Re: INFO: possible circular locking dependency detected

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jul 15, 2011 at 01:16:44PM -0400, Steven Rostedt wrote:
> On Fri, 2011-07-15 at 10:03 -0700, Paul E. McKenney wrote:
> > On Fri, Jul 15, 2011 at 12:55:57PM -0400, Steven Rostedt wrote:
> > > On Fri, 2011-07-15 at 15:07 +0200, Peter Zijlstra wrote:
> > > 
> > > > OK, so the latter case cannot happen (rcu_preempt_check_callbacks only
> > > > sets NEED_QS when rcu_read_lock_nesting), we need two interrupts for
> > > > this to happen.
> > > > 
> > > > rcu_read_lock()
> > > > 
> > > >  <IRQ>
> > > >    |= RCU_READ_UNLOCK_NEED_QS
> > > > 
> > > > rcu_read_unlock()
> > > >   __rcu_read_unlock()
> > > >    --rcu_read_lock_nesting;
> > > >      <IRQ>
> > > > 	ttwu()
> > > >           rcu_read_lock()
> > > > 	  rcu_read_unlock()
> > > > 	    rcu_read_unlock_special()
> > > > 	      *BANG*
> > > >    rcu_read_unlock_special()
> > > > 
> > > 
> > > What about this patch? Not even compiled tested.
> > 
> > This runs afoul of the restriction that ->rcu_read_unlock_special must
> > be updated with irqs disabled, please see below.
> 
> What about changing special into a local_t, then it could be updated
> atomically wrt interrupts (not for other CPUs).

I would like to avoid increasing the cost of the rcu_read_unlock()
fastpath.  I still believe that it is possible to fix this without
increasing that cost.

> > I am also missing what the goal is -- I don't immediatly see how this
> > prevents the scenario that Ed ran into, for example.
> 
> >From the example that Peter showed above:
> 
> The interrupt happens after decrementing lock_nesting, and then when it
> did the rcu_read_unlock(), it would call special() because the ->special
> variable was set. My patch makes it so that ->special will *not* be set.

But the rcu_read_unlock() called from within the irq handler would
take a second snapshot of ->special.  It could then enter
rcu_read_unlock_special().

> We will probably need to put a preempt_disable() in there too, to keep
> the ->special being zero and scheduled out.

But ->rcu_read_unlock_special is in the task structure, so would move
with the task.  But yes, that sort of thing is one reason that I would
like to keep ->rcu_read_unlock_special modifications under irq-disable.

> > 								Thanx, Paul
> > 
> > > -- Steve
> > > 
> > > diff --git a/kernel/rcutree_plugin.h b/kernel/rcutree_plugin.h
> > > index 14dc7dd..e3545fa 100644
> > > --- a/kernel/rcutree_plugin.h
> > > +++ b/kernel/rcutree_plugin.h
> > > @@ -284,18 +284,17 @@ static struct list_head *rcu_next_node_entry(struct task_struct *t,
> > >   * notify RCU core processing or task having blocked during the RCU
> > >   * read-side critical section.
> > >   */
> > > -static void rcu_read_unlock_special(struct task_struct *t)
> > > +static int rcu_read_unlock_special(struct task_struct *t, int special)
> > >  {
> > >  	int empty;
> > >  	int empty_exp;
> > >  	unsigned long flags;
> > >  	struct list_head *np;
> > >  	struct rcu_node *rnp;
> > > -	int special;
> > > 
> > >  	/* NMI handlers cannot block and cannot safely manipulate state. */
> > >  	if (in_nmi())
> > > -		return;
> > > +		return special;
> > > 
> > >  	local_irq_save(flags);
> > > 
> > > @@ -303,7 +302,6 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > >  	 * If RCU core is waiting for this CPU to exit critical section,
> > >  	 * let it know that we have done so.
> > >  	 */
> > > -	special = t->rcu_read_unlock_special;
> > >  	if (special & RCU_READ_UNLOCK_NEED_QS) {
> > >  		rcu_preempt_qs(smp_processor_id());
> > >  	}
> > > @@ -311,7 +309,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > >  	/* Hardware IRQ handlers cannot block. */
> > >  	if (in_irq()) {
> > >  		local_irq_restore(flags);
> > > -		return;
> > > +		return special;
> > >  	}
> > > 
> > >  	/* Clean up if blocked during RCU read-side critical section. */
> > > @@ -373,6 +371,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > >  	} else {
> > >  		local_irq_restore(flags);
> > >  	}
> > > +	return special;
> > >  }
> > > 
> > >  /*
> > > @@ -385,13 +384,21 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > >  void __rcu_read_unlock(void)
> > >  {
> > >  	struct task_struct *t = current;
> > > +	int special;
> > > 
> > > +	special = ACCESS_ONCE(t->rcu_read_unlock_special);
> > > +	/*
> > > +	 * Clear special here to prevent interrupts from seeing it
> > > +	 * enabled after decrementing lock_nesting and calling
> > > +	 * rcu_read_unlock_special().
> > > +	 */
> > 
> > Any change to ->rcu_read_unlock_special from an irq handler that happens
> > here is lost.  Changes to ->rcu_read_unlock_special must be done with
> > irqs disabled.  And I hope to avoid irq disabling on the rcu_read_unlock()
> > fastpath.
> 
> We can check if special changed afterwards. Hmm, would a xchg be bad to
> do?

I would really like to avoid that in the common rcu_read_unlock() fastpath.

> > > +	t->rcu_read_unlock_special = 0;
> > >  	barrier();  /* needed if we ever invoke rcu_read_unlock in rcutree.c */
> > >  	--t->rcu_read_lock_nesting;
> > >  	barrier();  /* decrement before load of ->rcu_read_unlock_special */
> > > -	if (t->rcu_read_lock_nesting == 0 &&
> > > -	    unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
> > > -		rcu_read_unlock_special(t);
> > > +	if (t->rcu_read_lock_nesting == 0 && special)
> > > +		special = rcu_read_unlock_special(t, special);
> > 
> > And changes to ->rcu_read_unlock_special from an irq handler that happens
> > here are also lost.
> 
> How expensive is xchg?
> 
> 	special = xchg(&t->rcu_read_lock_special, 0);
> 	[..]
> 	special = xchg(&t->rcu_read_lock_special, special);
> 	/* check special */
> 
> Or is xchg too expensive for rcu_read_unlock()?

It is a bit expensive for that fastpath.

							Thanx, Paul

> -- Steve
> 
> > 
> > > +	t->rcu_read_unlock_special = special;
> > >  #ifdef CONFIG_PROVE_LOCKING
> > >  	WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0);
> > >  #endif /* #ifdef CONFIG_PROVE_LOCKING */
> > > 
> > > 
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/