whitebox-tempest-plugin/whitebox_tempest_plugin/api/compute/test_vgpu.py

# Copyright 2020 Red Hat Inc.
# All Rights Reserved.
#
#    Licensed under the Apache License, Version 2.0 (the "License"); you may
#    not use this file except in compliance with the License. You may obtain
#    a copy of the License at
#
#         http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
#    WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
#    License for the specific language governing permissions and limitations
#    under the License.

from tempest.common.utils.linux import remote_client
from tempest.common import waiters
from tempest import config
from tempest.lib.common.utils import data_utils
from tempest.lib import decorators
from whitebox_tempest_plugin.api.compute import base
from whitebox_tempest_plugin.services import clients

from oslo_log import log as logging

CONF = config.CONF
LOG = logging.getLogger(__name__)


class VGPUTest(base.BaseWhiteboxComputeTest):

    # NOTE(jparker) as of Queens all hypervisors that support vGPUs accept
    # a single vGPU per instance, so this value is not exposed as a whitebox
    # hardware configurable at this time.
    vgpu_amount_per_instance = 1

    @classmethod
    def skip_checks(cls):
        super(VGPUTest, cls).skip_checks()
        if (CONF.whitebox_hardware.vgpu_vendor_id is None):
            msg = "CONF.whitebox_hardware.vgpu_vendor_id needs to be set."
            raise cls.skipException(msg)

    @classmethod
    def resource_setup(cls):
        # NOTE(jparker) Currently the inheritance tree for Whitebox test
        # classes for the test method create_flavor() does not resolve to a
        # classmethod for Whitebox. resource_setup expects setup methods to be
        # classmethods, so directly calling the flavors client to create the
        # necessary flavor for the tests
        super(VGPUTest, cls).resource_setup()
        flavor_name = data_utils.rand_name('vgpu_test_flavor')
        extra_specs = {"resources:VGPU": str(cls.vgpu_amount_per_instance)}
        cls.vgpu_flavor = cls.admin_flavors_client.create_flavor(
            name=flavor_name,
            ram=CONF.whitebox.flavor_ram_size,
            vcpus=2,
            disk=CONF.whitebox.flavor_volume_size,
            is_public='True')['flavor']
        cls.os_admin.flavors_client.set_flavor_extra_spec(
            cls.vgpu_flavor['id'],
            **extra_specs)

        cls.addClassResourceCleanup(
            cls.admin_flavors_client.wait_for_resource_deletion,
            cls.vgpu_flavor['id'])
        cls.addClassResourceCleanup(cls.admin_flavors_client.delete_flavor,
                                    cls.vgpu_flavor['id'])

    def _create_ssh_client(self, server, validation_resources):
        """Create an ssh client to execute commands on the guest instance

        :param server: the ssh client will be setup to interface with the
        provided server instance
        :param valdiation_resources: necessary validation information to setup
        an ssh session
        :return linux_client: the ssh client that allows for guest command
        execution
        """
        linux_client = remote_client.RemoteClient(
            self.get_server_ip(server, validation_resources),
            self.image_ssh_user,
            self.image_ssh_password,
            validation_resources['keypair']['private_key'],
            server=server,
            servers_client=self.servers_client)
        linux_client.validate_authentication()
        return linux_client

    @classmethod
    def setup_credentials(cls):
        cls.prepare_instance_network()
        super(VGPUTest, cls).setup_credentials()

    def _get_rp_uuid_from_hostname(self, hostname):
        """Given a provided compute host return its associated rp uuid

        :param hostname: str, compute hostname to check
        :return parent_rp_uuid: str, string representation of the rp uuid
        found on the compute host
        """
        resp = self.os_admin.resource_providers_client.list_resource_providers(
            name=hostname)
        return resp.get('resource_providers')[0].get('uuid')

    def _get_all_children_of_resource_provider(self, rp_uuid):
        """List all child RP UUIDs of provided resource provider

