Internet Engineering Task Force R. Shakir
Internet-Draft BT
Intended status: Informational July 2012
Expires: December 31, 2012

Operational Requirements for Enhanced Error Handling Behaviour in BGP-4
draft-ietf-grow-ops-reqs-for-bgp-error-handling-05

Abstract

BGP-4 is utilised as a key intra- and inter-Autonomous System routing protocol in modern IP networks. The failure modes as defined by the original protocol standards are based on a number of assumptions around the impact of session failure. Numerous incidents both in the global Internet routing table and within Service Provider networks have been caused by strict handling of a single invalid UPDATE message causing large-scale failures in one or more Autonomous Systems.

This memo describes the current use of BGP-4 within Service Provider networks, and outlines a set of requirements for further work to enhance the mechanisms available to a BGP-4 implementation when erroneous data is detected. Whilst this document does not provide specification of any standard, it is intended as an overview of a set of enhancements to BGP-4 to improve the protocol's robustness to suit its current deployment.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http:/⁠/⁠datatracker.ietf.org/⁠drafts/⁠current/⁠.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on December 31, 2012.

Copyright Notice

Copyright (c) 2012 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http:/⁠/⁠trustee.ietf.org/⁠license-⁠info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.


Table of Contents

1. Introduction

Where BGP-4 [RFC4271] is deployed in the Internet and Service Provider networks, numerous incidents have been recorded due to the manner in which [RFC4271] specifies errors in routing information should be handled. Whilst the behaviour defined in the existing standards retains utility, the deployments of the protocol have changed within modern networks, resulting in significantly different demands for protocol robustness. Whilst a number of Internet Drafts have been written to begin to enhance the behaviour of BGP-4 in terms of the handling of erroneous messages, this memo intends to define a set of requirements for ongoing work. These requirements are considered from the perspective of a Network Operator, and hence this draft does not intend to define the protocol mechanisms by which such error handling behaviour is to be implemented.

1.1. Role of BGP-4 in Service Provider Networks

BGP was designed as an inter-Autonomous System (AS) routing protocol and hence many of the error handling mechanisms within the protocol specification are designed to be conducive to this role. In general, this consideration as an inter-AS routing propagation mechanism results in the view that a BGP session propagates a relatively small amount of network-layer reachability information (NLRI) between two ASes. In this case, it is the expectation of session resilience for those adjacencies that are key to routing continuity (for example, it is expected that two networks peering via BGP would connect multiple times in order to safeguard equipment or protocol failure). In addition, there is some expectation of multiple paths to a particular NLRI being available - it would be expected that a network can fall back to utilising alternate, less direct, paths where a failure of a more direct path occurs.

Traditional network architectures would deploy an Interior Gateway Protocol (IGP) to carry infrastructure and customer routes, with an Exterior Gateway Protocol (EGP) such as BGP being utilised to propagate these routes to other Autonomous Systems. However, with the growth of IP-based services, this is no longer considered best practice. In order to ensure that convergence is within acceptable time bounds, the amount of routing information carried within the IGP is significantly reduced - and tends to be only infrastructure routes. iBGP is then utilised to propagate both customer, and external routes within an AS. As such, BGP has become an IGP, with traditional IGPs acting as a means by which to propagate the routing information which is required to establish a BGP session, and reach the egress node within the local routing domain. This change in role presents different requirements for the robustness of BGP as a routing protocol - with the expectation of similar level of robustness to that of an IGP being set.

Along with this change in role, the nature of the IP routing information that is carried has changed. BGP has become a ubiquitous means by which service information can be propagated between devices. For instance, BGP is utilised to carry routing information for IP/MPLS VPN services as described in [RFC4364]. Since there is an existing deployment of the protocol between PE devices in numerous networks, it has been adapted to propagate this routing information, as its use limits the number of routing protocols required on each device. This additional information being propagated represents a large change in requirement for the error handling of the protocol - where session failure occurs, it is likely a complete service outage for at least a subset of a network's customers is experienced where an erroneous packet may have occurred within a different sub-topology or even service (a different address family for example). For this reason, there is a significant demand to avoid service affecting failures that may be triggered by routing information within a single sub-topology or service.

The combination of the increased number of deployments of BGP-4 as an intra-AS routing protocol, its use for the propagation of additional types of routing and service information, and the growth of IP services has resulted in a substantial increase in the volume of information carried within BGP-4. In numerous networks, RIB sizes of the order of millions of entries exist within individual BGP speakers, with particularly high-scale points exhibited at BGP speakers performing aggregation or functionality designed improve utilisation of network resources (e.g., route reflector hierarchies). Clearly an increase in the amount routing information carried in BGP results in greater impact to services during failures, which is only amplified by a corresponding increase in recovery times. Following a failure, there is a substantial recovery time to learn, compute and distribute new paths, which results in a greater observed impact to services affected, and hence adds further weight to the requirement to avoid failures altogether or, at least, mitigate their impact to the narrowest scope possible, (e.g., a specific NLRI). Whilst an argument could be made that convergence time of BGP-4 could potentially be reduced through deployment of additional computational resource, it is notable that solution is not necessarily straightforward from an implementation or deployment perspective, (e.g., scaling computation resources within a single address-family is difficult). Thus, significant challenges continue to exist for operators when scaling BGP-4 deployments, and hence mechanisms which improve the scalability of BGP-4 are very important.

