WIP: Add limits.cpu.pin_strategy setting to disable CPU auto pinning by default

[kadinsayani]

Fixes #14133.

This pull request introduces a configuration option for instances, limits.cpu.pin_strategy. This option allows users to specify the strategy for CPU auto pinning. The possible values are none (disables CPU auto pinning - also the new default) and auto (enables CPU auto pinning).

[github-actions]

Heads up @mionaalex - the "Documentation" label was applied to this issue.

[kadinsayani]

If we modify the current tests in lxd-ci to launch a VM with the limits.cpu.pin_strategy set, there will be a regression since earlier LXD versions will return Error: Failed instance creation: Failed creating instance record: Unknown configuration key: limits.cpu.pin_strategy, causing the cpu-vm tests to fail.

Currently, the checks in lxd-ci for cpu pinning only run if the feature is present:

if journalctl --quiet --no-hostname --no-pager --boot=0 --unit=snap.lxd.daemon.service | grep "Scheduler: virtual-machine"; then
...

However, due to my implementation of checking the config setting in deviceTaskBalance, the log message Scheduler: virtual-machine is printed regardless of whether auto cpu pinning is enabled or not. Hence, I have added a new log message which we can use to check if the feature is present.

if c.Type() == instancetype.VM {
	if conf["limits.cpu.pin_strategy"] == "none" {
		continue
	} else if conf["limits.cpu.pin_strategy"] == "auto" {
		logger.Debugf("CPU auto pinning enabled for %q", c.Name())
	}
}

Therefore, I think the only change required (for now) to ensure we don't introduce a regression to lxd-ci is to change the check to look for the new log message.

For reference: canonical/lxd-ci@972ef98.

Please let me know if you have any suggestions 😄

[tomponline]

If we modify the current tests in lxd-ci to launch a VM with the limits.cpu.pin_strategy set, there will be a regression since earlier LXD versions will return Error: Failed instance creation: Failed creating instance record: Unknown configuration key: limits.cpu.pin_strategy, causing the cpu-vm tests to fail.

You will need to introduce a new API extension for the new config key and can test for support of that when testing to enable the functionality.

[simondeziel]

@kadinsayani I couldn't find any reference to back that but I think it's best for the commit adding the API extension to be the first in the PR. This way, when @tomponline does the backporting (if any, dunno for this one) he can quickly identify that a collection of commits need to be considered at once for the backport.

[kadinsayani]

@kadinsayani I couldn't find any reference to back that but I think it's best for the commit adding the API extension to be the first in the PR. This way, when @tomponline does the backporting (if any, dunno for this one) he can quickly identify that a collection of commits need to be considered at once for the backport.

Sounds good, thanks for letting me know!

[kadinsayani]

@tomponline will we be back porting the commit containing the new API extension to 6.1? Currently, the tests in lxd-ci are relying on LXD daemon logs for checking if the feature is present:

https://github.com/mihalicyn/lxd-ci/blob/972ef98d660c3c2efba3d8aa7408126a1b555ff9/tests/cpu-vm#L12

[tomponline]

@tomponline will we be back porting the commit containing the new API extension to 6.1? Currently, the tests in lxd-ci are relying on LXD daemon logs for checking if the feature is present:

yes we can because it defaults to disabled

[tomponline]

@kadinsayani please can you ensure that the change to disable auto pin for containers is separate so i can choose to backport it or not.