Troubleshooting Ironic

Nova returns “No valid host was found” Error

Sometimes Nova Conductor log file “nova-conductor.log” or a message returned from Nova API contains the following error:

NoValidHost: No valid host was found. There are not enough hosts available.

“No valid host was found” means that the Nova Scheduler could not find a bare metal node suitable for booting the new instance.

This in turn usually means some mismatch between resources that Nova expects to find and resources that Ironic advertised to Nova.

A few things should be checked in this case:

1. Make sure that enough nodes are in available state, not in maintenance mode and not already used by an existing instance. Check with the following command:

   baremetal node list --provision-state available --no-maintenance --unassociated
   

   If this command does not show enough nodes, use generic baremetal node list to check other nodes. For example, nodes in manageable state should be made available:

   baremetal node provide <IRONIC NODE>
   

   The Bare metal service automatically puts a node in maintenance mode if there are issues with accessing its management interface. See Power fault and recovery for details.

   The node validate command can be used to verify that all required fields are present. The following command should not return anything:

   baremetal node validate <IRONIC NODE> | grep -E '(power|management)\W*False'
   

   Maintenance mode will be also set on a node if automated cleaning has failed for it previously.

2. Make sure that you have Compute services running and enabled:

   $ openstack compute service list --service nova-compute
   +----+--------------+-------------+------+---------+-------+----------------------------+
   | ID | Binary       | Host        | Zone | Status  | State | Updated At                 |
   +----+--------------+-------------+------+---------+-------+----------------------------+
   |  7 | nova-compute | example.com | nova | enabled | up    | 2017-09-04T13:14:03.000000 |
   +----+--------------+-------------+------+---------+-------+----------------------------+
   

   By default, a Compute service is disabled after 10 consecutive build failures on it. This is to ensure that new build requests are not routed to a broken Compute service. If it is the case, make sure to fix the source of the failures, then re-enable it:

   openstack compute service set --enable <COMPUTE HOST> nova-compute
   

3. Starting with the Pike release, check that all your nodes have the resource_class field set using the following command:

   baremetal node list --fields uuid name resource_class
   

   Then check that the flavor(s) are configured to request these resource classes via their properties:

   openstack flavor show <FLAVOR NAME> -f value -c properties
   

   For example, if your node has resource class baremetal-large, it will be matched by a flavor with property resources:CUSTOM_BAREMETAL_LARGE set to 1. See Create flavors for use with the Bare Metal service for more details on the correct configuration.

4. Upon scheduling, Nova will query the Placement API service for the available resource providers (in the case of Ironic: nodes with a given resource class). If placement does not have any allocation candidates for the requested resource class, the request will result in a “No valid host was found” error. It is hence sensible to check if Placement is aware of resource providers (nodes) for the requested resource class with:

   $ openstack allocation candidate list --resource CUSTOM_BAREMETAL_LARGE='1'
   +---+-----------------------------+--------------------------------------+-------------------------------+
   | # | allocation                  | resource provider                    | inventory used/capacity       |
   +---+-----------------------------+--------------------------------------+-------------------------------+
   | 1 | CUSTOM_BAREMETAL_LARGE=1    | 2f7b9c69-c1df-4e40-b94e-5821a4ea0453 | CUSTOM_BAREMETAL_LARGE=0/1    |
   +---+-----------------------------+--------------------------------------+-------------------------------+
   

   For Ironic, the resource provider is the UUID of the available Ironic node. If this command returns an empty list (or does not contain the targeted resource provider), the operator needs to understand first, why the resource tracker has not reported this provider to placement. Potential explanations include:

   • the resource tracker cycle has not finished yet and the resource provider will appear once it has (the time to finish the cycle scales linearly with the number of nodes the corresponding nova-compute service manages);

   • the node is in a state where the resource tracker does not consider it to be eligible for scheduling, e.g. when the node has maintenance set to True; make sure the target nodes are in available and maintenance is False;

5. The Nova flavor that you are using does not match any properties of the available Ironic nodes. Use

   openstack flavor show <FLAVOR NAME>
   

   to compare. The extra specs in your flavor starting with capability: should match ones in node.properties['capabilities'].

   Note

   The format of capabilities is different in Nova and Ironic. E.g. in Nova flavor:

   $ openstack flavor show <FLAVOR NAME> -c properties
   +------------+----------------------------------+
   | Field      | Value                            |
   +------------+----------------------------------+
   | properties | capabilities:boot_mode='uefi'    |
   +------------+----------------------------------+
   

   But in Ironic node:

   $ baremetal node show <IRONIC NODE> --fields properties
   +------------+-----------------------------------------+
   | Property   | Value                                   |
   +------------+-----------------------------------------+
   | properties | {u'capabilities': u'boot_mode:uefi'}    |
   +------------+-----------------------------------------+
   

6. After making changes to nodes in Ironic, it takes time for those changes to propagate from Ironic to Nova. Check that

   openstack hypervisor stats show
   

   correctly shows total amount of resources in your system. You can also check openstack hypervisor show <IRONIC NODE> to see the status of individual Ironic nodes as reported to Nova.

7. Figure out which Nova Scheduler filter ruled out your nodes. Check the nova-scheduler logs for lines containing something like:

   Filter ComputeCapabilitiesFilter returned 0 hosts
   

   The name of the filter that removed the last hosts may give some hints on what exactly was not matched. See Nova filters documentation for more details.

8. If none of the above helped, check Ironic conductor log carefully to see if there are any conductor-related errors which are the root cause for “No valid host was found”. If there are any “Error in deploy of node <IRONIC-NODE-UUID>: [Errno 28] …” error messages in Ironic conductor log, it means the conductor run into a special error during deployment. So you can check the log carefully to fix or work around and then try again.

