feat: add mark health device function

internal/plugin/server.go

@@ -59,7 +59,7 @@ type nvidiaDevicePlugin struct {
 
 	socket string
 	server *grpc.Server
-	health chan *rm.Device
+	health chan *rm.DeviceEvent
 	stop   chan interface{}
 
 	imexChannels imex.Channels
@@ -105,7 +105,7 @@ func getPluginSocketPath(resource spec.ResourceName) string {
 
 func (plugin *nvidiaDevicePlugin) initialize() {
 	plugin.server = grpc.NewServer([]grpc.ServerOption{}...)
-	plugin.health = make(chan *rm.Device)
+	plugin.health = make(chan *rm.DeviceEvent, 10)
 	plugin.stop = make(chan interface{})
 }
 
@@ -272,11 +272,21 @@ func (plugin *nvidiaDevicePlugin) ListAndWatch(e *pluginapi.Empty, s pluginapi.D
 			return nil
 		case d := <-plugin.health:
 			// FIXME: there is no way to recover from the Unhealthy state.
-			d.Health = pluginapi.Unhealthy
-			klog.Infof("'%s' device marked unhealthy: %s", plugin.rm.Resource(), d.ID)
-			if err := s.Send(&pluginapi.ListAndWatchResponse{Devices: plugin.apiDevices()}); err != nil {
-				return nil
+			if d.Event == rm.DeviceUnHalthy {
+				d.Device.Health = pluginapi.Unhealthy
+				klog.Infof("'%s' device marked unhealthy: %s", plugin.rm.Resource(), d.Device.ID)
+				if err := s.Send(&pluginapi.ListAndWatchResponse{Devices: plugin.apiDevices()}); err != nil {
+					return nil
+				}
 			}
+			if d.Event == rm.DeviceHealthy {
+				d.Device.Health = pluginapi.Healthy
+				klog.Infof("'%s' device marked healthy: %s", plugin.rm.Resource(), d.Device.ID)
+				if err := s.Send(&pluginapi.ListAndWatchResponse{Devices: plugin.apiDevices()}); err != nil {
+					return nil
+				}
+			}
+
 		}
 	}
 }

internal/rm/devices.go

@@ -36,6 +36,18 @@ type Device struct {
 	Replicas int
 }
 
+type DeviceEventType int
+
+const (
+	DeviceUnHalthy DeviceEventType = iota
+	DeviceHealthy
+)
+
+type DeviceEvent struct {
+	Device *Device
+	Event  DeviceEventType
+}
+
 // deviceInfo defines the information the required to construct a Device
 type deviceInfo interface {
 	GetUUID() (string, error)

internal/rm/health.go

@@ -33,10 +33,14 @@ const (
 	// this is in addition to the Application errors that are already ignored.
 	envDisableHealthChecks = "DP_DISABLE_HEALTHCHECKS"
 	allHealthChecks        = "xids"
+
+	nvmlEventTypeGpuRecoveryAction = 0x0000000000008000 // from https://docs.nvidia.com/deploy/nvml-api/group__nvmlEventType.html?
+
+	nvmlEventTypeGpuUnavailableError = 0x0000000000004000
 )
 
 // CheckHealth performs health checks on a set of devices, writing to the 'unhealthy' channel with any unhealthy devices
-func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devices, unhealthy chan<- *Device) error {
+func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devices, unhealthy chan<- *DeviceEvent) error {
 	disableHealthChecks := strings.ToLower(os.Getenv(envDisableHealthChecks))
 	if disableHealthChecks == "all" {
 		disableHealthChecks = allHealthChecks
@@ -92,12 +96,15 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 	deviceIDToGiMap := make(map[string]uint32)
 	deviceIDToCiMap := make(map[string]uint32)
 
-	eventMask := uint64(nvml.EventTypeXidCriticalError | nvml.EventTypeDoubleBitEccError | nvml.EventTypeSingleBitEccError)
+	eventMask := uint64(nvml.EventTypeXidCriticalError | nvml.EventTypeDoubleBitEccError | nvml.EventTypeSingleBitEccError | nvmlEventTypeGpuUnavailableError | nvmlEventTypeGpuRecoveryAction)
 	for _, d := range devices {
 		uuid, gi, ci, err := r.getDevicePlacement(d)
 		if err != nil {
 			klog.Warningf("Could not determine device placement for %v: %v; Marking it unhealthy.", d.ID, err)
-			unhealthy <- d
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceUnHalthy,
+			}
 			continue
 		}
 		deviceIDToGiMap[d.ID] = gi
@@ -107,14 +114,20 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 		gpu, ret := r.nvml.DeviceGetHandleByUUID(uuid)
 		if ret != nvml.SUCCESS {
 			klog.Infof("unable to get device handle from UUID: %v; marking it as unhealthy", ret)
-			unhealthy <- d
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceUnHalthy,
+			}
 			continue
 		}
 
 		supportedEvents, ret := gpu.GetSupportedEventTypes()
 		if ret != nvml.SUCCESS {
 			klog.Infof("unable to determine the supported events for %v: %v; marking it as unhealthy", d.ID, ret)
-			unhealthy <- d
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceUnHalthy,
+			}
 			continue
 		}
 
