Cisco Systems

Santa Barbara 93117 California USA fred@cisco.com

Department of Computer Science University of Auckland PB 92019 Auckland 1142 New Zealand brian.e.carpenter@gmail.com

Internet IPv6 Maintenance This document describes expected IPv6 host behavior in a network that has more than one prefix, each allocated by an upstream network that implements BCP 38 ingress filtering, when the host has multiple routers to choose from. It also applies to other scenarios such as the usage of stateful firewalls that effectively act as address-based filters. This host behavior may interact with source address selection in a given implementation, but logically follows it. Given that the network or host is, or appears to be, multihomed with multiple provider-allocated addresses, that the host has elected to use a source address in a given prefix, and that some but not all neighboring routers are advertising that prefix in their Router Advertisement Prefix Information Options, this document specifies to which router a host should present its transmission. It updates RFC 4861.

This document describes the expected behavior of an IPv6 host in a network that has more than one prefix, each allocated by an upstream network that implements BCP 38 ingress filtering, and in which the host is presented with a choice of routers. It expects that the network will implement some form of egress routing, so that packets sent to a host outside the local network from a given ISP's prefix will go to that ISP. If the packet is sent to the wrong egress, it is liable to be discarded by the BCP 38 filter. However, the mechanics of egress routing once the packet leaves the host are out of scope. The question here is how the host interacts with that network. Various aspects of this issue, and possible solution approaches, are discussed in the document IPv6 Multihoming without Network Address Translation. BCP 38 filtering by ISPs is not the only scenario where such behavior is valuable. Implementations that combine existing recommendations, such as can also result in such filtering. Another case is when the connections to the upstream networks include stateful firewalls, such that return packets in a stream will be discarded if they do not return via the firewall that created state for the outgoing packets. A similar cause of such discards is unicast reverse path forwarding (uRPF) . In this document, the term "filter" is used for simplicity to cover all such cases. In any case, one cannot assume the host to be aware whether an ingress filter, a stateful firewall, or any other type of filter is in place. Therefore, the only consistent solution is to implement the features defined in this document. Note that, apart from ensuring that a message with a given source address is given to a first-hop router that appears to know about the prefix in question, this specification is consistent with . Nevertheless, implementers of Sections 5.2, 6.2.3, 6.3.4 and 8 of RFC 4861 will need to extend their implementations accordingly. This specification is fully consistent with and implementers will need to add support for its Rule 5.5. Hosts that do not support these features may fail to communicate in the presence of filters as described above.

It could be argued that the proposal of this document, which is to send messages using a source address in a given prefix to the router that advertised the prefix in its RA, is a form of 's Strong ES (e.g. Host) Model, discussed in section 3.3.4.2 of that document. In short, identifies two basic models, in which the "strong host" model models the host as a set of hosts in one chassis, each of which uses a single address on a single interface, and always both sends and receives on that interface, and the "weak host" model treats the host as one system with zero or more addresses on every interface, and capable of using any interface for any communication. As noted there, neither model is completely satisfactory. For example, a host with a link-local-only interface and a default route pointing to the interface will necessarily send packets using any address using that interface, and therefore be a de facto weak host. If the router upstream from such a host implements BCP 38 Ingress Filtering, such as by implementing uRPF on each interface, the router might prevent communication by weak hosts.

The proposal also differs slightly from 's language of the Strong Host Model. The statement is that the packet will go to the router that advertised a given prefix, but doesn't state what interface that might happen on. Hence, if the router is a MIF router and is using the same prefix on two or more LANs shared by the host (as in ), the host might use each of those LANs and meet the requirement. The Strong Host Model is not stated in those terms, but in terms of the interface used, and would find a MIF router quite confusing: (A) A host [MUST] silently discard an incoming datagram whose destination address does not correspond to the physical interface through which it is received. (B) A host [MUST] restrict itself to sending (non-source- routed) IP datagrams only through the physical interface that corresponds to the IP source address of the datagrams. However, comparing the presumptive route lookup mechanisms in each model, this proposal is indeed most similar to the Strong Host Model, as is any source/destination routing paradigm. route(src IP addr, dest IP addr, TOS) -> gateway route(dest IP addr, TOS) -> gateway, interface In the hypothetical MIF model suggested in , the address fails to identify a single interface, but it does identify a single gateway.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