Patching the Deploy Ramdisk

When debugging a problem with deployment and/or inspection you may want to quickly apply a change to the ramdisk to see if it helps. Of course you can inject your code and/or SSH keys during the ramdisk build (depends on how exactly you’ve built your ramdisk). But it’s also possible to quickly modify an already built ramdisk.

Create an empty directory and unpack the ramdisk content there:

$ mkdir unpack
$ cd unpack
$ gzip -dc /path/to/the/ramdisk | cpio -id

The last command will result in the whole Linux file system tree unpacked in the current directory. Now you can modify any files you want. The actual location of the files will depend on the way you’ve built the ramdisk.

Note

On a systemd-based system you can use the systemd-nspawn tool (from the systemd-container package) to create a lightweight container from the unpacked filesystem tree:

$ sudo systemd-nspawn --directory /path/to/unpacked/ramdisk/ /bin/bash

This will allow you to run commands within the filesystem, e.g. use package manager. If the ramdisk is also systemd-based, and you have login credentials set up, you can even boot a real ramdisk environment with

$ sudo systemd-nspawn --directory /path/to/unpacked/ramdisk/ --boot

After you’ve done the modifications, pack the whole content of the current directory back:

$ find . | cpio -H newc -o | gzip -c > /path/to/the/new/ramdisk

Note

You don’t need to modify the kernel (e.g. tinyipa-master.vmlinuz), only the ramdisk part.

API Errors

The debug_tracebacks_in_api config option may be set to return tracebacks in the API response for all 4xx and 5xx errors.

Retrieving logs from the deploy ramdisk

When troubleshooting deployments (specially in case of a deploy failure) it’s important to have access to the deploy ramdisk logs to be able to identify the source of the problem. By default, Ironic will retrieve the logs from the deploy ramdisk when the deployment fails and save it on the local filesystem at /var/log/ironic/deploy.

To change this behavior, operators can make the following changes to /etc/ironic/ironic.conf under the [agent] group:

• deploy_logs_collect: Whether Ironic should collect the deployment logs on deployment. Valid values for this option are:

  • on_failure (default): Retrieve the deployment logs upon a deployment failure.

  • always: Always retrieve the deployment logs, even if the deployment succeed.

  • never: Disable retrieving the deployment logs.

• deploy_logs_storage_backend: The name of the storage backend where the logs will be stored. Valid values for this option are:

  • local (default): Store the logs in the local filesystem.

  • swift: Store the logs in Swift.

• deploy_logs_local_path: The path to the directory where the logs should be stored, used when the deploy_logs_storage_backend is configured to local. By default logs will be stored at /var/log/ironic/deploy.

• deploy_logs_swift_container: The name of the Swift container to store the logs, used when the deploy_logs_storage_backend is configured to “swift”. By default ironic_deploy_logs_container.

• deploy_logs_swift_days_to_expire: Number of days before a log object is marked as expired in Swift. If None, the logs will be kept forever or until manually deleted. Used when the deploy_logs_storage_backend is configured to “swift”. By default 30 days.

When the logs are collected, Ironic will store a tar.gz file containing all the logs according to the deploy_logs_storage_backend configuration option. All log objects will be named with the following pattern:

<node>[_<instance-uuid>]_<timestamp yyyy-mm-dd-hh:mm:ss>.tar.gz

Note

The instance_uuid field is not required for deploying a node when Ironic is configured to be used in standalone mode. If present it will be appended to the name.

Accessing the log data

When storing in the local filesystem

When storing the logs in the local filesystem, the log files can be found at the path configured in the deploy_logs_local_path configuration option. For example, to find the logs from the node 5e9258c4-cfda-40b6-86e2-e192f523d668:

$ ls /var/log/ironic/deploy | grep 5e9258c4-cfda-40b6-86e2-e192f523d668
5e9258c4-cfda-40b6-86e2-e192f523d668_88595d8a-6725-4471-8cd5-c0f3106b6898_2016-08-08-13:52:12.tar.gz
5e9258c4-cfda-40b6-86e2-e192f523d668_db87f2c5-7a9a-48c2-9a76-604287257c1b_2016-08-08-14:07:25.tar.gz

Note

When saving the logs to the filesystem, operators may want to enable some form of rotation for the logs to avoid disk space problems.

When storing in Swift

When using Swift, operators can associate the objects in the container with the nodes in Ironic and search for the logs for the node 5e9258c4-cfda-40b6-86e2-e192f523d668 using the prefix parameter. For example:

$ swift list ironic_deploy_logs_container -p 5e9258c4-cfda-40b6-86e2-e192f523d668
5e9258c4-cfda-40b6-86e2-e192f523d668_88595d8a-6725-4471-8cd5-c0f3106b6898_2016-08-08-13:52:12.tar.gz
5e9258c4-cfda-40b6-86e2-e192f523d668_db87f2c5-7a9a-48c2-9a76-604287257c1b_2016-08-08-14:07:25.tar.gz

To download a specific log from Swift, do:

$ swift download ironic_deploy_logs_container "5e9258c4-cfda-40b6-86e2-e192f523d668_db87f2c5-7a9a-48c2-9a76-604287257c1b_2016-08-08-14:07:25.tar.gz"
5e9258c4-cfda-40b6-86e2-e192f523d668_db87f2c5-7a9a-48c2-9a76-604287257c1b_2016-08-08-14:07:25.tar.gz [auth 0.341s, headers 0.391s, total 0.391s, 0.531 MB/s]

The contents of the log file

The log is just a .tar.gz file that can be extracted as:

$ tar xvf <file path>

The contents of the file may differ slightly depending on the distribution that the deploy ramdisk is using:

• For distributions using systemd there will be a file called journal which contains all the system logs collected via the journalctl command.

• For other distributions, the ramdisk will collect all the contents of the /var/log directory.

For all distributions, the log file will also contain the output of the following commands (if present): ps, df, ip addr and iptables.

