Public bug reported: The nova.compute.claims.Claim object is used to test whether a set of requested resources can be satisfied by the compute node. In the constructor of the Claim object, the Claim._claim_test() object is called:
def __init__(self, context, instance, tracker, resources, overhead=None, limits=None): super(Claim, self).__init__() <snip> # Check claim at constructor to avoid mess code # Raise exception ComputeResourcesUnavailable if claim failed self._claim_test(resources, limits) If we take a look at _claim_test(), we see pretty clearly that resources are NOT supposed to be actually claimed -- instead, the method should only *check* to see if the request can be fulfilled: def _claim_test(self, resources, limits=None): """Test if this claim can be satisfied given available resources and optional oversubscription limits This should be called before the compute node actually consumes the resources required to execute the claim. :param resources: available local compute node resources :returns: Return true if resources are available to claim. """ <snip> reasons = [self._test_memory(resources, memory_mb_limit), self._test_disk(resources, disk_gb_limit), self._test_vcpus(resources, vcpus_limit), self._test_numa_topology(resources, numa_topology_limit), self._test_pci()] reasons = reasons + self._test_ext_resources(limits) reasons = [r for r in reasons if r is not None] if len(reasons) > 0: raise exception.ComputeResourcesUnavailable(reason= "; ".join(reasons)) Unfortunately, the PCI devices are *actually* claimed in the _test_pci() method: def _test_pci(self): pci_requests = objects.InstancePCIRequests.get_by_instance_uuid( self.context, self.instance.uuid) if pci_requests.requests: devs = self.tracker.pci_tracker.claim_instance(self.context, self.instance) if not devs: return _('Claim pci failed.') What this means is that if an instance is attempted to be launched on a compute node and that instance has PCI requests that can be satisfied by the compute host, but say, there isn't enough available RAM on the node, the Claim will raise ComputeResourcesUnavailable which will trigger a Retry operation to the scheduler, but the PCI devices will have already been marked as claimed by that instance in the PCI device tracker: devs = self.tracker.pci_tracker.claim_instance(self.context, self.instance) The above code actually marks one or more PCI devices on the compute host as claimed for the instance. This introduces inconsistent state into the system. Making things worse is the fact that the nova.pci.manager.PciDevTracker object uses the nova.pci.stats.PciDevStats object for tracking consumed quantities of "pools" of the PCI device types and both the stats aggregation AND the PciDevTracker.pci_devs PciDeviceList object have their state changed improperly. ** Affects: nova Importance: Undecided Status: New ** Tags: pci ** Tags added: pci -- You received this bug notification because you are a member of Yahoo! Engineering Team, which is subscribed to OpenStack Compute (nova). https://bugs.launchpad.net/bugs/1549984 Title: PCI devices claimed on compute node during _claim_test() Status in OpenStack Compute (nova): New Bug description: The nova.compute.claims.Claim object is used to test whether a set of requested resources can be satisfied by the compute node. In the constructor of the Claim object, the Claim._claim_test() object is called: def __init__(self, context, instance, tracker, resources, overhead=None, limits=None): super(Claim, self).__init__() <snip> # Check claim at constructor to avoid mess code # Raise exception ComputeResourcesUnavailable if claim failed self._claim_test(resources, limits) If we take a look at _claim_test(), we see pretty clearly that resources are NOT supposed to be actually claimed -- instead, the method should only *check* to see if the request can be fulfilled: def _claim_test(self, resources, limits=None): """Test if this claim can be satisfied given available resources and optional oversubscription limits This should be called before the compute node actually consumes the resources required to execute the claim. :param resources: available local compute node resources :returns: Return true if resources are available to claim. """ <snip> reasons = [self._test_memory(resources, memory_mb_limit), self._test_disk(resources, disk_gb_limit), self._test_vcpus(resources, vcpus_limit), self._test_numa_topology(resources, numa_topology_limit), self._test_pci()] reasons = reasons + self._test_ext_resources(limits) reasons = [r for r in reasons if r is not None] if len(reasons) > 0: raise exception.ComputeResourcesUnavailable(reason= "; ".join(reasons)) Unfortunately, the PCI devices are *actually* claimed in the _test_pci() method: def _test_pci(self): pci_requests = objects.InstancePCIRequests.get_by_instance_uuid( self.context, self.instance.uuid) if pci_requests.requests: devs = self.tracker.pci_tracker.claim_instance(self.context, self.instance) if not devs: return _('Claim pci failed.') What this means is that if an instance is attempted to be launched on a compute node and that instance has PCI requests that can be satisfied by the compute host, but say, there isn't enough available RAM on the node, the Claim will raise ComputeResourcesUnavailable which will trigger a Retry operation to the scheduler, but the PCI devices will have already been marked as claimed by that instance in the PCI device tracker: devs = self.tracker.pci_tracker.claim_instance(self.context, self.instance) The above code actually marks one or more PCI devices on the compute host as claimed for the instance. This introduces inconsistent state into the system. Making things worse is the fact that the nova.pci.manager.PciDevTracker object uses the nova.pci.stats.PciDevStats object for tracking consumed quantities of "pools" of the PCI device types and both the stats aggregation AND the PciDevTracker.pci_devs PciDeviceList object have their state changed improperly. To manage notifications about this bug go to: https://bugs.launchpad.net/nova/+bug/1549984/+subscriptions -- Mailing list: https://launchpad.net/~yahoo-eng-team Post to : yahoo-eng-team@lists.launchpad.net Unsubscribe : https://launchpad.net/~yahoo-eng-team More help : https://help.launchpad.net/ListHelp