Re: Warning in worker_enter_idle()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, May 07, 2012 at 12:40:42PM -0700, Tejun Heo wrote:
> Hello, Paul.
> 
> On Sun, May 06, 2012 at 08:38:14AM -0700, Paul E. McKenney wrote:
> > Hello!
> > 
> > The worker_enter_idle() is complaining that there all workers are idle,
> > but that there is work remaining:
> > 
> > 	/* sanity check nr_running */
> > 	WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
> > 		     atomic_read(get_gcwq_nr_running(gcwq->cpu)));
> > 
> > This is running on Power, .config attached.  I must confess that I don't
> > see any sort of synchronization or memory barriers that would keep the
> > counts straight on a weakly ordered system.  Or is there some clever
> > design constraint that prevents worker_enter_idle() from accessing other
> > CPUs' gcwq_nr_running variables?
> 
> Workers are tied to global cpu workqueues (gcwqs).  There's one gcwq
> per cpu and one unbound one, so yeah, workers access these counters
> under gcwq->lock.  Atomic accesses to nr_running is depended on only
> while nr_idle is adjusted under gcwq->lock, so there shouldn't be a
> discrepancy there.  Can you reproduce the problem?  What was going on
> the system?  Was CPU being brought up or down?

I was running rcutorture with CPU hotplug operations.  It has happened
a couple of times on the .config that I attached, but never under any
of the other 13 .configs that I run.

							Thanx, Paul