Kubelet Deivce Manager源码分析

2019独角兽企业重金招聘Python工程师标准>>>

Author: [email protected]

Create Device Manager Instance

Device Manager在何时创建

Device Manager和Volume Manager、QoS Container Manager等一样，都属于kubelet管理的众多Manager之一。Device Manager在kubelet启动时的NewContainerManager中创建。

pkg/kubelet/cm/container_manager_linux.go:197

func NewContainerManager(mountUtil mount.Interface, cadvisorInterface cadvisor.Interface, nodeConfig NodeConfig, failSwapOn bool, devicePluginEnabled bool, recorder record.EventRecorder) (ContainerManager, error) {
	
	...

	glog.Infof("Creating device plugin manager: %t", devicePluginEnabled)
	if devicePluginEnabled {
		cm.deviceManager, err = devicemanager.NewManagerImpl()
	} else {
		cm.deviceManager, err = devicemanager.NewManagerStub()
	}
	if err != nil {
		return nil, err
	}
	...
}	

ManagerImpl结构体

我们有必要先了解Device Manager的结构体：

// ManagerImpl is the structure in charge of managing Device Plugins.
type ManagerImpl struct {
	socketname string
	socketdir  string

	endpoints map[string]endpoint // Key is ResourceName
	mutex     sync.Mutex

	server *grpc.Server

	// activePods is a method for listing active pods on the node
	// so the amount of pluginResources requested by existing pods
	// could be counted when updating allocated devices
	activePods ActivePodsFunc

	// sourcesReady provides the readiness of kubelet configuration sources such as apiserver update readiness.
	// We use it to determine when we can purge inactive pods from checkpointed state.
	sourcesReady config.SourcesReady

	// callback is used for updating devices' states in one time call.
	// e.g. a new device is advertised, two old devices are deleted and a running device fails.
	callback monitorCallback

	// healthyDevices contains all of the registered healthy resourceNames and their exported device IDs.
	healthyDevices map[string]sets.String

	// unhealthyDevices contains all of the unhealthy devices and their exported device IDs.
	unhealthyDevices map[string]sets.String

	// allocatedDevices contains allocated deviceIds, keyed by resourceName.
	allocatedDevices map[string]sets.String

	// podDevices contains pod to allocated device mapping.
	podDevices podDevices
	store      utilstore.Store
	pluginOpts map[string]*pluginapi.DevicePluginOptions
}

下面是核心field的说明：

• socketname: 就是kubelet对外暴露的socket名，即 kubelet.sock。

• socketdir: device plugins' socket的存放的目录，/var/lib/kubelet/device-plugins/。

• endpoints: map对象，key为Resource Name，value为endpoint接口(包括run，stop，allocate，preStartContainer，getDevices，callback，isStoped，StopGracePeriodExpired)，每个endpoint接口对应一个已注册的device plugin，负责与device plugin的gRPC通信及缓存device plugin反馈的device states。

• server: Register服务暴露的gRPC Server。

• activePods: 用来获取该节点上所有active pods，即non-Terminated状态的Pods。在kubelet的initializeRuntimeDependentModules时会注册activePods Func为如下函数：

  	// GetActivePods returns non-terminal pods
  	func (kl *Kubelet) GetActivePods() []*v1.Pod {
  		allPods := kl.podManager.GetPods()
  		activePods := kl.filterOutTerminatedPods(allPods)
  		return activePods
  	}
  

• callback: 是kubelet收到device plugin的ListAndWatch gRCP stream中有devices state变更时的回调函数，包括有新设备增加、旧设备删除、设备状态变化，所以通过ListAndWatch接口的回调方式，可以实现设备的自动发现和热插拔。

