[Devel] [RFC rh7] ve: cgroups -- Allow to attach non-self into ve cgroups

[Kirill Tkhai]

В Вт, 16/06/2015 в 19:36 +0300, Cyrill Gorcunov пишет:
> On Tue, Jun 16, 2015 at 07:25:48PM +0300, Kirill Tkhai wrote:
> > > 
> > > May not we simply add into ve cgroup?
> > > 
> > > .can_attach
> > > 	...
> > > 	spin_lock(&task->sighand->siglock);
> > > .cance_attach
> > > 	...
> > > 	spin_unlock(&task->sighand->siglock);
> > > .attach
> > > 	...
> > > 	spin_unlock(&task->sighand->siglock);
> > 
> > It seems sighand lock is not good for this, because cgroup_attach_task()
> > iterating over all subsys:
> > 
> > for_each_subsys(root, ss) {
> >                 if (ss->can_attach) {
> > 
> > so we may bumped into wrong lock order in one of them. Also this solves
> > problem №1, but does not solve problem №2.
> 
> We gonna move only one task in container start time so I think
> this is not critical in timing.
> 
> As to #2 -- yes, I think rcu-readlock with sync should do the trick.
> Also I need to check in details what Vladimir suggested, maybe this
> will work even better?

Maybe it better, but it allows a situation, when parent's task_ve is
!ve0, and child's is ve0 for a some time. I'm afraid if there are hidden
places, which may work wrong because of this.

If we have any problems because of this, the solution is good.
 
> > Hm. How about stop_machine? It solves all of the problems. Also, it shouldn't
> > worsen performance, because this action is rare.
> 
> iirc stop-machine is a big hammer :/