Jive Communications

Quebec QC Canada sperreault@jive.com

France Telecom

Rennes 35000 France mohamed.boucadair@orange.com

Cisco

United States repenno@cisco.com

Cisco Systems, Inc.

170 West Tasman Drive San Jose California 95134 United States dwing@cisco.com

Apple Inc.

1 Infinite Loop Cupertino California 95014 United States +1 408 974 3207 cheshire@apple.com

This document specifies a new Port Control Protocol (PCP) functional element: the PCP proxy. The PCP proxy relays PCP requests received from PCP clients to upstream PCP server(s). A typical deployment usage of this function is to help establish successful PCP communications for PCP clients that cannot be configured with the address of a PCP server located more than one hop away.

This document defines a new Port Control Protocol (PCP) functional element: the PCP proxy. As shown in , the PCP proxy is logically equivalent to a PCP client back-to-back with a PCP server. The "glue" between the two is what is specified in this document. Other than that "glue", the server and the client behave exactly like their regular counterparts. The PCP proxy is responsible for relaying PCP messages received from PCP clients to upstream PCP servers and vice versa. Whether or not the PCP proxy is co-located with a flow-aware function (e.g., NAT, firewall) is deployment specific.

This document assumes a hop-by-hop PCP authentication scheme. That is, referring to , the leftmost PCP client authenticates with the PCP proxy, while the PCP proxy authenticates with the upstream server. Note that in some deployments, PCP authentication may only be enabled between the PCP proxy and an upstream PCP server (e.g., a customer premises host may not authenticate with the PCP proxy, but the PCP proxy may authenticate with the PCP server). The hop-by-hop authentication scheme is more suitable from a deployment standpoint. Furthermore, it allows implementations to easily support a PCP proxy that alters PCP messages (e.g., strips a PCP option, modifies a PCP field).

In today's world, with public routable IPv4 addresses becoming less readily available, it is increasingly common for customers to receive a private address from their Internet Service Provider (ISP), and the ISP uses a NAT gateway of its own to translate those packets before sending them out onto the public Internet. This means that there is likely to be more than one NAT on the path between client machines and the public Internet: If a residential customer receives a translated address from their ISP and then installs their own residential NAT gateway to share that address between multiple client devices in their home, then there are at least two NAT gateways on the path between client devices and the public Internet. If a mobile phone customer receives a translated address from their mobile phone carrier and uses "Personal Hotspot" or "Internet Sharing" software on their mobile phone to make Wireless LAN (WLAN) Internet access available to other client devices, then there are at least two NAT gateways on the path between those client devices and the public Internet. If a hotel guest connects a portable WLAN gateway to their hotel room's Ethernet port to share their room's Internet connection between their phone and their laptop computer, then packets from the client devices may traverse the hotel guest's portable NAT, the hotel network's NAT, and the ISP's NAT before reaching the public Internet. While it is possible, in theory, that client devices could somehow discover all the NATs on the path and communicate with each one separately using PCP , in practice it is not clear how client devices would reliably learn this information. Since the NAT gateways are installed and operated by different individuals and organizations, no single entity has knowledge of all the NATs on the path. Also, even if a client device could somehow know all the NATs on the path, requiring a client device to communicate separately with all of them imposes unreasonable complexity on PCP clients, many of which are expected to be simple low-cost devices. In addition, this goes against the spirit of NAT gateways. The main purpose of a NAT gateway is to make multiple downstream client devices appear, from the point of view of everything upstream of the NAT gateway, to be a single client device. In the same spirit, it makes sense for a PCP-capable NAT gateway to make multiple downstream client devices requesting port mappings appear, from the point of view of everything upstream of the NAT gateway, to be a single client device requesting port mappings.

Another envisioned use case of the PCP proxy is to help establish successful PCP communications for PCP clients that cannot be configured with the address of a PCP server located more than one hop away. A PCP proxy can, for instance, be embedded in a CPE (Customer Premises Equipment) while the PCP server is located in a network operated by an ISP. This is illustrated in .

