Merge "Enable cpus when an instance is spawning"

9 files changed, 625 insertions, 0 deletions

diff --git a/nova/conf/libvirt.py b/nova/conf/libvirt.py
index e32452c4ce..204fe5c4b8 100644
--- a/nova/conf/libvirt.py
+++ b/nova/conf/libvirt.py
@@ -1482,6 +1482,11 @@ libvirt_cpu_mgmt_opts = [
     cfg.BoolOpt('cpu_power_management',
                 default=False,
                 help='Use libvirt to manage CPU cores performance.'),
+    cfg.StrOpt('cpu_power_management_strategy',
+               choices=['cpu_state', 'governor'],
+               default='cpu_state',
+               help='Tuning strategy to reduce CPU power consumption when '
+                    'unused'),
     cfg.StrOpt('cpu_power_governor_low',
                default='powersave',
                help='Governor to use in order '
diff --git a/nova/tests/fixtures/__init__.py b/nova/tests/fixtures/__init__.py
index df254608fd..9ff4a2a601 100644
--- a/nova/tests/fixtures/__init__.py
+++ b/nova/tests/fixtures/__init__.py
@@ -16,6 +16,8 @@ from .cast_as_call import CastAsCallFixture  # noqa: F401
 from .cinder import CinderFixture  # noqa: F401
 from .conf import ConfFixture  # noqa: F401, F403
 from .cyborg import CyborgFixture  # noqa: F401
+from .filesystem import SysFileSystemFixture  # noqa: F401
+from .filesystem import TempFileSystemFixture  # noqa: F401
 from .glance import GlanceFixture  # noqa: F401
 from .libvirt import LibvirtFixture  # noqa: F401
 from .libvirt_imagebackend import LibvirtImageBackendFixture  # noqa: F401
diff --git a/nova/tests/fixtures/filesystem.py b/nova/tests/fixtures/filesystem.py
new file mode 100644
index 0000000000..932d42fe27
--- /dev/null
+++ b/nova/tests/fixtures/filesystem.py
@@ -0,0 +1,81 @@
+# 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.
+
+import os
+import shutil
+import tempfile
+from unittest import mock
+
+import fixtures
+
+from nova import filesystem
+from nova.virt.libvirt.cpu import core
+
+
+SYS = 'sys'
+
+
+class TempFileSystemFixture(fixtures.Fixture):
+    """Creates a fake / filesystem"""
+
+    def _setUp(self):
+        self.temp_dir = tempfile.TemporaryDirectory(prefix='fake_fs')
+        # NOTE(sbauza): I/O disk errors may raise an exception here, as we
+        # don't ignore them. If that's causing a problem in our CI jobs, the
+        # recommended solution is to use shutil.rmtree instead of cleanup()
+        # with ignore_errors parameter set to True (or wait for the minimum
+        # python version to be 3.10 as TemporaryDirectory will provide
+        # ignore_cleanup_errors parameter)
+        self.addCleanup(self.temp_dir.cleanup)
+
+
+class SysFileSystemFixture(TempFileSystemFixture):
+    """Creates a fake /sys filesystem"""
+
+    def __init__(self, cpus_supported=None):
+        self.cpus_supported = cpus_supported or 10
+
+    def _setUp(self):
+        super()._setUp()
+        self.sys_path = os.path.join(self.temp_dir.name, SYS)
+        self.addCleanup(shutil.rmtree, self.sys_path, ignore_errors=True)
+
+        sys_patcher = mock.patch(
+            'nova.filesystem.SYS',
+            new_callable=mock.PropertyMock(return_value=self.sys_path))
+        self.sys_mock = sys_patcher.start()
+        self.addCleanup(sys_patcher.stop)
+
+        avail_path_patcher = mock.patch(
+            'nova.virt.libvirt.cpu.core.AVAILABLE_PATH',
+            new_callable=mock.PropertyMock(
+                return_value=os.path.join(self.sys_path,
+                                          'devices/system/cpu/present')))
+        self.avail_path_mock = avail_path_patcher.start()
+        self.addCleanup(avail_path_patcher.stop)
+
+        cpu_path_patcher = mock.patch(
+            'nova.virt.libvirt.cpu.core.CPU_PATH_TEMPLATE',
+            new_callable=mock.PropertyMock(
+                return_value=os.path.join(self.sys_path,
+                                          'devices/system/cpu/cpu%(core)s')))
+        self.cpu_path_mock = cpu_path_patcher.start()
+        self.addCleanup(cpu_path_patcher.stop)
+
+        for cpu_nr in range(self.cpus_supported):