Both within Internet and multi-service routing architectures, a number of BGP sessions propagate a large proportion of the required routing information for network operation. For Internet routing, these are typically BGP sessions which propagate the global routing table to an AS - failure of these sessions may have a large impact on network service, based on a single erroneous update. In an multi-service environment, typical deployments utilise a small number of core-facing BGP sessions, typically towards route reflector devices. Failure of these sessions may also result in a large impact to network operation. Clearly, the avoidance of conditions requiring these sessions to fail is of great utility to any network operator, and provides further motivation for the revision of the existing behaviour.

Whilst the behaviour in [RFC4271] is suited to ensuring that BGP messages with erroneous routing information in are limited in scope (by means of session reset), with the above considerations, it is clear that this mechanism is not suited to all deployments. It should, however, be noted that the change in scope affects the handling only of errors occurring after BGP session establishment. There is no current operational requirement to amend the means by which error handling in session establishment, or liveliness detection, are performed.

1.2. Overview of Operator Requirements for BGP-4 Error Handling

It is the intention of this document to define a set of criteria for the manner in which a revised error handling mechanism in BGP-4 is required to conform. The motivation for the definition of these requirements can be summarised based on certain behaviour currently present in the protocol that is not deemed acceptable within current operational deployments, or where there is a short-fall in the tool set available to an operator. These key requirements can be summarised as follows:

This document describes each of these requirements in further depth, along with an overview of means by which they are expected to be achieved. In addition, the mechanism by which the enhancements meeting these requirements are to interact is discussed.

2. Errors within BGP-4 UPDATE Messages

Both through analysis of incidents occurring with the Internet DFZ, and multi-service environments utilising BGP-4 to signal service or routing information, a number of different classes of errors within BGP-4 UPDATE messages have been observed. In order to consider the applicability of enhanced error handling mechanisms, it is possible to divide these errors into a number of sub-classes, particularly focusing around the location of the error within the UPDATE message.

Where an UPDATE message is considered invalid by a BGP speaker due to an error within a path attribute that is not the NLRI (where the definition of NLRI includes reachability information encoded in the MP_REACH_NLRI and MP_UNREACH_NLRI attributes as specified in [RFC4760]) it is a requirement of any enhanced error handling mechanism to handle the error in a manner focused on the NLRI contained within the message found to be erroneous. Since in this case, the message received from the remote peer is syntactically valid, it is considered that such an UPDATE is indicative of erroneous data within one or more path attributes. The impact of the current behaviour defined within the protocol makes the implication that the BGP speaker from whom the message is received is now an invalid path for all NLRI announced via the session - which results in a disproportionate impact to overall network operation. In particular scenarios (such as networks with centralised BGP route reflection) such action can result in a loss of all reachability to a network. In other contexts (such as the Internet DFZ), it cannot be assumed that the BGP speaker from whom the UPDATE message is received is directly responsible for the erroneous information contained within the message.

Two further error cases exist within UPDATE messages, both of which are related to the mechanisms that are applicable to messages received where some difficulty exists in parsing the entire BGP message. The two cases concern those cases where a valid NLRI attribute can be extracted, and those where such an attribute is not able to be parsed. In these cases, errors in the packing of attributes within a BGP message may have occurred. Such errors are likely indicative of an error specifically caused by the remote BGP speaker. It is, however, desirable to an operator that such errors are handled without affecting all NLRI across a BGP session. As such, there is a key requirement to maximise the number of cases in which it is possible to extract NLRI from a BGP UPDATE message. To this end, it is required that where possible the MP_REACH_NLRI and MP_UNREACH_NLRI attributes are utilised for encoding all NLRI (including IPv4 Unicast), and that this attribute is included as the first attribute of a BGP UPDATE message (as originally recommended in [I-D.chen-ebgp-error-handling]). Such a change to the order of inclusion of this attribute maximises the number of cases in which NLRI can be extracted from an UPDATE. Where this is possible, it is again required that the error handling mechanisms utilised should be directly applied to the NLRI included in the UPDATE.

For all cases whereby NLRI can be obtained from an UPDATE message, it is expected that the requirements outlined in Section 3 should be considered by any enhancement to the BGP-4 protocol.

In the case that it is not possible to completely parse the NLRI attribute from the UPDATE message received from a peer, it is extremely likely that this is indicative of a serious error with either the process of attribute packing, or buffer usage on the remote BGP speaker. In this case, clearly, it is not possible to apply any error handling mechanism that is limited to a specific set of NLRI, since an implementation has no knowledge of the NLRI included within the UPDATE message. In addition, such errors are considered to be relatively fundamental to the operation of a BGP implementation, and hence may indicate a case whereby significant system errors have occurred. The current BGP-4 standard results in a BGP speaker restarting a session with the remote BGP speaker. However where such an error does occur, it is required that a graceful mechanism is utilised to provide a lower impact to network operation. The requirements for enhancements of this nature to BGP-4 are outlined in Section 5, with the requirements outlined therein focused on providing a means by which system integrity can be restored whilst allowing for continued network operation.

