| Harshit Aghera | 1180ed4 | 2025-09-30 16:55:39 +0530 | [diff] [blame] | 1 | /* |
| 2 | * SPDX-FileCopyrightText: Copyright OpenBMC Authors |
| 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | */ |
| 5 | |
| 6 | #include "NvidiaPciePortMetrics.hpp" |
| 7 | |
| 8 | #include "NvidiaUtils.hpp" |
| 9 | #include "Utils.hpp" |
| 10 | |
| 11 | #include <bits/basic_string.h> |
| 12 | |
| 13 | #include <MctpRequester.hpp> |
| 14 | #include <NvidiaGpuMctpVdm.hpp> |
| 15 | #include <NvidiaPcieDevice.hpp> |
| 16 | #include <OcpMctpVdm.hpp> |
| 17 | #include <phosphor-logging/lg2.hpp> |
| 18 | #include <sdbusplus/asio/connection.hpp> |
| 19 | #include <sdbusplus/asio/object_server.hpp> |
| 20 | #include <sdbusplus/message/native_types.hpp> |
| 21 | |
| 22 | #include <cstddef> |
| 23 | #include <cstdint> |
| 24 | #include <format> |
| 25 | #include <functional> |
| 26 | #include <memory> |
| 27 | #include <span> |
| 28 | #include <string> |
| 29 | #include <system_error> |
| 30 | #include <vector> |
| 31 | |
| 32 | using std::string; |
| 33 | |
| 34 | using namespace std::literals; |
| 35 | |
| 36 | constexpr const char* metricInterface = "xyz.openbmc_project.Metric.Value"; |
| 37 | |
| 38 | NvidiaPciePortMetrics::NvidiaPciePortMetrics( |
| 39 | std::shared_ptr<sdbusplus::asio::connection>& conn, |
| 40 | mctp::MctpRequester& mctpRequester, const std::string& name, |
| 41 | const std::string& pcieDeviceName, const std::string& path, uint8_t eid, |
| 42 | gpu::PciePortType portType, uint8_t upstreamPortNumber, uint8_t portNumber, |
| 43 | sdbusplus::asio::object_server& objectServer, uint8_t scalarGroupId, |
| 44 | const std::vector<NvidiaMetricInfo>& metricsInfo) : |
| 45 | eid(eid), portType(portType), upstreamPortNumber(upstreamPortNumber), |
| 46 | portNumber(portNumber), scalarGroupId(scalarGroupId), path(path), |
| 47 | conn(conn), mctpRequester(mctpRequester) |
| 48 | { |
| 49 | const std::string metricsDbusPathPrefix = |
| 50 | metricPath + std::format("port_{}_{}", pcieDeviceName, name); |
| 51 | |
| 52 | const sdbusplus::message::object_path portDbusPath = |
| 53 | sdbusplus::message::object_path(pcieDevicePathPrefix) / pcieDeviceName / |
| 54 | name; |
| 55 | |
| 56 | for (const auto& [id, name] : metricsInfo) |
| 57 | { |
| 58 | const std::string metricsDbusPath = metricsDbusPathPrefix + name; |
| 59 | |
| 60 | metricValueInterfaces[id] = |
| 61 | objectServer.add_interface(metricsDbusPath, metricInterface); |
| 62 | metricValueInterfaces[id]->register_property( |
| 63 | "Unit", "xyz.openbmc_project.Metric.Value.Unit.Count"s); |
| 64 | metricValueInterfaces[id]->register_property("Value", 0.0); |
| 65 | |
| 66 | std::vector<Association> associations; |
| 67 | associations.emplace_back("measuring", "measured_by", portDbusPath); |
| 68 | |
| 69 | metricAssociationInterfaces[id] = |
| 70 | objectServer.add_interface(metricsDbusPath, association::interface); |
| 71 | metricAssociationInterfaces[id]->register_property("Associations", |
| 72 | associations); |
| 73 | |
| 74 | if (!metricValueInterfaces[id]->initialize()) |
| 75 | { |
| 76 | lg2::error( |
| 77 | "Error initializing PCIe Port Metric Interface for EID={EID}, " |
