[Devel] Re: NFS causing oops when freeing namespace

Oleg Nesterov oleg at tv-sign.ru
Wed Jan 17 11:46:32 PST 2007


On 01/17, Cedric Le Goater wrote:
>
> Oleg Nesterov wrote:
> > On 01/17, Daniel Hokka Zakrisson wrote:
> >> It was the only semi-plausible explanation I could come up with. I added a
> >> printk in do_exit right before exit_task_namespaces, where sighand was
> >> still set, and one right before the spin_lock_irq in lockd_down, where it
> >> had suddenly been set to NULL.
> > 
> > I can't reproduce the problem, but
> 
> I did on a 2.6.20-rc4-mm1.
> 
> > 	do_exit:
> > 		exit_notify(tsk);
> > 		exit_task_namespaces(tsk);
> > 
> > the task could be reaped by its parent in between.
> 
> indeed. while it goes spleeping in lockd_down() just before it does
> 
> 	spin_lock_irq(&current->sighand->siglock);
> 
> current->sighand is valid before interruptible_sleep_on_timeout() and
> not after.
>  
> > We should not use ->signal/->sighand after exit_notify().
> > 
> > Can we move exit_task_namespaces() up?
> 
> yes but I moved it down because it invalidates ->nsproxy ... 

Well, we can fix the symptom if we change lockd_down() to use
lock_task_sighand(), or something like this,

	--- NFS/fs/lockd/svc.c~lockd_down	2006-11-27 21:20:11.000000000 +0300
	+++ NFS/fs/lockd/svc.c	2007-01-17 22:39:47.000000000 +0300
	@@ -314,6 +314,7 @@ void
	 lockd_down(void)
	 {
		static int warned;
	+	int sigpending;
	 
		mutex_lock(&nlmsvc_mutex);
		if (nlmsvc_users) {
	@@ -334,16 +335,15 @@ lockd_down(void)
		 * Wait for the lockd process to exit, but since we're holding
		 * the lockd semaphore, we can't wait around forever ...
		 */
	-	clear_thread_flag(TIF_SIGPENDING);
	+	sigpending = test_and_clear_thread_flag(TIF_SIGPENDING);
		interruptible_sleep_on_timeout(&lockd_exit, HZ);
		if (nlmsvc_pid) {
			printk(KERN_WARNING 
				"lockd_down: lockd failed to exit, clearing pid\n");
			nlmsvc_pid = 0;
		}
	-	spin_lock_irq(&current->sighand->siglock);
	-	recalc_sigpending();
	-	spin_unlock_irq(&current->sighand->siglock);
	+	if (sigpending)	/* can be wrong at this point, harmless */
	+		set_thread_flag(TIF_SIGPENDING);
	 out:
		mutex_unlock(&nlmsvc_mutex);
	 }

but this is not good anyway.

Oleg.

_______________________________________________
Containers mailing list
Containers at lists.osdl.org
https://lists.osdl.org/mailman/listinfo/containers




More information about the Devel mailing list