Re: [PATCH] netpoll: use non-BH variant of RCU

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

On Tue, Aug 10, 2010 at 04:43:58PM -0400, Herbert Xu wrote:
> On Tue, Aug 10, 2010 at 04:25:24PM -0400, John W. Linville wrote:
> > "netpoll: Fix RCU usage" switched netpoll_rx to use the BH variant
> > of RCU.  Unfortunately, calling netpoll_rx from netif_rx resulted in
> > the following backtrace:
> 
> Thanks for catching this John!
> 
> > diff --git a/include/linux/netpoll.h b/include/linux/netpoll.h
> > index 413742c..0bdd527 100644
> > --- a/include/linux/netpoll.h
> > +++ b/include/linux/netpoll.h
> > @@ -63,8 +63,8 @@ static inline bool netpoll_rx(struct sk_buff *skb)
> >  	unsigned long flags;
> >  	bool ret = false;
> >  
> > -	rcu_read_lock_bh();
> > -	npinfo = rcu_dereference_bh(skb->dev->npinfo);
> > +	rcu_read_lock();
> > +	npinfo = rcu_dereference(skb->dev->npinfo);
> 
> I really wanted to avoid mixing the two different RCU primitives
> because they require different synchronisations.
> 
> In this case, the problem is that we're being called in IRQ
> context, where BH is diabled anyway, so we don't actually need
> to do anything (assuming IRQ is off).
> 
> Paul, what could we do to resolve this (other than by switching
> to the non-BH variant of RCU)? Perhaps an additional variant
> of rcu_read_lock_bh that checks whether IRQ is off?

Hello, Herbert,

Your suggestion of providing another API rcu_read_lock_irqsoff()
and rcu_read_unlock_irqsoff() is the best I can think of right offhand.

What tree/commit do you need the patch against?

							Thanx, Paul