+            cpu_dir = os.path.join(self.cpu_path_mock % {'core': cpu_nr})
+            os.makedirs(os.path.join(cpu_dir, 'cpufreq'))
+            filesystem.write_sys(
+                os.path.join(cpu_dir, 'cpufreq/scaling_governor'),
+                data='powersave')
+        filesystem.write_sys(core.AVAILABLE_PATH,
+                             f'0-{self.cpus_supported - 1}')
diff --git a/nova/tests/functional/libvirt/test_power_manage.py b/nova/tests/functional/libvirt/test_power_manage.py
new file mode 100644
index 0000000000..fb1ac7d0cd
--- /dev/null
+++ b/nova/tests/functional/libvirt/test_power_manage.py
@@ -0,0 +1,270 @@
+#    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 unittest import mock
+
+import fixtures
+
+from nova import context as nova_context
+from nova import exception
+from nova import objects
+from nova.tests import fixtures as nova_fixtures
+from nova.tests.fixtures import libvirt as fakelibvirt
+from nova.tests.functional.libvirt import base
+from nova.virt import hardware
+from nova.virt.libvirt import cpu
+
+
+class PowerManagementTestsBase(base.ServersTestBase):
+
+    ADDITIONAL_FILTERS = ['NUMATopologyFilter']
+
+    ADMIN_API = True
+
+    def setUp(self):
+        super(PowerManagementTestsBase, self).setUp()
+
+        self.ctxt = nova_context.get_admin_context()
+
+        # Mock the 'NUMATopologyFilter' filter, as most tests need to inspect
+        # this
+        host_manager = self.scheduler.manager.host_manager
+        numa_filter_class = host_manager.filter_cls_map['NUMATopologyFilter']
+        host_pass_mock = mock.Mock(wraps=numa_filter_class().host_passes)
+        _p = mock.patch('nova.scheduler.filters'
+                        '.numa_topology_filter.NUMATopologyFilter.host_passes',
+                        side_effect=host_pass_mock)
+        self.mock_filter = _p.start()
+        self.addCleanup(_p.stop)
+
+        # for the sake of resizing, we need to patch the two methods below
+        self.useFixture(fixtures.MockPatch(
+            'nova.virt.libvirt.LibvirtDriver._get_instance_disk_info',
+             return_value=[]))
+        self.useFixture(fixtures.MockPatch('os.rename'))
+
+        self.useFixture(nova_fixtures.PrivsepFixture())
+
+        # Defining the main flavor for 4 vCPUs all pinned
+        self.extra_spec = {
+            'hw:cpu_policy': 'dedicated',
+            'hw:cpu_thread_policy': 'prefer',
+        }
+        self.pcpu_flavor_id = self._create_flavor(
+            vcpu=4, extra_spec=self.extra_spec)
+
+    def _assert_server_cpus_state(self, server, expected='online'):
+        inst = objects.Instance.get_by_uuid(self.ctxt, server['id'])
+        if not inst.numa_topology:
+            self.fail('Instance should have a NUMA topology in order to know '
+                      'its physical CPUs')
+        instance_pcpus = inst.numa_topology.cpu_pinning
+        self._assert_cpu_set_state(instance_pcpus, expected=expected)
+        return instance_pcpus
+
+    def _assert_cpu_set_state(self, cpu_set, expected='online'):
+        for i in cpu_set:
+            core = cpu.Core(i)
+            if expected == 'online':
+                self.assertTrue(core.online, f'{i} is not online')
+            elif expected == 'offline':
+                self.assertFalse(core.online, f'{i} is online')
+            elif expected == 'powersave':
+                self.assertEqual('powersave', core.governor)
+            elif expected == 'performance':
+                self.assertEqual('performance', core.governor)
+
+
+class PowerManagementTests(PowerManagementTestsBase):
+    """Test suite for a single host with 9 dedicated cores and 1 used for OS"""
+
+    def setUp(self):
+        super(PowerManagementTests, self).setUp()
+
+        self.useFixture(nova_fixtures.SysFileSystemFixture())
+
+        # Definining the CPUs to be pinned.
+        self.flags(cpu_dedicated_set='1-9', cpu_shared_set=None,
+                   group='compute')
+        self.flags(vcpu_pin_set=None)
+        self.flags(cpu_power_management=True, group='libvirt')
+
