On the command line, one can specify a NUMA node for NVDIMM devices. If we set up the topology to give NVDIMMs their own nodes, i.e. not containing any CPUs or regular memory, qemu doesn't populate SRAT memory affinity structures for these nodes. However the NFIT does reference those proximity domains.
As a result, Linux, while parsing the SRAT, fails to initialize node related structures for these nodes, and they never end up in the nodes_possible map. When these are onlined at a later point (via hotplug), this causes problems. I've followed the instructions in bios-tables-test.c to update the expected SRAT binary, and the tests (make check) pass. Patches 1 and 3 are the relevant ones for the binary update. Patch 2 is the main patch which changes SRAT generation. Vishal Verma (3): diffs-allowed: add the SRAT AML to diffs-allowed hw/acpi-build: account for NVDIMM numa nodes in SRAT tests/acpi: update expected SRAT files hw/i386/acpi-build.c | 20 ++++++++++++++++++++ tests/data/acpi/pc/SRAT.dimmpxm | Bin 392 -> 392 bytes tests/data/acpi/q35/SRAT.dimmpxm | Bin 392 -> 392 bytes 3 files changed, 20 insertions(+) -- 2.25.4
