[pci-passthrough] Failed to launch vm after restarting nova compute

[Moshe Levi]

my nova.conf contains pci_passthrough_whitelist parameter

launch vm after openstack installation the vm was successfully booted.
when restarting nova compute and then again trying to launch vm I can see the following error
"An object of type PciDevicePoolList is required"

please note it doesn't matter if it vm with normal or vm with direct port.

[Davanum Srinivas (DIMS) (dims-v)]

@Moshe,

Can you please specify which version of Nova? (Trunk?) and a reasonable traceback/log to better diagnose the issue?

[Moshe Levi (moshele)]

@Davanum
yes it is nova trunk.
nova-compute log is attached.
you can see the trace when I try to delete the vm in error.
in Line 968 you can see the failure on create
2015-03-23 22:30:11.080 11235 DEBUG nova.compute.utils [-] [instance: 0ddeb177-3abf-4477-9336-4d9cc8787f5d] An object of type PciDevicePoolList is required here notify_about_instance_usage /.autodirect/mtrswgwork/moshele/openstack/nova/nova/compute/utils.py:310
2015-03-23 22:30:11.081 11235 DEBUG nova.compute.manager [-] [instance: 0ddeb177-3abf-4477-9336-4d9cc8787f5d] Build of instance 0ddeb177-3abf-4477-9336-4d9cc8787f5d was re-scheduled: An object of type PciDevicePoolList is required here _do_build_and_run_instance /.autodirect/mtrswgwork/moshele/openstack/nova/nova/compute/manager.py:2189
(I think in this case the log level should be Error not Debug)

I started to debug this issue and I saw the in that one of resources sent to the conductor is "pci_device_pools" which is a list for example ('pci_device_pools': [{'count': 7, 'vendor_id': u'15b3', 'product_id': u'1004', 'tags': {u'numa_node': None, u'physical_network': u'physnet1'}}]) When this resource saved into the database I get the above error.

If I removing the pci_device_pools key from the resources (remove it from self.compute_node in the resource_tracker) it will fix the issue, but I am not sure that it is the correct way to go.

[Sean Dague (sdague)]

Please provide your nova.conf, there have been changes in the format here, and that might be part of the issue.

[Moshe Levi (moshele)]

nova.conf attached

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/167256
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=5b77c108f14f2bcd42fecfcd060331e57a2e07dd
Submitter: Jenkins
Branch: master

commit 5b77c108f14f2bcd42fecfcd060331e57a2e07dd
Author: Dan Smith <email address hidden>
Date: Tue Mar 24 07:49:41 2015 -0700

    Convert pci_device_pools dict to object before passing to scheduler

    This also fixes other issues in the resource tracker around inconsistent
    use of legacy pci_stats usage from when we were dealing with a db model
    compute node, to pci_device_pools to be in line with objects.

    Closes-Bug: #1435483
    Change-Id: I56a9aa8b3e81a22d57e83f956bbcf1e42599a954