        Given a parent resource provider uuid, get all in-tree child RP UUID
        that can provide the request resource amount
        API Reference:
        https://docs.openstack.org/api-ref/placement/#resource-providers

        :param rp_uuid: str, string representation of rp uuid to be searched
        :return rp_children_uuids: list of str, all rp uuids that match the
        resource=amount request
        """
        params = {'in_tree': rp_uuid}
        resp = self.os_admin.resource_providers_client.list_resource_providers(
            **params)
        # Create a list of uuids based on the return response of resource
        # providers from the rp client, exclude the parent uuid from this
        # list.
        child_uuids = [x.get('uuid') for x in resp.get('resource_providers')
                       if x != rp_uuid]
        return child_uuids

    def _get_vgpu_inventories_for_deployment(self):
        """Gathers the vGPU inventories from deployment

        :return inventories: dict, a dictionary mapping every rp_uuid with its
        current vGPU inventory
        """
        hostnames = self.list_compute_hosts()
        inventories = {}
        for hostname in hostnames:
            rp_uuid = self._get_rp_uuid_from_hostname(hostname)
            rp_children = self._get_all_children_of_resource_provider(
                rp_uuid=rp_uuid)
            for rp_uuid in rp_children:
                inventory = self.os_admin.resource_providers_client.\
                    list_resource_provider_inventories(rp_uuid)
                if 'VGPU' in [*inventory.get('inventories').keys()]:
                    inventories[rp_uuid] = inventory.get('inventories')
        return inventories

    def _get_usage_for_resource_class_vgpu(self, rp_uuids):
        """Total usage of resource class vGPU from provided list of RP UUIDs

        :param rp_uuids: list, comprised of str representing all RP UUIDs to
        query
        :return total_vgpu_usage: int, total usage of resource class VGPU from
        all provided RP UUIDS
        """
        total_vgpu_usage = 0
        for rp_uuid in rp_uuids:
            resp = self.os_admin.resource_providers_client.\
                list_resource_provider_usages(rp_uuid=rp_uuid)
            rp_usages = resp.get('usages')
            vgpu_usage = rp_usages.get('VGPU', 0)
            total_vgpu_usage += vgpu_usage
        return total_vgpu_usage

    def _get_vgpu_util_for_host(self, hostname):
        """Get the total usage of a vGPU resource class from the compute host

        :param hostname: str, compute hostname to gather usage data from
        :return resource_usage_count: int, the current total usage for the vGPU
        resource class
        """
        rp_uuid = self._get_rp_uuid_from_hostname(hostname)
        rp_children = self._get_all_children_of_resource_provider(
            rp_uuid=rp_uuid)
        resource_usage_count = \
            self._get_usage_for_resource_class_vgpu(rp_children)
        return resource_usage_count

    def _assert_vendor_id_in_guest(self, linux_client, expected_device_count):
        """Confirm vgpu vendor id is present in server instance sysfs

        :param pci_address: str when searching the guest's pci devices use
        provided pci_address value to parse for vendor id
        """
        cmd = "grep -nr 0x%s /sys/bus/pci/devices/*/vendor | grep -E " \
              "'[0-9a-e]{4}:[0-9a-e]{2}:[0-9a-e]{2}.[0-9a-e]{1}' | wc -l" % \
              CONF.whitebox_hardware.vgpu_vendor_id
        sys_out = linux_client.exec_command(cmd)
        self.assertIsNotNone(
            sys_out, 'Unable to find vendor id %s when checking the guest' %
            CONF.whitebox_hardware.vgpu_vendor_id)
        self.assertEqual(
            expected_device_count, int(sys_out), 'Should only find %s vGPU '
            'device in guest but instead found %s' %
            (expected_device_count, int(sys_out)))

    def _cold_migrate_server(self, server_id, target_host, revert=False):
        """Cold migrate a server with the option to revert the migration

