| Internet-Draft | Route Leak Solution Discussion | March 2022 |
| Sriram | Expires 8 September 2022 | [Page] |
This document captures the design rationale of the route leaks solution document (see draft-ietf-idr-route-leak-detection-mitigation, draft-ietf-grow-route-leak-detection-mitigation). The designers needed to balance many competing factors, and this document provides insights into the design questions and their resolution.¶
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 https://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 8 September 2022.¶
Copyright (c) 2022 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 (https://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 Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This document captures the design rationale of the route leaks solution document [I-D.ietf-idr-route-leak-detection-mitigation] [I-D.ietf-grow-route-leak-detection-mitigation]. The designers needed to balance many competing factors, and this document provides insights into the design questions and their resolution.¶
The solution described in [I-D.ietf-idr-route-leak-detection-mitigation] is based on setting an attribute in BGP route announcement to manage the transmission/receipt of the announcement based on the type of neighbor (e.g., customer, transit provider, etc.). Documented prior work related to this basic idea and mechanism dates back to at least the 1980's. Some examples of prior work are: (1) Information flow rules described in [proceedings-sixth-ietf] (see pp. 195-196); (2) Link Type described in [RFC1105-obsolete] (see pp. 4-5); (3) Hierarchical Recording described in [draft-kunzinger-idrp-ISO10747-01] (see Section 6.3.1.12). The problem of route leaks and possible solution mechanisms based on encoding peering-link type information, e.g., P2C (i.e., Transit-Provider to Customer), C2P (i.e., Customer to Transit-Provider), p2p (i.e., peer to peer) etc., in BGPsec updates and protecting the same under BGPsec path signatures have been discussed in IETF SIDR WG at least since 2011. [draft-dickson-sidr-route-leak-solns] attempted to describe these mechanisms in a BGPsec context. The draft expired in 2012. [draft-dickson-sidr-route-leak-solns] defined neighbor relationships on a per link basis, but in [I-D.ietf-idr-route-leak-detection-mitigation] the relationship is encoded per prefix, as routes for prefixes with different peering relationships may be sent over the same link. Also [draft-dickson-sidr-route-leak-solns] proposed a second signature block for the link type encoding, separate from the path signature block in BGPsec. By contrast, in [I-D.ietf-idr-route-leak-detection-mitigation], when BGPsec-based solution is considered, cryptographic protection is provided for Route-Leak Protection (RLP) encoding using the same signature block as that for path signatures (see Section 3.2.2 in [I-D.ietf-idr-route-leak-detection-mitigation]).¶
This section provides design justifications for the solution specified in [I-D.ietf-idr-route-leak-detection-mitigation], and also answers some questions that are anticipated or have been raised in the IETF IDR and SIDR working group meetings.¶
In Section 3.3 in [I-D.ietf-idr-route-leak-detection-mitigation], Rules 1 and 2 were stated and the route leak mitigation policy was based on these rules to preserve the property of stable route convergence (i.e., avoid possibility of persistent route oscillations). Rule 1 is stated as follows:¶
The rationale for Rule 1 can be developed as follows.¶
Preference condition for route stability: Prefer customer routes over peer or provider routes (see pp. 25-27 in [Gao-Rexford]).¶
Topology condition for route stability: No cycle of customer-provider relationships (see pp. 25-27 in [Gao-Rexford]).¶
Route-Leak Detection Theorem: Let it be given that ISP A receives a route r1 from customer AS C and another route r2 from provider AS B (for the same prefix), and each of the routes r1 and r2 contains AS C and AS X in its AS path and contains [X] in its RLP Attribute. Then, clearly r1 is in violation of [X]. It follows that r2 is also necessarily in violation of [X].¶
Proof: Let us suppose that r2 is not in violation of [X]. That implies that r2's path from C to B to A included only P2C links. That would mean that there is a cycle of customer-provider relationships involving the ASes in the AS path in r2. However, any such cycle is ruled out in practice by the topology condition for route stability as stated above. QED.¶
