Hi there,
Here's a little, rough patch to illustrate the idea we discussed a little at the last Wireless Battle Mesh, that is forwarding the noisy IPv6 Unsolicited Neighbor Advertisements and gratuitous ARP Replies via unicast.
Currently these ugly unsol. NAs cause about 40% of ICMPv6 multicast overhead in our networks here.
A little more detailed explanation with some pictures can be found here:
https://www.open-mesh.org/projects/batman-adv/wiki/Unicasting-unsolicited-ne...
In the few, simple scenarios with bridged-in hosts in VMs, the patch seemed to do what it was supposed to do uppon roaming. Unsol. NAs did only appear on the node the host roamed away from.
If I remember correctly, there were also some worries about how this might work together with BLA-II. I gave it a little more thought and had a look at the BLA-II source code, but couldn't find any issues.
The only BLA-II related non-ideal thing I could think of so far is noted in the code of this patch at the according place (but I think that might not matter much in practice).
Looking forward to hear about what you think about it, especially conceptually.
Cheers, Linus
With this patch, gratouitous ARP Replies and unsolicited neighbor advertisements are forwarded via unicast to the node the according host roamed from. If this host did not roam before, then these packet are simply dropped.
Roaming devices are causing an ugly amount of ICMPv6 unsolicited neighbour advertisements: Everytime they reconnect they reassign the same IPv6 address to their interface and flood the whole mesh network with according unsolicited neighbour advertisements, telling everyone who they are. Unfortunatly unsolicited neighbor advertisements are using the all-nodes multicast address, "ff02::1", so their overhead cannot be reduced with our previously introduced group-aware multicast optimizations.
Unsolicited neighbor advertisements are being used for two things:
a) Hot-swapping devices / changing link-layer addresses (rare case) b) Updating the mac table of switches in case of a roaming host, so that they are informed about the new port location of the roamed host, avoiding packet loss
For b) it is actually not necessary to inform all switches after roaming: It's enough to inform the switches behind the node the host roamed to and roamed from. For any other switch, the roamed host will still appear behind bat0, somewhere in the mesh, still in the same direction.
Not-yet-Signed-off-by: Linus Lüssing linus.luessing@web.de --- compat.h | 7 ++++ soft-interface.c | 115 +++++++++++++++++++++++++++++++++++++++++++++++++++ translation-table.c | 62 ++++++++++++++++++++++++++- translation-table.h | 3 ++ types.h | 1 + 5 files changed, 187 insertions(+), 1 deletion(-)
diff --git a/compat.h b/compat.h index ed5b815..b777a37 100644 --- a/compat.h +++ b/compat.h @@ -242,6 +242,13 @@ static inline void skb_reset_mac_len(struct sk_buff *skb)
#if LINUX_VERSION_CODE < KERNEL_VERSION(3, 3, 0)
+static inline int batadv_ipv6_skip_exthdr(const struct sk_buff *skb, int start, + u8 *nexthdrp, __be16 *frag_offp) +{ + return ipv6_skip_exthdr(skb, start, nexthdrp); +} +#define ipv6_skip_exthdr batadv_ipv6_skip_exthdr + #define batadv_interface_add_vid(x, y, z) \ __batadv_interface_add_vid(struct net_device *dev, __be16 proto,\ unsigned short vid);\ diff --git a/soft-interface.c b/soft-interface.c index 9bf382d..d3353a4 100644 --- a/soft-interface.c +++ b/soft-interface.c @@ -155,6 +155,113 @@ static void batadv_interface_set_rx_mode(struct net_device *dev) { }
+static bool +batadv_is_ipv4_grat_reply(struct batadv_priv *bat_priv, + struct sk_buff *skb) +{ + struct arphdr *arphdr; + __be32 ip_dst; + uint8_t *hw_dst; + int hdr_size = 0; + + /* pull the ARP payload */ + if (unlikely(!pskb_may_pull(skb, hdr_size + ETH_HLEN + + arp_hdr_len(skb->dev)))) + return false; + + arphdr = (struct arphdr *)(skb->data + hdr_size + ETH_HLEN); + + /* check whether the ARP packet carries a valid IP information */ + if (arphdr->ar_hrd != htons(ARPHRD_ETHER) || + arphdr->ar_pro != htons(ETH_P_IP) || + arphdr->ar_hln != ETH_ALEN || + arphdr->ar_pln != 4) + return false; + + hw_dst = (uint8_t *)arphdr + sizeof(*arphdr) + ETH_ALEN + 4; + ip_dst = *((__be32 *)(hw_dst + ETH_ALEN)); + + /* gratouitous ARP reply */ + if (arphdr->ar_op != htons(ARPOP_REPLY) || + !is_broadcast_ether_addr(hw_dst) || + !ipv4_is_lbcast(ip_dst)) + return false; + + return true; +} + +static bool +batadv_is_ipv6_unsol_neighadv(struct batadv_priv *bat_priv, + struct sk_buff *skb) +{ + const struct ipv6hdr *ip6h; + struct nd_msg *ndm; + __be16 frag_off; + u8 nexthdr; + unsigned int len = sizeof(struct ethhdr) + sizeof(*ip6h); + + if (!pskb_may_pull(skb, len)) + return false; + + ip6h = ipv6_hdr(skb); + + if (ip6h->version != 6 || ip6h->payload_len == 0 || + !ipv6_addr_is_ll_all_nodes(&ip6h->daddr)) + return false; + + if (skb->len < len + ntohs(ip6h->payload_len)) + return false; + + nexthdr = ip6h->nexthdr; + len = ipv6_skip_exthdr(skb, len, &nexthdr, &frag_off); + + /* TODO: validate checksum */ + + if (len < 0 || nexthdr != IPPROTO_ICMPV6) + return false; + + if (!pskb_may_pull(skb, len + sizeof(*ndm))) + return false; + + ndm = (struct nd_msg *)icmp6_hdr(skb); + + if (ndm->icmph.icmp6_type != NDISC_NEIGHBOUR_ADVERTISEMENT || + ndm->icmph.icmp6_solicited) + return false; + + return true; +} + +static bool +batadv_is_ip_neigh_bcast(struct batadv_priv *bat_priv, + struct sk_buff *skb, + struct batadv_orig_node **mcast_single_orig, + unsigned short vid) +{ + struct ethhdr *ethhdr = eth_hdr(skb); + int ret; + + switch (ntohs(ethhdr->h_proto)) { + case ETH_P_ARP: + ret = batadv_is_ipv4_grat_reply(bat_priv, skb); + break; + case ETH_P_IPV6: + ret = batadv_is_ipv6_unsol_neighadv(bat_priv, skb); + break; + default: + ret = false; + } + + if (!ret) + goto out; + + ethhdr = eth_hdr(skb); + *mcast_single_orig = batadv_tt_get_prev_orig(bat_priv, ethhdr->h_source, + vid); +out: + return ret; +} + static int batadv_interface_tx(struct sk_buff *skb, struct net_device *soft_iface) { @@ -267,6 +374,14 @@ send: if (forw_mode == BATADV_FORW_SINGLE) do_bcast = false; } + + if (do_bcast && + batadv_is_ip_neigh_bcast(bat_priv, skb, + &mcast_single_orig, vid)) + if (!mcast_single_orig) + goto dropped; + + do_bcast = false; }
batadv_skb_set_priority(skb, 0); diff --git a/translation-table.c b/translation-table.c index 5f59e7f..4cbe875 100644 --- a/translation-table.c +++ b/translation-table.c @@ -175,8 +175,23 @@ batadv_tt_global_hash_find(struct batadv_priv *bat_priv, const uint8_t *addr, static void batadv_tt_local_entry_free_ref(struct batadv_tt_local_entry *tt_local_entry) { - if (atomic_dec_and_test(&tt_local_entry->common.refcount)) + struct batadv_orig_node *best_orig; + + if (atomic_dec_and_test(&tt_local_entry->common.refcount)) { + rcu_read_lock(); + best_orig = rcu_dereference(tt_local_entry->prev_orig); + rcu_read_unlock(); + + if (best_orig) { + spin_lock_bh(&best_orig->tt_lock); + rcu_assign_pointer(tt_local_entry->prev_orig, NULL); + spin_unlock_bh(&best_orig->tt_lock); + + batadv_orig_node_free_ref(best_orig); + } + kfree_rcu(tt_local_entry, common.rcu); + } }
/** @@ -509,6 +524,7 @@ bool batadv_tt_local_add(struct net_device *soft_iface, const uint8_t *addr, unsigned short vid, int ifindex, uint32_t mark) { struct batadv_priv *bat_priv = netdev_priv(soft_iface); + struct batadv_orig_node *best_orig; struct batadv_tt_local_entry *tt_local; struct batadv_tt_global_entry *tt_global = NULL; struct batadv_softif_vlan *vlan; @@ -594,6 +610,8 @@ bool batadv_tt_local_add(struct net_device *soft_iface, const uint8_t *addr, tt_local->last_seen = jiffies; tt_local->common.added_at = tt_local->last_seen;
+ rcu_assign_pointer(tt_local->prev_orig, NULL); + /* the batman interface mac and multicast addresses should never be * purged */ @@ -620,6 +638,25 @@ check_roaming: * process has already been handled */ if (tt_global && !(tt_global->common.flags & BATADV_TT_CLIENT_ROAM)) { + /* This isn't optimal: over time the best_orig towards a + * bla-backbone might change; forwarding an unsol. Neighbor + * Advertisement / grat. ARP reply later could result in a + * temporary, quick route flap because of triggering a + * bla-claim switch + */ + best_orig = batadv_transtable_search(bat_priv, NULL, + tt_global->common.addr, + tt_global->common.vid); + if (best_orig) { + /* TODO: is this a good lock to synchronize with? + * i.e. check whether we will always get the same + * best_orig->tt_lock for a tt_local->prev_orig assignment + */ + spin_lock_bh(&best_orig->tt_lock); + rcu_assign_pointer(tt_local->prev_orig, best_orig); + spin_unlock_bh(&best_orig->tt_lock); + } + /* These node are probably going to update their tt table */ head = &tt_global->orig_list; rcu_read_lock(); @@ -3710,3 +3747,26 @@ bool batadv_tt_global_is_isolated(struct batadv_priv *bat_priv,
