-
Notifications
You must be signed in to change notification settings - Fork 190
Description
If a NUMA node contains CPUs but its memory is offline for now and exposed as DAX (either because its non-volatile or soft-reserved), we should be able to attach that DAX device directly near these CPUs. The current code uses the sysfs numa_node attribute which points to the closest online NUMA node at boot, so not the right CPUs.
DAX devices also have a "target_node" sysfs attribute that points to the right NUMA node. So we should attach there. But this node doesn't exist in hwloc and in /sys/devices/system/node/nodeX because that directory doesn't appear either when there's no memory, hence we can't read its cpumap or find its locality at all.
There was some discussion about exposing SLIT/HMAT/... locality info for those offline nodes but nothing seems available for now.
Fortunately this case should be rare since most platform will have some DRAM near each CPU.