libvirt: Create persistent mdevs

Related to blueprint persistent-mdevs

Change-Id: I7e1d10e66a260efd0a3f2d6522aeb246c7582178

doc/source/admin/virtual-gpu.rst

@@ -31,6 +31,61 @@ To enable virtual GPUs, follow the steps below:
 Enable GPU types (Compute)
 --------------------------
 
+#. Enable virtual functions on NVIDIA GPUs.
+
+   .. code-block:: bash
+
+      $ /usr/lib/nvidia/sriov-manage -e slot:bus:domain.function
+
+   For example, to enable the virtual functions for the GPU with
+   slot ``0000``, bus ``41``, domain ``00``, and function ``0``:
+
+   .. code-block:: bash
+
+      $ /usr/lib/nvidia/sriov-manage -e 0000:41:00.0
+
+   You may want to automate this process as it has to be done on each boot of
+   the host.
+
+   Given an example ``systemd`` template unit file named
+   ``nvidia-sriov-manage@.service``:
+
+   .. code-block:: text
+
+      [Unit]
+      After = nvidia-vgpu-mgr.service
+      After = nvidia-vgpud.service
+      Description = Enable Nvidia GPU virtual functions
+
+      [Service]
+      Type = oneshot
+      User = root
+      Group = root
+      ExecStart = /usr/lib/nvidia/sriov-manage -e %i
+      # Give a reasonable amount of time for the server to start up/shut down
+      TimeoutSec = 120
+      # This creates a specific slice which all services will operate from
+      #  The accounting options give us the ability to see resource usage
+      #  through the `systemd-cgtop` command.
+      Slice = system.slice
+      # Set Accounting
+      CPUAccounting = True
+      BlockIOAccounting = True
+      MemoryAccounting = True
+      TasksAccounting = True
+      RemainAfterExit = True
+      ExecStartPre = /usr/bin/sleep 30
+
+      [Install]
+      WantedBy = multi-user.target
+
+   To enable the virtual functions for the GPU with slot ``0000``, bus ``41``,
+   domain ``00``, and function ``0``:
+
+   .. code-block:: bash
+
+      $ systemctl enable nvidia-sriov-manage@0000:41:00.0.service
+
 #. Specify which specific GPU type(s) the instances would get.
 
    Edit :oslo.config:option:`devices.enabled_mdev_types`:

nova/tests/fixtures/libvirt.py

@@ -188,6 +188,8 @@ VIR_CONNECT_LIST_DOMAINS_INACTIVE = 2
 VIR_CONNECT_LIST_NODE_DEVICES_CAP_PCI_DEV = 2
 VIR_CONNECT_LIST_NODE_DEVICES_CAP_NET = 1 << 4
 VIR_CONNECT_LIST_NODE_DEVICES_CAP_VDPA = 1 << 17
+VIR_CONNECT_LIST_NODE_DEVICES_CAP_MDEV = 1 << 5
+VIR_CONNECT_LIST_NODE_DEVICES_INACTIVE = 1 << 8
 
 # secret type
 VIR_SECRET_USAGE_TYPE_NONE = 0
@@ -200,6 +202,12 @@ VIR_DOMAIN_METADATA_DESCRIPTION = 0
 VIR_DOMAIN_METADATA_TITLE = 1
 VIR_DOMAIN_METADATA_ELEMENT = 2
 
+# virNodeDeviceCreateXML flags
+VIR_NODE_DEVICE_CREATE_XML_VALIDATE = 4
+
+# virNodeDeviceDefineXML flags
+VIR_NODE_DEVICE_DEFINE_XML_VALIDATE = 5
+
 # Libvirt version to match MIN_LIBVIRT_VERSION in driver.py
 FAKE_LIBVIRT_VERSION = versionutils.convert_version_to_int(
     libvirt_driver.MIN_LIBVIRT_VERSION)
@@ -2126,6 +2134,12 @@ class Connection(object):
                     error_code=VIR_ERR_NO_NODE_DEVICE,
                     error_domain=VIR_FROM_NODEDEV)
 
