bk2.txt

Hello Benjamin,

Thank you very much for your additional comments!

On 19 Mar 2020, at 22:21, Benjamin Kaduk <kaduk@MIT.EDU> wrote:
> 
> I'll also reply inline.
> 
> On Thu, Mar 19, 2020 at 05:01:49PM +0100, Ragnar Sundblad wrote:
> > 
> > Hello Benjamin,
> > 
> > Thank you very much for your thorough review and all your comments!
> > This has helped us improve on the text in many places.
> > 
> > Please find our answers below.
> > 
> > > On 11 Mar 2020, at 01:22, Benjamin Kaduk via Datatracker <noreply@ietf.org> wrote:
> > > 
> > > Benjamin Kaduk has entered the following ballot position for
> > > draft-ietf-ntp-using-nts-for-ntp-23: 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-ntp-using-nts-for-ntp/
> > > 
> > > 
> > > 
> > > ----------------------------------------------------------------------
> > > DISCUSS:
> > > ----------------------------------------------------------------------
> > > 
> > > Thank you for this document; it has been a long time coming and is much
> > > awaited.  That said, I have a few points I'd like to discuss before I
> > > can comfortably ballot Yes:
> > > 
> > > I'm happy to see prominent references to RFCs 7525 and 6125.
> > > Unfortunately, merely citing RFC 6125 does not provide a usable
> > > specification for an application to implement; we need to additionally
> > > state what type of name (e.g., SRV-ID or DNS-ID) is used as input to the
> > > validation process and how the application obtains that name.  Given
> > > that we are defining a service name (and port number) for ntske, SRV-ID
> > > might be appropriate (DNS-ID is the most common form I see consumers
> > > of RFC 6125 using).
> > 
> > The service discovery is not in scope for this draft, and we anticipate
> > several different methods in the future. There are already discussions
> > going on on the ntp working group. We therefore do not specify these things
> > here, but want to refer to the relevant specifications for the different
> > methods. The rules for constructing a list of reference identifiers for
> > certificate validation is laid out in Section 6.2.1. of RFC 6125. This is what
> > we expect NTS-KE clients to do.
> 
> I don't think it's appropriate to leave the selection of type of identifier
> fully up to the application; even if we don't discuss server discovery, we
> can still discuss the framework used for naming servers, as naming is a key
> part of authentication and authorization.  RFC 6125 itself acknowledges
> this, noting in Section 2.1 that:
> 
>    [...] Ideally, protocol specifications that
>    reference this document will specify which identifiers are mandatory-
>    to-implement by servers and clients, which identifiers ought to be
>    supported by certificate issuers, and which identifiers ought to be
>    requested by application service providers.  Because these
>    requirements differ across applications, it is impossible to
>    categorically stipulate universal rules (e.g., that all software
>    implementations, service providers, and certification authorities for
>    all application protocols need to use or support DNS-IDs as a
>    baseline for the purpose of interoperability).
> 
>    However, it is preferable that each application protocol will at
>    least define a baseline that applies to the community of software
>    developers, application service providers, and CAs actively using or
>    supporting that technology (one such community, the CA/Browser Forum,
>    has codified such a baseline for "Extended Validation Certificates"
>    in [EV-CERTS]).
> 
> The checklist in Section 3 also gives guidance to application protocol
> designers as to how to select the type of identifier to be used for their
> protocol.

We have added a paragraph about that implementations must obey the
rules in RFC 5280 and RFC 6125 and that we recommend the use of
DNS-ID, and hope this will suffice for now.