+        self.flags(allow_resize_to_same_host=True)
+        self.host_info = fakelibvirt.HostInfo(cpu_nodes=1, cpu_sockets=1,
+                                         cpu_cores=5, cpu_threads=2)
+        self.compute1 = self.start_compute(host_info=self.host_info,
+                                           hostname='compute1')
+
+        # All cores are shutdown at startup, let's check.
+        cpu_dedicated_set = hardware.get_cpu_dedicated_set()
+        self._assert_cpu_set_state(cpu_dedicated_set, expected='offline')
+
+    def test_hardstop_compute_service_if_wrong_opt(self):
+        self.flags(cpu_dedicated_set=None, cpu_shared_set=None,
+                   group='compute')
+        self.flags(vcpu_pin_set=None)
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.assertRaises(exception.InvalidConfiguration,
+                          self.start_compute, host_info=self.host_info,
+                                              hostname='compute2')
+
+    def test_create_server(self):
+        server = self._create_server(
+            flavor_id=self.pcpu_flavor_id,
+            expected_state='ACTIVE')
+        # Let's verify that the pinned CPUs are now online
+        self._assert_server_cpus_state(server, expected='online')
+
+        # Verify that the unused CPUs are still offline
+        inst = objects.Instance.get_by_uuid(self.ctxt, server['id'])
+        instance_pcpus = inst.numa_topology.cpu_pinning
+        cpu_dedicated_set = hardware.get_cpu_dedicated_set()
+        unused_cpus = cpu_dedicated_set - instance_pcpus
+        self._assert_cpu_set_state(unused_cpus, expected='offline')
+
+    def test_stop_start_server(self):
+        server = self._create_server(
+            flavor_id=self.pcpu_flavor_id,
+            expected_state='ACTIVE')
+
+        server = self._stop_server(server)
+        # Let's verify that the pinned CPUs are now stopped...
+        self._assert_server_cpus_state(server, expected='offline')
+
+        server = self._start_server(server)
+        # ...and now, they should be back.
+        self._assert_server_cpus_state(server, expected='online')
+
+    def test_resize(self):
+        server = self._create_server(
+            flavor_id=self.pcpu_flavor_id,
+            expected_state='ACTIVE')
+        server_pcpus = self._assert_server_cpus_state(server,
+                                                      expected='online')
+
+        new_flavor_id = self._create_flavor(
+            vcpu=5, extra_spec=self.extra_spec)
+        self._resize_server(server, new_flavor_id)
+        server2_pcpus = self._assert_server_cpus_state(server,
+                                                       expected='online')
+        # Even if the resize is not confirmed yet, the original guest is now
+        # destroyed so the cores are now offline.
+        self._assert_cpu_set_state(server_pcpus, expected='offline')
+
+        # let's revert the resize
+        self._revert_resize(server)
+        # So now the original CPUs will be online again, while the previous
+        # cores should be back offline.
+        self._assert_cpu_set_state(server_pcpus, expected='online')
+        self._assert_cpu_set_state(server2_pcpus, expected='offline')
+
+    def test_changing_strategy_fails(self):
+        # As a reminder, all cores have been shutdown before.
+        # Now we want to change the strategy and then we restart the service
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+        # See, this is not possible as we would have offline CPUs.
+        self.assertRaises(exception.InvalidConfiguration,
+                          self.restart_compute_service, hostname='compute1')
+
+
+class PowerManagementTestsGovernor(PowerManagementTestsBase):
+    """Test suite for speific governor usage (same 10-core host)"""
+
+    def setUp(self):
+        super(PowerManagementTestsGovernor, self).setUp()
+
+        self.useFixture(nova_fixtures.SysFileSystemFixture())
+
+        # Definining the CPUs to be pinned.
+        self.flags(cpu_dedicated_set='1-9', cpu_shared_set=None,
