IANA Registry for the First Nibble Following a Label Stack

An MPLS packet consists of a label stack, an optional "post-stack header" and an optional embedded packet (in that order). However, in the data plane, there are scant clues regarding the post-stack header, and no clue as to the type of embedded packet; this information is communicated via other means, such as the routing protocols that signal the labels in the stack. Nonetheless, in order to better handle an MPLS packet in the data plane, it is common practice for network equipment to "guess" the type of embedded packet. Such equipment may also need to process the post-stack header. Both of these require parsing the data after the label stack. To do this, the "first nibble" (the top four bits of the first octet following the label stack) is often used. The semantics and usage of the first nibble is not well documented, nor are the assignments of values. This memo serves three purposes: To document the assignments already made To provide for the clear documentation of future assignments through the creation of an "MPLS First Nibble registry" Provide a method to tracking usage by requiring more consistent documentation

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 BCP14 when, and only when, they appear in all capitals, as shown here. label switching router. one whose Layer 2 header declares the type to be MPLS. For Ethernet, that means the Ethertype is 0x8847 or 0x8848. (of an MPLS packet) all labels (four octet fields) after the Layer 2 header, up to and including the label with the BoS bit set (). the most significant four bits of the first octet following the label stack. all data after the label stack, including the MFN, an optional post-stack header and the embedded packet. optional field of interest to the egress LSR (and possibly to transit LSRs). Examples include a control word or an associated channel. The PSH MUST indicate its length, so that a parser knows where the embedded packet starts. All octets beyond the PSH (if any). This could be an IPv4 or IPv6 packet (e.g., for traffic engineering of IP packets, or for a Layer 3 VPN ), an Ethernet packet (for VPLS (, ) or EVPN ), or some other type of Layer 2 frame .

shows an MPLS packet with Layer 2 header X and a label stack Y ending with Label-n. shows three examples of an MPLS payload, A, B and C. The full MPLS packets thus are: [X, Y, A], [X, Y, B], and [X, Y, C]. A. The first payload is a bare IP packet, i.e., no PSH. The MFN (MPLS First Nibble) in this case overlaps with the IP version number. B. The next payload is a bare non-IP packet; again, no PSH. The MFN here is the first nibble of the payload, whatever it happens to be. C. The last example is an MPLS Payload that starts with a PSH followed by the embedded packet.

An MPLS packet can contain many types of embedded packet. The most common types are: An IPv4 packet (whose IP header has version number 4). An IPv6 packet (whose IP header has version number 6). A Layer 2 Ethernet frame (i.e., not including the Preamble or the Start frame delimiter), starting with the destination MAC address. Many other packet types are possible, and in principle, any Layer 2 embedded packet is permissible; indeed, in the past, PPP, Frame Relay and ATM packets were reasonably common. In addition, there may be a post-stack header ahead of the embedded packet, and this needs be to parsed. The MPLS First Nibble is currently used for both of these purposes.

There are four common ways to load balance an MPLS packet: One can use the top label alone. One can do better by using all the (non-SPL) labels in the stack. One can do even better by "divining" the type of embedded packet, and using fields from the guessed header. One can do best by using either an Entropy Label or a FAT Pseudowire Label ; see .) Load balancing based on just the top label means that all packets with that top label will go the same way -- this is far from ideal. Load balancing based on the entire label stack (not including SPLs) is better, but may still be uneven. If, however, the embedded packet is an IP packet, then the combination of (<source IP address>, <dest IP address>, <transport protocol>, <source port>, and <dest port>) from the IP header of the embedded packet forms an excellent basis for load balancing. This is what is typically used for load balancing IP packets. An MPLS packet doesn't, however, carry a payload type identifier. There is a simple heuristic that is commonly used to guess the type of the embedded packet. The first nibble, i.e., the four most significant bits of the first octet, of an IP header contains the IP version number. This in turn indicates where to find the relevant fields for load balancing. The heuristic goes roughly as follows:

