libvirt: At start-up rework compareCPU() usage with a workaround

In this patch:

  - Remove the first compareCPU() call (called via the wrapper
    _compare_cpu()) in _check_cpu_compatibility(), and let libvirt
    handle it.  (QEMU >=2.9 and libvirt >= 4.4.0 are the mininum
    required versions, and upstream Nova satisfies them by a good
    margin.)

  - Validate the user-configured CPU models from
    _get_cpu_model_mapping().  And take into account all the CPU flags
    before calling _compare_cpu().

    (Suggested-by: Sean Mooney -- thanks!)

  - Add a workaround to allow skipping the remaining compareCPU() call
    in _check_cpu_compatibility() as a potential future-proof (because
    we cannot test all possible CPU models and hardware).  Unlike the
    removed first call, this call takes into account the extra CPU flags
    provided by the user into account when evaluating guest CPU model
    compatibility.

As a follow up comes the patch[1] that replaces the older libvirt CPU
API with the newer one.

[1] https://review.opendev.org/c/openstack/nova/+/869950 -- libvirt:
    Replace usage of compareCPU() with compareHypervisorCPU()

Change-Id: I8ef9db851b37c5249d2efbe09a15a1ddbae8205d
Signed-off-by: Kashyap Chamarthy <kchamart@redhat.com>

3 files changed, 46 insertions, 30 deletions

diff --git a/nova/conf/workarounds.py b/nova/conf/workarounds.py
index 1116664d36..e485ae673a 100644
--- a/nova/conf/workarounds.py
+++ b/nova/conf/workarounds.py
@@ -410,6 +410,13 @@ with the destination host. When using QEMU >= 2.9 and libvirt >=
 4.4.0, libvirt will do the correct thing with respect to checking CPU
 compatibility on the destination host during live migration.
 """),
+    cfg.BoolOpt('skip_cpu_compare_at_startup',
+               default=False,
+               help="""
+This will skip the CPU comparison call at the startup of Compute
+service and lets libvirt handle it.
+"""),
+
     cfg.BoolOpt(
         'skip_hypervisor_version_check_on_lm',
         default=False,
diff --git a/nova/tests/unit/virt/libvirt/test_driver.py b/nova/tests/unit/virt/libvirt/test_driver.py
index 1c5f79dc89..9d45ba017c 100644
--- a/nova/tests/unit/virt/libvirt/test_driver.py
+++ b/nova/tests/unit/virt/libvirt/test_driver.py
@@ -1333,6 +1333,22 @@ class LibvirtConnTestCase(test.NoDBTestCase,
     @mock.patch.object(libvirt_driver.LibvirtDriver,
                        '_register_all_undefined_instance_details',
                        new=mock.Mock())
+    @mock.patch('nova.virt.libvirt.host.libvirt.Connection.compareCPU')
+    def test__check_cpu_compatibility_skip_compare_at_init(
+            self, mocked_compare
+    ):
+        self.flags(group='workarounds', skip_cpu_compare_at_startup=True)
+        self.flags(cpu_mode="custom",
+                   cpu_models=["Icelake-Server-noTSX"],
+                   cpu_model_extra_flags = ["-mpx"],
+                   group="libvirt")
+        drvr = libvirt_driver.LibvirtDriver(fake.FakeVirtAPI(), True)
+        drvr.init_host("dummyhost")
+        mocked_compare.assert_not_called()
+
+    @mock.patch.object(libvirt_driver.LibvirtDriver,
+                       '_register_all_undefined_instance_details',
+                       new=mock.Mock())
     def test__check_cpu_compatibility_with_flag(self):
         self.flags(cpu_mode="custom",
                    cpu_models=["Penryn"],
@@ -1343,7 +1359,7 @@ class LibvirtConnTestCase(test.NoDBTestCase,
 
