Asymmetric Extended Route Optimization (AERO)

The following sections specify the operation of IP over Asymmetric Extended Route Optimization (AERO) links:

R1 | |(P1->S1; P2->S2)| | default->R1 | | P1->C1 | | ASP A1 | | P2->C2 | +-------+------+ +--------+-------+ +------+-------+ | | | X---+---+-------------------+------------------+---+---X | AERO Link | +-----+--------+ +--------+-----+ |AERO Client C1| |AERO Client C2| | Nbr: S1 | | Nbr: S2 | | default->S1 | | default->S2 | | ACP P1 | | ACP P2 | +------+-------+ +------+-------+ | | .-. .-. ,-( _)-. ,-( _)-. .-(_ IP )-. +-------+ +-------+ .-(_ IP )-. (__ EUN )--|Host H1| |Host H2|--(__ EUN ) `-(______)-' +-------+ +-------+ `-(______)-' ]]> presents the AERO link reference model. In this model: AERO Relay R1 aggregates AERO Service Prefix (ASP) A1, acts as a default router for its associated Servers S1 and S2, and connects the AERO link to the rest of the IP Internetwork. AERO Servers S1 and S2 associate with Relay R1 and also act as default routers for their associated Clients C1 and C2. AERO Clients C1 and C2 associate with Servers S1 and S2, respectively. They receive AERO Client Prefix (ACP) delegations P1 and P2, and also act as default routers for their associated physical or internal virtual EUNs. (Alternatively, Clients can act as multi-addressed hosts without serving any EUNs). Simple hosts H1 and H2 attach to the EUNs served by Clients C1 and C2, respectively. Each node on the AERO link maintains an AERO interface neighbor cache and an IP forwarding table the same as for any link. In common operational practice, there may be many additional Relays, Servers and Clients.

AERO Relays provide default forwarding services to AERO Servers. Each Relay also peers with each Server in a dynamic routing protocol instance to discover the Server's list of associated ACPs (see ). Relays forward packets between neighbors connected to the same AERO link and also forward packets between the AERO link and the native IP Internetwork. Relays present the AERO link to the native Internetwork as a set of one or more AERO Service Prefixes (ASPs) and serve as a gateway between the AERO link and the Internetwork. Relays maintain an AERO interface neighbor cache entry for each AERO Server, and maintain an IP forwarding table entry for each AERO Client Prefix (ACP). AERO Relays can also be configured to act as AERO Servers. AERO Servers provide default forwarding services to AERO Clients. Each Server also peers with each Relay in a dynamic routing protocol instance to advertise its list of associated ACPs (see ). Servers configure a DHCPv6 server function and act as delegating routers to facilitate Prefix Delegation (PD) exchanges with Clients. Each delegated prefix becomes an ACP taken from an ASP. Servers forward packets between AERO interface neighbors, and maintain an AERO interface neighbor cache entry for each Relay. They also maintain both neighbor cache entries and IP forwarding table entries for each of their associated Clients. AERO Servers can also be configured to act as AERO Relays. AERO Clients act as requesting routers to receive ACPs through DHCPv6 PD exchanges with AERO Servers over the AERO link. Each Client MAY associate with a single Server or with multiple Servers, e.g., for fault tolerance, load balancing, etc. Each IPv6 Client receives at least a /64 IPv6 ACP, and may receive even shorter prefixes. Similarly, each IPv4 Client receives at least a /32 IPv4 ACP (i.e., a singleton IPv4 address), and may receive even shorter prefixes. Clients maintain an AERO interface neighbor cache entry for each of their associated Servers as well as for each of their correspondent Clients.

The AERO routing system comprises a private instance of the Border Gateway Protocol (BGP) that is coordinated between Relays and Servers and does not interact with either the public Internet BGP routing system or the native IP Internetwork interior routing system. Relays advertise only a small and unchanging set of ASPs to the native routing system instead of the full dynamically changing set of ACPs. In a reference deployment, each AERO Server is configured as an Autonomous System Border Router (ASBR) for a stub Autonomous System (AS) using an AS Number (ASN) that is unique within the BGP instance, and each Server further peers with each Relay but does not peer with other Servers. Similarly, Relays do not peer with each other, since they will reliably receive all updates from all Servers and will therefore have a consistent view of the AERO link ACP delegations. Each Server maintains a working set of associated ACPs, and dynamically announces new ACPs and withdraws departed ACPs in its BGP updates to Relays. Clients are expected to remain associated with their current Servers for extended timeframes, however Servers SHOULD selectively suppress BGP updates for impatient Clients that repeatedly associate and disassociate with them in order to dampen routing churn. Each Relay configures a black-hole route for each of its ASPs. By black-holing the ASPs, the Relay will maintain forwarding table entries only for the ACPs that are currently active, and all other ACPs will correctly result in destination unreachable failures due to the black hole route. Relays do not send BGP updates for ACPs to Servers, but instead originate a default route. In this way, Servers have only partial topology knowledge (i.e., they know only about the ACPs of their directly associated Clients) and they forward all other packets to Relays which have full topology knowledge. Scaling properties of the AERO routing system are limited by the number of BGP routes that can be carried by Relays. Assuming O(10^6) as a maximum number of BGP routes, this means that O(10^6) Clients can be serviced by a single set of Relays. A means of increasing scaling would be to assign a different set of Relays for each set of ASPs. In that case, each Server still peers with each Relay, but the Server institutes route filters so that it only sends BGP updates to the specific set of Relays that aggregate the ASP. For example, if the ASP for the AERO link is 2001:db8::/32, a first set of Relays could service the ASP segment 2001:db8::/40, a second set of Relays could service 2001:db8:0100::/40, a third set could service 2001:db8:0200::/40, etc. Assuming up to O(10^3) sets of Relays, the AERO routing system can then accommodate O(10^9) ACPs with no additional overhead for Servers and Relays (for example, it should be possible to service 4 billion /64 ACPs taken from a /32 ASP and even more for shorter ASPs). In this way, each set of Relays services a specific set of ASPs that they advertise to the native routing system, and each Server configures ASP-specific routes that list the correct set of Relays as next hops. This arrangement also allows for natural incremental deployment, and can support small scale initial deployments followed by dynamic deployment of additional Clients, Servers and Relays without disturbing the already-deployed base. Note that in an alternate routing arrangement each set of Relays could advertise an aggregated ASP for the link into the native routing system even though each Relay services only a segment of the ASP. In that case, a Relay upon receiving a packet with a destination address covered by the ASP segment of another Relay can simply tunnel the packet to the correct Relay. The tradeoff then is the penalty for Relay-to-Relay tunneling compared with reduced routing information in the native routing system. Finally, Relays can determine preferences for ACPs learned from multiple Servers by examining a BGP weight associated with each Server's peering configuration. In this way Relays can choose the Server with the highest weight as the preferred path, and then fail over to a Server with lower weight in case of ACP withdrawal or Server failure.