        :param server_id: str, uuid of the server to migrate
        :param revert: bool, revert server migration action if true
        """
        src_host = self.get_host_for_server(server_id)
        self.admin_servers_client.migrate_server(server_id, host=target_host)
        waiters.wait_for_server_status(self.servers_client, server_id,
                                       'VERIFY_RESIZE')

        if revert:
            self.admin_servers_client.revert_resize_server(server_id)
            assert_func = self.assertEqual
        else:
            self.admin_servers_client.confirm_resize_server(server_id)
            assert_func = self.assertNotEqual

        waiters.wait_for_server_status(self.servers_client,
                                       server_id, 'ACTIVE')
        dst_host = self.get_host_for_server(server_id)
        assert_func(src_host, dst_host)

    def _validate_vgpu_instance(self, server, linux_client,
                                expected_device_count):
        """Confirm vgpu guest XML is correct and vendor id is present in guest

        :param server: dict, attributes describing the guest instance
        :param linux_client: ssh client capable of interacting with guest
        :param expected_device_count: int, expected number of XML vgpu devices
        in the guest
        """
        # Find all hostdev devices on the instance of type mdev in the provided
        # instance's XML
        vgpu_devices = self.get_server_xml(server['id']).findall(
            "./devices/hostdev[@type='mdev']"
        )

        # Validate the number of mdev host devices is equal to the expected
        # count provided to the method
        self.assertEqual(
            expected_device_count, len(vgpu_devices), "Expected %d "
            "xml hostdev vgpu element(s) on instance %s but instead found %d" %
            (expected_device_count, server['id'],
             len(vgpu_devices)))

        # If there are no expected mdev devices, additional verification of
        # the vgpu device is not necessary
        if expected_device_count == 0:
            return

        # Validate the vendor id is present in guest instance
        self._assert_vendor_id_in_guest(linux_client, expected_device_count)

    def create_validateable_instance(self, flavor, validation_resources):
        """Create a validateable instance based on provided flavor

        :param flavor: dict, attributes describing flavor
        :param validation_resources: dict, parameters necessary to setup ssh
        client and validate the guest
        """
        server = self.create_test_server(
            flavor=flavor['id'],
            validatable=True,
            validation_resources=validation_resources,
            wait_until='ACTIVE')

        return server

    def _validate_final_vgpu_rp_inventory(
            self, starting_rp_vgpu_inventory, end_rp_vgpu_inventory):
        """Validate that starting rp vgpu inventory matches the inventory after
        testing

        :param starting_rp_vgpu_inventory dict, vgpu resource provider
        inventories before test execution
        :param end_rp_vgpu_inventory dict, vgpu resource provider inventories
        after test execution
        """
        for rp_uuid in starting_rp_vgpu_inventory.keys():
            self.assertEqual(
                starting_rp_vgpu_inventory.get(rp_uuid),
                end_rp_vgpu_inventory.get(rp_uuid), 'The inventories have '
                'changed for RP %s between the start %s and the end %s' %
                (rp_uuid, starting_rp_vgpu_inventory.get(rp_uuid),
                 end_rp_vgpu_inventory.get(rp_uuid)))


class VGPUSmoke(VGPUTest):
    def test_boot_instance_with_vgpu(self):
        """Test creating an instance with a vGPU resource"""
        # Confirm vGPU guest XML contains correct number of vgpu devices. Then
        # confirm the vgpu vendor id is present in the sysfs for the guest
        starting_rp_vgpu_inventory = \
            self._get_vgpu_inventories_for_deployment()
        validation_resources = self.get_test_validation_resources(
            self.os_primary)
        server = self.create_validateable_instance(
            self.vgpu_flavor,
            validation_resources)
        linux_client = self._create_ssh_client(server, validation_resources)
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)

        # Delete the guest and confirm the inventory reverts back to the
        # original amount
        self.os_admin.servers_client.delete_server(server['id'])
        waiters.wait_for_server_termination(
            self.os_admin.servers_client, server['id'])
        end_rp_vgpu_inventory = self._get_vgpu_inventories_for_deployment()
        self._validate_final_vgpu_rp_inventory(
            starting_rp_vgpu_inventory, end_rp_vgpu_inventory)


class VGPUColdMigration(VGPUTest):