2.1. Classifying BGP Errors and Expected Error Handling

It is clearly of advantage for BGP-4 implementations to utilise a consistent set of error handling mechanisms for the different types of errors that are described in Section 2, and provide consistent nomenclature to refer to them. It is therefore suggested that errors that are indicative of larger scale failures of a BGP speaker, and hence require some error handling at the session level are referred to as 'critical' errors, whilst those errors that are identified based on incorrect content of one of more attributes of a message are referred to as 'semantic' errors.

The errors identified within the following sections consider only those errors within the specifications at the time of writing, it is recommended that in the definition of future extensions to the BGP-4 specification, the error handling behaviour (and the category within which errors within the extension should be considered by an implementation) is defined.

2.1.1. Critical BGP Errors

As described in this document, it is of advantage to limit the number of 'critical' errors that occur within the protocol, therefore, based on analysis of the processing of BGP UPDATE messages, it is required that 'critical' error handling behaviour is applied to: Section 5 are utilised to provide session-level handling of those errors identified as 'critical'.

It is expected that those requirements outlined in

2.1.2. Semantic BGP Errors

Where a BGP message is correctly formed, a number of cases exist whereby the contents of the UPDATE are not valid - in these cases, this represents errors that can be identified to affect specific NLRI. The following cases are expected to be classified as semantic errors: Section 3 and Section 4 of this memo.

In these cases, it is expected that these errors can be handled gracefully, following the requirements detailed in

3. Avoiding use of NOTIFICATION

The error handling behaviour defined in RFC4271 is problematic due to the limited options that are available to an implementation. When an erroneous BGP message is received, at the current time, the implementation must either ignore the error, or send a NOTIFICATION message, after which it is mandatory to terminate the BGP session. It is apparent that this requirement is at odds with that of protocol robustness.

There is significant complexity to this requirement. The mechanism defined in [I-D.chen-ebgp-error-handling] describes a means by which no NOTIFICATION message is generated for all cases whereby NLRI can be extracted from an UPDATE. The NLRI contained within the erroneous UPDATE message is considered as though the remote BGP speaker has provided an UPDATE marking it as withdrawn. This results in a limit in the propagation of the invalid routing information, whilst also ensuring that no traffic is forwarded via a previously-known path that may no longer be valid. This mechanism is referred to as "treat-as-withdraw".

Whilst this behaviour results in avoiding a NOTIFICATION message, keeping other routing information advertised by the remote BGP speaker within the RIB, it may result in unreachability for a sub-set of the NLRI advertised by the remote speaker. Two cases should be considered - that where the entry for a route in the Adj-RIB-In of the neighbour propagating an erroneous packet is utilised, and that where the route installed in the device's RIB is learnt from another BGP speaker. In the former case, should the identified NLRI not be treated as withdrawn, the original NLRI is utilised within the global RIB. However, this information is potentially now invalid (i.e. it no longer provides a valid forwarding path), whilst an alternate (valid) path may exist in another Adj-RIB-In. By continuing to utilise the NLRI for which the UPDATE was considered invalid, traffic may be forwarded via an invalid path, resulting in routing loops, or black-holing. In the second case, no impact to the forwarding of traffic, or global RIB, is incurred, yet where treat-as-withdraw is implemented, possibly stale routing information is purged from the Adj-RIB-In of the neighbour propagating errors.

Whilst mechanisms such as "treat-as-withdraw" are currently documented, the proposals are limited in their scope - particularly in terms of restrictions to implementation only on eBGP sessions. This limitation is made based on the view that the BGP RIB must be consistent across an autonomous system. By implementing treat-as-withdraw for a iBGP session, one or more routers within the Autonomous System may not have reachability to a route, and hence blackholing of traffic, or routing loops, may occur. It should, however, be considered if this view is valid, in light of the manner in which BGP is utilised within operator networks. Inconsistency in a RIB based on a single UPDATE being treated as withdrawn may cause a inconsistency in a single sub-topology (e.g. Layer 3 VPN service), or a service not operating completely (in the case of an UPDATE carrying service membership information). Where a NOTIFICATION and teardown is utilised this is destructive to all sub-topologies in all address family identifiers (AFIs) carried by the session in question. Even where mechanisms such as multi-session BGP are utilised, a whole AFI is affected by such a NOTIFICATION message. In terms of routing operation, it is therefore far less costly to endure a situation where a limited sub-set of routing information within an AS is invalid, than to consider all routing information as invalid based on a single trigger.

