This is the first in a series of articles describing my progress towards introducing EDNS(0) and DNSSEC support to Twisted Names.
In this article, I'll describe two new private classes designed to parse EDNS(0) messages.
But first a little background.
The EDNS protocol extensions depend on the use of a special DNS record which crucially contains the advertised maximum datagram size, extended return codes, a DNSSECOK flag, and a payload which contains zero or more extensible headers.
This special record is called an OPT record. It isn't a real record though; it doesn't convey any domain names or domain records. It's a hack to work around the lack of space in a standard 512 byte DNS message. All the header flags and fields that couldn't be shoehorned into the 96bit header section of a DNS message are instead encoded into an OPT "pseudo" record.
EDNS clients and servers append one OPT record to the "additional" records in a sent DNS message. This record is then extracted and interpreted by an EDNS aware receiver.
The first wrapper is a class called twisted.names.dns._OPTHeader, which wraps an instance of twisted.names.dns.RRHeader. This allows us to re-use the existing record parsing code, and means that the code in _OPTHeader is purely concerned with the special behaviour of an OPT record. Which seems neat! It also made the unit tests much more focused.
The second wrapper is a class called twisted.names.dns._EDNSMessage which is in trunk but not yet in a released version. This wraps an instance of twisted.names.dns.Message. And again, this allows us to re-use the existing parsing code while clearly separating the EDNS specific behaviour.
So Twisted Names can now interpret OPT records and EDNS messages, but we've decided to keep these private for the time being; until we are satisfied with the APIs.
This first version of _EDNSMessage has been written as a drop in replacement for dns.Message. But as a result it has inherited some of the ugliness of that class, namely:
- confusing abbreviated constructor argument names and attribute names, and
- a clunky decoding API which would be better implemented as a classmethod.
Tom has some ideas about introducing a new Interface for message constructors and I've done some work on improving the abbreviated names, but I got stuck trying to figure out how to introduce the changes while maintaining backwards compatibility.
I'd originally thought those things were blockers, because I intended that users would pass curried message constructors / factories to the various protocols and client APIs, as a way of configuring / overriding EDNS flags and fields in their requests. But that's not a good API (as Tom points out in another comment).
So I've decided that those cleanups would be nice to have, but probably not necessary right now.
A much simpler strategy – and my current priority – is to just switch all Twisted names code (client, protocols, server) to use the new dns._EDNSMessage class in place of dns.Message. This shouldn't introduce any backwards incompatibilities and I'll explicitly disable the use EDNS during encoding (for now). That way Twisted Names clients and servers should continue to behave exactly as before.
But I'll then be free to start work on a DNS client which can be configured to send EDNS(0) requests and which is capable of negotiating with EDNS(0) servers. That's the key requirement of the first project milestone.
(twisted.names.server and the twistd dns plugin will remain non-EDNS for the timebeing.)
In fact I've already started that work, but first I had to fix a bug in the way twisted.names.server generates response messages. My proposed fix includes new functions and methods for generating response messages from a request message. (Previously, the request message was being re-used as the basis for the response, and inevitably, client specific flags and fields were leaking into the response.)
Happily, these new APIs will also allow me to explicitly disable EDNS(0) in twisted.names.server, until such time as we implement server side EDNS(0) support.
So that's all for now. In my next post I hope to be able to report that Twisted Names is using the new _EDNSMessage class internally and summarise some of the other work that I've done on Twisted Names in support of the EDNS(0) and DNSSEC project.