Re: perf related lockdep bug

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Nov 04, 2015 at 11:28:00AM +0100, Peter Zijlstra wrote:
> On Wed, Nov 04, 2015 at 11:21:51AM +0100, Peter Zijlstra wrote:
> 
> > The problem appears to be due to the new RCU expedited grace period
> > stuff, with rcu_read_unlock() now randomly trying to acquire locks it
> > previously didn't.
> > 
> > Lemme go look at those rcu bits again..
> 
> Paul, I think this is because of:
> 
>   8203d6d0ee78 ("rcu: Use single-stage IPI algorithm for RCU expedited grace period")
> 
> What happens is that the IPI comes in and tags any random
> rcu_read_unlock() with the special bit, which then goes on and takes
> locks.
> 
> Now the problem is that we have scheduler activity inside this lock;
> the one reported lockdep seems easy enough to fix, see below.
> 
> I'll got and see if there's more sites than can cause this.

*sigh* yes, there's gobs more. In fact the very first one I looked at:

rcu_dump_cpu_stacks()
  raw_spin_lock_irqsave(&rnp->lock, flags);
  dump_cpu_task()
    pr_info()

That too will end up doing wakeups..

idem:
 - print_other_cpu_stall()
 - rcu_print_details_task_stall_rnp()


So just like you had to pull out all the rcu_gp_kthead_wake() calls from
under rnp->lock, so too must we pull out all implied wakeups, which very
much include printk().