nova boot GPU instance will attach more one GPU pci device when reschedule happened

[guolei]

Description
===========
When we boot a GPU instance, on nova-compute's instance_claim
https://github.com/openstack/nova/blob/7020196aaa2fedd537806fe229e237c91e4f0ca5/nova/compute/resource_tracker.py#L197-L217
input instance object's attribute 'pci_devices' had update from [] to [PciDevice], it include a calculated GPU PCI device object.

Ok, now we pay attention to claim's code flow:
https://github.com/openstack/nova/blob/7020196aaa2fedd537806fe229e237c91e4f0ca5/nova/compute/claims.py#L64
it cloned input instance object, set to self.instance
https://github.com/openstack/nova/blob/7020196aaa2fedd537806fe229e237c91e4f0ca5/nova/compute/claims.py#L78-L84
abort func will abort instance's claim with self.instance, it a cloned one, not the origin input instance object.

Now, we can see, if spawn instance failed, claim.abort will be called, it revert the cloned instance object's
 'pci_devices' attribute to [], and pci_device in db had reverted from allocate to free too. but the origin input instance object not, origin instance object's 'pci_devices' is still [PciDevice], and it will send to nova-conductor to do reschedule, and on next node, after claim, instance.pci_devices will be [PciDevice, PciDevice]

And then, spawn instance will have two GPU pci device, or raise a LibvirtError, "Device xxx is in used"

Steps to reproduce
==================
1. build libvirt error on all compute nodes
2. nova boot a GPU instance
3. show guest xml in nova-compute.log

Expected result
===============
on reschedule node, guest xml had just one GPU pci device

Actual result
=============
on reschedule node, guest xml had more then one GPU pci device

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.opendev.org/759544