Re: [RFC] cpu/hotplug: Modify lock status before making cpu hotplug callbacks

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 7 Jun 2017, John Allen wrote:
> A deadlock has been observed during a cpu hot add on powerpc machines.
> 
> The situation is as follows:
> First, in _cpu_up, we take the cpu_hotplug lock and towards the end of _cpu_up,
> we make the cpu hotplug callbacks, one of which is arch_update_cpu_topology.
> For most other architectures, making this call while the parent thread has the
> cpu_hotplug lock seems harmless as most implementations of
> arch_update_cpu_topology appear to be quite simple. However, the powerpc
> implementation is significantly more complex and requires us to make a call to
> stop_machine which in turn attempts to take the cpu_hotplug lock in
> get_online_cpus. We then deadlock as the parent thread is waiting on the child
> to complete and the child is waiting on the parent to release the lock.
> 
> This solution attempts to resolve the issue by incrementing the cpu_hotplug
> refcount and releasing the cpu_hotplug mutex in _cpu_up so that the subsequent
> call to get_online_cpus can take the mutex while other invocations of _cpu_up
> are still prevented from executing as the refcount is non-zero. From my testing,
> this seems to resolve the deadlock, but I'm not sure if there is a reason that
> get_online_cpus *should* be locked out at this point.

This is horrible, really.

We have code pending in -next which reworks the hotplug lock to a per cpu
reader writer semaphore, so this hack is not going to work anymore.

See git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git smp/hotplug

This addresses the issue of calling stomp_machine() from inside a hotplug
locked region, by introducing stop_machine_cpuslocked(). That should solve
your problem nicely.

Thanks,

	tglx