+    def nodeDeviceCreateXML(self, xml, flags):
+        pass
+
+    def nodeDeviceDefineXML(self, xml, flags):
+        pass
+
     def listDevices(self, cap, flags):
         if cap == 'pci':
             return self.pci_info.get_all_devices()

nova/tests/functional/libvirt/test_vgpu.py

@@ -19,6 +19,7 @@ import os_resource_classes as orc
 from oslo_config import cfg
 from oslo_log import log as logging
 from oslo_utils import uuidutils
+from oslo_utils import versionutils
 
 from nova.compute import instance_actions
 import nova.conf
@@ -55,9 +56,16 @@ class VGPUTestBase(base.ServersTestBase):
             'used': 44,
             'free': 84,
         }
-        self.useFixture(fixtures.MockPatch(
-            'nova.privsep.libvirt.create_mdev',
-            side_effect=self._create_mdev))
+        # Persistent mdevs in libvirt >= 7.3.0
+        if self.FAKE_LIBVIRT_VERSION < versionutils.convert_version_to_int(
+                libvirt_driver.MIN_LIBVIRT_PERSISTENT_MDEV):
+            create_mdev_str = 'nova.privsep.libvirt.create_mdev'
+        else:
+            create_mdev_str = (
+                'nova.virt.libvirt.driver.LibvirtDriver._create_mdev')
+            self._create_mdev = self._create_mdev_7_3
+        self.useFixture(
+            fixtures.MockPatch(create_mdev_str, side_effect=self._create_mdev))
 
         # for the sake of resizing, we need to patch the two methods below
         self.useFixture(fixtures.MockPatch(
@@ -114,6 +122,23 @@ class VGPUTestBase(base.ServersTestBase):
                                                    parent=libvirt_parent)})
         return uuid
 
+    def _create_mdev_7_3(self, dev_name, mdev_type, uuid=None):
+        # We need to fake the newly created sysfs object by adding a new
+        # FakeMdevDevice in the existing persisted Connection object so
+        # when asking to get the existing mdevs, we would see it.
+        if not uuid:
+            uuid = uuidutils.generate_uuid()
+        mdev_name = libvirt_utils.mdev_uuid2name(uuid)
+        # Here, we get the right compute thanks by the self.current_host that
+        # was modified just before
+        connection = self.computes[
+            self._current_host].driver._host.get_connection()
+        connection.mdev_info.devices.update(
+            {mdev_name: fakelibvirt.FakeMdevDevice(dev_name=mdev_name,
+                                                   type_id=mdev_type,
+                                                   parent=dev_name)})
+        return uuid
+
     def start_compute_with_vgpu(self, hostname, pci_info=None):
         if not pci_info:
             pci_info = fakelibvirt.HostPCIDevicesInfo(
@@ -751,3 +776,10 @@ class DifferentMdevClassesTests(VGPUTestBase):
                                expected_rc='CUSTOM_NOTVGPU')
         self.assert_mdev_usage(self.compute2, expected_amount=1,
                                expected_rc='CUSTOM_NOTVGPU')
+
+
+class VGPUTestsLibvirt7_3(VGPUTests):
+
+    # Minimum version supporting persistent mdevs is 7.3.0.
+    # https://libvirt.org/drvnodedev.html#mediated-devices-mdevs
+    FAKE_LIBVIRT_VERSION = 7003000

nova/tests/functional/regressions/test_bug_1951656.py

@@ -11,37 +11,13 @@
 # License for the specific language governing permissions and limitations
 # under the License.
 
