Re: Lockdep violation in pcmcia

[Dominik Brodowski <linux@dominikbrodowski.net>]

Hey,

On Mon, Nov 02, 2009 at 04:21:23PM -0500, Alan Stern wrote:
> I've been getting these warnings for a long, long time, and finally 
> decided to report them:

Thanks!

> [ 1893.036023] other info that might help us debug this:
> [ 1893.036023] 2 locks held by cardmgr/1457:

Wow, cardmgr still in use... I had hoped it had already disappeared on
non-embedded systems...

> The cause is easy enough to track down.  In pcmcia_ioctl.c,
> pcmcia_adjust_resource_info() does a down_read() on
> pcmcia_socket_list_rwsem.  While holding the rwsem, one of the pathways
> calls the s->resource_ops->add_mem method.  On my system this method is
> realized by adjust_memory() in rsrc_nonstatic.c, which does its own
> down_read() on the same rwsem -- i.e., a recursive locking attempt.
> 
> The reason lockdep warns about this behavior is that it can lead to
> deadlock in the wrong circumstances, namely, if another thread were to
> do a down_write() in between the two down_read() calls.
> 
> Would it be correct simply to omit the down_read() in adjust_memory()?

No, for there are other code paths reaching adjust_memory() not holding a
(read) lock on this semaphore. As pcmcia_ioctl.c is on its way out anyways:
would you mind keeping it as it is, for a down_write() call is quite
unlikely during the only time cardmgr actually does this call (system
startup)?

Best,
	Dominik