Re: [PATCH 3/7] memcg: immigrate charges only when a threadgroup leader is moved

[Oleg Nesterov <oleg-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On 05/22, Michal Hocko wrote:
>
> On Thu 21-05-15 21:27:16, Oleg Nesterov wrote:
> > On 05/20, Michal Hocko wrote:
> > >
> > > On Wed 20-05-15 19:53:02, Oleg Nesterov wrote:
> > > >
> > > > Yes, yes, the group leader can't go away until the whole thread-group dies.
> > >
> > > OK, then we should have a guarantee that mm->owner is always thread
> > > group leader, right?
> >
> > No, please note that the exiting leader does exit_mm()->mm_update_next_owner()
> > and this changes mm->owner.
>
> I am confused now. Yeah it changes the owner but the new one will be
> again the thread group leader, right?

Why?

In the likely case (if CLONE_VM without CLONE_THREAD was not used) the
last for_each_process() in mm_update_next_owner() will find another thread
from the same group.

Oh. I think mm_update_next_owner() needs some cleanups. Perhaps I'll send
the patch today.

> > Btw, this connects to other potential cleanups... task_struct->mm looks
> > a bit strange, we probably want to move it into signal_struct->mm and
> > make exit_mm/etc per-process. But this is not trivial, and off-topic.
>
> I am not sure this is a good idea but I would have to think about this
> some more. Let's not distract this email thread and discuss it in a
> separate thread please.

Yes, yes. I mentioned this to explain that we can't keep the exited leader
as mm->owner in any case. And report that get_mem_cgroup_from_mm() looks
racy.

Oleg.