  	type monitorCallback func(resourceName string, added, updated, deleted []pluginapi.Device)
  	```	
  
  

• healthyDevices: map对象，key为Resource Name，value为对应的健康的device IDs。

• unhealthyDevices: map对象，key为Resource Name，value为对应的不健康的device IDs。

• allocatedDevices: map对象，key为Resource Name，value为已经分配出去的device IDs。

• podDevices: 记录每个pod中每个容器的device分配情况。

  	// ContainerAllocateResponse为容器内某个device对应的分配信息，包括注入的环境变量、挂载信息、Annotations。
  	type ContainerAllocateResponse struct {
  		Envs map[string]string 
  		Mounts []*Mount 
  		Devices []*DeviceSpec 
  		Annotations map[string]string 
  	}
  
  	// deviceAllocateInfo
  	type deviceAllocateInfo struct {
  		deviceIds sets.String
  		allocResp *pluginapi.ContainerAllocateResponse
  	}
  
  	type resourceAllocateInfo map[string]deviceAllocateInfo // Keyed by resourceName.
  	type containerDevices map[string]resourceAllocateInfo   // Keyed by containerName.
  	type podDevices map[string]containerDevices             // Keyed by podUID.
  

• store: 是对checkpointData的文件存储(/var/lib/kubelet/device-plugins/kubelet_internal_checkpoint)，具体存储了每个Pod分配的Devices信息PodDeviceEntries, 以及已经注册的Resource Name及对应的Devices IDs。

  	type checkpointData struct {
  		PodDeviceEntries  []podDevicesCheckpointEntry
  		RegisteredDevices map[string][]string // key为Resource Name，value为DeviceIDs
  	}
  
  	type podDevicesCheckpointEntry struct {
  		PodUID        string
  		ContainerName string
  		ResourceName  string
  		DeviceIDs     []string
  		AllocResp     []byte
  	}
  

  

• pluginOpts: map对象，key为Resource Name，value为DevicePluginOptions，目前只有一项内容，就是PreStartRequired bool，表示是否在容器启动前要调用device plugin的PreStartContiner接口。在nvidia-k8s-plugin中，PreStartContainer为空实现。

NewManagerImpl

我们再来看看Device Manager的具体创建实现NewManagerImpl。

pkg/kubelet/cm/devicemanager/manager.go:97

// NewManagerImpl creates a new manager.
func NewManagerImpl() (*ManagerImpl, error) {

	// 通过/var/lib/kubelet/device-plugins/kubelet.sock与device plugin交互
	return newManagerImpl(pluginapi.KubeletSocket)
}

func newManagerImpl(socketPath string) (*ManagerImpl, error) {
	glog.V(2).Infof("Creating Device Plugin manager at %s", socketPath)

	if socketPath == "" || !filepath.IsAbs(socketPath) {
		return nil, fmt.Errorf(errBadSocket+" %v", socketPath)
	}

	dir, file := filepath.Split(socketPath)
	manager := &ManagerImpl{
		endpoints:        make(map[string]endpoint),
		socketname:       file,
		socketdir:        dir,
		healthyDevices:   make(map[string]sets.String),
		unhealthyDevices: make(map[string]sets.String),
		allocatedDevices: make(map[string]sets.String),
		pluginOpts:       make(map[string]*pluginapi.DevicePluginOptions),
		podDevices:       make(podDevices),
	}
	manager.callback = manager.genericDeviceUpdateCallback

	// The following structs are populated with real implementations in manager.Start()
	// Before that, initializes them to perform no-op operations.
	manager.activePods = func() []*v1.Pod { return []*v1.Pod{} }
	manager.sourcesReady = &sourcesReadyStub{}
	var err error
	
	// 在/var/lib/kubelet/device-plugins/目录下创建file store类型的key-value存储文件kubelet_internal_checkpoint，用来作为kubelet的device plugin的checkpoint。
	manager.store, err = utilstore.NewFileStore(dir, utilfs.DefaultFs{})
	if err != nil {
		return nil, fmt.Errorf("failed to initialize device plugin checkpointing store: %+v", err)
	}

	return manager, nil
}

• kubelet Device Manager通过/var/lib/kubelet/device-plugins/kubelet.sock与device plugin交互。
• 注册callback为genericDeviceUpdateCallback，用来处理对应devices的add,delete,update事件。
• 在/var/lib/kubelet/device-plugins/目录下创建file store类型的key-value存储文件kubelet_internal_checkpoint，用来作为kubelet的device plugin的checkpoint。
  • 当监听到devices add/delete/update事件发生时，会更新到kubelet_internal_checkpoint文件中。
