Image handlers (in essex)

George asks in the comments on my previous post about loop and nbd devices an interesting question about the behavior of this code on essex. I figured the question was worth bringing out into its own post so that its more visible. I’ve edited George’s question lightly so that this blog post flows reasonably.

Can you please explain the order (and conditions) in which the three methods are used? In my Essex installation, the “img_handlers” is not defined in nova.conf, so it takes the default value “loop,nbd,guestfs”. However, nova is using nbd as the chose method.

The handlers will be used in the order specified — with the caveat that loop doesn’t support Copy On Write (COW) images and will therefore be skipped if the libvirt driver is trying to create a COW image. Whether COW images are used is configured with the use_cow_images flag, which defaults to True. So, loop is being skipped because you’re probably using COW images.

My ssh keys are obtained by cloud-init, and still whenever I start a new instance I see in the nova-compute.logs this sequence of events:

qemu-nbd -c /dev/nbd15 /var/lib/nova/instances/instance-0000076d/disk
kpartx -a /dev/nbd15
mount /dev/mapper/nbd15p1 /tmp/tmpxGBdT0
umount /dev/mapper/nbd15p1
kpartx -d /dev/nbd15
qemu-nbd -d /dev/nbd15

I don’t understand why the mount of the first partition is necessary and what it happens when the partition is mounted.

This is a bit harder than the first bit of the question. What I think is happening is that there are files being injected, and that’s causing the mount. Just because the admin password isn’t being inject doesn’t mean that other things aren’t being injected still. You’d be able to tell what’s happening by grepping your logs for “Injecting .* into image” and seeing what shows up.

Some quick operational notes for users of loop and nbd devices

A quick note for OpenStack operators — if you are using loop or nbd devices to mount disk images in nova-compute, then you would be well served to have plenty of device files hanging around to reduce contention. For loop devices, that seems as simple as making more of them with MAKEDEV. With nbd, you’ll also need to increase the value of the max_nbd_devices flag to nova-compute. The latter is improved in grizzly, where we will autodetect nbd devices.