2001-10-29 01:01:46

by Robert Kuebel

Subject: [PATCH] 8139too reparent_to_init() race

hello all,

lately i noticed this message during boot-up (when the network
interfaces were being configured) ...

"task `ifconfig' exit_signal 17 in reparent_to_init"

this happens only about 1/2 of the time.

after some digging this is what i found...
sometimes ifconfig's parent exits before ifconfig reaches
rtl8139_thread(). when this happens, ifconfig's exit_signal is set to
SIGCHLD (in forget_original_parent), because its new parent is init.
then rlt8139_thread() is reached it calls reparent_to_init(), which
complains that exit_signal is already non-zero.

basically this patch stops rtl8139_thread() from calling
reparent_to_init() when its parent is already init.

is this the right way to fix the problem?
should reparent_to_init() check that the parent is not already init?

as a budding kernel hacker i would appreciate any comment on this
change.

please, cc me on replies. i can only handle the digest form of this
list.

thanks.
rob.

--- linux-2.4.13/drivers/net/8139too.orig.c Sun Oct 28 18:16:48 2001
+++ linux-2.4.13/drivers/net/8139too.c Sun Oct 28 18:18:47 2001
@@ -1654,7 +1654,8 @@
unsigned long timeout;

daemonize ();
- reparent_to_init();
+ if (current->p_opptr != child_reaper)
+ reparent_to_init();
spin_lock_irq(&current->sigmask_lock);
sigemptyset(&current->blocked);
recalc_sigpending(current);

2001-10-29 01:29:02

by Andrew Morton

Subject: Re: [PATCH] 8139too reparent_to_init() race

Robert Kuebel wrote:
>
> hello all,
>
> lately i noticed this message during boot-up (when the network
> interfaces were being configured) ...
>
> "task `ifconfig' exit_signal 17 in reparent_to_init"
>
> this happens only about 1/2 of the time.
>
> after some digging this is what i found...
> sometimes ifconfig's parent exits before ifconfig reaches
> rtl8139_thread(). when this happens, ifconfig's exit_signal is set to
> SIGCHLD (in forget_original_parent), because its new parent is init.
> then rlt8139_thread() is reached it calls reparent_to_init(), which
> complains that exit_signal is already non-zero.
>
> basically this patch stops rtl8139_thread() from calling
> reparent_to_init() when its parent is already init.
>

Thanks - that's a useful analysis.

The check in reparent_to_init() was to warn about the situation
where someone had deliberately set the exit signal to some
non-zero value and then the child calls reparent_to_init() - it's
telling you that we're about to stomp on your chosen exit signal.
I hadn't thought about the forget_original_parent() case.

So the fix should be to change the debug code in reparent_to_init()
so it doesn't complain if the exit signal is already SIGCHLD. Or
just kill it off altogether.


--- linux-2.4.14-pre3/kernel/sched.c Tue Oct 23 23:09:48 2001
+++ linux-akpm/kernel/sched.c Sun Oct 28 17:23:26 2001
@@ -1250,11 +1250,6 @@ void reparent_to_init(void)
SET_LINKS(this_task);

/* Set the exit signal to SIGCHLD so we signal init on exit */
- if (this_task->exit_signal != 0) {
- printk(KERN_ERR "task `%s' exit_signal %d in "
- __FUNCTION__ "\n",
- this_task->comm, this_task->exit_signal);
- }
this_task->exit_signal = SIGCHLD;

/* We also take the runqueue_lock while altering task fields

-