At the time of writing, error handling mechanisms related to optional, transitive attributes - such as [I-D.ietf-idr-optional-transitive] are restricted to handling only a subset of attribute errors - whereas the operational requirement is to expand this coverage to the widest set of errors possible (i.e., all semantic errors within UPDATE messages). Additionally, where approaches applicable to a greater number of attributes are proposed (e.g., [I-D.chen-ebgp-error-handling]), these are limited to deployment in eBGP applications only, where requirements also exist in intra-domain cases. As such, it is envisaged that if extended to cover these expanded cases, these mechanisms provide a means to avoid the transmission of a NOTIFICATION message to a remote BGP speaker, based on a single erroneous message, where at all possible, and hence meet this requirement. Critical errors, including those whereby the NLRI cannot be extracted from the UPDATE message, represent cases whereby the receiving system cannot handle the error gracefully based on this mechanism.

4. Recovering RIB Consistency

The recommendations described in Section 3 may result in the RIB for a topology within an AS being inconsistent across the AS' internal routers. Alternatively, where such mechanisms are deployed at an AS boundary, interconnects between two ASes may be inconsistent with each other. There are therefore risks of traffic blackholing, due to missing routing information, or forwarding loops. Whilst this is deemed an acceptable compromise in the short term, clearly, it is suboptimal. Therefore, a requirement exists to provide mechanisms by which a BGP speaker is able to recover the consistency of the Adj-RIB-In for a particular neighbour.

In the general case, the consistency of the BGP RIB can be recovered by re-requesting the entire Adj-RIB-Out of a remote BGP speaker is re-advertised. A mechanism to achieve this re-advertisement is defined within the ROUTE-REFRESH specification [RFC2918]. It is envisaged that by requesting a refresh of all NLRI advertised by a BGP speaker, any NLRI which has been withdrawn due to being contained within an invalid UPDATE message is re-learnt. Where a ROUTE REFRESH is used to directly perform a consistency check between the Adj-RIB-Out of a remote device, and the Adj-RIB-In of the local BGP speaker, a demarcation between the ROUTE-REFRESH, and normal UPDATE messages is required (in order that an "end" of the refresh can be used to identify any 'stale' NLRI) - [I-D.ietf-idr-bgp-enhanced-route-refresh] provides a means by which the ROUTE-REFRESH mechanism can be extended to meet this requirement.

Whilst re-advertisement of the whole BGP RIB provides a means by which withdrawn NLRI can be re-advertised, there are some scaling implications that must be considered. In the case that a ROUTE-REFRESH is generated, all NLRI must be re-packed into UPDATE messages and advertised by one speaker on the BGP session, whilst the other must receive all UPDATE messages, and validate the RIB's consistency. In order to avoid the control-plane load, it is therefore a requirement to utilise targeted mechanisms where possible, rather than incurring the additional load on both the advertising and receiving speaker of building and processing UPDATEs for the entire contents of the RIB.

It is envisaged that during routing inconsistencies caused by utilising the 'treat-as-withdraw' mechanism, the local BGP speaker is aware that some routing information was not able to be processed - due to the fact that an UPDATE message was not parsed correctly. Since this mechanism (as discussed in Section 3) requires the local BGP speaker to have determined the set of NLRI for which an erroneous UPDATE message was received, it is possible to use a targeted mechanisms to re-request the specific NLRI that was contained within the erroneous UPDATE message. By re-requesting, this provides the remote BGP speaker an opportunity to re-transmit the NLRI - possibly providing an opportunity to leverage alternative methods to build the UPDATE message. Such a request requires extension to the existing BGP-4 protocol, in terms of specific UPDATE generation filters with a transient lifetime. It is envisaged that the work within [I-D.zeng-idr-one-time-prefix-orf] provides a mechanism allowing targeted elements of the Adj-RIB-In for a BGP neighbour to be recovered.

It is of particular note for both means of recovering RIB consistency described that these are effective only when considering transient errors within an implementation - for instance, should an RFC interpretation error within an implementation be present, regardless of the number of times a specific UPDATE is generated, it is likely that this error condition will persist (as it may with the existing behaviour defined by [RFC4271]). For this reason, there is an requirement to consider the means by which such consistency recovery mechanisms are utilised. It is not advisable that a dynamic filter and advertisement mechanism is triggered by all error handling events due to the load this is likely to place on the neighbour receiving such a request. Where this BGP speaker is a relatively centralised device - a route reflector (as described by [RFC4456]) for example - the act of generation of UPDATE messages with such frequency is likely to cause disproportionate load. It is therefore an operational requirement of such mechanisms that means of request dampening be required by any such extension.

In cases whereby the consistency of the Adj-RIB-In is to be restored (e.g., following the 'treat-as-withdraw' behaviour described in Section 3), and mechanisms such as those described herein are triggered, such a condition should be noted to an operator by means of a specific flag, SNMP trap, or other logging mechanism. In order to identify the subset of NLRI that are considered to be inconsistent, this information is of operational benefit and hence should be logged.

5. Reducing the Impact of Session Reset

Even where protocol enhancements allow errors in the BGP-4 protocol to cease to trigger NOTIFICATION messages, and hence reset a BGP session, it is clear that some error conditions may not be exited. In particular, errors due to existing state, or memory structures, associated with a specific BGP session will not be handled. It is therefore important to consider how these error conditions are currently handled by the protocol. It should be noted that the following discussion and analysis considers only those NOTIFICATION messages generated in response to errors in UPDATE messages (as defined by Section 6.3 in [RFC4271]).

