[6lo] Benjamin Kaduk's Discuss on draft-ietf-6lo-ap-nd-15: (with DISCUSS and COMMENT)

[Benjamin Kaduk via Datatracker <[e-mail]>]

Benjamin Kaduk has entered the following ballot position for
draft-ietf-6lo-ap-nd-15: Discuss

When responding, please keep the subject line intact and reply to all
email addresses included in the To and CC lines. (Feel free to cut this
introductory paragraph, however.)


Please refer to https://www.ietf.org/iesg/statement/discuss-criteria.html
for more information about IESG DISCUSS and COMMENT positions.


The document, along with other ballot positions, can be found here:
https://datatracker.ietf.org/doc/draft-ietf-6lo-ap-nd/



----------------------------------------------------------------------
DISCUSS:
----------------------------------------------------------------------

The CIPO description specifies that a 5-bit Reserved1 field and a 13-bit
Public Key Length field are combined to fit into a 16 (not 18)-bit space.
(The Figure shows the Public Key Length field as 11 bits.)

Why do we need to allow ambiguity/flexibility as to the point representation
within a given Crypto-Type value (as opposed to fixing the representation as
a parameter of the Crypto-Type)?  Alternately, what does "(un)compressed"
mean, as it's not really clarified directly?  Furthermore, Table 2 lists an
"(un)compressed" representation for type 0 (over P-256), but RFC 7518 notes
that the compressed representation is not supported for P-256 (Section
6.2.1.1).  I also am not finding the needed codepoints registered in the JOSE
registries to use ECDSA25519 (crypto-type 2) -- do we need to register
anything there?

Per my comment on Section 4.4, there may be some implicit
expectation/requirement of 128+-bit ROVRs for AP-ND, but I didn't find where
such a requirement was stated.


----------------------------------------------------------------------
COMMENT:
----------------------------------------------------------------------

Thanks for this well-written document!  It will be good to see this stuff
getting deployed.

What's the risk of DoS via (short) address exhaustion in terms of 6LBR
resources (as distinct from 6LR resources mentioned at the end of Section
7.1)?

Section 1

   In this specification, a 6LN generates a cryptographic ID (Crypto-ID)
   and places it in the ROVR field during the registration of one (or
   more) of its addresses with the 6LR(s).  Proof of ownership of the
   Crypto-ID is passed with the first registration exchange to a new
   6LR, and enforced at the 6LR.  The 6LR validates ownership of the
   cryptographic ID before it creates any new registration state, or
   changes existing information.

I'll read on and see, but how much trust does this imply that we require of
all 6LRs (as opposed to the 6LBR, which is kind of required to be pretty
trusted)?  [ed.: basically complete trust is required]

   The protected address registration protocol proposed in this document
   enables Source Address Validation (SAVI) [RFC7039].  This ensures
   that only the actual owner uses a registered address in the IPv6
   source address field.  A 6LN can only use a 6LR for forwarding
   packets only if it has previously registered the address used in the
   source field of the IPv6 packet.

nit: this last sentence is written as if we not only enable, but also
require, SAVI.

   The 6lo adaptation layer in [RFC4944] and [RFC6282] requires a device
   to form its IPv6 addresses based on its Layer-2 address to enable a
   better compression.  This is incompatible with Secure Neighbor
   Discovery (SeND) [RFC3971] and Cryptographically Generated Addresses
   (CGAs) [RFC3972], since they derive the Interface ID (IID) in IPv6
   addresses with cryptographic keys.

Are we going to have some further discussion of the scope of that loss of
functionality and/or alternative mechanisms to achieve those goals, or is
this statement just intended to note the incompatibility?

Section 3

   The proof requires the support of Elliptic Curve Cryptography (ECC)
   and that of a hash function as detailed in Section 6.2.  To enable
   the verification of the proof, the registering node needs to supply
   certain parameters including a Nonce and a signature that will
   demonstrate that the node has the private-key corresponding to the
   public-key used to build the Crypto-ID.

nit: the proof itself may only indirectly involve a hash, since we use
non-prehash Ed25519 (but the hash function is still needed to generate the
Crypto-ID itself).

nit(?): I think we tend to see "possesses the private key" more often than "has
the private-key", for parallelism with the "proof of possession" term of
art.

