Even more than 20 years on, people are finding new ways to improve kernel internals. This is the sort of subtle, intelligent work that distinguishes excellent software from the rest.

As long as people keep finding ways to regress the kernel, there will be opportunities created to fix those regressions.

AFAICT the performance win is coming from undoing a performance loss stemming from relatively recently added CONFIG_SCHED_MM_CID mechanism[0] (2022).

[0] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/lin...