Re: [manet] draft-ietf-manet-aodvv2-13 review - a couple of big ticket Items

[Justin Dean <[e-mail]>]

A few comments from me inline.

*Loops*
> Just to bring this out: I share Chris’ worry about conflicting and
> concurrent statements from the authors on “There are no loops possible” and
> “We need to fix two situations where loops can occur” and “we are still
> investigating some loop conditions”
>
> I particularly worry that this is not a discussion had in public, but
> apparently in some other forum…
>
> Discussions on list are key.  This needs to happen more.


>
> *Intermediate Route Replies, and all of section 10*
> Section 10 contains a set of “vaguely specified extensions”, which is
> incoherent with the intended status indicated for this document.
>
> Specifically, and this is not unrelated to the point about loops above,
> intermediate RREPs (section 10.3) are a potential source for loops.
>
> Expanding Ring Multicast (section 10.1) is not documented in a way that
> can be implemented (and also, see “Forwarding-vs-regeneration” below, it is
> in the present form of this protocol impossible), etc.
>
> I'm all for removal of vaguely specified extensions.  The use of reduced
relay sets is one of those that should be cut at this time.  Expanding Ring
Multicast could be kept IMO.  It's not much different than the way OLSR
specifies that MPR selectors can be chosen in different ways but must cover
all the 2 hop neighbors.  Something as simple as.  If limited scope RREQ
are sent each successive RREQ for the same address MUST at least double the
hop count.

>
> *Forwarding-vs-regeneration*
> Recent exchanges on the list made me understand that protocol control
> messages are *not* forwarded, but are consumed at each hop, then a new
> message with (almost-but-not-quite) the same content is generated and
> transmitted.
>
> I have thought some more on this (& read some of the exchanges on the list
> on this topic by Chris, Ulrich, and others), and I am convinced that this
> is not the right way to go, *at least* for the following reasons:
>
>
> o *It renders the hop-limit/hop-count fields in the RFC5444 message
> header useless.*
> This would not be bad if the functionality offered by those fields was not
> useful
> — sadly, it is. For example, for scope limited flooding (expanding ring
> search, and
> such) which may be of interest, and which require hop-limit.
> A hop-count field may also provide a “cheap” (in terms of overhead)
> additional piece
> of information to use conjunctively with a “real” metric.
>
> The only practical solution would be to re-introduce these functions by
> way of inserting a
> MessageTLV — which (i) is not specified in this document, and (ii) which
> would just
> serve to render messages bigger than strictly needed.
>
> Scope limited flooding  does seem to be a necessary requirement, if for no
> other reason than to prevent information from “circulating forever in the
> network”.
>
> JWD: I partly disagree that it can't be used for this purpose.  RFC5444
specifies:
 <msg-hop-limit> is omitted if the mhashoplimit flag is cleared

      '0'); otherwise, is an 8-bit unsigned integer field that can
      contain the maximum number of hops that the message should be
      further transmitted.


which in retrospect was a mistake as 5444 was meant to be a format and not
a protocol specification as we mentioned many times.  No other  protocol
message is going to forward these messages as AODVv2 "owns" the message
type so I don't see an issue with bending the terms in 5444 to allow this
forwarding of "information"  The way multiplexers work messages can get
torn apart and sent back out in a different way.  My OLSRv2 with another
version will likely do just that.

o *It makes end-to-end authentication unnecessarily hard.*
> I think Chris called this out already, but it bears repeating: S generates
> a message
> (say, a RREQ), and includes an ICV calculated over the content of the
> message.
> For any recipient to be able to validate the ICV, the message has to be
> exactly
> the same — not just in content, but in structure — as what was generated.
>
> “Regenerating” rather than “forwarding” messages means, that the
> intermediate
> router “regenerating” the RREQ may chose a different structure (e.g.,
> include TLVs
> in a different order).
>
> The proposal from
> https://tools.ietf.org/html/draft-perkins-manet-aodv-e2esec-00
> is to add constraints on (i) the set of elements to include in a signature
> and (ii) the
> order of these elements.
>
> One problem with that approach is (i): if an extension adds a message TLV,
> or an
> Address TLV, to a message, then that will not be “covered” by the proposed
> e2esec TLV.
> Rather for *each* extension developed, an “updates e2esec” clause needs to
> be done.
>
> I’d say that this approach would be prone to errors — and add entropy to
> the process
> of designing protocol extensions. The alternative, a message being
> generated by the
> source and *forwarded* (as we do in OLSRv2, for example) would allow ICV
> TLVs
> (even, allow reuse of those specified for OLSRv2) for covering a message
> and
> extensions.
>
> “But what about the metrics value which will change on each hop”, you may
> say?
> Fortunately, that is relatively easy to handle: simply zero out the value
> of that TLV when
> generating or verifying the ICV MessageTLV. And use Packet-TLVs for
> hop-by-hop
> authentication, if needed (but, see below).
>
> o *It prevents the use of more clever/advanced signature schemes/ICVs*
> Aggregate signature algorithms (
> https://crypto.stanford.edu/~dabo/papers/aggsurvey.pdf)
> exist, and an interesting use-case can be found in also reactive
> protocols, allowing verifying
> not “just” the end points, but also the intermediaries (again, with the
> appropriate “zero out”
> discussed above, or something smarter).
> Regeneration of messages, rather than forwarding, renders that impossible
> (or, if used,
> updating to
> https://tools.ietf.org/html/draft-perkins-manet-aodv-e2esec-00)
>
> JWD: this is IMO the big issue.


