In RFC 1123 it is stipulated that:

a DNS resolver or server that is sending a non-zone-transfer query MUST send a UDP query first.

However, if you are doing a DNS UPDATE you really want the reliability that TCP offers, even if you don’t expect truncation to be an issue. The update is sent to the relevant authority server, so the arguments about load on root servers in the RFC aren’t applicable.

I’ve made the UDP implementation retry by default, but I think if you need more than 2 retries, you should be considering using TCP with its (much more advanced) retransmission algorithms.

Peter also found a bug in glibc’s res_send (actually in their send_dg function) whereby the resolver interprets the lack of the DNS “recursion available” flag in the header as an error. However, that flag isn’t even meaningful for DNS UPDATE responses; according to RFC 2136, those bits:

Should be zero (0) in all requests and responses. A non-zero Z field should be ignored by implementations of this specification.

As a result, glibc was setting errno to ECONNREFUSED or ETIMEDOUT even when the update was successful. I’ve hacked dudders to double-check after res_send, but it’s making me question the wisdom of using res_send at all, given that I’m constantly working around it.

Update: submitted glibc bug report #11950

To get dudders-1.04 on OpenWRT, simply update the official package feed and select dudders from the Net > DNS > dudders menu in the buildroot config. For systems other than OpenWRT, you can grab the source from sourceforge, or even github.

#adding the rule to MINIUPNPD
$IPTABLES -t nat -A PREROUTING -d $EXTIP -i $EXTIF -j MINIUPNPD

Unfortunately, this rule isn’t replaced when the connection is dropped (and the router gets a new dynamic IP). OpenWRT does have a hotplug script:

[ "$ACTION" = "ifup" -a "$INTERFACE" = "wan" ] && \
/etc/init.d/miniupnpd enabled && \
/etc/init.d/miniupnpd restart

Problem here is that /etc/init.d/miniupnpd enabled returns false! It’s a function defined in /etc/rc.common that checks for /etc/rc.d/S95miniupnpd. So the solution is:

ln -s /etc/init.d/miniupnpd ./S95miniupnpd

Realistically, the ipkg should come with this file (effectively enabling miniupnpd by default). A more robust solution would be to modify the hotplug script so that it restarts the daemon without the enabled check, and rely on the restart function not to start the daemon if it wasn’t already running.

As I try to use cool crypto wherever possible, the public-key based SIG(0) from RFC2931 seemed like a nifty protocol to use. The usual tool for generating keys is dnssec-keygen, and for performing updates it’s nsupdate, both part of ISC’s BIND. I set this up and tested it from scud, where it worked fine.

However, since spade controls my PPP connection, and thus knows when my IP address changes, the logical place to be running nsupdate was on spade. Unfortunately OpenWRT’s nsupdate (packaged as bind-client) is relatively large at 439186 octects; add to this libopenssl and bind-libs, and I’d run out of room on my JFFS2 partition. If you have the room, it’ll work for TSIGs: Ignace Mouzannar describes how to do it. However, even when I got it installed, I couldn’t make it sign a SIG(0) request. Chip Rosenthal has also noticed problems, and came up with an HTTP-based solution to get another host to run nsupdate, called web-nsupdate.

My solution? Write my own client in C, that can read the “dnssec-keygen“ key files. Presenting dudders: my lightweight DNS UPDATE client. It even comes as an OpenWRT ipkg. It still needs a crypto library, but it can use libgcrypt which (even after dependencies) still weighs in lighter than libopenssl.