Re: [PATCH UPDATED 03/10] threadgroup: extend threadgroup_lock() to cover exit and exec

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Dec 05, 2011 at 10:43:15AM -0800, Tejun Heo wrote:
> Hello, Frederic.
> 
> On Fri, Dec 02, 2011 at 05:28:03PM +0100, Frederic Weisbecker wrote:
> > But I don't think it's very useful to protect against irq_exit_thread(),
> > what happens there is purely of internal irq interest.
> > 
> > Then right after, PF_EXITING is set before any interesting change.
> > Isn't it possible to simply lock this flag setting? IIRC, as soon
> > as the PF_EXITING flag is set, you ignore the task for attachment.
> 
> I think that's technically possible but it does introduce another
> class of tasks - the dying ones.  e.g. If a task has PF_EXITING set
> and the containing process is migrating, we'll have to migrate all
> tasks but the dying one and cgroup ->exit callbacks can be called on
> the lonely task after the migration is complete.  It's kinda messy and
> if someone makes a wrong assumption there, the bug is gonna be even
> more difficult to reproduce / track down than now.  Yes, smaller scope
> locking is nicer but I would like to avoid api weirdities like that.

I don't understand what you mean. On your patches, you only process tasks
that don't have PF_EXITING, ie: you don't include these in the flex array
on cgroup_attach_proc(). So that still applies in my proposal.

>From the exit path we would have:

	exit_signal() {
		lock_threadgroup_change(task);
		task->flags |= PF_EXITING;
		lock_threadgroup(task);
	}

	exit all the rest: mm, etc...

Then from cgroup_attach_proc():

	lock_threadgroup(task);
	for_each_thread(task) {
		if (!(task->flags & PF_EXITING))
			include in flex array
	}

Am I forgetting something?