| 78 | "PortType={PT}, PortNumber={PN}, ScalarGroup={SG}, Metric={MN}", |
| 79 | "EID", eid, "PT", static_cast<uint8_t>(portType), "PN", |
| 80 | portNumber, "EID", eid, "PN", portNumber, "SG", scalarGroupId, |
| 81 | "MN", name); |
| 82 | } |
| 83 | |
| 84 | if (!metricAssociationInterfaces[id]->initialize()) |
| 85 | { |
| 86 | lg2::error( |
| 87 | "Error initializing PCIe Port Metric Association Interface for EID={EID}, " |
| 88 | "PortType={PT}, PortNumber={PN}, ScalarGroup={SG}, Metric={MN}", |
| 89 | "EID", eid, "PT", static_cast<uint8_t>(portType), "PN", |
| 90 | portNumber, "EID", eid, "PN", portNumber, "SG", scalarGroupId, |
| 91 | "MN", name); |
| 92 | } |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | void NvidiaPciePortMetrics::processResponse( |
| 97 | const std::error_code& sendRecvMsgResult, std::span<const uint8_t> response) |
| 98 | { |
| 99 | if (sendRecvMsgResult) |
| 100 | { |
| 101 | lg2::error( |
| 102 | "Error updating PCIe Port Metrics: sending message over MCTP failed, " |
| 103 | "rc={RC}, EID={EID}, PortType={PT}, PortNumber={PN}, ScalarGroup={SG}", |
| 104 | "RC", sendRecvMsgResult.message(), "EID", eid, "PT", |
| 105 | static_cast<uint8_t>(portType), "PN", portNumber, "SG", |
| 106 | scalarGroupId); |
| 107 | return; |
| 108 | } |
| 109 | |
| 110 | ocp::accelerator_management::CompletionCode cc{}; |
| 111 | uint16_t reasonCode = 0; |
| 112 | size_t numTelemetryValue = 0; |
| 113 | |
| 114 | int rc = gpu::decodeQueryScalarGroupTelemetryV2Response( |
| 115 | response, cc, reasonCode, numTelemetryValue, telemetryValues); |
| 116 | |
| 117 | if (rc != 0 || cc != ocp::accelerator_management::CompletionCode::SUCCESS) |
| 118 | { |
| 119 | lg2::error( |
| 120 | "Error updating PCIe Port Errors: decode failed, " |
| 121 | "rc={RC}, cc={CC}, reasonCode={RESC}, EID={EID}, PortType={PT}, PortNumber={PN}, ScalarGroup={SG}", |
| 122 | "RC", rc, "CC", static_cast<uint8_t>(cc), "RESC", reasonCode, "EID", |
| 123 | eid, "PT", static_cast<uint8_t>(portType), "PN", portNumber, "SG", |
| 124 | scalarGroupId); |
| 125 | return; |
| 126 | } |
| 127 | |
| 128 | for (size_t i = 0; i < numTelemetryValue; ++i) |
| 129 | { |
| 130 | if (metricValueInterfaces[i] != nullptr) |
| 131 | { |
| 132 | metricValueInterfaces[i]->set_property( |
| 133 | "Value", static_cast<double>(telemetryValues[i])); |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | void NvidiaPciePortMetrics::update() |
| 139 | { |
| 140 | auto rc = gpu::encodeQueryScalarGroupTelemetryV2Request( |
| 141 | 0, portType, upstreamPortNumber, portNumber, scalarGroupId, request); |
| 142 | |
| 143 | if (rc != 0) |
| 144 | { |
| 145 | lg2::error( |
| 146 | "Error updating PCIe Port Errors: encode failed, rc={RC}, EID={EID}, PortType={PT}, PortNumber={PN}, ScalarGroup={SG}", |
| 147 | "RC", rc, "EID", eid, "PT", static_cast<uint8_t>(portType), "PN", |
| 148 | portNumber, "SG", scalarGroupId); |
| 149 | return; |
| 150 | } |
| 151 | |
| 152 | mctpRequester.sendRecvMsg( |
| 153 | eid, request, |
| 154 | [weak{weak_from_this()}](const std::error_code& ec, |
| 155 | std::span<const uint8_t> buffer) { |
| 156 | std::shared_ptr<NvidiaPciePortMetrics> self = weak.lock(); |