Here’s one example when extracting the content of a log file for a distribution that uses systemd:

$ tar xvf 5e9258c4-cfda-40b6-86e2-e192f523d668_88595d8a-6725-4471-8cd5-c0f3106b6898_2016-08-08-13:52:12.tar.gz
df
ps
journal
ip_addr
iptables

DHCP during PXE or iPXE is inconsistent or unreliable

This can be caused by the spanning tree protocol delay on some switches. The delay prevents the switch port moving to forwarding mode during the nodes attempts to PXE, so the packets never make it to the DHCP server. To resolve this issue you should set the switch port that connects to your baremetal nodes as an edge or PortFast type port. Configured in this way the switch port will move to forwarding mode as soon as the link is established. An example on how to do that for a Cisco Nexus switch is:

$ config terminal
$ (config) interface eth1/11
$ (config-if) spanning-tree port type edge

Why does X issue occur when I am using LACP bonding with iPXE?

If you are using iPXE, an unfortunate aspect of its design and interaction with networking is an automatic response as a Link Aggregation Control Protocol (or LACP) peer to remote switches. iPXE does this for only the single port which is used for network booting.

In theory, this may help establish the port link-state faster with some switch vendors, but the official reasoning as far as the Ironic Developers are aware is not documented for iPXE. The end result of this is that once iPXE has stopped responding to LACP messages from the peer port, which occurs as part of the process of booting a ramdisk and iPXE handing over control to a full operating-system, switches typically begin a timer to determine how to handle the failure. This is because, depending on the mode of LACP, this can be interpreted as a switch or network fabric failure.

This may demonstrate as any number of behaviors or issues from ramdisks finding they are unable to acquire DHCP addresses over the network interface to downloads abruptly stalling, to even minor issues such as LLDP port data being unavailable in introspection.

Overall:

• Ironic’s agent doesn’t officially support LACP and the Ironic community generally believes this may cause more problems than it would solve. During the Victoria development cycle, we added retry logic for most actions in an attempt to navigate the worst-known default hold-down timers to help ensure a deployment does not fail due to a short-lived transitory network connectivity failure in the form of a switch port having moved to a temporary blocking state. Where applicable and possible, many of these patches have been backported to supported releases. These patches also require that the switchport has an eventual fallback to a non-bonded mode. If the port remains in a blocking state, then traffic will be unable to flow and the deployment is likely to time out.

• If you must use LACP, consider passive LACP negotiation settings in the network switch as opposed to active. The difference being with passive the connected workload is likely a server where it should likely request the switch to establish the Link Aggregate. This is instead of being treated as if it’s possibly another switch.

• Consult your switch vendor’s support forums. Some vendors have recommended port settings for booting machines using iPXE with their switches.

IPMI errors

When working with IPMI, several settings need to be enabled depending on vendors.

Enable IPMI over LAN

Machines may not have IPMI access over LAN enabled by default. This could cause the IPMI port to be unreachable through ipmitool, as shown:

$ ipmitool -I lan -H ipmi_host -U ipmi_user -P ipmi_pass chassis power status
Error: Unable to establish LAN session

To fix this, enable IPMI over lan setting using your BMC tool or web app.

Troubleshooting lanplus interface

When working with lanplus interfaces, you may encounter the following error:

$ ipmitool -I lanplus -H ipmi_host -U ipmi_user -P ipmi_pass power status
Error in open session response message : insufficient resources for session
Error: Unable to establish IPMI v2 / RMCP+ session

To fix that issue, please enable RMCP+ Cipher Suite3 Configuration setting using your BMC tool or web app.

Why are my nodes stuck in a “-ing” state?

The Ironic conductor uses states ending with ing as a signifier that the conductor is actively working on something related to the node.

Often, this means there is an internal lock or reservation set on the node and the conductor is downloading, uploading, or attempting to perform some sort of Input/Output operation - see Why does API return “Node is locked by host”? for details.

In the case the conductor gets stuck, these operations should timeout, but there are cases in operating systems where operations are blocked until completion. These sorts of operations can vary based on the specific environment and operating configuration.

What can cause these sorts of failures?

Typical causes of such failures are going to be largely rooted in the concept of iowait, either in the form of downloading from a remote host or reading or writing to the disk of the conductor. An operator can use the iostat tool to identify the percentage of CPU time spent waiting on storage devices.

The fields that will be particularly important are the iowait, await, and tps ones, which can be read about in the iostat manual page.

In the case of network file systems, for backing components such as image caches or distributed tftpboot or httpboot folders, IO operations failing on these can, depending on operating system and underlying client settings, cause threads to be stuck in a blocking wait state, which is realistically undetectable short the operating system logging connectivity errors or even lock manager access errors.

For example with nfs, the underlying client recovery behavior, in terms of soft, hard, softreval, nosoftreval, will largely impact this behavior, but also NFS server settings can impact this behavior. A solid sign that this is a failure, is when an ls /path/to/nfs command hangs for a period of time. In such cases, the Storage Administrator should be consulted and network connectivity investigated for errors before trying to recover to proceed.

The bad news for IO related failures

If the node has a populated reservation field, and has not timed out or proceeded to a fail state, then the conductor process will likely need to be restarted. This is because the worker thread is hung with-in the conductor.

Manual intervention with-in Ironic’s database is not advised to try and “un-wedge” the machine in this state, and restarting the conductor is encouraged.

Note

Ironic’s conductor, upon restart, clears reservations for nodes which were previously managed by the conductor before restart.

If a distributed or network file system is in use, it is highly recommended that the operating system of the node running the conductor be rebooted as the running conductor may not even be able to exit in the state of an IO failure, again dependent upon site and server configuration.

File Size != Disk Size

An easy to make misconception is that a 2.4 GB file means that only 2.4 GB is written to disk. But if that file’s virtual size is 20 GB, or 100 GB things can become very problematic and extend the amount of time the node spends in deploying and deploy wait states.