AERO interface link-local address types include administratively-provisioned addresses and AERO addresses. Administratively-provisioned addresses are allocated from the range fe80::/96 and assigned to a Server or Relay's AERO interface. Administratively-provisioned addresses MUST be managed for uniqueness by the administrative authority for the AERO link. (Note that fe80:: is the IPv6 link-local subnet router anycast address, and fe80::ffff:ffff is the address used by Clients to bootstrap AERO address autoconfiguration. These special addresses are therefore not available for administrative provisioning.) An AERO address is an IPv6 link-local address with an embedded prefix based on an ACP and associated with a Client's AERO interface. AERO addresses remain stable as the Client moves between topological locations, i.e., even if its link-layer addresses change. For IPv6, AERO addresses begin with the prefix fe80::/64 and include in the interface identifier (i.e., the lower 64 bits) a 64-bit prefix taken from one of the Client's IPv6 ACPs. For example, if the AERO Client receives the IPv6 ACP: 2001:db8:1000:2000::/56 it constructs its corresponding AERO addresses as: fe80::2001:db8:1000:2000 fe80::2001:db8:1000:2001 fe80::2001:db8:1000:2002 ... etc. ... fe80::2001:db8:1000:20ff For IPv4, AERO addresses are based on an IPv4-mapped IPv6 address formed from an IPv4 ACP and with a Prefix Length of 96 plus the ACP prefix length. For example, for the IPv4 ACP 192.0.2.32/28 the IPv4-mapped IPv6 ACP is: 0:0:0:0:0:FFFF:192.0.2.16/124 The Client then constructs its AERO addresses with the prefix fe80::/64 and with the lower 64 bits of the IPv4-mapped IPv6 address in the interface identifier as: fe80::FFFF:192.0.2.16 fe80::FFFF:192.0.2.17 fe80::FFFF:192.0.2.18 ... etc. ... fe80:FFFF:192.0.2.31 When the Server delegates ACPs to the Client, both the Server and Client use the lowest-numbered AERO address from the first ACP delegation as the "base" AERO address. (For example, for the ACP 2001:db8:1000:2000::/56 the base address is 2001:db8:1000:2000.) The Client then assigns the base AERO address to the AERO interface and uses it for the purpose of maintaining the neighbor cache entry. If the Client has multiple AERO addresses (i.e., when there are multiple ACPs and/or ACPs with short prefix lengths), the Client originates IPv6 ND messages using the base AERO address as the source address and accepts and responds to IPv6 ND messages destined to any of its AERO addresses as equivalent to the base AERO address. In this way, the Client maintains a single neighbor cache entry with multiple AERO addresses.

AERO interfaces use encapsulation (see: ) to exchange packets with neighbors attached to the AERO link. AERO interfaces maintain a neighbor cache, and use both DHCPv6 and IPv6 ND control messaging to manage the creation, modification and deletion of neighbor cache entries. AERO interfaces use standard DHCPv6 messaging for prefix delegation, prefix management, admission control and neighbor cache entry establishment. AERO interfaces use unicast IPv6 ND Neighbor Solicitation (NS), Neighbor Advertisement (NA), Router Solicitation (RS) and Router Advertisement (RA) messages for neighbor cache management the same as for any IPv6 link. AERO interfaces use two IPv6 ND redirection message types -- the first known as a Predirect message and the second being the standard Redirect message (see ). AERO interface ND messages include one or more Source/Target Link-Layer Address Options (S/TLLAOs) formatted as shown in :

In this format: Type is set to '1' for SLLAO or '2' for TLLAO the same as for IPv6 ND. Length is set to the constant value '5' (i.e., 5 units of 8 octets). Reserved is set to the value '0' on transmission and ignored on receipt. Interface ID is set to an integer value between 0 and 65535 corresponding to an underlying interface of the AERO node. UDP Port Number and IP Address are set to the addresses used by the AERO node when it sends encapsulated packets over the underlying interface. When UDP is not used as part of the encapsulation, UDP Port Number is set to the value '0'. When the encapsulation IP address family is IPv4, IP Address is formed as an IPv4-mapped IPv6 address as specified in . P[i] is a set of 64 Preference values that correspond to the 64 Differentiated Service Code Point (DSCP) values . Each P(i) is set to the value '0' ("disabled"), '1' ("low"), '2' ("medium") or '3' ("high") to indicate a preference level for packet forwarding purposes. AERO interfaces may be configured over multiple underlying interfaces. For example, common mobile handheld devices have both wireless local area network ("WLAN") and cellular wireless links. These links are typically used "one at a time" with low-cost WLAN preferred and highly-available cellular wireless as a standby. In a more complex example, aircraft frequently have many wireless data link types (e.g. satellite-based, cellular, terrestrial, air-to-air directional, etc.) with diverse performance and cost properties. If a Client's multiple underlying interfaces are used "one at a time" (i.e., all other interfaces are in standby mode while one interface is active), then IPv6 ND messages include only a single S/TLLAO with Interface ID set to a constant value. If the Client has multiple active underlying interfaces, then from the perspective of IPv6 ND it would appear to have multiple link-layer addresses. In that case, IPv6 ND messages MAY include multiple S/TLLAOs -- each with an Interface ID that corresponds to a specific underlying interface of the AERO node.

When a Relay enables an AERO interface, it first assigns an administratively-provisioned link-local address fe80::ID to the interface. Each fe80::ID address MUST be unique among all AERO nodes on the link, and is taken from the range fe80::/96 but excluding the special addresses fe80:: and fe80::ffff:ffff. The Relay then engages in a dynamic routing protocol session with all Servers on the link (see: ), and advertises its assigned ASPs into the native IP Internetwork. Each Relay subsequently maintains an IP forwarding table entry for each active ACP covered by its ASP(s), and maintains a neighbor cache entry for each Server on the link. Relays exchange NS/NA messages with AERO link neighbors the same as for any AERO node, however they typically do not perform explicit Neighbor Unreachability Detection (NUD) (see: ) since the dynamic routing protocol already provides reachability confirmation.

When a Server enables an AERO interface, it assigns an administratively-provisioned link-local address fe80::ID the same as for Relays. The Server further configures a DHCPv6 server function to facilitate DHCPv6 PD exchanges with AERO Clients. The Server maintains a neighbor cache entry for each Relay on the link, and manages per-Client neighbor cache entries and IP forwarding table entries based on control message exchanges. Each Server also engages in a dynamic routing protocol with each Relay on the link (see: ). When the Server receives an NS/RS message from a Client on the AERO interface it returns an NA/RA message. The Server further provides a simple link-layer conduit between AERO interface neighbors. In particular, when a packet sent by a source Client arrives on the Server's AERO interface and is destined to another AERO node, the Server forwards the packet at the link layer without ever disturbing the network layer and without ever leaving the AERO interface.

When a Client enables an AERO interface, it uses the special administratively-provisioned link-local address fe80::ffff:ffff as the source network-layer address in DHCPv6 PD messages to obtain one or more ACPs from an AERO Server. Next, the Client assigns the base AERO address to the AERO interface and sends an RS to the Server to receive an RA. In this way, the DHCPv6 PD exchange securely bootstraps autoconfiguration of unique link-local address(es) while the RS/RA exchange establishes link-layer addresses and autoconfigures AERO link parameters. The Client maintains a neighbor cache entry for each of its Servers and each of its active correspondent Clients. When the Client receives IPv6 ND messages on the AERO interface it updates or creates neighbor cache entries, including link-layer address information.

Each AERO interface maintains a conceptual neighbor cache that includes an entry for each neighbor it communicates with on the AERO link, the same as for any IPv6 interface . AERO interface neighbor cache entires are said to be one of "permanent", "static" or "dynamic". Permanent neighbor cache entries are created through explicit administrative action; they have no timeout values and remain in place until explicitly deleted. AERO Relays maintain a permanent neighbor cache entry for each Server on the link, and AERO Servers maintain a permanent neighbor cache entry for each Relay. Each entry maintains the mapping between the neighbor's fe80::ID network-layer address and corresponding link-layer address. Static neighbor cache entries are created and maintained through DHCPv6 PD and IPv6 ND exchanges as specified in , and remain in place for durations bounded by prefix delegation lifetimes. AERO Servers maintain static neighbor cache entries for the ACPs of each of their associated Clients, and AERO Clients maintain a static neighbor cache entry for each of their associated Servers. When an AERO Server delegates prefixes via DHCPv6 PD, it creates a static neighbor cache entry for the Client using the Client's base AERO address as the network-layer address and associates all of the Client's other AERO addresses with the neighbor cache entry. When the Client receives the prefix delegation, it creates a static neighbor cache entry for the Server based on the DHCPv6 Reply message link-local source address as the network-layer address and the encapsulation IP source address and UDP source port number as the link-layer address. The Client then sends an RS message to inform the Server of its link-layer addresses and to solicit an RA. When the Server returns an RA message, the Client uses the autoconfiguration information in the RA message to configure AERO interface parameters. Dynamic neighbor cache entries are created or updated based on receipt of a Predirect/Redirect message as specified in , and are garbage-collected when keepalive timers expire. AERO Clients maintain dynamic neighbor cache entries for each of their active correspondent Clients with lifetimes based on IPv6 ND messaging constants. When an AERO Client receives a valid Predirect message it creates or updates a dynamic neighbor cache entry for the Predirect target network-layer and link-layer addresses. The node then sets an "AcceptTime" variable in the neighbor cache entry to ACCEPT_TIME seconds and uses this value to determine whether packets received from the correspondent can be accepted. When an AERO Client receives a valid Redirect message it creates or updates a dynamic neighbor cache entry for the Redirect target network-layer and link-layer addresses. The Client then sets a "ForwardTime" variable in the neighbor cache entry to FORWARD_TIME seconds and uses this value to determine whether packets can be sent directly to the correspondent. The Client also sets a "MaxRetry" variable to MAX_RETRY to limit the number of keepalives sent when a correspondent may have gone unreachable. It is RECOMMENDED that FORWARD_TIME be set to the default constant value 30 seconds to match the default REACHABLE_TIME value specified for IPv6 ND . It is RECOMMENDED that ACCEPT_TIME be set to the default constant value 40 seconds to allow a 10 second window so that the AERO redirection procedure can converge before AcceptTime decrements below FORWARD_TIME. It is RECOMMENDED that MAX_RETRY be set to 3 the same as described for IPv6 ND address resolution in Section 7.3.3 of . Different values for FORWARD_TIME, ACCEPT_TIME, and MAX_RETRY MAY be administratively set, if necessary, to better match the AERO link's performance characteristics; however, if different values are chosen, all nodes on the link MUST consistently configure the same values. Most importantly, ACCEPT_TIME SHOULD be set to a value that is sufficiently longer than FORWARD_TIME to allow the AERO redirection procedure to converge. When there may be a Network Address Translator (NAT) between the Client and the Server, or if the path from the Client to the Server should be tested for reachability, the Client can send periodic RS messages to the Server to receive RA replies. The RS/RA messaging will keep NAT state alive and test Server reachability without disturbing the DHCPv6 server.