-from oslo_utils import uuidutils
-
-
 from nova.tests.fixtures import libvirt as fakelibvirt
 from nova.tests.functional.libvirt import test_vgpu
-from nova.virt.libvirt import utils as libvirt_utils
 
 
 class VGPUTestsLibvirt7_7(test_vgpu.VGPUTestBase):
 
-    def _create_mdev(self, physical_device, mdev_type, uuid=None):
-        # We need to fake the newly created sysfs object by adding a new
-        # FakeMdevDevice in the existing persisted Connection object so
-        # when asking to get the existing mdevs, we would see it.
-        if not uuid:
-            uuid = uuidutils.generate_uuid()
-        mdev_name = libvirt_utils.mdev_uuid2name(uuid)
-        libvirt_parent = self.pci2libvirt_address(physical_device)
-
-        # Libvirt 7.7 now creates mdevs with a parent_addr suffix.
-        new_mdev_name = '_'.join([mdev_name, libvirt_parent])
-
-        # Here, we get the right compute thanks by the self.current_host that
-        # was modified just before
-        connection = self.computes[
-            self._current_host].driver._host.get_connection()
-        connection.mdev_info.devices.update(
-            {mdev_name: fakelibvirt.FakeMdevDevice(dev_name=new_mdev_name,
-                                                   type_id=mdev_type,
-                                                   parent=libvirt_parent)})
-        return uuid
+    FAKE_LIBVIRT_VERSION = 7007000
 
     def setUp(self):
         super(VGPUTestsLibvirt7_7, self).setUp()

nova/tests/unit/virt/libvirt/test_config.py

@@ -3350,6 +3350,9 @@ class LibvirtConfigNodeDeviceTest(LibvirtConfigBaseTest):
         self.assertEqual(12, obj.mdev_information.iommu_group)
         self.assertIsNone(obj.mdev_information.uuid)
 
+        xmlout = obj.to_xml()
+        self.assertXmlEqual(xmlin, xmlout)
+
     def test_config_mdev_device_uuid(self):
         xmlin = """
         <device>
@@ -3375,6 +3378,19 @@ class LibvirtConfigNodeDeviceTest(LibvirtConfigBaseTest):
         self.assertEqual("b2107403-110c-45b0-af87-32cc91597b8a",
                          obj.mdev_information.uuid)
 
+        expected_xml = """
+        <device>
+          <name>mdev_b2107403_110c_45b0_af87_32cc91597b8a_0000_41_00_0</name>
+          <parent>pci_0000_41_00_0</parent>
+          <capability type='mdev'>
+            <type id='nvidia-442'/>
+            <uuid>b2107403-110c-45b0-af87-32cc91597b8a</uuid>
+            <iommuGroup number='57'/>
+          </capability>
+        </device>"""
+        xmlout = obj.to_xml()
+        self.assertXmlEqual(expected_xml, xmlout)
+
     def test_config_vdpa_device(self):
         xmlin = """
         <device>

nova/tests/unit/virt/libvirt/test_driver.py

@@ -27596,6 +27596,142 @@ class LibvirtDriverTestCase(test.NoDBTestCase, TraitsComparisonMixin):
         self.assertRaises(exception.InvalidLibvirtMdevConfig,
                           drvr.init_host, host='foo')
 
