Session Traversal Utilities for NAT (STUN)
draft-ietf-tram-stunbis-21

Balloting NoObj, trusting AD.

[Thanks for addressing my DISCUSS -- the original text of my DISCUSS follows]

Thanks to everyone who worked on this revision of the STUN protocol. Thanks in
particular to the ARTART reviewer and to the authors for actively engaging on
the points he raised.

I have one concern about interoperability and another about the IANA changes
that I believe require changes to the document prior to publication.

§14.2:

>  X-Port is computed by taking the mapped port in host byte order,
>  XOR'ing it with the most significant 16 bits of the magic cookie, and
>  then the converting the result to network byte order.  If the IP
>  address family is IPv4, X-Address is computed by taking the mapped IP
>  address in host byte order, XOR'ing it with the magic cookie, and
>  converting the result to network byte order.  If the IP address
>  family is IPv6, X-Address is computed by taking the mapped IP address
>  in host byte order, XOR'ing it with the concatenation of the magic
>  cookie and the 96-bit transaction ID, and converting the result to
>  network byte order.

The discussion of performing operations "in host byte order" is very confusing,
and seems likely to cause issues communicating between machines of different
endianness. As an implementor, based on this description, I cannot tell
whether, given a port of 0x1234 (and operating on a little-endian machine),
I'm supposed to do:

Port (host order):   34 12
Magic Cookie Prefix: 21 12
Result (host order): 15 00
X-Port (net order):  00 15

or:

Port (host order):   34 12
Magic Cookie Prefix: 12 21
Result (host order): 26 33
X-Port (net order):  33 26

One of these is clearly wrong. I think it's the first one, but I *also* think
that the first one is the most straightforward interpretation of the quoted
paragraph.

The following would seem to be a complete description of the
operation without introducing possible confusion about the difference between
host and network order:

   X-Port is computed by XOR'ing the mapped port with the most significant 16
   bits of the magic cookie.  If the IP address family is IPv4, X-Address is
   computed XOR'ing the mapped IP with the magic cookie.  If the IP address
   family is IPv6, X-Address is computed by XOR'ing the mapped IP address with
   the concatenation of the magic cookie and the 96-bit transaction ID. In all
   cases, the XOR operation works on its inputs in network byte order (that is,
   the order they will be encoded in the message).

This makes it clear that the proper operation is:

Port:                12 34
Magic Cookie Prefix: 21 12
Result / X-Port:     33 26

---------------------------------------------------------------------------

§17.3.1:

>  IANA is requested to update the names for attributes 0x0002, 0x0003,
>  0x0004, 0x0005, 0x0007, and 0x000B, and the reference from RFC 5389
>  to RFC-to-be for the following STUN methods:
...
>  0x0003: (Reserved; prior to [RFC5389] this was CHANGE-REQUEST)

The attribute 0x0003 is registered by RFC 5780, and should not be removed by this document.
---------------------------------------------------------------------------

§3:

This is almost, but not quite, the boilerplate defined by RFC 8174. Please
update to match the text in RFC 8174.

---------------------------------------------------------------------------

§5:

>  server to detect if the client will understand certain attributes
>  that were added in this revised specification.

This is a *bit* misleading, as these attributes were actually added in RFC 5389.
Consider: "...that were added to STUN by [RFC5389]."

---------------------------------------------------------------------------

§6.1:

>  If the agent is sending a request, it SHOULD add a SOFTWARE attribute
>  to the request.

I believe this would benefit from a pointer to the security section; e.g.:
"Note that the inclusion of a SOFTWARE attribute may have security
implications; see Section 15.1.2 for details."

---------------------------------------------------------------------------

§6.2:

>  STUN may be used
>  with anycast addresses, but only with UDP and in usages where
>  authentication is not used.

This bit of operational advice seems out of place in the middle of an
implementation discussion, and is quite likely to be missed by its intended
audience. Consider relocating it to an "Operational Considerations" section.

---------------------------------------------------------------------------

§6.2.1:

>  As with TCP, the usage of Karn's
>  algorithm is RECOMMENDED [KARN87].

This normative language means that [KARN87] needs to be a normative reference.

---------------------------------------------------------------------------

§6.2.3 says:

>  Alternatively, a
>  client MAY be configured with a set of IP addresses that are trusted;
>  if a certificate is received that identifies one of those IP
>  addresses, the client considers the identity of the server to be
>  verified.

Presumably, this means the server supplies a certificate with SubjectAltName
containing an iPAddress? Please add text to clarify whether that's the
intention.