@@ -124,7 +137,10 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 		}
 		if ret != nvml.SUCCESS {
 			klog.Infof("Marking device %v as unhealthy: %v", d.ID, ret)
-			unhealthy <- d
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceUnHalthy,
+			}
 		}
 	}
 
@@ -142,16 +158,14 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 		if ret != nvml.SUCCESS {
 			klog.Infof("Error waiting for event: %v; Marking all devices as unhealthy", ret)
 			for _, d := range devices {
-				unhealthy <- d
+				unhealthy <- &DeviceEvent{
+					Device: d,
+					Event:  DeviceUnHalthy,
+				}
 			}
 			continue
 		}
 
-		if e.EventType != nvml.EventTypeXidCriticalError {
-			klog.Infof("Skipping non-nvmlEventTypeXidCriticalError event: %+v", e)
-			continue
-		}
-
 		if skippedXids[e.EventData] {
 			klog.Infof("Skipping event %+v", e)
 			continue
@@ -163,7 +177,10 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 			// If we cannot reliably determine the device UUID, we mark all devices as unhealthy.
 			klog.Infof("Failed to determine uuid for event %v: %v; Marking all devices as unhealthy.", e, ret)
 			for _, d := range devices {
-				unhealthy <- d
+				unhealthy <- &DeviceEvent{
+					Device: d,
+					Event:  DeviceUnHalthy,
+				}
 			}
 			continue
 		}
@@ -173,6 +190,30 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 			klog.Infof("Ignoring event for unexpected device: %v", eventUUID)
 			continue
 		}
+		// nvmlEventTypeRecovery is a special case, where we mark the device as healthy.
+		if e.EventType == nvmlEventTypeGpuRecoveryAction {
+			klog.Infof("Gpu recovery event: %+v", e)
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceHealthy,
+			}
+			continue
+
+		}
+
+		if e.EventType == nvmlEventTypeGpuUnavailableError {
+			klog.Infof("Gpu unavailable event: %+v", e)
+			unhealthy <- &DeviceEvent{
+				Device: d,
+				Event:  DeviceUnHalthy,
+			}
+			continue
+		}
+
+		if e.EventType != nvml.EventTypeXidCriticalError {
+			klog.Infof("Skipping non-nvmlEventTypeXidCriticalError event: %+v", e)
+			continue
+		}
 
 		if d.IsMigDevice() && e.GpuInstanceId != 0xFFFFFFFF && e.ComputeInstanceId != 0xFFFFFFFF {
 			gi := deviceIDToGiMap[d.ID]
@@ -184,7 +225,10 @@ func (r *nvmlResourceManager) checkHealth(stop <-chan interface{}, devices Devic
 		}
 
 		klog.Infof("XidCriticalError: Xid=%d on Device=%s; marking device as unhealthy.", e.EventData, d.ID)
-		unhealthy <- d
+		unhealthy <- &DeviceEvent{
+			Device: d,
+			Event:  DeviceUnHalthy,
+		}
 	}
 }
 

internal/rm/nvml_manager.go

@@ -91,7 +91,7 @@ func (r *nvmlResourceManager) GetDevicePaths(ids []string) []string {
 }
 
 // CheckHealth performs health checks on a set of devices, writing to the 'unhealthy' channel with any unhealthy devices
-func (r *nvmlResourceManager) CheckHealth(stop <-chan interface{}, unhealthy chan<- *Device) error {
+func (r *nvmlResourceManager) CheckHealth(stop <-chan interface{}, unhealthy chan<- *DeviceEvent) error {
 	return r.checkHealth(stop, r.devices, unhealthy)
 }
 

internal/rm/rm.go

@@ -42,7 +42,7 @@ type ResourceManager interface {
 	Devices() Devices
 	GetDevicePaths([]string) []string
 	GetPreferredAllocation(available, required []string, size int) ([]string, error)
-	CheckHealth(stop <-chan interface{}, unhealthy chan<- *Device) error
+	CheckHealth(stop <-chan interface{}, unhealthy chan<- *DeviceEvent) error
 	ValidateRequest(AnnotatedIDs) error
 }
 

internal/rm/tegra_manager.go

@@ -71,6 +71,6 @@ func (r *tegraResourceManager) GetDevicePaths(ids []string) []string {
 }
 
 // CheckHealth is disabled for the tegraResourceManager
-func (r *tegraResourceManager) CheckHealth(stop <-chan interface{}, unhealthy chan<- *Device) error {
+func (r *tegraResourceManager) CheckHealth(stop <-chan interface{}, unhealthy chan<- *DeviceEvent) error {
 	return nil
 }