Again, these sorts of cases will depend upon the exact configuration of the deployment, but hopefully these are areas where these actions can occur.

• Conversion to raw image files upon download to the conductor, from the DEFAULT.force_raw_images option. Users using Glance may also experience issues here as the conductor will cache the image to be written which takes place when the agent.image_download_source is set to http instead of swift.

Note

The QCOW2 image conversion utility does consume quite a bit of memory when converting images or writing them to the end storage device. This is because the files are not sequential in nature, and must be re-assembled from an internal block mapping. Internally Ironic limits this to 1GB of RAM. Operators performing large numbers of deployments may wish to disable raw images in these sorts of cases in order to minimize the conductor becoming a limiting factor due to memory and network IO.

Why are my nodes stuck in a “wait” state?

The Ironic conductor uses states containing wait as a signifier that the conductor is waiting for a callback from another component, such as the Ironic Python Agent or the Inspector. If this feedback does not arrive, the conductor will time out and the node will eventually move to a failed state. Depending on the configuration and the circumstances, however, a node can stay in a wait state for a long time or even never time out. The list of such wait states includes:

• clean wait for cleaning,

• inspect wait for introspection,

• rescue wait for rescuing, and

• wait call-back for deploying.

Communication issues between the conductor and the node

One of the most common issues when nodes seem to be stuck in a wait state occur when the node never received any instructions or does not react as expected: the conductor moved the node to a wait state but the node will never call back. Examples include wrong ciphers which will make ipmitool get stuck or BMCs in a state where they accept commands, but don’t do the requested task (or only a part of it, like shutting off, but not starting). It is useful in these cases to see via a ping or the console if and which action the node is performing. If the node does not seem to react to the requests sent be the conductor, it may be worthwhile to try the corresponding action out-of-band, e.g. confirm that power on/off commands work when directly sent to the BMC. The section on IPMI errors. above gives some additional points to check. In some situations, a BMC reset may be necessary.

Ironic Python Agent stuck

Nodes can also get remain in a wait state when the component the conductor is waiting for gets stuck, e.g. when a hardware manager enters a loop or is waiting for an event that is never happening. In these cases, it might be helpful to connect to the IPA and inspect its logs, see the trouble shooting guide of the ironic-python-agent (IPA) on how to do this.

Stopping the operation

Cleaning, inspection and rescuing can be stopped while in clean wait, inspect wait and rescue wait states using the abort command. It will move the node to the corresponding failure state (clean failed, inspect failed or rescue failed):

baremetal node abort <node>

Deploying can be aborted while in the wait call-back state by starting an undeploy (normally resulting in cleaning):

baremetal node undeploy <node>

See Bare Metal State Machine for more details.

Note

Since the Bare Metal service is not doing anything actively in waiting states, the nodes are not moved to failed states on conductor restart.

Deployments fail with “failed to update MAC address”

The design of the integration with the Networking service (neutron) is such that once virtual ports have been created in the API, their MAC address must be updated in order for the DHCP server to be able to appropriately reply.

This can sometimes result in errors being raised indicating that the MAC address is already in use. This is because at some point in the past, a virtual interface was orphaned either by accident or by some unexpected glitch, and a previous entry is still present in Neutron.

This error looks something like this when reported in the ironic-conductor log output.:

Failed to update MAC address on Neutron port 305beda7-0dd0-4fec-b4d2-78b7aa4e8e6a.: MacAddressInUseClient: Unable to complete operation for network 1e252627-6223-4076-a2b9-6f56493c9bac. The mac address 52:54:00:7c:c4:56 is in use.

Because we have no idea about this entry, we fail the deployment process as we can’t make a number of assumptions in order to attempt to automatically resolve the conflict.

How did I get here?

Originally this was a fairly easy issue to encounter. The retry logic path which resulted between the Orchestration (heat) and Compute (nova) services, could sometimes result in additional un-necessary ports being created.

Bugs of this class have been largely resolved since the Rocky development cycle. Since then, the way this can become encountered is due to Networking (neutron) VIF attachments not being removed or deleted prior to deleting a port in the Bare Metal service.

Ultimately, the key of this is that the port is being deleted. Under most operating circumstances, there really is no need to delete the port, and VIF attachments are stored on the port object, so deleting the port CAN result in the VIF not being cleaned up from Neutron.

Under normal circumstances, when deleting ports, a node should be in a stable state, and the node should not be provisioned. If the baremetal port delete command fails, this may indicate that a known VIF is still attached. Generally if they are transitory from cleaning, provisioning, rescuing, or even inspection, getting the node to the available state will unblock your delete operation, that is unless there is a tenant VIF attahment. In that case, the vif will need to be removed from with-in the Bare Metal service using the baremetal node vif detach command.

A port can also be checked to see if there is a VIF attachment by consulting the port’s internal_info field.

Warning

The maintenance flag can be used to force the node’s port to be deleted, however this will disable any check that would normally block the user from issuing a delete and accidentally orphaning the VIF attachment record.

How do I resolve this?

Generally, you need to identify the port with the offending MAC address. Example:

$ openstack port list --mac-address 52:54:00:7c:c4:56

From the command’s output, you should be able to identify the id field. Using that, you can delete the port. Example:

$ openstack port delete <id>

Warning

Before deleting a port, you should always verify that it is no longer in use or no longer seems applicable/operable. If multiple deployments of the Bare Metal service with a single Neutron, the possibility that a inventory typo, or possibly even a duplicate MAC address exists, which could also produce the same basic error message.

My test VM image does not deploy – mount point does not exist

What is likely occurring

The image attempting to be deployed likely is a partition image where the file system that the user wishes to boot from lacks the required folders, such as /dev and /proc, which are required to install a bootloader for a Linux OS image

It should be noted that similar errors can also occur with whole disk images where we are attempting to setup the UEFI bootloader configuration. That being said, in this case, the image is likely invalid or contains an unexpected internal structure.

