The whole vector stuff is racy versus cpu hotplug and Jins patch
merily addresses a small part of it and by doing that it breaks stuff
as well.
With that patch we move the vector setup after marking the cpu online,
which is wrong because the vector array on that cpu is not up to date
until we call __setup_vector_irq(). Proper patch below.
We still have an issue in the cpu_disable() patch, but I haven't yet
wrapped my head around it completely.
Thanks,
tglx