I'm currently working on a node affine scheduler extension for NUMA
machines and the load balancer behaves a bit different from the original.
So after a few boot failures with those slowly booting 16 CPU IA64
machines I thought there must be a simpler solution than synchronizing and
waiting for the load balancer: just let migration_CPU0 do what it is
designed for. So my proposal is:
   - start all migration threads on CPU#0
   - initialize migration_CPU0 (trivial, reliable, as it already is on
     the right CPU)
   - let all other migration threads use set_cpus_allowed() to get to the
     right place

The only synchronization needed is the non-zero migration threads waiting
for migration_CPU0 to start working, which it will, as it is already on
the right CPU. This saves quite some lines of code.

I first posted this to LKML on March 6th (BTW, the fix #1, too) and since
then it was tested on several big NUMA platforms: 16 CPU NEC AzusA (IA64)
(also known as HP rx....), up to 32 CPU SGI IA64, 16 CPU IBM NUMA-Q
(IA32). No more lock-ups at boot since then. So I consider it working.

There is another good reason for this approach: the integration of the CPU
hotplug patch with the new scheduler becomes easier. One just needs to
create the new migration thread, it will move itself to the right CPU
without any additional magic (which you otherwise need because of the
synchronizations which won't be there at hotplug). Kimi Suganuma in the
neighboring cube is fiddling this out currently.