Store and allow libvirt instance device buses and models to be updated

https://blueprints.launchpad.net/nova/+spec/libvirt-device-bus-model-update

QEMU support for device buses and models can come and go dependent on the underlying instance machine type and QEMU version used within an environment. The defaults provided by libosinfo and currently hardcoded in to the libvirt driver are not persisted by each instance at present.

This spec aims to outline a basic set of nova-manage commands to allow operators to move instances between specific device bus and model types without requiring a rebuild.

Problem description

At present device bus and model types defined as image properties associated with an instance are always used when launching instances with the libvirt driver. When these types are not defined as image properties their values either come from libosinfo or those directly hardcoded into the libvirt driver.

Support for each device bus and model is dependent on the machine type used and version of QEMU available on the underlying compute host.

As such any changes to the machine type of an instance or version of QEMU on a host might suddenly invalidate the stashed device bus or model image properties with no way of updating outside of a complete instance rebuild against a new image defining new image properties.

Additionally any changes to the defaults provided by libosinfo or the libvirt driver could result in unforeseen changes to existing instances. This has been encountered in the past as libosinfo assumes that libvirt domain definitions are static when OpenStack Nova specifically rewrites and redefines these domains during a hard reboot or migration allowing changes to possibly occur.

Use Cases

  • As a user I want the device buses and models used by my instance to remain stable for as long as possible and not be changed by new defaults in libosinfo or the OpenStack Nova libvirt driver.

  • As an operator I want to change the device bus or model of an instance without forcing users to fully rebuild the instance in order to accommodate changing machine types or QEMU deprecations for certain types.

Proposed change

Register existing device buses and models within system_metadata

As with hw_machine_type we first want to ensure the current device bus and model types associated with an instance are stashed ensuring they remain stable during the lifetime of the instance. This already happens when these buses or models are defined by image properties so we only need to capture their value when these are not defined at either service startup or instance creation time.

The following list of image properties outline the list of device buses and models this spec will aim to cover:

  • hw_cdrom_bus

  • hw_disk_bus

  • hw_floppy_bus

  • hw_input_bus

  • hw_pointer_model

  • hw_video_model

  • hw_vif_model

Note

hw_rng_model, hw_scsi_model and hw_rescue_bus are not included here as they have no default values. They must be defined to be used negating the need for us to capture them here.

Provide nova-manage commands to update existing device buses and models

With the bus and model types stored we can now provide commands to operators to inspect and update only the list of allowed image properties above:

$ nova-manage image-property list $instance

$ nova-manage image-property show $instance $property

Will list or show the stashed image properties of an instance.

$ nova-manage image-property set \
    --property hw_disk_bus=scsi \
    --property hw_scsi_model=virtio-scsi $instance

Will update image properties of an instance, only accepting the previously defined list of image properties for the time being.

Prerequisites

The following prerequisites apply when attempting to update the image properties of an instance:

  • The instance must be in a STOPPED, SHELVED or SHELVED_OFFLOADED vm_sate.

  • The provided type will be validated against the corresponding versioned object fields for the bus or model.

Once updated the user or admin can power on or unshelve the instance, causing the underlying libvirt domain to be redefined using the new bus or model type.

Alternatives

None, other than providing a generic API to allow stashed image properties to be updated by users over time without requiring a rebuild but that’s out of scope for this basic nova-manage command spec.

Data model impact

None

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

Users should now find that the device bus and models used by their instances remain stable throughout their lifetime unless a move is forced upon them by the operator, QEMU support deprecations etc.

Performance Impact

Stashing these values will incur a slight overhead at compute service start time when using the libvirt driver and additionally when spawning new instances.

Other deployer impact

Operators should have more control over when and how they move users to different machine types and versions of QEMU.

Developer impact

None

Upgrade impact

None

Implementation

Assignee(s)

Primary assignee:

lyarwood

Other contributors:

Feature Liaison

Feature liaison:

lyarwood

Work Items

  • Register existing device buses and models within system_metadata

  • Provide nova-manage commands to update existing device buses and models

Dependencies

None

Testing

Extensive unit and functional tests will be written to validate this.

Documentation Impact

Operator/admin facing documentation will be written outlining the usecase for these commands as well as the normal documentation for the commands themselves.

References

History

Revisions

Release Name

Description

Yoga

Reproposed

Xena

Introduced