On Fri, 2023-03-24 at 10:46 +0100, Thomas Gleixner wrote:
> 
> It's even worse. Any topology relevant change _must_ be serialized by
> holding cpu_hotplug_lock write locked. So this needs fundamentally more
> thought

Yes. Yes, it does. Which is why it isn't being done in parallel in this
patch series. The topology update happens from smp_callin() which
happens from do_wait_cpu_initialized(), which is still in the
CPUHP_BRINGUP_CPU stage for now, being done one at a time.

When we come to look at the next stage, doing do_wait_cpu_initialized()
in parallel for multiple APs too, we will be happy to hear specifics of
why this "must" be serialized by the cpu_hotplug_lock and no other. It
will be great to check that we haven't missed anything.

But we aren't there yet.