Re: TASK_DEAD && ttwu() again (Was: ensure arch_ptrace/ptrace_request can never race with SIGKILL)

[Tejun Heo <tj@kernel.org>]

Hello, Oleg.

On Wed, Jan 23, 2013 at 08:19:46PM +0100, Oleg Nesterov wrote:
> Let me remind the problem. To oversimplify, we have
> 
> 	try_to_wake_up(task, state)
> 	{
> 		lock(task->pi_lock);
> 
> 		if (task->state & state)
> 			task->state = TASK_RUNNING;
> 
> 		unlock(task->pi_lock);
> 	}
> 
> And this means that a task doing
> 
> 	current->state = STATE_1;
> 	// no schedule() in between
> 	current->state = STATE_2;
> 	schedule();
> 
> can be actually woken up by try_to_wake_up(STATE_1) even if it already
> sleeps in STATE_2.

Hmmm... nasty.

...
> and we have the same problem again. So _I think_ that we we need another
> mb() after unlock_wait() ?

Seems so, or, maybe we should add barrier semantics to unlock_wait()?
As it currently stands, it kinda invites misusages.

> And, afaics, in theory we can't simply move the current mb() down, after
> unlock_wait().  (again, only in theory, if nothing else we should have
> the implicit barrrers after we played with ->state in the past).
> 
> Or perhaps we should modify ttwu_do_wakeup() to not blindly set RUNNING,
> say, cmpxchg(old_state, RUNNING). But this is not simple/nice.

I personally think this is the right thing to do short of requiring
locking on current->state changes.  The situation is a bit muddy
because we're generally requiring sleepers to loop while still having
cases where things don't work that way.  It's a little scary that we
require looping to protect against stray wakeups, which can be very
rare, without any way to verify/test.

The waker would be acquiring the cacheline exclusively one way or the
other, so I don't think doing cmpxchg would add much overhead.  We
would definitely want to do comparisons tho.

Thanks.

-- 
tejun