Re: NUMA topology question wrt. d4edc5b6

[Nishanth Aravamudan <nacc@linux.vnet.ibm.com>]

On 09.06.2014 [14:38:26 -0700], David Rientjes wrote:
> On Fri, 23 May 2014, Srivatsa S. Bhat wrote:
> 
> > diff --git a/arch/powerpc/include/asm/topology.h b/arch/powerpc/include/asm/topology.h
> > index c920215..58e6469 100644
> > --- a/arch/powerpc/include/asm/topology.h
> > +++ b/arch/powerpc/include/asm/topology.h
> > @@ -18,6 +18,7 @@ struct device_node;
> >   */
> >  #define RECLAIM_DISTANCE 10
> >  
> > +#include <linux/nodemask.h>
> >  #include <asm/mmzone.h>
> >  
> >  static inline int cpu_to_node(int cpu)
> > @@ -30,7 +31,7 @@ static inline int cpu_to_node(int cpu)
> >  	 * During early boot, the numa-cpu lookup table might not have been
> >  	 * setup for all CPUs yet. In such cases, default to node 0.
> >  	 */
> > -	return (nid < 0) ? 0 : nid;
> > +	return (nid < 0) ? first_online_node : nid;
> >  }
> >  
> >  #define parent_node(node)	(node)
> 
> I wonder what would happen on ppc if we just returned NUMA_NO_NODE here 
> for cpus that have not been mapped (they shouldn't even be possible).  

Well, with my patch (Ben sent it to Linus in the last pull request, I
think), powerpc uses the generic per-cpu stuff, so this function is
gone. Dunno if it makes sense to initialize the per-cpu data to
NUMA_NO_NODE (rather than 0?).

For powerpc, it's a timing thing. We can call cpu_to_node() quite early,
and we may not have set up the mapping information yet.

> This would at least allow callers that do
> kmalloc_node(..., cpu_to_node(cpu)) to be allocated on the local cpu 
> rather than on a perhaps offline or remote node 0.
> 
> It would seem better to catch callers that do 
> cpu_to_node(<not-possible-cpu>) rather than blindly return an online node.

Agreed, but I've not seen such a case.

Thanks,
Nish