Users performing testing may choose something that they believe will work based on it working for virtual machines. These images are often attractive for testing as they are generic and include basic support for establishing networking and possibly installing user keys. Unfortunately, these images often lack drivers and firmware required for many different types of physical hardware which makes using them very problematic. Additionally, images such as Cirros do not have any contents in the root filesystem (i.e. an empty filesystem), as they are designed for the ramdisk to write the contents to disk upon boot.

How do I not encounter this issue?

We generally recommend using diskimage-builder or vendor supplied images. Centos, Ubuntu, Fedora, and Debian all publish operating system images which do generally include drivers and firmware for physical hardware. Many of these published “cloud” images, also support auto-configuration of networking AND population of user keys.

Issues with autoconfigured TLS

These issues will manifest as an error in ironic-conductor logs looking similar to (lines are wrapped for readability):

ERROR ironic.drivers.modules.agent_client [-]
Failed to connect to the agent running on node d7c322f0-0354-4008-92b4-f49fb2201001
for invoking command clean.get_clean_steps. Error:
HTTPSConnectionPool(host='192.168.123.126', port=9999): Max retries exceeded with url:
/v1/commands/?wait=true&agent_token=<token> (Caused by
SSLError(SSLError(1, '[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:897)'),)):
requests.exceptions.SSLError: HTTPSConnectionPool(host='192.168.123.126', port=9999):
Max retries exceeded with url: /v1/commands/?wait=true&agent_token=<token>
(Caused by SSLError(SSLError(1, '[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:897)'),))

The cause of the issue is that the Bare Metal service cannot access the ramdisk with the TLS certificate provided by the ramdisk on first heartbeat. You can inspect the stored certificate in /var/lib/ironic/certificates/<node>.crt.

You can try connecting to the ramdisk using the IP address in the log message:

curl -vL https://<IP address>:9999/v1/commands \
    --cacert /var/lib/ironic/certificates/<node UUID>.crt

You can get the detailed information about the certificate using openSSL:

openssl x509 -text -noout -in /var/lib/ironic/certificates/<node UUID>.crt

Clock skew

One possible source of the problem is a discrepancy between the hardware clock on the node and the time on the machine with the Bare Metal service. It can be detected by comparing the Not Before field in the openssl output with the timestamp of a log message.

The recommended solution is to enable the NTP support in ironic-python-agent by passing the ipa-ntp-server argument with an address of an NTP server reachable by the node.

If it is not possible, you need to ensure the correct hardware time on the machine. Keep in mind a potential issue with timezones: an ability to store timezone in hardware is pretty recent and may not be available. Since ironic-python-agent is likely operating in UTC, the hardware clock should also be set in UTC.

Note

Microsoft Windows uses local time by default, so a machine that has previously run Windows will likely have wrong time.

I changed ironic.conf, and now I can’t edit my nodes.

Whenever a node is created in ironic, default interfaces are identified as part of driver composition. This maybe sourced from explicit default values which have been set in ironic.conf or by the interface order for the enabled interfaces list. The result of this is that the ironic-conductor cannot spawn a task using the composed driver, as a portion of the driver is no longer enabled. This makes it difficult to edit or update the node if the settings have been changed.

For example, with networking interfaces, if you have default_network_interface=neutron or enabled_network_interfaces=neutron,flat in your ironic.conf, nodes would have been created with the neutron network interface.

This is because default_network_interface overrides the setting for new nodes, and that setting is saved to the database nodes table.

Similarly, the order of enabled_network_interfaces takes priority, and the first entry in the list is generally set to the default for the node upon creation, and that record is saved to the database nodes table.

The only case where driver composition does not calculate a default is if an explicit value is provided upon the creation of the node.

Example failure

A node in this state, when the network_interface was saved as neutron, yet the neutron interface is no longer enabled will fail basic state transition requests:

$ baremetal node manage 7164efca-37ab-1213-1112-b731cf795a5a
Could not find the following interface in the 'ironic.hardware.interfaces.network' entrypoint: neutron. Valid interfaces are ['flat']. (HTTP 400)

How to fix this?

Revert the changes you made to ironic.conf.

This applies to any changes to any default_*_interface options or the order of interfaces in the for the enabled_*_interfaces options.

Once the conductor has been restarted with the updated configuration, you should now be able to update the interface using the baremetal node set command. In this example we use the network_interface as this is most commonly where it is encountered:

$ baremetal node set $NAME_OR_UUID --network-interface flat

Note

There are additional paths one can take to remedy this sort of issue, however we encourage operators to be mindful of operational consistency when making major configuration changes.

Once you have updated the saved interfaces, you should be able to safely return the ironic.conf configuration change in changing what interfaces are enabled by the conductor.

I’m getting Out of Memory errors

This issue, also known as the “the OOMKiller got my conductor” case, is where your OS system memory reaches a point where the operating system engages measures to shed active memory consumption in order to prevent a complete failure of the machine. Unfortunately this can cause unpredictable behavior.

How did I get here?

One of the major consumers of memory in a host running an ironic-conductor is transformation of disk images using the qemu-img tool. This tool, because the disk images it works with are both compressed and out of linear block order, requires a considerable amount of memory to efficiently re-assemble and write-out a disk to a device, or to simply convert the format such as to a raw image.

By default, ironic’s configuration limits this conversion to 1 GB of RAM for the process, but each conversion does cause additional buffer memory to be used, which increases overall system memory pressure. Generally memory pressure alone from buffers will not cause an out of memory condition, but the multiple conversions or deployments running at the same time CAN cause extreme memory pressure and risk the system running out of memory.

How do I resolve this?

This can be addressed a few different ways:

• Use raw images, however these images can be substantially larger and require more data to be transmitted “over the wire”.

• Add more physical memory.

• Add swap space.

• Reduce concurrency, possibly via another conductor or changing the nova-compute.conf max_concurrent_builds parameter.