----------|Server| +------+ | +-----+ +------+ |Client|--+ CPE +------+ | | LAN ]]>

This works because the proxy's server side is listening on the address used as a default gateway by the clients. The clients use that address as a fallback when discovering the PCP server's address. The proxy picks up the requests and forwards them upstream to the ISP's PCP server, with whose address it has been provisioned through regular PCP client provisioning means. This particular use case assumes that provisioning the server's address on the CPE is feasible while doing it on the clients in the LAN is not, which is what makes the PCP proxy valuable. An alternative way to contact an upstream PCP server that may be several hops away is to use a well-known anycast address , but that technique can be problematic when multiple PCP servers are to be contacted .

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in "Key words for use in RFCs to Indicate Requirement Levels". Where this document uses the terms "upstream" and "downstream", the term "upstream" refers to the direction outbound packets travel towards the public Internet, and the term "downstream" refers to the direction inbound packets travel from the public Internet towards client systems. Typically, when a home user views a web site, their computer sends an outbound TCP SYN packet upstream towards the public Internet, and an inbound downstream TCP SYN ACK reply comes back from the public Internet.

Upon receipt of a PCP mapping-creation request from a downstream PCP client, a PCP proxy first examines its local mapping table to see if it already has a valid active mapping matching the internal address and internal port (and in the case of PEER requests, the remote peer) given in the request. If the PCP proxy does not already have a valid active mapping for this mapping-creation request, then it allocates an available port on its external interface. We assume for the sake of this description that the address of its external interface is itself a private address, subject to translation by an upstream NAT. The PCP proxy then constructs an appropriate corresponding PCP request of its own (as described below) and sends it to its upstream NAT, and the newly created local mapping is considered temporary until a confirming reply is received from the upstream PCP server. If the PCP proxy does already have a valid active mapping for this mapping-creation request and the lifetime remaining on the local mapping is at least 3/4 of the lifetime requested by the PCP client, then the PCP proxy SHOULD send an immediate reply giving the outermost external address and port (previously learned using PCP recursively, as described below) and the actual lifetime remaining for this mapping. If the lifetime remaining on the local mapping is less than 3/4 of the lifetime requested by the PCP client, then the PCP proxy MUST generate an upstream request as described below. For mapping-deletion requests (lifetime = 0), the local mapping, if any, is deleted, and then (regardless of whether or not a local mapping existed) a corresponding upstream request is generated. The PCP proxy knows the destination IP address for its upstream PCP request using the same means that are available for provisioning a PCP client. In particular, the PCP proxy MUST follow the procedure defined in Section 8.1 of the PCP specification to discover its PCP server. This does not preclude other means from being used in addition. In the upstream PCP request: The PCP client's IP address and internal port are the PCP proxy's own external address and port just allocated for this mapping. The suggested external address and port in the upstream PCP request SHOULD be copied from the original PCP request. On a typical renewal request, this will be the outermost external address and port previously learned by the client. The requested lifetime is as requested by the client if it falls within the acceptable range for this PCP server; otherwise, it SHOULD be capped to appropriate minimum and maximum values configured for this PCP server. The mapping nonce is copied from the original PCP request. For PEER requests, the remote peer IP address and port are copied from the original PCP request. Upon receipt of a PCP reply giving the outermost (i.e., publicly routable) external address, port, and lifetime, the PCP proxy records this information in its own mapping table and relays the information to the requesting downstream PCP client in a PCP reply. The PCP proxy therefore records, among other things, the following information in its mapping table: Client's internal address and port. External address and port allocated by this PCP proxy. Outermost external address and port allocated by the upstream PCP server. Mapping lifetime (also dictated by the upstream PCP server). Mapping nonce. In the downstream PCP reply: The lifetime is as granted by the upstream PCP server, or less if the granted lifetime exceeds the maximum lifetime this PCP server is configured to grant. If the proxy chooses to grant a downstream lifetime greater than the lifetime granted by the upstream PCP server (which is NOT RECOMMENDED), then this PCP proxy MUST take responsibility for renewing the upstream mapping itself. The Epoch Time is this PCP proxy's Epoch Time, not the Epoch Time of the upstream PCP server. Each PCP server has its own independent Epoch Time. However, if the Epoch Time received from the upstream PCP server indicates a loss of state in that PCP server, the PCP proxy can either (1) recreate the lost mappings itself or (2) reset its own Epoch Time to cause its downstream clients to perform such state repairs themselves. A PCP proxy MUST NOT simply copy the upstream PCP server's Epoch Time into its downstream PCP replies, because if it suffers its own state loss it needs the ability to communicate that state loss to clients. Thus, each PCP server has its own independent Epoch Time. However, as a convenience, a downstream PCP proxy may simply choose to reset its own Epoch Time whenever it detects that its upstream PCP server has lost state. Thus, in this case, the PCP proxy's Epoch Time always resets whenever its upstream PCP server loses state; it may reset at other times as well. The mapping nonce is copied from the reply received from the upstream PCP server. The assigned external port and assigned external IP address are copied from the reply received from the upstream PCP server (i.e., they are the outermost external IP address and port, not the locally assigned external address and port). By recursive application of this procedure, the outermost external IP address and port are relayed from the outermost NAT, through one or more intervening PCP proxies, until they ultimately reach the downstream client. For PEER requests, the remote peer IP address and port are copied from the reply received from the upstream PCP server.