If that *is* the intended meaning, then this behavior in §8.1 seems
unnecessarily restrictive:

>  A "stuns" URI
>  containing an IP address MUST be rejected, unless the domain name is
>  provided by the same mechanism that provided the STUN URI, and that
>  domain name can be passed to the verification code.

Presumably, this is done because certs with iPAddress-form SubjectAltNames are
pretty rare (although CAB Forum baseline requirements do explicitly allow
their issuance) -- but if the text in §6.2.3 is based on allowing the use of
such certs in a TURN deployment, then it seems that URI resolution should be
also.

---------------------------------------------------------------------------

§9.1.3:

>     If the MESSAGE-
>     INTEGRITY-SHA256 attribute is not present, and using the same
>     password, compute the value for the message integrity as described
>     in Section 14.5.

I can't figure out what is meant by "and using the same password" here. The
structure of the sentence would imply that the subject of "using" is "attribute"
(the one that is not present), but that doesn't make sense. Is it supposed to
be something like this?

      If the MESSAGE-INTEGRITY-SHA256 attribute is not present, compute the
      value for the message integrity as described in Section 14.5, using the
      same password as would have been used for MESSAGE-INTEGRITY-SHA256.

---------------------------------------------------------------------------

§9.1.5:

>  A client sending subsequent requests to the same server MUST send
>  only the MESSAGE-INTEGRITY-SHA256 or the MESSAGE-INTEGRITY attribute
>  that matches the attribute that was received in the response to the
>  initial request.

How long does it continue to do so? It seems sensible to do this for the
length of validity of the credentials (and no longer) -- and probably even
scoped to the credentials --  but the document really should call that out
explicitly.

---------------------------------------------------------------------------

§9.2.2:

>  The structure of the key when used with long-term credentials
>  facilitates deployment in systems that also utilize SIP [RFC3261].
>  Typically, SIP systems utilizing SIP's digest authentication
>  mechanism do not actually store the password in the database.
>  Rather, they store a value called H(A1), which is equal to the key
>  defined above.

Suggest adding something like: "For example, this mechanism can be used with the
authentication extensions defined in [RFC5090]."

---------------------------------------------------------------------------

§9.2.3.2:

>  Once a request/response transaction has completed successfully, the

The use of "successfully" here is kind of confusing, since the transaction in
question was completed with an Error Response. Suggest dropping "successfully".

---------------------------------------------------------------------------

§9.2.4:

>  o  If the NONCE is no longer valid, the server MUST generate an error
>     response with an error code of 438 (Stale Nonce).  This response
>     MUST include NONCE, REALM, and PASSWORD-ALGORITHMS attributes and
>     SHOULD NOT include the USERNAME, USERHASH attribute, The NONCE
>     attribute value MUST be valid.

Nit: The comma before the final "The" should be a period.

---------------------------------------------------------------------------

§9.2.4:

>     USERHASH attribute.  The response MAY include a MESSAGE-INTEGRITY
>     or MESSAGE-INTEGRITY-SHA256 attribute, using the previous password
>     to calculate it.

I think this means to say "using the previous key," right?

---------------------------------------------------------------------------

§14:

>  The
>  padding bits are ignored, and may be any value.

This seems to be encouraging implementors to use uninitialized memory for these
bits, which could be a security issue. It would be far safer to do the more
traditional "must be set to zero on sending, must be ignored by receiver" thing
here.

---------------------------------------------------------------------------

§14.5:

>  Petit-Huguenin, et al.  Expires September 6, 2018 [Page 40]
>
>  Internet-Draft Session Traversal Utilities for NAT (STUN) March 2018
>  Similarly, when validating the MESSAGE-INTEGRITY, the length field in

This appears to be an errant header/footer pair. Note that the
correctly-formatted footer for page 40 appears several paragraphs later.

---------------------------------------------------------------------------

§14.7:

>  The 32-bit CRC is the one defined in ITU V.42 [ITU.V42.2002], which
>  has a generator polynomial of
>  x32+x26+x23+x22+x16+x12+x11+x10+x8+x7+x5+x4+x2+x+1.

Please correct the polynomial representation:

x^32 + x^26 + x^23 + x^22 + x^16 + x^12 + x^11 + x^10 + x^8 + x^7 + x^5 + x^4
+ x^2 + x + 1.

---------------------------------------------------------------------------

§14.8:

>   The reason phrase is meant for user consumption

This is a very tricky assertion to maintain, unless you want to start working
with language tags and reason phrase localization. I think you're far safer
saying something like "The reason phrase is meant for diagnostic purposes."
I would not encourage implementations to just display its contents to users
without any localization discussion.

---------------------------------------------------------------------------

§15.2.4:

It's probably worth additionally noting that this attack can be trivially
launched by the STUN server itself -- so users of STUN servers should have the
same level of trust in them as any other node that can insert themselves into
the communication flow.