+                   group='compute')
+        self.flags(vcpu_pin_set=None)
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+
+        self.flags(allow_resize_to_same_host=True)
+        self.host_info = fakelibvirt.HostInfo(cpu_nodes=1, cpu_sockets=1,
+                                         cpu_cores=5, cpu_threads=2)
+        self.compute1 = self.start_compute(host_info=self.host_info,
+                                           hostname='compute1')
+
+    def test_create(self):
+        cpu_dedicated_set = hardware.get_cpu_dedicated_set()
+        # With the governor strategy, cores are still online but run with a
+        # powersave governor.
+        self._assert_cpu_set_state(cpu_dedicated_set, expected='powersave')
+
+        # Now, start an instance
+        server = self._create_server(
+            flavor_id=self.pcpu_flavor_id,
+            expected_state='ACTIVE')
+        # When pinned cores are run, the governor state is now performance
+        self._assert_server_cpus_state(server, expected='performance')
+
+    def test_changing_strategy_fails(self):
+        # Arbitratly set a core governor strategy to be performance
+        cpu.Core(1).set_high_governor()
+        # and then forget about it while changing the strategy.
+        self.flags(cpu_power_management_strategy='cpu_state', group='libvirt')
+        # This time, this wouldn't be acceptable as some core would have a
+        # difference performance while Nova would only online/offline it.
+        self.assertRaises(exception.InvalidConfiguration,
+                          self.restart_compute_service, hostname='compute1')
+
+
+class PowerManagementMixedInstances(PowerManagementTestsBase):
+    """Test suite for a single host with 6 dedicated cores, 3 shared and one
+    OS-restricted.
+    """
+
+    def setUp(self):
+        super(PowerManagementMixedInstances, self).setUp()
+
+        self.useFixture(nova_fixtures.SysFileSystemFixture())
+
+        # Definining 6 CPUs to be dedicated, not all of them in a series.
+        self.flags(cpu_dedicated_set='1-3,5-7', cpu_shared_set='4,8-9',
+                   group='compute')
+        self.flags(vcpu_pin_set=None)
+        self.flags(cpu_power_management=True, group='libvirt')
+
+        self.host_info = fakelibvirt.HostInfo(cpu_nodes=1, cpu_sockets=1,
+                                         cpu_cores=5, cpu_threads=2)
+        self.compute1 = self.start_compute(host_info=self.host_info,
+                                           hostname='compute1')
+
+        # Make sure only 6 are offline now
+        cpu_dedicated_set = hardware.get_cpu_dedicated_set()
+        self._assert_cpu_set_state(cpu_dedicated_set, expected='offline')
+
+        # cores 4 and 8-9 should be online
+        self._assert_cpu_set_state({4, 8, 9}, expected='online')
+
+    def test_standard_server_works_and_passes(self):
+
+        std_flavor_id = self._create_flavor(vcpu=2)
+        self._create_server(flavor_id=std_flavor_id, expected_state='ACTIVE')
+
+        # Since this is an instance with floating vCPUs on the shared set, we
+        # can only lookup the host CPUs and see they haven't changed state.
+        cpu_dedicated_set = hardware.get_cpu_dedicated_set()
+        self._assert_cpu_set_state(cpu_dedicated_set, expected='offline')
+        self._assert_cpu_set_state({4, 8, 9}, expected='online')
+
+        # We can now try to boot an instance with pinned CPUs to test the mix
+        pinned_server = self._create_server(
+            flavor_id=self.pcpu_flavor_id,
+            expected_state='ACTIVE')
+        # We'll see that its CPUs are now online
+        self._assert_server_cpus_state(pinned_server, expected='online')
+        # but it doesn't change the shared set
+        self._assert_cpu_set_state({4, 8, 9}, expected='online')
diff --git a/nova/tests/unit/virt/libvirt/cpu/test_api.py b/nova/tests/unit/virt/libvirt/cpu/test_api.py
index d47b3690a3..b5bcb762f3 100644
--- a/nova/tests/unit/virt/libvirt/cpu/test_api.py
+++ b/nova/tests/unit/virt/libvirt/cpu/test_api.py
@@ -12,6 +12,8 @@
 