A PCP proxy SHOULD implement optimized hairpin routing. What this means is the following: If a PCP proxy observes an outgoing packet arriving on its internal interface that is addressed to an external address and port appearing in the NAT gateway's own mapping table, then the NAT gateway SHOULD (after creating a new outbound mapping if one does not already exist) rewrite the packet appropriately and deliver it to the internal client to which that external address and port are currently allocated. Similarly, if a PCP proxy observes an outgoing packet arriving on its internal interface that is addressed to an *outermost* external address and port appearing in the NAT gateway's own mapping table, then the NAT gateway SHOULD do as described above: create a new outbound mapping if one does not already exist, and then rewrite the packet appropriately and deliver it to the internal client to which that outermost external address and port are currently allocated. This is not necessary for successful communication, but it provides efficiency. Without this optimized hairpin routing, the packet will be delivered all the way to the outermost NAT gateway, which will then perform standard hairpin translation and send it back. Using knowledge of the outermost external address and port, this rewriting can be anticipated and performed locally. This rewriting technique will typically offer higher throughput and lower latency than sending packets all the way to the outermost NAT gateway and back. Note that traffic counters maintained by an upstream PCP server will differ from the counters of a PCP proxy implementing optimized hairpin routing.

Any recursive algorithm needs a mechanism to terminate the recursion at the appropriate point. This termination of recursion can be achieved in a variety of ways. The following (non-exhaustive) examples are provided for illustration purposes: An ISP's PCP-controlled gateway (which may embed a NAT, firewall, or any function that can be controlled with PCP) could be configured to know that it is the outermost PCP-controlled gateway, and consequently it does not need to relay PCP requests upstream. A PCP-controlled gateway could determine automatically that if its external address is not one of the known private addresses , then its external address is a public routable IP address, and consequently it does not need to relay PCP requests upstream. Recursion may be terminated if there is no explicit list of PCP servers configured (manually, using DHCP, or otherwise) or if its default router is not responsive to PCP requests. Recursion may also be terminated if the upstream PCP-controlled device does not embed a PCP proxy.

As with a regular PCP server, the PCP-controlled device can be a NAT, a firewall, or even some sort of hybrid. In particular, a PCP proxy that simply relays all requests upstream can be thought of as the degenerate case of a PCP server controlling a wide-open firewall back-to-back with a regular PCP client. One important property of the PCP-controlled device will affect the PCP proxy's behavior: when the proxy's server part instructs the device to create a mapping, that mapping's external address may or may not be one that belongs to the proxy node. When the mapping's external address belongs to the proxy node, as would presumably be the case for a NAT, then the proxy's client side sends out an upstream PCP request using the mapping's external IP address as the source. When the mapping's external address does not belong to the proxy node, as would presumably be the case for a firewall, then the proxy's client side needs to install upstream mappings on behalf of its downstream clients. To do this, it MUST insert a THIRD_PARTY option in its upstream PCP request carrying the mapping's external address. Note that hybrid PCP-controlled devices may create NAT-like mappings in some circumstances and firewall-like mappings in others. A proxy controlling such a device would adjust its behavior dynamically, depending on the kind of mapping created.

