[RFC v3] sched/topology: fix kernel crash when a CPU is hotplugged in a memoryless node

[Peter Zijlstra]

On Mon, Mar 04, 2019 at 08:59:52PM +0100, Laurent Vivier wrote:
> This happens because initially powerpc code computes
> sched_domains_numa_masks of offline nodes as if they were merged with
> node 0 (because firmware doesn't provide the distance information for
> memoryless/cpuless nodes):
> 
>   node   0   1   2   3
>     0:  10  40  10  10
>     1:  40  10  40  40
>     2:  10  40  10  10
>     3:  10  40  10  10

*groan*... what does it do for things like percpu memory? ISTR the
per-cpu chunks are all allocated early too. Having them all use memory
out of node-0 would seem sub-optimal.

> We should have:
> 
>   node   0   1   2   3
>     0:  10  40  40  40
>     1:  40  10  40  40
>     2:  40  40  10  40
>     3:  40  40  40  10

Can it happen that it introduces a new distance in the table? One that
hasn't been seen before? This example only has 10 and 40, but suppose
the new node lands at distance 20 (or 80); can such a thing happen?

If not; why not?

> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> index 3f35ba1d8fde..24831b86533b 100644
> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -1622,8 +1622,10 @@ void sched_init_numa(void)
>  				return;
>  
>  			sched_domains_numa_masks[i][j] = mask;
> +			if (!node_state(j, N_ONLINE))
> +				continue;
>  
> -			for_each_node(k) {
> +			for_each_online_node(k) {
>  				if (node_distance(j, k) > sched_domains_numa_distance[i])
>  					continue;
>  

So you're relying on sched_domain_numa_masks_set/clear() to fix this up,
but that in turn relies on the sched_domain_numa_levels thing to stay
accurate.

This all seems very fragile and unfortunate.