On Mon, 2007-07-30 at 11:22 +0200, Ingo Molnar wrote:

* Daniel Walker [off-list ref] wrote:

quoted

+	/*
+	 * If our rcu_read_lock_nesting went negative, likely
+	 * something is wrong..
+	 */
+	WARN_ON(current->rcu_read_lock_nesting < 0);

have you actually caught any rcu locking problem this way? Double 
unlocks should be caught by lockdep already, at a higher level.

in any case i've added a slightly different form of this change to the 
-rt queue that will also check for counter overflows. But i'm not sure 
we want to litter the code with trivial checks like this, so i'm keeping 
it separate and if it does not trigger anything real i'll remove it.

I haven't caught anything with it, but this code would have made it much
easier to catch the single rcu unlock in sys_sched_yield() which was
silent in PREEMPT_RT, and hung !PREEMPT_RT ..

It's fine with me, if you have another method.

Daniel