On Thu, Nov 07, 2013 at 01:41:59PM +0100, Vasilis Liaskovitis wrote:
This patch adds a _PXM method to ACPI CPU objects for the pc machine. The _PXM value is derived from the passed in guest info, same way as CPU SRAT entries.
The motivation for this patch is a CPU hot-unplug/hot-plug bug observed when using a 3.11 linux guest kernel on a multi-NUMA node qemu/kvm VM. The linux guest kernel parses the SRAT CPU entries at boot time and stores them in the array __apicid_to_node. When a CPU is hot-removed, the linux guest kernel resets the removed CPU's __apicid_to_node entry to NO_NUMA_NODE (kernel commit c4c60524). When the removed cpu is hot-added again, the linux kernel looks up the hot-added cpu object's _PXM method instead of somehow re-discovering the SRAT entry info. With current qemu/seabios, the _PXM method is not found, and the CPU is thus hot-plugged in the default NUMA node 0. (The problem does not show up on initial hotplug of a cpu; the PXM method is still not found in this case, but the kernel still has the correct proximity value from the CPU's SRAT entry stored in __apicid_to_node)
ACPI spec mentions that the _PXM method is the correct way to determine proximity information at hot-add time.
Where does it say this? I found this: If the Local APIC ID / Local SAPIC ID / Local x2APIC ID of a dynamically added processor is not present in the System Resource Affinity Table (SRAT), a _PXM object must exist for the processor’s device or one of its ancestors in the ACPI Namespace.
Does this mean that linux is buggy, and should be fixed up to look up the apic ID in SRAT?
So far, qemu/seabios do not provide this method for CPUs. So regardless of kernel behaviour, it is a good idea to add this _PXM method. Since ACPI table generation has recently been moved from seabios to qemu, we do this in qemu.
Note that the above hot-remove/hot-add scenario has been tested on an older qemu + non-upstreamed patches for cpu hot-removal support, and not on qemu master (since cpu-del support is still not on master). The only testing done with qemu/seabios master and this patch, are successful boots of multi-node linux and windows8 guests.
For the initial discussion on seabios and linux-acpi lists see http://www.spinics.net/lists/linux-acpi/msg47058.html
Signed-off-by: Vasilis Liaskovitis vasilis.liaskovitis@profitbricks.com Reviewed-by: Thilo Fromm t-lo@thilo-fromm.de
Even if this is a linux bug, I have no issue with working around it in qemu.
But I think proper testing needs to be done with rebased upport for cpu-del.
hw/i386/acpi-build.c | 2 ++ hw/i386/ssdt-proc.dsl | 2 ++ 2 files changed, 4 insertions(+)
diff --git a/hw/i386/acpi-build.c b/hw/i386/acpi-build.c index 6cfa044..9373f5e 100644 --- a/hw/i386/acpi-build.c +++ b/hw/i386/acpi-build.c @@ -603,6 +603,7 @@ static inline char acpi_get_hex(uint32_t val) #define ACPI_PROC_OFFSET_CPUHEX (*ssdt_proc_name - *ssdt_proc_start + 2) #define ACPI_PROC_OFFSET_CPUID1 (*ssdt_proc_name - *ssdt_proc_start + 4) #define ACPI_PROC_OFFSET_CPUID2 (*ssdt_proc_id - *ssdt_proc_start) +#define ACPI_PROC_OFFSET_CPUPXM (*ssdt_proc_pxm - *ssdt_proc_start) #define ACPI_PROC_SIZEOF (*ssdt_proc_end - *ssdt_proc_start) #define ACPI_PROC_AML (ssdp_proc_aml + *ssdt_proc_start)
@@ -724,6 +725,7 @@ build_ssdt(GArray *table_data, GArray *linker, proc[ACPI_PROC_OFFSET_CPUHEX+1] = acpi_get_hex(i); proc[ACPI_PROC_OFFSET_CPUID1] = i; proc[ACPI_PROC_OFFSET_CPUID2] = i;
proc[ACPI_PROC_OFFSET_CPUPXM] = guest_info->node_cpu[i]; } /* build this code:diff --git a/hw/i386/ssdt-proc.dsl b/hw/i386/ssdt-proc.dsl index 8229bfd..7eef8b2 100644 --- a/hw/i386/ssdt-proc.dsl +++ b/hw/i386/ssdt-proc.dsl @@ -47,6 +47,8 @@ DefinitionBlock ("ssdt-proc.aml", "SSDT", 0x01, "BXPC", "BXSSDT", 0x1)
- also updating the C code.
*/ Name(_HID, "ACPI0007")
ACPI_EXTRACT_NAME_BYTE_CONST ssdt_proc_pxmName(_PXM, 0xAA)
The ACPI spec says this should be a DWORD value:
Return Value: An Integer (DWORD) containing a proximity domain identifier.
External(CPMA, MethodObj) External(CPST, MethodObj) External(CPEJ, MethodObj)-- 1.7.10.4