Re: [PATCH] cgroup: Move cgroup_task_dead() to task_struct clean up

[Tejun Heo <tj@kernel.org>]

Hello,

On Sat, Mar 14, 2026 at 10:17:52AM +0100, Sebastian Andrzej Siewior wrote:
> On 2026-03-13 18:33:14 [+0100], Michal Koutný wrote:
> > Hello.
> > 
> > On Wed, Mar 11, 2026 at 01:08:29PM +0100, Sebastian Andrzej Siewior <bigeasy@linutronix.de> wrote:
> > > @@ -7084,6 +7031,7 @@ void cgroup_task_free(struct task_struct *task)
> > >  {
> > >  	struct css_set *cset = task_css_set(task);
> > >  
> > > +	cgroup_task_dead(task);
> > >  	if (!list_empty(&task->cg_list)) {
> > >  		spin_lock_irq(&css_set_lock);
> > >  		css_set_skip_task_iters(task_css_set(task), task);
> > 
> > Erm, isn't this way too late?
> 
> My understanding is that the point is that the cgroup must not be
> exposed to userland once the task is about to die. This has been moved
> after the fine schedule() which is too late because the parent is
> signaled before that and the removal happens after that.
> So now we hide the tasks in the iterator once the task is PF_EXITING.

Yeah, I think it's too late. That's where we tell the cgroup is becoming
empty among other things. cgroup_task_free() waits for all task refcnts to
drop which can take any amount of time. You can also imagine situations
where e.g. userspace pins a specific task through pidfd unitl the cgroup it
belongs becomes empty, which would have worked before but now would
deadlock.

Thanks.

-- 
tejun