On Tue, May 01, 2018 at 04:10:53PM +0530, Kohli, Gaurav wrote:
Yes with loop, it will reset TASK_PARKED but that is not happening in the
dumps we have seen.
But was that with or without the fixed wait-loop? I don't care about
stuff you might have seen with the current code, that is clearly broken.
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.
So during next unpark
__kthread_unpark -> __kthread_bind -> wait_task_inactive (this got failed,
as current state is running so failed on below call:
Aah, yes, I seem to have mis-remembered how wait_task_inactive() works.
And it is indeed still a problem..
Let me ponder what the best solution is, it's a bit of a mess.