[SWDEV-542718] Correct socket_affinity (#760)

* [SWDEV-542718] Correct socket_affinity

Updated Socket affinity to show bitmask and expanded cpu list.

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

* Update per-device local_cpulist for socket_affinity

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

* Added amdsmi_get_cpu_affinity_from_local_cpulist API.
Updated the wrapper.

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

* Revert "Added amdsmi_get_cpu_affinity_from_local_cpulist API."

This reverts commit 9a2ef934b1787f8aa09d3e4efe02f897b4295215.

* Moved the changes to C API.
In case of SOCKET_SCOPE, use local_cpulist first.
If it is unavailable or not readable, fallback to
numa.

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

* Addressed review comments

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

---------

Signed-off-by: Bindhiya Kanangot Balakrishnan <[email protected]>

Author: bkanango

amdsmi_cli/amdsmi_commands.py

@@ -892,16 +892,24 @@ def static_gpu(self, args, multiple_devices=False, gpu=None, asic=None, bus=None
                     logging.debug("Failed to get cpu affinity for gpu %s | %s", gpu_id, e.get_error_info())
 
                 try:
-                    cpusockets = amdsmi_interface.amdsmi_get_cpu_affinity_with_scope(args.gpu, amdsmi_interface.AmdSmiAffinityScope.SOCKET_SCOPE)
-                    cpusockets = {f'socket_{i}': socket for i, socket in enumerate(set(cpusockets))}
+                    socket_set = amdsmi_interface.amdsmi_get_cpu_affinity_with_scope(args.gpu, amdsmi_interface.AmdSmiAffinityScope.SOCKET_SCOPE)
+                    socket_set = [f"{cpus:016X}" for cpus in socket_set]
+                    socket_set = {f'cpu_list_{i}': f"{cpus}" for i, cpus in enumerate(socket_set)}
+                    socket_bitmask_ranges = self.helpers.get_bitmask_ranges(socket_set)
+                    socket_affinity = {}
+                    for key in socket_set:
+                        socket_affinity[key] = {
+                            "bitmask": socket_set[key],
+                            "cpu_cores_affinity": socket_bitmask_ranges.get(key, "N/A")
+                        }
                 except amdsmi_exception.AmdSmiLibraryException as e:
-                    cpusockets = {}
+                    socket_affinity = "N/A"
                     logging.debug("Failed to get socket affinity for gpu %s | %s", gpu_id, e.get_error_info())
 
                 static_dict['numa'] = { 'node' : numa_node_number,
                                         'affinity' : numa_affinity,
                                         'cpu_affinity' : cpu_affinity,
-                                        'socket_affinity' : cpusockets if cpusockets else "N/A"}
+                                        'socket_affinity' : socket_affinity}
         if args.vram:
             vram_info_dict = {"type" : "N/A",
                               "vendor" : "N/A",

include/amd_smi/impl/amd_smi_gpu_device.h

@@ -72,6 +72,7 @@ class AMDSmiGPUDevice: public AMDSmiProcessor {
 // New methods for -e feature
     std::string bdf_to_string() const;     // -e feature
     std::vector<uint64_t> get_bitmask_from_numa_node(int32_t node_id, uint32_t size) const;
+    std::vector<uint64_t> get_bitmask_from_local_cpulist(uint32_t drm_card, uint32_t size) const;
 
  private:
     uint32_t gpu_id_;

src/amd_smi/amd_smi.cc

@@ -5187,27 +5187,12 @@ amdsmi_status_t amdsmi_get_cpu_affinity_with_scope(amdsmi_processor_handle proce
 
         case AMDSMI_AFFINITY_SCOPE_SOCKET:
         {
-            std::vector<uint32_t> sockets = amd::smi::AMDSmiSystem::getInstance().get_cpu_sockets_from_numa_node(node_id);
-
-            if(sockets[0] == std::numeric_limits<int32_t>::max()){
+            uint32_t drm_card = gpu_device->get_card_id();
+            std::vector<uint64_t> bitmask = gpu_device->get_bitmask_from_local_cpulist(drm_card, cpu_set_size);
+            if(bitmask[0] == std::numeric_limits<int32_t>::max()){
                 return AMDSMI_STATUS_REFCOUNT_OVERFLOW;
             } else {
-            for (uint32_t idx : sockets) {
-                cpu_set[idx] = idx;
-            }
-
-            std::sort(cpu_set, cpu_set + cpu_set_size);
-
-            // Discard duplicates
-            uint32_t temp_size = 0;
-            for (uint32_t i = 0; i < cpu_set_size; ++i) {
-                if (i == 0 || cpu_set[i] != cpu_set[i - 1]) {
-                    cpu_set[temp_size++] = cpu_set[i];
-                }
-            }
-
-            // Update the size to the temp size after discarding duplicates
-            cpu_set_size = temp_size;
+                std::memcpy(cpu_set, bitmask.data(), cpu_set_size * sizeof(uint64_t));
             }
             break;
         }

src/amd_smi/amd_smi_gpu_device.cc

@@ -310,4 +310,38 @@ std::vector<uint64_t> AMDSmiGPUDevice::get_bitmask_from_numa_node(int32_t node_i
     return bitmask;
 }
 
+std::vector<uint64_t> AMDSmiGPUDevice::get_bitmask_from_local_cpulist(uint32_t drm_card, uint32_t size) const {
+    std::vector<uint64_t> bitmask(size, 0);
+
+    if (drm_card < 0) {
+        bitmask[0] = std::numeric_limits<int32_t>::max();
+        return bitmask;
+    }
+
+    std::string path = "/sys/class/drm/card" + std::to_string(drm_card) + "/device/local_cpulist";
+    std::ifstream file(path);
+
+    if (file.is_open()) {
+        std::string info;
+        while (std::getline(file, info)) {
+            std::istringstream sstr(info);
+            std::string node_cpus;
+            while (std::getline(sstr, node_cpus, ',')) {
+                size_t hyphen = node_cpus.find('-');
+                if (hyphen != std::string::npos) {
+                    int start = std::stoi(node_cpus.substr(0, hyphen));
+                    int end = std::stoi(node_cpus.substr(hyphen + 1));
+                    for (int i = start; i <= end; ++i) {
+                        bitmask[i / 64] |= (1ULL << (i % 64));
+                    }
+                } else {
+                    int core = std::stoi(node_cpus);
+                    bitmask[core / 64] |= (1ULL << (core % 64));
+                }
+            }
+        }
+    }
+    return bitmask;
+}
+
 } // namespace amd::smi