PCI passthrough reschedule race condition

Bug #1860555 reported by Mark Goddard on 2020-01-22
14
This bug affects 1 person
Affects Status Importance Assigned to Milestone
OpenStack Compute (nova)
High
Mark Goddard

Bug Description

Steps to reproduce
------------------

Create multiple instances concurrently using a flavor with a PCI passthrough request (--property "pci_passthrough:alias"="<alias>:<count>"), and a scheduler hint with some anti-affinity constraint.

Expected result
---------------

The instances are created successfully, and each have the expected number of PCI devices attached.

Actual result
-------------

Sometimes, instances may fail during creation, or may be created with more PCI devices than requested.

Environment
-----------

Nova 18.2.2 (rocky), CentOS 7, libvirt, deployed by kolla-ansible.

Analysis
--------

If an instance with PCI passthrough devices is rescheduled (e.g. due to
affinity violation), the instance can end up with extra PCI devices attached.
If the devices selected on the original and subsequent compute nodes have the
same address, the instance will fail to create, with the following error:

libvirtError: internal error: Device 0000:89:00.0 is already in use

However, if the devices are different, and all available on the first and
second compute nodes, the VM may end up with additional hostdevs.

On investigation, when the node is rescheduled, the instance object passed to
the conductor RPC API contains the PCI devices that should have been freed.
This is because the claim object holds a clone of the instance that is used to
perform the abort on failure [1][2], and the PCI devices removed from its list are not
reflected in the original object. There is a secondary issue that the PCI
manager was not passing through the instance to the PCI object's free() method
in all cases [3], resulting in the PCI device not being removed from the
instance.pci_devices list.

I have two alternative fixes for this issue, but they will need a little time to work their way out of an organisation. Essentially:

1. pass the original instance (not the clone) to the abort function in the Claim.
2. refresh the instance from DB when rescheduling

The former is a more general solution, but I don't know the reasons for using a clone in the first place. The second works for reschedules, but may leave a hole for resize or migration. I haven't reproduced the issue in those cases but it seems possible that it would be present.

[1] https://opendev.org/openstack/nova/src/branch/master/nova/compute/claims.py#L64
[2] https://opendev.org/openstack/nova/src/branch/master/nova/compute/claims.py#L83
[3] https://opendev.org/openstack/nova/src/branch/master/nova/pci/manager.py#L309

Changed in nova:
importance: Undecided → High

Fix proposed to branch: master
Review: https://review.opendev.org/710847

Changed in nova:
assignee: nobody → Mark Goddard (mgoddard)
status: New → In Progress

Fix proposed to branch: master
Review: https://review.opendev.org/710848

To post a comment you must log in.
This report contains Public information  Edit
Everyone can see this information.

Other bug subscribers