Corollary 1: The route leak detection theorem holds also when "provider AS B" in the theorem is replaced by "peer AS B". (Here peer means a lateral peer.)¶
Proof: Since r2 contains [X] in the RLP Attribute set by an AS prior to peer AS B, it follows that r2 is in violation of [X]. QED.¶
It can be observed that Rule 1 follows from the combination of the Theorem, Corollary 1 and the preference condition for route stability (stated above).¶
In Section 3.3 in [I-D.ietf-idr-route-leak-detection-mitigation], Rule 2 is stated as follows:¶
The rationale for Rule 2 can be developed as follows.¶
Corollary 2: The route leak detection theorem holds also when "customer AS C" in the theorem is replaced by "peer AS C".¶
Proof for Corollary 2: Let us suppose that r2 is not in violation of [X]. That implies that r2's path from C to B to A included only P2C links. This results in a topology in which A's lateral peer B is also A's transit provider's transit provider. This gives rise to possibility of looping of routes since A can send routes to its transit B, B can forward the routes to its transit C, and C can forward the routes to its peer A. But such looping is forbidden by the topology condition stated above.¶
It can be observed that Rule 2 follows from Corollary 2. In essence, if the provider route (r2) is a detoured (longer) version of the lateral peer route (r1), and violates the same RLP [X] as does the peer route, then prefer the shorter route (r2) via the peer.¶
Rules 1 and 2 are¶
It has been asked if a route-leak solution without BGPsec, i.e., when RLP Fields are not protected, can turn into a new attack vector. The answer seems to be: not really! Even the NLRI and AS_PATH in BGP updates are attack vectors, and RPKI/OV/BGPsec seek to fix that. Consider the following. Say, if 99% of route leaks are accidental and 1% are malicious, and if route-leak solution without BGPsec eliminates the 99%, then perhaps it is worth it (step in the right direction). When BGPsec comes into deployment, the route-leak protection (RLP) bits can be mapped into BGPsec (using the Flags field) and then necessary security will be in place as well (within each BGPsec island as and when they emerge).¶
Further, let us consider the worst-case damage that can be caused by maliciously manipulating the RLP Field values in an implementation without cryptographic protection (i.e., sans BGPsec). Manipulation of the RLP bits can result in one of two types of attacks: (a) Upgrade attack and (b) Downgrade attack. Descriptions and discussions about these attacks follow. In what follows, P2C stands for transit provider to customer (Down); C2P stands for customer to transit provider (Up), and p2p stands for peer to peer (lateral or non-transit relationship).¶
(a) Upgrade attack: An AS that wants to intentionally leak a route would alter the RLP encodings for the preceding hops from 1 (i.e., 'Do not Propagate Up or Lateral') to 0 (default) wherever applicable. This poses no problem for a route that keeps propagating in the 'Down' (P2C) direction. However, for a route that propagates 'Up' (C2P) or 'Lateral' (p2p), the worst that can happen is that a route leak goes undetected. That is, a receiving router would not be able to detect the leak for the route in question by the RLP mechanism described here. However, the receiving router may still detect and mitigate it in some cases by applying other means such as prefix filters [RFC7454] [NIST-800-54]. If some malicious leaks go undetected (when RLP is deployed without BGPsec) that is possibly a small price to pay for the ability to detect the bulk of route leaks that are accidental.¶
(b) Downgrade attack: RLP encoding is set to 1 (i.e., 'Do not Propagate Up or Lateral') when it should be set to 0 (default). This would result in a route being mis-detected and marked as a route leak. By default, RLP encoding is set to 0, and that helps reduce errors of this kind (i.e., accidental downgrade incidents). Every AS or ISP wants reachability for prefixes it originates and for its customer prefixes. So, an AS or ISP is not likely to change an RLP value 0 to 1 intentionally. If a route leak is detected (due to intentional or accidental downgrade) by a receiving router, it would prefer an alternate 'clean' route from a transit provider or peer over a 'marked' route from a customer. It may end up with a suboptimal path. In order to have reachability, the receiving router would accept a 'marked' route if there is no alternative that is 'clean'. So, RLP downgrade attacks (intentional or accidental) would be quite rare, and the consequences do not appear to be grave.¶
