Re: [PATCH tip/core/rcu 11/15] rcu: Break more call_rcu() deadlock involving scheduler and perf

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

On Mon, Sep 01, 2014 at 01:29:34PM +0200, Peter Zijlstra wrote:
> On Thu, Aug 28, 2014 at 11:26:30AM -0700, Paul E. McKenney wrote:
> 
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index b6acb9340192..a1af86099a67 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -2121,16 +2121,23 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
> >  			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
> >  					    TPS("WakeEmpty"));
> >  		} else {
> > -			rdp->nocb_defer_wakeup = true;
> > +			rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
> >  			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
> >  					    TPS("WakeEmptyIsDeferred"));
> >  		}
> >  		rdp->qlen_last_fqs_check = 0;
> >  	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
> >  		/* ... or if many callbacks queued. */
> > -		wake_nocb_leader(rdp, true);
> > +		if (!irqs_disabled_flags(flags)) {
> > +			wake_nocb_leader(rdp, true);
> > +			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
> > +					    TPS("WakeOvf"));
> > +		} else {
> > +			rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
> > +			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
> > +					    TPS("WakeOvfIsDeferred"));
> > +		}
> >  		rdp->qlen_last_fqs_check = LONG_MAX / 2;
> > -		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
> >  	} else {
> >  		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
> >  	}
> 
> Is it possible for the RCU_NOCP_WAKE write to overwrite a WAKE_FORCE ?
> If not, why not? (Would make a good comment thereabouts).

Good point.

So RCU_NOGP_WAKE only happens on the empty-to-non-empty transition,
and that RCU_NOGP_WAKE_FORCE only happens once the queue has a large
number of entries (10,000 by default).  Because this code runs with irqs
disabled (and must because you can do call_rcu() from interrupt handlers),
RCU_NOGP_WAKE_FORCE could overwrite RCU_NOGP_WAKE, but not vice versa.

However, the same effect can be obtained by having the setting of
RCU_NOGP_WAKE_FORCE race with the call to do_nocb_deferred_wakeup().
However, this would require that the call to do_nocb_deferred_wakeup()
was delayed long enough for 9,999 additional callbacks be registered,
and the next 10,000 callbacks will do another RCU_NOGP_WAKE_FORCE.
In addition, the RCU_NOGP_WAKE_FORCE shortens delays in the rcuo kthreads
rather than waking a rcuo kthread that would otherwise sleep indefinitely.
So I didn't see the point of closing this window.

							Thanx, Paul