Re: [DNSOP] CNSRRSIG (was: Re: [Ext] draft-fujiwara-dnsop-delegation-information-signer))

[Mark Andrews <[e-mail]>]

Before going on I would really like to know what operational problem is being
attempted to be solved by signing delegating information?

Fujiwara-san has presented the draft without specifying what problem it is
attempting to solve.  The fact the records are not signed is a observation
not a problem per say.

Mark


> On 11 Dec 2020, at 10:48, Brian Dickson <[e-mail]> wrote:
> 
> 
> 
> On Thu, Dec 10, 2020 at 1:19 PM Peter van Dijk <[e-mail]> wrote:
> Hello Paul,
> 
> On Mon, 2020-11-30 at 15:43 +0000, Paul Hoffman wrote:
> > The more I think about draft-fujiwara-dnsop-delegation-information-signer, the more I think that it is much more complex than what we are doing now in DNSSEC, and therefore undesirable.
> 
> My feelings are similar but not identical - the conversation already
> shows that the glue story will be almost impossible to implement. Next
> to that, I haven't figured out why we'd want to authenticate glue.
> However, authenticating the delegation NSset fills an obvious gap in
> various suggested answers to the dprive phase2 question (privacy
> between resolvers and authoritatives).
> 
> > If the goal is "a way for a signer in a parent to sign child NS in a way that does not affect validators that have not been updated (or don't care)", a significantly easier solution would be a new RRtype (maybe called "CNSRRSIG") that closely mimics RRSIG but only allows child NS for signing. An aware signer included the CNSRRSIG in the zone, and an aware authoritative server includes in in the Authority section when serving child NS records. An aware resolver can use this, an unware resolver would treat it like any other unknown RRtype that appears in the Authority section.
> 
> 
> I think this is close to the mark, but still slightly off.
> My take on this is roughly as follows:
> 	• The goal here (as I understand it, or perhaps how I interpret it) is to have NS data, corresponding to the child apex NS set, published on the parent zone and signed by the parent (and used by upgraded resolvers without breaking anything not upgraded).
> 	• This involves at minimum the following elements, presented in a kind of "reverse Polish" order:
> 		• An RRSIG (or RRSIG-like thing) in the parent zone over something new (TBD1)
> 		• A way of publishing that "something new" to the parent (TBD2)
> 		• A way of authenticating/validating that "something new" in the process of publishing to the parent (TBD3)
> 		• A way of linking the published "authenticating/validating" data with data signed in the child zone, when being used/validated by the resolver (or client) (TBD4)
> 	• For "TBD1" If we keep the existing RRSIG signatures themselves, the only question is when to include the "something new" in the delegation response, along with the matching RRSIG. 
> 		• This suggests no need for a new RRSIG-like record type per se.
> 		• Existing 403x/5155 logic handles adding RRSIGs to answers
> 	• For "TBD2", the concern over what "something new" could be, and how it gets published to the parent. 
> 		• Options to consider for "something new" itself are:
> 			• A new RRTYPE with owner name of "apex"
> 			• An existing RRTYPE with owner name of "apex", with new parameters
> 			• An existing RRTYPE with different owner name
> 			• (For completeness, A new RRTYPE with different owner name)
> 		• Options to consider for "how it gets published" are basically:
> 			• EPP
> 			• This suggests the fewer "new" things involved, the better. Zero new things would be ideal.
> 		• This itself leads to preference for "An existing RRTYPE with owner name of 'apex', with new parameters", specifically:
> 			• Published RRTYPE == DS
> 			• DS value is hash of DNSKEY with new parameters, specifically a new algorithm
> 			• The new algorithm is "concatenation of child apex NS RRSET in canonical form and order" (i.e. that is what the DNSKEY 'public key' field would be comprised of)
> 			• NB: The existing specifications handle this -- new (unknown) algorithms MUST be treated as if the record didn't exist. 
> 				• Some implementations may need to be fixed, based on some experiments thus far.
> 	• I think TBD3 and TBD4 are flip sides of the same coin.
> 		• RRTYPE == DNSKEY or DS (depending on TLD) transferred via EPP to the parent
> 		• If RRTYPE transferred is DNSKEY, perform hash to produce DS.
> 		• TBD3 is probably optional as it related to EPP
> 		• TBD4 is the resolver/client validation, in two parts:
> 			• Construct DNSKEY from input data (NS RRSET)
> 			• Hash to create DS
> 			• Compare to published DS
> 			• Check RRSIG over DS
> 	• The use case of getting the glue (and validating it) without actually sending queries to the auth server itself prior to validating the glue, is a classic "catch 22". The owner name of the NS set in the child zone, is "privacy sensitive", at least in most use cases. However:
> 		• Keeping the apex and delegation NS set in sync, and the DNSKEY and DS in sync, can address this problem:
> 			• Use CDS/CDNSKEY to publish DS/DNSKEY records that match the current AND updated NS RRSET, prior to changing the NS RRSET
> 			• Use the delegation NS RRSET rather than the child apex NS RRSET for the validation (ie for recreation of DNSKEY and DS records)
> 			• This would add extra steps when changing NS, but would still be deterministic and possible to build into a state machine, with an extra moving part (or two)
> 			• It would then be feasible to get validated glue prior to querying the authority server (e.g. for its TLSA record) and/or to establish a TLS connection to the correct IP and name.
> The TL;DR: of this is: no EPP changes needed, no new RRTYPEs needed, slightly more steps in rolling the NS set during a change.
>  
> This makes a ton of sense to me.
> 
> Compared to DiS, registrar complexity is identical (because the
> complexity is also hidden in the signer here); signer complexity is
> potentially lower. The only real complexity change vs. DiS is in the
> auths, that now need to know to serve CNSRRSIG from the parent side in
> the additional part of a delegation response. For resolvers, this vs.
> DiS is again pretty much moot.
> 
> The CNSRRSIG would also require delegation auths (i.e. TLDs) to make changes, and I think also require EPP changes.
> See above for a very slightly different take on this idea, that avoids these issues.
> (EPP and TLDs are definitely the "long pole in the tent".)
>  
> 
> > There are probably a few other diffs from the RRSIG definition in RFC 403x, but they should be minor. This might make implementation more likely to be correct for signers, servers, and resolvers.
> 
> Yes, this calls for some experiments, but I suspect the outcome will be
> as you described above - which means no hurdles to incremental rollout.
> 
> (I am not even convinced this needs to be a new type, vs. reusing RRSIG
> under the specific semantics you described, but a new type feels
> cleaner to me.)
> 
> > Thoughts?
> 
> +1 !
> 
> +1 for me too.
> 
> Brian 
> _______________________________________________
> DNSOP mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/dnsop

-- 
Mark Andrews, ISC
1 Seymour St., Dundas Valley, NSW 2117, Australia
PHONE: [phone]              INTERNET: [e-mail]