The existing NOTIFICATION behaviour triggers a reset of all elements of the BGP-4 session, as described in Section 6 of [RFC4271]. It is expected that session teardown requires an implementation to re-initialise all structures and state required for session maintenance. Clearly, there is some utility to this requirement, as error conditions in BGP are, in general, exited from. However, this definition is responsible for the forwarding outages within networks utilising BGP for propagation of routing or service when each error is experienced. The requirement described in Section 3 is intended to reduce the cases whereby a NOTIFICATION is required, however, any mechanism implemented as a response to this requirement by definition cannot provide a session reset to the extent of that achieved by the current behaviour.

In order to address this, there is a requirement for a means by which a BGP speaker can signal that an unhandled error condition in an UPDATE message occurred - requiring a session reset - yet also continue to utilise the paths advertised by the neighbour that are currently in use within the RIB. In this case, the Adj-RIB-In received from the neighbour is not considered invalid, despite a NOTIFICATION, and session reset, being required. This set of requirements is akin to those answered by the BGP Graceful Restart mechanism described in [RFC4724]. Since the operational requirement in this case is to provide a means to achieve a complete session restart without disrupting the forwarding path of those routes in use within a BGP speaker's RIB, it is expected that utilising a procedure similar to the Graceful Restart mechanism meets the error handling requirement. By responding to an error condition (repeated or otherwise) with a message indicating that an error that cannot be handled has occurred, forcing session reset, whilst retaining forwarding information within the RIB allows forwarding to all routes within a system's RIB to continue during the period in which the session restarts. It is envisaged that the additional complexity introduced by the introduction of such a mechanism can be limited by extending existing BGP messages - one such approach is proposed in [I-D.ietf-idr-bgp-gr-notification]. By placing a time bound on the restart lifetime, should an error condition not be transient - for example, should an error have occurred with the BGP process, rather than a specific of the BGP session - the remote BGP speaker is still detected as an invalid device for forwarding.

In some cases, the erroneous condition may be due to corruption of the Adj-RIB-Out on the advertising BGP speaker - rather than caused by the receiving speaker's state. In these cases, where existing structures are replayed whilst performing graceful restart functionality, the error condition is not necessarily resolved. Therefore, it is recommended that during a session restart event, as described within this section, the advertising speaker purge and rebuild RIB structures, in order to resolve any corruption within these structures.

It should be noted that a protocol enhancement meeting this requirement is not able to solve all error conditions - however, a complete restart of the BGP and TCP session between two BGP speakers implements an identical recovery mechanism to that which is achieved by the existing behaviour. Where an error condition such as memory or configuration corruption has occurred in a BGP implementation, it is expected that a mechanism meeting this requirement continues to detect this, by means of a bound on time for session restart to occur. Whilst there may be some consideration that packets continue to be forwarded through a device which can be in an failure mode of this nature for a longer period due to this requirement, the architecture of modern IP routers should be considered. A divided forwarding and control plane is common in many devices, as well as process separation for software-based devices - corruption of a specific protocol daemon does not necessarily imply forwarding is affected. Indeed, where forwarding behaviour of a device is affected, it is envisaged that a failure detection mechanism (be it Bidirectional Forwarding Detection, or indeed BGP KEEPALIVE packets) will detect such a failure in almost all cases, with the symptomatic behaviour of such a failure being an invalid UPDATE message in very few other cases.

6. Operational Toolset for Monitoring BGP

A significant complexity that is introduced through the requirements defined in this document is that of monitoring BGP session status for an operator. Although the existing error handling behaviour causes a disproportionate failure, session failure is extremely visible to most operational personnel within a Network Operator due to both existing definitions of SNMP trap mechanisms for BGP, along with the forwarding impact typically caused by such a failure. By introducing mechanisms by which errors of this nature are not as visible, this is no longer the case. There is a requirement that where subsets of the RIB on a device are no longer reachable from a BGP speaker, or indeed an AS, that some visibility of this situation, alongside a mechanism to determine the cause is available to an operator. Whilst, to some extent, this can be solved by mandating a sub-requirement of each of the aforementioned requirements that a BGP speaker must log where such errors occur, and are hence handled, this does not solve all cases. In order to clarify this requirement, the example of the transmission of an erroneous Optional Transitive attribute can be considered. Since, by definition, there is no requirement for all BGP speakers to parse such an attribute, a receiving router may treat NLRI as withdrawn based on an erroneous attribute not examined by its neighbour. In this case, the upstream device or network, propagating the UPDATE, has no visibility of this error. Operationally, however, it is of interest to the upstream router operator that such invalid information was propagated.

