Merge amd-staging into amd-master 20231116

This merge skips Ibdaf031be9d916020b4049544dbd725858c7711d as that change introduces a bug in gpu-metrics Change-Id: Ied8447affd5ed3c847734d75517b04c073dc44b4 Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>
2023-11-16 11:17:50 -06:00
@@ -35,7 +35,7 @@ find_program (GIT NAMES git)

 ## Setup the package version based on git tags.
 set(PKG_VERSION_GIT_TAG_PREFIX "rsmi_pkg_ver")
-get_package_version_number("6.0.0" ${PKG_VERSION_GIT_TAG_PREFIX} GIT)
+get_package_version_number("6.1.0" ${PKG_VERSION_GIT_TAG_PREFIX} GIT)
 message("Package version: ${PKG_VERSION_STR}")
 set(${ROCM_SMI_LIBS_TARGET}_VERSION_MAJOR "${VERSION_MAJOR}")
 set(${ROCM_SMI_LIBS_TARGET}_VERSION_MINOR "${VERSION_MINOR}")
@@ -72,7 +72,7 @@ endif()

 ## Compiler flags
 set(CMAKE_CXX_FLAGS
- "${CMAKE_CXX_FLAGS} -Wall -Wextra -fno-rtti -std=c++17")
+ "${CMAKE_CXX_FLAGS} -Wall -Wextra -fno-rtti")
 if (${CMAKE_HOST_SYSTEM_PROCESSOR} STREQUAL "x86_64")
    set(CMAKE_CXX_FLAGS
     "${CMAKE_CXX_FLAGS} -m64 -msse -msse2")
@@ -363,16 +363,16 @@ typedef rsmi_clk_type_t rsmi_clk_type;
 */
 typedef enum {
  RSMI_COMPUTE_PARTITION_INVALID = 0,
-  RSMI_COMPUTE_PARTITION_CPX = 1,    //!< Core mode (CPX)- Per-chip XCC with
-                                     //!< shared memory
-  RSMI_COMPUTE_PARTITION_SPX  = 2,   //!< Single GPU mode (SPX)- All XCCs work
-                                     //!< together with shared memory
-  RSMI_COMPUTE_PARTITION_DPX  = 3,   //!< Dual GPU mode (DPX)- Half XCCs work
-                                     //!< together with shared memory
-  RSMI_COMPUTE_PARTITION_TPX = 4,    //!< Triple GPU mode (TPX)- One-third XCCs
-                                     //!< work together with shared memory
-  RSMI_COMPUTE_PARTITION_QPX = 5,    //!< Quad GPU mode (QPX)- Quarter XCCs
-                                     //!< work together with shared memory
+  RSMI_COMPUTE_PARTITION_CPX, //!< Core mode (CPX)- Per-chip XCC with
+                              //!< shared memory
+  RSMI_COMPUTE_PARTITION_SPX, //!< Single GPU mode (SPX)- All XCCs work
+                              //!< together with shared memory
+  RSMI_COMPUTE_PARTITION_DPX, //!< Dual GPU mode (DPX)- Half XCCs work
+                              //!< together with shared memory
+  RSMI_COMPUTE_PARTITION_TPX, //!< Triple GPU mode (TPX)- One-third XCCs
+                              //!< work together with shared memory
+  RSMI_COMPUTE_PARTITION_QPX, //!< Quad GPU mode (QPX)- Quarter XCCs
+                              //!< work together with shared memory
 } rsmi_compute_partition_type_t;
 /// \cond Ignore in docs.
 typedef rsmi_compute_partition_type_t rsmi_compute_partition_type;
@@ -680,8 +680,8 @@ typedef enum {
 */
 typedef enum _RSMI_IO_LINK_TYPE {
  RSMI_IOLINK_TYPE_UNDEFINED      = 0,          //!< unknown type.
-  RSMI_IOLINK_TYPE_PCIEXPRESS     = 1,          //!< PCI Express
-  RSMI_IOLINK_TYPE_XGMI           = 2,          //!< XGMI
+  RSMI_IOLINK_TYPE_PCIEXPRESS,                  //!< PCI Express
+  RSMI_IOLINK_TYPE_XGMI,                        //!< XGMI
  RSMI_IOLINK_TYPE_NUMIOLINKTYPES,              //!< Number of IO Link types
  RSMI_IOLINK_TYPE_SIZE           = 0xFFFFFFFF  //!< Max of IO Link types
 } RSMI_IO_LINK_TYPE;
@@ -1503,6 +1503,23 @@ rsmi_status_t rsmi_dev_subsystem_vendor_id_get(uint32_t dv_ind, uint16_t *id);
 */
 rsmi_status_t rsmi_dev_unique_id_get(uint32_t dv_ind, uint64_t *id);