   The elliptic curves and the hash functions that can be used with this
   specification are listed in Table 2 in Section 8.3.  The signature
   scheme that specifies which combination is used is signaled by a

nit: this could be misread as saying that the Table is authoritative, not
the registry.

Section 4.1

   The Crypto-ID is derived from the public key and a modifier as
   follows:

   1.  The hash function indicated by the Crypto-Type is applied to the
       CIPO.  Note that all the reserved and padding bits MUST be set to
       zero.
   2.  The leftmost bits of the resulting hash, up to the size of the
       ROVR field, are used as the Crypto-ID.

This construction seems to allow for some malleability, in that an attacker
who observes a given Crypto-ID and CIPO could produce a related, valid,
Crypto-ID by reducing the ROVR length and truncating what was received.  I
haven't attempted to explore what (if any) potential consequences there are
of such malleability, and would like to know if anyone else has already done
so.  It looks like the NDPSO construction includes the length of the
Crypto-ID, so it would be hard for an attacker to do much with such a
truncated Crypto-ID, but attacks only get better...

Section 4.2

nit: I confess I'm missing what compels us to produce a de novo definition
of the "Status" field (and Length field) as opposed to deferring to the RFC
8505 definition, as we do for most of the other fields.

Section 4.3

   Crypto-Type:  8-bit unsigned integer.  The type of cryptographic
      algorithm used in calculation Crypto-ID (see Table 2 in
      Section 8.3).  Although the different signature schemes target
      similar cryptographic strength, they rely on different curves,
      hash functions, signature algorithms, and/or representation
      conventions.

While the currently-defined signature schemes target similar cryptographic
strength, do we need to imply that all future ones will also target a
similar strength (especially since Section 8.3 admits the possibility of
"better security")?

   Modifier:  8-bit unsigned integer.  Set to an arbitrary value by the
      creator of the Crypto-ID.  The role of the modifier is to enable
      the formation of multiple Crypto-IDs from a same key pair, which
      reduces the traceability and thus improves the privacy of a
      constrained node that could not maintain many key-pairs.

It may be worth digging in a little further to the threat model being
addressed here -- while I laud the ability to have multiple Crypto-IDs from
a single keypair, it seems that the tracking-prevention this affords implies
an attacker who can observe Crypto-IDs, most likely at multiple points in
the network.  Are Crypto-IDs used for anything other than EAROs?  I don't
have a great sense for how likely it is that an attacker in position to
observe (a subset of) EAROs would not also be in a position to observe at
least one associated CIPO, which includes all the information needed to
generate the Crypto-ID other than the Modifier.  With only an 8-bit Modifier
space, brute-forcing the other possible Crypto-IDs for that keypair is not
terribly strenuous, whereas allowing for a (potentially variable length?)
larger Modifier, say 64 bits, would substantially increase the work-factor
needed to predict (and store!) other Crypto-IDs that might be used for this
keypair.  Given that the CIPO already carries the public key itself, it's
not entirely clear how space-constrained we are so as to limit this to just
8 bits of modifier.

   Padding:  A variable-length field completing the Public Key field to
      align to the next 8-bytes boundary.

(MUST be set to zero by sender and ignored by receiver, as well?)

   The implementation of multiple hash functions in a constrained
   devices may consume excessive amounts of program memory.

But multiple signature algorithm implementations will not?

Section 4.4

Is there some more discussion to have about how NDPSO includes only a
specific fixed set of information under the signature but RSAO signs all
options that appear prior to it in the message?

nit: style suggests parallelism between the Section 4.3 and 4.4 introductory
remarks (e.g., "this specification defines the NDP Signature Option
(NDPSO)").

   As opposed to the RSA Signature Option (RSAO) defined in section 5.2.
   of SEND [RFC3971], the NDPSO does not have a key hash field.  The
   hash that can be used as index is the 128 leftmost bits of the ROVR
   field in the EARO.

nit: this last sentence perhaps leaves more context implicit than the reader
might like; I suggest "Instead, the leftmost 128 bits of the ROVR field in
the sibling EARO are used to lookup the key used for signature verification,
[left-padded with zeros if needed]".  (I added the part in square brackets
because I'm not sure that anything guarantees the ROVR will actually contain
128 bits -- when AP-ND is not in use its typically just 64 bits, right?)