  • 当device plugin的stop time超过grace period time(代码写死为5min，不可配置),会从checkpoint中删除对应的devices。在这个时间范围内，Device Manager会继续缓存该endpoint及对应的devices。
  • 为Container Allocate Devices后，也会将PodDevices更新到checkpoint中。

我们来看看callback的实现genericDeviceUpdateCallback的实现，了解Device Manager是如何处理devices的add/delete/update消息的。

pkg/kubelet/cm/devicemanager/manager.go:134

func (m *ManagerImpl) genericDeviceUpdateCallback(resourceName string, added, updated, deleted []pluginapi.Device) {
	kept := append(updated, added...)
	m.mutex.Lock()
	if _, ok := m.healthyDevices[resourceName]; !ok {
		m.healthyDevices[resourceName] = sets.NewString()
	}
	if _, ok := m.unhealthyDevices[resourceName]; !ok {
		m.unhealthyDevices[resourceName] = sets.NewString()
	}
	for _, dev := range kept {
		if dev.Health == pluginapi.Healthy {
			m.healthyDevices[resourceName].Insert(dev.ID)
			m.unhealthyDevices[resourceName].Delete(dev.ID)
		} else {
			m.unhealthyDevices[resourceName].Insert(dev.ID)
			m.healthyDevices[resourceName].Delete(dev.ID)
		}
	}
	for _, dev := range deleted {
		m.healthyDevices[resourceName].Delete(dev.ID)
		m.unhealthyDevices[resourceName].Delete(dev.ID)
	}
	m.mutex.Unlock()
	m.writeCheckpoint()
}

• 将callback中收到的devices状态是Healthy，那么将device ID插入到ManagerImpl中healthDevices中，并从unhealthyDevices中删除。
• 将callback中收到的devices状态是Unhealthy，那么将device ID插入到ManagerImpl中unhealthDevices中，并从healthyDevices中删除。
• 将device plugin反馈的需要delete的devices从healthDevices和unhealthDevices中一并删除。
• 将ManagerImpl中的数据更新到checkpoint文件中。

Device Manager的启动

前面把Device Manager的创建流程分析了一下，还涉及到checkpoint和callback的分析。接下来，我们继续对Device Manager的Start流程进行分析。

Start Device Manager

Device Manager是在containerManagerImpl的Start时启动的。

pkg/kubelet/cm/container_manager_linux.go:527

func (cm *containerManagerImpl) Start(node *v1.Node,
	activePods ActivePodsFunc,
	sourcesReady config.SourcesReady,
	podStatusProvider status.PodStatusProvider,
	runtimeService internalapi.RuntimeService) error {

	...
	
	// Starts device manager.
	if err := cm.deviceManager.Start(devicemanager.ActivePodsFunc(activePods), sourcesReady); err != nil {
		return err
	}

	return nil
}

• deviceManager.Start的第一个参数是获取该节点的active(non-terminated)Pods的函数。
• SourcesReady是用来跟踪kubelet配置的Pod Sources，这些Sources包括：
  • file: 通过static file创建静态Pods。
  • http: 通过http接口来获取Pods信息。
  • api: 从Kubernetes API Server获取Pods信息，是Kubernetes默认的内部机制。
  • *: 表示包含以上全部的Sources类型。

ManagerIml Start

ManagerIml.Start负责启动Device Manager，对外提供gRPC服务。

pkg/kubelet/cm/devicemanager/manager.go:204

// Start starts the Device Plugin Manager amd start initialization of
// podDevices and allocatedDevices information from checkpoint-ed state and
// starts device plugin registration service.
func (m *ManagerImpl) Start(activePods ActivePodsFunc, sourcesReady config.SourcesReady) error {

	m.activePods = activePods
	m.sourcesReady = sourcesReady

	// Loads in allocatedDevices information from disk.
	err := m.readCheckpoint()
	...

	socketPath := filepath.Join(m.socketdir, m.socketname)
	os.MkdirAll(m.socketdir, 0755)

	// Removes all stale sockets in m.socketdir. Device plugins can monitor
	// this and use it as a signal to re-register with the new Kubelet.
	if err := m.removeContents(m.socketdir); err != nil {
		glog.Errorf("Fail to clean up stale contents under %s: %+v", m.socketdir, err)
	}

	s, err := net.Listen("unix", socketPath)
	if err != nil {
		glog.Errorf(errListenSocket+" %+v", err)
		return err
	}

	m.server = grpc.NewServer([]grpc.ServerOption{}...)

	pluginapi.RegisterRegistrationServer(m.server, m)
	go m.server.Serve(s)

	glog.V(2).Infof("Serving device plugin registration server on %q", socketPath)

	return nil
}

• 首先读取checkpoint file中数据，恢复ManagerImpl的相关数据，包括：
  • podDevices；
  • allocatedDevices；
  • healthyDevices；
  • unhealthyDevices；
  • endpoints，注意这里会将endpoint的stop time设置为当前时间，意味着kubelet restart后，需要等待device plugin进行re-register后，才认为这些resource是可用的。
