bonding: fix a race in IGMP handling
Eric Dumazet [Thu, 18 Nov 2010 17:30:42 +0000 (09:30 -0800)]
RCU conversion in IGMP code done in net-next-2.6 raised a race in
__bond_resend_igmp_join_requests().

It iterates in_dev->mc_list without appropriate protection (RTNL, or
read_lock on in_dev->mc_list_lock).

Another cpu might delete an entry while we use it and trigger a fault.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/bonding/bond_main.c

index bdb68a6..71a1697 100644 (file)
@@ -878,8 +878,10 @@ static void __bond_resend_igmp_join_requests(struct net_device *dev)
        rcu_read_lock();
        in_dev = __in_dev_get_rcu(dev);
        if (in_dev) {
+               read_lock(&in_dev->mc_list_lock);
                for (im = in_dev->mc_list; im; im = im->next)
                        ip_mc_rejoin_group(im);
+               read_unlock(&in_dev->mc_list_lock);
        }
 
        rcu_read_unlock();