Hello Tejun, On Wed, May 20, 2026 at 09:52 AM Tejun Heo wrote: > I'm not sure about complicating dmem control model without implementing > reclaim. What are we slowing them down for if the only recovery action is > killing them?
Thank you for the feedback. Your point about the lack of a reclaim path is well-taken. Simple throttling without a way to recover resources is indeed incomplete and inconsistent with the cgroup v2 philosophy. To address this from several perspectives in v2: 1. Recovery Path: As suggested by Maarten Lankhorst, we will pivot to a reclaim-centric model. Exceeding `dmem.high` will trigger a prioritized eviction process, where memory objects from over-limit cgroups are targeted first for reclaim. This provides the meaningful "recovery action" you mentioned. 2. Backpressure: Throttling will then serve as a secondary tool to synchronize user-space demand with the kernel's reclaim speed, preventing bursty workloads from overwhelming the system before reclaim can finish. 3. Graceful Degradation: For GPU compute jobs, this model provides a managed "pressure point" that allows transient peaks to be handled via rebalancing rather than immediate, fatal allocation failures (max/OOM). The goal for v2 is to achieve convergence with the `memory.high` model, pairing prioritized reclaim with backpressure. Thanks, Qiliang