> There are other reasons, but the above are those that jump at me as
> immediate show-stoppers.
>
> I do honestly not see what possible benefit there is from “regeneration” —
> but I see very clear inconveniences, and security is not the least of
> these. Insisting on “regeneration” requires development of “non-general
> workarounds” as pointed out above.
>
> *Security Considerations*
> This is an always thorny subject. When OLSRv2 was going through the
> process we got a thorough education in how little we knew about security
> from the SEC-ADs, and had to spend about a year or so developing RFC7183.
> The bottom line is, that this protocol needs its “RFC7183  equivalent”,
> either as part of the main document, or as an independent document.
> currently, that is not the case.
>
> A minima, looking at BCP72 and BCP107 — taking inspiration from RFC7183
> might be aw good idea, as that was the most recent that went through the
> SEC AD. Regardless of how, however, a “mandatory to implement” security
> mechanism most be specified (I think the right term was “MUST implement,
> SHOULD use”), in sufficient detail to ensure interoperable implementations.
>
> As an example, both [RFC6130] and [RFC7181] set out that that:
>
>       On receiving a ... message, a router MUST first check if the
>       message is invalid for processing by this router
>
> and then proceed to give a number of conditions that, each, will lead to a
> rejection of the message as "badly formed and therefore invalid for
> processing” — a list which RFC7183 then amended. That gave a “hook” for
> RFC7183 for inserting “rejection”. Idem for message generation.
>
> If I was to do RFC7181/RFC6130 today, I would include that directly into
> the protocol specifications. It turned out to be more overhead (and slow
> down publication anyways) to do it as separate documents.
>
> Secondly, we need to be a lot more rigid in terms of what ICVs,
> Timestamps, etc. are added/removed, and what that brings.
>
> For example (with the assumption that messages are *forwarded* and *not*
> regenerated), this could be one option:
>
> o When a RREQ, RREP message is generated, add an ICV Message TLV, which
> is calculated <this way>
>  …(take inspiration from RFC7183 here)
>
> o When a RREQ, RREP message is transmitted  (by the source, or forwarded
> by an intermediary)
> in an RFC5444 packet, an ICV Packet TLV is added to the packet, which is
> calculated <this way> …
>
> o When an RFC5444 packet is received, validate the ICV Packet TLV <this
> way> and if the ICV doesn’t validate,
> discard the contained messages.
>
> o When receiving a RREQ/RREP (either as destination, or as an
> intermediary), validate the ICV Message TLV
> <this way> and if the ICV doesn’t validate, discard the message, do not
> process, do not forward.
>
> This is a different (and more flexible?) model than “just” transitive
> trust and “just” end-to-end trust discussed previously, and is enabled by
> the “forwarding” bit described previously.
>
> *Metrics*
> I have asked this numerous times, but have not gotten any answer so I am
> escalating this again. Why is this document having a normative reference to
> RFC6551?
>
>    [RFC6551]  Vasseur, JP., Ed., Kim, M., Ed., Pister, K., Dejean, N.,
>               and D. Barthel, "Routing Metrics Used for Path Calculation
>               in Low-Power and Lossy Networks", RFC 6551 <https://tools.ietf.org/html/rfc6551>, DOI 10.17487/
>               RFC6551 <https://tools.ietf.org/html/rfc6551>, March 2012,
>               <http://www.rfc-editor.org/info/rfc6551>.
>
>
>
> Unless I am mistaken, this is not the ROLL WG. And the abstract to RFC6551
> explicitly reads:
>
>    Low-Power and Lossy Networks (LLNs) have unique characteristics
>    compared with traditional wired and ad hoc networks that require the
>    specification of new routing metrics and constraints.
>
> Which explicitly is disclaiming that these metrics are* not applicable in
> ad hoc networks* — and, I am wondering, what has changed since the
> publication of RFC6551 to suddenly make these applicable?
>
>  I am also wondering if looking at (and, perhaps, using?) the same metric
> type TLV type as in OLSRv2 is not a possibility?
>
> Either way, referencing RFC6551 here is wrong.
>
> *RFC6621 (SMF) usage*
> As late as earlier this month, RFC6621 reared its head again on the
> mailing list. I understand that the authors are set on using flooding
> reduction as part of RREQ distribution. If that is the case, then this
> protocol MUST specify that flooding reduction.
>
> Saying “Use RFC6621” won’t work, for reasons Chris point out in an email
> earlier to the list:
>
> The network does not and fundamentally cannot offer multicast suppression
> to AODVv2. This isn't a hard concept, but seems to keep having to be said.
> AODVv2 messages are carried in 5444 packets, and those only travel one hop.
> That doesn't allow for suppression. Furthermore AODVv2 is creating new
> messages each hop. That needs AODVv2 specific handling.
>
>
> And quoting an email from myself:
>
> If protocol X relies on using a flooding operation, then it certainly is
> protocol X's responsibility to  specify the flooding operation (or, at
> least, the requirements to the flooding operation ) for correct functioning
> of protocol X.
>
>
> I’ve previously pointed out that RFC6621 requires some degree of
> configuration (can’t just, for example, pick two different relay set
> selection algorithms, as that may produce a disconnected flooding domain).
>
> In other words, what the authors propose in
> https://tools.ietf.org/html/draft-perkins-manet-using-rfc6621-00 is
> insufficient.
>
> JWD: I agree references to RFC6621 should be removed and use of reduced
relay sets should be work for an extension draft.  All that's is needed to
to allow AODVv2 routers to not regenerate RREQ which is already the case.


