From: Andy Gospodarek Subject: Re: [PATCH net-next-2.6] bonding: allow arp_ip_targets to be on a separate vlan from bond device Date: Wed, 2 Dec 2009 16:24:49 -0500 Message-ID: <20091202212449.GL1639@gospo.rdu.redhat.com> References: <20091130201453.GF1639@gospo.rdu.redhat.com> <29892.1259625638@death.nxdomain.ibm.com> <20091201012145.GH1639@gospo.rdu.redhat.com> <6611.1259632635@death.nxdomain.ibm.com> <20091201144410.GI1639@gospo.rdu.redhat.com> <21331.1259702893@death.nxdomain.ibm.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: netdev@vger.kernel.org To: Jay Vosburgh Return-path: Received: from mx1.redhat.com ([209.132.183.28]:41632 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753385AbZLBVYz (ORCPT ); Wed, 2 Dec 2009 16:24:55 -0500 Content-Disposition: inline In-Reply-To: <21331.1259702893@death.nxdomain.ibm.com> Sender: netdev-owner@vger.kernel.org List-ID: On Tue, Dec 01, 2009 at 01:28:13PM -0800, Jay Vosburgh wrote: > Andy Gospodarek wrote: > [...] > >I am using arp_validate, actually. I forgot that the arp_validate > >option doesn't show up in the output of /proc/net/bonding/bondX and I > >intended to have that in the subject, but somehow dropped it. > > Ok, I was doing it wrong earlier; it works with arp_validate. > I'm seeing one problem with tcpdump, though, which I'll get to in a > minute. > > Could you update the summary / changelog message to mention that > this patch fixes the specific case of arp_validate + arp_ip_target on > VLAN? > > Second, in regards to this: > > --- a/net/core/dev.c > +++ b/net/core/dev.c > @@ -2439,8 +2439,8 @@ int netif_receive_skb(struct sk_buff *skb) > skb->skb_iif = skb->dev->ifindex; > > null_or_orig = NULL; > - orig_dev = skb->dev; > - if (orig_dev->master) { > + orig_dev = __dev_get_by_index(dev_net(skb->dev),skb->skb_iif); > + if (orig_dev->master && !(skb->dev->priv_flags & IFF_802_1Q_VLAN)) { > if (skb_bond_should_drop(skb)) > null_or_orig = orig_dev; /* deliver only exact match */ > else > > Would it be useful to add a comment to the effect that VLAN > packets are run through skb_bond_should_drop at the VLAN layer? > > Lastly, in regards to this: > > @@ -2492,7 +2492,7 @@ ncls: > &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) { > if (ptype->type == type && > (ptype->dev == null_or_orig || ptype->dev == skb->dev || > - ptype->dev == orig_dev)) { > + ptype->dev == orig_dev || ptype->dev == orig_dev->master)) { > if (pt_prev) > ret = deliver_skb(skb, pt_prev, orig_dev); > pt_prev = ptype; > > This is presumably here because orig_dev will now be the actual > slave the packet arrived on, but we want to additionally deliver to the > master, correct? > > Lastly, tcpdump. > > This patch appears to affect what traffic tcpdump of a slave or > the bonding master itself will capture. Previously, tcpdump of the > active slave would see only the transmitted packets sent over the bond, > and tcpdump of the inactive slave would see incoming Ethernet-layer > multicast or broadcasts sent to its switch port. Tcpdump on the master > would see all sent and non-VLAN received traffic, and tcpdump of the > VLAN interface over the master would see just the VLAN traffic. > > After this change, tcpdump of the active slave or of the bonding > master (bond0) sees both sent and received traffic for the VLAN, but > nothing for the non-VLAN traffic other than incoming broadcast / > multicasts. This holds true whether or not a VLAN is configured. > > I added a "ptype->dev == orig_dev->master" test to the ptype_all > receive block in netif_receive_skb, but