The requirement for logging of error conditions in transmitted BGP messages, which are visible to only the receiver, cannot be achieved by any existing BGP message, or capability. It is envisaged that each erroneous event should be transmitted to the remote peer - including the information as to the set of NLRI that were considered invalid. Whilst with some mechanisms this is achieved by default (for example, One-Time Prefix ORF [I-D.zeng-idr-one-time-prefix-orf] (Outbound Route Filtering) will transmit the set of routes that are required), the operator requirement is to know which routes may have been unreachable in all cases. It is envisaged that an extension to meet this requirement will allow for such information to be transmitted between peers, and hence logged. Such a mechanism may provide further utility as a either a diagnostic, or logging toolset.

As such, it is possible to divide the messages that are required in order to provide further visibility into BGP for an operator. Such a division can be made both due to the required means of message transmission, alongside the criticality of each request.

Whilst the operational requirement for such monitoring tools to allow for visibility into BGP is clearly agreed upon, the means by which such messages are transmitted between two BGP speakers is likely to be dependent upon both the positions of the speakers in question (for instances, the requirements for such a protocol may differ where a session is between two ASBRs under separate administration). The introduction of additional message types to the BGP protocol clearly introduces further complexity - and leaves room for further implementation and standardisation errors that may compromise the robustness of the BGP protocol. In addition, the queuing and scheduling of these BGP messages must be interleaved with the transmission of the key protocol messages - such as KEEPALIVE and UPDATE packets. It is therefore a concern that should a large number of messages specifically for operational visibility be transmitted, this will delay the transmission of UPDATE packets, and hence adversely affect the end-to-end convergence time for NLRI carried within BGP. The operational requirement for why messages are advantageous to be in-band to a protocol should also be considered. In particular, it should be noted that where such information is to be transmitted between administrative boundaries a BGP session represents an existing channel between the two ASes. This channel is considered to be secure insofar as the routing information, and requests sent via the session are considered to come from a trusted source. Since error information relates to both a particular attachment, and is key to ensuring that such a session is operating as expected, it is considered of great operational benefit that this information is transmitted over this channel. In addition, the overall system scalability is improved by such in-band transmission. It is expected that erroneous information resulting in the 'treat-as-withdraw' mechanism being utilised is relatively infrequently transmitted between two peers (when compared to the frequency of UPDATE messages transmission). The impact of including an additional BGP message type for such operational visibility is relatively small from a resource utilisation perspective - additional processing overhead is only experienced when such a message is received. Where a separate session is maintained, particular network elements within a service provider topology may require hundreds, or thousands, of additional sessions for the transmission of this information. Such an resource consumption overhead is likely to be unacceptable to some network operators.

For the reasons explained above, it is expected that mechanisms specified to meet the requirements for event visibility consider the relative impacts of additional monitoring sessions, or message inclusion in band to BGP in order not to compromise the security, scalability and robustness of the BGP-4 protocol.

7. Operational Complexities Introduced by Altering RFC4271

The existing NOTIFICATION and subsequent teardown of a BGP session upon encountering an error has the advantage that a consistent approach to error handling is required of all implementations of the BGP-4 protocol. This is of operational advantage as it provides a clear expectation of the behaviour of the protocol. The requirements defined herein add further complexity to the error-handling within BGP, and hence are liable to compromise the existing deterministic protocol behaviour. It is therefore deemed that there is a further requirement to define a set of recommended behaviours based on the reception of a particular class of erroneous UPDATE message, alongside highlighting some of the implementation complexities that may need to be handled in the case that particular recommendations made within this memo are deployed.

Utilising the classes of erroneous UPDATE message described in Section 2, the recommended behaviour for a BGP-4 implementation can be divided into two branches. Primarily, where a semantic error is identified, an implementation is expected to utilise the reduced-impact error handling approach, as described in Section 3. In the case that such an approach results in known NLRI being withdrawn from the BGP speaker's RIB, and an implementation provides functionality such that these errors are recovered from through an automatically triggered means, such as those described within Section 4, some consideration of the scalability of these recovery mechanisms is required. Clearly, there is an computational and bandwidth overhead associated with the re-advertisement of NLRI between two BGP speakers - both due to the generation of UPDATE messages, their transmission between the two speakers, and the parsing and processing into the RIB required. This overhead is directly proportional to the number of UPDATE messages that are required. Where a semantic error is experienced, by definition the NLRI contained within the UPDATE can be extracted. It is therefore possible to minimise the proportion of the RIB that is re-advertised by targeting any recovery mechanism on the NLRI contained within the erroneous UPDATE. Such a targeted mechanism can be achieved through a means such as One-Time ORF, or other means of targeting UPDATE messages not discussed within this memo. It is recommended that where available, any automatic (or manual) triggered recovery mechanism behaviour utilises such targeted means in preference to any whole RIB refresh mechanism (such as ROUTE-REFRESH).

