Block Device Mapping in Nova

Nova has a concept of block devices that can be exposed to cloud instances. There are several types of block devices an instance can have (we will go into more details about this later in this document), and which ones are available depends on a particular deployment and the usage limitations set for tenants and users. Block device mapping is a way to organize and keep data about all of the block devices an instance has.

When we talk about block device mapping, we usually refer to one of two things

  1. API/CLI structure and syntax for specifying block devices for an instance boot request

  2. The data structure internal to Nova that is used for recording and keeping, which is ultimately persisted in the block_device_mapping table. However, Nova internally has several “slightly” different formats for representing the same data. All of them are documented in the code and or presented by a distinct set of classes, but not knowing that they exist might trip up people reading the code. So in addition to BlockDeviceMapping [1] objects that mirror the database schema, we have:

    2.1 The API format - this is the set of raw key-value pairs received from the API client, and is almost immediately transformed into the object; however, some validations are done using this format. We will refer to this format as the ‘API BDMs’ from now on.

    2.2 The virt driver format - this is the format defined by the classes in :mod: nova.virt.block_device. This format is used and expected by the code in the various virt drivers. These classes, in addition to exposing a different format (mimicking the Python dict interface), also provide a place to bundle some functionality common to certain types of block devices (for example attaching volumes which has to interact with both Cinder and the virt driver code). We will refer to this format as ‘Driver BDMs’ from now on.

Data format and its history

In the early days of Nova, block device mapping general structure closely mirrored that of the EC2 API. During the Havana release of Nova, block device handling code, and in turn the block device mapping structure, had work done on improving the generality and usefulness. These improvements included exposing additional details and features in the API. In order to facilitate this, a new extension was added to the v2 API called BlockDeviceMappingV2Boot [2], that added an additional block_device_mapping_v2 field to the instance boot API request.

Block device mapping v1 (aka legacy)

This was the original format that supported only cinder volumes (similar to how EC2 block devices support only EBS volumes). Every entry was keyed by device name (we will discuss why this was problematic in its own section later on this page), and would accept only:

  • UUID of the Cinder volume or snapshot
  • Type field - used only to distinguish between volumes and Cinder volume snapshots
  • Optional size field
  • Optional delete_on_termination flag

While all of Nova internal code only uses and stores the new data structure, we still need to handle API requests that use the legacy format. This is handled by the Nova API service on every request. As we will see later, since block device mapping information can also be stored in the image metadata in Glance, this is another place where we need to handle the v1 format. The code to handle legacy conversions is part of the :mod: nova.block_device module.

Intermezzo - problem with device names

Using device names as the primary per-instance identifier, and exposing them in the API, is problematic for Nova mostly because several hypervisors Nova supports with its drivers can’t guarantee that the device names the guest OS assigns are the ones the user requested from Nova. Exposing such a detail in the public API of Nova is obviously not ideal, but it needed to stay for backwards compatibility. It is also required for some (slightly obscure) features around overloading a block device in a Glance image when booting an instance [3].

The plan for fixing this was to allow users to not specify the device name of a block device, and Nova will determine it (with the help of the virt driver), so that it can still be discovered through the API and used when necessary, like for the features mentioned above (and preferably only then).

Another use for specifying the device name was to allow the “boot from volume” functionality, by specifying a device name that matches the root device name for the instance (usually /dev/vda).

Currently (mid Liberty) users are discouraged from specifying device names for all calls requiring or allowing block device mapping, except when trying to override the image block device mapping on instance boot, and it will likely remain like that in the future. Libvirt device driver will outright override any device names passed with it’s own values.

Block device mapping v2

New format was introduced in an attempt to solve issues with the original block device mapping format discussed above, and also to allow for more flexibility and addition of features that were not possible with the simple format we had.

New block device mapping is a list of dictionaries containing the following fields (in addition to the ones that were already there):

  • source_type - this can have one of the following values:

    • image
    • volume
    • snapshot
    • blank

  • dest_type - this can have one of the following values:

    • local
    • volume

Combination of the above two fields would define what kind of block device the entry is referring to. We currently support the following combinations:

  • image -> local - this is only currently reserved for the entry referring to the Glance image that the instance is being booted with (it should also be marked as a boot device). It is also worth noting that an API request that specifies this, also has to provide the same Glance uuid as the image_ref parameter to the boot request (this is done for backwards compatibility and may be changed in the future). This functionality might be extended to specify additional Glance images to be attached to an instance after boot (similar to kernel/ramdisk images) but this functionality is not supported by any of the current drivers.
  • volume -> volume - this is just a Cinder volume to be attached to the instance. It can be marked as a boot device.
  • snapshot -> volume - this works exactly as passing type=snap does. It would create a volume from a Cinder volume snapshot and attach that volume to the instance. Can be marked bootable.
  • image -> volume - As one would imagine, this would download a Glance image to a cinder volume and attach it to an instance. Can also be marked as bootable. This is really only a shortcut for creating a volume out of an image before booting an instance with the newly created volume.
  • blank -> volume - Creates a blank Cinder volume and attaches it. This will also require the volume size to be set.
  • blank -> local - Depending on the guest_format field (see below), this will either mean an ephemeral blank disk on hypervisor local storage, or a swap disk (instances can have only one of those).
  • guest_format - Tells Nova how/if to format the device prior to attaching, should be only used with blank local images. Denotes a swap disk if the value is swap.
  • device_name - See the previous section for a more in depth explanation of this - currently best left empty (not specified that is), unless the user wants to override the existing device specified in the image metadata. In case of Libvirt, even when passed in with the purpose of overriding the existing image metadata, final set of device names for the instance may still get changed by the driver.
  • disk_bus and device_type - low level details that some hypervisors (currently only libvirt) may support. Some example disk_bus values can be: ide, usb, virtio, scsi, while device_type may be disk, cdrom, floppy, lun. This is not an exhaustive list as it depends on the virtualization driver, and may change as more support is added. Leaving these empty is the most common thing to do.
  • boot_index - Defines the order in which a hypervisor will try devices when attempting to boot the guest from storage. Each device which is capable of being used as boot device should be given a unique boot index, starting from 0 in ascending order. Some hypervisors may not support booting from multiple devices, so will only consider the device with boot index of 0. Some hypervisors will support booting from multiple devices, but only if they are of different types - eg a disk and CD-ROM. Setting a negative value or None indicates that the device should not be used for booting. The simplest usage is to set it to 0 for the boot device and leave it as None for any other devices.

Nova will not allow mixing of two formats in a single request, and will do basic validation to make sure that the requested block device mapping is valid before accepting a boot request.

[1]In addition to the BlockDeviceMapping Nova object, we also have the BlockDeviceDict class in :mod: nova.block_device module. This class handles transforming and validating the API BDM format.
[2]This work predates API microversions and thus the only way to add it was by means of an API extension.
[3]This is a feature that the EC2 API offers as well and has been in Nova for a long time, although it has been broken in several releases. More info can be found on this bug <https://launchpad.net/bugs/1370250>