it didn't help. At the moment, > I'm not exactly sure why tcpdump breaks. > Jay, The issue was that that orig_dev was getting set to the active slave, so your running tcpdump on the active slave made the conditional inside this loop: list_for_each_entry_rcu(ptype, &ptype_all, list) { if (ptype->dev == null_or_orig || ptype->dev == skb->dev || ptype->dev == orig_dev) { if (pt_prev) ret = deliver_skb(skb, pt_prev, orig_dev); pt_prev = ptype; } } hit and deliver_skb was being called for all traffic coming toward bond0.. I'm not completely happy with this solutoin, but I think it resolves both the original problem I was trying to solve and the regression you discovered with your original patch. Let me know if you see everything working now like I do. diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c index 726bd75..b1e3b2f 100644 --- a/drivers/net/bonding/bond_main.c +++ b/drivers/net/bonding/bond_main.c @@ -2697,6 +2697,19 @@ static int bond_arp_rcv(struct sk_buff *skb, struct net_device *dev, struct pack bond = netdev_priv(dev); read_lock(&bond->lock); + /* + * We may have dev passed in as a vlan device, so make sure to get to the + * core netdev before continuing. + */ + if (dev->priv_flags & IFF_802_1Q_VLAN) { + dev = vlan_dev_real_dev(dev); + /* + * Don't necessarily trust passed in orig_dev since vlan accelerated + * netdevs and bonding don't play well together. + */ + orig_dev = __dev_get_by_index(dev_net(skb->dev),skb->skb_iif); + } + pr_debug("bond_arp_rcv: bond %s skb->dev %s orig_dev %s\n", bond->dev->name, skb->dev ? skb->dev->name : "NULL", orig_dev ? orig_dev->name : "NULL"); diff --git a/net/8021q/vlan_core.c b/net/8021q/vlan_core.c index e75a2f3..8d8a778 100644 --- a/net/8021q/vlan_core.c +++ b/net/8021q/vlan_core.c @@ -14,6 +14,7 @@ int __vlan_hwaccel_rx(struct sk_buff *skb, struct vlan_group *grp, if (skb_bond_should_drop(skb)) goto drop; + skb->skb_iif = skb->dev->ifindex; __vlan_hwaccel_put_tag(skb, vlan_tci); skb->dev = vlan_group_get_device(grp, vlan_tci & VLAN_VID_MASK); @@ -85,6 +86,7 @@ vlan_gro_common(struct napi_struct *napi, struct vlan_group *grp, if (skb_bond_should_drop(skb)) goto drop; + skb->skb_iif = skb->dev->ifindex; __vlan_hwaccel_put_tag(skb, vlan_tci); skb->dev = vlan_group_get_device(grp, vlan_tci & VLAN_VID_MASK); diff --git a/net/core/dev.c b/net/core/dev.c index 5d131c2..9c3ba0d 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -2421,6 +2421,7 @@ int netif_receive_skb(struct sk_buff *skb) { struct packet_type *ptype, *pt_prev; struct net_device *orig_dev; + struct net_device *bond_dev; struct net_device *null_or_orig; int ret = NET_RX_DROP; __be16 type; @@ -2487,12 +2488,24 @@ ncls: if (!skb) goto out; + /* + * A bonding interface with a VLAN on top doesn't play nicely when the + * netdev in the bond is capable of stripping the VLAN tag for us. + * Knowing the base bond device is important in the event that bond + * control frames arrive with a VLAN tag, but need to be serviced by + * a hook installed for the base bond device. + */ + bond_dev = skb->dev; + if ((bond_dev->priv_flags & IFF_802_1Q_VLAN) && + (vlan_dev_real_dev(bond_dev)->priv_flags & IFF_BONDING)) + bond_dev = vlan_dev_real_dev(bond_dev); + type = skb->protocol; list_for_each_entry_rcu(ptype, &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) { if (ptype->type == type && (ptype->dev == null_or_orig || ptype->dev == skb->dev || - ptype->dev == orig_dev)) { + ptype->dev == orig_dev || ptype->dev == bond_dev)) { if (pt_prev) ret = deliver_skb(skb, pt_prev, orig_dev); pt_prev = ptype;