Troubleshooting Node Management Failures

Troubleshooting Node Management Failures

Where Are the Logs?

Some logs are stored in journald, but most are stored as text files in /var/log. They are only accessible by the root user.

ironic-inspector

The introspection logs (from ironic-inspector) are located in /var/log/ironic-inspector. If something fails during the introspection ramdisk run, ironic-inspector stores the ramdisk logs in /var/log/ironic-inspector/ramdisk/ as gz-compressed tar files. File names contain date, time and IPMI address of the node if it was detected (only for bare metal).

To collect introspection logs on success as well, set always_store_ramdisk_logs = true in /etc/ironic-inspector/inspector.conf, restart the openstack-ironic-inspector service and retry the introspection.

ironic

The deployment logs (from ironic) are located in /var/log/ironic. If something goes wrong during deployment or cleaning, the ramdisk logs are stored in /var/log/ironic/deploy. See ironic logs retrieving documentation for more details.

Node Registration and Management Problems

Nodes in enroll state after registration

If you see your nodes staying in the enroll provision state after the registration process (which may hang due to this), it means that Ironic is unable to verify power management credentials, and you need to fix them. Check the pm_addr, pm_user and pm_password fields in your instackenv.json. In some cases (e.g. when using VirtualBMC) you also need a correct pm_port. Update the node as explained in Fixing invalid node information.

Fixing invalid node information

Any problems with node data registered into Ironic can be fixed using the Ironic CLI.

For example, a wrong MAC can be fixed in two steps:

  • Find out the assigned port UUID by running

    openstack baremetal port list --node <NODE UUID>
    
  • Update the MAC address by running

    openstack baremetal port set --address <NEW MAC> <PORT UUID>
    

A Wrong IPMI address can be fixed with the following command:

openstack baremetal node set <NODE UUID> --driver-info ipmi_address=<NEW IPMI ADDRESS>

Node power state is not enforced by Ironic

By default Ironic will not forcibly sync the power state of the nodes, because in our HA (high availability) model Pacemaker is the one responsible for controlling the power state of the nodes when fencing. If you are using a non-HA setup and want Ironic to take care of the power state of the nodes please change the value of the force_power_state_during_sync configuration option in the /etc/ironic/ironic.conf file to True and restart the openstack-ironic-conductor service.

Also, note that if openstack undercloud install is re-run the value of the force_power_state_during_sync configuration option will be set back to the default, which is False.

How do I repair broken nodes

Usually, the nodes should only be deleted when the hardware is decommissioned. Before that, you’re expected to remove instances from them using scale-down. However, in some cases, it may be impossible to repair a node with e.g. broken power management, and it gets stuck in an abnormal state.

Warning

Before proceeding with this section, always try to decommission a node normally, by scaling down your cloud. Forcing node deletion may cause unpredicable results.

Ironic requires that nodes that cannot be operated normally are put in the maintenance mode. It is done by the following command:

openstack baremetal node maintenance set <NODE UUID> --reason "<EXPLANATION>"

Ironic will stop checking power and health state for such nodes, and Nova will not pick them for deployment. Power command will still work on them, though.

After a node is in the maintenance mode, you can attempt repairing it, e.g. by Fixing invalid node information. If you manage to make the node operational again, move it out of the maintenance mode:

openstack baremetal node maintenance unset <NODE UUID>

If repairing is not possible, you can force deletion of such node:

openstack baremetal node delete <NODE UUID>

Forcing node removal will leave it powered on, accessing the network with the old IP address(es) and with all services running. Before proceeding, make sure to power it off and clean up via any means.

After that, the associated Nova instance is orphaned, and must be deleted. You can do it normally via the scale down procedure.

Hardware Introspection Problems

Introspection hangs and times out

ironic-inspector times out introspection process after some time (defaulting to 1 hour) if it never gets response from the introspection ramdisk. This can be a sign of a bug in the introspection ramdisk, but usually it happens due to environment misconfiguration, particularly BIOS boot settings. Please refer to ironic-inspector troubleshooting documentation for information on how to detect and fix such problems.

Accessing the ramdisk

Note that the introspection ramdisk is by default built with the dynamic-login element, so you can set up an SSH key and log into it for debugging.

First, think of a temporary root password. Generate a hash by feeding it into openssl passwd -1 command. Edit /httpboot/inspector.ipxe manually. Find the line starting with “kernel” and append rootpwd=”HASH” to it. Do not append the real password. Alternatively, you can append sshkey=”PUBLIC_SSH_KEY” with your public SSH key.

Warning

In both cases quotation marks are required!

When ramdisk is running, figure out its IP address by checking arp utility or DHCP logs from

sudo journalctl -u openstack-ironic-inspector-dnsmasq

SSH as a root user with the temporary password or the SSH key.

Note

Some operating systems, such as RHEL and CentOS, require SELinux to be in permissive or disabled mode so that you can log in to the image. This is achieved by building the image with the selinux-permissive element for diskimage-builder or by passing selinux=0 in the kernel command line.

Refusing to introspect node with provision state “available”

If you’re running introspection directly using ironic-inspector CLI (or in case of bugs in our scripts), a node can be in the “AVAILABLE” state, which is meant for deployment, not for introspection. You should advance node to the “MANAGEABLE” state before introspection and move it back before deployment. Please refer to upstream node states documentation for information on how to fix it.

How can introspection be stopped?

Introspection for a node can be stopped with the following command:

openstack baremetal introspection abort <NODE UUID>
Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.