+    @mock.patch('oslo_utils.uuidutils.generate_uuid')
+    def test_create_mdev(self, mock_generate_uuid, uuid=None, drvr=None):
+        if drvr is None:
+            drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), True)
+            drvr._host = mock.Mock()
+
+        r = drvr._create_mdev(
+            mock.sentinel.dev_name, mock.sentinel.mdev_type, uuid=uuid)
+
+        drvr._host.device_create.assert_called_once()
+        dev_conf = drvr._host.device_create.call_args.args[0]
+        self.assertIsInstance(dev_conf, vconfig.LibvirtConfigNodeDevice)
+        self.assertEqual(mock.sentinel.dev_name, dev_conf.parent)
+        self.assertEqual(
+            mock.sentinel.mdev_type, dev_conf.mdev_information.type)
+        expected_uuid = uuid or mock_generate_uuid.return_value
+        self.assertEqual(expected_uuid, dev_conf.mdev_information.uuid)
+        drvr._host.device_define.assert_called_once_with(dev_conf)
+        drvr._host.device_set_autostart.assert_called_once_with(
+            drvr._host.device_define.return_value, autostart=True)
+        self.assertEqual(expected_uuid, r)
+
+    def test_create_mdev_with_uuid(self):
+        self.test_create_mdev(uuid=uuids.mdev)
+
+    @mock.patch('nova.virt.libvirt.driver.LOG.info')
+    def test_create_mdev_autostart_error(self, mock_log_info):
+        drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), True)
+        drvr._host = mock.Mock()
+        drvr._host.device_set_autostart.side_effect = test.TestingException(
+            'error')
+
+        self.test_create_mdev(uuid=uuids.mdev, drvr=drvr)
+
+        mock_log_info.assert_called_once_with(
+            'Failed to set autostart to True for mdev '
+            f'{drvr._host.device_define.return_value.name.return_value} with '
+            f'UUID {uuids.mdev}: error.')
+
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver,
+        '_register_all_undefined_instance_details', new=mock.Mock())
+    def test_start_inactive_mediated_devices_on_init_host(self):
+        drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), True)
+        device1 = mock.MagicMock()
+        device2 = mock.MagicMock()
+        drvr._host = mock.Mock()
+        drvr._host.list_all_devices.return_value = [device1, device2]
+
+        drvr.init_host(host='foo')
+
+        flags = (
+            fakelibvirt.VIR_CONNECT_LIST_NODE_DEVICES_CAP_MDEV |
+            fakelibvirt.VIR_CONNECT_LIST_NODE_DEVICES_INACTIVE)
+        drvr._host.list_all_devices.assert_called_once_with(flags)
+        self.assertEqual(
+            [mock.call(device1), mock.call(device2)],
+            drvr._host.device_start.mock_calls)
+
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver, '_get_all_assigned_mediated_devices',
+        new=mock.Mock(return_value={}))
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver, 'destroy', new=mock.Mock())
+    @mock.patch('oslo_utils.fileutils.ensure_tree', new=mock.Mock())
+    @mock.patch(
+        'nova.virt.libvirt.blockinfo.get_disk_info',
+        new=mock.Mock(return_value=mock.sentinel.disk_info))
+    @mock.patch.object(libvirt_driver.LibvirtDriver, '_allocate_mdevs')
+    @mock.patch('nova.objects.Instance.image_meta')
+    @mock.patch.object(libvirt_driver.LibvirtDriver, '_get_guest_xml')
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver, '_create_images_and_backing',
+        new=mock.Mock())
+    @mock.patch(
+        'oslo_service.loopingcall.FixedIntervalLoopingCall', new=mock.Mock())
+    def _test_hard_reboot_allocate_missing_mdevs(
+            self, mock_get_xml, mock_image_meta, mock_allocate_mdevs):
+        mock_compute = mock.Mock()
+        mock_compute.reportclient.get_allocations_for_consumer.return_value = (
+            mock.sentinel.allocations)
+        virtapi = manager.ComputeVirtAPI(mock_compute)
+        drvr = libvirt_driver.LibvirtDriver(virtapi, True)
+        ctxt = context.get_admin_context()
+        instance = objects.Instance(
+            uuid=uuids.instance,
+            system_metadata={},
+            image_ref=uuids.image,
+            flavor=objects.Flavor(extra_specs={'resources:VGPU': 1}))
+
+        drvr._hard_reboot(ctxt, instance, mock.sentinel.network_info)
+
+        (mock_compute.reportclient.get_allocations_for_consumer.
+            assert_called_once_with(ctxt, instance.uuid))
+        mock_allocate_mdevs.assert_called_once_with(mock.sentinel.allocations)
+        mock_get_xml.assert_called_once_with(
+            ctxt, instance, mock.sentinel.network_info,
+            mock.sentinel.disk_info, mock_image_meta, block_device_info=None,
+            mdevs=mock_allocate_mdevs.return_value, accel_info=None)
+
+        return ctxt, mock_get_xml, instance
+
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver, '_create_guest_with_network',
+        new=mock.Mock())
+    def test_hard_reboot_allocate_missing_mdevs(self):
+        # Test a scenario where the instance's flavor requests VGPU but for
+        # whatever reason (example: libvirt error raised after the domain was
+        # undefined) it is missing assigned mdevs.
+        self._test_hard_reboot_allocate_missing_mdevs()
+
+    @mock.patch.object(
+        libvirt_driver.LibvirtDriver, '_create_guest_with_network')
+    def test_hard_reboot_allocate_missing_mdevs_fail(self, mock_create_guest):
+        # Test the scenario where a libvirt error is raised the first time we
+        # try to create the guest after allocating missing mdevs.
+        err_msg = (
+            'error getting device from group 0: Input/output error '
+            'Verify all devices in group 0 are bound to vfio-<bus> or pci-stub'
+            'and not already in use')
+        error = fakelibvirt.make_libvirtError(
+            fakelibvirt.libvirtError, err_msg, error_message=err_msg,
+            error_code=fakelibvirt.VIR_ERR_INTERNAL_ERROR)
+        # First attempt to create the guest fails and the second succeeds.
+        mock_create_guest.side_effect = [error, None]
+
+        ctxt, mock_get_xml, instance = (
+            self._test_hard_reboot_allocate_missing_mdevs())
+
+        call = mock.call(
+            ctxt, mock_get_xml.return_value, instance,
+            mock.sentinel.network_info, None, vifs_already_plugged=True,
+            external_events=[])
+        # We should have tried to create the guest twice.
+        self.assertEqual([call, call], mock_create_guest.mock_calls)
+
     @mock.patch.object(libvirt_guest.Guest, 'detach_device')
     def _test_detach_mediated_devices(self, side_effect, detach_device):
 