> > > In a related vein, if ALPN is necessary for confirming ntske operation,
> > > it's not entirely clear to me that a dedicated port (in addition to
> > > service name) is required, and RFC 6335 discourages fixed port numbera
> > > llocations.  Perhaps it's not intended that DNS-SD is used to discover
> > > NTS-KE servers, though -- there's no reference to RFC 6763 in this
> > > document, or other discussion of server discovery that I can see.
> > 
> > Again, we anticipate several different methods and there is ongoing work in the
> > ntp working group.
> > 
> > > We seem to be internally inconsistent about whether the Cookie extension
> > > can be encrypted -- Section 5.7 says "MUST NOT be encrypted", but
> > > Section 5.2 implies that it could be encrypted:
> > > 
> > >  Always included among the authenticated or authenticated-and-
> > >  encrypted extension fields are a cookie extension field and a unique
> > >  identifier extension field.  [...]
> > 
> > The cookie is not encrypted from client to server (since the server
> > does not yet know what C2S key to use, it is in the cookie), but
> > encrypted when sent from the server to the client, for unlinkability.
> > We added a reference to Section 5.7. to emphasize how this should be done.
> > 
> > > Section 7.6 says that applications for new record types need to specify
> > > the contents of the "Set Critical Bit" field, but this field is not
> > > included in the table of initial entries.  Additionally, there doesn't
> > > seem to be a clear description of what the semantics of the "Set
> > > Critical Bit" field should be.
> > 
> > Thanks - there used to be a Critical bit in the registry, but as you
> > saw is not any more. It is now removed from the text too. The entire paragraph
> > in question will be removed in draft 25.
> 
> Ah, excellent.
> 
> > > ----------------------------------------------------------------------
> > > COMMENT:
> > > ----------------------------------------------------------------------
> > > 
> > > Section 1.1
> > > 
> > >  o  Identity: Through the use of the X.509 public key infrastructure,
> > > 
> > > It's not exactly accurate to say that there's one privileged "X.509 PKI"
> > > (even if that would have been true about the original X.500 vision); we
> > > have a large set of stuff on the "Web PKI" but there are other
> > > X.509-based PKIs in use in, e.g., the cable industry.  I'd suggest
> > > either dropping "the" or (if appropriate, which is not entirely clear)
> > > adding "Web".
> > 
> > Indeed, fixed!
> > 
> > >  o  Replay prevention: Client implementations may detect when a
> > >     received time synchronization packet is a replay of a previous
> > >     packet.
> > > 
> > > Given that a replay of time synchronization information kind of defeats
> > > the point of a time synchronization protocol, do we want to use
> > > something stronger than "may" here?
> > 
> > The reasoning behind this was that the protocol allows for this,
> > but can not guarantee that an implementation does what it is
> > supposed to do.
> 
> I think that using "can" would avoid any risk of misunderstanding as
> "sometimes possible" vs. "at their option"

This time we got your point, and have now changed them to "can".
Thanks!

> > >  o  Unlinkability: For mobile clients, NTS will not leak any
> > >     information additional to NTP which would permit a passive
> > >     adversary to determine that two packets sent over different
> > >     networks came from the same client.
> > > 
> > > I guess it's kind of implicit that server mobility isn't really a thing
> > > and thus we don't need to be concerned about linkability of servers.
> > 
> > We think so too.
> > 
> > >  o  Non-amplification: Implementations (especially server
> > >     implementations) may avoid acting as distributed denial-of-service
> > >     (DDoS) amplifiers by never responding to a request with a packet
> > >     larger than the request packet.
> > > 
> > > [similar comment about "may"]
> > 
> > (See answer above.)
> > 
> > > Section 1.2
> > > 
> > >  The typical protocol flow is as follows: The client connects to an
> > >  NTS-KE server on the NTS TCP port and the two parties perform a TLS
> > > 
> > > (Section 7.1 requests a port for "ntske"; I don't know how tightly we
> > > need to constrain descriptive usage like this.)
> > 
> > It says "typical", so our feeling is that this is ok.
> > 
> > >  handshake.  Via the TLS channel, the parties negotiate some
> > >  additional protocol parameters and the server sends the client a
> > >  supply of cookies along with an IP address to the NTP server for
> > >  which the cookies are valid.  The parties use TLS key export
> > > 
> > > nit: "IP address of the NTP server"
> > 
> > Fixed!
> > 
> > >  phase of the protocol.  This negotiation takes only a single round
> > >  trip, after which the server closes the connection and discards all
> > >  associated state.  At this point the NTS-KE phase of the protocol is
> > > 
> > > nit(?): I'm not sure if there's a preference between "only takes a single
> > > round trip" and "consists of a single request/response exchange".
> > 
> > We neither, so we decided to let it be (there are a lot of last
> > minute changes anyway).
> > 
> > >  Figure 1 provides an overview of the high-level interaction between
> > >  the client, the NTS-KE server, and the NTP server.  Note that the
> > >  cookies' data format and the exchange of secrets between NTS-KE and
> > >  NTP servers are not part of this specification and are implementation
> > >  dependent.  However, a suggested format for NTS cookies is provided
> > >  in Section 6.
> > > 
> > > This is analogous to the RFC 5077 treatment of TLS session tickets that
> > > gives a "recommended" construction; however, while RFC 8446 to some
> > > extent supersedes RFC 5077's specification, 8446 specifically does not
> > > give a "recommended" construction, since there's not a whole lot that
> > > specifically needs to be recommended.  I wonder if it would be better to
> > > discuss this as an "example" construction rather than a "recommended"
> > > one.
> > 
> > We could not see a big advantage of changing it, it would have to be
> > changed in several places, so we let it be.
> > 
> > > Section 3
> > > 
> > >  Since the NTS protocol is new as of this publication, no backward-
> > >  compatibility concerns exist to justify using obsolete, insecure, or
> > >  otherwise broken TLS features or versions.  Implementations MUST
> > > 
> > > So we can mandate TLS 1.3?
> > > (Among other things, this would also obviate any need to worry about
> > > requiring extended master secret before using the TLS Exporter
> > > interface.)
> > 
> > This has been debated back and forth several times in the WG, and the last time
> > it was still an argument that there were a lot of installations that would have
> > had trouble supporting TLS 1.3, so demanding 1.3 would have slowed down the
> > adoption rate. Considering the current state of TLS 1.3 adoption in operating
> > systems, we are not entirely sure those arguments still hold.
> 
> I guess the key question is whether any OSes would be willing to take an
> updated NTP implementation that supports NTS but get stymied by the lack of
> support for TLS 1.3 in the system TLS implementation.  I know that FreeBSD,
> for example, will not upgrade the openssl major version within a stable release
> branch, but I believe similar arguments apply to the ntpd major version.
> 
> If we do retain support for TLS 1.2, I do think we need to say that the
> extended-master-secret extension must be used for those connections,
> though.

