Boot From Volume

Overview

The Bare Metal service supports booting from a Cinder iSCSI volume as
of the Pike release. This guide will primarily deal with this use case,
but will be updated as more paths for booting from a volume, such as
FCoE, are introduced.

The boot from volume is supported on both legacy BIOS and UEFI (iPXE
binary for EFI booting) boot mode. We need to perform with suitable
images which will be created by diskimage-builder tool.

How this works –
From Ironic’s point of view

In essence, ironic sets the stage for the process, by providing the
required information to the boot interface to facilitate the
configuration of the the node OR the iPXE boot templates such that the
node CAN be booted.

In this example, the boot interface does the heavy lifting. For
drivers the irmc and ilo hardware types with
hardware type specific boot interfaces, they are able to signal via an
out of band mechanism to the baremetal node’s BMC that the integrated
iSCSI initiators are to connect to the supplied volume target
information.

In most hardware this would be the network cards of the machine.

In the case of the ipxe boot interface, templates are
created on disk which point to the iscsi target information that was
either submitted as part of the volume target, or when integrated with
Nova, what was requested as the baremetal’s boot from volume disk upon
requesting the instance.

In terms of network access, both interface methods require
connectivity to the iscsi target. In the vendor driver specific path,
additional network configuration options may be available to allow
separation of standard network traffic and instance network traffic. In
the iPXE case, this is not possible as the OS userspace re-configures
the iSCSI connection after detection inside the OS ramdisk boot.

An iPXE user may be able to leverage multiple VIFs, one
specifically set to be set with pxe_enabled to handle the
initial instance boot and back-end storage traffic where as external
facing network traffic occurs on a different interface. This is a common
pattern in iSCSI based deployments in the physical realm.

Prerequisites

Currently booting from a volume requires:

• Bare Metal service version 9.0.0
• Bare Metal API microversion 1.33 or later
• A driver that utilizes the PXE boot mechanism </install/configure-pxe>.
  Currently booting from a volume is supported by the reference drivers
  that utilize PXE boot mechanisms when iPXE is enabled.
• iPXE is an explicit requirement, as it provides the mechanism that
  attaches and initiates booting from an iSCSI volume.
• Metadata services need to be configured and available for the
  instance images to obtain configuration such as keys. Configuration
  drives are not supported due to minimum disk extension sizes.

Conductor Configuration

In ironic.conf, you can specify a list of enabled storage interfaces.
Check [DEFAULT]enabled_storage_interfaces in your
ironic.conf to ensure that your desired interface is enabled. For
example, to enable the cinder and noop storage
interfaces:

[DEFAULT]
enabled_storage_interfaces = cinder,noop

If you want to specify a default storage interface rather than
setting the storage interface on a per node basis, set
[DEFAULT]default_storage_interface in ironic.conf. The
default_storage_interface will be used for any node that
doesn’t have a storage interface defined.

Node Configuration

Storage Interface

You will need to specify what storage interface the node will use to
handle storage operations. For example, to set the storage interface to
cinder on an existing node:

baremetal node set --storage-interface cinder $NODE_UUID

A default storage interface can be specified in ironic.conf. See the
Conductor Configuration section
for details.

iSCSI Configuration

In order for a bare metal node to boot from an iSCSI volume, the
iscsi_boot capability for the node must be set to
True. For example, if you want to update an existing node
to boot from volume:

baremetal node set --property capabilities=iscsi_boot:True $NODE_UUID

You will also need to create a volume connector for the node, so the
storage interface will know how to communicate with the node for storage
operation. In the case of iSCSI, you will need to provide an iSCSI
Qualifying Name (IQN) that is unique to your SAN. For example, to create
a volume connector for iSCSI:

baremetal volume connector create \
         --node $NODE_UUID --type iqn --connector-id iqn.2017-08.org.openstack.$NODE_UUID

Image Creation

We use disk-image-create in diskimage-builder tool to
create images for boot from volume feature. Some required elements for
this mechanism for corresponding boot modes are as following:

• Legacy BIOS boot mode: iscsi-boot element.
• UEFI boot mode: iscsi-boot and
  block-device-efi elements.

An example below:

export IMAGE_NAME=<image_name>
export DIB_CLOUD_INIT_DATASOURCES="ConfigDrive, OpenStack"
disk-image-create centos7 vm cloud-init-datasources dhcp-all-interfaces iscsi-boot dracut-regenerate block-device-efi -o $IMAGE_NAME

Advanced Topics

Use without the Compute
Service

As discussed in other sections, the Bare Metal service has a concept
of a connector that is used to represent
an interface that is intended to be utilized to attach the remote
volume.

In addition to the connectors, we have a concept of a target that can be defined via the API. While a
user of this feature through the Compute service would automatically
have a new target record created for them, it is not explicitly
required, and can be performed manually.

A target record can be created using a command similar to the example
below:

baremetal volume target create \
          --node $NODE_UUID --type iscsi --boot-index 0 --volume $VOLUME_UUID

Use Without Cinder

In the Rocky release, an external storage interface is
available that can be utilized without a Block Storage Service
installation.

Under normal circumstances the cinder storage interface
interacts with the Block Storage Service to orchestrate and manage
attachment and detachment of volumes from the underlying block service
system.

The external storage interface contains the logic to
allow the Bare Metal service to determine if the Bare Metal node has
been requested with a remote storage volume for booting. This is in
contrast to the default noop storage interface which does
not contain logic to determine if the node should or could boot from a
remote volume.

It must be noted that minimal configuration or value validation
occurs with the external storage interface. The
cinder storage interface contains more extensive
validation, that is likely un-necessary in a external
scenario.

Setting the external storage interface:

baremetal node set --storage-interface external $NODE_UUID

Setting a volume:

baremetal volume target create --node $NODE_UUID \
    --type iscsi --boot-index 0 --volume-id $VOLUME_UUID \
    --property target_iqn="iqn.2010-10.com.example:vol-X" \
    --property target_lun="0" \
    --property target_portal="192.168.0.123:3260" \
    --property auth_method="CHAP" \
    --property auth_username="ABC" \
    --property auth_password="XYZ" \

Ensure that no image_source is defined:

baremetal node unset \
    --instance-info image_source $NODE_UUID

Deploy the node:

baremetal node deploy $NODE_UUID

Upon deploy, the boot interface for the baremetal node will attempt
to either create iPXE configuration OR set boot parameters out-of-band
via the management controller. Such action is boot interface specific
and may not support all forms of volume target configuration. As of the
Rocky release, the bare metal service does not support writing an
Operating System image to a remote boot from volume target, so that also
must be ensured by the user in advance.

Records of volume targets are removed upon the node being undeployed,
and as such are not persistent across deployments.

Cinder Multi-attach

Volume multi-attach is a function that is commonly performed in
computing clusters where dedicated storage subsystems are utilized. For
some time now, the Block Storage service has supported the concept of
multi-attach. However, the Compute service, as of the Pike release, does
not yet have support to leverage multi-attach. Concurrently,
multi-attach requires the backend volume driver running as part of the
Block Storage service to contain support for multi-attach volumes.

When support for storage interfaces was added to the Bare Metal
service, specifically for the cinder storage interface, the
concept of volume multi-attach was accounted for, however has not been
fully tested, and is unlikely to be fully tested until there is Compute
service integration as well as volume driver support.

The data model for storage of volume targets in the Bare Metal
service has no constraints on the same target volume from being
utilized. When interacting with the Block Storage service, the Bare
Metal service will prevent the use of volumes that are being reported as
in-use if they do not explicitly support multi-attach.