Boot configuration API

Problem description

Currently, an Ironic conductor generates various boot files (e.g. iPXE scripts) in the public directory of the HTTP server associated with it. Then dynamic DHCP configuration with Neutron is used to point a node at the right HTTP server.

For standalone installations, the DHCP configuration is usually static. It works well with a single conductor, but is not operational out-of-box in a multi-conductor setup: a node simply does not know which HTTP server to boot from.

Proposed change

New API will be added to Ironic to serve boot configuration files, with iPXE being the reference implementation. This approach will allow us to rely on the existing hash ring to direct the request to the right conductor. See REST API impact for further details.

This is how the boot process will work in the case of iPXE, standalone Ironic and static DHCP configuration:

  1. The Node’s iPXE firmware initiates a new DHCP session.

  2. The DHCP request reaches the Ironic’s DHCP server (usually dnsmasq).

  3. The DHCP server responds with an IP address and the catch-all boot script. This boot script will be located next to an arbitrary conductor, quite likely the one co-existing with the DHCP server.

  4. The boot script will direct the iPXE firmware to the URL in the form of http://<IP>:6385/v1/boot/<MAC>/boot.ipxe where <IP> is the IP of any Ironic API instance (most likely the one co-existing with the HTTP server), <MAC> is the node’s MAC address.

  5. Ironic API will find the node by MAC and direct the request to the correct conductor.

  6. The conductor responsible for the node returns the iPXE script.


One proposed approach is to enable coordination between conductors. While potentially useful in more cases, that proposal has a JSON RPC multiplication problem in a standalone setup.

The other approach to the problem is to introduce managed standalone DHCP. The first step towards it has already been done: Ironic can manage its dnsmasq server. In a multi-conductor setup, it may imply having several dnsmasq instances on the same network, which is a potentially problematic setup. The current implementation still requires some sort of coordination between conductor or a periodic task to disable access to the DHCP servers of unrelated conductors.

A much better, although also the most complex, option would be to use an existing DHCP server with API access. One such server is Kea. The main problem: such a complex change may be too much for Ironic right now. The contributors are spread thin already. On top of that, I’ve been told that only paid addons give us what we need.

Data model impact


State Machine Impact


REST API impact

GET /v1/boot/<MAC>/<NAME>


MAC address of any NIC of a node.


Configuration name to request (e.g. boot.ipxe).

The API will find the node by the MAC address, check its provision state and call into the get_boot_config RPC on success. On failure, a generic HTTP 404 will be returned to avoid disclosing any further information.


The whole API branch /v1/boot will not be versioned since we don’t expect firmware implementations to support extra headers or any sort of reasonable version negotiation.

Client (CLI) impact

None. The API is not for end users.

“openstack baremetal” CLI




RPC API impact

def get_boot_config(self, context, node_id, name):
    """Request a boot configuration file."""

Driver API impact

The boot interface will get a new call:

def get_boot_config(self, task, name):
    """Request a boot configuration file.

    :param task: a TaskManage instance with a shared lock.
    :param name: configuration name.
    :raises: NotFound if the configuration cannot be produced.
    raise exception.UnsupportedDriverExtension(
        driver=task.node.driver, extension='get_boot_config')

A reference implementation will be added to the iPXE boot interface, supporting a configuration called boot.ipxe - the iPXE script.

Nova driver impact


Ramdisk impact


Security impact

The new API will allow enumerating nodes in certain states by their MAC addresses. Some information may potentially be exposed by the boot configuration. The enumeration is already possible with the lookup API, and the configuration can be leaked by the boot scripts in the HTTP server. We will advise operators to limit access to the new API endpoints.

On top of that, the boot interface’s validate will not be called to avoid exposing information about the node fields.

Other end user impact


Scalability impact

Serving boot scripts via the API is somewhat less efficient than from an HTTP server. Operators concerned about the impact can opt into using the old approach.

We will avoid using an exclusive lock or launching additional threads in the implementation. The initial version will just read the existing files from disk.

Performance Impact


Other deployer impact

The feature will be configured with a few new options:

[pxe]ipxe_use_boot_config_api = False

Enables the feature. If true, the generated root iPXE script (boot.ipxe) will contain links to the boot configuration API, not to the HTTP server.

[pxe]ipxe_config_api_root_url = <None>

Specifies the root URL to use for links to the boot configuration API. An example use case is an Ironic deployment with TLS: iPXE does not support custom certificates without recompiling the firmware, so e.g. a proxy must be established instead.

[api]restrict_boot_config = True

Instructs the API to be restricted to only nodes in * WAIT states. Operators using fast-track may want to set this to False.

Developer impact




Primary assignee:

Dmitry Tantsur (dtantsur)

Work Items





We can switch Bifrost to the new API. It’s highly likely that Metal3 will also switch to it in the near future as we develop its HA story.

Upgrades and Backwards Compatibility


Documentation Impact

Installation guide may need to be adjusted.