This and other issues showed that keeping 1.2 support would have
required handling several special cases and considerations for the use
of 1.2 vs >=1.3. It would also risk getting any of them wrong, or
missing some, in the specification or the implementations.

We have now removed support for 1.2 and require >=1.3.
A big step, but it is now or never (at least not in a long time).

For running on older OS:es without native 1.3 support, the
implementations will have to be linked to a separate library
supporting 1.3. This is inconvenient, but doable.

> > >  conform with [RFC7525] or with a later revision of BCP 195.  In
> > >  particular, failure to use cipher suites that provide forward secrecy
> > >  will make all negotiated NTS keys recoverable by anyone that gains
> > >  access to the NTS-KE server's private key.  Furthermore:
> > > 
> > > So forward secret ciphersuites are REQUIRED?
> > 
> > Since TLS 1.3 requires PFS ciphers, this is only an issue with TLS 1.2
> > implementations. Our wording in Section 3 should be interpreted that we
> > advocate PFS ciphers for TLS 1.2 implementations in the spirit of Section 6.3 in
> > RFC 7525: "This document therefore advocates strict use of forward-secrecy-only
> > ciphers."
> > 
> > > Section 4
> > > 
> > > Is there any significance to the ordering of records within a message
> > > (other than End of Message)?
> > 
> > No.
> > 
> > >  The NTS key establishment protocol is conducted via TCP port
> > >  [[TBD1]].  The two endpoints carry out a TLS handshake in conformance
> > > 
> > > [See discuss point about server discovery]
> > > 
> > >  in the TLS-protected channel.  Then, the server SHALL send a single
> > >  response followed by a TLS "Close notify" alert and then discard the
> > >  channel state.
> > > 
> > > [side note: since we define NTS-KE to be the single request/response
> > > pair and the response is self-framing, the close_notify is kind of
> > > redundant.  Which is not to say that it's bad to talk about, though it
> > > does make one wonder why we only say the server sends it and don't have
> > > the client also send one in response.  (Note that RFC 8446 has "Each
> > > party MUST send a "close_notify" alert before closing its write side of
> > > the connection, unless it has already sent some error alert.")]
> > 
> > Thank you! This has been addressed in draft 25.
> > 
> > > While the prospect of a reflection attack seens unlikely (would a
> > > machine be acting as both NTS-KE client and server?), it may be worth
> > > introducing some marker of whether a record is being sent by a client
> > > vs. a server.
> > 
> > We don't really see how that attack would work.
> > 
> > Yes, a machine could be both a NTS-KE client and server, for example an
> > NTS enabled stratum 2 server box. With bad configuration, it could
> > potentially use itself as a server, including negotiating cookies
> > and keys for itself, just as can be done with an NTP server today
> > (but if nothing else the NTP stratum mechanism should make that
> > time link useless).
> 
> IMO explicitly noting the message direction is more robust and future-proof
> (should any groundbreaking extensions appear), but I cannot insist on a
> change here.

Since the protocol currently only supports a single request which gets
a single reply, this should not be a problem until someone makes
changes to do more interactions per session. We opted to not change
this now, since this would be a larger change that should be well
thought through, and since this can be handled by that future change.

