On 24 Jul 2014, at 14:58, Viktor Dukhovni <[email protected]> wrote:

> On Thu, Jul 24, 2014 at 10:19:08AM +0200, Olle E. Johansson wrote:
> 
>> Section 3:
>> 
>> I really like this here, but would like it to be expanded a bit to help
>> implementors. The DNS SRV RFC is not easy to understand and in the IP
>> world RFC 3263 has confused a lot of people. DNS SRV says that all
>> candidates for a given priority should be attempted before one moves to
>> the next priority. It doesn't say anything about order, which this document
>> does when you say "SHOULD be performed in parallel".
> 
> Note, the text at the top of Section 3:
> 
>   To expedite connection to the intended service, where possible the
>   queries described in the following sections SHOULD be performed in
>   parallel (this is similar to the "happy eyeballs" approach for IPv4
>   and IPv6 connections described in [RFC6555]).
> 
> is encouraging parallel *DNS lookup*, not parallel connection
> attempts.  And in fact parallel connection attempts would be wrong
> unless all the weights are equal, and, as a blanket policy, questionable
> even then.
> 
>> If one given priority have two hosts with two IPv4 and three IPv6 each -
>> should all of them be tried in parallell?
> 
> The answer is NO, because the document does not suggest connection
> concurrency, and because even that is I thing a mistake.  When the
> weights are unequal, the client is supposed to employ some sort of
> RNG to choose candidate servers with a probability based on the
> weights.  If the client is correctly employing weighted selection,
> parallel address lookup is generally inappropriate, in most cases
> the first (randomly) chosen server will in fact be available and
> lookups for other server addresses are I believe wasteful.
> 
> Thus parallel lookups are only useful when the first server does
> not respond and further lookups are required to find more servers,
> and could arguably have been performed in parallel with the first
> lookup.  However, after incurring a connection timeout with the
> first server, additional DNS lookup latency for the second server
> is not a major increment in the latency.
> 
> Also one needs TLSA lookups which should only follow the address
> lookups because the TLSA lookup should not be made when the address
> records are not secure.  (Also when CNAMEs are supported on the
> RHS of SRV records, the TLSA base domain is not known until the
> original address record is CNAME expanded).
> 
> Bottom line I think the "parallel" language should be withdrawn.
Or rewritten to make it olle-compatible ;-)

> 
>> This is a big issue and may be too big an issue for this document to cover,
>> especially with normative language.
> 
> This document definitely cannot specify anything about connection
> concurrency.  I would like to suggest strongly it also should *not*
> specify lookup concurrency.
+1

The reference to Happy Eyeballs, which is all about connection attempts,
mislead me I guess. I think we may add "the handling of network connection
attempts is not defined in this document".
> 
>> According to the SIP domain cert RFC the CN should be disregarded
>> and NOT used if there are any SANs. I don't know the
>> reasoning behind this. Anyone? Should we do that here too or just forget it?
> 
> This is IIRC either from 5280, 6125 or both.  DNS names in the
> subject CN are deprecated, SANs are preferred, and trump the CN
> when both (at least one SAN of type DNS and a CN in the subject
> DN) are present.

Then we need to clarify this again.

Thanks for your response, Viktor!

/O
_______________________________________________
dane mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/dane

Reply via email to