Ingo explains:

The condition is 'impossible', but the whole balancing code is
(intentionally) a bit racy:

                cpus_and(tmp, group->cpumask, cpu_online_map);
                if (!cpus_weight(tmp))
                        goto next_group;

                for_each_cpu_mask(i, tmp) {
                        if (!idle_cpu(i))
                                goto next_group;
                        push_cpu = i;
                }

                rq = cpu_rq(push_cpu);
                double_lock_balance(busiest, rq);
                move_tasks(rq, push_cpu, busiest, 1, sd, IDLE);

in the for_each_cpu_mask() loop we specifically check for each CPU in
the target group to be idle - so push_cpu's runqueue == busiest [==
current runqueue] cannot be true because the current CPU is not idle, we
are running in the migration thread ... But this is not a real problem,
load-balancing we do in a racy way to reduce overhead [and it's all
statistics anyway so absolute accuracy is impossible], and active
balancing itself is somewhat racy due to the migration-thread wakeup
(and the active_balance flag) going outside the runqueue locks [for
similar reasons].

so it all looks quite plausible - the normal SMP boxes dont trigger it,
but Bjorn's 128-CPU setup with a non-trivial domain hiearachy triggers
it.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>