 from unittest import mock
 
+from nova import exception
+from nova import objects
 from nova import test
 from nova.virt.libvirt.cpu import api
 from nova.virt.libvirt.cpu import core
@@ -23,6 +25,13 @@ class TestAPI(test.NoDBTestCase):
         super(TestAPI, self).setUp()
         self.core_1 = api.Core(1)
 
+        # Create a fake instance with two pinned CPUs but only one is on the
+        # dedicated set
+        numa_topology = objects.InstanceNUMATopology(cells=[
+            objects.InstanceNUMACell(cpu_pinning_raw={'0': '0', '2': '2'}),
+        ])
+        self.fake_inst = objects.Instance(numa_topology=numa_topology)
+
     @mock.patch.object(core, 'get_online')
     def test_online(self, mock_get_online):
         mock_get_online.return_value = True
@@ -61,3 +70,125 @@ class TestAPI(test.NoDBTestCase):
         self.core_1.set_high_governor()
         mock_set_governor.assert_called_once_with(self.core_1.ident,
                                                   'fake_high_gov')
+
+    @mock.patch.object(core, 'set_online')
+    def test_power_up_online(self, mock_online):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_up(self.fake_inst)
+        # only core #0 can be set as core #2 is not on the dedicated set
+        # As a reminder, core(i).online calls set_online(i)
+        mock_online.assert_called_once_with(0)
+
+    @mock.patch.object(core, 'set_governor')
+    def test_power_up_governor(self, mock_set_governor):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_up(self.fake_inst)
+        # only core #0 can be set as core #2 is not on the dedicated set
+        # As a reminder, core(i).set_high_governor calls set_governor(i)
+        mock_set_governor.assert_called_once_with(0, 'performance')
+
+    @mock.patch.object(core, 'set_online')
+    def test_power_up_skipped(self, mock_online):
+        self.flags(cpu_power_management=False, group='libvirt')
+        api.power_up(self.fake_inst)
+        mock_online.assert_not_called()
+
+    @mock.patch.object(core, 'set_online')
+    def test_power_up_skipped_if_standard_instance(self, mock_online):
+        self.flags(cpu_power_management=True, group='libvirt')
+        api.power_up(objects.Instance(numa_topology=None))
+        mock_online.assert_not_called()
+
+    @mock.patch.object(core, 'set_offline')
+    def test_power_down_offline(self, mock_offline):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_down(self.fake_inst)
+        # only core #0 can be set as core #2 is not on the dedicated set
+        # As a reminder, core(i).online calls set_online(i)
+        mock_offline.assert_called_once_with(0)
+
+    @mock.patch.object(core, 'set_governor')
+    def test_power_down_governor(self, mock_set_governor):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_down(self.fake_inst)
+        # only core #0 can be set as core #2 is not on the dedicated set
+        # As a reminder, core(i).set_high_governor calls set_governor(i)
+        mock_set_governor.assert_called_once_with(0, 'powersave')
+
+    @mock.patch.object(core, 'set_offline')
+    def test_power_down_skipped(self, mock_offline):
+        self.flags(cpu_power_management=False, group='libvirt')
+        api.power_down(self.fake_inst)
+        mock_offline.assert_not_called()
+
+    @mock.patch.object(core, 'set_offline')
+    def test_power_down_skipped_if_standard_instance(self, mock_offline):
+        self.flags(cpu_power_management=True, group='libvirt')
+        api.power_down(objects.Instance(numa_topology=None))
+        mock_offline.assert_not_called()
+
