Revised Error Handling for BGP UPDATE Messages

According to the base BGP specification , a BGP speaker that receives an UPDATE message containing a malformed attribute is required to reset the session over which the offending attribute was received. This behavior is undesirable, because a session reset impacts not only routes with the offending attribute, but also other, valid routes exchanged over the session. In the case of optional transitive attributes, the behavior is especially troublesome and may present a potential security vulnerability. This is because attributes may have been propagated without being checked by intermediate routers that don't recognize the attributes. In effect, the attributes may have been tunneled; and when they reach a router that recognizes and checks the attributes, the session that is reset may not be associated with the router that is at fault. To make matters worse, in such cases although the problematic attributes may have originated with a single update transmitted by a single BGP speaker, by the time they encounter a router that checks them they may have been replicated many times, and thus may cause the reset of many peering sessions. Thus the damage inflicted may be multiplied manyfold. The goal for revising the error handling for UPDATE messages is to minimize the impact on routing by a malformed UPDATE message, while maintaining protocol correctness to the extent possible. This can be achieved largely by maintaining the established session and keeping the valid routes exchanged, but removing the routes carried in the malformed UPDATE from the routing system. This document partially revises the error handling for UPDATE messages, and provides guidelines for the authors of documents defining new attributes. Finally, it revises the error handling procedures for a number of existing attributes. Specifically, the error handling procedures of , , , , , , , and are revised.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

In this document we refer to four different approaches to handle errors found in BGP path attributes. They are as follows (listed in order, from the one with the "strongest" action to the one with the "weakest" action): Session reset: This is the approach used throughout the base BGP specification, where a NOTIFICATION is sent and the session terminated. AFI/SAFI disable: specifies a procedure for disabling a particular AFI/SAFI. Treat-as-withdraw: In this approach, the UPDATE message containing the path attribute in question MUST be treated as though all contained routes had been withdrawn just as if they had been listed in the WITHDRAWN ROUTES field (or in the MP_UNREACH_NLRI attribute if appropriate) of the UPDATE message, thus causing them to be removed from the Adj-RIB-In according to the procedures of . Attribute discard: In this approach the malformed attribute MUST be discarded and the UPDATE message continues to be processed. This approach MUST NOT be used except in the case of an attribute that has no effect on route selection or installation.

This specification amends Section 6.3 in a number of ways. See also for treatment of specific path attributes. The first paragraph is revised as follows: Old Text: All errors detected while processing the UPDATE message MUST be indicated by sending the NOTIFICATION message with the Error Code UPDATE Message Error. The error subcode elaborates on the specific nature of the error. New Text: An error detected while processing the UPDATE message for which a session reset is specified MUST be indicated by sending the NOTIFICATION message with the Error Code UPDATE Message Error. The error subcode elaborates on the specific nature of the error. Error handling for the following case remains unchanged: If the Withdrawn Routes Length or Total Attribute Length is too large (i.e., if Withdrawn Routes Length + Total Attribute Length + 23 exceeds the message Length), then the Error Subcode MUST be set to Malformed Attribute List. Attribute Flag error handling is revised as follows: Old Text: If any recognized attribute has Attribute Flags that conflict with the Attribute Type Code, then the Error Subcode MUST be set to Attribute Flags Error. The Data field MUST contain the erroneous attribute (type, length, and value). New Text: If the value of either the Optional or Transitive bits in the Attribute Flags is in conflict with their specified values, then the attribute MUST be treated as malformed and the treat-as-withdraw approach used, unless the specification for the attribute mandates different handling for incorrect Attribute Flags. If any of the well-known mandatory attributes are not present in an UPDATE message, then "treat-as-withdraw" MUST be used. (Note that reclassifies NEXT_HOP as what is effectively discretionary.) "Treat-as-withdraw" MUST be used for the cases that specify a session reset and involve any of the attributes ORIGIN, AS_PATH, NEXT_HOP, MULTI_EXIT_DISC, or LOCAL_PREF. "Attribute discard" MUST be used for any of the cases that specify a session reset and involve ATOMIC_AGGREGATE or AGGREGATOR. If the MP_REACH_NLRI attribute or the MP_UNREACH_NLRI attribute appears more than once in the UPDATE message, then a NOTIFICATION message MUST be sent with the Error Subcode "Malformed Attribute List". If any other attribute (whether recognized or unrecognized) appears more than once in an UPDATE message, then all the occurrences of the attribute other than the first one SHALL be discarded and the UPDATE message continue to be processed. When multiple attribute errors exist in an UPDATE message, if the same approach (either "session reset", "treat-as-withdraw" or "attribute discard") is specified for the handling of these malformed attributes, then the specified approach MUST be used. Otherwise the approach with the strongest action MUST be used. The Withdrawn Routes field MUST be checked for syntactic correctness in the same manner as the NLRI field. This is discussed further below, and in . Finally, we observe that in order to use the approach of "treat-as-withdraw", the entire NLRI field and/or the MP_REACH_NLRI and MP_UNREACH_NLRI attributes need to be successfully parsed -- what this entails is discussed in more detail in . If this is not possible, the procedures of and/or continue to apply, meaning that the "session reset" approach (or the "AFI/SAFI disable" approach) MUST be followed.