+/**
+ *  @brief Get the XGMI physical id associated with the device
+ *
+ *  @details Given a device index @p dv_ind and a pointer to a uint32_t to
+ *  which the XGMI physical id will be written
+ *
+ *  @param[in] dv_ind a device index
+ *
+ *  @param[inout] revision a pointer to uint32_t to which the XGMI physical id
+ *  will be written
+ *
+ *  @retval ::RSMI_STATUS_SUCCESS is returned upon successful call.
+ *
+ */
+rsmi_status_t rsmi_dev_xgmi_physical_id_get(uint32_t dv_ind, uint16_t *id);
+
+
 /** @} */  // end of IDQuer

 /*****************************************************************************/
@@ -103,6 +103,7 @@ enum DevInfoTypes {
  kDevOverDriveLevel,
  kDevMemOverDriveLevel,
  kDevDevID,
+  kDevXGMIPhysicalID,
  kDevDevRevID,
  kDevDevProdName,
  kDevDevProdNum,
@@ -836,6 +836,21 @@ rsmi_dev_id_get(uint32_t dv_ind, uint16_t *id) {
  return ret;
 }

+rsmi_status_t
+rsmi_dev_xgmi_physical_id_get(uint32_t dv_ind, uint16_t *id) {
+  std::ostringstream ss;
+  rsmi_status_t ret;
+  ss << __PRETTY_FUNCTION__ << "| ======= start =======";
+  LOG_TRACE(ss);
+  CHK_SUPPORT_NAME_ONLY(id)
+
+  ret = get_id(dv_ind, amd::smi::kDevXGMIPhysicalID, id);
+  ss << __PRETTY_FUNCTION__ << " | ======= end ======="
+     << ", reporting " << amd::smi::getRSMIStatusString(ret);
+  LOG_TRACE(ss);
+  return ret;
+}
+
 rsmi_status_t
 rsmi_dev_revision_get(uint32_t dv_ind, uint16_t *revision) {
  std::ostringstream outss;
@@ -2776,9 +2791,6 @@ rsmi_dev_od_volt_info_get(uint32_t dv_ind, rsmi_od_volt_freq_data_t *odv) {
  ss << __PRETTY_FUNCTION__ << "| ======= start =======";
  LOG_TRACE(ss);
  DEVICE_MUTEX
-  if (odv == nullptr) {
-    return RSMI_STATUS_INVALID_ARGS;
-  }
  CHK_SUPPORT_NAME_ONLY(odv)
  rsmi_status_t ret = get_od_clk_volt_info(dv_ind, odv);

@@ -82,6 +82,7 @@ static const char *kDevPerfLevelFName = "power_dpm_force_performance_level";
 static const char *kDevDevProdNameFName = "product_name";
 static const char *kDevDevProdNumFName = "product_number";
 static const char *kDevDevIDFName = "device";
+static const char* kDevXGMIPhysicalIDFName = "xgmi_physical_id";
 static const char *kDevDevRevIDFName = "revision";
 static const char *kDevVendorIDFName = "vendor";
 static const char *kDevSubSysDevIDFName = "subsystem_device";
@@ -238,6 +239,7 @@ static const std::map<DevInfoTypes, const char *> kDevAttribNameMap = {
    {kDevDevProdName, kDevDevProdNameFName},
    {kDevDevProdNum, kDevDevProdNumFName},
    {kDevDevID, kDevDevIDFName},
+    {kDevXGMIPhysicalID, kDevXGMIPhysicalIDFName},
    {kDevDevRevID, kDevDevRevIDFName},
    {kDevVendorID, kDevVendorIDFName},
    {kDevSubSysDevID, kDevSubSysDevIDFName},
@@ -379,6 +381,7 @@ static const std::map<const char *, dev_depends_t> kDevFuncDependsMap = {
    // Functions with only mandatory dependencies
  {"rsmi_dev_vram_vendor_get",           {{kDevVramVendorFName}, {}}},
  {"rsmi_dev_id_get",                    {{kDevDevIDFName}, {}}},
+  {"rsmi_dev_xgmi_physical_id_get",      {{kDevXGMIPhysicalIDFName}, {}}},
  {"rsmi_dev_revision_get",              {{kDevDevRevIDFName}, {}}},
  {"rsmi_dev_vendor_id_get",             {{kDevVendorIDFName}, {}}},
  {"rsmi_dev_name_get",                  {{kDevVendorIDFName,
@@ -956,6 +959,7 @@ int Device::readDevInfo(DevInfoTypes type, uint64_t *val) {
    case kDevSubSysVendorID:
    case kDevVendorID:
    case kDevErrCntFeatures:
+    case kDevXGMIPhysicalID:
      ret = readDevInfoStr(type, &tempStr);
      RET_IF_NONZERO(ret);

