![Vladislav Belogrudov](/assets/img/avatar_default.png)
Migration from and to RBD devices works for in-use volumes since Ussuri release. Also as of now a migration of in-use volumes for non-active VMs is prohibited for any backends. Related-Bug: #1673090 Related-Bug: #1868996 Closes-Bug: #1969913 Change-Id: Id639c412a6f5ad24ab9c80416c56c1d99b3f104c
13 KiB
Migration
Introduction to volume migration
Cinder provides the volume migration support within the same deployment, which means the node of cinder volume service, c-vol node where the source volume is located, is able to access the c-vol node where the destination volume is located, and both of them share the same Cinder API service, scheduler service, message queue service, etc.
As a general rule migration is possible for volumes in 'available' or 'in-use' status, for the driver which has implemented volume migration. So far, we are confident that migration will succeed for 'available' volumes, whose drivers implement the migration routines. However, the migration of 'in-use' volumes is driver dependent. It depends on different drivers involved in the operation. It may fail depending on the source or destination driver of the volume.
For example, for older releases (before Ussuri), the migration of 'in-use' volume from RBD to LVM will succeed, but from LVM to RBD, it will fail. Currently, migration of 'in-use' volumes will not succeed for any backend when they are attached to an instance in any of these states: SHUTOFF, SUSPENDED, or SOFT-DELETED.
There are two major scenarios, which volume migration supports in Cinder:
Scenario 1: Migration between two back-ends with the same volume type, regardless if they are located on the same c-vol node or not.
Scenario 2: Migration between two back-ends with different volume types, regardless if the back-ends are located on the same c-vol node or not.
Note
Retyping an unencrypted volume to the same size encrypted volume will most likely fail. Even though the volume is the same size as the source volume, the encrypted volume needs to store additional encryption information overhead. This results in the new volume not being large enough to hold all data. Please do not try this in older releases.
How to do volume migration via CLI
Scenario 1 of volume migration is done via the following command from the CLI:
cinder migrate [--force-host-copy [<True|False>]]
[--lock-volume [<True|False>]]
<volume> <host>
Mandatory arguments:
<volume> ID of volume to migrate.
<host> Destination host. The format of host is
host@backend#POOL, while 'host' is the host
name of the volume node, 'backend' is the back-end
name and 'POOL' is a logical concept to describe
a set of storage resource, residing in the
back-end. If the back-end does not have specified
pools, 'POOL' needs to be set with the same name
as 'backend'.
Optional arguments:
--force-host-copy [<True|False>]
Enables or disables generic host-based force-
migration, which bypasses the driver optimization.
Default=False.
--lock-volume [<True|False>]
Enables or disables the termination of volume
migration caused by other commands. This option
applies to the available volume. True means it locks
the volume state and does not allow the migration to
be aborted. The volume status will be in maintenance
during the migration. False means it allows the volume
migration to be aborted. The volume status is still in
the original status. Default=False.
Important note: Currently, error handling for failed migration operations is under development in Cinder. If we would like the volume migration to finish without any interruption, please set --lock-volume to True. If it is set to False, we cannot predict what will happen, if other actions like attach, detach, extend, etc, are issued on the volume during the migration. It all depends on which stage the volume migration has reached and when the request of another action comes.
Scenario 2 of volume migration can be done via the following command from the CLI:
cinder retype --migration-policy on-demand
<volume> <volume-type>
Mandatory arguments:
<volume> Name or ID of volume for which to modify type.
<volume-type> New volume type.
Source volume type and destination volume type must be different and they must refer to different back-ends.
Configurations
To set up an environment to try the volume migration, we need to configure at least two different back-ends on the same node of cinder volume service, c-vol node or two back-ends on two different volume nodes of cinder volume service, c-vol nodes. Which command to use, 'cinder migrate' or 'cinder retype', depends on which type of volume we would like to test.
Scenario 1 for migration
To configure the environment for Scenario 1 migration, e.g. a volume is migrated from back-end <driver-backend> on Node 1 to back-end <driver-backend> on Node 2, cinder.conf needs to contains the following entries for the same back-end on both of source and the destination nodes:
- For Node 1:
-
... [<driver-backend>] volume_driver=xxxx volume_backend_name=<driver-backend> ...
- For Node 2:
-
... [<driver-backend>] volume_driver=xxxx volume_backend_name=<driver-backend> ...
If a volume with a predefined volume type is going to migrate, the back-end drivers from Node 1 and Node 2 should have the same value for volume_backend_name, which means <driver-backend> should be the same for Node 1 and Node 2. The volume type can be created with the extra specs {volume_backend_name: driver-biz}.
If we are going to migrate a volume with a volume type of none, it is not necessary to set the same value to volume_backend_name for both Node 1 and Node 2.
