Re: [RFC,PATCH 2/2] change __wake_up_parent() to use filtered wakeup

[Oleg Nesterov <oleg@redhat.com>]

On 06/24, Roland McGrath wrote:
>
> do_notify_parent_cldstop:
>
> 	if (task_ptrace(tsk))
> 		parent = tsk->parent;
> 	else {
> 		tsk = tsk->group_leader;
> 		parent = tsk->real_parent;
> 	}
> 	...
> 	__wake_up_parent(tsk, parent);
>
> In the "else" case, parent is not necessarily tsk->parent.  That is, if
> an untraced thread calls do_notify_parent_cldstop() but its group_leader
> is ptrace_reparented().  Then waitee->group_leader->real_parent is who
> gets the wakeup, but __wake_up_parent->child_wait_wakeup would check
> only waiter == waitee->group_leader->parent.

... and in this case we do not wake up ->group_leader->real_parent.

But this is fine? It doesn't make sense to wake up, wait_consider_task()
will notice task_ptrace() and do nothing?

I really need to think with a fresh head, but it seems to me we could add
BUG_ON(p->parent != parent) into wait_consider_task() after "if (ptrace...)"
check.

And this "proves" your check in child_wait_callback() is correct, with
__WNOTHREAD do_wait_thread(parent) is always called with parent ==
sleeper, the caller of do_wait().

No?


Btw, this reminds me that wait_consider_task() doesn't need the "parent"
argument. I noticed this after adding wait_opts, but forgot to send a
patch.

Oleg.