There are two error cases in which the Total Attribute Length value can be in conflict with the enclosed path attributes, which themselves carry length values. In the "overrun" case, as the enclosed path attributes are parsed, the length of the last encountered path attribute would cause the Total Attribute Length to be exceeded. In the "underrun" case, as the enclosed path attributes are parsed, after the last successfully-parsed attribute, fewer than three octets remain, or fewer than four octets, if the Attribute Flags field has the Extended Length bit set -- that is, there remains unconsumed data in the path attributes but yet insufficient data to encode a single minimum-sized path attribute. In either of these cases, an error condition exists and the treat-as-withdraw approach MUST be used (unless some other, more severe error is encountered dictating a stronger approach), and the Total Attribute Length MUST be relied upon to enable the beginning of the NLRI field to be located. For all path attributes other than those specified as having an attribute length that may be zero, it SHALL be considered a syntax error for the attribute to have a length of zero. (Of the path attributes considered in this specification, only AS_PATH and ATOMIC_AGGREGATE may validly have an attribute length of zero.)

To facilitate the determination of the NLRI field in an UPDATE with a malformed attribute: The MP_REACH_NLRI or MP_UNREACH_NLRI attribute (if present) SHALL be encoded as the very first path attribute in an UPDATE. An UPDATE message MUST NOT contain more than one of the following: non-empty Withdrawn Routes field, non-empty Network Layer Reachability Information field, MP_REACH_NLRI attribute, and MP_UNREACH_NLRI attribute. Since older BGP speakers may not implement these restrictions, an implementation MUST still be prepared to receive these fields in any position or combination. If the encoding of is used, the NLRI field for the IPv4 unicast address family is carried immediately following all the attributes in an UPDATE. When such an UPDATE is received, we observe that the NLRI field can be determined using the "Message Length", "Withdrawn Route Length" and "Total Attribute Length" (when they are consistent) carried in the message instead of relying on the length of individual attributes in the message.

specifies an End-of-RIB message ("EoR") that can be encoded as an UPDATE message that contains only a MP_UNREACH_NLRI attribute that encodes no NLRI (it can also be a completely empty UPDATE message in the case of the "legacy" encoding). In all other well-specified cases, an UPDATE either carries only withdrawn routes (either in the Withdrawn Routes field, or the MP_UNREACH_NLRI attribute), or it advertises reachable routes (either in the Network Layer Reachability Information field, or the MP_REACH_NLRI attribute). Thus, if an UPDATE message is encountered that does contain path attributes other than MP_UNREACH_NLRI and doesn't encode any reachable NLRI, we cannot be confident that the NLRI have been successfully parsed as Section 3 (j) requires. For this reason, if any path attribute errors are encountered in such an UPDATE message, and if any encountered error specifies an error-handling approach other than "attribute discard", then the "session reset" approach MUST be used.