Combining the results of route-leak detection, OV, and BGPsec validation for path selection decision is up to local policy in a receiving router. As an example, a router may always give precedence to outcomes of OV and BGPsec validation over that of route-leak detection. That is, if an update fails OV or BGPsec validation, then the update is not considered a candidate for path selection. Instead, an alternate update is chosen that passed OV and BGPsec validation and additionally was not marked as route leak.¶
If only OV is deployed (and not BGPsec), then there are six possible combinations between OV and route-leak detection outcomes. Because there are three possible outcomes for OV (NotFound, Valid, and Invalid) and two possible outcomes for route-leak detection (marked as leak and not marked). If OV and BGPsec are both deployed, then there are twelve possible combinations between OV, BGPsec validation, and route-leak detection outcomes. As stated earlier, since BGPsec protects the RLP encoding, there would be added certainty in route-leak detection outcome if an update is BGPsec valid (see Section 3.2).¶
There are studies in the literature [Anwar] [Giotsas] [Wijchers] observing and analyzing the behavior of routes announced in BGP updates using data gathered from the Internet. The studies have focused on how often there appear to be valley-free (e.g., Gao-Rexford [Gao] model) violations, and if they can be explained [Anwar]. One important consideration for explanation of the violations is per-prefix routing policies, i.e., routes for prefixes with different peering relationships may be sent over the same link. One encouraging result reported in [Anwar] is that when per-prefix routing policies are taken into consideration in the data analysis, more than 80% of the observed routing decisions fit the valley-free model (see Section 4.3 and SPA-1 data in Figure 2). [Anwar] also observes, "it is well known that this model [the basic Gao-Rexford model and some variations of it] fails to capture many aspects of the interdomain routing system. These aspects include AS relationships that vary based on the geographic region or destination prefix, and traffic engineering via hot-potato routing or load balancing." So, there may be potential for explaining the remaining (20% or less) violations of valley-free as well.¶
One major design factor is that the Route-Leak Protection (RLP) encoding is per prefix. Hence, the solution is consistent with ISPs' per-prefix routing policies. Large global and other major ISPs will be the likely early adopters, and they are expected to have expertise in setting policies (including per prefix policies, if applicable), and make proper use of the RLP indications on a per prefix basis. When the large ISPs participate in this solution deployment, it is envisioned that they would form a ring of protection against route leaks, and co-operatively avoid many of the common types of route leaks that are observed. Route leaks may still happen occasionally within the customer cones (if some customer ASes are not participating or not diligently implementing RLP), but such leaks are unlikely to propagate from one large participating ISP to another.¶
It is reasonable to ask if techniques considered in BCPs such as[RFC7454] (BGP Operations and Security) and [NIST-800-54] may be adequate to address route leaks. The prefix filtering recommendations in the BCPs may be complementary but not adequate. The difficulty is in ISPs' ability to construct prefix filters that represent their customer cones (CC) accurately, especially when there are many levels in the hierarchy within the CC. In the RLP-encoding based solution described here, each AS sets RLP for each route propagated and thus signals if it must not be subsequently propagated to a transit provider or peer.¶