> > >  For clarity regarding bit-endianness: the Critical Bit is the most-
> > >  significant bit of the first octet.  In C, given a network buffer
> > >  `unsigned char b[]` containing an NTS-KE record, the critical bit is
> > >  `b[0] >> 7` while the record type is `((b[0] & 0x7f) << 8) + b[1]`.
> > > 
> > > When people use C syntax for examples we tend to ask for a reference or
> > > at least indication of a language version, though I can't imaging this
> > > particular aspect changing in a future version of C...
> > 
> > Neither can we. Draft 24 was updated to emphasize that we are refering to the C
> > programming language. (There is a bit called C on the same page of the draft.)
> 
> (I didn't even notice that there was a 'C' bit nearby!)
> 
> > > Please mention somewhere that the angle brackets in Figure 3 are used to
> > > denote optional records.
> > 
> > We could not figure out a reasonable way to fix this to make it
> > entirely right, and it is meant just an illustration to the text
> > which has the details.
> > 
> > > Section 4.1.2
> > > 
> > > Just to check: because of the "close_notify" requirement, any future
> > > "Next Protocol" is going to only be using the key material established
> > > through this NTS-KE session, but cannot be using this TLS connection
> > > (akin to STARTTILS ... or should I say STARTNTS?).
> > 
> > Correct, the protocol that uses the key material will not use this
> > TLS connection.
> > 
> > > Section 4.1.3
> > > 
> > > What do I do if I get an Error record without the critical bit set?
> > > Treat it as ... an error?
> > 
> > The critical bit should only be checked if the receiver does not
> > recognise the record id, see Section 4, page 9 of the draft.
> > 
> > > Section 4.1.5
> > > 
> > > The AEAD Algorithms registry just gives a listing of "raw" ciphers; do
> > > we expect there to be a need for additional specifications that cover
> > > how to integrate the AEAD cipher into a given NTS Next Protocol?
> > > (Obviously this document does so for NTPv4.)
> > > 
> > >  If the NTS Next Protocol Negotiation record offers Protocol ID 0 (for
> > >  NTPv4), then this record MUST be included exactly once.  Other
> > >  protocols MAY require it as well.
> > > 
> > > (This seems to in practice force any future NPN protocols to not try to
> > > define semantics for including AEAD Algorithm Negotiation more than once
> > > ... which is probably fine.)
> > 
> > Yes, that is pretty much the expectation, though details from the
> > NTPv4 next protocol may be reused.
> > 
> > > Section 4.1.7
> > > 
> > >  The NTPv4 Server Negotiation record has a Record Type number of 6.
> > >  Its body consists of an ASCII-encoded [ANSI.X3-4.1986] string.  The
> > > 
> > > Is the ANSI ref preferred over RFC 20?
> > 
> > No, it is now changed to RFC 20.
> > 
> > >  contents of the string SHALL be either an IPv4 address, an IPv6
> > >  address, or a fully qualified domain name (FQDN).  IPv4 addresses
> > > 
> > > I note that up in Section 1.2 we just said that the NTS-KE server sends
> > > "an IP address to [sic] the NTP server for which the cookies are valid",
> > > which is internally inconsistent.
> > 
> > Right, fixed with a small change of the text.
> > 
> > >  [RFC4291] and MUST NOT include zone identifiers [RFC6874].  If a
> > >  label contains at least one non-ASCII character, an A-LABEL MUST be
> > >  used as defined in section 2.3.2.1 of RFC 5890 [RFC5890].
> > > 
> > > It's probably implicit, but this of course only applies in the "FQDN"
> > > case, and the previous sentences were talking about the IP-address
> > > cases...
> > 
> > You are right. We have updated this to get it more explicit.
> > 
> > >     The disambiguating label string MUST be "EXPORTER-network-time-
> > >     security/1".
> > > 
> > > I guess it's a matter of taste (though maybe we had some IAB guidance?),
> > > but the version suffix "/1" may not be needed.
> > 
> > We intend this to be a version number, should it ever be needed.
> 
> I think that was clear.  The argument (IIRC) I refer to is basically that
> the protocol-level difference between going from "foo/1" to "foo/2" and
> going from "foo" to "foov2" is negligible, and in the latter case you get
> prettier identifiers for the initial batch of time.

Right, and this also seems to be the contemporary way of doing it.
We changed this. Again, now or never. Thanks for the suggestion!