IP packets enter a node's AERO interface either from the network layer (i.e., from a local application or the IP forwarding system) or from the link layer (i.e., from the AERO tunnel virtual link). Packets that enter the AERO interface from the network layer are encapsulated and forwarded into the AERO link, i.e., they are tunnelled to an AERO interface neighbor. Packets that enter the AERO interface from the link layer are either re-admitted into the AERO link or forwarded to the network layer where they are subject to either local delivery or IP forwarding. In all cases, the AERO interface itself MUST NOT decrement the network layer TTL/Hop-count since its forwarding actions occur below the network layer. AERO interfaces may have multiple underlying interfaces and/or neighbor cache entries for neighbors with multiple Interface ID registrations (see ). The AERO node uses each packet's DSCP value to select an outgoing underlying interface based on the node's own preference values, and also to select a destination link-layer address based on the neighbor's underlying interface with the highest preference value. If multiple outgoing interfaces and/or neighbor interfaces have a preference of "high", the AERO node sends one copy of the packet via each of the (outgoing interface, neighbor interface) pairs; otherwise, the node sends a single copy of the packet. The following sections discuss the AERO interface forwarding algorithms for Clients, Servers and Relays. In the following discussion, a packet's destination address is said to "match" if it is a non-link-local address with a prefix covered by an ASP/ACP, or if it is an AERO address that embeds an ACP, or if it is the same as an administratively-provisioned link-local address.

When an IP packet enters a Client's AERO interface from the network layer the Client searches for a neighbor cache entry that matches the destination. If there is a match, the Client uses one or more link-layer addresses in the entry as the link-layer addresses for encapsulation and admits the packet into the AERO link. Otherwise, the Client uses the link-layer address in a static neighbor cache entry for a Server as the encapsulation address. When an IP packet enters a Client's AERO interface from the link-layer, if the destination matches one of the Client's ACPs or link-local addresses the Client decapsulates the packet and delivers it to the network layer. Otherwise, the Client drops the packet silently.

When an IP packet enters a Server's AERO interface from the network layer, the Server searches for a static or dynamic neighbor cache entry that matches the destination. If there is a match, the Server uses one or more link-layer addresses in the entry as the link-layer addresses for encapsulation and admits the packet into the AERO link. Otherwise, the Server uses the link-layer address in a permanent neighbor cache entry for a Relay (selected through longest-prefix match) as the link-layer address for encapsulation. When an IP packet enters a Server's AERO interface from the link layer, the Server processes the packet as follows: if the destination matches one of the Server's link-local addresses the Server decapsulates the packet and forwards it to the network layer for local delivery. else, if the destination matches a dynamic neighbor cache entry the Server first determines whether the neighbor is the same as the one it received the packet from. If so, the Server MUST drop the packet silently to avoid looping; otherwise, the Server uses the neighbor's link-layer address(es) as the destination for encapsulation and re-admits the packet into the AERO link. else, the Server uses the link-layer address in a permanent neighbor cache entry for a Relay (selected through longest-prefix match) as the link-layer address for encapsulation.

When an IP packet enters a Relay's AERO interface from the network layer, the Relay searches its IP forwarding table for an ACP entry that matches the destination and otherwise searches for a neighbor cache entry that matches the destination. If there is a match, the Relay uses the link-layer address in the corresponding neighbor cache entry as the link-layer address for encapsulation and forwards the packet into the AERO link. Otherwise, the Relay drops the packet and (for non-link-local addresses) returns an ICMP Destination Unreachable message subject to rate limiting (see: ). When an IP packet enters a Relay's AERO interface from the link-layer, the Relay processes the packet as follows: if the destination does not match an ASP, or if the destination matches one of the Relay's link-local addresses, the Relay decapsulates the packet and forwards it to the network layer where it will be subject to either local delivery or IP forwarding. else, if the destination matches an ACP entry in the IP forwarding table, or if the destination matches the link-local address in a permanent neighbor cache entry, the Relay first determines whether the neighbor is the same as the one it received the packet from. If so the Relay MUST drop the packet silently to avoid looping; otherwise, the Relay uses the neighbor's link-layer address as the destination for encapsulation and re-admits the packet into the AERO link. else, the Relay drops the packet and (for non-link-local addresses) returns an ICMP Destination Unreachable message subject to rate limiting (see: ).

AERO interfaces encapsulate IP packets according to whether they are entering the AERO interface from the network layer or if they are being re-admitted into the same AERO link they arrived on. This latter form of encapsulation is known as "re-encapsulation". The AERO interface encapsulates packets per the Generic UDP Encapsulation (GUE) procedures in , or through an alternate encapsulation format (see: ). For packets entering the AERO interface from the network layer, the AERO interface copies the "TTL/Hop Limit", "Type of Service/Traffic Class" , "Flow Label" .(for IPv6) and "Congestion Experienced" values in the packet's IP header into the corresponding fields in the encapsulation IP header. For packets undergoing re-encapsulation, the AERO interface instead copies these values from the original encapsulation IP header into he new encapsulation header, i.e., the values are transferred between encapsulation headers and *not* copied from the encapsulated packet's network-layer header. (Note especially that by copying the TTL/Hop Limit between encapsulation headers the value will eventually decrement to 0 if there is a (temporary) routing loop.) For IPv4 encapsulation/re-encapsulation, the AERO interface sets the DF bit as discussed in . When GUE encapsulation is used, the AERO interface next sets the UDP source port to a constant value that it will use in each successive packet it sends, and sets the UDP length field to the length of the encapsulated packet plus 8 bytes for the UDP header itself plus the length of the GUE header (or 0 if GUE direct IP encapsulation is used). For packets sent to a Server or Relay, the AERO interface sets the UDP destination port to 8060, i.e., the IANA-registered port number for AERO. For packets sent to a Client, the AERO interface sets the UDP destination port to the port value stored in the neighbor cache entry for this Client. The AERO interface then either includes or omits the UDP checksum according to the GUE specification.

AERO interfaces decapsulate packets destined either to the AERO node itself or to a destination reached via an interface other than the AERO interface the packet was received on. Decapsulation is per the procedures specified for the appropriate encapsulation format.