    # Requires at least placement microversion 1.14 in order search through
    # nested resources providers via the 'in_tree=<UUID>' parameter
    placement_min_microversion = '1.14'
    placement_max_microversion = 'latest'

    @classmethod
    def skip_checks(cls):
        super(VGPUColdMigration, cls).skip_checks()
        if CONF.compute.min_compute_nodes < 2:
            msg = "Need two or more compute nodes to execute cold migration"
            raise cls.skipException(msg)
        if not CONF.compute_feature_enabled.vgpu_cold_migration_supported:
            msg = "vGPU Cold Migration support needed in order to run tests"
            raise cls.skipException(msg)

    def test_vgpu_cold_migration(self):
        starting_rp_vgpu_inventory = \
            self._get_vgpu_inventories_for_deployment()
        validation_resources = self.get_test_validation_resources(
            self.os_primary)
        server = self.create_validateable_instance(
            self.vgpu_flavor, validation_resources)
        linux_client = self._create_ssh_client(server, validation_resources)

        # Determine the host the vGPU enabled guest is currently on. Next
        # get another potential compute host to serve as the migration target
        src_host = self.get_host_for_server(server['id'])
        dest_host = self.get_host_other_than(server['id'])

        # Get the current VGPU usage from the resource providers on
        # the source and destination compute hosts.
        pre_src_usage = self._get_vgpu_util_for_host(src_host)
        pre_dest_usage = self._get_vgpu_util_for_host(dest_host)

        # Cold migrate the the instance to the target host
        self._cold_migrate_server(server['id'],
                                  target_host=dest_host,
                                  revert=False)

        LOG.info('Guest %(server)s was just cold migrated to %(dest_host)s, '
                 'guest will now be validated after operation',
                 {'server': server['id'], 'dest_host': dest_host})
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)

        # Regather the VGPU resource usage on both compute hosts involved in
        # the cold migration. Confirm the original source host's VGPU usage has
        # updated to no longer report original usage for the vGPU resource and
        # the destination is now accounting for the resource.
        post_src_usage = self._get_vgpu_util_for_host(src_host)
        post_dest_usage = self._get_vgpu_util_for_host(dest_host)
        expected_src_usage = pre_src_usage - self.vgpu_amount_per_instance
        self.assertEqual(
            expected_src_usage,
            post_src_usage, 'After migration, host %s expected to have %s '
            'usage for resource class VGPU but instead found %d' %
            (src_host, expected_src_usage, post_src_usage))
        expected_dest_usage = pre_dest_usage + self.vgpu_amount_per_instance
        self.assertEqual(
            expected_dest_usage, post_dest_usage, 'After migration, Host '
            '%s expected to have resource class VGPU usage totaling %d but '
            'instead found %d' %
            (dest_host, expected_dest_usage, post_dest_usage))

        # Delete the guest and confirm the inventory reverts back to the
        # original amount
        self.os_admin.servers_client.delete_server(server['id'])
        waiters.wait_for_server_termination(
            self.os_admin.servers_client, server['id'])
        end_rp_vgpu_inventory = self._get_vgpu_inventories_for_deployment()
        self._validate_final_vgpu_rp_inventory(
            starting_rp_vgpu_inventory, end_rp_vgpu_inventory)

    def test_revert_vgpu_cold_migration(self):
        starting_rp_vgpu_inventory = \
            self._get_vgpu_inventories_for_deployment()
        validation_resources = self.get_test_validation_resources(
            self.os_primary)
        server = self.create_validateable_instance(
            self.vgpu_flavor,
            validation_resources)
        linux_client = self._create_ssh_client(server, validation_resources)

        # Determine the host the vGPU enabled guest is currently on. Next
        # get another potential compute host to serve as the migration target
        src_host = self.get_host_for_server(server['id'])
        dest_host = self.get_host_other_than(server['id'])