@@ -1102,6 +1106,7 @@ int Device::readDevInfo(DevInfoTypes type, std::string *val) {
    case kDevComputePartition:
    case kDevMemoryPartition:
    case kDevNumaNode:
+    case kDevXGMIPhysicalID:
      return readDevInfoStr(type, val);
      break;

@@ -53,6 +53,7 @@
 #include <functional>
 #include <iostream>
 #include <memory>
+#include <algorithm>
 #include <set>
 #include <sstream>
 #include <string>
@@ -85,6 +86,7 @@ amd::smi::RocmSMI::devInfoTypesStrings = {
  {amd::smi::kDevOverDriveLevel, amdSMI + "kDevOverDriveLevel"},
  {amd::smi::kDevMemOverDriveLevel, amdSMI + "kDevMemOverDriveLevel"},
  {amd::smi::kDevDevID, amdSMI + "kDevDevID"},
+  {amd::smi::kDevXGMIPhysicalID, amdSMI + "kDevXGMIPhysicalID"},
  {amd::smi::kDevDevRevID, amdSMI + "kDevDevRevID"},
  {amd::smi::kDevDevProdName, amdSMI + "kDevDevProdName"},
  {amd::smi::kDevDevProdNum, amdSMI + "kDevDevProdNum"},
@@ -383,9 +385,28 @@ RocmSMI::Initialize(uint64_t flags) {
         << "\n | final update: device->bdfid() holds correct device bdf";
      LOG_TRACE(ss);
  }
-  if (ret != 0) {
-    throw amd::smi::rsmi_exception(RSMI_INITIALIZATION_ERROR,
-            "Failed to initialize rocm_smi library (amdgpu node discovery).");
+
+  std::shared_ptr<amd::smi::Device> dev;
+  // Sort index based on the BDF, collect BDF id firstly.
+  std::vector<std::pair<uint64_t, std::shared_ptr<amd::smi::Device>>> dv_to_id;
+  dv_to_id.reserve(devices_.size());
+  for (uint32_t dv_ind = 0; dv_ind < devices_.size(); ++dv_ind) {
+      dev = devices_[dv_ind];
+      uint64_t bdfid = dev->bdfid();
+      dv_to_id.push_back({bdfid, dev});
+    }
+  ss << __PRETTY_FUNCTION__ << " Sort index based on BDF.";
+  LOG_DEBUG(ss);
+
+  // Stable sort to keep the order if bdf is equal.
+  std::stable_sort(dv_to_id.begin(), dv_to_id.end(), []
+  (const std::pair<uint64_t, std::shared_ptr<amd::smi::Device>>& p1,
+    const std::pair<uint64_t, std::shared_ptr<amd::smi::Device>>& p2) {
+        return p1.first < p2.first;
+  });
+  devices_.clear();
+  for (uint32_t dv_ind = 0; dv_ind < dv_to_id.size(); ++dv_ind) {
+    devices_.push_back(dv_to_id[dv_ind].second);
  }

  std::map<uint64_t, std::shared_ptr<KFDNode>> tmp_map;
@@ -406,7 +427,6 @@ RocmSMI::Initialize(uint64_t flags) {
  for (it = io_link_map_tmp.begin(); it != io_link_map_tmp.end(); it++)
    io_link_map_[it->first] = it->second;

-  std::shared_ptr<amd::smi::Device> dev;

  // Remove any drm nodes that don't have  a corresponding readable kfd node.
  // kfd nodes will not be added if their properties file is not readable.
@@ -451,6 +471,7 @@ RocmSMI::Initialize(uint64_t flags) {
  if (ROCmLogging::Logger::getInstance()->isLoggerEnabled()) {
    logSystemDetails();
  }
+
  // Leaving below to help debug temp file issues
  // displayAppTmpFilesContent();
  std::string amdGPUDeviceList = displayAllDevicePaths(devices_);