> > > Section 5.1
> > > 
> > > My instincts are telling me to try to include, e.g., the TLS SNI value
> > > and the NTPv4 Server/Port negotiation results in the exporter input, but
> > > I think that is not actually needed (the TLS handshake transcript
> > > protects the SNI value and the TLS connection itself protects the
> > > others).
> > 
> > Indeed, this should be handled by the TLS layer.
> > 
> > > Section 5.2
> > > 
> > >  field is to enable the server to offload storage of session state
> > >  onto the client.  The purpose of the unique identifier extension
> > >  field is to protect the client from replay attacks.
> > > 
> > > Just to check: this is a "unique identifier" of the NTP packet (i.e., a
> > > nonce), not a "unique identifier" of the client or some other persistent
> > > entity?
> > 
> > Yes. It is all described in the following section (5.3).
> > 
> > > Section 5.5
> > > 
> > > It's probably worth mentioning here that the Cookie Placeholder is
> > > allowed to appear multiple times (Section 5.7 gives the recommended
> > > procedure for doing so).
> > > (Interestingly, the TLS WG is currently discussing a TLS extension to
> > > allow a client to request a specific number of cookies on a given
> > > handshake, which RFC 8446 left entirely to the server's discretion.)
> > > [As another side note, this "single-use cookie with a new one returned
> > > on each exchange" mechanism reminds me a lot of the ACME nonce usage,
> > > though ACME places the burden of single-use on the server whereas for
> > > NTS the client can enforce or choose to not enforce single-use.]
> > 
> > We have elected to specify all details regarding that in Section 5.7. only.
> > 
> > >  prevent DDoS amplification.  The contents of the NTS Cookie
> > >  Placeholder extension field's body are undefined and, aside from
> > >  checking its length, MUST be ignored by the server.
> > > 
> > > It seems like whenever we leave something's contents undefined, some
> > > clever person in the future is going to want to repurpose it for
> > > something...if we require it to be all-zeros (or some other value) for
> > > now, such future extensions can be more reliable.
> > 
> > Good point, we have now changed it to "SHOULD be all zeros".
> > (It also avoids sending data from uninitialised buffers.)
> > 
> > > Section 5.6
> > > 
> > >  not exceed the length of the request.  For mode 4 (server) packets,
> > >  no Additional Padding field is ever required.  For mode 3 (client)
> > > 
> > > nit: s/ever/never/
> > 
> > Here it actually should be "ever".
> 
> Indeed, my apologies -- I must have missed the "no".
> 
> > >  Senders are always free to include more Additional Padding than
> > >  mandated by the above paragraph.  Theoretically, it could be
> > > 
> > > Is there a practical upper bound on how much padding could be included?
> > > Would we expect servers to discard packets with "too much" additional
> > > padding?
> > 
> > It will be limited by the maximum UDP packet size. We can't see any
> > reason to put any upper boundary on the padding, since having more
> > than necessary will mainly harm the sender, and the harm to the
> > receiver (e.g. filling packet reassembly buffers) could likely be
> > accomplished anyway with or without padding.
> > 
> > > Section 5.7
> > > 
> > >     Exactly one Unique Identifier extension field which MUST be
> > >     authenticated, MUST NOT be encrypted, and whose contents MUST NOT
> > >     duplicate those of any previous request.
> > > 
> > > What's the scope of the "MUST NOT duplicate"?  Per C2S key?  Per client?
> > 
> > Text updated to clarify this.
> > 
> > >     and MUST NOT be encrypted.  The cookie MUST be one which has been
> > >     previously provided to the client; either from the key exchange
> > >     server during the NTS-KE handshake or from the NTP server in
> > >     response to a previous NTS-protected NTP request.
> > > 
> > > nit: comma instead of semicolon (the second half is not a complete
> > > sentence).
> > 
> > Fixed.
> > 
> > >  preference.  Section 10.1 describes the privacy considerations of
> > >  this in further detail.
> > > 
> > > nit(?): maybe "this tradeoff"?
> > 
> > It may not be a "tradeoff" in all situations, so we let it be.
> > 
> > >  fields which MUST be authenticated and MAY be encrypted.  The number
> > >  of NTS Cookie Placeholder extension fields that the client includes
> > >  SHOULD be such that if the client includes N placeholders and the
> > >  server sends back N+1 cookies, the number of unused cookies stored by
> > >  the client will come to eight.  The client SHOULD NOT include more
> > >  than seven NTS Cookie Placeholder extension fields in a request.
> > >  When both the client and server adhere to all cookie-management
> > >  guidance provided in this memo, the number of placeholder extension
> > >  fields will equal the number of dropped packets since the last
> > >  successful volley.
> > > 
> > > Choosing the number of Cooke Placeholders to include in this way can
> > > introduce a side channel of information for (essentially) the amount of
