Re: setns() && PR_SET_CHILD_SUBREAPER

[Oleg Nesterov <oleg@redhat.com>]

On 01/24, Eric W. Biederman wrote:
>
> Oleg Nesterov <oleg@redhat.com> writes:
>
> > Suppose we have a process P in the root namespace and another namespace X.
> >
> > P does setns() and enters the X namespace.
> > P forks a child C.
> >
> > C forks a grandchild G.
> > C exits.
> >
> > The question is, where should we reparent the grandchild G? In the normal
> > case it will be reparented to X->child_reaper and this looks correct.
> >
> > But lets suppose that P runs with the ->has_child_subreaper bit set. In
> > this case it will be reparented to P's sub-reaper or a global init, and
> > given that P can't control its ->has_child_subreaper flag this does not
> > look right to me.
> >
> > I can make a simple patch but perhaps I missed something or we actually
> > want this (imo strange) behaviour?
>
> We definitely do not want a child to be repareted out of a pid namespace
> when the pid namespace has a perfectly fine child_reaper.
>
> The special case for the init_task in find_new_reaper appears to be the
> instance of this problem that was considered in the code.

Actually we should blame the same_thread_group(reaper, child_reaper) check,
it should had ensured we could not cross the namespaces, but it is not
enough. Because this logic predates setns().

> Semantically what we want to do is walk up the parents in the process
> tree.  If a parent has is_child_subreaper we stop at it.  If the
> transition from one parent to the next we are switching pid namespaces
> we want the reaper from the pid namespace.

Yes, this is what I have in mind, see the patch below. I need to re-check
it and update the comment to explain why we can't simply check child_reaper
as we currently do.

This way we can start the search from father->real_parent, but the comment
above the "reaper == &init_task" is no longer correct, we always need this
check although perhaps is_idle_task(reaper) would be better.

> As I recall has_child_subreaper was just supposed to be an optimization
> so the common case would not have to walk up the process tree when
> finding it's parent.

Yep.

> If we retain any optimizations such as has_child_subreaper please
> consider the case where a process with is_child_subreaper set exits,
> and what happens to it's children.

Yes, in this case it should not have any effect. Well, there is another
corner case, perhaps we should turn

		if (!reaper->signal->is_child_subreaper)
			continue;

into
		if (!reaper->signal->is_child_subreaper) {
			if (!reaper->signal->has_child_subreaper)
				break;
			continue;
		}

this looks a bit more correct if the exited "is_child_subreaper" process
was forked, and after that its parent called prctl(SET_CHILD_SUBREAPER).
But I think we do not care and Pavel is going to eliminate the case when
a child of is_child_subreaper task can run without has_child_subreaper
flag set.

So what do you think about the patch below?

Oleg.

--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -569,15 +569,15 @@ static struct task_struct *find_new_reaper(struct task_struct *father,
 		return thread;
 
 	if (father->signal->has_child_subreaper) {
+		unsigned int level = task_pid(father)->level;
 		/*
 		 * Find the first ->is_child_subreaper ancestor in our pid_ns.
-		 * We start from father to ensure we can not look into another
-		 * namespace, this is safe because all its threads are dead.
+		 * We check pid->level, this is slightly more efficient than
+		 * task_active_pid_ns(reaper) != task_active_pid_ns(father).
 		 */
-		for (reaper = father;
-		     !same_thread_group(reaper, child_reaper);
+		for (reaper = father->real_parent;
+		     task_pid(reaper)->level == level;
 		     reaper = reaper->real_parent) {
-			/* call_usermodehelper() descendants need this check */
 			if (reaper == &init_task)
 				break;
 			if (!reaper->signal->is_child_subreaper)