Re: [PATCH] cgroup: Don't expose dead tasks in cgroup

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2026-03-03 10:22:38 [-1000], Tejun Heo wrote:
> So, I think we can fix this in the iterator without moving the unlink.
> css_task_iter_advance() already skips dying leaders w/ no live threads
> but only on the dying_tasks list, which gets populated too late. We can
> extend it to catch PF_EXITING tasks on the regular tasks list too:
> 
>   if ((task->flags & PF_EXITING) &&
>       !atomic_read(&task->signal->live))
>           goto repeat;
> 
> PF_EXITING is set in exit_signals() which is before exit_notify(), so
> by the time the parent wakes up, the flag is already set. The
> signal->live check keeps zombie leaders with live threads visible. I
> can't see anything that would break by this being checked earlier than
> cgroup_task_exit() - everything between PF_EXITING and
> cgroup_task_exit() is just teardown (mm, files, etc.) and it should
> close the race window.
> 
> Haven't tested this yet. What do you think?

So this

--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -5108,6 +5108,8 @@ static void css_task_iter_advance(struct css_task_iter *it)
 		return;
 
 	task = list_entry(it->task_pos, struct task_struct, cg_list);
+	if ((task->flags & PF_EXITING) && !atomic_read(&task->signal->live))
+		goto repeat;
 
 	if (it->flags & CSS_TASK_ITER_PROCS) {
 		/* if PROCS, skip over tasks which aren't group leaders */

does work.
So we delay the removal due to sched_ext and then hide due to userspace.
Nice ;)
The signal check is to see the zombies, so you they pop up in the list
until a waitpid()?
Anyway, do you want me make a proper patch out of it?

> Thanks.
> 

Sebastian