        # Get the current VGPU usage from the resource providers on
        # the source and destination compute hosts.
        pre_src_usage = self._get_vgpu_util_for_host(src_host)
        pre_dest_usage = self._get_vgpu_util_for_host(dest_host)

        # Cold migrate the the instance to the target host
        self._cold_migrate_server(server['id'],
                                  target_host=dest_host,
                                  revert=True)

        # Sanity check the guest, confirming the vgpu XML device is present
        # and the vendor id is present in the sysfs
        LOG.info(
            'Cold migration of guest %(server)s has been reverted back to '
            '%(src_host)s, vGPU guest will now be validated after revert '
            'operation', {'server': server['id'], 'src_host': src_host})
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)

        # Regather the VGPU resource usage on both compute hosts involved in
        # the cold migration. Due to the migration revert, confirm the target
        # host is not reporting vGPU usage for the instance and the source
        # host accurately reports current usage based on flavor request
        current_src_usage = self._get_vgpu_util_for_host(src_host)
        current_dest_usage = self._get_vgpu_util_for_host(dest_host)
        self.assertEqual(
            pre_dest_usage, current_dest_usage, 'After migration revert, host '
            '%s expected to not have any usage for resource class VGPU but '
            'instead found %d' % (dest_host, current_dest_usage))
        self.assertEqual(
            pre_src_usage, current_src_usage, 'After migration revert '
            'host %s expected to have resource class VGPU usage totaling %d '
            'but instead found %d' %
            (src_host, pre_src_usage, current_dest_usage))

        # Do a final sanity check of the guest after the revert to confirm the
        # vgpu device is present in the XML and vendor id is present in sysfs
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)

        # Delete the guest and confirm the inventory reverts back to the
        # original amount
        self.os_admin.servers_client.delete_server(server['id'])
        waiters.wait_for_server_termination(
            self.os_admin.servers_client, server['id'])
        end_rp_vgpu_inventory = self._get_vgpu_inventories_for_deployment()
        self._validate_final_vgpu_rp_inventory(
            starting_rp_vgpu_inventory, end_rp_vgpu_inventory)


class VGPUResizeInstance(VGPUTest):

    # Requires at least placement microversion 1.14 in order search through
    # nested resources providers via the 'in_tree=<UUID>' parameter
    placement_min_microversion = '1.14'
    placement_max_microversion = 'latest'

    @classmethod
    def skip_checks(cls):
        super(VGPUResizeInstance, cls).skip_checks()
        if not CONF.compute_feature_enabled.vgpu_cold_migration_supported:
            msg = "vGPU Cold Migration support needed in order to run " \
                  "resize tests"
            raise cls.skipException(msg)
        if not CONF.compute_feature_enabled.resize:
            msg = 'Resize not available.'
            raise cls.skipException(msg)

    def test_vgpu_to_standard_resize(self):
        # Create a vGPU instance and get the vGPU resource utilization from
        # its compute host
        server = self.create_test_server(flavor=self.vgpu_flavor['id'],
                                         wait_until='ACTIVE')
        host = self.get_host_for_server(server['id'])
        pre_resize_usage = self._get_vgpu_util_for_host(host)
        standard_flavor = self.create_flavor()
        self.resize_server(server['id'], standard_flavor['id'])

        # Check the guest's XML and confirm the mdev device is no longer
        # present
        self._validate_vgpu_instance(
            server,
            linux_client=None,
            expected_device_count=0)

        if CONF.compute_feature_enabled.console_output:
            # Confirm there are no errors when interacting with the guest
            # after it was resized from vgpu to standard
            self.servers_client.get_console_output(server['id'])

        # Gather the vGPU resource utilization from the compute host. The
        # instance will either land on a new compute host or remain on
        # the same source host but will be resized to a standard flavor.
        # In either action the source host should always report a vGPU usage
        # that is less than what the guest was originally utilizing
        post_resize_usage = self._get_vgpu_util_for_host(host)
        expected_usage = pre_resize_usage - self.vgpu_amount_per_instance

