Re: [TLS] Eric Rescorla's Discuss on draft-ietf-tls-dnssec-chain-extension-06: (with DISCUSS and COMMENT)

[Willem Toorop <[e-mail]>]

Op 26-02-18 om 15:26 schreef Paul Wouters:
> On Thu, 22 Feb 2018, Shumon Huque wrote:
> 
>> On Wed, Feb 21, 2018 at 2:48 PM, Paul Wouters <[e-mail]> wrote:
>>       On Wed, 21 Feb 2018, Shumon Huque wrote:
>>
>>             Okay, got it. For people that have already implemented
>> this, I think
>>             there has been an implicit understanding that the format
>> of the chain
>>             data is a sequence of concatenated wire format RRs exactly
>> as they appear
>>             in a DNS message section
>>
>>
>>       Note, I would not call it "sequence of concatenated wire format
>> RRs", as
>>       it is simply the wireformat of a DNS message.
>>
>>       The fact that some parts
>>       are concatenated RR's is just part of the wireformat of the DNS
>> reply
>>       message. That is, this isn't introducing some new form of
>> concatenated
>>       content - it is _just_ a regular DNS wire format message, and the
>>       document should not go deeper into its explanation.
>>
>>
>> It would _not_ be correct to say that this is a "DNS wire format
>> message" - that 
>> would mean there is a DNS header section with flags and response codes, 
>> and other sections. The chain data structure as currently specified
>> really is 
>> concatenated wire-format RRs (as they appear _within_ a DNS message 
>> _section_). Let me know if that is unclear in the draft (or my
>> proposed edits).
> 
> So it was decided to not use a full DNS packet format? And then since you
> miss the structure of the Answer Section and Additional/Authority
> Section, you require the "answer RR's" come first where you basically
> emulate an Answer Section?

No Paul, the division in sections is irrelevant for a verifier.  The
only bit of information in a DNS message that is used by a verifier is
the question.  From the question, validation starts and the relevant
records are followed and verified.  But the question section is also not
needed as the question can be derived from the name and port of the
service, i.e. <port>._tcp.<name>. TLSA

The order described in the draft is both an optimization to reduce the
number of times a verifier has to go over the RRs, and it makes the
content easier to read (and understand) for humans too.

Also, for non existence answers, DNSSEC validators (and thus also a
verifier for the chain extension) simply ignore the DNS message header.
Proof of non-existence can and must be derived from the set of RRs in
the message body/sections too.

The extension already supports Denial of Existence proof b.t.w., because
it is also needed for wildcard expansions (which are supported).

Personally, I do not have very strong feelings whether or not the
extension content is a DNS message or not, but a verifier would ignore
the header and the question anyway, so why bother with that overhead?
Also, you *can* use the content of the RFC7901 chain query.  You just
have to skip the header and question section, and then include the rest.

> 
> Isn't that an indication that we should really use the wireformat of an
> entire DNS message here? Maybe some DNS library/tools people can chime
> in here to tell us if this matters much to them (assuming they are
> adding support for creating/consuming these chains of RRsets in wire
> format.
> 
> I am personally a little sad we cannot have a dig +chainquery command
> where we write out the entire answer packet format to a blob, to be
> loaded by
> the TLS server. And where a TLS client cannot just hand over the blob
> "as if it came in as a reply from a DNS server" to its DNS/cache
> receiving code path.
> 
>> I recall at one point way back there was a discussion about whether
>> the chain
>> data should just be a fully formed DNS message, but I don't believe
>> that idea
>> had much support in the working group (personally I would have been
>> fine with 
>> that choice too, if it did have support).
> 
> Do you remember why not? I'll see about checking the archives, but to me
> the hint that you are losing information and require some kind of
> ordering seems to suggest there is a need for using the full DNS message
> reply format
> 
>> There is some residual wording in the draft about ordering of CNAMEs
>> etc in
>> the answer records part. Assuming the WG agrees, I am fine with relaxing
>> that requirement - that ended up in there because although there is no
>> defined
>> ordering of RRs within a DNS message section, as a practical matter
>> CNAMEs
>> are almost always ordered since there are some DNS queriers that get
>> confused 
>> otherwise. This is a new protocol though, so we can be more faithful
>> to the DNS 
>> spec.
> 
> I would prefer the residual wording to go away. Any hints at order being
> important should be squashed.
> 
>>       I don't think getting unrelated DNSSEC records would be an
>> issue. TLS
>>       has its maximum sizes for the handshake. In fact, it could allow
>> the
>>       extension to have some useful data in the case of MX or SRV. 
>> (and could
>>       be a feature to build a nice resolver over TLS using 1 tor
>> circuit).
>>
>> The draft currently doesn't address the MX or SRV use case. I suggest
>> that we tackle that in a future version.
> 
> Jus make sure the document doesn't forbid any such data, and allow the
> client to ignore these or put them in the cache as it sees fit.
> 
>>       I'm also not sure about the talking of unsigned CNAME records from
>>       DNAME. The above pseudo code (extended with special cases)
>> should be
>>       in some DNS library, and that library will know what records to
>> expect
>>       unsigned which are proven by the DNAME (or wildcard) synthesis
>> and knows
>>       when/if to add it to the validated cache. I don't think that
>> should be
>>       explained in this RFC at all. The DNS implementation does not need
>>       to be specified in this document and it should just focus on saying
>>       that "the DNS message response is validated and upon validation the
>>       content can be considered DANE validated".
>>
>>
>> Where we ended up, is that WG participants asked for some level of DNSSEC
>> detail to be included in this doc.
> 
> Again, that will only lead to bad implementations. There should be no
> need to update this document once something like ANAME, ALIAS or ZNAME
> is introduced to the DNS. So you should not talk about these things at
> all. TLS client should use a dns library that knows these things, or if
> they want to write one from scratch, these implementors should look at
> the DNS RFCs and not this RFC for guidance on how the DNS protocol
> works.
> 
> Paul