A move_pages() equivalent for physical memory
summary
Proposal adds a syscall that migrates physical pages between NUMA nodes by reusing move_pages infrastructure, accepting device or host PFNs and requiring CAP_SYS_ADMIN; prototype targets x86 Linux 6.6 and still performs rmap/VMA checks (cgroup/cpuset validation) before moving. It exists because devices report only device physical addresses and reverse-translation is costly — 90 seconds to reverse-map 512GB under load — and offloaded faulting or prefetch behavior can hide hot pages from IBS/PEBS, so policy, security and standardization remain unresolved. move_phys_pages syscall
AI-written summary · about
chapters (9)
- 0:00Introduction and thesis
- 1:05Memory tiering challenges
- 3:54Current monitoring mechanisms
- 8:11Hardware vendor proposals
- 13:26Reverse mapping overhead
- 17:18Move Fiz Pages implementation
- 22:40Kernel vs user space debate
- 25:47Prototyping and standardization
- 28:56Use cases and future
related talks
30mFlexible Paging in Linux: Per-process Page Size
44mLarge Pages in Linux
47mLinux's Leap Beyond Fixed Pages
11mAdvanced Parallel Memory Virtualization
12mPageFlex: Flexible and Efficient User-space Delegation of Linux Paging Policies...
26mCooperative DMA in a memory-oversubscribed environment
15mNomad: Non-Exclusive Memory Tiering via Transactional Page Migration
39mMM101: Introduction to Linux Memory Management