AERO nodes employ simple data origin authentication procedures for encapsulated packets they receive from other nodes on the AERO link. In particular: AERO Servers and Relays accept encapsulated packets with a link-layer source address that matches a permanent neighbor cache entry. AERO Servers accept authentic encapsulated DHCPv6 and IPv6 ND messages from Clients, and create or update a static neighbor cache entry for the Client based on the specific message type. AERO Clients and Servers accept encapsulated packets if there is a static neighbor cache entry with a link-layer address that matches the packet's link-layer source address. AERO Clients and Servers accept encapsulated packets if there is a dynamic neighbor cache entry with an AERO address that matches the packet's network-layer (link-local or non-link-local) source address, with a link-layer address that matches the packet's link-layer source address, and with a non-zero AcceptTime. Note that this simple data origin authentication is effective in environments in which link-layer addresses cannot be spoofed. In other environments, each AERO message must include a signature that the recipient can use to authenticate the message origin, e.g., as for common VPN systems such as OpenVPN . In environments where end systems use end-to-end security, however, it may be sufficient to require signatures only for AERO DHCPv6, IPv6 ND and ICMP control plane messages and omit signatures for data plane messages.

The AERO interface is the node's attachment to the AERO link. The AERO interface acts as a tunnel ingress when it sends a packet to an AERO link neighbor and as a tunnel egress when it receives a packet from an AERO link neighbor. AERO interfaces observe the packet sizing considerations for tunnels discussed in and as specified below. IPv6 specifies a minimum link Maximum Transmission Unit (MTU) of 1280 bytes . Although IPv4 specifies a smaller minimum link MTU of 68 bytes , AERO interfaces also observe the IPv6 minimum for IPv4 even if the packet may incur fragmentation in the network. IPv6 specifies a minimum Maximum Reassembly Unit (MRU) of 1500 bytes , while the minimum MRU for IPv4 is only 576 bytes (note that common IPv6 over IPv4 tunnels already assume a larger MRU than the IPv4 minimum). Original sources expect that IP packets will either be delivered to the final destination or a suitable Packet Too Big (PTB) message returned. However, PTB messages may be crafted for malicious purposes such as denial of service, or lost in the network resulting in failure of the IP Path MTU Discovery (PMTUD) mechanisms . For these reasons, AERO links employ operational procedures that avoid all interactions with PMTUD. AERO Servers advertise an MTU that MUST NOT be smaller than 1280 bytes, MUST NOT be larger than the minimum MRU among all nodes on the AERO link minus the encapsulation overhead ("ENCAPS"), and SHOULD NOT be smaller than 1500 bytes. AERO Servers advertise a Maximum Fragment Unit (MFU) as the maximum size for the fragments of an encapsulated packet that require fragmentation. The MFU value MUST NOT be larger than (MTU+ENCAPS) and MUST NOT be larger than 1280 bytes unless there is operational assurance that a larger size can traverse the link along all paths without fragmentation. All AERO nodes MUST set the AERO interface MTU/MFU according to the values advertised by Servers. All AERO nodes MUST configure the same MTU/MFU values for reasons cited in (in particular, multicast support requires a common MTU value among all nodes on the link). All AERO nodes MUST configure an MRU large enough to reassemble packets up to (MTU+ENCAPS) bytes in length; nodes that cannot configure a large-enough MRU MUST NOT enable an AERO interface. In accordance with these requirements, the ingress accommodates packets of various sizes as follows: First, for each original IPv4 packet that is larger than the AERO interface MTU and with the DF bit set to 0, the ingress uses IPv4 fragmentation to break the packet into a minimum number of non-overlapping fragments where the first fragment is no larger than (MFU-ENCAPS) bytes and the remaining fragments are no larger than the first. Next, for each original IP packet or fragment that is no larger than (MFU-ENCAPS) bytes, the ingress encapsulates the packet and admits it into the tunnel. For IPv4 encapsulation, the ingress sets the Don't Fragment (DF) bit to 0 so that these packets will be delivered to the egress even if some fragmentation occurs in the network. For all other original IP packets or fragments, if the packet is larger than the AERO interface MTU, the ingress drops the packet and returns a PTB message to the original source. Otherwise, the ingress encapsulates the packet and fragments the encapsulated packet into a minimum number of non-overlapping fragments where the first fragment is no larger than MFU bytes and the remaining fragments are no larger than the first. The ingress then admits the fragments into the tunnel, and for IPv4 sets the DF bit to 0 in the IP encapsulation header. These fragmented encapsulated packets will be delivered to the egress, which reassembles them into a whole packet. Several factors must be considered when fragmentation of the encapsulated packet is needed. For AERO links over IPv4, the IP ID field is only 16 bits in length, meaning that fragmentation at high data rates could result in data corruption due to reassembly misassociations . For AERO links over both IPv4 and IPv6, studies have also shown that IP fragments are dropped unconditionally over some network paths [I-D.taylor-v6ops-fragdrop]. In environments where IP fragmentation issues could result in operational problems, the ingress SHOULD employ intermediate-layer fragmentation (see: and ) before appending the outer encapsulation headers to each fragment. Since the encapsulation fragment header reduces the room available for packet data, but the original source has no way to control its insertion, the ingress MUST include the fragment header length in the ENCAPS length even for packets in which the header is absent.

When an AERO node admits encapsulated packets into the AERO interface, it may receive link-layer or network-layer error indications. A link-layer error indication is an ICMP error message generated by a router on the path to the neighbor or by the neighbor itself. The message includes an IP header with the address of the node that generated the error as the source address and with the link-layer address of the AERO node as the destination address. The IP header is followed by an ICMP header that includes an error Type, Code and Checksum. Valid type values include "Destination Unreachable", "Time Exceeded" and "Parameter Problem" . (AERO interfaces ignore all link-layer IPv4 "Fragmentation Needed" and IPv6 "Packet Too Big" messages since they only emit packets that are guaranteed to be no larger than the IP minimum link MTU as discussed in .) The ICMP header is followed by the leading portion of the packet that generated the error, also known as the "packet-in-error". For ICMPv6, specifies that the packet-in-error includes: "As much of invoking packet as possible without the ICMPv6 packet exceeding the minimum IPv6 MTU" (i.e., no more than 1280 bytes). For ICMPv4, specifies that the packet-in-error includes: "Internet Header + 64 bits of Original Data Datagram", however Section 4.3.2.3 updates this specification by stating: "the ICMP datagram SHOULD contain as much of the original datagram as possible without the length of the ICMP datagram exceeding 576 bytes". The link-layer error message format is shown in (where, "L2" and "L3" refer to link-layer and network-layer, respectively):

The AERO node rules for processing these link-layer error messages are as follows: When an AERO node receives a link-layer Parameter Problem message, it processes the message the same as described as for ordinary ICMP errors in the normative references . When an AERO node receives persistent link-layer Time Exceeded messages, the IP ID field may be wrapping before earlier fragments awaiting reassembly have been processed. In that case, the node SHOULD begin including integrity checks and/or institute rate limits for subsequent packets. When an AERO node receives persistent link-layer Destination Unreachable messages in response to encapsulated packets that it sends to one of its dynamic neighbor correspondents, the node SHOULD test the path to the correspondent using Neighbor Unreachability Detection (NUD) (see ). If NUD fails, the node SHOULD set ForwardTime for the corresponding dynamic neighbor cache entry to 0 and allow future packets destined to the correspondent to flow through a default route. When an AERO Client receives persistent link-layer Destination Unreachable messages in response to encapsulated packets that it sends to one of its static neighbor Servers, the Client SHOULD test the path to the Server using NUD. If NUD fails, the Client SHOULD associate with a new Server and send a DHCPv6 Release message to the old Server as specified in . When an AERO Server receives persistent link-layer Destination Unreachable messages in response to encapsulated packets that it sends to one of its static neighbor Clients, the Server SHOULD test the path to the Client using NUD. If NUD fails, the Server SHOULD cancel the DHCPv6 PD for the Client's ACP, withdraw its route for the ACP from the AERO routing system and delete the neighbor cache entry (see and ). When an AERO Relay or Server receives link-layer Destination Unreachable messages in response to an encapsulated packet that it sends to one of its permanent neighbors, it treats the messages as an indication that the path to the neighbor may be failing. However, neighbor reachability will be determined by the dynamic routing protocol. When an AERO Relay receives a packet for which the network-layer destination address is covered by an ASP, if there is no more-specific routing information for the destination the Relay drops the packet and returns a network-layer Destination Unreachable message subject to rate limiting. The Relay first writes the network-layer source address of the original packet as the destination address of the message and determines the next hop to the destination. If the next hop is reached via the AERO interface, the Relay uses the IPv6 address "::" or the IPv4 address "0.0.0.0" as the source address of the message, then encapsulates the message and forwards it to the next hop within the AERO interface. Otherwise, the Relay uses one of its non link-local addresses as the source address of the message and forwards it via a link outside the AERO interface. When an AERO node receives an encapsulated packet for which the reassembly buffer it too small, it drops the packet and returns an network-layer Packet Too Big (PTB) message. The node first writes the MRU value into the PTB message MTU field, writes the network-layer source address of the original packet as the destination address of the message and determines the next hop to the destination. If the next hop is reached via the AERO interface, the node uses the IPv6 address "::" or the IPv4 address "0.0.0.0" as the source address of the message, then encapsulates the message and forwards it to the next hop within the AERO interface. Otherwise, the node uses one of its non link-local addresses as the source address of the message and forwards it via a link outside the AERO interface. When an AERO node receives any network-layer error message via the AERO interface, it examines the network-layer destination address. If the next hop toward the destination is via the AERO interface, the node re-encapsulates and forwards the message to the next hop within the AERO interface. Otherwise, if the network-layer source address is the IPv6 address "::" or the IPv4 address "0.0.0.0", the node writes one of its non link-local addresses as the source address, recalculates the IP and/or ICMP checksums then forwards the message via a link outside the AERO interface.

