Ceph » rbd

While trying to reproduce a null pointer messenger problem,
I kept hitting messages like this after some (fairly random)
number of iterations:

[ 885.866666] =====================================
[ 885.866667] [ BUG: bad unlock balance detected! ]
[ 885.866669] 3.5.0-rc1-ceph-00037-g07accf4 #1 Not tainted
[ 885.866670] -------------------------------------
[ 885.866673] flush-250:0/9378 is trying to release lock (&(&wb->list_lock)->r:
[ 885.866679] [<ffffffff811ac309>] wb_writeback+0x199/0x2d0
[ 885.866679] but there are no more locks to release!

(full dump below)

This appears to be a writeback-related problem, but I think we
need to investigate to see whether the problem has something to
do with RBD.

I have been running xfstests over RBD, and test 049 is the one
that has been reliable in reproducing problems. This may
complicate understanding this issue a bit, because this test
runs some tests on various combinations of one file system
mounted on a loop device backed by a file on another file
system, which is then backed by an RBD device. And the
loop-mounted and underlying file systems are ext2 or xfs.