Mike,

I think there is a bug either in this ia64 patch, or in the related
generic arch patch: Convert cpu_sibling_map to be a per cpu variable
(v3).

It dies early in boot on me, on the SGI internal 8 processor IA64
system that you and I know as 'margin'.  The death is a hard hang, due
to a corrupt stack, due to a bogus cpu index.

I haven't tracked it down all the way, but have gotten this far.  If I add
the following patch, I get a panic on the BUG_ON if I have these two patches
in 2.6.23-rc8-mm1, but it boots just fine if I don't have these two patches.

It seems that the "cpu_sibling_map[cpu]" cpumask_t is empty (all zero
bits) with your two patches applied, but has some non-zero bits
otherwise, which leads to 'group' being NR_CPUS instead of a useful CPU
number.  Unfortunately, I have no idea why the "cpu_sibling_map[cpu]"
cpumask_t is empty -- good luck on that part.

The patch that catches this bug earlier is this:

--- 2.6.23-rc8-mm1.orig/kernel/sched.c	2007-09-28 01:42:20.144561024 -0700
+++ 2.6.23-rc8-mm1/kernel/sched.c	2007-09-28 02:27:14.239075497 -0700

@@ -5905,6 +5905,7 @@ static int cpu_to_phys_group(int cpu, co
 #else
 	group = cpu;
 #endif
+	BUG_ON(group == NR_CPUS);
 	if (sg)
 		*sg = &per_cpu(sched_group_phys, group);
 	return group;