[dns-operations] DNS cookies in a mixed resolver anycast environment
marka at isc.org
Fri May 31 08:15:29 UTC 2019
Well having different implementations serving the same client is
always a dangerous idea, whether is is different vendors or not.
Resolvers need to track server characteristics.
* Does the server support EDNS?
* Does EDNS buffer size X work?
* Does the server support DNS COOKIE?
RFC 7873 had BIND’s algorithms described so that other implementations
could implement them if they were intending to be configured in a
anycast scenario with BIND. Doing so reduces the number of BAD COOKIE
responses and subsequent fallback to TCP. DNS COOKIE clients will handle
different algorithms and/or different shared secrets.
As for discarding replies without DNS COOKIEs once the recursive server
learns that the server supports DNS COOKIEs there is only a issue if the
operator has configured a anycast server cluster for a given client with
servers that both support and don’t support DNS COOKIE. The client thinks
it is seeing spoofed responses when it see non-DNS COOKIE responses. If
you are configuring a anycast server it’s just one more thing you need to
ensure is consistent across the server set for a given client.
Having mixed support is much more of a issue for local anycast instances
than it is for global anycast instances.
> On 31 May 2019, at 5:11 pm, Patrik Lundin <patrik at sigterm.se> wrote:
> I have been trying to figure out how to best deal with DNS cookies in an
> environment where you are running multiple resolver implementations. From
> what I can tell, out of BIND, Knot Resolver, PowerDNS Recursor and
> Unbound only BIND is currently implementing cookie support. Knot seemed
> to have done so previously, but as of 3.0.0 the cookie support was
> removed (https://www.knot-resolver.cz/2018-08-20-knot-resolver-3.0.0.html)
> because of some ongoing work in the IETF DNSOP.
> Reading RFC 7873 it states "If the client is expecting the response to
> contain a COOKIE option and it is missing, the response MUST be
> discarded.", which leads me to believe that having a anycast cluster of
> a set of BIND servers where cookies are enabled together with a set of
> servers where the cookies are not supported would be a bad thing,
> causing clients to discard answers.
> Yet, when looking up how one would go about to disable the sending of
> cookies in responses to clients for BIND, the documentation for
> "answer-cookie" (https://ftp.isc.org/isc/bind9/cur/9.15/doc/arm/Bv9ARM.ch05.html)
> states the following:
> "answer-cookie no is intended as a temporary measure, for use when named
> shares an IP address with other servers that do not yet support DNS
> COOKIE. A mismatch between servers on the same address is not expected
> to cause operational problems, but the option to disable COOKIE
> responses so that all servers have the same behavior is provided out of
> an abundance of caution. DNS COOKIE is an important security mechanism,
> and should not be disabled unless absolutely necessary."
> If clients are instructed to discard replies where the cookie are
> missing, how can this not cause operational problems? Am I missing
> On a related note, given that a set of BIND servers are already having
> the default cookies enabled, what is the expected fallout of setting
> "answer-cookie no" if this turns out to be the favorable approach in
> this case?
> Patrik Lundin
> dns-operations mailing list
> dns-operations at lists.dns-oarc.net
> dns-operations mailing list
Mark Andrews, ISC
1 Seymour St., Dundas Valley, NSW 2117, Australia
PHONE: +61 2 9871 4742 INTERNET: marka at isc.org
More information about the dns-operations