nova/virt/libvirt/config.py

@@ -3367,6 +3367,16 @@ class LibvirtConfigNodeDevice(LibvirtConfigObject):
         self.vdpa_capability = None
         self.vpd_capability = None
 
+    def format_dom(self):
+        dev = super().format_dom()
+        if self.name:
+            dev.append(self._text_node('name', str(self.name)))
+        if self.parent:
+            dev.append(self._text_node('parent', str(self.parent)))
+        if self.mdev_information:
+            dev.append(self.mdev_information.format_dom())
+        return dev
+
     def parse_dom(self, xmldoc):
         super(LibvirtConfigNodeDevice, self).parse_dom(xmldoc)
 
@@ -3520,6 +3530,21 @@ class LibvirtConfigNodeDeviceMdevInformation(LibvirtConfigObject):
         self.iommu_group = None
         self.uuid = None
 
+    def format_dom(self):
+        dev = super().format_dom()
+        dev.set('type', 'mdev')
+        if self.type:
+            mdev_type = self._new_node('type')
+            mdev_type.set('id', self.type)
+            dev.append(mdev_type)
+        if self.uuid:
+            dev.append(self._text_node('uuid', self.uuid))
+        if self.iommu_group:
+            iommu_group = self._new_node('iommuGroup')
+            iommu_group.set('number', str(self.iommu_group))
+            dev.append(iommu_group)
+        return dev
+
     def parse_dom(self, xmldoc):
         super(LibvirtConfigNodeDeviceMdevInformation,
               self).parse_dom(xmldoc)

nova/virt/libvirt/driver.py

@@ -243,6 +243,14 @@ VGPU_RESOURCE_SEMAPHORE = 'vgpu_resources'
 MIN_MDEV_LIVEMIG_LIBVIRT_VERSION = (8, 6, 0)
 MIN_MDEV_LIVEMIG_QEMU_VERSION = (8, 1, 0)
 
