drm and cgroup memory management - two great tastes
summary
Integration work patches add memcg counters for GPU active and reclaimable memory, port TTM pools to list_lru and make the TTM shrinker NUMA-aware, and account system-memory GEM objects on first allocation (driver opt‑in). Target use case is unified‑memory servers like MI300A where DMA‑buf hacks were a stopgap. Eviction semantics for objects that move between device and system memory remain unresolved, and DMA/CMA corner cases are deferred. Patchset is posted and awaiting broader review.
AI-written summary · about
chapters (6)
- 0:00Introduction and background
- 3:57Why cgroups and NUMA matter now
- 10:35DRM memory management basics
- 14:10Challenges with cgroup integration
- 21:11Merging DRM and cgroup memory
- 27:55Q&A and next steps
related talks
43mLive in a world with multiple memory types
26mCXL Memory Tiering for heterogenous computing
24mConsolidating representations of the physical memory
26mCooperative DMA in a memory-oversubscribed environment
24mA Virtual IOMMU With Cooperative DMA Buffer Tracking
39mTaming DMA: Tales Wrestling Memory Corruption
21mCgroup Slab Memory Controller and Time Namespace
43mHybrid THP Mechanism. Selective Use of Huge Pages by Hot Applications