When no NAT is co-located with the PCP proxy, the port numbers included in received PCP messages (from the PCP server or PCP client(s)) are not altered by the PCP proxy. The PCP proxy relays to the PCP server unknown options and Opcodes because there is no reachability failure risk.

By default, the proxy MUST relay unknown Opcodes and mandatory-to-process unknown options. Rejecting unknown options and Opcodes has the drawback of preventing a PCP client from making use of new capabilities offered by the PCP server but not supported by the PCP proxy, even if no IP address and/or port is included in the option/Opcode. Because PCP messages with an unknown Opcode or mandatory-to-process unknown options can carry a hidden internal address or internal port that will not be translated, a PCP proxy MUST be configurable to disable relaying unknown Opcodes and mandatory-to-process unknown options. If the PCP proxy is configured to disable relaying unknown Opcodes and mandatory-to-process unknown options, the PCP proxy MUST behave as follows: a PCP proxy co-located with a NAT MUST reject, via an UNSUPP_OPCODE error response, a received request with an unknown Opcode. a PCP proxy co-located with a NAT MUST reject, via an UNSUPP_OPTION error response, a received request with a mandatory-to-process unknown option.

ANNOUNCE requests received from PCP clients are handled locally; as such, these requests MUST NOT be relayed to the provisioned PCP server. Upon receipt of an unsolicited ANNOUNCE response from a PCP server, the PCP proxy proceeds to renew the mappings and checks to see whether or not there are changes compared to a local cache if it is maintained by the PCP proxy. If no change is detected, no unsolicited ANNOUNCE is generated towards PCP clients. If a change is detected, the PCP proxy MUST generate unsolicited ANNOUNCE message(s) to appropriate PCP clients. If the PCP proxy does not maintain a local cache for the mappings, unsolicited multicast ANNOUNCE messages are sent to PCP clients. Upon change of its external IP address, the PCP proxy SHOULD renew the mappings it maintained. If the PCP server assigns a different external port, the PCP proxy SHOULD follow the PCP mapping repair procedure . This can be achieved only if a full state table is maintained by the PCP proxy.

A PCP proxy MAY handle multiple PCP servers at the same time. Each PCP server is associated with its own epoch value. PCP clients are not aware of the presence of multiple PCP servers. Following the PCP Server Selection process, if several PCP servers are configured to the PCP proxy, it will contact in parallel all these PCP servers. In some contexts (e.g., PCP-controlled Carrier-Grade NATs (CGNs)), the PCP proxy MAY load-balance the PCP clients among available PCP servers. The PCP proxy MUST ensure that requests of a given PCP client are relayed to the same PCP server. The PCP proxy MAY rely on some fields (e.g., Zone-ID ) in the PCP request to redirect the request to a given PCP server.

The PCP proxy MUST follow the security considerations detailed in the PCP specification for both the client and server side. specifies the cases where a THIRD_PARTY option is inserted by the PCP proxy. In those cases, ways to prevent a malicious user from creating mappings on behalf of a third party must be employed as discussed in Section 13.1 of the PCP specification. In particular, THIRD_PARTY options MUST NOT be enabled unless the network on which the PCP messages are to be sent is fully trusted (via physical or cryptographic security, or both) -- for example, if access control lists (ACLs) are installed on the PCP proxy, the PCP server, and the network between them so that those ACLs allow only communications from a trusted PCP proxy to the PCP server. A received request carrying an unknown Opcode or option SHOULD be dropped (or, in the case of an unknown option that is not mandatory to process, the option SHOULD be removed) if it is not compatible with security controls provisioned to the PCP proxy. The device embedding the PCP proxy MAY block PCP requests directly sent to the upstream PCP server(s). This can be enforced using ACLs.

Many thanks to C. Zhou, T. Reddy, and D. Thaler for their review and comments. Special thanks to F. Dupont, who contributed to this document.