take (a) and (d) as more unlikely as they tend to get used much
more and I would have heard more bug reports already.

The cryptoloop code hasn't changed at all since 2.6.29.

The loop code howver has a very interesting commit just after 2.6.39:

commit 68db1961bbf4e16c220ccec4a780e966bc1fece3
Author: Nikanth Karthikesan <knikanth at suse.de>
Date:   Tue Mar 24 12:29:54 2009 +0100

    loop: support barrier writes

Can you try reverting this one (it cleanly reverse-applies against
2.6.30 and current mainline) and see if that makes a difference?