MGLRU is classic example of how the mm's review process has been extremely poor.
It's basically unmaintained, it's got questionable performance characteristics beyond google-specific stuff.
It was merged with only tags from people internal to implementing it and random people, nobody core to reclaim.
It had pushback about refactoring which was ignored.
MM review is in a MUCH better place now but there are several such examples of this, stuff we now have to maintain.
https://lore.kernel.org/linux-mm/aaBsrrmV25FTIkVX@casper.infradead.org/
3
0
1
MGLRU _as a whole_ might be fine by the way, it's just there was no way we should have accepted it as-is at the time.
1
0
1
@haoli yeah, it's super heuristic and very very sensitive to even the smallest thing.
Recently saw a patch that was done a while ago that significantly altered behaviour for compaction on NUMA systems for instance.
Reclaim is difficult code to work on/with :)
0
0
2
https://lore.kernel.org/lkml/67f1588e-a094-413c-9746-9c28bdf6cc56@lucifer.local/ well time to be more direct about this :)
1
0
0
Vlastimil Babka 🇨🇿🇪🇺🇺🇦
vbabka@mastodon.social@ljs I don't think it's a classic example however. Here it's not about the lack of review flags. The relevant people discussed it at length at lsf/mm 2022 and, knowing the downsides, decided to give it a chance. https://lwn.net/Articles/894859/
1
0
0
@vbabka *tears in eyes* fuck michal then *weeps* sorry michal forgive me FORGIVE ME *downs pint*
0
0
0