A host receives prefixes in a Router Advertisement, which goes on to identify whether they are usable by SLAAC . When no prefixes are usable for SLAAC, the Router Advertisement would normally signal the availability of DHCPv6 and the host would use it to configure its addresses. In the latter case (or if both SLAAC and DHCPv6 are used on the same link for some reason) it will generally be the case that the configured addresses match one of the prefixes advertised in a Router Advertisement that are supposed to be on-link for that link. The simplest multihomed network implementation in which a host makes choices among routers might be a LAN with one or more hosts on it and two or more routers, one for each upstream network, or a host that is served by disjoint networks on separate interfaces. In such a network, especially the latter, there is not necessarily a routing protocol, and the two routers may not even know that the other is a router as opposed to a host, or may be configured to ignore its presence. One might expect that the routers may or may not receive each other's RAs and form an address in the other router's prefix (which is not per , but is implemented by some stub router implementations). However, all hosts in such a network might be expected to create an address in each prefix so advertised.

If there is no routing protocol among those routers, there is no mechanism by which packets can be deterministically forwarded between the routers (as described in BCP 84) in order to avoid filters. Even if there was routing, it would result in an indirect route, rather than a direct route originating with the host; this is not "wrong", but can be inefficient. Therefore the host would do well to select the appropriate router itself. Since the host derives fundamental default routing information from the Router Advertisement, this implies that, in any network with hosts using multiple prefixes, each prefix SHOULD be advertised via a Prefix Information Option (PIO) by one of the attached routers, even if addresses are being assigned using DHCPv6. A router that advertises a prefix indicates that it is able to appropriately route packets with source addresses within that prefix, regardless of the setting of the L and A flags in the PIO. In some circumstances both L and A might be zero. If SLAAC is not wanted (A=0) and there is no reason to announce an on-link prefix (L=0), a PIO SHOULD be sent to inform hosts that the prefix is source-routed by the router in question. Although this does not violate the existing standard , such a PIO has not previously been common, and it is possible that existing host implementations simply ignore such a PIO or that a router implementation rejects such a PIO as a configuration error. Newer implementations that support this mechanism will need to be updated accordingly: a host SHOULD NOT ignore a PIO simply because both L and A flags are cleared; a router SHOULD be able to send such a PIO.

The direct implication of is that, if the network uses a routing protocol, the routing protocols used in multihomed networks SHOULD implement source-prefix based egress routing, for example as described in . Network designs exist that can usefully limit themselves to static routing (such as a simple tree network), or may internally use no routers at all, such as a single LAN with two CE routers, each of which leads to a different upstream network.

As described in and , a Router Advertisement may contain zero or more Prefix information Options (PIOs), or zero or more Route Information Options (RIOs). In their original intent, these indicate general information to a host: "the router whose address is found in the source address field of this packet is one of your default routers", "you might create an address in this prefix", or "this router would be a good place to send traffic directed to a given destination prefix". In a multi-homed network implementing source/destination routing, the interpretation of default router or an RIO has to be modified with the words "if the source address is in one of the prefixes I advertise in a PIO". Additionally, the PIO must be reinterpreted to also imply that the advertising router would be a reasonable first hop for any packet using a source address in any advertised prefix.

The implications bear consideration. Imagine, , that hosts Alice and Bob are in communication, Bob's network is multihomed, and Bob's first hop routers are Bob-A (to upstream ISP A advertising prefix PA) and Bob-B (to upstream network B and advertising prefix PB). If Bob is responding to a message from Alice, his choice of source address is forced to be the address Alice used as a destination (which we may presume to have been in prefix PA). Hence, Bob created or was assigned an address in PA, and can only reasonably send traffic using it to Bob-A as a first hop router. If there were several instances of Bob-A and one had advertised itself as a default router or as having a route to Alice, that is the router Bob should choose. If none of Bob-A have advertised that but Bob-B has, it is irrelevant; Bob is using the address allocated in PA and courts a BCP 38 discard if he doesn't send the packet to Bob-A. In the special case that Bob is initiating the conversation, an RIO might, however, influence source address choice. Bob could presumably use any address allocated to him, in this case his address in PA or PB. If Bob-B has advertised an RIO for Alice's prefix and Bob-A has not, Bob MAY take that fact into account in address selection - choosing an address that would allow him to make use of the RIO.

