Re: [PATCH, RFC] v7 scalable classic RCU implementation

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

On Fri, Oct 17, 2008 at 02:04:52PM +0530, Gautham R Shenoy wrote:
> On Fri, Oct 10, 2008 at 09:09:30AM -0700, Paul E. McKenney wrote:
> > +static void __cpuinit rcu_online_cpu(int cpu)
> > +{
> > +#ifdef CONFIG_NO_HZ
> > +	struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
> > +
> > +	rdtp->dynticks_nesting = 1;
> > +	rdtp->dynticks |= 1; 	/* need consecutive #s even for hotplug. */
> > +	rdtp->dynticks_nmi = (rdtp->dynticks + 1) & ~0x1;
>
> =>	rdtp->dynticks is odd. Hence rdtp->dynticks + 1 should be even.
> 	Why is the additional & ~0x1 ?

Because this line should instead be:

	rdtp->dynticks_nmi = (rdtp->dynticks_nmi + 1) & ~0x1;

Well spotted, even if it did take me a good long time to figure out
that this really was a bug in my code!  ;-)

That said, you would have to really work to exercise this one...  Near as
I can tell, you would need to wrap the ->dynticks counter, which would
then cause the dynticks_nmi counter to appear to go backwards.  And then
you would have to prevent the newly onlined CPU from ever passing through
a quiescent state, which would cause a failure in any case.

Still, good to fix, even if I can't figure out how it would result in
a failure.  Real hardware and software tends to be -much- better than me
at finding such failures!

							Thanx, Paul