Re: [PATCH] account_group_exec_runtime: fix the racy usage of ->signal

[Oleg Nesterov <oleg@redhat.com>]

On 11/07, Doug Chapman wrote:
> 
> On Fri, 2008-11-07 at 17:21 +0100, Ingo Molnar wrote:
> > * Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > > @@ -351,10 +351,12 @@ static inline void account_group_exec_ru
> > >  					      unsigned long long ns)
> > >  {
> > >  	struct signal_struct *sig;
> > > +	unsigned long flags;
> > >
> > > -	sig = tsk->signal;
> > > -	if (unlikely(!sig))
> > > +	if (unlikely(!lock_task_sighand(tsk, &flags)))
> > >  		return;
> >
> > i think this will lock up: the signal lock must not nest inside the rq 
> > lock, and these accounting functions are called from within the 
> > scheduler.
>
> I can confirm that this does hang on bootup.

Thanks a lot Doug.

If only I could understand what happens. I am running the 2.6.27 kernel
with the patch below just fine.

Ingo, could you please explain?

OK, perhaps we can check ->exit_state... I'll return on Monday.


--- linux-2.6.27/kernel/sched_fair.c~DBG	2008-10-10 00:13:53.000000000 +0200
+++ linux-2.6.27/kernel/sched_fair.c	2008-11-07 19:15:28.000000000 +0100
@@ -484,6 +484,16 @@ __update_curr(struct cfs_rq *cfs_rq, str
 	curr->vruntime += delta_exec_weighted;
 }
 
+static void ttt(struct task_struct *tsk)
+{
+	unsigned long flags;
+
+	if (unlikely(!lock_task_sighand(tsk, &flags)))
+		return;
+
+	unlock_task_sighand(tsk, &flags);
+}
+
 static void update_curr(struct cfs_rq *cfs_rq)
 {
 	struct sched_entity *curr = cfs_rq->curr;
@@ -507,6 +517,7 @@ static void update_curr(struct cfs_rq *c
 		struct task_struct *curtask = task_of(curr);
 
 		cpuacct_charge(curtask, delta_exec);
+		ttt(curtask);
 	}
 }