Inside Linux cgroup v2 memory.high Throttling and Direct Reclaim Backpressure

How exceeding memory.high injects proportional schedule_timeout delays into allocating threads before memory.max triggers the OOM killer.

1. Why Your Service Froze Without an OOM Kill

Unlike `memory.max` (which invokes the cgroup OOM killer when reclaim fails), `memory.high` in cgroup v2 is a soft throttle boundary. When a container's page cache + anonymous memory crosses `memory.high`, the kernel's `mem_cgroup_handle_over_high()` hook penalizes the allocating task on return to userspace, forcing synchronous page reclaim and sleeping up to 2 seconds per allocation call if overage persists.

$ cat /sys/fs/cgroup/system.slice/storage.service/memory.pressure
some avg10=18.42 avg60=12.10 avg300=4.05 total=84920144
full avg10=9.15 avg60=5.40 avg300=1.12 total=31204912

Monitoring Pressure Stall Information (`memory.pressure`) and setting `memory.high` at 92% of `memory.max` only alongside background cache eviction callbacks prevents silent multi-second allocation freezes.

← Older Note (2024-05-26)CRC32C vs. XXH3 vs. BLAKE3 for Storage Block Checksumming on x86_64 and ARM64Newer Note (2024-08-30) →Adaptive Concurrency Limits Using Little's Law and Gradient2 Queue Controllers

← Back to all notes · View full chronological archive (24)