AERO Router Discovery, Prefix Delegation and Autoconfiguration are coordinated by the DHCPv6 and IPv6 ND control messaging protocols as discussed in the following Sections.

Each AERO Server configures a DHCPv6 server function to facilitate PD requests from Clients. Each Server is provisioned with a database of ACP-to-Client ID mappings for all Clients enrolled in the AERO system, as well as any information necessary to authenticate each Client. The Client database is maintained by a central administrative authority for the AERO link and securely distributed to all Servers, e.g., via the Lightweight Directory Access Protocol (LDAP) , via static configuration, etc. Therefore, no Server-to-Server DHCPv6 PD state synchronization is necessary, and Clients can optionally hold separate PDs for the same ACPs from multiple Servers. In this way, Clients can associate with multiple Servers, and can receive new PDs from new Servers before deprecating PDs received from existing Servers. This provides the Client with a natural fault-tolerance and/or load balancing profile. AERO Clients and Servers use unicast IPv6 ND messages to maintain neighbor cache entries the same as for any link. AERO Servers act as default routers for AERO Clients, and therefore send unicast RA messages with configuration information in response to a Client's RS message. The following sections specify the Client and Server behavior.

AERO Clients discover the link-layer addresses of AERO Servers via static configuration (e.g., from a flat-file map of Server addresses and locations), or through an automated means such as DNS name resolution. In the absence of other information, the Client resolves the FQDN "linkupnetworks.[domainname]" where "linkupnetworks" is a constant text string and "[domainname]" is a DNS suffix for the Client's underlying network (e.g., "example.com"). After discovering the link-layer addresses, the Client associates with one or more of the corresponding Servers. To associate with a Server, the Client acts as a requesting router to request ACPs through a DHCPv6 PD exchange . The Client's DHCPv6 Solicit message includes fe80::ffff:ffff as the IPv6 source address, 'All_DHCP_Relay_Agents_and_Servers' as the IPv6 destination address, the address of the Client's underlying interface as the link-layer source address and the link-layer address of the Server as the link-layer destination address. The Client also includes a Client Identifier option with the Client's DUID, and an Identity Association for Prefix Delegation (IA_PD) option. If the Client is pre-provisioned with ACPs associated with the AERO service, it MAY also include the ACPs in the IA_PD to indicate its preferences to the DHCPv6 server. The Client finally includes any additional DHCPv6 options (including any necessary authentication options to identify itself to the DHCPv6 server), and sends the encapsulated Solicit message via any available underlying interface. When the Client attempts to perform a DHCPv6 PD exchange with a Server that is too busy to service the request, the Client may receive an error status code such as "NoPrefixAvail" in the Server's Reply or no Reply at all. In that case, the Client SHOULD discontinue DHCPv6 PD attempts through this Server and try another Server. When the Client receives a Reply from the AERO Server it creates a static neighbor cache entry with the Server's link-local address as the network-layer address and the Server's encapsulation address as the link-layer address. Next, the Client autoconfigures AERO addresses for each of the delegated ACPs and assigns the base AERO address to the AERO interface. The Client then prepares a unicast RS message to send to the Server in order to obtain a solicited RA. The Client includes its base AERO address as the network-layer source address, the Server's link-local address as the network-layer destination address, the Client's link-layer address as the link-layer source address, and Server's link-layer address as the link-layer destination address. The Client also includes one or more SLLAOs formatted as described in to register its link-layer address(es) with the Server. The first SLLAO MUST correspond to the underlying interface over which the Client will send the RS. The Client MAY include additional SLLAOs specific to other underlying interfaces, but if so it MUST have assurance that there will be no NATs on the paths to the Server via those interfaces (otherwise, the Client can register additional link-layer addresses with the Server by sending subsequent unsolicited NA messages after the initial RS/RA exchange). The Server will use the S/TLLAOs to populate its link-layer address information for the Client. When the Client receives an RA from the AERO Server (see ), it configures a default route with the Server as the next hop via the AERO interface. The Client also caches any ASPs included in Prefix Information Options (PIOs) as ASPs to associate with the AERO link, and assigns the MTU/MFU values in the MTU options to its AERO interface while configuring an appropriate MRU. This configuration information applies to the AERO link as a whole, and all AERO nodes will use the same values.. Following autoconfiguration, the Client sub-delegates the ACPs to its attached EUNs and/or the Client's own internal virtual interfaces. In the former case, the Client acts as a router for nodes on its attached EUNs. In the latter case, the Client acts as a host and can configure as many addresses as it wants from /64 prefixes taken from the ACPs and assign them to either an internal virtual interface ("weak end-system") or to the AERO interface itself ("strong end-system") while black-holing the remaining portions of the /64s. The Client subsequently renews its ACP delegations through each of its Servers by sending DHCPv6 Renew messages. After the Client registers its Interface IDs and their associated 'P(i)' values, it may wish to change one or more Interface ID registrations, e.g., if an underlying interface becomes unavailable, if cost profiles change, etc. To do so, the Client prepares an unsolicited NA message to send over any available underlying interface. The NA MUST include a S/TLLAO specific to the selected available underlying interface as the first S/TLLAO and MAY include any additional S/TLLAOs specific to other underlying interfaces. The Client includes fresh 'P(i)' values in each S/TLLAO to update the Server's neighbor cache entry. If the Client wishes to disable some or all DSCPs for an underlying interface, it includes an S/TLLAO with 'P(i)' values set to 0 ("disabled"). If the Client wishes to discontinue use of a Server it issues a DHCPv6 Release message to both delete the Server's neighbor cache entry and release the DHCPv6 PD.

