
Previously, the libvirt driver defaulted to 1024 * (# of CPUs) for the value of domain/cputune/shares in the libvirt XML. This value is then passed directly by libvirt to the cgroups API. Cgroups v2 imposes a maximum value of 10000 that can be passed in. This makes Nova unable to launch instances with more than 9 CPUs on hosts that run cgroups v2, like Ubuntu Jammy or RHEL 9. Fix this by just removing the default entirely. Because there is no longer a guarantee that domain/cputune will contain at least a shares element, we can stop always generating the former, and only generate it if it will actually contain something. We can also make operators's lives easier by leveraging the fact that we update the XML during live migration, so this patch also adds a method to remove the shares value from the live migration XML if one was not set as the quota:cpu_shares flavor extra spec. For operators that *have* set this extra spec to something greater than 10000, their flavors will have to get updates, and their instances resized. Partial-bug: 1978489 Change-Id: I49d757f5f261b3562ada27e6cf57284f615ca395
16 lines
888 B
YAML
16 lines
888 B
YAML
upgrade:
|
|
- |
|
|
In the libvirt driver, the default value of the ``<cputune><shares>``
|
|
element has been removed, and is now left to libvirt to decide. This is
|
|
because allowed values are platform dependant, and the previous code was
|
|
not guaranteed to be supported on all platforms. If any of your flavors are
|
|
using the quota:cpu_shares extra spec, you may need to resize to a
|
|
supported value before upgrading.
|
|
|
|
To facilitate the transition to no Nova default for ``<cputune><shares>``,
|
|
its value will be removed during live migration unless a value is set in
|
|
the ``quota:cpu_shares`` extra spec. This can cause temporary CPU
|
|
starvation for the live migrated instance if other instances on the
|
|
destination host still have the old default ``<cputune><shares>`` value. To
|
|
fix this, hard reboot, cold migrate, or live migrate the other instances.
|