   Pad Length:  8-bit unsigned integer.  The length of the Padding
      field.

(In octets?)

   Padding:  A variable-length field making the option length a multiple
      of 8, containing as many octets as specified in the Pad Length
      field.  Typically there is no need of a padding and the field is
      NULL.

Just to check: nothing requires the use of minimal-length padding to achieve
the required alignment, and padding of 200+ bytes is permitted by the spec,
albeit silly?

Section 6

   This specification enables the 6LR to verify the ownership of the
   binding at any time assuming that the "C" flag is set.  The
   verification prevents other nodes from stealing the address and
   trying to attract traffic for that address or use it as their source
   address.

nit: I think that the verification cannot be done unilaterally by the 6LR,
so maybe "enables the 6LR to challenge the node to verify its ownership of
the binding at any time" is better.

   A node may use multiple IPv6 addresses at the same time.  The node
   MAY use the same Crypto-ID, to prove the ownership of multiple IPv6
   addresses.  The separation of the address and the cryptographic

nit: no comma

   material avoids the constrained device to compute multiple keys for

nit: "avoids the need for"

   multiple addresses.  The registration process allows the node to use
   the same Crypto-ID for all of its addresses.

nit(?): We could potentially mention again the ability to use different
Crypto-IDs with a single keypair, though there's not any clear need to do
so.

Section 6.1

Mostly just for my own elucidation, where is it first specified to put the
address being registered in the Target Address field of the NS message?

   detrimental to the battery lifetime.  Alternatively, the device may
   use an always-incrementing value saved in the same stable storage as
   the key, so they are lost together, and starting at a best effort
   random value, either as Nonce value or as a component to its
   computation.

nit: "as the Nonce value".

   and the NDPSO containing the signature.  The information associated
   to a Crypto-ID stored by the 6LR on the first NS exchange where it
   appears.  The 6LR MUST store the CIPO parameters associated with the

nit: I think there's a verb missing in this sentence ("MUST be"?)

        |------- NS with EARO, CIPO, NonceLN and NDPSO -------->|

(This still includes the Crypto-ID, even if there's not room in the figure
to explicition indicate that, right?)

   *  Upon the first exchange with a 6LR, a 6LN will be challenged to
      prove ownership of the Crypto-ID and the Target Address being

How is the "first exchange" tracked?  Just whether the 6LR knows about the
registered IPv6 address, or the L2 address, or something else?  In
particular, if a 6LR has cached an IP address/ROVR binding, and an attacker
tries to join that 6LR with that address/ROVR, what controls whether the 6LR
will challenge the attacker's 6LN?  Is it purely at the 6LR's discretion?

   *  The challenge is triggered when the registration for a Source
      Link-Layer Address is not verifiable either at the 6LR or the
      6LBR.  In the latter case, the 6LBR returns a status of
      "Validation Requested" in the DAR/DAC exchange, which is echoed by
      the 6LR in the NA (EARO) back to the registering node.  The
      challenge MUST NOT alter a valid registration in the 6LR or the
      6LBR.

nit(?): the previous bullet describes a case where the 6LR MUST issue a
challenge; my understanding is that this bullet describes an additional way
in which challenges can occur (i.e., when the 6LBR doesn't know about the
address registration either).  Stylistically, it might be more clear to
reword this as being an "additional case" on top of the previous one, as
opposed to the current wording which tries to be very general about when a
challenge can happen, which is neither a parallel construction to the
previous bullet nor particularly clear about what new information is being
conveyed by this bullet.

   *  In order to validate the ownership, the 6LR performs the same
      steps as the 6LN and rebuilds the Crypto-ID based on the
      parameters in the CIPO.  If the rebuilt Crypto-ID matches the
      ROVR, the 6LN also verifies the signature contained in the NDPSO
      option.  If at that point the signature in the NDPSO option can be
      verified, then the validation succeeds.  Otherwise the validation
      fails.

I worry a little bit that we haven't fully specified the entire list of
steps the 6LR takes to follow the cryptographic chain and validate that the
6LN holds the key in question and registers the address in question to the
Crypto-ID in question.  I don't have a specific thing in mind that the 6LR
might forget, but if we explicitly state what things are bound to what other
things, the 6LR is less likely to make a security-relevant mistake.