> *Multiple interfaces configured with the same address(es) - & RFC6621*
> I think that it has been established that the ability to use the same
> address on multiple interfaces is a requirement. It’s not clear to me that
> a single way of doing so has been proposed, nor that the “operating
> conditions” are called out (for example, will it work  if router A and
> router B both have two interfaces, and that all 4 interfaces use the same
> IP address?)
>
> This is not entirely unrelated to flooding reduction and RFC6621. One of
> the complexities is to get that just right: OLSRv2 - for example -
> calculates Flooding-MPRs per interface, and RFC6130 is careful in how it
> detects bidirectionality of links.
>
> I have not worked out all the details of how this impacts this protocol —
> but, it requires attention. Especially when wanting to be able to say “…use
> RFC6621”, then the interface configuration constraints when faced with
> multiple interfaces must be worked out.
>
> JWD: The super simple dumb way of doing this is to mandate that any
interface which shares an address with another much be bridged.  Then you
really only have a single interface again. Or just say it's not allowed.
Other more complex methods may be possible but we are honestly past time
for that and as Thomas points out there are corner cases that we will
likely miss and its past time for that.


> *Prefixes & Gateways*
> A question to my mind, and I am not sure I know how to answer that from
> reading the draft since it is not discussed in section 9 (at least), is:
> “What happens if I send a RREQ to a prefix” (i.e., a /<128 in IPv6)?
>
> JWD: I believe Lotte answered this in that it's not allowed.  Simple fix.

> o If the whole prefix is part of the attached network set of a single
> router,
> then I expect to get a single RREP back, is that the case?
>
> o If "half the prefix"  is part of the attached network set of a single
> router,
> and the other half is part of the attached network set of another  single
> router,
> then I expect two RREPs, is that the case?
>
> o More generally, will I get a tree build through the network here?
>
> Section 9 (or, the whole spec) is also not explicit enough about multiple
>  "default gateways" — i.e., the external network being “the rest of the
> Internet”, and there being multiple gateways to the Internet. I understand
> how the RREQ gets to the gateway, and how the gateway determines if it
> should respond (the destination is not part of “the MANET prefix, with
> which it is configured”).
>
> But, if A is close to GW1, and B is close to GW2, but A and B being far
> far apart (according to the metric used) within the MANET, then it would be
> preferential for the route from A to B to go via GW1-“the-Internet”-GW2?
>
> Is this a supported configuration?
> o If yes, how is it (seq# wise, I would expect some complications)
> handled?
> o If not, where is it stated that this configuration is not possible with
> this protocol?
>
>
> The applicability statement says:
>
>    AODVv2 is designed for stub or disconnected mobile ad hoc networks,
>    i.e., non-transit networks or those not connected to the internet.
>    AODVv2 can, however, be configured to perform gateway functions when
>    attached to external networks, as discussed in Section 9 <https://tools.ietf.org/html/draft-ietf-manet-aodvv2-13#section-9>.
>
>
> I believe that the scenario I describe is indeed a stub network (and not a
> transit network)?
>
> JWD: I'd change it to be configured to perform limited gateway functions
when attached to external networks.

> — — —
>
> As I indicated, these are the big-picture issues that I have gotten around
> to thinking about enough to write up. I’m not through working through the
> document, but those were the ones that came to mind, and which seem
> imperative to resolve.
>
> Best,
>
> Thomas
>
>
>
>
> _______________________________________________
> manet mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/manet
>
>