The NLRI field or Withdrawn Routes field SHALL be considered "syntactically incorrect" if either of the following are true: The length of any of the included NLRI is greater than 32, When parsing NLRI contained in the field, the length of the last NLRI found exceeds the amount of unconsumed data remaining in the field. Similarly, the MP_REACH_NLRI or MP_UNREACH_NLRI attribute of an update SHALL be considered to be incorrect if any of the following are true: The length of any of the included NLRI is inconsistent with the given AFI/SAFI (for example, if an IPv4 NLRI has a length greater than 32 or an IPv6 NLRI has a length greater than 128), When parsing NLRI contained in the attribute, the length of the last NLRI found exceeds the amount of unconsumed data remaining in the attribute. The attribute flags of the attribute are inconsistent with those specified in . The length of the MP_UNREACH_NLRI attribute is less than 3, or the length of the MP_REACH_NLRI attribute is less than 5.

Certain address families, for example MCAST-VPN, MCAST-VPLS and EVPN have NLRI that are typed. Since supported type values within the address family are not expressed in the MP-BGP capability, it is possible for a BGP speaker to advertise support for the given address family and sub-address family while still not supporting a particular type of NLRI within that AFI/SAFI. A BGP speaker advertising support for such a typed address family MUST handle routes with unrecognized NLRI types within that address family by discarding them, unless the relevant specification for that address family specifies otherwise.

Although the "treat-as-withdraw" error-handling behavior defined in Section 2 makes every effort to preserve BGP's correctness, we note that if an UPDATE received on an IBGP session is subjected to this treatment, inconsistent routing within the affected Autonomous System may result. The consequences of inconsistent routing can include long-lived forwarding loops and black holes. While lamentable, this issue is expected to be rare in practice, and, more importantly, is seen as less problematic than the session-reset behavior it replaces. When a malformed attribute is indeed detected over an IBGP session, we RECOMMEND that routes with the malformed attribute be identified and traced back to the ingress router in the network where the routes were sourced or received externally, and then a filter be applied on the ingress router to prevent the routes from being sourced or received. This will help maintain routing consistency in the network. Even if inconsistent routing does not arise, the "treat-as-withdraw" behavior can cause either complete unreachability or sub-optimal routing for the destinations whose routes are carried in the affected UPDATE message. Note that "treat-as-withdraw" is different from discarding an UPDATE message. The latter violates the basic BGP principle of incremental update, and could cause invalid routes to be kept. Because of these potential issues, a BGP speaker MUST provide debugging facilities to permit issues caused by a malformed attribute to be diagnosed. At a minimum, such facilities MUST include logging an error listing the NLRI involved, and containing the entire malformed UPDATE message when such an attribute is detected. The malformed UPDATE message SHOULD be analyzed, and the root cause SHOULD be investigated. mentions that attribute discard should not be used in cases where "the attribute in question has or may have an effect on route selection." Although all cases that specify attribute discard in this document do not affect route selection by default, in principle routing policies could be written that affect selection based on such an attribute. Operators should take care when writing such policies to consider the possible consequences of an attribute discard. (In general, as long as such policies are only applied to external BGP sessions, correctness issues are not expected to arise.)

In the following subsections, we elaborate on the conditions for error-checking various path attributes, and specify what approach(es) should be used to handle malformations. It is possible that implementations may apply other error checks not contemplated here. If so, the error handling approach given here should generally be applied. This section addresses all path attributes that are defined at the time of this writing, that were not defined with error-handling consistent with , and that are not marked as "deprecated" in . Attributes 17 (AS4_PATH), 18 (AS4_AGGREGATOR), 22 (PMSI_TUNNEL), 23 (Tunnel Encapsulation Attribute), 26 (AIGP), 27 (PE Distinguisher Labels) and 29 (BGP-LS Attribute) do have error-handling consistent with and thus are not further discussed herein. Attributes 11 (DPA), 12 (ADVERTISER), 13 (RCID_PATH / CLUSTER_ID), 19 (SAFI Specific Attribute), 20 (Connector Attribute), 21 (AS_PATHLIMIT) and 28 (BGP Entropy Label Capability Attribute) are deprecated and thus are not further discussed herein.

