Re: [PATCH] autofs4 deadlock during expire - kernel 2.6

[Mike Waychison]

Ian Kent wrote:
> On Wed, 24 Sep 2003, Arjan van de Ven wrote:
>
>
> > On Wed, 2003-09-24 at 15:01, Ian Kent wrote:
> >
> > > This is a corrected patch for the autofs4 daedlock problem I posted about 
> > > @@ -206,6 +207,11 @@
> > >
> > > 		interruptible_sleep_on(&wq->queue);
> > >
> > > +		if (waitqueue_active(&wq->queue) && current != wq->owner) {
> > > +			set_current_state(TASK_INTERRUPTIBLE);
> > > +			schedule_timeout(wq->wait_ctr * (HZ/10));
> > > +		}
> > > +
> >
> > this really really looks like you're trying to pamper over a bug by
> > changing the timing somewhere instead of fixing it...
>
>
> Agreed.
>
>
> > also are you sure the deadlock isn't because of the racey use of
> > interruptible_sleep_on ?

I think the deadlock itself needs to be properly identified.

Could you explain where the deadlock is actually occuring?  I briefed 
over the automount 4 code as well as autofs4 and I don't see the 
deadlock.  The 'owner' in the case of an expiry will be a child process 
of the daemon, within a call to ioctl(EXPIRE_MULTI), correct?  Having it 
be released from the waitqueue first should not affect flow of execution 
and released from deadlock.

I don't see how having it wake up before before any other racing 
processes solves anything.

I think Arjan is right in that the race is do to the nautilus process 
entering the sleep_on after the a call to wake_up(&wq->queue).  I don't 
know if a change to using a workqueue is best..   how about refactoring 
that chunk of code to use wait_event_interruptible on the queue, which 
should be clear of any waitqueue/sleep_on races.

> OK so maybe I should have suggestions instead of comments.
>
> Please elaborate.

How about you try out this quick patch I threw together.

Mike Waychison
===== waitq.c 1.6 vs edited =====
--- 1.6/fs/autofs4/waitq.c	Fri Feb  7 12:25:20 2003
+++ edited/waitq.c	Wed Sep 24 15:48:30 2003
@@ -204,7 +204,7 @@
 		recalc_sigpending();
 		spin_unlock_irqrestore(&current->sighand->siglock, irqflags);
 
-		interruptible_sleep_on(&wq->queue);
+		wait_event_interruptible(wq->queue, wq->name == NULL);
 
 		spin_lock_irqsave(&current->sighand->siglock, irqflags);
 		current->blocked = oldset;