AS path based Outbound Route Filter (ORF) described in [I-D.ietf-idr-aspath-orf] is also an interesting complementary technique. It can be used as an automated collaborative messaging system (implemented in BGP) for ISPs to try to develop a complete view of the ASes and AS paths in their CCs. Once an ISP has that view, then AS path filters can be possibly used to detect route leaks. One limitation of this technique is that it cannot duly take into account the fact that routes for prefixes with different peering relationships may be sent over the same link between ASes. Also, the success of AS path based ORF depends on whether ASes at all levels of the hierarchy in a CC participate and provide accurate information (in the ORF messages) about the AS paths they expect to have in their BGP updates.¶
The route-leak detection and mitigation mechanism described in [I-D.ietf-idr-route-leak-detection-mitigation] is based on setting RLP Fields on a per-hop basis. There is another possible mechanism based on a single RLP flag per update.¶
Method A - Per-Hop RLP Field: The sender (eBGP router) on each hop in the AS path sets its RLP Field = 1 if sending the update to a customer or lateral peer (see Section 3.2 in [I-D.ietf-idr-route-leak-detection-mitigation]). No AS (if operating correctly) would rewrite the RLP Field set by any preceding AS.¶
Method Z - Single RLP Flag per Update: As it propagates, the update would have at most one RLP flag. Once an eBGP router (in the update path) determines that it is sending an update towards a customer or lateral peer AS, it sets the RLP flag. The flag value equals the AS number of the eBGP router that is setting it. Once the flag is set, subsequent ASes in the path must propagate the flag as is.¶
To compare Methods A and Z, consider the example illustrated in Figure 1. Consider a partial deployment scenario in which AS1, AS2, AS3 and AS5 participate in RLP, and AS4 does not. AS1 (2 levels deep in AS3's customer cone) has imperfect RLP operation. Each complying AS's route leak mitigation policy is to prefer an update not marked as route leak (see Section 3.3 in [I-D.ietf-idr-route-leak-detection-mitigation]). If there is no alternative, then a transit-provider may accept and propagate a marked update from a customer to avoid unreachability. In this example, multi-homed AS4 leaks a route received for prefix Q from transit-provider AS3 to transit-provider AS5.¶
+-----------+ RLP=1 +-----------+
| AS3 |---------->| AS5 |
|(Major ISP)| U2 |(Major ISP)|
+-----------+ +-----------+
/\ \ /\ U1
Route for Q propagated / \RLP=1 /
due to lack of /RLP=0 \ / (route leak;
alternate route / \/ / bad behavior)
+---------+ +-------------+
| AS2 | | AS4 |
+---------+ +-------------+
/\ (legacy; does not support RLP)
/
/
/RLP=1 (set incorrectly)
/
+----------+
| AS1 |
+----------+
/\
/
/ Prefix Q
If Method A is implemented in the network, the two BGP updates for prefix Q received at AS5 are (note that AS4 is not participating in RLP):¶
Alternatively, if Method Z is implemented in the network, the two BGP updates for prefix Q received at AS5 are:¶
All received routes for prefix Q at AS5 are marked as route leak in either case (Method A or Z). In the case of Method A, AS5 can use additional information gleaned from the RLP fields in the updates to possibly make a better best path selection. For example, AS5 can determine that U1A update received from its customer AS4 exhibits violation of two RLP fields (those set by AS1 and AS3) and one of them was set just two hops away. But U2A update exhibits that only one RLP field was violated and that was set three hops back. Based on this logic, AS5 may prefer U2A over U1A (even though U1A is a customer route). This would be a good decision. However, Method Z does not facilitate this kind of more rational decision process. With Method Z, both updates U1B and U2B exhibit that they violated only one RLP field (set by AS1 several hops away). AS5 may then prefer U1B over U2B since U1B is from a customer, and that would be bad decision. This illustrates that, due to more information in per-hop RLP Fields, Method A seems to be operationally more beneficial than Method Z.¶
Further, for detection and notification of neighbor AS's non-compliance, Method A (per-hop RLP) is better than Method Z (single RLP). With Method A, the bad behavior of AS4 would be explicitly evident to AS5 since it violated AS3's (only two hops away) RLP field as well. AS5 would alert AS4 and AS2 would alert AS1 about lack of compliance (when Method A is used). With Method Z, the alerting process may not be as expeditious.¶