Scenario 2 for migration
To configure the environment for Scenario 2 migration: For example, a volume is migrated from driver-biz back-end on Node 1 to driver-net back-end on Node 2, cinder.conf needs to contains the following entries:
- For Node 1:
-
... [driver-biz] volume_driver=xxxx volume_backend_name=driver-biz ...
- For Node 2:
-
... [driver-net] volume_driver=xxxx volume_backend_name=driver-net ...
For example, a volume is migrated from driver-biz back-end on Node 1 to driver-biz back-net on the same node, cinder.conf needs to contains the following entries:
... [driver-biz] volume_driver=xxxx volume_backend_name=driver-biz ...
... [driver-net] volume_driver=xxxx volume_backend_name=driver-net ...
Two volume types need to be created. One is with the extra specs: {volume_backend_name: driver-biz}. The other is with the extra specs: {volume_backend_name: driver-net}.
What can be tracked during volume migration
The volume migration is an administrator only action and it may take a relatively long time to finish. The property 'migration status' will indicate the stage of the migration process for the volume. The administrator can check the 'migration status' via the 'cinder list' or 'cinder show <volume-id>' command. The 'cinder list' command presents a list of all the volumes with some properties displayed, including the migration status, only to the administrator. However, the migration status is not included if 'cinder list' is issued by an ordinary user. The 'cinder show <volume-id>' will present all the detailed information of a specific volume, including the migration status, only to the administrator.
If the migration status of a volume shows 'starting', 'migrating' or 'completing', it means the volume is in the process of a migration. If the migration status is 'success', it means the migration has finished and the previous migration of this volume succeeded. If the migration status is 'error', it means the migration has finished and the previous migration of this volume failed.
How to implement volume migration for a back-end driver
There are two kinds of implementations for the volume migration currently in Cinder.
The first is the generic host-assisted migration, which consists of two different transfer modes, block-based and file-based. This implementation is based on the volume attachment to the node of cinder volume service, c-vol node. Any back-end driver supporting iSCSI will be able to support the generic host-assisted migration for sure. The back-end driver without iSCSI supported needs to be tested to decide if it supports this kind of migration. The block-based transfer mode is done by 'dd' command, applying to drivers like LVM, Storwize, etc, and the file-based transfer mode is done by file copy, typically applying to the RBD driver.
The second is the driver specific migration. Since some storage back-ends have their special commands to copy the volume, Cinder also provides a way for them to implement in terms of their own internal commands to migrate.
If the volume is migrated between two nodes configured with the same storage back-end, the migration will be optimized by calling the method migrate_volume in the driver, if the driver provides an implementation for it to migrate the volume within the same back-end, and will fallback to the generic host-assisted migration provided in the manager, if no such implementation is found or this implementation is not applicable for this migration.
If your storage driver in Cinder provides iSCSI support, it should naturally work under the generic host-assisted migration, when --force-host-copy is set to True from the API request. Normally you do not need to change any code, unless you need to transfer the volume from your driver via a different way from the block-based transfer or the file-based transfer.
If your driver uses a network connection to communicate the block data itself, you can use file I/O to participate in migration. Please take the RBD driver as a reference for this implementation.
If you would like to implement a driver specific volume migration for your driver, the API method associated with the driver specific migration is the following admin only method:
migrate_volume(self, ctxt, volume, host)
If your driver is taken as the destination back-end for a generic host-assisted migration and your driver needs to update the volume model after a successful migration, you need to implement the following method for your driver:
update_migrated_volume(self, ctxt, volume, new_volume, original_volume_status)
Required methods
There is one mandatory method that needs to be implemented for the driver to implement the driver specific volume migration.
migrate_volume
Used to migrate the volume directly if source and destination are managed by same storage.
There is one optional method that could be implemented for the driver to implement the generic host-assisted migration.
update_migrated_volume
Used to return the key-value pairs to update the volume model after a successful migration. The key-value pairs returned are supposed to be the final values your driver would like to be in the volume model, if a migration is completed.
This method can be used in a generally wide range, but the most common use case covered in this method is to rename the back-end name to the original volume id in your driver to make sure that the back-end still keeps the same id or name as it is before the volume migration. For this use case, there are two important fields: _name_id and provider_location.
The field _name_id is used to map the cinder volume id and the back-end id or name. The default value is None, which means the cinder volume id is the same to the back-end id or name. If they are different, _name_id is used to saved the back-end id or name.
The field provider_location is used to save the export information, created by the volume attach. This field is optional, since some drivers support the export creation and some do not. It is the driver maintainer's responsibility to decide what this field needs to be.
If the back-end id or name is renamed successfully, this method can return {'_name_id': None, 'provider_location': None}. It is the choice for your driver to implement this method and decide what use cases should be covered.