If the MFN is 0x4 (0100b), treat the payload as an IPv4 packet, and find the relevant fields for load balancing on that basis. If the MFN is 0x6 (0101b), treat the payload as an IPv6 packet, and find the relevant fields for load balancing on that basis. If the MFN is anything else, the MPLS payload is not an IP packet; fall back to load balancing using the label stack. This heuristic has been implemented in many (legacy) routers, and performs well in the case of , A. However, this heuristic can work very badly for , B. For example, if payload B is an Ethernet frame, then the MFN is the first nibble of the OUI of the destination MAC address, which can be 0x4 or 0x6, and if so would lead to very bad load balancing. This behavior can happen to other types of non-IP payload as well. This in turn led to the idea of inserting a PSH (e.g., a pseudowire control word , a DetNet control word or a BIER header ) where the MPLS First Nibble is NOT 0x4 or 0x6, to explicitly prevent forwarding engines from confusing the MPLS payload with an IP packet. recommends the use of a control word when the embedded packet is an Ethernet frame. RFC 8469 was published at the request of the operator community and the IEEE RAC as a result of operational difficulties with pseudowires that did not contain the control word. This memo introduces a requirement and a recommendation, the first building on the above; the second deprecating the use of the heuristic in . The intent of both of these is that legacy routers continue to operate as they have, with no new problems introduced as a result of this memo. However, new implementations SHOULD follow these recommendations for more robust operation.

Going forward, network equipment MUST use a post-stack header with an MPLS First Nibble value that is not 0x4 or 0x6 in all cases when the MPLS payload is not an IP packet. Effectively, , B is disallowed. [AGREED???] This replaces the following text from , section 3, paragraph 3: "It is REQUIRED, however, that applications depend upon in-order packet delivery restrict the first nibble values to 0x0 and 0x1. This will ensure that their traffic flows will not be affected if some future routing equipment does similar snooping on some future version(s) of IP." This also replaces the following text from , section 4, paragraph 1: "This document updates [RFC4448] to state that both the ingress provider edge (PE) and the egress PE SHOULD support the Ethernet PW CW and that, if supported, the CW MUST be used."

It is RECOMMENDED that, going forward, if good load balancing of MPLS packets is desired, either an Entropy Label or a FAT Pseudowire Label SHOULD be used; furthermore, going forward, the heuristic in MUST NOT be used. [AGREED???] A consequence of Recommendation 2 is that, while legacy routers may look for a MPLS First Nibble of 0x4 or 0x6, no router will look for a MPLS First Nibble of 0x7 (or whatever the next IP version number will be) for load balancing purposes. This means that the values 0x4 and 0x6 are used to (sometimes incorrectly) identify IPv4 and IPv6 packets, but no other First Nibble values will be used to identify IP packets. This obviates the need for paragraph 4, section 3 in : "This behavior implies that if in the future an IP version is defined with a version number of 0x0 or 0x1, then equipment complying with this BCP would be unable to look past one or more MPLS headers, and loadsplit traffic from a single LSP across multiple paths based on a hash of specific fields in the IPv0 or IPv1 headers. That is, IP traffic employing these version numbers would be safe from disturbances caused by inappropriate loadsplitting, but would also not be able to get the performance benefits." This also expands the MFN Registry to all 16 possible values, not just 0x0 and 0x1.

Given the above recommendations on the use of a post-stack header and future non-use of the heuristic () via the use of Entropy or FAT Pseudowire Labels, the main reason for creating a First Nibble registry is to document the types of post-stack headers that may follow a label stack, and to simplify their parsing.

The MPLS WG is currently engaged in updating the MPLS architecture; part of this work involves the use of post-stack headers. This is not possible if post-stack header values are allocated on an ad hoc basis, and their parsing and semantics is ill-specified. Consider that the MPLS First Nibble value of 0x0 has two different formats, depending on whether the post-stack header is a pseudowire control word or a DetNet control word; disambiguation requires the context of the service label. This was a considered decision; documenting this would be helpful to future implementors. With a registy, post-stack headers become easier to parse; the values are unique, not needing means outside the data plane to interpret them correctly; and their semantics and usage are documented. (Thank you, IANA!)

The use of the MPLS First Nibble stemmed from the desire to heuristically identify IP packets for load balancing purposes. It was then discovered that non-IP packets, misidentified as IP when the heuristic failed, were being badly load balanced, leading to . This situation may confuse some as to relationship between the MPLS First Nibble Registry and the IP Version Numbers registry. These registries are quite different: The IP Version Numbers registry's explicit purpose is to track IP version numbers in an IP header. The MPLS First Nibble registry's purpose is to track post-stack header types. The only intersection points between the two registries is for values 0x4 and 0x6 (for backward compatibility). There is no need to track future IP version number allocations in the MPLS First Nibble registry.