This bug affects AMD family 12h and newer.

The device numa_node information in the sysfs for each PCI device (e.g.
/devices/pci0000:50/0000:50:00.2/numa_node) is showing value -1.

In file arch/x86/pci/acpi.c, in function pci_acpi_scan_root(), it is
queries the node information as following:

#ifdef CONFIG_ACPI_NUMA
      pxm = acpi_get_pxm(device->handle);
      if (pxm >= 0)
          node = pxm_to_node(pxm);
      if (node != -1)
          set_mp_bus_to_node(busnum, node);
      else
#endif
          node = get_mp_bus_to_node(busnum);

In this case, I see that the acpi_get_pxm() returns -1.  Therefore, it
falls back to using the node information in mp_bus_to_node[].  For AMD systems, the arch/x86/pci/amd_bus.c only probe the hostbridge for family10h and 11h, and showing the correct numa_node information.  For the newer systems, it does not fill the mp_bus_to_node[], and result in the value -1.

Please refer to the following email thread for more information:
http://marc.info/?l=linux-kernel&m=139405365524526&w=2