pci passthrough: incorrect device allocation

[Baodong (Robert) Li]

The PCI filter for scheduling runs on the basis of pci stats pools available in a compute node. If the PCI requests match one or more pools, the number of devices will be subtracted from the total number of devices available in those pools.

On the compute node, PCI device allocation is performed based on the list of free devices that are available on the node. The PCI requests are used to match against the device itself, rather than the pci stats pools.

The unsymmetrical handling of scheduling versus allocating could cause incorrect pci stats, and thus incorrect scheduling, and nova instances failed to boot.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/81954

[Baodong (Robert) Li (baoli)]

I'd like to bring this to the core developer's attention for approval. Thanks.

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/81954
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=fa2f139e22ed1317f4afe01faaf7ee3943444715
Submitter: Jenkins
Branch: master

commit fa2f139e22ed1317f4afe01faaf7ee3943444715
Author: Robert Li <email address hidden>
Date: Fri Jul 11 14:26:28 2014 -0400

    makes sure correct PCI device allocation

    With this patch, on the compute node, a stats pool will be associated with a
    list of devices that belongs to the pool. This makes sure that PCI devices are
    allocated out of the same stats pools that are used by the nova scheduler to
    satisfy the PCI requests. And therefore, stats pools on the compute nodes will
    be kept in sync with their counterparts in the nova scheduler.

    Change-Id: I2d97c6354215e2ac5ff659e3203c33771abe1c09
    Closes-bug: #1288809