OpenStack Nova's `enabled_perf_events`, RHEL 7.5, and Intel CMT

Issue

If you have enabled performance events for Intel Cache Monitoring Technology in Nova, instances will fail to start after upgrading to RHEL 7.5.

Details of the issue

Nova supports monitoring performance of guests via specifying events (e.g. 'cmt' to get the usage of CPU L3 cache by an application; full list of possible events is here) through a config attribute, enabled_perf_events in the [libvirt] section of /etc/nova/nova.conf. So, if you are using that to monitor events for Intel Cache Monitoring Technology (CMT)[1] as follows:

    [libvirt]
    enabled_perf_events = cmt, mbml, mbmt

The impact is two fold:

• Once you have upgraded to RHEL 7.5 kernel, it will result in Nova instances not starting anymore.

• It also affects live migration: I.e. if you have performance events enabled on a RHEL 7.4 Compute node, you will not be able to live migrate an instance from that RHEL 7.4 Compute node to a RHEL 7.5 Compute node -- until you have both disabled perf events and restarted the Nova instance.

Note that by default Nova does not enable the Intel CMT via enabled_perf_events -- i.e. the out-of-the-box behaviour is not affected. So, if you have not used enabled_perf_events in your OpenStack environment, you can completely disregard this article.

[1] Intel's Cache Monitoring Technology has two acronyms: "CQM" in the Linux kernel parlance and "CMT" in the libvirt XML elements, which is what OpenStack Nova uses in its config attribute.)