Package: kernel
Version: 6.12.94-1
After upgrading from 6.1 series kernel to 6.12.x series kernel, a
performance regression was encountered.
The customer is experiencing:
~2x slower execution times for write-heavy workloads under tight memory
cgroup limits (~256Mi)
Significant latency spikes (median 14s vs 7.1s on kernel 6.1; max 23.5s
vs 7.6s)
Occasional premature OOM conditions
Root Cause (Identified by Customer)
The issue is an upstream Linux kernel MGLRU regression:
Regression commit: 14aa8b2d5c2e ("mm/mglru: don't sync disk for each
aging cycle") – removed waking flusher threads during MGLRU reclaim,
causing dirty page accumulation and reclaim stalls under cgroup memory
pressure.
Fix commit: 1bc542c6a0d1 ("mm/vmscan: wake up flushers conditionally to
avoid cgroup OOM") – available in mainline ~6.13/6.14 but not backported
to the 6.12.x stable series (no Cc: stable tag). Ubuntu had to manually
SRU it.
Validation: Disabling MGLRU (/sys/kernel/mm/lru_gen/enabled=0) on kernel
6.12 nodes restores near-6.1 performance and eliminates the latency
tail, confirming the MGLRU reclaim path as the cause.
Please include the fix commit in the 6.12 series kernel.