AERO Servers configure a DHCPv6 server function on their AERO links. AERO Servers arrange to add their encapsulation layer IP addresses (i.e., their link-layer addresses) to a static map of Server addresses for the link and/or the DNS resource records for the FQDN "linkupnetworks.[domainname]" before entering service. When an AERO Server receives a prospective Client's Solicit on its AERO interface, and the Server is too busy to service the message, it SHOULD return a Reply with status code "NoPrefixAvail" per . Otherwise, the Server authenticates the message. If authentication succeeds, the Server determines the correct ACPs to delegate to the Client by searching the Client database. Next, the Server prepares a Reply message to send to the Client while using fe80::ID as the network-layer source address, the link-local address taken from the Client's Solicit as the network-layer destination address, the Server's link-layer address as the source link-layer address, and the Client's link-layer address as the destination link-layer address. The Server also includes an IA_PD option with the delegated ACPs. For IPv4 ACPs, the prefix included in the IA_PD option is in IPv4-mapped IPv6 address format and with prefix length set as specified in . When the Server sends the Reply message, it creates a static neighbor cache entry for the Client using the base AERO address as the network-layer address and with lifetime set to no more than the smallest PD lifetime. The Client will subsequently issue an RS message with one or more SLLAO options and with the Client's base AERO address as the source address. When the Server receives the RS message, it first verifies that a neighbor cache entry for the Client exists (otherwise, it discards the RS). The Server then updates the neighbor cache entry link-layer address(es) by recording the information in each SLLAO option indexed by the Interface ID and including the UDP port number, IP address and P(i) values. For the first SLLAO in the list, however, the Server records the actual encapsulation source UDP and IP addresses instead of those that appear in the SLLAO in case there was a NAT in the path. The Server then prepares a unicast RA message to send back to the Client using fe80::ID as the network-layer source address, the Client's base AERO address as the network-layer destination address, the Server's link-layer address as the source link-layer address, and the source link-layer address of the RS message as the destination link-layer address. In the RA message, the Server includes one or more PIOs that encode the ASPs for the AERO link, and with flags A=0; L=1. The Server also includes two MTU options - the first MTU option includes the MTU for the link and the second MTU option includes the MFU for the link (see ). When the Server delegates the ACPs, it also creates an IP forwarding table entry for each ACP so that the AERO BGP-based routing system will propagate the ACPs to all Relays that aggregate the corresponding ASP (see: ). After the initial DHCPv6 PD Solicit/Reply and IPv6 ND RS/RA exchanges, the AERO Server maintains the neighbor cache entry for the Client until the PD lifetimes expire. If the Client issues a Renew, the Server extends the PD lifetimes. If the Client issues a Release, or if the Client does not issue a Renew before the lifetime expires, the Server deletes the neighbor cache entry for the Client and withdraws the IP routes from the AERO routing system.

AERO Clients and Servers are always on the same link (i.e., the AERO link) from the perspective of DHCPv6. However, in some implementations the DHCPv6 server and AERO interface driver may be located in separate modules. In that case, the Server's AERO interface driver module can act as a Lightweight DHCPv6 Relay Agent (LDRA) to relay DHCPv6 messages to and from the DHCPv6 server module. When the LDRA receives a DHCPv6 message from a Client addressed to either 'All_DHCP_Relay_Agents_and_Servers' or the Server's fe80::ID unicast address, it wraps the message in a Relay-Forward message header and includes an Interface-ID option that includes enough information to allow the LDRA to forward the resulting Reply message back to the Client (this information may include the Client's UDP and IP addresses, a security association identifier, etc). The LDRA then forwards the message to the DHCPv6 server. When the DHCPv6 server prepares a Reply message, it wraps the message in a Relay-Reply message and echoes the Interface-ID option. The DHCPv6 server then delivers the Relay-Reply message to the LDRA, which discards the Relay-Reply wrapper and delivers the DHCPv6 message to the Client based on the information in the Interface ID option.

When a source Client forwards packets to a prospective correspondent Client within the same AERO link domain (i.e., one for which the packet's destination address is covered by an ASP), the source Client MAY initiate an AERO link route optimization procedure. The procedure is based on an exchange of IPv6 ND messages using a chain of AERO Servers and Relays as a trust basis. Although the Client is responsible for initiating route optimization, the Server is the policy enforcement point that determines whether route optimization is permitted. For example, on some AERO links route optimization would allow traffic to circumvent critical network-based traffic interception points. In those cases, the Server can simply discard any route optimization messages instead of forwarding them. The following sections specify the AERO link route optimization procedure.

depicts the AERO link route optimization reference operational scenario, using IPv6 addressing as the example (while not shown, a corresponding example for IPv4 addressing can be easily constructed). The figure shows an AERO Relay ('R1'), two AERO Servers ('S1', 'S2'), two AERO Clients ('C1', 'C2') and two ordinary IPv6 hosts ('H1', 'H2'):

In , Relay ('R1') assigns the administratively-provisioned link-local address fe80::1 to its AERO interface with link-layer address L2(R1), Server ('S1') assigns the address fe80::2 with link-layer address L2(S1),and Server ('S2') assigns the address fe80::3 with link-layer address L2(S2). Servers ('S1') and ('S2') next arrange to add their link-layer addresses to a published list of valid Servers for the AERO link. AERO Client ('C1') receives the ACP 2001:db8:0::/48 in a DHCPv6 PD exchange via AERO Server ('S1') then assigns the address fe80::2001:db8:0:0 to its AERO interface with link-layer address L2(C1). Client ('C1') configures a default route and neighbor cache entry via the AERO interface with next-hop address fe80::2 and link-layer address L2(S1), then sub-delegates the ACP to its attached EUNs. IPv6 host ('H1') connects to the EUN, and configures the address 2001:db8:0::1. AERO Client ('C2') receives the ACP 2001:db8:1::/48 in a DHCPv6 PD exchange via AERO Server ('S2') then assigns the address fe80::2001:db8:1:0 to its AERO interface with link-layer address L2(C2). Client ('C2') configures a default route and neighbor cache entry via the AERO interface with next-hop address fe80::3 and link-layer address L2(S2), then sub-delegates the ACP to its attached EUNs. IPv6 host ('H2') connects to the EUN, and configures the address 2001:db8:1::1.

Again, with reference to , when source host ('H1') sends a packet to destination host ('H2'), the packet is first forwarded over the source host's attached EUN to Client ('C1'). Client ('C1') then forwards the packet via its AERO interface to Server ('S1') and also sends a Predirect message toward Client ('C2') via Server ('S1'). Server ('S1') then re-encapsulates and forwards both the packet and the Predirect message out the same AERO interface toward Client ('C2') via Relay ('R1'). When Relay ('R1') receives the packet and Predirect message, it consults its forwarding table to discover Server ('S2') as the next hop toward Client ('C2'). Relay ('R1') then forwards both the packet and the Predirect message to Server ('S2'), which then forwards them to Client ('C2'). After Client ('C2') receives the Predirect message, it process the message and returns a Redirect message toward Client ('C1') via Server ('S2'). During the process, Client ('C2') also creates or updates a dynamic neighbor cache entry for Client ('C1'). When Server ('S2') receives the Redirect message, it re-encapsulates the message and forwards it on to Relay ('R1'), which forwards the message on to Server ('S1') which forwards the message on to Client ('C1'). After Client ('C1') receives the Redirect message, it processes the message and creates or updates a dynamic neighbor cache entry for Client ('C2'). Following the above Predirect/Redirect message exchange, forwarding of packets from Client ('C1') to Client ('C2') without involving any intermediate nodes is enabled. The mechanisms that support this exchange are specified in the following sections.

AERO Redirect/Predirect messages use the same format as for IPv6 ND Redirect messages depicted in Section 4.5 of . AERO Redirect/Predirect messages formats are identical except that Redirect messages use Code=0, while Predirect messages use Code=1. The Redirect/Predirect message format is shown in :

When a Client forwards a packet with a source address from one of its ACPs toward a destination address covered by an ASP (i.e., toward another AERO Client connected to the same AERO link), the source Client MAY send a Predirect message forward toward the destination Client via the Server. In the reference operational scenario, when Client ('C1') forwards a packet toward Client ('C2'), it MAY also send a Predirect message forward toward Client ('C2'), subject to rate limiting (see Section 8.2 of ). Client ('C1') prepares the Predirect message as follows: the link-layer source address is set to 'L2(C1)' (i.e., the link-layer address of Client ('C1')). the link-layer destination address is set to 'L2(S1)' (i.e., the link-layer address of Server ('S1')). the network-layer source address is set to fe80::2001:db8:0:0 (i.e., the base AERO address of Client ('C1')). the network-layer destination address is set to the AERO address corresponding to the destination address of Client ('C2'). the Type is set to 137. the Code is set to 1 to indicate "Predirect". the Target Address is set to fe80::2001:db8:0:0 (i.e., the base AERO address of Client ('C1')). the Destination Address is set to the source address of the originating packet that triggered the Predirection event. (If the originating packet is an IPv4 packet, the address is constructed in IPv4-mapped IPv6 address format). the message includes one or more TLLAOs set to appropriate values for Client ('C1')'s underlying interfaces. the message includes one or more Route Information Options (RIOs) that include Client ('C1')'s ACPs. the message SHOULD include a Timestamp option and a Nonce option. the message includes a Redirected Header Option (RHO) that contains the originating packet truncated if necessary to ensure that at least the network-layer header is included but the size of the message does not exceed 1280 bytes. Note that the act of sending Predirect messages is cited as "MAY", since Client ('C1') may have advanced knowledge that the direct path to Client ('C2') would be unusable or otherwise undesirable. If the direct path later becomes unusable after the initial route optimization, Client ('C1') simply allows packets to again flow through Server ('S1').

