Re: [PATCH 1/2] workqueue: rescuer_thread() processes all pwqs before exit

From: Tejun Heo <tj@kernel.org>
To: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 1/2] workqueue: rescuer_thread() processes all pwqs before exit
Date: Thu, 17 Apr 2014 12:27:20 -0400	[thread overview]
Message-ID: <20140417162720.GQ15326@htj.dyndns.org> (raw)
In-Reply-To: <CACvQF53DOj0+sn5EQKLA+mCBdtuceO29gBmH2HCqCOXBmRMWAQ@mail.gmail.com>

On Fri, Apr 18, 2014 at 12:21:51AM +0800, Lai Jiangshan wrote:
> > Hmmm?  Isn't this the same race condition that you tried to remove by
> > relocating the test?  It doesn't matter what destroy_workqueue() does,
> > the rescuer may get preempted inbetween and anything can happen
> > inbetween including someone maydaying and initiation of
> > destroy_workqueue().  Your patch doesn't change the situation at all.
> > It can still return with non-empty mayday list.
> 
> You are right. We need a additional atomic check.

I don't think increasing the number of checks will do anything.  We
just need to keep the test where it is now and exit after doing the
loop.  ie.

	set TASK_INTRRUPTIBLE
	test should_stop and cache the result

	spin_lock_irq();

	while () {
		process the list
	}

	spin_unlock_irq();

	if (was should_stop asserted?)
		exit;

	schedule();

Should do, no?

-- 
tejun