+# Minimum version supporting persistent mdevs.
+# https://libvirt.org/drvnodedev.html#mediated-devices-mdevs
+MIN_LIBVIRT_PERSISTENT_MDEV = (7, 3, 0)
+
+# Autostart appears to be available starting in 7.8.0
+# https://github.com/libvirt/libvirt/commit/c6607a25b93bd6b0188405785d6608fdf71c8e0a
+MIN_LIBVIRT_NODEDEV_AUTOSTART = (7, 8, 0)
+
 LIBVIRT_PERF_EVENT_PREFIX = 'VIR_PERF_PARAM_'
 
 # Maxphysaddr minimal support version.
@@ -837,9 +845,17 @@ class LibvirtDriver(driver.ComputeDriver):
         # wrongly modified.
         libvirt_cpu.power_down_all_dedicated_cpus()
 
-        # TODO(sbauza): Remove this code once mediated devices are persisted
-        # across reboots.
-        self._recreate_assigned_mediated_devices()
+        if not self._host.has_min_version(MIN_LIBVIRT_PERSISTENT_MDEV):
+            # TODO(sbauza): Remove this code once mediated devices are
+            # persisted across reboots.
+            self._recreate_assigned_mediated_devices()
+        else:
+            # NOTE(melwitt): We shouldn't need to do this with libvirt 7.8.0
+            # and newer because we're setting autostart=True on the devices --
+            # but if that fails for whatever reason and any devices become
+            # inactive, we can start them here. With libvirt version < 7.8.0,
+            # this is needed because autostart is not available.
+            self._start_inactive_mediated_devices()
 
         self._check_cpu_compatibility()
 
@@ -1088,6 +1104,25 @@ class LibvirtDriver(driver.ComputeDriver):
 
         LOG.debug('Enabling emulated TPM support')
 
+    def _start_inactive_mediated_devices(self):
+        # Get a list of inactive mdevs so we can start them and make them
+        # active. We need to start inactive mdevs even if they are not
+        # currently assigned to instances because attempting to use an inactive
+        # mdev when booting a new instance, for example, will raise an error:
+        # libvirt.libvirtError: device not found: mediated device '<uuid>' not
+        # found.
+        # An inactive mdev is an mdev that is defined but not created.
+        flags = (
+            libvirt.VIR_CONNECT_LIST_NODE_DEVICES_CAP_MDEV |
+            libvirt.VIR_CONNECT_LIST_NODE_DEVICES_INACTIVE)
+        inactive_mdevs = self._host.list_all_devices(flags)
+        if inactive_mdevs:
+            names = [mdev.name() for mdev in inactive_mdevs]
+            LOG.info(f'Found inactive mdevs: {names}')
+        for mdev in inactive_mdevs:
+            LOG.info(f'Starting inactive mdev: {mdev.name()}')
+            self._host.device_start(mdev)
+
     @staticmethod
     def _is_existing_mdev(uuid):
         # FIXME(sbauza): Some kernel can have a uevent race meaning that the
@@ -4059,6 +4094,19 @@ class LibvirtDriver(driver.ComputeDriver):
                                             instance,
                                             instance.image_meta,
                                             block_device_info)
+        # NOTE(melwitt): It's possible that we lost track of the allocated
+        # mdevs of an instance if, for example, a libvirt error was encountered
+        # after the domain XML was undefined in a previous hard reboot.
+        # Try to get existing mdevs that are created but not assigned so they
+        # will be added into the generated domain XML.
+        if instance.flavor.extra_specs.get('resources:VGPU') and not mdevs:
+            LOG.info(
+                'The instance flavor requests VGPU but no mdevs are assigned '
+                'to the instance. Attempting to re-assign mdevs.',
+                instance=instance)
+            allocs = self.virtapi.reportclient.get_allocations_for_consumer(
+                    context, instance.uuid)
+            mdevs = self._allocate_mdevs(allocs)
         # NOTE(vish): This could generate the wrong device_format if we are
         #             using the raw backend and the images don't exist yet.
         #             The create_images_and_backing below doesn't properly
@@ -4110,10 +4158,32 @@ class LibvirtDriver(driver.ComputeDriver):
 
