What is to be done about MGLRU?

7 hours ago

Memory reclaim in the Linux kernel is fragmented with two separate algorithms (traditional LRU and MGLRU) in mm/vmscan.c, causing duplication and maintenance issues.
Developers propose unifying the two reclaim subsystems by separating code into files, defining benchmarking workloads, identifying common features, and comparing implementations.
Kairui Song has pending MGLRU improvements, and debate exists over whether to split code first or integrate changes, with concerns about Git history and regression risks.
MGLRU shows benefits like lower locking contention and efficient lazy promotion but has issues with page-cache performance, jumpy metrics, and excessive page-flag usage on 32-bit systems.
On Android, MGLRU is used on millions of devices but causes problems like aggressive file-backed page reclaim and difficulty controlling reclaim timing, requiring vendor-specific hooks.
Future plans include unifying refault distance calculations, adding BPF hooks for customizable reclaim policies, and exposing MGLRU parameters via sysfs for better system awareness.

Hasty Briefsbeta