Note: The intra-AS messaging for route leak prevention can be done using a non-transitive BGP Community or Attribute. The Community-based method is described below. For the BGP Attribute-based method, see [I-D.ietf-idr-bgp-open-policy].¶
The following procedure (or similar) for intra-AS messaging (i.e., between ingress and egress routers) for prevention of route leaks is a fairly common practice used by large ISPs. (Note: This information was gathered from discussions on the NANOG mailing list [Nanog-thread-June2016] as well as through private discussions with operators of large ISP networks.)¶
Routes are tagged on ingress to an AS with communities for origin, including the type of eBGP peer it was learned from (customer, provider or lateral peer), geographic location, etc. The community attributes are carried across the AS with the routes. These communities are used along with additional logic in route policies to determine which routes are to be announced to which eBGP peers and which are to be dropped. In this process, the ISP's AS also ensures that routes learned from a transit-provider or a lateral peer (i.e., non-transit) at an ingress router are not leaked at an egress router to another transit-provider or lateral peer.¶
Additionally, in many cases, ISP network operators' outbound policies require explicit matches for expected communities before passing routes. This helps ensure that that if an update has been entered into the RIB-in but has missed its ingress community tagging (due to a missing/misapplied ingress policy), it will not be inadvertently leaked.¶
The above procedure (or a simplified version of it) is also applicable when an AS consists of a single eBGP router. It is recommended that all AS operators SHOULD implement the procedure described above (or similar that is appropriate for their network) to prevent route leaks that they have direct control over.¶
A stopgap method is described here for detection and mitigation of route leaks for the intermediate phase when OV is deployed but BGP protocol on the wire is unchanged. The stopgap solution can be in the form of construction of a prefix filter list from ROAs. A suggested procedure for constructing such a list comprises of the following steps:¶
Special considerations regarding the above procedure may be needed for DDoS mitigation service providers. They typically originate or announce a DDoS victim's prefix to their own ISP on a short notice during a DDoS emergency. Some provisions would need to be made for such cases, and they can be determined with the help of inputs from DDoS mitigation service providers.¶
For developing a list of all the ASes (Cust_AS_List) that are in the customer cone of an ISP, the AS path based Outbound Route Filter (ORF) technique [I-D.ietf-idr-aspath-orf] can be helpful (see discussion in Section 3.5).¶
Another technique based on AS_PATH filters is described in [Snijders]. This method is applicable to very large ISPs that have lateral peering. For a pair of such very large ISPs, say A and B, the method depends on ISP A communicating out-of-band (e.g., by email) with ISP B about whether or not it (ISP A) has any transit providers. This out-of-band knowledge enables ISP B to apply suitable AS_PATH filtering criteria for routes involving the presence of ISP A in the path and prevent certain kinds of route leaks (see [Snijders] for details).¶
This document requires no security considerations. See [I-D.ietf-idr-route-leak-detection-mitigation] for security considerations for the solution for route leaks detection and mitigation.¶
This document has no IANA actions.¶
The authors wish to thank Jared Mauch, Jeff Haas, Job Snijders, Warren Kumari, Amogh Dhamdhere, Jakob Heitz, Geoff Huston, Ruediger Volk, Sue Hares, John Scudder, Wes George, Chris Morrow, Sandy Murphy, Danny McPherson, and Eric Osterweil for comments, suggestions, and critique. The authors are also thankful to Padma Krishnaswamy, Oliver Borchert, and Okhee Kim for their review and comments.¶
The following people made significant contributions to this document and should be considered co-authors:¶
Alexander Azimov Yandex Email: a.e.azimov@gmail.com Brian Dickson Independent Email: brian.peter.dickson@gmail.com Doug Montgomery USA National Institute of Standards and Technology Email: dougm@nist.gov Keyur Patel Arrcus Email: keyur@arrcus.com Andrei Robachevsky Internet Society Email: robachevsky@isoc.org Eugene Bogomazov Qrator Labs Email: eb@qrator.net Randy Bush Internet Initiative Japan Email: randy@psg.com¶