Section 6.2

Might we want to include the Modifier or the Crypto-ID itself in the signed
data, in addition to the public key, to avoid any potential misbinding?

   2.  JWK-encoded public key

[just to check: JWK-encoding means that we have an injective map and aren't
subject to some sort of extension attack?  Not that such a thing would be
easy to pull off, with a strict registry of Crypto-Types, of course]

   6.  The length of the ROVR field in the NS message containing the
       Crypto-ID that was sent.

nit: is this really the NS "that was sent" as opposed to the NS under
construction that contains the NDPSO?  (This is related to my question up in
Section 6.1.)  Tying the initial NS to the followup one requires more state
on the 6LR (though the 6LR probably already had to keep state for the Nonce)
and makes some of the verification steps a bit tougher to implement
correctly.

   7.  1-byte (in network byte order) Crypto-Type value sent in the CIPO
       option.

nit: Is there any byte order with distinct representation from network byte
order, for a 1-byte quantity?  ;)

   *  Depending on the Crypto-Type, apply the hash function on this
      concatenation.

nit: is this clearer as "Apply the hash function (if any) specified by the
Crypto-Type to the concatenated data"?

   *  Depending on the Crypto-Type, sign the hash output with ECDSA (if
      curve P-256 is used) or sign the hash with EdDSA (if curve Ed25519
      (PureEdDSA)).

nit: this text is not very consistent with the idea of an extensible
registry of Crypto-Types; similar "apply the signature algorithm specified
by the Crypto-Type" language to what I suggested above may be appropriate.

   The 6LR on receiving the NDPSO and CIPO options first regenerates the
   Crypto-ID based on the CIPO option to make sure that the leftmost
   bits up to the size of the ROVR match.  If and only if the check is

In the vein of my previous questions about where the Crypto-ID appears and
truncation/malleability, this text seems to leave some room for confusion
about how many bits of output are being compared, and compared to what.

   *  Depending on the Crypto-Type indicated by the (6LN) in the CIPO,
      apply the hash function on this concatenation.

[similar comment as above]

   *  Verify the signature with the public-key received and the locally
      computed values.  If the verification succeeds, the 6LR and 6LBR
      add the state information about the Crypto-ID, public-key and
      Target Address being registered to their database.

This might be a place to reiterate the FCFS nature of address registration,
though it's hardly necessary.

Section 6.3

   In a multihop 6LoWPAN, a 6LBR sends RAs with prefixes downstream and
   the 6LR receives and relays them to the nodes. 6LR and 6LBR
   communicate using ICMPv6 Duplicate Address Request (DAR) and
   Duplicate Address Confirmation (DAC) messages.  The DAR and DAC use
   the same message format as NS and NA, but have different ICMPv6 type
   values.

   In AP-ND we extend DAR/DAC messages to carry cryptographically
   generated ROVR.  In a multihop 6LoWPAN, the node exchanges the
   messages shown in Figure 6.  The 6LBR must identify who owns an
   address (EUI-64) to defend it, if there is an attacker on another
   6LR.

This description feels a little vague to me -- I'm interested in knowing how
the 6LBR gains confidence that a proof-of-possession was provided to the
6LR, and how the ROVR gets validated at 6LBR and 6LR when the 6LN in
question moves to a new 6LR.  What information remains local to (each) 6LR
and what is communicated between 6LR and 6LBR?
This also relates to the "how trusted are 6LRs?" question I mentioned
earlier.

Section 7

It might be appropriate to talk about how the NonceLR and NonceLN combine to
ensure contributory behavior from both 6LR and 6LN, and that each protects
against a different type of replay attack.

Section 7.1

   Duplicate Address Detection DoS Attack:  Inside the LLN, Duplicate
      Addresses are sorted out using the ROVR, which differentiates it
      from a movement.  DAD coming from the backbone are not forwarded
      over the LLN, which provides some protection against DoS attacks
      inside the resource-constrained part of the network.  Over the
      backbone, the EARO option is present in NS/NA messages.  This
      protects against misinterpreting a movement for a duplication, and
      enables the backbone routers to determine which one has the
      freshest registration and is thus the best candidate to validate
      the registration for the device attached to it.  But this
      specification does not guarantee that the backbone router claiming
      an address over the backbone is not an attacker.