The attribute is considered malformed if its length is not 1, or it has an undefined value . An UPDATE message with a malformed ORIGIN attribute SHALL be handled using the approach of "treat-as-withdraw".

An AS_PATH is considered malformed if an unrecognized segment type is encountered, or if it contains a malformed segment. A segment is considered malformed if any of the following obtains: There is an overrun, where the path segment length field of the last segment encountered would cause the Attribute Length to be exceeded. There is an underrun, where after the last successfully-parsed segment, there is only a single octet remaining (that is, there is not enough unconsumed data to provide even an empty segment header). It has a path segment length field of zero. An UPDATE message with a malformed AS_PATH attribute SHALL be handled using the approach of "treat-as-withdraw". also says that an implementation optionally "MAY check whether the leftmost ... AS in the AS_PATH attribute is equal to the autonomous system number of the peer that sent the message". A BGP implementation SHOULD also handle routes that violate this check using "treat-as-withdraw", but MAY follow the session reset behavior if configured to do so.

The attribute is considered malformed if its length is not 4 . A length of 16 is also permissible if is in use and all conditions given in that specification have been satisfied. An UPDATE message with a malformed NEXT_HOP attribute SHALL be handled using the approach of "treat-as-withdraw".

The attribute is considered malformed if its length is not 4 . An UPDATE message with a malformed MULTI_EXIT_DISC attribute SHALL be handled using the approach of "treat-as-withdraw".

The error handling of is revised as follows. If the LOCAL_PREF attribute is received from an external neighbor, it SHALL be discarded using the approach of "attribute discard", or if received from an internal neighbor, it SHALL be considered malformed if its length is not equal to 4. If malformed, the UPDATE SHALL be handled using the approach of "treat-as-withdraw".

The attribute SHALL be considered malformed if its length is not 0 . An UPDATE message with a malformed ATOMIC_AGGREGATE attribute SHALL be handled using the approach of "attribute discard".

The error conditions specified in for the attribute are revised as follows: The AGGREGATOR attribute SHALL be considered malformed if any of the following applies: Its length is not 6 (when the "4-octet AS number capability" is not advertised to, or not received from the peer ). Its length is not 8 (when the "4-octet AS number capability" is both advertised to, and received from the peer). An UPDATE message with a malformed AGGREGATOR attribute SHALL be handled using the approach of "attribute discard".

The error handling of is revised as follows: The Community attribute SHALL be considered malformed if its length is not a nonzero multiple of 4. An UPDATE message with a malformed Community attribute SHALL be handled using the approach of "treat-as-withdraw".

The error handling of is revised as follows. If the ORIGINATOR_ID attribute is received from an external neighbor, it SHALL be discarded using the approach of "attribute discard", or if received from an internal neighbor, it SHALL be considered malformed if its length is not equal to 4. If malformed, the UPDATE SHALL be handled using the approach of "treat-as-withdraw".

The error handling of is revised as follows. If the CLUSTER_LIST attribute is received from an external neighbor, it SHALL be discarded using the approach of "attribute discard", or if received from an internal neighbor, it SHALL be considered malformed if its length is not a nonzero multiple of 4. If malformed, the UPDATE SHALL be handled using the approach of "treat-as-withdraw".