When Server ('S1') receives a Predirect message from Client ('C1'), it first verifies that the TLLAOs in the Predirect are a proper subset of the Interface IDs in Client ('C1')'s neighbor cache entry. If the Client's TLLAOs are not acceptable, Server ('S1') discards the message. Otherwise, Server ('S1') validates the message according to the Redirect message validation rules in Section 8.1 of , except that the Predirect has Code=1. Server ('S1') also verifies that Client ('C1') is authorized to use the ACPs encoded in the RIOs of the Predirect. If validation fails, Server ('S1') discards the Predirect; otherwise, it copies the correct UDP Port number and IP Address for Client ('C1')'s underlying link into the first TLLAO in case the addresses have been subject to NAT. Server ('S1') then examines the network-layer destination address of the Predirect to determine the next hop toward Client ('C2') by searching for the AERO address in the neighbor cache. Since Client ('C2') is not one of its neighbors, Server ('S1') re-encapsulates the Predirect and relays it via Relay ('R1') by changing the link-layer source address of the message to 'L2(S1)' and changing the link-layer destination address to 'L2(R1)'. Server ('S1') finally forwards the re-encapsulated message to Relay ('R1') without decrementing the network-layer TTL/Hop Limit field. When Relay ('R1') receives the Predirect message from Server ('S1') it determines that Server ('S2') is the next hop toward Client ('C2') by consulting its forwarding table. Relay ('R1') then re-encapsulates the Predirect while changing the link-layer source address to 'L2(R1)' and changing the link-layer destination address to 'L2(S2)'. Relay ('R1') then relays the Predirect via Server ('S2'). When Server ('S2') receives the Predirect message from Relay ('R1') it determines that Client ('C2') is a neighbor by consulting its neighbor cache. Server ('S2') then re-encapsulates the Predirect while changing the link-layer source address to 'L2(S2)' and changing the link-layer destination address to 'L2(C2)'. Server ('S2') then forwards the message to Client ('C2').

When Client ('C2') receives the Predirect message, it accepts the Predirect only if the message has a link-layer source address of one of its Servers (e.g., L2(S2)). Client ('C2') further accepts the message only if it is willing to serve as a redirection target. Next, Client ('C2') validates the message according to the Redirect message validation rules in Section 8.1 of , except that it accepts the message even though Code=1 and even though the network-layer source address is not that of it's current first-hop router. In the reference operational scenario, when Client ('C2') receives a valid Predirect message, it either creates or updates a dynamic neighbor cache entry that stores the Target Address of the message as the network-layer address of Client ('C1') , stores the link-layer addresses found in the TLLAOs as the link-layer addresses of Client ('C1'), and stores the ACPs encoded in the RIOs of the Predirect as the ACPs for Client ('C1'). Client ('C2') then sets AcceptTime for the neighbor cache entry to ACCEPT_TIME. After processing the message, Client ('C2') prepares a Redirect message response as follows: the link-layer source address is set to 'L2(C2)' (i.e., the link-layer address of Client ('C2')). the link-layer destination address is set to 'L2(S2)' (i.e., the link-layer address of Server ('S2')). the network-layer source address is set to fe80::2001:db8:1:0 (i.e., the base AERO address of Client ('C2')). the network-layer destination address is set to fe80::2001:db8:0:0 (i.e., the base AERO address of Client ('C1')). the Type is set to 137. the Code is set to 0 to indicate "Redirect". the Target Address is set to fe80::2001:db8:1:0 (i.e., the base AERO address of Client ('C2')). the Destination Address is set to the destination address of the originating packet that triggered the Redirection event. (If the originating packet is an IPv4 packet, the address is constructed in IPv4-mapped IPv6 address format). the message includes one or more TLLAOs set to appropriate values for Client ('C2')'s underlying interfaces. the message includes one or more Route Information Options (RIOs) that include Client ('C2')'s ACPs. the message SHOULD include a Timestamp option and MUST echo the Nonce option received in the Predirect (i.e., if a Nonce option is included). the message includes as much of the RHO copied from the corresponding Predirect message as possible such that at least the network-layer header is included but the size of the message does not exceed 1280 bytes. After Client ('C2') prepares the Redirect message, it sends the message to Server ('S2').

When Server ('S2') receives a Redirect message from Client ('C2'), it first verifies that the TLLAOs in the Redirect are a proper subset of the Interface IDs in Client ('C2')'s neighbor cache entry. If the Client's TLLAOs are not acceptable, Server ('S2') discards the message. Otherwise, Server ('S2') validates the message according to the Redirect message validation rules in Section 8.1 of . Server ('S2') also verifies that Client ('C2') is authorized to use the ACPs encoded in the RIOs of the Redirect message. If validation fails, Server ('S2') discards the Redirect; otherwise, it copies the correct UDP Port number and IP Address for Client ('C2')'s underlying link into the first TLLAO in case the addresses have been subject to NAT. Server ('S2') then examines the network-layer destination address of the Redirect to determine the next hop toward Client ('C1') by searching for the AERO address in the neighbor cache. Since Client ('C1') is not a neighbor, Server ('S2') re-encapsulates the Redirect and relays it via Relay ('R1') by changing the link-layer source address of the message to 'L2(S2)' and changing the link-layer destination address to 'L2(R1)'. Server ('S2') finally forwards the re-encapsulated message to Relay ('R1') without decrementing the network-layer TTL/Hop Limit field. When Relay ('R1') receives the Redirect message from Server ('S2') it determines that Server ('S1') is the next hop toward Client ('C1') by consulting its forwarding table. Relay ('R1') then re-encapsulates the Redirect while changing the link-layer source address to 'L2(R1)' and changing the link-layer destination address to 'L2(S1)'. Relay ('R1') then relays the Redirect via Server ('S1'). When Server ('S1') receives the Redirect message from Relay ('R1') it determines that Client ('C1') is a neighbor by consulting its neighbor cache. Server ('S1') then re-encapsulates the Redirect while changing the link-layer source address to 'L2(S1)' and changing the link-layer destination address to 'L2(C1)'. Server ('S1') then forwards the message to Client ('C1').

When Client ('C1') receives the Redirect message, it accepts the message only if it has a link-layer source address of one of its Servers (e.g., 'L2(S1)'). Next, Client ('C1') validates the message according to the Redirect message validation rules in Section 8.1 of , except that it accepts the message even though the network-layer source address is not that of it's current first-hop router. Following validation, Client ('C1') then processes the message as follows. In the reference operational scenario, when Client ('C1') receives the Redirect message, it either creates or updates a dynamic neighbor cache entry that stores the Target Address of the message as the network-layer address of Client ('C2'), stores the link-layer addresses found in the TLLAOs as the link-layer addresses of Client ('C2') and stores the ACPs encoded in the RIOs of the Redirect as the ACPs for Client ('C2').. Client ('C1') then sets ForwardTime for the neighbor cache entry to FORWARD_TIME. Now, Client ('C1') has a neighbor cache entry with a valid ForwardTime value, while Client ('C2') has a neighbor cache entry with a valid AcceptTime value. Thereafter, Client ('C1') may forward ordinary network-layer data packets directly to Client ('C2') without involving any intermediate nodes, and Client ('C2') can verify that the packets came from an acceptable source. (In order for Client ('C2') to forward packets to Client ('C1'), a corresponding Predirect/Redirect message exchange is required in the reverse direction; hence, the mechanism is asymmetric.)

