Virtual instance rescue with stable disk devices¶
This will provide the ability to indicate that the rescue disk image should be attached as a transient disk device (ie USB stick), so that existing storage attached to an instance doesn’t change its device address during rescue mode.
When an instance is booted normally there are a number of possible disks that will be attached to the instance
An ephemeral or persistent cinder volume root disk
Zero or more ephemeral non-root disks
Zero or more persistent non-root cinder volumes
An optional swap disk
An optional config drive disk
When the instance is booted in rescue mode though, this storage setup changes significantly, and differently depending on virt drivers. In the Libvirt driver, the rescue instance gets:
A rescue root disk
The original root disk
An optional config drive disk
There are multiple problems with this. First of all several of the disks are missing entirely, eg the ephemeral non-root disks, all cinder volumes and the swap disk. This missing storage limits the scope of work the admin can do in rescue mode.
The rescue root disk is put on a device that previously held the real root disk. For example the rescue root is /dev/vda and the real root image is now shifted to a different device /dev/vdb. Although a well designed OS setup should not rely on the root device appearing at a fixed device name, some OSes none the less do depend on this. Moving the root disk during rescue mode can thus introduce problems of its own, and in fact contribute to mistakes in rescue mode. For example it may confuse the admin into setting up their fstab to refer to /dev/vdb, when the root disk will go back to /dev/vda after rescue mode is finished.
This change in disk presence during rescue mode is very different to what happens to disks on a baremetal machine when booted from rescue media. This means that admin knowledge from working in a bare metal world needs to be re-learned for OpenStack rescue mode, which adds an undesirable learning burden for the admin.
When disks change what address they appear at, this can cause upset licensing checks of some guests OS too. For example, if hardware devices change their address too frequently, Windows may decide to ask for license re-activation. This is again an undesirable thing for admins in general.
When the tenant user boots a VM in rescue mode they expect the existing storage device configuration to be identical to that seen when running in normal mode, but with an extra transient disk hotplugged to represent the rescue media.
This spec will not cover the removal of the current boot from volume check in the compute API that currently blocks any attempt to rescue an instance using a root cinder volume. The removal of this check and subsequent impact on the overall API will be covered in a follow up spec.
The compute manager code will be changed such that when rescue is performed the full block device mapping will be present. This will allow instances to be configured with the full set of non-root cinder volumes that would appear during normal boot.
New image properties have already been introduced during Ocata  that will be used to indicate the type of device and associated bus to use as the rescue device.
If omitted, the virt driver will default to whatever behaviour it currently has for setting up the rescue disk. For the Libvirt driver, this means the default bus would match the hw_disk_bus, and the device type would be “disk”.
The expected recommended setup would be to tag the rescue image in glance with hw_rescue_bus=usb, which would indicate to the virt driver that it should attach a USB flash drive to the guest, containing the rescue image. For hypervisors which can’t support this an alternative recommendation would be to tag the rescue image with hw_rescue_bus=ide and hw_rescue_device=cdrom to cause a new CDROM device to be exposed with the rescue media.
The Libvirt nova driver will be changed so that when booting in rescue mode, all the non-root cinder volumes, local ephemeral non-root disks and swap disks are present in rescue mode. The rescue root device will be added as the last device in the configuration, but will be marked as bootable for the BIOS, so it takes priority over the existing root device. This relies on KVM/QEMU supporting the “bootindex” parameter, which all supported versions do. This new rescue mode would not be supported by Xen, nor LXC.
Other virt driver maintainers may wish to also implement this blueprint, so approval should be considered to give blessing to all virt drivers. If other virt driver maintainers wish to commit to doing this in this cycle the list of assignees will be updated.
Do nothing is always an option, but the current setup has a number of undesirable characteristics described earlier.
An alternative might be to simply hardcode a different approach. eg when using KVM simply always use a USB flash device as the rescue media, and don’t bother with supporting an image property. This is certainly a viable option, and if it were not for the sake of maintaining backwards compatibility with earlier OpenStack, it might even be the preferred approach.
Data model impact¶
None, as the ImageMetaProps object changes have already landed in Ocata .
REST API impact¶
None, as support for BFV instances will be covered in a separate spec.
Other end user impact¶
The tenant user will gain the ability to set a new image meta property against rescue disk images which will indicate the type of disk bus and device to use when rescuing instances.
Other deployer impact¶
If the admin pre-populates any rescue disk images, they may wish to set the disk bus and device type to override the historic default behaviour.
Virt driver maintainers can continue to silently ignore the newly introduced image properties or optionally start using them by implementing this new stable device approach.
Older Libvirt based computes that are not able to honour the stable device rescue image properties will continue to silently ignore them as they have since these were introduced during Ocata . Once upgraded to Ussuri they will then start rescusing instances with a stable device layout.
- Primary assignee:
lyarwood (Libvirt impl)
- Other contributors:
Extend the compute manager rescue code to handle the full block device mapping including non-root cinder volume attachments.
Extend the nova Libvirt driver to setup all disks when running in rescue mode.
Extend the nova Libvirt driver to honour the new image meta properties in rescue mode disk config.
A new tempest Libvirt feature configurable and test will be used to validate correct operation of the new code.
The new image properties should be documented, and any information about rescue mode should be updated to explain how disks appear.