nvidia-gpu: add support for communication to the endpoint
The commit uses MCTP VDM protocol to read temperature sensor value from
the gpu.
The MCTP VDM protocol is an extension of the OCP Accelerator Management
Interface specification. [1]
Tested: Build an image for gb200nvl-obmc machine with the following
patches cherry picked. This patches are needed to enable the mctp stack.
https://gerrit.openbmc.org/c/openbmc/openbmc/+/79422
Restart the nvidiagpusensor service.
```
root@gb200nvl-obmc:~# systemctl start xyz.openbmc_project.nvidiagpusensor.service
```
The app is detecting entity-manager configuration on gb200nvl-obmc
machine. The app is also able to detect all the endpoints from the mctp
service dbus tree. The app is reading temperature sensor value from gpu
correctly and the temperature sensor is also present on redfish.
```
$ curl -k -u 'root:0penBmc' https://10.137.203.137/redfish/v1/Chassis/NVIDIA_GB200_1/Sensors/temperature_NVIDIA_GB200_GPU
{
"@odata.id": "/redfish/v1/Chassis/NVIDIA_GB200_1/Sensors/temperature_NVIDIA_GB200_GPU",
"@odata.type": "#Sensor.v1_2_0.Sensor",
"Id": "temperature_NVIDIA_GB200_GPU",
"Name": "NVIDIA GB200 GPU",
"Reading": 36.4375,
"ReadingRangeMax": 127.0,
"ReadingRangeMin": -128.0,
"ReadingType": "Temperature",
"ReadingUnits": "Cel",
"Status": {
"Health": "OK",
"State": "Enabled"
}
}%
root@gb200nvl-obmc:~# busctl tree xyz.openbmc_project.GpuSensor
└─ /xyz
└─ /xyz/openbmc_project
└─ /xyz/openbmc_project/sensors
└─ /xyz/openbmc_project/sensors/temperature
└─ /xyz/openbmc_project/sensors/temperature/NVIDIA_GB200_GPU
root@gb200nvl-obmc:~# busctl introspect xyz.openbmc_project.GpuSensor /xyz/openbmc_project/sensors/temperature/NVIDIA_GB200_GPU
NAME TYPE SIGNATURE RESULT/VALUE FLAGS
org.freedesktop.DBus.Introspectable interface - - -
.Introspect method - s -
org.freedesktop.DBus.Peer interface - - -
.GetMachineId method - s -
.Ping method - - -
org.freedesktop.DBus.Properties interface - - -
.Get method ss v -
.GetAll method s a{sv} -
.Set method ssv - -
.PropertiesChanged signal sa{sv}as - -
xyz.openbmc_project.Association.Definitions interface - - -
.Associations property a(sss) 1 "chassis" "all_sensors" "/xyz/openbmc… emits-change
xyz.openbmc_project.Sensor.Value interface - - -
.MaxValue property d 127 emits-change
.MinValue property d -128 emits-change
.Unit property s "xyz.openbmc_project.Sensor.Value.Unit.… emits-change
.Value property d 36.3125 emits-change writable
xyz.openbmc_project.Sensor.ValueMutability interface - - -
.Mutable property b true emits-change
xyz.openbmc_project.State.Decorator.Availability interface - - -
.Available property b true emits-change writable
xyz.openbmc_project.State.Decorator.OperationalStatus interface - - -
.Functional property b true emits-change
```
[1] https://www.opencompute.org/documents/ocp-gpu-accelerator-management-interfaces-v1-pdf
Change-Id: Ied938b9e5c19751ee283b4b948e16c905c78fb48
Signed-off-by: Harshit Aghera <haghera@nvidia.com>
diff --git a/src/nvidia-gpu/OcpMctpVdm.cpp b/src/nvidia-gpu/OcpMctpVdm.cpp
new file mode 100644
index 0000000..6497621
--- /dev/null
+++ b/src/nvidia-gpu/OcpMctpVdm.cpp
@@ -0,0 +1,88 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION &
+ * AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "OcpMctpVdm.hpp"
+
+#include <endian.h>
+
+#include <cerrno>
+#include <cstdint>
+#include <cstring>
+#include <span>
+
+namespace ocp
+{
+namespace accelerator_management
+{
+
+int packHeader(uint16_t pciVendorId, const BindingPciVidInfo& hdr,
+ BindingPciVid& msg)
+{
+ if (hdr.ocp_accelerator_management_msg_type !=
+ static_cast<uint8_t>(MessageType::RESPONSE) &&
+ hdr.ocp_accelerator_management_msg_type !=
+ static_cast<uint8_t>(MessageType::REQUEST))
+ {
+ return EINVAL;
+ }
+
+ if (hdr.instance_id > instanceMax)
+ {
+ return EINVAL;
+ }
+
+ msg.instance_id = hdr.instance_id & instanceIdBitMask;
+
+ if (hdr.ocp_accelerator_management_msg_type ==
+ static_cast<uint8_t>(MessageType::REQUEST))
+ {
+ msg.instance_id |= requestBitMask;
+ }
+ else
+ {
+ msg.instance_id &= ~requestBitMask;
+ }
+
+ msg.pci_vendor_id = htobe16(pciVendorId);
+ msg.instance_id &= ~instanceIdReservedBitMask;
+ msg.ocp_version = ocpVersion & ocpVersionBitMask;
+ msg.ocp_version |= (ocpType << ocpTypeBitOffset) & ocpTypeBitMask;
+ msg.ocp_accelerator_management_msg_type = hdr.msg_type;
+
+ return 0;
+}
+
+int decodeReasonCodeAndCC(const std::span<const uint8_t> buf,
+ CompletionCode& cc, uint16_t& reasonCode)
+{
+ if (buf.size() <
+ sizeof(ocp::accelerator_management::CommonNonSuccessResponse))
+ {
+ return EINVAL;
+ }
+
+ // These expression decodes data communicated over the network
+ // The use of reinterpret_cast enables direct memory access to raw byte
+ // buffers without doing unnecessary data copying
+ // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+ const auto* response = reinterpret_cast<
+ const ocp::accelerator_management::CommonNonSuccessResponse*>(
+ buf.data());
+
+ cc = static_cast<CompletionCode>(response->completion_code);
+ if (cc == CompletionCode::SUCCESS)
+ {
+ reasonCode = 0;
+ return 0;
+ }
+
+ // reason code is expected to be present if CC != SUCCESS
+ reasonCode = le16toh(response->reason_code);
+
+ return 0;
+}
+} // namespace accelerator_management
+} // namespace ocp