Re: 3.14.0+/x86: lockdep and mutexes not getting along

[Jason Low <jason.low2@hp.com>]

On Thu, 2014-04-10 at 11:18 +0200, Peter Zijlstra wrote:
> On Wed, Apr 09, 2014 at 10:42:59PM -0700, Jason Low wrote:
> > As a starting point, would either of you like to test the following
> > patch to see if it fixes the issue? This patch essentially generates the
> > same code as in older kernels in the debug case. This applies on top of
> > kernels with both commits 6f008e72cd11 and 1d8fe7dc8078.
> 
> 
> So I managed to reproduce, and the below makes it go away. I just don't
> understand why though. will stare more.

So one thing I noticed that is different in the current code is that in
debug_mutex_unlock(), there is is a possibility that it does not unlock
the mutex (when !debug_locks). May be interesting to try out this
patch too:

-----
diff --git a/kernel/locking/mutex-debug.c b/kernel/locking/mutex-debug.c
index e1191c9..97f8df0 100644
--- a/kernel/locking/mutex-debug.c
+++ b/kernel/locking/mutex-debug.c
@@ -72,7 +72,7 @@ void mutex_remove_waiter(struct mutex *lock, struct mutex_waiter *waiter,
 void debug_mutex_unlock(struct mutex *lock)
 {
 	if (unlikely(!debug_locks))
-		return;
+		goto out;
 
 	DEBUG_LOCKS_WARN_ON(lock->magic != lock);
 
@@ -84,6 +84,7 @@ void debug_mutex_unlock(struct mutex *lock)
 	DEBUG_LOCKS_WARN_ON(!lock->wait_list.prev && !lock->wait_list.next);
 	mutex_clear_owner(lock);
 
+out:
 	/*
 	 * __mutex_slowpath_needs_to_unlock() is explicitly 0 for debug
 	 * mutexes so that we can do it here after we've verified state.