MAAS/curtin generate invalid /e/n/i and failed deployment for nodes with long (biosdevname) interface names, which in turn have VLANs

[Dimiter Naydenov]

I have MAAS 2.0.0 (beta4+bzr4944) and an Intel NUC node with 2 NICs (second one usb3-to-ethernet adapter), commissioned with xenial: http://paste.ubuntu.com/16052205/.

If I try to deploy this, it never completes, but since the console gets cleared soon after the initial boot messages, it took me some time to figure out what's wrong. It turned out the /e/n/i maas (or curtin) generates for the node (using this result from 'get-curtin-config <id>': http://paste.ubuntu.com/16052334/) contains stanzas like 'iface enx00e1000015de.200 inet manual'. That name is too long for ifup (or ip link) to handle, as it exceeds the Linux device name length limit (16 characters).

It's relatively easy to workaround this, by renaming the second NIC from 'enx00e1000015de' to e.g. 'eno2' after commissioning. MAAS automatically renames all its children (VLAN) NICs correctly, and the deployment completes successfully, as expected.

[Mike Pontillo (mpontillo)]

This was discussed on IRC, and one potential workaround is to set "net.ifnames=0" during commissioning, so that the interfaces get the names assigned by the kernel.

Otherwise, the only solution I can think of is for MAAS to shorten the names to 10 characters post-commissioning, and eliminate any duplicates.

[Andres Rodriguez (andreserl)]

This is not an issue with MAAS. See:

https://bugs.launchpad.net/ubuntu/+source/systemd/+bug/1567744

[Dimiter Naydenov (dimitern)]

Thanks for the link Andres! I've added a comment about how that affects Juju and the bridges we create.

There are ways to to get shorter names - net.ifnames=0 is one option, comissioning with trusty by default also works. As far as Juju is concerned, we can detect such cases and use the MAAS API to rename long interfaces between node acquisition and deployment.

It's a real pity in 2016 the Linux kernel still has such ridiculous limits on network device names.