         # NOTE(efried): The instance should already have a vtpm_secret_uuid
         # registered if appropriate.
-        self._create_guest_with_network(
-            context, xml, instance, network_info, block_device_info,
-            vifs_already_plugged=vifs_already_plugged,
-            external_events=external_events)
+        try:
+            self._create_guest_with_network(
+                context, xml, instance, network_info, block_device_info,
+                vifs_already_plugged=vifs_already_plugged,
+                external_events=external_events)
+        except libvirt.libvirtError as e:
+            errcode = e.get_error_code()
+            errmsg = e.get_error_message()
+            # NOTE(melwitt): If we are reassigning mdevs, we might hit the
+            # following error on the first attempt to create the guest:
+            #   error getting device from group <group>: Input/output error
+            #   Verify all devices in group <group> are bound to vfio-<bus> or
+            #   pci-stub and not already in use
+            # Retry the guest creation once in this case as it usually succeeds
+            # on the second try.
+            if (mdevs and errcode == libvirt.VIR_ERR_INTERNAL_ERROR and
+                    'error getting device from group' in errmsg):
+                LOG.info(
+                    f'Encountered error {errmsg}, reattempting creation of '
+                    'the guest.', instance=instance)
+                self._create_guest_with_network(
+                    context, xml, instance, network_info, block_device_info,
+                    vifs_already_plugged=vifs_already_plugged,
+                    external_events=external_events)
+            else:
+                raise
 
         def _wait_for_reboot():
             """Called at an interval until the VM is running again."""
@@ -8726,6 +8796,33 @@ class LibvirtDriver(driver.ComputeDriver):
         LOG.info('Available mdevs at: %s.', available_mdevs)
         return available_mdevs
 
+    def _create_mdev(self, dev_name, mdev_type, uuid=None):
+        if uuid is None:
+            uuid = uuidutils.generate_uuid()
+        conf = vconfig.LibvirtConfigNodeDevice()
+        conf.parent = dev_name
+        conf.mdev_information = (
+            vconfig.LibvirtConfigNodeDeviceMdevInformation())
+        conf.mdev_information.type = mdev_type
+        conf.mdev_information.uuid = uuid
+        # Create the transient device.
+        self._host.device_create(conf)
+        # Define it to make it persistent.
+        mdev_dev = self._host.device_define(conf)
+        if self._host.has_min_version(MIN_LIBVIRT_NODEDEV_AUTOSTART):
+            # Set it to automatically start when the compute host boots or the
+            # parent device becomes available.
+            # NOTE(melwitt): Make this not fatal because we can try to manually
+            # start mdevs in init_host() if they didn't start automatically
+            # after a host reboot.
+            try:
+                self._host.device_set_autostart(mdev_dev, autostart=True)
+            except Exception as e:
+                LOG.info(
+                    'Failed to set autostart to True for mdev '
+                    f'{mdev_dev.name()} with UUID {uuid}: {str(e)}.')
+        return uuid
+
     def _create_new_mediated_device(self, parent, uuid=None):
         """Find a physical device that can support a new mediated device and
         create it.
@@ -8755,8 +8852,12 @@ class LibvirtDriver(driver.ComputeDriver):
                 # We need the PCI address, not the libvirt name
                 # The libvirt name is like 'pci_0000_84_00_0'
                 pci_addr = "{}:{}:{}.{}".format(*dev_name[4:].split('_'))
-                chosen_mdev = nova.privsep.libvirt.create_mdev(
-                    pci_addr, dev_supported_type, uuid=uuid)
+                if not self._host.has_min_version(MIN_LIBVIRT_PERSISTENT_MDEV):
+                    chosen_mdev = nova.privsep.libvirt.create_mdev(
+                        pci_addr, dev_supported_type, uuid=uuid)
+                else:
+                    chosen_mdev = self._create_mdev(
+                        dev_name, dev_supported_type, uuid=uuid)
                 LOG.info('Created mdev: %s on pGPU: %s.',
                          chosen_mdev, pci_addr)
                 return chosen_mdev