+    @mock.patch.object(core, 'set_offline')
+    def test_power_down_all_dedicated_cpus_offline(self, mock_offline):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_down_all_dedicated_cpus()
+        # All dedicated CPUs are turned offline
+        mock_offline.assert_has_calls([mock.call(0), mock.call(1)])
+
+    @mock.patch.object(core, 'set_governor')
+    def test_power_down_all_dedicated_cpus_governor(self, mock_set_governor):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+
+        api.power_down_all_dedicated_cpus()
+        # All dedicated CPUs are turned offline
+        mock_set_governor.assert_has_calls([mock.call(0, 'powersave'),
+                                            mock.call(1, 'powersave')])
+
+    @mock.patch.object(core, 'set_offline')
+    def test_power_down_all_dedicated_cpus_skipped(self, mock_offline):
+        self.flags(cpu_power_management=False, group='libvirt')
+        api.power_down_all_dedicated_cpus()
+        mock_offline.assert_not_called()
+
+    def test_power_down_all_dedicated_cpus_wrong_config(self):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set=None, group='compute')
+        self.assertRaises(exception.InvalidConfiguration,
+                          api.power_down_all_dedicated_cpus)
+
+    @mock.patch.object(core, 'get_governor')
+    @mock.patch.object(core, 'get_online')
+    def test_validate_all_dedicated_cpus_for_governor(self, mock_get_online,
+                                                      mock_get_governor):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+        self.flags(cpu_power_management_strategy='governor', group='libvirt')
+        mock_get_governor.return_value = 'performance'
+        mock_get_online.side_effect = (True, False)
+        self.assertRaises(exception.InvalidConfiguration,
+                          api.validate_all_dedicated_cpus)
+
+    @mock.patch.object(core, 'get_governor')
+    @mock.patch.object(core, 'get_online')
+    def test_validate_all_dedicated_cpus_for_cpu_state(self, mock_get_online,
+                                                       mock_get_governor):
+        self.flags(cpu_power_management=True, group='libvirt')
+        self.flags(cpu_dedicated_set='0-1', group='compute')
+        self.flags(cpu_power_management_strategy='cpu_state', group='libvirt')
+        mock_get_online.return_value = True
+        mock_get_governor.side_effect = ('powersave', 'performance')
+        self.assertRaises(exception.InvalidConfiguration,
+                          api.validate_all_dedicated_cpus)
diff --git a/nova/virt/libvirt/cpu/__init__.py b/nova/virt/libvirt/cpu/__init__.py
index 962c9469a0..4410a4e579 100644
--- a/nova/virt/libvirt/cpu/__init__.py
+++ b/nova/virt/libvirt/cpu/__init__.py
@@ -14,3 +14,9 @@ from nova.virt.libvirt.cpu import api
 
 
 Core = api.Core
+
+
+power_up = api.power_up
+power_down = api.power_down
+validate_all_dedicated_cpus = api.validate_all_dedicated_cpus
+power_down_all_dedicated_cpus = api.power_down_all_dedicated_cpus
diff --git a/nova/virt/libvirt/cpu/api.py b/nova/virt/libvirt/cpu/api.py
index e0b0a277d1..1c17458d6b 100644
--- a/nova/virt/libvirt/cpu/api.py
+++ b/nova/virt/libvirt/cpu/api.py
@@ -15,6 +15,10 @@ from dataclasses import dataclass
 from oslo_log import log as logging
 
 import nova.conf
+from nova import exception
+from nova.i18n import _
+from nova import objects
+from nova.virt import hardware
 from nova.virt.libvirt.cpu import core
 
 LOG = logging.getLogger(__name__)
@@ -51,6 +55,9 @@ class Core:
     def __eq__(self, other):
         return self.ident == other.ident
 