---------------------------------------------------------------------------

§15.3:

I'm surprised that there is no plan in here for cryptoagility of the USERHASH
TLV. While a practical preimage attack on SHA-256 seems pretty implausible
today, it would be nice to know that we have a remedy if we ever decide that we
need to change things.

Thank you for addressing my DISCUSS points and comments.

Hi, thanks for this work. For the record, I have mainly reviewed the diff. 

General: I think Peter's ART-ART review has some points worth addressing.

§1, last paragraph: This seems out of place. Also, I wonder if it should be stated more strongly, perhaps with a SHOULD or even MUST.

§3: RFC 8174 offers it's own recommended boilerplate; is there a reason not to use it?

§6.2.1, third paragraph: " The value SHOULD be considered stale and discarded after
   10 minutes without any transactions to the same server. "
This is a bit ambiguous. I think it means the value is stale if no transactions have occurred in the last 10 minutes, But I think it could be read to say that there should be no transactions to the same server once one considers the value as stale.

§9.1.4: Does it make sense to signal that "an attack took place" rather than signaling "integrity protection was violated" and let the human operators decide if that implies an attack?

§14.3: "A compliant implementation MUST
   be able to parse UTF-8 encoded sequence of less than 763."

763 whats?

§15.3, 2nd bullet: I'm not sure I understand what is meant by "get an updated external mechanism"

I'm balloting No Objection and not DISCUSS because no individual
item rises quite to that level of criticality, but there still are
quite a number of things that ought to be addressed prior to
publication.

In addition to the excellent points raised in the secdir review
(which are mostly addressed but only in the editor's copy?), I
would call out a few more key observations:

If I understand correctly, the server's PASSWORD-ALGORITHMS array
from the response are echoed back by the client in the subsequent
request in order to provide downgrade protection -- the initial
response is at least sometimes not authenticated, and by having the
client repeat it back under an authenticated scheme, the server can
detect if the list was tampered to remove any new, stronger,
algorithms.  This is probably an important enough concept to be
called out explicitly, either where the behavior is described or in
Section 9.2.1 ("Bid Down Attack Prevention").

Section 8.1 has some text that could be read as requiring a STUN
client to implement dual-stack IPv4/IPv6, which presumably is not
the intention:
   In addition, instead of querying either the A or the AAAA resource
   records for a domain name, the client MUST query both and try the
   requests with all the IP addresses received, as specified in
   [RFC8305].

There are still some places where the text talks about
MESSAGE-INTEGRITY when it really means "message integrity
protection" (i.e., either the SHA1 or SHA-256 variants, for now);
I've attempted to note them in the notes below.

There is text in both Sections 9.1 and 9.2 that talks of replay
protection being provided, by the time-limited nature of the
short-term credentials or the nonce for long-term credentials.  This
seems to only, strictly speaking, be true if any given
password/nonce can only be used once, but for the short-term
credentials the password is used for the duration of the
"session-equivalent", and for the long-term credentials a nonce can
be reused for some amount of time as well.  So while these are both
replay countermeasures that can limit replay attacks, it seems a
stretch to claim that they prevent replay attacks (entirely).

The Stale Nonce behavior seems potentially worrisome, in that it
opens up a side channel for a distinguishing attack,
between a 401 and 438 response.  (That is, "password correct" vs.
"password incorrect".)  The impact seems rather muted, though, since
the gain to the attacker is to be able to precompute a bunch of
requests using a nonce of the attacker's choosing and blindly replay
the precomputed object against (possibly multiple) servers looking
for a guess.  The realm and username are still in play, so the scope
for the attacker to gain from the precomputation seems limited.
(The same level of brute-force guessing can be obtained "live" just
by computing the trial responses against a live server, using a
valid nonce.)  So, while it might be nice to give guidance that 438
should only be used when the server can validate that it did
generate the nonce and the nonce was valid "recently", and to treat
other cases as authentication failures, it's not clear to me that
there's enough of a benefit from the change to make it worth doing.


Section 6.2.3

Maybe just cite BCP 195/RFC 7525 instead of inlining requirements?

Also, the "client SHOULD verify ..." and "client MUST verify ..." are
similar enough that some clarification of certificate vs. identity
verification would be in order, if both statements are believed
accurate.


Section 6.3

   [...] Known-but-unexpected attributes SHOULD be ignored by
   the agent. [...]

Why is this not an error?


Section 6.3.1, 6.3.2, etc.

Should the second paragraph start with "Otherwise, "?  Presumably
once an error response is sent or processing ceases, processing stops...