In the case that an erroneous UPDATE has been processed through a means such as treat-as-withdraw (described within Section 3), a recovering mechanism may be considered superfluous, if the assumption is made that the RIB inconsistency will only be recovered from based on a path re-convergence (or change in BGP attribute) for the advertising BGP speaker. However, where this assumption is not considered to provide adequate recovery behaviour, and a mechanism to restore RIB consistency automatically is implemented, some consideration must be made for where repeated erroneous messages occur. In this case, in order to limit the impact to the BGP speaker's network operation, at a pre-defined point it is recommended that such automatic recovery mechanisms towards the BGP speaker from which erroneous UPDATEs are repeatedly received are suppressed, and the fact that such suppression has occurred is highlighted to an operator. The point at which such behaviour is suppressed is to be defined on a per-implementation basis, taking into account feedback from the Network Operator community based on the deployment of the recommendations described in this document. It is expected that such trigger points are dependent upon the mechanisms implemented for a particular BGP-4 implementations, and the impact upon the speaker of these means of RIB recovery.

Where critical errors are experienced, such that a session reset is required, the mechanism discussed in Section 5 should be used. Again, since such a mechanism results in a restart of a BGP session, it expected that all NLRI carried over the session is re-advertised as it is re-established, incurring processing overhead on both the advertising and receiving BGP speaker. In order to minimise the consumption of control-plane computational resource on both speakers, it is recommended that mechanisms allowing a reduced set of BGP UPDATE messages to be re-transmitted between two speakers are employed wherever possible - for instance through employing mechanisms such as those described in [I-D.ietf-idr-enhanced-gr].

In the case that repeated critical errors occur, the overhead of performing any mechanism implemented based on the requirements in Section 5 is incurred following each erroneous UPDATE message. Since these mechanisms are, by definition, performed automatically in response to the erroneous message being received similar considerations as to the impact to the BGP speaker must be taken into account. As such, it is expected that after a certain trigger level, the ongoing receipt of critical errors within BGP UPDATE messages is deemed to be indicative of a long-lasting failure, and a session no longer considered viable. Where such an case is experienced, it is expected that the BGP session reverts to the standard session failure behaviour, as described in [RFC4271] and documents updating this base standard. Where such a reversion is implemented this condition should be flagged to an network operator. The number of restart attempts before the session reverts to being shut down should be determined based on the overhead of the recovery mechanisms implemented (for instance, where [I-D.ietf-idr-enhanced-gr] is implemented, the impact of session restart may be significantly lower), and operational experience of the deployment of the recommendations described in this document.

Since repeated erroneous UPDATE messages which experience critical errors may be indicative of long-lasting failure modes, it is recommended that a back-off from restarting BGP sessions experiencing such behaviour is implemented. As such, this is not applicable to restart behaviour through means such as those described in Section 5 since such restarts are time-bound based on the period for which the Adj-RIB-In from a BGP speaker is maintained as valid (e.g., when considering BGP Graceful Restart, such restarts are time-bound by the Restart Time described in [RFC4724]). However, following a session reverting to being pulled down based on repeated error conditions, it is recommended that following restart attempts are subject to an exponentially increasing interval between subsequent attempts. It is therefore recommended that in such cases an implementation implements the increasing values of IdleHoldTimer as described in the BGP-4 FSM documented in [RFC4271].

7.1. Reducing the Network Impact of Session Teardown

As discussed within the preceding section, where repeated critical UPDATE message errors are received, it is recommended that the impact to the both advertising and receiving BGP-4 speakers be limited by reverting to tearing the BGP-4 session experiencing such errors down. The BGP-4 specification presented in [RFC4271] achieves such a session shutdown by sending a NOTIFICATION message, however, this has the net result that all downstream BGP speakers (i.e. those to whom the routes carried over the now ceased BGP session was readvertised) must withdraw this route from their RIB, and perform a best-path selection if required. In some cases, there may be no alternate path available, and hence a period of time for which no valid BGP route exists. Particularly, this is very likely to occur where an upstream BGP speaker performs a best-path selection and advertises only a single path to its neighbours - there is a requirement for the upstream speaker to perform a best-path selection, and re-advertise a new set of NLRI before the downstream system is able to converge to a new path. It should be noted that where UPDATE messages withdrawing NLRI are not subject to the BGP session's configured MinRouteAdvertisementInterval (MRAI) [RFC4271], but re-advertisements are, this may result in a BGP speaker being without a path for a period up to the MRAI.

Clearly, it is advantageous to avoid this period of time for which there may be no reachability for a set of routes, especially since the BGP speaker terminating a particular session is doing so due to a particular error handling policy. The graceful shutdown mechanism detailed in [I-D.ietf-grow-bgp-gshut] provides a mechanism by which a BGP speaker is able to signal that a set of routes are to be withdrawn, and hence allow downstream systems to pre-emptively perform a best-path selection, and hence advertise new reachability information in a make-before-break manner.

It is therefore envisaged, that where a session is to be shutdown, based on a trigger relating to erroneous UPDATE messages being received (be they repeated or not) that the graceful shutdown procedure in utilised, so as to reduce the forwarding impact of routes received on the session being withdrawn.

8. IANA Considerations

This memo includes no request to IANA.