+    def __str__(self):
+        return str(self.ident)
+
     @property
     def governor(self) -> str:
         return core.get_governor(self.ident)
@@ -60,3 +67,91 @@ class Core:
 
     def set_low_governor(self) -> None:
         core.set_governor(self.ident, CONF.libvirt.cpu_power_governor_low)
+
+
+def power_up(instance: objects.Instance) -> None:
+    if not CONF.libvirt.cpu_power_management:
+        return
+    if instance.numa_topology is None:
+        return
+
+    cpu_dedicated_set = hardware.get_cpu_dedicated_set() or set()
+    pcpus = instance.numa_topology.cpu_pinning
+    powered_up = set()
+    for pcpu in pcpus:
+        if pcpu in cpu_dedicated_set:
+            pcpu = Core(pcpu)
+            if CONF.libvirt.cpu_power_management_strategy == 'cpu_state':
+                pcpu.online = True
+            else:
+                pcpu.set_high_governor()
+            powered_up.add(str(pcpu))
+    LOG.debug("Cores powered up : %s", powered_up)
+
+
+def power_down(instance: objects.Instance) -> None:
+    if not CONF.libvirt.cpu_power_management:
+        return
+    if instance.numa_topology is None:
+        return
+
+    cpu_dedicated_set = hardware.get_cpu_dedicated_set() or set()
+    pcpus = instance.numa_topology.cpu_pinning
+    powered_down = set()
+    for pcpu in pcpus:
+        if pcpu in cpu_dedicated_set:
+            pcpu = Core(pcpu)
+            if CONF.libvirt.cpu_power_management_strategy == 'cpu_state':
+                pcpu.online = False
+            else:
+                pcpu.set_low_governor()
+            powered_down.add(str(pcpu))
+    LOG.debug("Cores powered down : %s", powered_down)
+
+
+def power_down_all_dedicated_cpus() -> None:
+    if not CONF.libvirt.cpu_power_management:
+        return
+    if (CONF.libvirt.cpu_power_management and
+        not CONF.compute.cpu_dedicated_set
+    ):
+        msg = _("'[compute]/cpu_dedicated_set' is mandatory to be set if "
+                "'[libvirt]/cpu_power_management' is set."
+                "Please provide the CPUs that can be pinned or don't use the "
+                "power management if you only use shared CPUs.")
+        raise exception.InvalidConfiguration(msg)
+
+    cpu_dedicated_set = hardware.get_cpu_dedicated_set() or set()
+    for pcpu in cpu_dedicated_set:
+        pcpu = Core(pcpu)
+        if CONF.libvirt.cpu_power_management_strategy == 'cpu_state':
+            pcpu.online = False
+        else:
+            pcpu.set_low_governor()
+    LOG.debug("Cores powered down : %s", cpu_dedicated_set)
+
+
+def validate_all_dedicated_cpus() -> None:
+    if not CONF.libvirt.cpu_power_management:
+        return
+    cpu_dedicated_set = hardware.get_cpu_dedicated_set() or set()
+    governors = set()
+    cpu_states = set()
+    for pcpu in cpu_dedicated_set:
+        pcpu = Core(pcpu)
+        # we need to collect the governors strategy and the CPU states
+        governors.add(pcpu.governor)
+        cpu_states.add(pcpu.online)
+    if CONF.libvirt.cpu_power_management_strategy == 'cpu_state':
+        # all the cores need to have the same governor strategy
+        if len(governors) > 1:
+            msg = _("All the cores need to have the same governor strategy"
+                    "before modifying the CPU states. You can reboot the "
+                    "compute node if you prefer.")
+            raise exception.InvalidConfiguration(msg)
+    elif CONF.libvirt.cpu_power_management_strategy == 'governor':
+        # all the cores need to be online
+        if False in cpu_states:
+            msg = _("All the cores need to be online before modifying the "
+                    "governor strategy.")
+            raise exception.InvalidConfiguration(msg)
diff --git a/nova/virt/libvirt/driver.py b/nova/virt/libvirt/driver.py
index 905ec0f306..869996f615 100644
--- a/nova/virt/libvirt/driver.py
+++ b/nova/virt/libvirt/driver.py
@@ -114,6 +114,7 @@ from nova.virt.image import model as imgmodel
 from nova.virt import images
 from nova.virt.libvirt import blockinfo
 from nova.virt.libvirt import config as vconfig
+from nova.virt.libvirt import cpu as libvirt_cpu
 from nova.virt.libvirt import designer
 from nova.virt.libvirt import event as libvirtevent
 from nova.virt.libvirt import guest as libvirt_guest
@@ -817,6 +818,18 @@ class LibvirtDriver(driver.ComputeDriver):
                         "force_raw_images to True.")
                 raise exception.InvalidConfiguration(msg)
 
