aboutsummaryrefslogtreecommitdiff
path: root/samples/pidfd
diff options
context:
space:
mode:
authorYazen Ghannam2023-05-15 11:35:37 +0000
committerBorislav Petkov (AMD)2023-06-19 13:01:44 +0200
commit4251566ebc1cf95ae26a1e5a24cdac1ac25e942f (patch)
treec1e5ce69dee3f22fbbd81987520ff72945b1a82c /samples/pidfd
parent9c42edd571aa4f8b2125b71e3924eeb0f6a54af1 (diff)
EDAC/amd64: Cache and use GPU node map
AMD systems have historically provided an "AMD Node ID" that is a unique identifier for each die in a multi-die package. This was associated with a unique instance of the AMD Northbridge on a legacy system. And now it is associated with a unique instance of the AMD Data Fabric on modern systems. Each instance is referred to as a "Node"; this is an AMD-specific term not to be confused with NUMA nodes. The data fabric provides a number of interfaces accessible through a set of functions in a single PCI device. There is one PCI device per Data Fabric (AMD Node), and multi-die systems will see multiple such PCI devices. The AMD Node ID matches a Node's position in the PCI hierarchy. For example, the Node 0 is accessed using the first PCI device, Node 1 is accessed using the second, and so on. A logical CPU can find its AMD Node ID using CPUID. Furthermore, the AMD Node ID is used within the hardware fabric, so it is not purely a logical value. Heterogeneous AMD systems, with a CPU Data Fabric connected to GPU data fabrics, follow a similar convention. Each CPU and GPU die has a unique AMD Node ID value, and each Node ID corresponds to PCI devices in sequential order. However, there are two caveats: 1) GPUs are not x86, and they don't have CPUID to read their AMD Node ID like on CPUs. This means the value is more implicit and based on PCI enumeration and hardware-specifics. 2) There is a gap in the hardware values for AMD Node IDs. Values 0-7 are for CPUs and values 8-15 are for GPUs. For example, a system with one CPU die and two GPUs dies will have the following values: CPU0 -> AMD Node 0 GPU0 -> AMD Node 8 GPU1 -> AMD Node 9 EDAC is the only subsystem where this has a practical effect. Memory errors on AMD systems are commonly reported through MCA to a CPU on the local AMD Node. The error information is passed along to EDAC where the AMD EDAC modules use the AMD Node ID of reporting logical CPU to access AMD Node information. However, memory errors from a GPU die will be reported to the CPU die. Therefore, the logical CPU's AMD Node ID can't be used since it won't match the AMD Node ID of the GPU die. The AMD Node ID of the GPU die is provided as part of the MCA information, and the value will match the hardware enumeration (e.g. 8-15). Handle this situation by discovering GPU dies the same way as CPU dies in the AMD NB code. But do a "node id" fixup in AMD64 EDAC where it's needed. The GPU data fabrics provide a register with the base AMD Node ID for their local "type", i.e. GPU data fabric. This value is the same for all fabrics of the same type in a system. Read and cache the base AMD Node ID from one of the GPU devices during module initialization. Use this to fixup the "node id" when reporting memory errors at runtime. [ bp: Squash a fix making gpu_node_map static as reported by Tom Rix <trix@redhat.com>. Link: https://lore.kernel.org/r/20230610210930.174074-1-trix@redhat.com ] Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com> Co-developed-by: Muralidhara M K <muralidhara.mk@amd.com> Signed-off-by: Muralidhara M K <muralidhara.mk@amd.com> Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de> Link: https://lore.kernel.org/r/20230515113537.1052146-6-muralimk@amd.com
Diffstat (limited to 'samples/pidfd')
0 files changed, 0 insertions, 0 deletions