• Or finally, adjust the DEFAULT.minimum_required_memory parameter in your ironic.conf file. The default should be considered a “default of last resort” and you may need to reserve additional memory. You may also wish to adjust the DEFAULT.minimum_memory_wait_retries and DEFAULT.minimum_memory_wait_time parameters.

Why does API return “Node is locked by host”?

This error usually manifests as HTTP error 409 on the client side:

Node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 is locked by host 192.168.122.1, please retry after the current operation is completed.

It happens, because an operation that modifies a node is requested, while another such operation is running. The conflicting operation may be user requested (e.g. a provisioning action) or related to the internal processes (e.g. changing power state during Power Synchronization). The reported host name corresponds to the conductor instance that holds the lock.

Normally, these errors are transient and safe to retry after a few seconds. If the lock is held for significant time, these are the steps you can take.

First of all, check the current provision_state of the node:

verifying

means that the conductor is trying to access the node’s BMC. If it happens for minutes, it means that the BMC is either unreachable or misbehaving. Double-check the information in driver_info, especially the BMC address and credentials.

If the access details seem correct, try resetting the BMC using, for example, its web UI.

deploying/inspecting/cleaning

means that the conductor is doing some active work. It may include downloading or converting images, executing synchronous out-of-band deploy or clean steps, etc. A node can stay in this state for minutes, depending on various factors. Consult the conductor logs.

available/manageable/wait call-back/clean wait

means that some background process is holding the lock. Most commonly it’s the power synchronization loop. Similarly to the verifying state, it may mean that the BMC access is broken or too slow. The conductor logs will provide you insights on what is happening.

To trace the process using conductor logs:

1. Isolate the relevant log parts. Lock messages come from the ironic.conductor.task_manager module. You can also check the ironic.common.states module for any state transitions:

   $ grep -E '(ironic.conductor.task_manager|ironic.common.states|NodeLocked)' \
       conductor.log > state.log
   

2. Find the first instance of NodeLocked. It may look like this (stripping timestamps and request IDs here and below for readability):

   DEBUG ironic.conductor.task_manager [-] Attempting to get exclusive lock on node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 (for node update) __init__ /usr/lib/python3.6/site-packages/ironic/conductor/task_manager.py:233
   DEBUG ironic_lib.json_rpc.server [-] RPC error NodeLocked: Node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 is locked by host 192.168.57.53, please retry after the current operation is completed. _handle_error /usr/lib/python3.6/site-packages/ironic_lib/json_rpc/server.py:179
   

   The events right before this failure will provide you a clue on why the lock is held.

3. Find the last successful exclusive locking event before the failure, for example:

   DEBUG ironic.conductor.task_manager [-] Attempting to get exclusive lock on node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 (for provision action manage) __init__ /usr/lib/python3.6/site-packages/ironic/conductor/task_manager.py:233
   DEBUG ironic.conductor.task_manager [-] Node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 successfully reserved for provision action manage (took 0.01 seconds) reserve_node /usr/lib/python3.6/site-packages/ironic/conductor/task_manager.py:350
   DEBUG ironic.common.states [-] Exiting old state 'enroll' in response to event 'manage' on_exit /usr/lib/python3.6/site-packages/ironic/common/states.py:307
   DEBUG ironic.common.states [-] Entering new state 'verifying' in response to event 'manage' on_enter /usr/lib/python3.6/site-packages/ironic/common/states.py:313
   

   This is your root cause, the lock is held because of the BMC credentials verification.

4. Find when the lock is released (if at all). The messages look like this:

   DEBUG ironic.conductor.task_manager [-] Successfully released exclusive lock for provision action manage on node d7e2aed8-50a9-4427-baaa-f8f595e2ceb3 (lock was held 60.02 sec) release_resources /usr/lib/python3.6/site-packages/ironic/conductor/task_manager.py:447
   

   The message tells you the reason the lock was held (for provision action manage) and the amount of time it was held (60.02 seconds, which is way too much for accessing a BMC).

Unfortunately, due to the way the conductor is designed, it is not possible to gracefully break a stuck lock held in *-ing states. As the last resort, you may need to restart the affected conductor. See Why are my nodes stuck in a “-ing” state?.

What is ConcurrentActionLimit?

ConcurrentActionLimit is an exception which is raised to clients when an operation is requested, but cannot be serviced at that moment because the overall threshold of nodes in concurrent “Deployment” or “Cleaning” operations has been reached.

These limits exist for two distinct reasons.

The first is they allow an operator to tune a deployment such that too many concurrent deployments cannot be triggered at any given time, as a single conductor has an internal limit to the number of overall concurrent tasks, this restricts only the number of running concurrent actions. As such, this accounts for the number of nodes in deploy and deploy wait states. In the case of deployments, the default value is relatively high and should be suitable for most larger operators.

The second is to help slow down the ability in which an entire population of baremetal nodes can be moved into and through cleaning, in order to help guard against authenticated malicious users, or accidental script driven operations. In this case, the total number of nodes in deleting, cleaning, and clean wait are evaluated. The default maximum limit for cleaning operations is 50 and should be suitable for the majority of baremetal operators.

These settings can be modified by using the conductor.max_concurrent_deploy and conductor.max_concurrent_clean settings from the ironic.conf file supporting the ironic-conductor service. Neither setting can be explicitly disabled, however there is also no upper limit to the setting.

Note

This was an infrastructure operator requested feature from actual lessons learned in the operation of Ironic in large scale production. The defaults may not be suitable for the largest scale operators.

Why do I have an error that an NVMe Partition is not a block device?

In some cases, you can encounter an error that suggests a partition that has been created on an NVMe block device, is not a block device.

Example:

lsblk: /dev/nvme0n1p2: not a block device

What has happened is the partition contains a partition table inside of it which is confusing the NVMe device interaction. While basically valid in some cases to have nested partition tables, for example, with software raid, in the NVMe case the driver and possibly the underlying device gets quite confused. This is in part because partitions in NVMe devices are higher level abstracts.

