Re: sched/autogroup: race if !sysctl_sched_autogroup_enabled ?

From: Oleg Nesterov
Date: Thu Nov 10 2016 - 08:10:07 EST


On 11/09, Peter Zijlstra wrote:
>
> On Wed, Nov 09, 2016 at 05:59:33PM +0100, Oleg Nesterov wrote:
>
> > We need to ensure that autogroup/tg returned by autogroup_task_group()
> > can't go away if we race with autogroup_move_group(), and unless the
> > caller holds ->siglock we rely on fact that autogroup_move_group()
> > will a) see this task and b) do sched_move_task() which needs the same
> > same rq->lock.
> >
> > However. autogroup_move_group() skips for_each_thread/sched_move_task
> > if sysctl_sched_autogroup_enabled == 0.
> >
> > So. Doesn't this mean that cgroup migration to the root cgroup can race
> > with autogroup_move_group() and use the soon-to-be-freed autogroup->tg?
>
> Argh, its too late for this, also jet-lag. But maybe, I can sort of feel
> a hole here but cannot for the life of me still think.

And the 3rd case which I didn't think about yesterday. And now I really hope
it can explain the vmcore we have.

If sysctl_sched_autogroup_enabled was enabled and then disabled, it is
possible that the "autogrouped" process runs with ag->kref.refcount == 1,
and if it does setsid() it frees its active task_group.

> > although this is a bit off-topic. Another question is that I fail to
> > understand why sched_autogroup_create_attach() does autogroup_create()
> > and changes signal->autogroup even if !sysctl_sched_autogroup_enabled.
>
> I really cannot remember back that far, but it could be to allow
> flipping it back on.

Yes, I thought about this too, but I think it is hardly possible to explain
what do we actually want when sysctl_sched_autogroup_enabled changes from 0
to 1.

So I am going to send the patch which simply moves the sysctl check from
autogroup_move_group() to sched_autogroup_create_attach(), but perhaps I
should split this change?

I mean, the first patch for -stable could just remove the current check,
the 2nd one will add it into sched_autogroup_create_attach().

Oleg.