> > > packet loss the client is observing.  In the interest of reducing the
> > > number of side channels available, perhaps we should give some positive
> > > guidance to prefer encrypting the placeholders rather than leaving them
> > > authenticated but unencrypted?
> > 
> > The number of cookie placeholders can be seen from the packet size
> > anyway, so there is no real gain in hiding them.
> > 
> > >  1280 octets if no non-NTS-related extensions are used; 1280 octets is
> > >  the minimum prescribed MTU for IPv6 and is in practice also safe for
> > >  avoiding IPv4 fragmentation.  Nonetheless, senders SHOULD include
> > > 
> > > (If I held the pen this would be "generally safe", but I do not have
> > > specific grounds to object to this phrasing.)
> > 
> > That is better, changed.
> > 
> > > I know that RFC 5116 discusses AEAD decryption as producing either
> > > plaintext or FAIL, but when Figure 5 has the server merely "verify time
> > > request message", should we be saying anything about "decrypt encrypted
> > > extension fields EF" as well?  [Similarly for the client's "verify time
> > > response message".]
> > 
> > We saw little gain in adding more text to the figure, so we didn't.
> > 
> > >  present.  Servers MAY implement exceptions to this requirement for
> > >  particular extension fields if their specification explicitly
> > >  provides for such.
> > > 
> > > [the specification of the server or the specification of the extension
> > > field?]
> > 
> > The specification of the extension field.
> > 
> > >  Upon receiving an NTS-protected request, the server SHALL (through
> > >  some implementation-defined mechanism) use the cookie to recover the
> > >  AEAD Algorithm, C2S key, and S2C key associated with the request, and
> > >  then use the C2S key to authenticate the packet and decrypt the
> > >  ciphertext.  If the cookie is valid and authentication and decryption
> > >  succeed, the server SHALL include the following extension fields in
> > >  its response:
> > > 
> > > [This is not consistent with the RFC 5116 AEAD terminology, though it's
> > > hard to make a strong case that it needs to be.]
> > 
> > We believe it is clear enough as it is.
> > 
> > >  The server MAY include additional (non-NTS-related) extension fields
> > >  which MAY appear prior to the NTS Authenticator and Encrypted
> > >  Extension Fields extension field (therefore authenticated but not
> > >  encrypted), within it (therefore encrypted and authenticated), or
> > >  after it (therefore neither encrypted nor authenticated).  In
> > >  general, however, the client MUST discard any unauthenticated
> > >  extension fields and process the packet as though they were not
> > >  present.  Clients MAY implement exceptions to this requirement for
> > >  particular extension fields if their specification explicitly
> > >  provides for such.
> > > 
> > > [same as for the other direction]
> > 
> > (Same answer here.)
> > 
> > >  If the NTP server has previously responded with authentic NTS-
> > >  protected NTP packets (i.e., packets containing the NTS Authenticator
> > >  and Encrypted Extension Fields extension field), the client MUST
> > > 
> > > I do see "authentic" outside the parenthetical, but still wonder if
> > > "authentic" or "properly validated" or something should also appear
> > > within the parenthetical.
> > 
> > Right, the parenthesis is now removed.
> > 
> > >  protocol upon forced disassociation from an NTP server.  In that
> > >  case, it MUST be able to detect and stop looping between the NTS-KE
> > >  and NTP servers by rate limiting the retries using e.g. exponential
> > >  retry intervals.
> > > 
> > > [I frequently express a desire to see the base of the exponent indicated
> > > when "exponential retry intervals" are specified, though it does not
> > > seem mandatory for correcness of operation here.]
> > 
> > Here, exponential is just an example, it may be dependant on the
> > application, e.g. something with very low available bandwidth
> > and/or low time dependency may not want to waste cycles on this,
> > and the exact rate limiting methods and constants may need to change 
> > when we get more knowledge or other external circumstances change.
> > 
> > > Section 6
> > > 
> > >  nonce reuse.  It need not be the same as the one that was negotiated
> > >  with the client.  Servers should randomly generate and store a master
> > >  AEAD key `K`. Servers should additionally choose a non-secret, unique
> > > 
> > > And in general should be the same for all cookies generated by a given
> > > server?
> > 
> > Yes, until it is switched to the next 'K'.
> > 
> > > Also, we should say that 'K' should (MUST?) not be used for any other
> > > purpose.
> > 
> > Do we really have to explicitly specify that for a secret key?
> 
> Unfortunately, it seems that we do, in general.  Too many instances of
> software authors reusing a key because it was convenient :(
> 
> > Also, Section 6 is non-normative and contains general suggestions.
> > Specific details are implementation dependent.
> 
> Fair enough.

