Unifying sources of page hotness information
summary
A centralized PG hot subsystem collects access reports from NUMA hint faults, PTE access-bit scans (KLRU/KLRU scanD), device samples like AMD IBS and CXL HMU, and folio-mark-access hooks, records one unsigned long per PFN (about 2% overhead per TB of lower-tier memory), stores a 3-bit access count plus timestamp and a ready-to-migrate bit, and hands hottest PFNs to a K_migr kernel thread for migration. Early results match existing NUMA-B2 promotion when thresholds are tuned (default is two accesses, settable to one), hardware hints give better speed, and section-level hotness tracking cuts scan overhead.
AI-written summary · about