Re: [PATCH 1/3] stop_machine: ensure that a queued callback will be called before cpu_stop_park()

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 14, 2015 at 09:03:56PM +0200, Oleg Nesterov wrote:
> On 10/14, Peter Zijlstra wrote:
> >
> > On Thu, Oct 08, 2015 at 04:51:31PM +0200, Oleg Nesterov wrote:
> > > cpu_stop_queue_work() checks stopper->enabled before it queues the
> > > work, but ->enabled == T can only guarantee cpu_stop_signal_done()
> > > if we race with cpu_down().
> > >
> > > This is not enough for stop_two_cpus() or stop_machine(), they will
> > > deadlock if multi_cpu_stop() won't be called by one of the target
> > > CPU's. stop_machine/stop_cpus are fine, they rely on stop_cpus_mutex.
> > > But stop_two_cpus() has to check cpu_active() to avoid the same race
> > > with hotplug, and this check is very unobvious and probably not even
> > > correct if we race with cpu_up().
> > >
> > > Change cpu_down() pass to clear ->enabled before cpu_stopper_thread()
> > > flushes the pending ->works and returns with KTHREAD_SHOULD_PARK set.
> > >
> > > Note also that smpboot_thread_call() calls cpu_stop_unpark() which
> > > sets enabled == T at CPU_ONLINE stage, so this CPU can't go away until
> > > cpu_stopper_thread() is called at least once. This all means that if
> > > cpu_stop_queue_work() succeeds, we know that work->fn() will be called.
> >
> > This hard relies on the fact that cpu_down uses stop machine, right?
> 
> Not really.
> 
> > IIRC part of the hotplug rework Thomas is doing is geared towards
> > breaking away from stop machine. There is nothing fundamental about
> > hot-unplug that requires stop machine.
> 
> cpu_down() should park/kill/whatever the percpu stopper thread anyway.
> And this path should clear ->enabled, it can also flush the pending
> works.

So the proposed patch does: ->enabled=false; park();, which can race
with if (->enabled) wake();

smpboot_thread_fn() will not call ->thread_fn() when should_park(), and
thus any pending work will not get flushed.

It only works now because the stopper task calls park(), which means
cpu_stopper_thread() will flush, but that very much relies on the
stopper thread calling park in itself.

Or I'm just terminally confused.. which is entirely possible.