Since this is just one of probably an infinite set of ways that
someone could make bad decisions in their designs, and since this is
not part of the NTS protocol, we decided to not changes this.

We added an extra "secret" before "key K" in the text though, to
emphasise that it is.

> > >  periods prior.  Servers should continue to accept any cookie
> > >  generated using keys that they have not yet erased, even if those
> > > 
> > > nit: "been erased"
> > 
> > We think the text is correct as it is.
> 
> And I agree, sigh.  (I assume I read past "they".)
> 
> > >  function of its predecessor.  A recommended concrete implementation
> > >  of this approach is to use HKDF [RFC5869] to derive new keys, using
> > >  the key's predecessor as Input Keying Material and its key identifier
> > >  as a salt.
> > > 
> > > Do we need to say to generate the appropriate amount of output material
> > > from HKDF as needed for the next-generation key?  (I kind of hope not,
> > > but...)
> > 
> > Section 6 is non-normative and contains general suggestions. Specific
> > details are implementation dependent.
> > 
> > >  The cookie should consist of the tuple `(I,N,C)`.
> > > 
> > > Are 'I' and 'N' fixed-length or length-prefixed?
> > 
> > (Same answer as above.)
> > 
> > >  and nonce `N` with no associated data.  If decryption or verification
> > >  fails, reply with an NTS NAK.  Otherwise, parse out the contents of
> > > 
> > > [same comment about RFC 5116 terminology]
> > 
> > (Same answer as above.)
> > 
> > > Section 7.2
> > > 
> > > In a similar vein as my comment on the
> > > "EXPORTER-network-time-security/1" string, perhaps the "/1" is not
> > > necessary here?
> > 
> > Here too it is meant to be a version number.
> > 
> > > Section 7.3
> > > 
> > > I will send a note to the TLS WG giving them a heads-up about the
> > > "Recommended" value of "Y", though my sense is that this should be in
> > > keeping with the spirit of recommended values.
> > 
> > We have changed this to an "N" in draft 24 since we thought it was more correct
> > for this type of label. We immediately learned it was not. Thus, it will be
> > changed back to 'Y' in draft 25.
> 
> Sorry to have caused any confusion!
> 
> > > Section 9
> > > 
> > > We should probably have some discussion of the consequences of
> > > compromise of the server's cookie-encryption key.  Possibly also for
> > > when the key associated with a cookie is compromised from a client,
> > > though if cookies are single-use the scope may be quite limited.  (That
> > > is, since we are using the shared symmetric key for authentication by
> > > virtue of "I didn't send it so it must be from the other party that has
> > > the key".)
> > 
> > NTP and NTS-KE server operators are expected to remove compromised keys as soon
> > as the compromise is discovered. This would cause the NTP servers to respond
> > with NTS NAK, thus forcing key renegotiation. We are aware that this does
> > not protect against MITM attacks where the attacker has access to a
> > compromised cookie encryption key.
> 
> I did not doubt that you are aware of this; the question is whether all
> future readers of this document will similarly be aware.

Right, we added a new section ("Key Compromise") to describe the
problem and how it should be handled.

> > > Section 9.2
> > > 
> > >  NTS users should also consider that they are not fully protected
> > >  against DDoS attacks by on-path adversaries.  In addition to dropping
> > >  packets and attacks such as those described in Section 9.5, an on-
> > >  path attacker can send spoofed kiss-o'-death replies, which are not
> > >  authenticated, in response to NTP requests.  This could result in
> > > 
> > > Wouldn't these on-path attacks just be DoS attacks, not DDoS ones?
> > 
> > Right, changed.
> > 
> > >  significantly increased load on the NTS-KE server.  Implementers have
> > >  to weigh the user's need for unlinkability against the added
> > >  resilience that comes with cookie reuse in cases of NTS-KE server
> > >  unavailability.
> > > 
> > > It's perhaps worth mentioning the requirement for kiss-o'-death replies
> > > to echo the message identifier once a server is known to have sent valid
> > > NTS-protected messages, which limits the scope of this attack to on-path
> > > attackers.
> > 
> > Since we already state that the attack is limited to on-path attackers in the
> > same paragraph, we have elected not to make further changes.
> 
> Okay.
> 
> > > Section 9.3
> > > 
> > > It might be worth some discussion of the potential tradeoffs involved in
> > > encoding a client IP address into a cookie, with respect to allowing for
> > > some form of source-address validation (and thus protection against
> > > serving as a DDoS reflector).  QUIC, for example, has facility for
> > > encoding source-address validation state into cookie-like artifacts.
> > > Such IP address recording comes with concordant privacy and tracking
> > > considerations, though (hence "tradeoffs").
> > 
> > This would prevent mobile clients from using stored cookies. In the
> > end, it's an implementation issue.
> 
> True.
> 
> > > Additionally, it seems that if the "NTP Server Negotiation" record is
> > > optional for the client to send but can be set by the server, that would
> > > result in some modest packet expansion, which would be topical to
> > > mention here.
> > 
> > The NTP Server Negotiation record is used by NTS-KE and is not an
> > NTP extension field. There is no need to protect against packet
> > expansion when using the TCP/TLS based NTS-KE part of NTS.
> > 
> > > Section 9.5
> > > 
> > >  mitigate this attack.  However, the maximum error that an adversary
> > >  can introduce is bounded by half of the round trip delay.
> > > 
> > > Half of the round-trip after the delay attack or in its absence?
> > 
> > Half of the round-trip time after the delay attack. The round-trip
> > time is typically limited to 1 second in many implementations.
> > We have added a link to the referenced paper where this is explained.
> 
> Ah, thanks.
> 
> > > Section 9.6
> > > 
> > >  At various points in NTS, the generation of cryptographically secure
> > >  random numbers is required.  Whenever this draft specifies the use of
> > > 
> > > The string "random" appears only in the construction of the unique
> > > identifier and the non-normative recommended cookie construction, and
> > > the RFC 5116 recommended nonce construction also takes a random input;
> > > do we want to mention specific examples instead of or in addition to
> > > saying "various points"?  (I might consider mentioning that the SIV mode
> > > helps reduce the dependence on strong random numbers as wel.)
> > 
> > This section was removed and Section 5.7 updated to clarify this.
> > 
> > > Section 9.7
> > > 
> > > Is there anything to say about a target end-state where some peers can
> > > successfully require NTS for all connections (i.e., no fallback of any
> > > sort)?
> 
> (I'll take that as a "no".)