If the Length of Next Hop Network Address field of the MP_REACH attribute is inconsistent with that which was expected, the attribute is considered malformed. Since the next hop precedes the NLRI field in the attribute, in this case it will not be possible to reliably locate the NLRI, and thus the "session reset" or "AFI/SAFI disable" approach MUST be used. "That which was expected", while somewhat vague, is intended to encompass the next hop specified for the Address Family Identifier and Subsequent Address Family Identifier fields and potentially modified by any extensions in use. For example, if is in use then the next hop would have to have a length of 4 or 16. and provide further discussion of the handling of this attribute.

and discuss the handling of this attribute.

We note that does not detail what constitutes "malformation" for the Traffic Engineering path attribute. A future update to that specification may provide more guidance. In the interim, an implementation that determines (for whatever reason) that an UPDATE message contains a malformed Traffic Engineering path attribute MUST handle it using the approach of "treat-as-withdraw".

The error handling of is revised as follows: The Extended Community attribute SHALL be considered malformed if its length is not a nonzero multiple of 8. An UPDATE message with a malformed Extended Community attribute SHALL be handled using the approach of "treat-as-withdraw". Note that a BGP speaker MUST NOT treat an unrecognized Extended Community Type or Sub-Type as an error.

The error handling of is revised as follows: The IPv6 Address Specific Extended Community attribute SHALL be considered malformed if its length is not a nonzero multiple of 20. An UPDATE message with a malformed IPv6 Address Specific Extended Community attribute SHALL be handled using the approach of "treat-as- withdraw". Note that a BGP speaker MUST NOT treat an unrecognized IPv6 Address Specific Extended Community Type or Sub-Type as an error.

The final paragraph of Section 5 of is revised as follows: Old Text: An UPDATE message with a malformed ATTR_SET attribute SHALL be handled as follows. If its Partial flag is set and its Neighbor-Complete flag is clear, the UPDATE is treated as a route withdraw as discussed in [OPT-TRANS-BGP]. Otherwise (i.e., Partial flag is clear or Neighbor-Complete is set), the procedures of the BGP-4 base specification [RFC4271] MUST be followed with respect to an Optional Attribute Error. New Text: An UPDATE message with a malformed ATTR_SET attribute SHALL be handled using the approach of "treat as withdraw". Furthermore, the normative reference to [OPT-TRANS-BGP] in is removed.

A document that specifies a new BGP attribute MUST provide specifics regarding what constitutes an error for that attribute and how that error is to be handled. Allowable error-handling approaches are detailed in . The treat-as-withdraw approach is generally preferred. The document SHOULD also provide consideration of what debugging facilities may be required to permit issues caused by a malformed attribute to be diagnosed. For any malformed attribute that is handled by the "attribute discard" instead of the "treat-as-withdraw" approach, it is critical to consider the potential impact of doing so. In particular, if the attribute in question has or may have an effect on route selection or installation, the presumption is that discarding it is unsafe, unless careful analysis proves otherwise. The analysis should take into account the tradeoff between preserving connectivity and potential side effects. Authors can refer to for examples.

This document makes no request of IANA.

This specification addresses the vulnerability of a BGP speaker to a potential attack whereby a distant attacker can generate a malformed optional transitive attribute that is not recognized by intervening routers (which thus propagate the attribute unchecked) but that causes session resets when it reaches routers that do recognize the given attribute type. In other respects, this specification does not change BGP's security characteristics.

The authors wish to thank Juan Alcaide, Deniz Bahadir, Ron Bonica, Mach Chen, Andy Davidson, Bruno Decraene, Rex Fernando, Jeff Haas, Chris Hall, Joel Halpern, Dong Jie, Akira Kato, Miya Kohno, Warren Kumari, Tony Li, Alton Lo, Shin Miyakawa, Tamas Mondal, Jonathan Oddy, Tony Przygienda, Robert Raszuk, Yakov Rekhter, Eric Rosen, Shyam Sethuram, Rob Shakir, Naiming Shen, Adam Simpson, Ananth Suryanarayana, Kaliraj Vairavakkalai, Lili Wang and Ondrej Zajicek for their observations and discussion of this topic, and review of this document.