In some environments, the Server nearest the target Client may need to serve as a proxy redirection target, e.g., if direct Client-to-Client communications are not possible. In that case, when the source Client sends a Predirect message the target Server prepares a corresponding Redirect the same as if it were the target Client (see: ), except that it writes its own link-layer address in the TLLAO option. The Server must then maintain a dynamic neighbor cache entry for the redirected source Client. Similarly, when the source Client must send all packets via its own Server and cannot act on a route optimization request, the source Server can send a Predirect message toward the target Client. The target Client then prepares a corresponding Redirect message the same as for Client-to-Client route optimization and sends the Redirect message back to the source Server. Thereafter, if a Client moves to a new Server, the old Server sends ICMP "Destination Unreachable" messages subject to rate limiting in response to data packets received from a correspondent node to report that the route optimization ForwardTime should be set to 0. The correspondent Client (or Server) then allows future packets destined to the departed Client to again flow through its own Server (or Relay). In any case, a Server MUST NOT send a BGP update to its Relays for Clients discovered through dynamic route optimization redirection. BGP updates are only to be sent for the Server's working set of statically-associated Clients.

If neither the source nor target Clients are capable of sending packets other than via their own Servers, a Server-to-Server route optimization can still be employed. In that case, the source Client's Server can send a Predirect message via a Relay to the AERO address of the target Client, and the Relay will forward the message to the target Client's Server. The target Server prepares the Redirect message the same as if it were the target Client, except that it writes its own link-layer address in the TLLAO option then sends a Redirect message back to the source Server via the Relay. Both Servers must then maintain a dynamic neighbor cache entry for the redirected Clients. Thereafter, if a Client moves to a new Server, the old Server sends ICMP "Destination Unreachable" messages subject to rate limiting in response to data packets forwarded by the correspondent Server to report that the route optimization ForwardTime should be set to 0. The correspondent Server then allows future packets destined to the departed Client to again flow through its own Relay. In any case, a Server MUST NOT send a BGP update to its Relays for Clients discovered through dynamic route optimization redirection. BGP updates are only to be sent for the Server's working set of statically-associated Clients..

AERO nodes perform Neighbor Unreachability Detection (NUD) by sending unicast NS messages with SLLAOs to elicit solicited NA messages from neighbors the same as described in . NUD is performed either reactively in response to persistent L2 errors (see ) or proactively to test existing neighbor cache entries and/or update link-layer information. When an AERO node sends an NS/NA message, it MUST use one of its link-local addresses as the IPv6 source address and a link-local address of the neighbor as the IPv6 destination address. When an AERO node receives an NS message or a solicited NA message, it accepts the message if it has a neighbor cache entry for the neighbor; otherwise, it ignores the message. When a source Client is redirected to a target Client it SHOULD proactively test the direct path by sending an initial NS message to elicit a solicited NA response. While testing the path, the source Client can optionally continue sending packets via the Server, maintain a small queue of packets until target reachability is confirmed, or (optimistically) allow packets to flow directly to the target. The source Client SHOULD thereafter periodically test the direct path to the target Client (see Section 7.3 of ) in order to keep dynamic neighbor cache entries alive. If the source Client is unable to elicit a solicited NA response from the target Client after MaxRetry attempts, it SHOULD set ForwardTime to 0 and resume sending packets via one of its Servers. Otherwise, the source Client considers the path usable and SHOULD thereafter process any link-layer errors as a hint that the direct path to the target Client has either failed or has become intermittent. When ForwardTime for a dynamic neighbor cache entry expires, the source Client resumes sending any subsequent packets via a Server and may (eventually) attempt to re-initiate the AERO redirection process. When AcceptTime for a dynamic neighbor cache entry expires, the target Client discards any subsequent packets received directly from the source Client. When both ForwardTime and AcceptTime for a dynamic neighbor cache entry expire, the Client deletes the neighbor cache entry.

When a Client needs to change its link-layer addresses, e.g., due to a mobility event, it sends unsolicited NAs to its neighbors using the new link-layer address as the source address and with TLLAOs that include the updated Client link-layer information. The Client MAY send up to MaxRetry unsolicited NA messages in parallel with sending actual data packets in case one or more NAs are lost. If all NAs are lost, the Client will eventually invoke NUD by sending NS messages that include SLLAOs.

When a Client needs to bring new underlying interfaces into service (e.g., when it activates a new data link), it sends unsolicited NAs to its neighbors using the new link-layer address as the source address and with TLLAOs that include the new Client link-layer information.

When a Client needs to remove existing underlying interfaces from service (e.g., when it de-activates an existing data link), it sends unsolicited NAs to its neighbors with TLLAOs that include P(i) values set to "disabled". If the Client needs to send the unsolicited NAs over a link other than the one being removed from service, it MUST include a TLLAO for the sending link as the first TLLAO and include the TLLAO for the link being removed from service as an additional TLLAO.

AERO interface neighbors MAY include a configuration knob that allows them to perform implicit mobility management in which no IPv6 ND messaging is used. In that case, the Client only transmits packets over a single interface at a time, and the neighbor always observes packets arriving from the Client from the same link-layer source address. If the Client's underlying interface address changes (either due to a readdressing of the original interface or switching to a new interface) the neighbor immediately updates the neighbor cache entry for the Client and begins accepting and sending packets to the Client's new link-layer address. This implicit mobility method applies to use cases such as cellphones with both WiFi and Cellular interfaces where only one of the interfaces is active at a given time, and the Client automatically switches over to the backup interface if the primary interface fails.

When a Client associates with a new Server, it performs the Client procedures specified in . When a Client disassociates with an existing Server, it sends a DHCPv6 Release message via a new Server with its base AERO address as the network-layer source address and the unicast link-local address of the old Server as the network-layer destination address. The new Server then encapsulates the Release message in a DHCPv6 Relay-Forward message header, writes the Client's source address in the 'peer-address' field, and writes its own link-local address in the IP source address (i.e., the new Server acts as a DHCPv6 relay agent). The new Server then forwards the message to a Relay, which forwards the message to the old Server based on the network-layer destination address. When the old Server receives the Release, it first authenticates the message then releases the DHCPv6 PDs and deletes the Client's ACP routes. The old Server then deletes the Client's neighbor cache entry so that any in-flight packets will be forwarded via a Relay to the new Server, which will forward them to the Client. The old Server finally returns a DHCPv6 Relay-Reply message via an AERO Relay to the new Server, which will decapsulate the DHCPv6 Reply message and forward it to the Client. When the new Server forwards the Reply message, the Client can delete both the default route and the neighbor cache entry for the old Server. (Note that since Release/Reply messages may be lost in the network the Client SHOULD retry until it gets a Reply indicating that the Release was successful. If the Client does not receive a Reply after MaxRetry attempts, the old Server may have failed and the Client should discontinue its Release attempts.) Note that this DHCPv6 relay-chaining approach is necessary to avoid failures, e.g., due to temporary routing fluctuations. In particular, the Client should always be able to forward messages via its new Server but may not always be able to send messages directly to an old Server. But, the new Server and Old Server should always be able to exchange messages with one another. Finally, Clients SHOULD NOT move rapidly between Servers in order to avoid causing excessive oscillations in the AERO routing system. Such oscillations could result in intermittent reachability for the Client itself, while causing little harm to the network. Examples of when a Client might wish to change to a different Server include a Server that has gone unreachable, topological movements of significant distance, etc.

AERO Clients and Servers should maintain a small queue of packets they have recently sent to an AERO neighbor, e.g., a 1 second window. If the AERO neighbor moves, the AERO node MAY retransmit the queued packets to ensure that they are delivered to the AERO neighbor's new location. Note that this may have performance implications for asymmetric paths. For example, if the AERO neighbor moves from a 50Mbps link to a 128Kbps link, retransmitting a 1 second window could cause significant congestion. However, any retransmission bursts will subside after the 1 second window.

When the underlying network does not support multicast, AERO Clients map link-scoped multicast addresses to the link-layer address of a Server, which acts as a multicast forwarding agent. The AERO Client also serves as an IGMP/MLD Proxy for its EUNs and/or hosted applications per while using the link-layer address of the Server as the link-layer address for all multicast packets. When the underlying network supports multicast, AERO nodes use the multicast address mapping specification found in for IPv4 underlying networks and use a TBD site-scoped multicast mapping for IPv6 underlying networks. In that case, border routers must ensure that the encapsulated site-scoped multicast packets do not leak outside of the site spanned by the AERO link.