nova/virt/libvirt/host.py

@@ -1254,6 +1254,65 @@ class Host(object):
         """
         return self.get_connection().nodeDeviceLookupByName(name)
 
+    def device_create(self, conf, validate=False):
+        """Create a node device from specified device XML
+
+        This creates the device as transient.
+
+        :param conf: A LibvirtConfigObject of the device to create
+        :param validate: whether to validate the XML document against schema
+
+        :returns: a virNodeDevice instance if successful, else None
+        """
+        flag = libvirt.VIR_NODE_DEVICE_CREATE_XML_VALIDATE
+        flags = validate and flag or 0
+        device_xml = conf.to_xml()
+        return self.get_connection().nodeDeviceCreateXML(device_xml, flags)
+
+    def device_define(self, conf, validate=False):
+        """Define a node device from specified device XML
+
+        This defines the device to make it persistent.
+
+        :param conf: A LibvirtConfigObject of the device to create
+        :param validate: whether to validate the XML document against schema
+
+        :returns: a virNodeDevice instance if successful, else None
+        """
+        flag = libvirt.VIR_NODE_DEVICE_DEFINE_XML_VALIDATE
+        flags = validate and flag or 0
+        device_xml = conf.to_xml()
+        return self.get_connection().nodeDeviceDefineXML(device_xml, flags)
+
+    def device_start(self, dev):
+        """Start a defined node device
+
+        :param dev: The virNodeDevice instance to start
+        """
+        # extra flags; not used yet, so callers should always pass 0
+        #   https://libvirt.org/html/libvirt-libvirt-nodedev.html
+        flags = 0
+        result = dev.create(flags)
+        if result == -1:
+            msg = f'Failed to start node device {dev.name()}'
+            raise exception.InternalError(_(msg))
+
+    def device_set_autostart(self, dev, autostart=True):
+        """Set a node device to automatically start when the host boots
+
+        This can set whether the node device should automatically start when
+        the host machine boots or when the parent device becomes available.
+
+        :param dev: The virNodeDevice instance to set the autostart value
+        :param autostart: Whether to set the device to automatically start
+        """
+        result = dev.setAutostart(autostart=autostart)
+        if result == -1:
+            msg = (
+                f'Failed to set autostart to {autostart} for node device '
+                f'{dev.name()}')
+            raise exception.InternalError(_(msg))
+
     def _get_pcinet_info(
         self,
         dev: 'libvirt.virNodeDevice',
@@ -1611,7 +1670,7 @@ class Host(object):
         """Lookup devices.
 
         :param flags: a bitmask of flags to filter the returned devices.
-        :returns: a list of virNodeDevice xml strings.
+        :returns: a list of virNodeDevice instances.
         """
         try:
             return self.get_connection().listAllDevices(flags) or []

releasenotes/notes/persistent-mdevs-2a69e44baae9d5ca.yaml

@@ -0,0 +1,11 @@
+features:
+  - |
+    With the libvirt driver and libvirt version 7.3.0 or newer, mediated
+    devices for vGPUs are now persisted across reboots of a compute host.
+
+    .. note::
+
+       After a compute host reboots, the virtual functions for the GPU must be
+       enabled again before instances will be able to use their vGPUs.
+       Please see https://docs.openstack.org/nova/latest/admin/virtual-gpu.html
+       for more information.