[PATCH 1/2] powerpc/numa: use cached value of update->cpu in update_cpu_topology

There isn't any need to keep referring to update->cpu, as we've already
checked cpu == update->cpu at this point.

Signed-off-by: Nishanth Aravamudan <redacted>

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 649666d5d1c2..86fdb004ad2f 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -1491,8 +1491,8 @@ static int update_cpu_topology(void *data)
 		if (cpu != update->cpu)
 			continue;
 
-		unmap_cpu_from_node(update->cpu);
-		map_cpu_to_node(update->cpu, update->new_nid);
+		unmap_cpu_from_node(cpu);
+		map_cpu_to_node(cpu, update->new_nid);
 		vdso_getcpu_init();
 	}
 

We received a report of warning in kernel/sched/core.c where the sched
group was NULL on an LPAR after a topology update. This seems to occur
because after the topology update has moved the CPUs, cpu_to_node is
returning the old value still, which ends up breaking the consistency of
the NUMA topology in the per-cpu maps. Ensure that we update the per-cpu
fields when we re-map CPUs.
    
Signed-off-by: Nishanth Aravamudan <redacted>

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 86fdb004ad2f..048be62dc979 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -1488,11 +1488,14 @@ static int update_cpu_topology(void *data)
 	cpu = smp_processor_id();
 
 	for (update = data; update; update = update->next) {
+		int new_nid = update->new_nid;
 		if (cpu != update->cpu)
 			continue;
 
 		unmap_cpu_from_node(cpu);
-		map_cpu_to_node(cpu, update->new_nid);
+		map_cpu_to_node(cpu, new_nid);
+		set_cpu_numa_node(cpu, new_nid);
+		set_cpu_numa_mem(cpu, local_memory_node(new_nid));
 		vdso_getcpu_init();
 	}
 

On Sat, 2014-18-10 at 00:50:40 UTC, Nishanth Aravamudan wrote:

We received a report of warning in kernel/sched/core.c where the sched
group was NULL on an LPAR after a topology update. This seems to occur
because after the topology update has moved the CPUs, cpu_to_node is
returning the old value still, which ends up breaking the consistency of
the NUMA topology in the per-cpu maps. Ensure that we update the per-cpu
fields when we re-map CPUs.

This looks like a bug fix, I assume you want it to go in for 3.18 ?

cheers

Hi Michael,

On 21.10.2014 [15:36:27 +1100], Michael Ellerman wrote:

On Sat, 2014-18-10 at 00:50:40 UTC, Nishanth Aravamudan wrote:

quoted

We received a report of warning in kernel/sched/core.c where the sched
group was NULL on an LPAR after a topology update. This seems to occur
because after the topology update has moved the CPUs, cpu_to_node is
returning the old value still, which ends up breaking the consistency of
the NUMA topology in the per-cpu maps. Ensure that we update the per-cpu
fields when we re-map CPUs.

This looks like a bug fix, I assume you want it to go in for 3.18 ?

Yes, please!

Thanks,
Nish