[regression] 3.0-rc boot failure -- bisected to cd4ea6ae3982

[Peter Zijlstra]

On Fri, 2011-07-15 at 10:45 +1000, Anton Blanchard wrote:
> Hi,
> 
> > Urgh.. so those spans are generated by sched_domain_node_span(), and
> > it looks like that simply picks the 15 nearest nodes to the one we've
> > got without consideration for overlap with previously generated spans.
> 
> I do wonder if we need this extra level at all on ppc64. From memory
> SGI added it for their massive setups, but our largest setup is 32 nodes
> and breaking that down into 16 node chunks seems overkill.
> 
> I just realised we were setting NEWIDLE on our node definition and that
> was causing large amounts of rebalance work even with
> SD_NODES_PER_DOMAIN=16.
> 
> After removing it and bumping SD_NODES_PER_DOMAIN to 32, things look
> pretty good.
> 
> Perhaps we should allow an arch to override SD_NODES_PER_DOMAIN so this
> extra level is only used by SGI boxes.

We can certainly remove the whole topology layer that causes this
problem for 3.0 and try to fix up for 3.1 again.

But I was rather hoping to introduce more of those layers in the near
future, I was hoping to create a layer per node_distance() value, such
that the load-balancing is aware of the interconnects.

Now for that I ran into the exact same problem, and at the time didn't
come up with a solution, but I think I now see a way out.

Something like the below ought to avoid the problem.. makes SGI sad
though :-)

---
 kernel/sched.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/kernel/sched.c b/kernel/sched.c
index 8fb4245..877b9f1 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -7203,7 +7203,7 @@ static struct sched_domain_topology_level default_topology[] = {
 #endif
 	{ sd_init_CPU, cpu_cpu_mask, },
 #ifdef CONFIG_NUMA
-	{ sd_init_NODE, cpu_node_mask, },
+//	{ sd_init_NODE, cpu_node_mask, },
 	{ sd_init_ALLNODES, cpu_allnodes_mask, },
 #endif
 	{ NULL, },