nova volume-detach fails silently, compute log shows libvirtError: invalid argument: no target device vdb

I'm trying to detach a cinder volume from a nova instance.

Nova volume-list shows it as in-use:

| 34b0ea26-f85c-4b62-8ebd-884b0e63e2d5 | in-use | filestore | 256  | None| 4d05ffe4-d30a-4c93-b710-c9ec80dad1c2 |

The volume is connected via iscsi:

# iscsiadm -m session
tcp: [5] 10.3.40.10:3260,1 iqn.2010-10.org.openstack:volume-34b0ea26-f85c-4b62-8ebd-884b0e63e2d5 (non-flash)

It's visible as /dev/vdb within the instance, and is mountable and read/writable.

However, after unmounting it in the instance and issuing

# nova volume-detach 4d05ffe4-d30a-4c93-b710-c9ec80dad1c2 34b0ea26-f85c-4b62-8ebd-884b0e63e2d5

it does not detach. The volume stays labeled "in-use". I get this error in the compute.log:

libvirtError: invalid argument: no target device vdb

When I run "virsh edit 4" on the compute node, the disk device for vdb is in fact missing. However, when I run "virsh dumpxml 4" it is still there!

How can I untangle this?

Virsh dumpxml vs. edit diff: http://ift.tt/1EQVnxc

Compute.log of the error: http://ift.tt/18maBx0