Re: [PATCH v1] kthread/smpboot: Serialize kthread parking against wakeup

[Peter Zijlstra]

On Tue, May 01, 2018 at 01:20:26PM +0530, Kohli, Gaurav wrote:
> But In our older case, where we have seen failure below is the wake up path
> and ftraces, Wakeup occured  and completed before schedule call only.
> 
> So final state of CPUHP is running not parked. I have also pasted debug
> ftraces that we got during issue reproduction.
> 
> Here wakeup for cpuhp is below:
> 
> takedown_cpu-> kthread_park-> wake_up_process
> 
> 
>  39,034,311,742,395  apps (10240)        Trace Printk cpuhp/0  (16)  [000]
> 39015.625000: <debug> __kthread_parkme state=512 task=ffffffcc7458e680
> flags: 0x5 -> state 5 -> state is parked inside parkme function
> 
> 39,034,311,846,510  apps (10240)        Trace Printk cpuhp/0  (16)  [000]
> 39015.625000: <debug> before schedule __kthread_parkme state=0
> task=ffffffcc7458e680 flags: 0xd  ->  just before schedule call, state is
> running
> 
> tatic void __kthread_parkme(struct kthread *self)
> 
> {
> 
>         __set_current_state(TASK_PARKED);
> 
>         while (test_bit(KTHREAD_SHOULD_PARK, &self->flags)) {
> 
>                 if (!test_and_set_bit(KTHREAD_IS_PARKED, &self->flags))
> 
>                         complete(&self->parked);
> 
>                 schedule();
> 
>                 __set_current_state(TASK_PARKED);
> 
>         }
> 
>         clear_bit(KTHREAD_IS_PARKED, &self->flags);
> 
>         __set_current_state(TASK_RUNNING);
> 
> }
> 
> So my point is here also, if it is reschedule then it can set TASK_PARKED,
> but it seems after takedown_cpu call this thread never get a chance to run,
> So final state is TASK_RUNNING.
> 
> In our current fix also can't we observe same scenario where final state is
> TASK_RUNNING.

I'm not sure I understand your concern. Loosing the TASK_PARKED store
with the above code is obviously bad. But with the loop as proposed I
don't see a problem.

takedown_cpu() can proceed beyond smpboot_park_threads() and kill the
CPU before any of the threads are parked -- per having the complete()
before hitting schedule().

And, afaict, that is harmless. When we go offline, sched_cpu_dying() ->
migrate_tasks() will migrate any still runnable threads off the cpu.
But because at this point the thread must be in the PARKED wait-loop, it
will hit schedule() and go to sleep eventually.

Also note that kthread_unpark() does __kthread_bind() to rebind the
threads.

Aaaah... I think I've spotted a problem there. We clear SHOULD_PARK
before we rebind, so if the thread lost the first PARKED store,
does the completion, gets migrated, cycles through the loop and now
observes !SHOULD_PARK and bails the wait-loop, then __kthread_bind()
will forever wait.

Is that what you had in mind?