Hi, Dino. I have additional responses inline.
Kyle
> For the internet core (DFZ RIB) use-case, LISP proposes replacing BGP sessions
> and global eventually-consistent state sharing with a global control plane and
LISP *does not propse to eliminate BGP*, in fact it needs it so RLOC reachability across the network is available, or there would be no underlay for the LISP overlay.
The whole point of LISP is to create a routing overlay for the EID address space, the RIB of which is managed by a global mapping system, not BGP sessions. If control plane traffic managed by BGP (or static routes, or whatever networks use once the DFZ RIB is limited to entities in the core) continues to flow, that is of small comfort to end users trying to get data over the data plane. >From the perspective of end users, BGP is being replaced routing of the traffic that matters to them.
Maybe this is just splitting hairs: I certainly don't want to rathole on this point.
You are missing pieces of the design and hence why you came to the conclusion you did. There are three documents that enhance the security of LISP, they are:
(1) RFC 8061 - Locator/ID Separation Protocol (LISP) Data-Plane Confidentiality
I did not review this document because I was following up on authentication of control plane messages. If that is covered in a document titled "Data-Plane Confidentiality", some reorganization may be in order.
(2) RFC 8111 - Locator/ID Separation Protocol Delegated Database Tree (LISP-DDT)
(3) draft-farinacci-lisp-ecdsa-auth
I take it from the name that this has not even been adopted by the WG yet? Evaluating something I'm supposed to review for IETF LC should not depend on an unadopted draft.
That said, I have been trying to make sense of this document for the past few days. I have a few basic observations:
* It does not resolve the trust anchor problem. Instead of proposing a PKI, you seem to be proposing a trusted third party authoritative for the Hash-EID namespace. (Q.v.. section 2, the Hash-EID definition: "Another entity")
* 128 bits is already too short for a cryptographic hash, and this document proposes reducing that to 80 bits for a /48..
* The entire scheme relies on the integrity of the Hash-EID mappings: if an attacker can compromise a Hash-EID mapping with a different public key that hashes to the same truncated hash, they can sign anything for that space of Crypto-EIDs.
* The only protection provided by this scheme is for Map-Register and Map-Request, and not for the response to Map-Request, which (to me) is the critical gap. The problem I noted in my original review is that the ITR-OTK only allows the Map-Server to prove that a response is associated with a Map-Resolver's request, not that the Map-Server is the proper authority for the response. What the system needs is a global trust network that does not rely on either universal transitive trust (e.g., I'll just trust whatever my trusted-but-potentially-mistaken neighbors tell me about the rest of the world based on what they heard) or n^2 relationships (e.g., symmetric keys with every other entity).
* Why a scheme like this for addressing Map-Register? Presumably an ETR will either be owned by the same entity running the Map-Server authoritative for its own EID space, or it will contract with one. Either case is a single, pre-established relationship per ETR, which scales fine.
* What is the use case for Map-Request authentication? Its only use seems to be to keep address space -> network topology mapping hidden.. From whom? The end user doesn't need to know, but every entity in the core needs to have access to a Crypto-EID's associated RLOC or it won't be able to determine the next hop.
* My general recommendation for design documents is that the overall architecture should be clear to a reader with a relevant technical background without having to first consume a pile of other documents. To the authentication point specifically, you may wish to include some text regarding the properties generically offered by an asymmetric authentication scheme with the appropriate feature set, and reference those properties in the security considerations section when addressing certain classes of attacks. You can then use these requirements as input to a document describing a specific authentication scheme.
* This document doesn't present a clear description of the problem it's trying to solve in the introduction. I understand what it's getting at now that I've read the whole thing, but a better introduction (for the actual problem you need to solve, I'd add) would say something like, q( The LISP architecture introduces two address spaces along with a mapping from one to the other. The integrity of this mapping is critical for achieving the intended routing of traffic through a LISP network, being directly related to service availability and end user privacy. This document describes a mechanism for implementation of strong authentication of this mapping. ) The security considerations should then discuss the turtles-all-the-way-down problem of BGP hijacking on the underlay (should BGP be employed for this purpose) so operators are clear on the limitations of the scheme.
You find strong asymmetric authenticaiton and authorization in (3). And you’ll find authentication of the mapping system nodes in (2). And note that (1) can be used and layered under the control-plane to give encrypted control-plane flows (or use DTLS as LISP-SEC refers to for the messages it requires for its functionality).
"TLS" does not appear anywhere in the draft of LISP-SEC I reviewed:
> One area of concern, of which I have not been able to find discussion, is that
> of the implications of shared capacity for the control and data planes, and how
> this can allow a volumetric data plane attack to deny a router access to the
> global mapping system, slowly choking off service to uncached portions of the
Well yes, this happens with all our IETF protocols. It is a valid concern and there are many operational techniques in network infrastructure that *help* solve (but not eliminate) these problems.
I would like to see a discussion of whether and how the nature and scale of this problem differs from that of the status quo. BGP sessions and RIB push have properties that are well-established from decades of experience: surely LISP does not have exactly the same properties. The security considerations should make clear, for instance, how a loss of control plane connectivity differs from the loss of a BGP session, and how this impacts visibility and behavior of the data plane.
> I would also like clarification on what defines the separation between the
> control plane and data plane, and whether authentication itself is used to
A control-plane obtains information to store in a table. The data-plane uses that table. That is the definition in the simpliest form.
I mean specific to LISP, not generically. For instance, does "LISP Control-Plane signaling" include only valid messages, or valid + inauthentic (and presumably dropped) messages? Traditional attack traffic (e.g., a DDoS attack against a website) is part of the same data plane as all legitimate end user traffic; is attack traffic directed at control plane endpoints considered part of the control plane, or is it a third category of traffic? If the latter, then what does an operator need to do to ensure that control plane is always available?