Section 8.1

   [...] A "stuns" URI
   containing an IP address MUST be rejected, unless the domain name is
   provided by the same mechanism that provided the STUN URI, and that
   domain name can be passed to the verification code.

"the verifcation code" is probably unnecessarily vague; we're just
talking about the (D)TLS SNI and certifciate verification code,
right?



Section 9.2

   [...[ The nonce provides the replay protection.  It is a cookie,
   selected by the server, and encoded in such a way as to indicate a
   duration of validity or client identity from which it is valid.

Is this like a TLS cookie, where only the server needs to know about
the internal structure (viz. "encoded" in the above)?  If so, the
text could probably be more clear.

When mentioning just "a message-integrity", maybe forward-ref both
versions of it to be more clear that we don't mean just the old one?

It's unclear why indications can't use a nonce, if one is known
from a previous request/response on the same connection.

When encoding the value of 0 for the security feature, it may be
worth clarifying that it is still encoded as a 24-bit integer and
base64-encoded, though the "13 character string" text ought to make
this clear to the reader.


Section 9.2.4

      *  If the request contains neither PASSWORD-ALGORITHMS nor
         PASSWORD- ALGORITHM, then the request is processed as though
         PASSWORD- ALGORITHM were MD5 (Note that if the original
         PASSWORD-ALGORITHMS attribute did not contain MD5, this will
         result in a 400 Bad Request in a later step below).

This Note seems a bit odd, since we're predicated on
PASSWORD-ALGORITHMS being absent.

      *  Otherwise, unless (1) PASSWORD-ALGORITHM and PASSWORD-
         ALGORITHMS are both present, (2) PASSWORD-ALGORITHMS matches
         the value sent in the response that sent this NONCE, and (3)
         PASSWORD-ALGORITHM matches one of the entries in PASSWORD-
         ALGORITHMS, the server MUST generate an error response with an
         error code of 400 (Bad Request).

Maybe we could be more clear about "the value of the
PASSWORD-ALGORITHMS attribute in the message matches the value of
the PASSWORD-ALGORITHMS attribute sent in the response that sent the
NONCE used with this message"?


Section 9.2.5

   If the response contains a PASSWORD-ALGORITHMS attribute, the
   subsequent request MUST be authenticated using MESSAGE-INTEGRITY-
   SHA256 only.

This means "all subsequent requests to this server"?  Maybe the
wording could be more clear.  (It also means that if we get a new
scheme when SHA256 is broken, the document introducing it may need
to Update this document to remove the MUST.)

Also, I think the Security Feature bit is "Password algorithms"
plural (in the next paragraph and maybe multiple locations in the
document).


Section 10

   [...] If the transaction uses TLS or DTLS and if the
   transaction is authenticated by a MESSAGE-INTEGRITY-SHA256 attribute
   and if the server wants to redirect to a server that uses a different
   certificate, then it MUST include an ALTERNATE-DOMAIN attribute
   containing the subjectAltName of that certificate.

The (definite article), implying the entire subjectAltName?  Or just
one name that is contained as a SAN entry?


Section 14.8

Need to not just say MESSAGE-INTEGRITY since the SHA256 variant
should be fine as well.


Section 14.10

Maybe the nonce cookie should get a shout-out here?


Section 14.12

Do we need to talk about the padding, since this is the the sole
contents of the attribute and the attribute padding suffices?

Also, "derive the long-term password" doesn't seem quite right;
we're deriving the key from the long-term password, per Section
17.15.1.


Section 15.1.2

   There is no amplification of the number of packets with this attack
   (the STUN server sends one packet for each packet sent by the
   client), though there is a small increase in the amount of data,
   since STUN responses are typically larger than requests.

I usually hear amplification in terms of bytes or bandwidth, not
packets.  So maybe just "amplification is modest"?


Section 15.3

   This specification uses both HMAC-SHA1 and HMAC-SHA256 for
   computation of the message integrity.

Could perhaps be reworded, since the number of situations when both
are used on the same message is small, and we try to push SHA256
when we can.

   o  STUN Client and Server using the Short Term Credential Mechanism

Clients and Servers, plural.


Section 17.32.

I'm confused why we have MUST-level requirements about using
PASSWORD-ALGORITHMS when it is in the comprehension-optional range
(e.g., echoing from response to repeated request as in Section
9.2.5).  The server cannot actually enforce this MUST when talking
to an old client if we want to retain the comprehension-optional
semantics, IIUC.


Section 17.6

If we're adding support for DTLS-over-udp (per the next section),
should we be registering stuns/udp?

"password encryption algorithm" is not quite the right statement to
make, since we're hashing, not encrypting.

Thank you for addressing my DISCUSS