| 157 | if (!self) |
| 158 | { |
| 159 | lg2::error("Invalid reference to NvidiaPciePortErrors"); |
| 160 | return; |
| 161 | } |
| 162 | self->processResponse(ec, buffer); |
| 163 | }); |
| 164 | } |
| 165 | |
| 166 | std::shared_ptr<NvidiaPciePortMetrics> makeNvidiaPciePortErrors( |
| 167 | std::shared_ptr<sdbusplus::asio::connection>& conn, |
| 168 | mctp::MctpRequester& mctpRequester, const std::string& name, |
| 169 | const std::string& pcieDeviceName, const std::string& path, uint8_t eid, |
| 170 | gpu::PciePortType portType, uint8_t upstreamPortNumber, uint8_t portNumber, |
| 171 | sdbusplus::asio::object_server& objectServer) |
| 172 | { |
| 173 | static constexpr uint8_t nvidiaPciePortErrorScalarGroupId = 2; |
| 174 | |
| 175 | return std::make_shared<NvidiaPciePortMetrics>( |
| 176 | conn, mctpRequester, name, pcieDeviceName, path, eid, portType, |
| 177 | upstreamPortNumber, portNumber, objectServer, |
| 178 | nvidiaPciePortErrorScalarGroupId, |
| 179 | std::vector<NvidiaMetricInfo>{ |
| 180 | {0, "/pcie/non_fatal_error_count"}, |
| 181 | {1, "/pcie/fatal_error_count"}, |
| 182 | {2, "/pcie/unsupported_request_count"}, |
| 183 | {3, "/pcie/correctable_error_count"}, |
| 184 | }); |
| 185 | } |
| 186 | |
| 187 | std::shared_ptr<NvidiaPciePortMetrics> makeNvidiaPciePortCounters( |
| 188 | std::shared_ptr<sdbusplus::asio::connection>& conn, |
| 189 | mctp::MctpRequester& mctpRequester, const std::string& name, |
| 190 | const std::string& pcieDeviceName, const std::string& path, uint8_t eid, |
| 191 | gpu::PciePortType portType, uint8_t upstreamPortNumber, uint8_t portNumber, |
| 192 | sdbusplus::asio::object_server& objectServer) |
| 193 | { |
| 194 | static constexpr uint8_t nvidiaPciePortCounterScalarGroupId = 4; |
| 195 | |
| 196 | return std::make_shared<NvidiaPciePortMetrics>( |
| 197 | conn, mctpRequester, name, pcieDeviceName, path, eid, portType, |
| 198 | upstreamPortNumber, portNumber, objectServer, |
| 199 | nvidiaPciePortCounterScalarGroupId, |
| 200 | std::vector<NvidiaMetricInfo>{ |
| 201 | {1, "/pcie/nak_received_count"}, |
| 202 | {2, "/pcie/nak_sent_count"}, |
| 203 | {4, "/pcie/replay_rollover_count"}, |
| 204 | {6, "/pcie/replay_count"}, |
| 205 | }); |
| 206 | } |
| 207 | |
| 208 | std::shared_ptr<NvidiaPciePortMetrics> makeNvidiaPciePortL0ToRecoveryCount( |
| 209 | std::shared_ptr<sdbusplus::asio::connection>& conn, |
| 210 | mctp::MctpRequester& mctpRequester, const std::string& name, |
| 211 | const std::string& pcieDeviceName, const std::string& path, uint8_t eid, |
| 212 | gpu::PciePortType portType, uint8_t upstreamPortNumber, uint8_t portNumber, |
| 213 | sdbusplus::asio::object_server& objectServer) |
| 214 | { |
| 215 | static constexpr uint8_t nvidiaPciePortL0ToRecoveryCountScalarGroupId = 3; |
| 216 | |
| 217 | return std::make_shared<NvidiaPciePortMetrics>( |
| 218 | conn, mctpRequester, name, pcieDeviceName, path, eid, portType, |
| 219 | upstreamPortNumber, portNumber, objectServer, |
| 220 | nvidiaPciePortL0ToRecoveryCountScalarGroupId, |
| 221 | std::vector<NvidiaMetricInfo>{ |
| 222 | {0, "/pcie/l0_to_recovery_count"}, |
| 223 | }); |
| 224 | } |