9. Security Considerations

The requirements outlined in this document provide mechanisms by which erroneous BGP messages may be responded to with limited impact to forwarding operation. This is of benefit to the security of a BGP speaker in general. Where UPDATE messages may have been propagated by a single malicious Autonomous System or router within a network (or the Internet default free zone - DFZ), which are then propagated to all devices within the same routing domain, all other NLRI available over the same session become unreachable. This mechanism may provide means by which an Autonomous System can be isolated from required routing domains (such as the Internet), should the relevant UPDATE messages be propagated via specific paths. By reducing the impact of such failures, it is envisaged that this possibility may be constrained to a specific set of NLRI, or a specific topology.

Some mechanisms meeting the requirements specified in this document, particularly those within Section 6 may provide further security concerns, however, it is envisaged that these are addressed in per-enhancement memos.

10. Acknowledgements

The author would like to thank the following network operators for their insight, and valuable input in defining the requirements for a variety of operational deployments of the BGP-4 protocol; Shane Amante, Bruno Decraene, Rob Evans, David Freedman, Wes George, Tom Hodgson, Sven Huster, Jonathan Newton, Neil McRae, Thomas Mangin, Tom Scholl and Ilya Varlashkin.

In addition, many thanks are extended to Jeff Haas, Wim Hendrickx, Tony Li, Alton Lo, Keyur Patel, John Scudder, Adam Simpson and Robert Raszuk for their expertise relating to implementations of the BGP-4 protocol.

11. References

11.1. Normative References

[RFC2918] Chen, E., "Route Refresh Capability for BGP-4", RFC 2918, September 2000.
[RFC2858] Bates, T., Rekhter, Y., Chandra, R. and D. Katz, "Multiprotocol Extensions for BGP-4", RFC 2858, June 2000.
[RFC4271] Rekhter, Y., Li, T. and S. Hares, "A Border Gateway Protocol 4 (BGP-4)", RFC 4271, January 2006.
[RFC4364] Rosen, E. and Y. Rekhter, "BGP/MPLS IP Virtual Private Networks (VPNs)", RFC 4364, February 2006.
[RFC4456] Bates, T., Chen, E. and R. Chandra, "BGP Route Reflection: An Alternative to Full Mesh Internal BGP (IBGP)", RFC 4456, April 2006.
[RFC4724] Sangli, S., Chen, E., Fernando, R., Scudder, J. and Y. Rekhter, "Graceful Restart Mechanism for BGP", RFC 4724, January 2007.
[RFC4760] Bates, T., Chandra, R., Katz, D. and Y. Rekhter, "Multiprotocol Extensions for BGP-4", RFC 4760, January 2007.

11.2. Informational References

[RFC5881] Katz, D. and D. Ward, "Bidirectional Forwarding Detection (BFD) for IPv4 and IPv6 (Single Hop)", RFC 5881, June 2010.
[I-D.ietf-idr-bgp-enhanced-route-refresh] Patel, K, Chen, E and B Venkatachalapathy, "Enhanced Route Refresh Capability for BGP-4", Internet-Draft draft-ietf-idr-bgp-enhanced-route-refresh-01, December 2011.
[I-D.ietf-idr-optional-transitive] Scudder, J, Chen, E, Mohapatra, P and K Patel, "Revised Error Handling for BGP UPDATE Messages", Internet-Draft draft-ietf-idr-optional-transitive-04, October 2011.
[I-D.chen-ebgp-error-handling] Chen, E, Mohapatra, P and K Patel, "Revised Error Handling for BGP Updates from External Neighbors", Internet-Draft draft-chen-ebgp-error-handling-01, September 2011.
[I-D.zeng-idr-one-time-prefix-orf] Zeng, Q, Dong, J, Heitz, J, Patel, K, Shakir, R and Z Huang, "One-time Address-Prefix Based Outbound Route Filter for BGP-4", Internet-Draft draft-zeng-idr-one-time-prefix-orf-02, July 2012.
[I-D.ietf-idr-operational-message] Freedman, D, Raszuk, R and R Shakir, "BGP OPERATIONAL Message", Internet-Draft draft-ietf-idr-operational-message-00, March 2012.
[I-D.ietf-idr-enhanced-gr] Patel, K, Chen, E, Fernando, R and J Scudder, "Accelerated Routing Convergence for BGP Graceful Restart", Internet-Draft draft-ietf-idr-enhanced-gr-00, December 2011.
[I-D.ietf-grow-bmp] Scudder, J, Fernando, R and S Stuart, "BGP Monitoring Protocol", Internet-Draft draft-ietf-grow-bmp-06, December 2011.
[I-D.ietf-grow-bgp-gshut] Francois, P, Decraene, B, Pelsser, C, Patel, K and C Filsfils, "Graceful BGP session shutdown", Internet-Draft draft-ietf-grow-bgp-gshut-03, December 2011.
[I-D.ietf-idr-bgp-gr-notification] Patel, K, Fernando, R and J Scudder, "Notification Message support for BGP Graceful Restart", Internet-Draft draft-ietf-idr-bgp-gr-notification-00, December 2011.

Author's Address

Rob Shakir BT pp C3L BT Centre 81, Newgate Street London, EC1A 7AJ UK EMail: rob.shakir@bt.com URI: http://www.bt.com/