Re: 3.0-git15 Atomic scheduling in pidmap_init

[Josh Boyer]

On Sun, Aug 14, 2011 at 04:04:53PM -0700, Paul E. McKenney wrote:
> > > > > Now rcu_init_percpu_data() still sets rdp->qs_pending to 1, and that
> > > > > is going to stay as is as long as preemption is disabled.
> > > > 
> > > > But setting rdp->qs_pending to 1 in rcu_init_percpu_data() has no effect
> > > > until a grace period starts.  So, if grace periods are prevented from
> > 
> > Er... really?  Because it gets set and __rcu_pending looks at it
> > unconditionally in the case that is calling set_need_resched.  It
> > doesn't check if there is anything about a grace period going on or not.
> 
> Frederic noted the condition that prevents this at boot time, but it
> appears that newly onlined CPUs might send themselves needless resched
> IPIs at runtime if RCU is idle.
> 
> > > > starting, no need to mess with rcu_init_percpu_data().  Especially given
> > > > that rcu_init_percpu_data() is also used at late boot and runtime for
> > > > CPU hotplug.
> > > 
> > > Ok.
> > > 
> > > > 
> > > > So I believe that it is sufficient to change cpu_needs_another_gp()
> > > > to check for boot being far enough along to allow grace periods.
> > > 
> > > Yep, sounds good.
> > 
> > I looked at doing this but got lost as to 1) how it would help the
> > situtation I've reported, and 2) exactly how to do that.
> 
> It would prevent control from reaching that point, and that might
> well be needed for other reasons.  (This bit about RCU needing to
> work differently at boot time is, well, "interesting".)
> 
> > I'd be happy to test, but at the moment the proposed solution is
> > confusing to me.
> 
> Please see the attached.

Fixed it up quickly to apply on top of -rc2 and it seems to solve the
problem nicely.  Thanks for the patch.

josh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/