Well, I’ve spend the better part of the last week or so debugging a problem with out new Xen VMs (on one host) not getting DHCP leases. Since our DHCP servers (ISC DHCPd 3.0.6) are using Brian Masney’s LDAP patch (which has recently been included in mainline DHCPd), I assumed that this might be the source of some of the problems.
The Xen client booting process uses Etherboot
(specifically 5.4.2 on my machines) to get DHCP and PXE for the client.
The Etherboot ROM is generated online with
ROM-o-Matic (by the Xen devs) and is then
hvmloader (though I would not find this out until much
later in my investigation, given the piss-poor documentation about it).
So, aside from pulling the SRPM for Xen and looking around, I decided I
couldn’t really debug much on the client end (I wasn’t getting any
BIOS-like messages from Xen, so the client end seemed to be pretty much
a black box).
After a week of re-examining our DHCP servers, moving the client between pieces of hardware and different networks, tracing out the DHCP logs, and debugging everything in-between, I finally resorted to doing packet captures and analyzing them. Last Thursday (my last work day of the week), I popped my laptop on the same network, set it up in DHCP, and did some captures of the laptop (which worked correctly) and the problem VM.
Sometime around 10:00 PM, long after I’d gotten home, I had opened captures of both the good and bad hosts in separate Wireshark windows, and was going through them line-by-line. I’d also written a nifty little Perl script (my weakest language) using Net::DHCP::Packet to craft packets identical to the ones from the working and FUBAR hosts, and inject them into the network. The only thing I could find different was the value of the “secs” field of the DHCPDISCOVER packets (octets 9 and 10), which are supposed to contain the number of seconds that have passed since the host started booting (RFC 1541). My laptop (the working host) started getting replies from the server at 21 seconds. I took my Perl packet-injecting script, and started adjusting the “secs” values of both the working and bad packets. Sure enough, with identical packets from each host, the values converged. Anything with “secs” below 2 got no response from the DHCP server, anything with 2 or greater got a correct lease.
Then it hit me. When we used to have a primary/secondary DHCP server
setup (with manual failover), we’d configured the secondary server with
min-secs: 2, instructing it to not give out any leases to clients with
a “secs” value of under 2, to prevent the secondary server from
answering if, for some reason, the primary was still online. Bingo.
At this point, I’m waiting to have a meeting of the binds before I add a network-level override of “min-secs: 0” for the networks in question. But I’m relatively confident that everything will go smoothly from there on.
This experience highlighted that one small “bug” can confuse 3 people for the better part of a week:
Etherboot (at least 5.4.2) doesn’t increment the “secs” field in its DISCOVER packets as RFC 1541 suggests. Therefore anyone with “min-secs” in their dhcpd configuration won’t ever give out a lease.
In hindsight, it really was a brain-dead moment. There was a little note in the dhcpd logs that I, and the others, totally overlooked: “DHCPDISCOVER (…) 0 secs < 2”. I guess I should have turned my brain on and realized that those few characters were probably important, and there for a reason, no matter how unassuming (and un-error-like) they may be…