+        # NOTE(sbauza): We verify first if the dedicated CPU performances were
+        # modified by Nova before. Note that it can provide an exception if
+        # either the governor strategies are different between the cores or if
+        # the cores are offline.
+        libvirt_cpu.validate_all_dedicated_cpus()
+        # NOTE(sbauza): We powerdown all dedicated CPUs but if some instances
+        # exist that are pinned for some CPUs, then we'll later powerup those
+        # CPUs when rebooting the instance in _init_instance()
+        # Note that it can provide an exception if the config options are
+        # 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()
@@ -1512,6 +1525,8 @@ class LibvirtDriver(driver.ComputeDriver):
             # NOTE(GuanQiang): teardown container to avoid resource leak
             if CONF.libvirt.virt_type == 'lxc':
                 self._teardown_container(instance)
+            # We're sure the instance is gone, we can shutdown the core if so
+            libvirt_cpu.power_down(instance)
 
     def destroy(self, context, instance, network_info, block_device_info=None,
                 destroy_disks=True, destroy_secrets=True):
@@ -3164,6 +3179,7 @@ class LibvirtDriver(driver.ComputeDriver):
 
         current_power_state = guest.get_power_state(self._host)
 
+        libvirt_cpu.power_up(instance)
         # TODO(stephenfin): Any reason we couldn't use 'self.resume' here?
         guest.launch(pause=current_power_state == power_state.PAUSED)
 
@@ -7646,6 +7662,7 @@ class LibvirtDriver(driver.ComputeDriver):
                 post_xml_callback()
 
             if power_on or pause:
+                libvirt_cpu.power_up(instance)
                 guest.launch(pause=pause)
 
             return guest
diff --git a/releasenotes/notes/bp-libvirt-cpu-state-mgmt-fbc9c1f9f473003c.yaml b/releasenotes/notes/bp-libvirt-cpu-state-mgmt-fbc9c1f9f473003c.yaml
new file mode 100644
index 0000000000..95422fce67
--- /dev/null
+++ b/releasenotes/notes/bp-libvirt-cpu-state-mgmt-fbc9c1f9f473003c.yaml
@@ -0,0 +1,18 @@
+---
+features:
+  - |
+    This is now possible to configure nova-compute services using libvirt driver
+    by setting ``[libvirt]cpu_power_management`` to ``True`` in order to let the
+    service to powering down or up physical CPUs depending on whether those CPUs
+    are pinned or not to instances. In order on to support this feature, the
+    compute service needs to be set with ``[compute]cpu_dedicated_set``. If so,
+    all the related CPUs will be powering down until they are used by an
+    instance where the related pinned CPU will be powering up just before
+    starting the guest. If ``[compute]cpu_dedicated_set`` isn't set, then the
+    compute service will refuse to start.
+    By default the power strategy will offline CPUs when powering down and
+    online the CPUs on powering up but another strategy is possible by using
+    ``[libvirt]cpu_power_management_strategy=governor`` which will rather modify
+    the related CPU governor using ``[libvirt]cpu_power_governor_low`` and
+    ``[libvirt]cpu_power_governor_high`` configuration values (respective
+    defaults being  ``powersave`` and ``performance``)