        # Confirm the original host's vGPU resource usage now accounts for the
        # guest resizing to a flavor without any vGPU resources
        self.assertEqual(
            expected_usage, post_resize_usage, 'After guest resize, host '
            '%s should be reporting total vGPU usage of %d, but instead is'
            'reporting %d' % (host, expected_usage, post_resize_usage))

    def test_standard_to_vgpu_resize(self):
        # Create a standard instance and then resize the instance to a flavor
        # that uses a vGPU resource
        validation_resources = self.get_test_validation_resources(
            self.os_primary)
        standard_flavor = self.create_flavor()
        server = self.create_validateable_instance(
            standard_flavor,
            validation_resources)
        linux_client = self._create_ssh_client(server, validation_resources)
        self.resize_server(server['id'], self.vgpu_flavor['id'])

        # Check the guest's XML and confirm that the correct number of vGPU
        # devices are present and the devices are present in the guest sysfs
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)


class VGPUMultiTypes(VGPUTest):

    @classmethod
    def skip_checks(cls):
        super(VGPUMultiTypes, cls).skip_checks()
        if not CONF.whitebox_hardware.vgpu_type_mapping:
            msg = "Requires custom vgpu trait to subsystem product id mapping"
            raise cls.skipException(msg)

    def _get_vgpu_mdev_type_on_guest(self, server_id):
        """Find the mdev_type of the vgpu device for the provided server id

        :param server_id: str, the guest uuid to check
        :return mdev_type: str, the mdev_type of vgpu provided to guest
        """
        host = self.get_host_for_server(server_id)
        ssh_client = clients.SSHClient(host)
        root = self.get_server_xml(server_id)
        vgpu_dev = root.find("./devices/hostdev[@type='mdev']")
        source_element = vgpu_dev.find(".source/address")
        mdev_uuid = source_element.get('uuid')

        cmd = "ls -al /sys/bus/mdev/devices/%s/ | awk -F'/' " \
              "'/mdev_type/ {print $3}'" % mdev_uuid
        mdev_type = ssh_client.execute(cmd)
        return mdev_type.rstrip()

    def _generate_vgpu_custom_trait_flavor(self, trait):
        """Creates a vGPU flavor that requests a specific type

        :param trait str, the custom trait used by the flavor to distinguish
        a specific vGPU type to request.
        """
        flavor_name = data_utils.rand_name(
            'vgpu_test_flavor_{}'.format(trait))
        extra_specs = {
            "resources:VGPU": str(self.vgpu_amount_per_instance),
            "trait:{}".format(trait): 'required'}
        vgpu_flavor = self.admin_flavors_client.create_flavor(
            name=flavor_name,
            ram=CONF.whitebox.flavor_ram_size,
            vcpus=2,
            disk=CONF.whitebox.flavor_volume_size,
            is_public='True',)['flavor']
        self.os_admin.flavors_client.set_flavor_extra_spec(
            vgpu_flavor['id'],
            **extra_specs)
        self.addClassResourceCleanup(
            self.admin_flavors_client.wait_for_resource_deletion,
            vgpu_flavor['id'])
        self.addClassResourceCleanup(self.admin_flavors_client.delete_flavor,
                                     vgpu_flavor['id'])
        return vgpu_flavor

    def test_deploy_multiple_vgpu_types(self):
        """Test creating multiple guests with different vGPU types