return ret; } + +struct batadv_orig_node *batadv_tt_get_prev_orig(struct batadv_priv *bat_priv, + const uint8_t *addr, + unsigned short vid) +{ + struct batadv_tt_local_entry *tt; + struct batadv_orig_node *orig; + + tt = batadv_tt_local_hash_find(bat_priv, addr, vid); + if (!tt) + return NULL; + + rcu_read_lock(); + + orig = rcu_dereference(tt->prev_orig); + if (!orig || !atomic_inc_not_zero(&orig->refcount)) + orig = NULL; + + rcu_read_unlock(); + + batadv_tt_local_entry_free_ref(tt); + return orig; +} diff --git a/translation-table.h b/translation-table.h index ad84d7b..e71ca8e 100644 --- a/translation-table.h +++ b/translation-table.h @@ -52,5 +52,8 @@ bool batadv_tt_add_temporary_global_entry(struct batadv_priv *bat_priv, unsigned short vid); bool batadv_tt_global_is_isolated(struct batadv_priv *bat_priv, const uint8_t *addr, unsigned short vid); +struct batadv_orig_node *batadv_tt_get_prev_orig(struct batadv_priv *bat_priv, + const uint8_t *addr, + unsigned short vid);
#endif /* _NET_BATMAN_ADV_TRANSLATION_TABLE_H_ */ diff --git a/types.h b/types.h index 8854c05..0c78994 100644 --- a/types.h +++ b/types.h @@ -934,6 +934,7 @@ struct batadv_tt_common_entry { */ struct batadv_tt_local_entry { struct batadv_tt_common_entry common; + struct batadv_orig_node __rcu *prev_orig; /* rcu protected pointer */ unsigned long last_seen; };
Hello Linus,
thank you very much for the patch proposal.
Hi there,
Here's a little, rough patch to illustrate the idea we discussed a little at the last Wireless Battle Mesh, that is forwarding the noisy IPv6 Unsolicited Neighbor Advertisements and gratuitous ARP Replies via unicast.
Currently these ugly unsol. NAs cause about 40% of ICMPv6 multicast overhead in our networks here.
I think it's quite a good idea - to motivate your proposal, I'd suggest to add some information on how many broadcasts you think you can save in your networks - 40% of ICMPv6 traffic is a little vague, since we don't know how much of your traffic is ICMPv6 and what you have already blocked.
A little more detailed explanation with some pictures can be found here:
https://www.open-mesh.org/projects/batman-adv/wiki/Unicasting-unsolicited-n eighbor-advertisements
Nice page. :)
In the few, simple scenarios with bridged-in hosts in VMs, the patch seemed to do what it was supposed to do uppon roaming. Unsol. NAs did only appear on the node the host roamed away from.
If I remember correctly, there were also some worries about how this might work together with BLA-II. I gave it a little more thought and had a look at the BLA-II source code, but couldn't find any issues.
I don't see any issues either - as you've implemented, it should be similarly working as the DHCP requests from client.
The only BLA-II related non-ideal thing I could think of so far is noted in the code of this patch at the according place (but I think that might not matter much in practice).
You describe that this feature could be used for hot-swapping, and as far as I understand, turning on that feature could break IPv6 hotswapping? I don't know how much this is actually used and its probably not the most common thing for our networks (and isn't reliable anyway), but you should make you feature optional with a sysfs option or similar. Then users or firmware engineers can turn that feature on as required, just as BLA and others.
Thanks, Simon
On Fri, Aug 08, 2014 at 04:06:00PM +0200, Simon Wunderlich wrote:
Currently these ugly unsol. NAs cause about 40% of ICMPv6 multicast overhead in our networks here.
I think it's quite a good idea - to motivate your proposal, I'd suggest to add some information on how many broadcasts you think you can save in your networks - 40% of ICMPv6 traffic is a little vague, since we don't know how much of your traffic is ICMPv6 and what you have already blocked.
Hrmpf, had spend some time to set up some live-monitoring plus dumping into RRDs. And now figured out, that from day to it varies a lot, the NA traffic at Freifunk Hamburg for instance is very "peaky", so only at certain minutes but then quite a lot. Some days I see 40% again, some days it's maybe just 5-10%.
So at least along half of a day NA overhead is actually marginal.
I'll need to figure out first what's causing these spikes. Whether it's some broken or malicious client devices, the bus and train tables or something else causing these extreme spikes. Funnily, any other multicast traffic is smoothly following the #number-of-clients curve along the day (max at ~13:00, min at ~04:00)...
Cheers, Linus
b.a.t.m.a.n@lists.open-mesh.org
-
Linus Lüssing -
Simon Wunderlich