The way this occurs is you likely had a whole-disk image, and it was configured as a partition image. If using glance, your image properties may have a img_type field, which should be whole-disk, or you have a kernel_id and ramdisk_id value in the glance image properties field. Definition of a kernel and ramdisk value also indicates that the image is of a partition image type. This is because a whole-disk image is bootable from the contents within the image, and partition images are unable to be booted without a kernel, and ramdisk.

If you are using Ironic in standalone mode, the optional instance_info/image_type setting may be advisable to be checked. Very similar to Glance usage above, if you have set Ironic’s node level instance_info/kernel and instance_info/ramdisk parameters, Ironic will proceed with deploying an image as if it is a partition image, and create a partition table on the new block device, and then write the contents of the image into the newly created partition.

Note

As a general reminder, the Ironic community recommends the use of whole disk images over the use of partition images.

Why can’t I use Secure Erase/Wipe with RAID controllers?

Situations have been reported where an infrastructure operator is expecting particular device types to be Secure Erased or Wiped when they are behind a RAID controller.

For example, the server may have NVMe devices attached to a RAID controller which could be in pass-through or single disk volume mode. The same scenario exists basically regardless of the disk/storage medium/type.

The basic reason why is that RAID controllers essentially act as command translators with a buffer cache. They tend to offer a simplified protocol to the Operating System, and interact with the storage device in whatever protocol is native to the device. This is the root of the underlying problem.

Protocols such as SCSI are rooted in quite a bit of computing history, but never evolved to include primitives like Secure Erase which evolved in the ATA protocol.

The closest primitives in SCSI to ATA Secure Erase is the FORMAT UNIT and UNMAP commands.

FORMAT UNIT might be a viable solution, and a tool named sg_format exists, but there has not been a sufficient call upstream to implement this and test it sufficiently that the Ironic community would be comfortable shipping such a capability. The possibility also exists that a RAID controller might not translate this command through to an end device, just as some RAID controllers know how to handle and pass through ATA commands to disk devices which support them. It is entirely dependent upon the hardware configuration scenario.

The UNMAP command is similar to the ATA TRIM command. Unfortunately the SCSI protocol requires this be performed at block level, and similar to FORMAT UNIT, it may not be supported or just passed through.

If your interested in working on this area, or are willing to help test, please feel free to contact the Ironic development community. An additional option is the creation of your own custom Hardware Manager which can contain your preferred logic, however this does require some Python development experience.

One last item of note, depending on the RAID controller, the BMC, and a number of other variables, you may be able to leverage the RAID configuration interface to delete volumes/disks, and recreate them. This may have the same effect as a clean disk, however that too is RAID controller dependent behavior.

I’m in “clean failed” state, what do I do?

There is only one way to exit the clean failed state. But before we visit the answer as to how, we need to stress the importance of attempting to understand why cleaning failed. On the simple side of things, this may be as simple as a DHCP failure, but on a complex side of things, it could be that a cleaning action failed against the underlying hardware, possibly due to a hardware failure.

As such, we encourage everyone to attempt to understand why before exiting the clean failed state, because you could potentially make things worse for yourself. For example if firmware updates were being performed, you may need to perform a rollback operation against the physical server, depending on what, and how the firmware was being updated. Unfortunately this also borders the territory of “no simple answer”.

This can be counter balanced with sometimes there is a transient networking failure and a DHCP address was not obtained. An example of this would be suggested by the last_error field indicating something about “Timeout reached while cleaning the node”, however we recommend following several basic troubleshooting steps:

• Consult the last_error field on the node, utilizing the baremetal node show <uuid> command.

• If the version of ironic supports the feature, consult the node history log, baremetal node history list and baremetal node history get <uuid>.

• Consult the actual console screen of the physical machine. If the ramdisk booted, you will generally want to investigate the controller logs and see if an uploaded agent log is being stored on the conductor responsible for the baremetal node. Consult Retrieving logs from the deploy ramdisk. If the node did not boot for some reason, you can typically just retry at this point and move on.

How to get out of the state, once you’ve understood why you reached it in the first place, is to utilize the baremetal node manage <node_id> command. This returns the node to manageable state, from where you can retry “cleaning” through automated cleaning with the provide command, or manual cleaning with clean command. or the next appropriate action in the workflow process you are attempting to follow, which may be ultimately be decommissioning the node because it could have failed and is being removed or replaced.

I can’t seem to introspect newly added nodes in a large cluster

With larger clusters, the act of synchronizing DHCP for introspection and hardware discovery can take quite a bit of time because of the operational overhead. What happens is we spend so much time trying to perform the update that the processes stay continuously busy, which can have a side effect such as impacting the ability to successfully introspect nodes which were very recently added to the cluster.

To remedy this, try setting [pxe_filter]sync_period to be less frequent, i.e. a larger value to enable conductors to have time between running syncs.

Note

It is anticipated that as part of the 2024.1 release, Ironic will have this functionality also merged into Ironic directly as part of the merge of the ironic-inspector service into ironic itself. This merger will result in a slightly more performant implementation, which may necessitate re-evaluation and tuning of the [pxe_filter]sync_period parameter.

Some or all of my baremetal nodes disappeared! Help?!

If you just upgraded, and this has occurred:

1. Don’t Panic

2. Don’t try to re-enroll the nodes. They should still be there, you just can’t see them at the moment.

Over the past few years, Ironic and OpenStack project as a whole has been working to improve the model of Role Based Access Control. For users of Ironic, this means an extended role based access control model allowing delineation of nodes and the ability for projects to both self-manage.

The result is that users inside of a project are only permitted to see baremetal nodes, through the owner and lessee field, which has been granted access to the project.

