When assigning an SR-IOV Virtual Function (VF) device to an instance, its corresponding Physical Function (PF) device is correctly masked as unavailable in the database. However, in past releases, deleting the instance did not update the PF as available. As a result, PCI devices were never released from the database after instances which used them were deleted.
With this update, nova now maintains an in-memory tree of PCI devices, then periodically flushes it into the database. This helps update the database on information about available devices.
DescriptionSahid Ferdjaoui
2016-06-03 14:53:15 UTC
Assigning an SR-IOV VF device to an instance when PFs are whitelisted too correctly marks the PF as unavailable if one of it's VFs got assigned. However when we delete the instance, the PF is not makred as available.
Steps to reproduce:
1) Whitelist PFs and VFs in nova.conf (as explained in the docs) for example
pci_passthrough_whitelist = [{"product_id":"1520", "vendor_id":"8086", "physical_network":"phynet"}, {"product_id":"1521", "vendor_id":"8086", "physical_network":"phynet"}] # Both pfs and vfs are whitelisted
2) Add an alias to assign a VF pci_alias = {"name": "vf", "device_type": "type-VF"}
3) Set up a flavor with an alias extra_spec
$ nova flavor-key 2 set "pci_passthrough:alias"="vf:1"
4) Boot an instance with the said flavor and observe a VF being set to 'allocated' and a PF being set to 'unavailable'
select * from pci_devices where deleted=0;
5) Delete the instance from step 4 and observe that the VF has been made available but the PF is still 'unavailable'. Both should be back to available if this was the only VF used.
Ok I finally found the problem. When we compare PCIdevice objects with a NULL value the code is raising an exception whereas returning False. I have addressed a patch upstream which also needs to be backported [1].
https://review.openstack.org/#/c/318159/
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.
https://rhn.redhat.com/errata/RHEA-2016-1597.html
Assigning an SR-IOV VF device to an instance when PFs are whitelisted too correctly marks the PF as unavailable if one of it's VFs got assigned. However when we delete the instance, the PF is not makred as available. Steps to reproduce: 1) Whitelist PFs and VFs in nova.conf (as explained in the docs) for example pci_passthrough_whitelist = [{"product_id":"1520", "vendor_id":"8086", "physical_network":"phynet"}, {"product_id":"1521", "vendor_id":"8086", "physical_network":"phynet"}] # Both pfs and vfs are whitelisted 2) Add an alias to assign a VF pci_alias = {"name": "vf", "device_type": "type-VF"} 3) Set up a flavor with an alias extra_spec $ nova flavor-key 2 set "pci_passthrough:alias"="vf:1" 4) Boot an instance with the said flavor and observe a VF being set to 'allocated' and a PF being set to 'unavailable' select * from pci_devices where deleted=0; 5) Delete the instance from step 4 and observe that the VF has been made available but the PF is still 'unavailable'. Both should be back to available if this was the only VF used.