Re: [PATCH net-next 3/4] bridge: multicast port group RCU fix

[Stephen Hemminger <shemminger@vyatta.com>]

On Wed, 28 Apr 2010 11:07:09 +0800
Herbert Xu <herbert@gondor.apana.org.au> wrote:

> On Tue, Apr 27, 2010 at 06:01:06PM -0700, Stephen Hemminger wrote:
> > The recently introduced bridge mulitcast port group list was only
> > partially using RCU correctly. It was missing rcu_dereference()
> > and missing the necessary barrier on deletion.
> > 
> > The code should have used one of the standard list methods (list or hlist)
> > instead of open coding a RCU based link list.
> > 
> > Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
> > 
> > --- a/net/bridge/br_forward.c	2010-04-27 17:51:27.909588950 -0700
> > +++ b/net/bridge/br_forward.c	2010-04-27 17:53:18.790721091 -0700
> > @@ -217,7 +217,7 @@ static void br_multicast_flood(struct ne
> >  	prev = NULL;
> >  
> >  	rp = rcu_dereference(br->router_list.first);
> > -	p = mdst ? mdst->ports : NULL;
> > +	p = mdst ? rcu_dereference(mdst->ports) : NULL;
> >  	while (p || rp) {
> >  		lport = p ? p->port : NULL;
> >  		rport = rp ? hlist_entry(rp, struct net_bridge_port, rlist) :
> > @@ -231,7 +231,7 @@ static void br_multicast_flood(struct ne
> >  			goto out;
> >  
> >  		if ((unsigned long)lport >= (unsigned long)port)
> > -			p = p->next;
> > +			p = rcu_dereference(p->next);
> >  		if ((unsigned long)rport >= (unsigned long)port)
> >  			rp = rcu_dereference(rp->next);
> >  	}
> 
> Thanks for catching this!
> 
> > --- a/net/bridge/br_multicast.c	2010-04-27 17:51:31.509593914 -0700
> > +++ b/net/bridge/br_multicast.c	2010-04-27 17:52:48.209243982 -0700
> > @@ -259,7 +259,7 @@ static void br_multicast_del_pg(struct n
> >  		if (p != pg)
> >  			continue;
> >  
> > -		*pp = p->next;
> > +		rcu_assign_pointer(*pp, p->next);
> 
> But this is bogus.  br_multicast_del_pg is removing an entry from
> the RCU list.
> 
> You only need write barriers when you're putting a new entry in
> it, and only if there is no other barrier between the code filling
> in the new entry and the line adding it to the RCU list.

Yeah, it is extra barrier (one more reason to stick to hlist_del_rcu)




--