        1. Create multiple guests by iterating over the vgpu_type_mapping
        provided in tempest.conf.
        2. For each iteration create a flavor using a CUSTOM_TRAIT from
        vgpu_type_mapping and then launch an instance with the flavor.
        3. Once the guest is ACTIVE, ssh into the guest's respective compute
        host. From there check the uuid of the vgpu device assigned to the
        guest. Using the uuid, check mdev type of the host via
        /sys/bus/mdev/devices/<uuid> and confirm the mdev_type there matches
        the expected mdev_type provided to the guest.
        """
        custom_traits = CONF.whitebox_hardware.vgpu_type_mapping
        for trait, expected_mdev_type in custom_traits.items():
            vgpu_flavor = self._generate_vgpu_custom_trait_flavor(trait)
            validation_resources = self.get_test_validation_resources(
                self.os_primary)
            server = self.create_validateable_instance(
                vgpu_flavor,
                validation_resources)
            found_mdev_type = self._get_vgpu_mdev_type_on_guest(server['id'])
            self.assertEqual(
                expected_mdev_type, found_mdev_type,
                "The found mdev_type %s does not match the expected mdev_type "
                "%s for %s" % (found_mdev_type, expected_mdev_type, trait))


@decorators.skip_because(bug='1874664')
class VGPUServerEvacuation(VGPUTest):

    min_microversion = '2.95'

    @classmethod
    def skip_checks(cls):
        super(VGPUServerEvacuation, cls).skip_checks()
        if CONF.compute.min_compute_nodes < 2:
            msg = "Need two or more compute nodes to execute evacuate."
            raise cls.skipException(msg)

    def test_evacuate_server_having_vgpu(self):
        starting_rp_vgpu_inventory = \
            self._get_vgpu_inventories_for_deployment()
        validation_resources = self.get_test_validation_resources(
            self.os_primary)
        server = self.create_validateable_instance(
            self.vgpu_flavor, validation_resources)
        linux_client = self._create_ssh_client(server, validation_resources)

        # Determine the host the vGPU enabled guest is currently on. Next
        # get another potential compute host to serve as the migration target
        src_host = self.get_host_for_server(server['id'])
        dest_host = self.get_host_other_than(server['id'])

        # Get the current VGPU usage from the resource providers on
        # the source and destination compute hosts.
        pre_src_usage = self._get_vgpu_util_for_host(src_host)
        pre_dest_usage = self._get_vgpu_util_for_host(dest_host)

        host_a_svc = clients.NovaServiceManager(
            src_host, 'nova-compute', self.os_admin.services_client)

        with host_a_svc.stopped():
            self.shutdown_server_domain(server, src_host)
            self.evacuate_server(server['id'])

        self.assertEqual(self.get_host_for_server(server['id']), dest_host)

        LOG.info('Guest %(server)s was just evacuated to %(dest_host)s, '
                 'guest will now be validated after operation',
                 {'server': server['id'], 'dest_host': dest_host})
        self._validate_vgpu_instance(
            server,
            linux_client=linux_client,
            expected_device_count=self.vgpu_amount_per_instance)

        # Regather the VGPU resource usage on both compute hosts involved.
        # Confirm the original source host's VGPU usage has
        # updated to no longer report original usage for the vGPU resource and
        # the destination is now accounting for the resource.
        post_src_usage = self._get_vgpu_util_for_host(src_host)
        post_dest_usage = self._get_vgpu_util_for_host(dest_host)
        expected_src_usage = pre_src_usage - self.vgpu_amount_per_instance
        self.assertEqual(
            expected_src_usage,
            post_src_usage, 'After evacuation, host %s expected to have %s '
            'usage for resource class VGPU but instead found %d' %
            (src_host, expected_src_usage, post_src_usage))
        expected_dest_usage = pre_dest_usage + self.vgpu_amount_per_instance
        self.assertEqual(
            expected_dest_usage, post_dest_usage, 'After evacuation, Host '
            '%s expected to have resource class VGPU usage totaling %d but '
            'instead found %d' %
            (dest_host, expected_dest_usage, post_dest_usage))

        # Delete the guest and confirm the inventory reverts back to the
        # original amount
        self.os_admin.servers_client.delete_server(server['id'])
        waiters.wait_for_server_termination(
            self.os_admin.servers_client, server['id'])
        end_rp_vgpu_inventory = self._get_vgpu_inventories_for_deployment()
        self._validate_final_vgpu_rp_inventory(
            starting_rp_vgpu_inventory, end_rp_vgpu_inventory)