Where do we specify the behavior of the 6LBR to reject EAROs when the 6LBR
is not the one with the freshest registration?  Or do I misunderstand the
discussion of "freshest registration"?

   Replay Attacks  Using Nonces (NonceLR and NonceLN) generated by both
      the 6LR and 6LN provides an efficient protection against replay
      attacks of challenge response flow.  The quality of the protection
      still depends on the quality of the Nonce, in particular of a
      random generator if they are computed that way.

It's probably worth noting explicitly that we do not require unpredictable
nonces, just non-repeating ones.

   Neighbor Discovery DoS Attack:  A rogue node that managed to access
      the L2 network may form many addresses and register them using AP-
      ND.  The perimeter of the attack is all the 6LRs in range of the
      attacker.  The 6LR MUST protect itself against overflows and
      reject excessive registration with a status 2 "Neighbor Cache
      Full".  This effectively blocks another (honest) 6LN from
      registering to the same 6LR, but the 6LN may register to other
      6LRs that are in its range but not in that of the rogue.

Are there L2 technologies where L2 media keys are tied to L2 address, so a
rogue node might be rate-limited or given a quota based on L2 address?

Section 7.2

   address.  This specification frees the device to form its addresses
   in any fashion, thereby enabling not only 6LoWPAN compression which
   derives IPv6 addresses from Layer-2 addresses but also privacy
   addresses.

We noted in Section 1 that the 6lo adaptation layer requires use of
L2-address-based IPv6 address selection, so it's not clear how it would be
valid to use privacy addresses with this specification.

Section 7.3

We seem to be assuming for these calculations/discussion that the hash used
for calculating the Crypto-ID is a well-behaved cryptographic hash and thus
that random collisions are the only ones possible.  It would be worth
mentioning that weaknesses in the hashes associated with a given Crypto-Type
could allow attackers to make more targetted collision attempts.

   The formula for calculating the probability of a collision is 1 -
   e^{-k^2/(2n)} where n is the maximum population size (2^64 here,
   1.84E19) and K is the actual population (number of nodes).  If the

nit: minuscule vs. majuscule 'k'/'K'

Section 7.4

   The signature schemes referenced in this specification comply with
   NIST [FIPS186-4] or Crypto Forum Research Group (CFRG) standards
   [RFC8032] and offer strong algorithmic security at roughly 128-bit
   security level.  These signature schemes use elliptic curves that

Do we want to make any forward-looking statement about the expected strength
of future-defined Crypto-Types?  ("No" is a fine answer...)

Section 7.5

The text here is more about cross-algorithm attacks than cross-protocol
ones.  It would be great to see some text about cross-protocol attacks
added, too (e.g., prohibiting use of the keypair for anything other than
AP-ND).

Section 7.6

   This specification distributes the challenge and its validation at
   the edge of the network, between the 6LN and its 6LR.  The central
   6LBR is offloaded, which avoids DOS attacks targeted at that central
   entity.  This also saves back and forth exchanges across a

nit: "the central 6LBR is offloaded" reads very oddly, as the offloading is
of work from the 6LBR, but the 6LBR itself remains in place.

   The downside is that the 6LBR needs to trust the 6LR for performing
   the checking adequately, and the communication between the 6LR and
   the 6LBR must be protected to avoid tempering with the result of the
   test.

I'd prefer to see some more details about the nature of the required
protection for 6LR/6LBR communications.

   If a 6LR is compromised, it may pretend that it owns any address and
   defeat the protection.  It may also admit any rogue and let it take
   ownership of any address in the network, provided that the 6LR knows
   the ROVR field used by the real owner of the address.

nit: this language could probably be tightened up to be more precise about
the hazards.

Section 8.3

   New Crypto-Type values providing similar or better security (with
   less code) may be defined in the future.

I'm not sure what purpose the "with less code" requirement serves.  It seems
to needlessly constrain future actions.

Section 11

Curve25519 support is a MAY, so I think RFC 7748 belongs as normative;
likewise for RFCs 8032 and 6234.

It's probably better to cite RFC 7696 as BCP 201 directly.

Section B.3

It would be great if the document shepherd could diff the paramters listed
here against [CURVE-REPRESENTATIONS].