As reported here, there’s currently a bug that makes kworker events_unbound eating 100% CPU as a single thread, which according to perf is caused by function btrfs_preempt_reclaim_metadata_space called repeatedly.
In the same report, there’s a list of patches that seem to have solved this problem. What I want to know is whether 5.12.1-2 already contains them (thus invalidating the claim) or not (chance of having them in the next version)?
Yep, none. As all of those patches I point has btrfs in their name. Guess it can be considered for addition if upstream doesn’t include them in the next release. Right now, I think the most obvious way to avoid this problem is to go back to 5.10.X.