However, as with any complex effort, there can be hiccups, and you have encountered one. Specifically that based upon large scale operator feedback, Ironic kept logic behind System scoped user usage, which OpenStack largely avoided due to concerns over effort.

As such, you have a couple different paths you can take, and your ideal path is also going to vary upon your model of usage and comfort level. We recommend reading the rest of this answer section before taking any further action.

A good starting point is obtaining a system scoped account with an admin or member role. Either of those roles will permit a node’s owner or lessee fields to be changed. Executing baremetal node list commands with this account should show you all baremetal nodes across all projects. Alternatively, if you just want to enable the legacy RBAC policies temporarily to change the fields, that is also an option, although not encouraged, and can be done utilizing the [oslo_policy] enforce_scope and [oslo_policy] enforce_new_defaults settings.

System Scoped Accounts

A system scoped account is one which has access and authority over the whole of the of an OpenStack deployment. A simplified way to think of this is when deployed, a username and password is utilized to “bootstrap” keystone. The rights granted to that user are inherently a system scoped admin role level of access. You can use this level of access to check the status, or run additional commands.

In this example below, which if successful, should return a list of all baremetal nodes known to Ironic, once the executing user supplies the valid password. In this case the “admin” account keystone was bootstrapped with. As a minor note, you will not be able to have any “OS_*” environment variables loaded into your current command shell, including “OS_CLOUD” for this command to be successful.

$ openstack --os-username=admin --os-user-domain-name=default --os-system-scope all baremetal node list

You can alternatively issue a system-scoped token and reuse further commands with that token, or even generate a new system scoped account with a role of member.

Changing/Assigning an Owner

Ironic performs matching based upon Project ID. The owner field can be set to a project’s ID value, which allows baremetal nodes to be visible.

$ PROJECT_ID=$(openstack project show -c id -f value $PROJECT_NAME)
$ baremetal node set --owner $PROJECT_ID $NODE_UUID_OR_NAME

Why am I only seeing some of the nodes?

During the Zed development cycle of Ironic, Ironic added an option which defaulted to True, which enabled project scoped admin users to be able to create their own baremetal nodes without needing higher level access. This default enabled option, [api] project_admin_can_manage_own_nodes, automatically stamps the requestor’s project ID on to a baremetal node if an owner is not otherwise specified upon creation. Obviously, this can create a mixed perception if an operator never paid attention to the owner field before now.

If your bare metal management processes require that full machine management is made using a project scoped account, please configure an appropriate node owner for the nodes which need to be managed. Ironic recognizes this is going to vary based upon processes and preferences.

Config Drives in Swift, but rebuilds fails?

When deploying instances, Ironic can be configured such that configuration drives are stored in Swift. The pointer to the configuration drive is saved in Ironic as a Temporary URL which has a time expiration.

When you issue the rebuild request for a node, Ironic expects that you will supply new configuration drive contents with your request, however this is also optional.

Because Swift has been set as the optional configuration drive storage location, a rebuild can fail if the prior configuration drive file is no longer accessible and no new configuration drive has been supplied to Ironic.

To resolve this case, you can either supply new configuration drive contents with your request, or disable configuration from being stored in Swift for new baremetal node deployments by changing setting deploy.configdrive_use_object_store to false.

Ironic says my Image is Invalid

As a result of security fixes which were added to Ironic, resulting from the security posture of the qemu-img utility, Ironic enforces certain aspects related to image files.

• Enforces that the file format of a disk image matches what Ironic is told by an API user. Any mismatch will result in the image being declared as invalid. A mismatch with the file contents and what is stored in the Image service will necessitate uploading a new image as that property cannot be changed in the image service after creation of an image.

• Enforces that the input file format to be written is qcow2 or raw. This can be extended by modifying [conductor]permitted_image_formats in ironic.conf.

• Performs safety and sanity check assessment against the file, which can be disabled by modifying [conductor]disable_deep_image_inspection and setting it to True. Doing so is not considered safe and should only be done by operators accepting the inherent risk that the image they are attempting to use may have a bad or malicious structure. Image safety checks are generally performed as the deployment process begins and stages artifacts, however a late stage check is performed when needed by the ironic-python-agent.

Using /dev/sda does not write to the first disk

Alternative name: I chose /dev/sda but I found it as /dev/sdb after rebooting.

Historically, Linux users have grown accustom to a context where /dev/sda is the first device in a physical machine. Meaning, if you look at the device by_path information or the HCTL, or device LUN, the device ends with a zero.

For example, assuming 3 disks, two controllers, with a single disk on the second controller would look something like this:

• /dev/sda maps to a device with lun 0, HCTL 0:0:0:0

• /dev/sdb maps to a device with lun 1, HCTL 0:0:1:0

• /dev/sdc maps to a device with lun 2, HCTL 0:1:0:0

However, this was a pattern we grew accustom to because the order of device discovery was sequential and synchronous. In other words the kernel stepped through all possible devices one at a time. Where this breaks is when the kernel is operating in a mode where device initialization is asynchronous as some distributions have decided to adopt.

The result of a move to an asynchronous initialization is /dev/sda has always been the first device to initialize, not the first device in the system. As a result, we can end up with something looking like:

• /dev/sda maps to a device with lun 1, HCTL 0:0:1:0

• /dev/sdb maps to a device with lun 2, HCTL 0:1:0:0

• /dev/sdc maps to a device with lun 0, HCTL 0:0:0:0

Generally, most operators might then consider referencing the /dev/disk/by-path structure to match disk devices because that seems to imply a static order, however a kernel operating with asynchronous device initialization will order everything, including PCI devices the same way, meaning by-path can also be unreliable. Furthermore, if your server hardware is using multipath IO, you should be operating with multipath enabled such that the device is used.

The net result is the best criteria to match on is:

• Serial Number

• World Wide Name

• Device HCTL, which does appear to be static in these cases, but is not applicable for hosts using multipathing. It may, ultimately, not be static enough, just depending on the hardware in use.