Default Router Selection is modified as follows: A host SHOULD select default routers for each prefix it is assigned an address in. Routers that have advertised the prefix in its Router Advertisement message SHOULD be preferred over routers that do not advertise the prefix. As a result of doing so, when a host sends a packet using a source address in one of those prefixes and has no history directing it otherwise, it SHOULD send it to the indicated default router. In the "simplest" network described in , that would get it to the only router that is directly capable of getting it to the right ISP. This will also apply in more complex networks, even when more than one physical or virtual interface is involved. In more complex cases, wherein routers advertise RAs for multiple prefixes whether or not they have direct or isolated upstream connectivity, the host is dependent on the routing system already. If the host gives the packet to a router advertising its source prefix, it should be able to depend on the router to do the right thing.

There is an interaction with Default Address Selection. A host following the recommendation in the previous section will store information about which next-hops advertised which prefixes. Rule 5.5 of RFC 6724 states that the source address used to send to a given destination address should if possible be chosen from a prefix known to be advertised by the next-hop router for that destination. This selection rule would therefore be applicable in a host following the recommendation in the previous section.

There is potential for adverse interaction with any off-link Redirect (Redirect for a destination that is not on-link) message sent by a router in accordance with Section 8 of . Hosts SHOULD apply off-link redirects only for the specific pair of source and destination addresses concerned, so the host's Destination Cache may need to contain appropriate source-specific entries.

Some modern hosts maintain history, in terms of what has previously worked or not worked for a given address or prefix and in some cases the effective window and MSS values for TCP or other protocols. This might include a next hop address for use when a packet is sent to the indicated address. When such a host makes a successful exchange with a remote destination using a particular address pair, and the host has previously received a PIO that matches the source address, then the host SHOULD include the prefix in such history, whatever the setting of the L and A flags in the PIO. On subsequent attempts to communicate with that destination, if it has an address in that prefix at that time, a host MAY use an address in the remembered prefix for the session.

Consider a network where routers on a link run a routing protocol and are configured with the same information. Thus, on each link all routers advertise all prefixes on the link. The assumption that packets will be forwarded to the appropriate egress by the local routing system might cause at least one extra hop in the local network (from the host to the wrong router, and from there to another router on the same link). In a slightly more complex situation such as the disjoint LAN case of , which happens to be one of the authors' home plus corporate home-office configuration, the two upstream routers might be on different LANs and therefore different subnets (e.g., the host is itself multi-homed). In that case, there is no way for the "wrong" router to detect the existence of the "right" router, or to route to it. In such a case it is particularly important that hosts take the responsibility to memorize and select the best first-hop as described in .

This memo asks the IANA for no new parameters.

This document does not create any new security or privacy exposures. It is intended to avoid connectivity issues in the presence of BCP 38 ingress filters or stateful firewalls combined with multihoming. There might be a small privacy improvement, however: with the current practice, a multihomed host that sends packets with the wrong address to an upstream router or network discloses the prefix of one upstream to the other upstream network. This practice reduces the probability of that occurrence.

Comments were received from Jinmei Tatuya and Ole Troan, who have suggested important text, plus Mikael Abrahamsson, Steven Barth, Juliusz Chroboczek, Toerless Eckert, David Farmer, Dusan Mudric, Tadahisa Okimoto, Pierre Pfister, Behcet Sarikaya, Mark Smith, Bob Hinden, and James Woodyatt.

2015-08-05 Update text on PIOs, added text on Redirects, and clarified the concept of a "simple" network, 2015-08-13. Clarifications after WG discussions, 2015-08-19. More clarifications after more WG discussions, especially adding stateful firewalls, uRPF, and more precise discussion of RFC 4861, 2015-09-03. Responds to various comments including Questions regarding RFC 1122's strong and weak host models. This model is, strictly speaking, neither, but is most similar to the strong host model. Some wording errors. Requests for discussion of the handling of the RIO, PIO, and Default Router List in an RA. More clarifications after more WG discussions, 2015-11-03. A final clarification re uRPF, 2015-12-15.