Solving Zombie Cgroups - Simple Fix or Smart Solution?
summary
Kernel fixes reparent stuck pages to parent memory cgroup to free zombie cgroup metadata, implemented by moving page charges and updating folio locking so folio→mmc binding stays stable. The simple approach works upstream-minded but causes lost page hotness on LRU (pages may be mis-ordered young/old), which can harm reclamation under repeated create/destroy workloads. Prior recharging proposal was rejected for user-semantic and consistency reasons. The session seeks real workload data where this reparent-only tradeoff is harmful.
AI-written summary · about
chapters (11)
- 0:00Introduction to cgroups
- 1:05Memory cgroups and shared pages
- 3:33Problem of zombie cgroups
- 4:43Proposed solution overview
- 10:10Discussion on reparenting approach
- 18:32Alternative solutions and concerns
- 22:49Implementation details and challenges
- 26:29Q&A on page sharing and accounting
- 31:43Metadata overhead and accuracy
- 34:11Reclaiming pages and trade-offs
- 37:10Summary and closing remarks
related talks
27mWhen Linux Memory Accounting Goes Wrong
18mIO and cgroups, the Current and Future Work
35mFuture-Proof Your Applications: Avoid Breakdowns With Cgroup Chan...
38mUnderstanding and Working with the Cgroups Interface
34mCgroups and Enterprise Users
43mHybrid THP Mechanism. Selective Use of Huge Pages by Hot Applications
42mcgroup2 Resource Control strategies w/ resctl-demo
23mImproving data placement for Zoned Linux File systems