Red Hat Bugzilla – Bug 1468625
[pci]/alias option is not present in the nova configuration on the computes
Last modified: 2017-07-21 09:44:22 EDT
Description of problem:
[pci]/alias option is not being configured by the compute manifest. This option is needed in the nova configuration file on compute nodes starting with Ocata. It is instead being configured by the nova::api class when it should havebeen configured by nova::compute.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
Assigning to openstack-nova for now until we determine conclusively where the pci passthough conf should be done (compute config vs scheduler config vs both)
Seems like [pci]/alias option is needed on the controllers. On removing [pci]/alias from the controllers, restarting the nova services and trying to launch an instance with the alias in flavor I see PCI alias nvme is not defined (HTTP 400) . However, I have not confirmed if [pci]/alias is needed on the compute nodes.
Link to Error: https://gist.github.com/smalleni/e84c6955675445adf1f750dc7a8eb01a#file-gistfile1-txt-L1
The are two important options related to PCI; passthrought_whitelist and alias.
The passthrought_whitelist is related to the hardwares installed on the compute hosts. There is nothing in relation with the controllers. In the case you find a relation you should open a bug report.
The alias is related to the API, it provides an easy way for users to request hardwares. That option should only be relevant in the controllers.
Based on Sahid's comment, which was my expectation after briefly reviewing the code, I'll close this as not a bug. However we should correct the (upstream?) docs the mention setting pci alias on the computes.