added the ecc error sensor to the nvml plugin

dcdbpusher/sensors/nvml/nvml.conf

@@ -92,9 +92,16 @@ group nvml_utilisation {
     }
 }
 
+group nvml_g2 {
 
+    default  def1
 
+    sensor gpu_ecc_errors {
+        mqttsuffix   /ecc_errors
+        feature      GPU_ECC_ERR
+    }
 
+}
 
 
 

dcdbpusher/sensors/nvml/nvmlConfigurator.cpp

@@ -45,6 +45,7 @@ nvmlConfigurator::nvmlConfigurator() {
 	_gpuFeatureMAP["GPU_CLK_MEM"] = GPU_CLK_MEM;
 	_gpuFeatureMAP["GPU_UTL_MEM"] = GPU_UTL_MEM;
 	_gpuFeatureMAP["GPU_UTL_GPU"] = GPU_UTL_GPU;
+	_gpuFeatureMAP["GPU_ECC_ERR"] = GPU_ECC_ERR;
 
 	_groupName = "group";
 	_baseName = "sensor";

dcdbpusher/sensors/nvml/nvmlSensorBase.h

@@ -50,6 +50,7 @@ enum GPU_FEATURE {
 	GPU_CLK_MEM		= 9,
 	GPU_UTL_MEM             = 10,
 	GPU_UTL_GPU             = 11,
+	GPU_ECC_ERR		= 13,
 };
 
 /**
@@ -141,6 +142,9 @@ class nvmlSensorBase : public SensorBase {
 				case GPU_UTL_GPU:
 					feature = "GPU_UTL_GPU";
 					break;
+                                case GPU_ECC_ERR:
+                                        feature = "GPU_ECC_ERR";
+                                        break;
 			}  
 			LOG_VAR(ll) << leading << "    Feature type:  " << feature;    
 		}

dcdbpusher/sensors/nvml/nvmlSensorGroup.cpp

@@ -51,7 +51,7 @@ struct counters_t {
 	unsigned int clockspeed_mem;
 	nvmlMemory_t memory;
 	unsigned int power;
-	//	unsigned long long ecc_counts;
+	unsigned long long ecc_counts;
 	nvmlUtilization_t utilization;
 } counters;
 
@@ -186,6 +186,10 @@ void nvmlSensorGroup::read() {
 					err = nvmlDeviceGetUtilizationRates (env.device, &(counters.utilization));
 					reading.value = counters.utilization.gpu;
 					break;
+				case(GPU_ECC_ERR):
+					err = nvmlDeviceGetTotalEccErrors (env.device, NVML_MEMORY_ERROR_TYPE_CORRECTED,NVML_VOLATILE_ECC,&(counters.ecc_counts));
+					reading.value = counters.ecc_counts;
+					break;
 
 			}
 			s->storeReading(reading);