     @mock.patch('nova.virt.libvirt.host.libvirt.Connection.compareCPU')
     def test__check_cpu_compatibility_advance_flag(self, mocked_compare):
-        mocked_compare.side_effect = (2, 0)
+        mocked_compare.side_effect = (-1, 0)
         self.flags(cpu_mode="custom",
                    cpu_models=["qemu64"],
                    cpu_model_extra_flags = ["avx", "avx2"],
@@ -1356,7 +1372,7 @@ class LibvirtConnTestCase(test.NoDBTestCase,
     def test__check_cpu_compatibility_wrong_flag(self, mocked_compare):
         # here, and in the surrounding similar tests, the non-zero error
         # code in the compareCPU() side effect indicates error
-        mocked_compare.side_effect = (2, 0)
+        mocked_compare.side_effect = (-1, 0)
         self.flags(cpu_mode="custom",
                    cpu_models=["Broadwell-noTSX"],
                    cpu_model_extra_flags = ["a v x"],
@@ -1369,7 +1385,7 @@ class LibvirtConnTestCase(test.NoDBTestCase,
     def test__check_cpu_compatibility_enabled_and_disabled_flags(
             self, mocked_compare
     ):
-        mocked_compare.side_effect = (2, 0)
+        mocked_compare.side_effect = (-1, 0)
         self.flags(
             cpu_mode="custom",
             cpu_models=["Cascadelake-Server"],
diff --git a/nova/virt/libvirt/driver.py b/nova/virt/libvirt/driver.py
index d7fe0fcc93..ff503de7e6 100644
--- a/nova/virt/libvirt/driver.py
+++ b/nova/virt/libvirt/driver.py
@@ -984,33 +984,26 @@ class LibvirtDriver(driver.ComputeDriver):
             msg = _("The cpu_models option is required when cpu_mode=custom")
             raise exception.Invalid(msg)
 
-        cpu = vconfig.LibvirtConfigGuestCPU()
-        for model in models:
-            cpu.model = self._get_cpu_model_mapping(model)
-            try:
-                self._compare_cpu(cpu, self._get_cpu_info(), None)
-            except exception.InvalidCPUInfo as e:
-                msg = (_("Configured CPU model: %(model)s is not "
-                         "compatible with host CPU. Please correct your "
-                         "config and try again. %(e)s") % {
-                            'model': model, 'e': e})
-                raise exception.InvalidCPUInfo(msg)
-
-        # Use guest CPU model to check the compatibility between guest CPU and
-        # configured extra_flags
-        cpu = vconfig.LibvirtConfigGuestCPU()
-        cpu.model = self._host.get_capabilities().host.cpu.model
-        for flag in set(x.lower() for x in CONF.libvirt.cpu_model_extra_flags):
-            cpu_feature = self._prepare_cpu_flag(flag)
-            cpu.add_feature(cpu_feature)
-            try:
-                self._compare_cpu(cpu, self._get_cpu_info(), None)
-            except exception.InvalidCPUInfo as e:
-                msg = (_("Configured extra flag: %(flag)s it not correct, or "
-                         "the host CPU does not support this flag. Please "
-                         "correct the config and try again. %(e)s") % {
-                            'flag': flag, 'e': e})
-                raise exception.InvalidCPUInfo(msg)
+        if not CONF.workarounds.skip_cpu_compare_at_startup:
+            # Use guest CPU model to check the compatibility between
+            # guest CPU and configured extra_flags
+            for model in models:
+                cpu = vconfig.LibvirtConfigGuestCPU()
+                cpu.model = self._get_cpu_model_mapping(model)
+                for flag in set(x.lower() for
+                                x in CONF.libvirt.cpu_model_extra_flags):
+                    cpu_feature = self._prepare_cpu_flag(flag)
+                    cpu.add_feature(cpu_feature)
+                try:
+                    self._compare_cpu(cpu, self._get_cpu_info(), None)
+                except exception.InvalidCPUInfo as e:
+                    msg = (_("Configured CPU model: %(model)s "
+                             "and CPU Flags %(flags)s ar not "
+                             "compatible with host CPU. Please correct your "
+                             "config and try again. %(e)s") % {
+                                'model': model, 'e': e,
+                                'flags': CONF.libvirt.cpu_model_extra_flags})
+                    raise exception.InvalidCPUInfo(msg)
 
     def _check_vtpm_support(self) -> None:
         # TODO(efried): A key manager must be configured to create/retrieve