(The missing reply was an oversight, by me.)

We believe that this will be handled by the ongoing work on NTS
service discovery etc. in the WG. In addition, it is a question
about implementation and application.

> > >  For the reasons described here, implementations SHOULD NOT revert
> > >  from NTS-protected to unprotected NTP with any server without
> > >  explicit user action.
> > > 
> > > Which would then force a choice between tolerating bad time and reusing
> > > cookies, until user action can be obtained?
> > 
> > This is about not accepting a downgrade from NTS-protected to
> > unprotected (traditional) NTP. Whether cookies are reused or not is
> > another issue, and is only about unlinkability of the client vs
> > resilience of the time link.
> 
> If there's an attacker stripping NTS so I get unprotected responses, I'll
> quickly use up my cookie pool since there's no NTS return traffic to
> replenish it.  I could try NTS-KE again, though maybe the attacker can
> block that as well.  With no cookie pool, I either have no time service at
> all, an insecure time service, or a secure time service with cookie reuse.
> I expect most people to choose the last one, but perhaps in some cases the
> first one would be preferred.

The client MUST only accept responses that are authentic under the S2C
key associated with the corresponding request.

So stripping of NTS extensions from the NTP response (or sending a
fake response) should (=MUST) just lead to the client discarding the
response without further processing.
We added some more explicit text in "Protocol Details" to emphasise
this. We also changed a word in "NTS Stripping" to make it more clear
that this should not be possible to do.

If the attacker can block the KE server traffic, it could likely block
any traffic, and then the client could find itself without time
service or any other service, with or without NTS.
(We would love to be able to solve this though, it would fix so many
things, from attacks to broken cables, we regret to say that NTS is
not the answer to this. :-) )

> > > Section 10.1
> > > 
> > >  The unlinkability objective only holds for time synchronization
> > >  traffic, as opposed to key exchange traffic.  This implies that it
> > >  cannot be guaranteed for devices that function not only as time
> > >  clients, but also as time servers (because the latter can be
> > >  externally triggered to send authentication data).
> > > 
> > > Are we just talking about the TLS server certificate?
> > 
> > Likely that was what was meant, but not necessarily, there are
> > other ways, to identify a server, the time scale, for example.
> > 
> > The text has been updated to better reflect this.
> > 
> > > Section 12.1
> > > 
> > > I'm not sure that a normative reference is needed for "MUST NOT do this"
> > > (to RFC 6874).
> > 
> > Implementers are required to know about zone identifiers and how they
> > are formatted in order to be able to detect them correctly if they
> > are received.
> > 
> > > RFC 7507 seems used only in the definition of SCSV, which is itself
> > > unused in the document.
> > 
> > Right, removed now.
> > 
> > > Section 12.2
> > > 
> > > Why is RFC 5280 only informative but RFC 6125 is normative?
> > 
> > Changed to be normative.
> > 
> > > The "SHALL" regarding HKDF usage in TLS 1.3 exporters would seem to make
> > > RFC 5869 a normative reference per
> > > https://www.ietf.org/about/groups/iesg/statements/normative-informative-references/,
> > > though in my opinion the RFC 8446 reference would suffice for that
> > > usage.
> > 
> > Changed to be normative.
> > 
> > > Also, Daniel should update (or remove) the address in his Author entry.
> > 
> > He has. :-)
> 
> 
> Thanks for all the updates!
> 
> -Ben