Re: [PATCH] kthread: Enhance kthread_stop to abort interruptible sleeps

[ebiederm@xmission.com (Eric W. Biederman)]

Oleg Nesterov <oleg@tv-sign.ru> writes:

> On 04/24, Oleg Nesterov wrote:
>>
>> Hm, mm_release() clears ->vfork_done before complete().

Duh.  Yes somehow I had a blind spot there.  I clearly
need to handle that case.

>> 	mm_release:
>> 
>> 		struct completion *vfork_done = tsk->vfork_done;
>> 
>> 		if (vfork_done) {
>> 			tsk->vfork_done = NULL;
>> 			complete(vfork_done);
>> 		}
>> 
>> 
>> 	kthread_stop:
>> 
>> 		set_tsk_thread_flag(tsk, TIF_KTHREAD_STOP);
>> 		signal_wake_up(tsk, 1);
>> 
>> 		// tsk exits, sets ->vfork_done == NULL
>> 
>> 		wait_for_completion(tsk->vfork_done);
>
> Since the task_struct should be pinned anyway, I think kthread_stop()
> should do:
>
> 	vfork_done = tsk->vfork_done;
> 	barrier();
> 	if (vfork_done)
> 		wait_for_completion(vfork_done);

That should work, and this may explain what is going on. I was trying
to figure out how this could happen but if the thread is running on
another cpu there is a race and it may exit quickly enough to cause us
problems before we get to kthread_stop.

I don't know if this is the problem but it certainly needs to be fixed.

Eric