diff --git a/nova/compute/manager.py b/nova/compute/manager.py
index 10a38511dad2..0b5d201ba6d5 100644
--- a/nova/compute/manager.py
+++ b/nova/compute/manager.py
@@ -4931,6 +4931,13 @@ class ComputeManager(manager.Manager):
         # terminate all the connections with the volume server and the host
         self._terminate_volume_connections(context, instance, bdms)
 
+        # Free up the resource allocations in the placement service.
+        # This should happen *before* the vm_state is changed to
+        # SHELVED_OFFLOADED in case client-side code is polling the API to
+        # schedule more instances (or unshelve) once this server is offloaded.
+        rt = self._get_resource_tracker()
+        rt.delete_allocation_for_shelve_offloaded_instance(context, instance)
+
         instance.power_state = current_power_state
         # NOTE(mriedem): The vm_state has to be set before updating the
         # resource tracker, see vm_states.ALLOW_RESOURCE_REMOVAL. The host/node
@@ -4944,9 +4951,6 @@ class ComputeManager(manager.Manager):
         # NOTE(ndipanov): Free resources from the resource tracker
         self._update_resource_tracker(context, instance)
 
-        rt = self._get_resource_tracker()
-        rt.delete_allocation_for_shelve_offloaded_instance(context, instance)
-
         # NOTE(sfinucan): RPC calls should no longer be attempted against this
         # instance, so ensure any calls result in errors
         self._nil_out_instance_obj_host_and_node(instance)
diff --git a/nova/tests/unit/compute/test_shelve.py b/nova/tests/unit/compute/test_shelve.py
index 78a2a252637f..2ac38c2c3fda 100644
--- a/nova/tests/unit/compute/test_shelve.py
+++ b/nova/tests/unit/compute/test_shelve.py
@@ -225,16 +225,24 @@ class ShelveComputeManagerTestCase(test_compute.BaseTestCase):
         fake_bdms = objects.BlockDeviceMappingList()
         mock_get_bdms.return_value = fake_bdms
 
-        with mock.patch.object(instance, 'save'):
-            self.compute.shelve_offload_instance(self.context, instance,
-                                                 clean_shutdown=clean_shutdown)
-            mock_notify.assert_has_calls([
-                mock.call(self.context, instance, 'fake-mini',
-                          action='shelve_offload', phase='start',
-                          bdms=fake_bdms),
-                mock.call(self.context, instance, 'fake-mini',
-                          action='shelve_offload', phase='end',
-                          bdms=fake_bdms)])
+        def stub_instance_save(inst, *args, **kwargs):
+            # If the vm_state is changed to SHELVED_OFFLOADED make sure we
+            # have already freed up allocations in placement.
+            if inst.vm_state == vm_states.SHELVED_OFFLOADED:
+                self.assertTrue(mock_delete_alloc.called,
+                                'Allocations must be deleted before the '
+                                'vm_status can change to shelved_offloaded.')
+
+        self.stub_out('nova.objects.Instance.save', stub_instance_save)
+        self.compute.shelve_offload_instance(self.context, instance,
+                                             clean_shutdown=clean_shutdown)
+        mock_notify.assert_has_calls([
+            mock.call(self.context, instance, 'fake-mini',
+                      action='shelve_offload', phase='start',
+                      bdms=fake_bdms),
+            mock.call(self.context, instance, 'fake-mini',
+                      action='shelve_offload', phase='end',
+                      bdms=fake_bdms)])
 
         self.assertEqual(vm_states.SHELVED_OFFLOADED, instance.vm_state)
         self.assertIsNone(instance.task_state)