• 然后将/var/lib/kubelet/device-plugins/下面的所有文件清空，当然checkpiont文件除外，也就是清空所有的socket文件，包括自己的kubelet.sock,以及其他所有之前的device plugin的socket文件。device plugin会监控kubelet.sock文件是否被删除，如果删除，则会触发自己的向kubelet重新注册自己。
• 创建kubelet.sock并启动gRPC Server对外提供gRPC服务，目前只注册了Register服务，用于Device plugin调用进行插件注册。

Register服务

我们就来看看kubelet Device Manager对外提供的唯一gRPC接口Register。

Register

pkg/kubelet/cm/devicemanager/manager.go:289

// Register registers a device plugin.
func (m *ManagerImpl) Register(ctx context.Context, r *pluginapi.RegisterRequest) (*pluginapi.Empty, error) {
	glog.Infof("Got registration request from device plugin with resource name %q", r.ResourceName)
	metrics.DevicePluginRegistrationCount.WithLabelValues(r.ResourceName).Inc()
	var versionCompatible bool
	for _, v := range pluginapi.SupportedVersions {
		if r.Version == v {
			versionCompatible = true
			break
		}
	}
	if !versionCompatible {
		errorString := fmt.Sprintf(errUnsupportedVersion, r.Version, pluginapi.SupportedVersions)
		glog.Infof("Bad registration request from device plugin with resource name %q: %v", r.ResourceName, errorString)
		return &pluginapi.Empty{}, fmt.Errorf(errorString)
	}

	if !v1helper.IsExtendedResourceName(v1.ResourceName(r.ResourceName)) {
		errorString := fmt.Sprintf(errInvalidResourceName, r.ResourceName)
		glog.Infof("Bad registration request from device plugin: %v", errorString)
		return &pluginapi.Empty{}, fmt.Errorf(errorString)
	}

	// TODO: for now, always accepts newest device plugin. Later may consider to
	// add some policies here, e.g., verify whether an old device plugin with the
	// same resource name is still alive to determine whether we want to accept
	// the new registration.
	go m.addEndpoint(r)

	return &pluginapi.Empty{}, nil
}

• 注册请求是device plugin向kubelet发送的，注册请求RegisterRequest为：

  	type RegisterRequest struct {
  		Version string  // Kubernetes 1.10对应的device plugin api version为v1beta1
  		Endpoint string // device plugin对应的socket name
  		ResourceName string 
  		Options *DevicePluginOptions 
  	}
  

• 这里会检查注册的Resource Name是否符合Extended Resource的规则：

  • Resource Name不能属于kubernetes.io，得有自己的domain，比如nvidia.com。
  • Resource Name中不能包含requests.前缀。
  • 对应的Resource value只能是整数值。

• 调用addEndpoint进行插件注册。

addEndpoint进行device plugin注册

从上面Register的方法中可见，真正插件注册的逻辑是在addEndpoint中实现的。

pkg/kubelet/cm/devicemanager/manager.go:332

func (m *ManagerImpl) addEndpoint(r *pluginapi.RegisterRequest) {
	existingDevs := make(map[string]pluginapi.Device)
	m.mutex.Lock()
	old, ok := m.endpoints[r.ResourceName]
	if ok && old != nil {
		// Pass devices of previous endpoint into re-registered one,
		// to avoid potential orphaned devices upon re-registration
		devices := make(map[string]pluginapi.Device)
		for _, device := range old.getDevices() {
			devices[device.ID] = device
		}
		existingDevs = devices
	}
	m.mutex.Unlock()

	socketPath := filepath.Join(m.socketdir, r.Endpoint)
	e, err := newEndpointImpl(socketPath, r.ResourceName, existingDevs, m.callback)
	if err != nil {
		glog.Errorf("Failed to dial device plugin with request %v: %v", r, err)
		return
	}
	m.mutex.Lock()
	if r.Options != nil {
		m.pluginOpts[r.ResourceName] = r.Options
	}
	// Check for potential re-registration during the initialization of new endpoint,
	// and skip updating if re-registration happens.
	// TODO: simplify the part once we have a better way to handle registered devices
	ext := m.endpoints[r.ResourceName]
	if ext != old {
		glog.Warningf("Some other endpoint %v is added while endpoint %v is initialized", ext, e)
		m.mutex.Unlock()
		e.stop()
		return
	}
	// Associates the newly created endpoint with the corresponding resource name.
	// Stops existing endpoint if there is any.
	m.endpoints[r.ResourceName] = e
	glog.V(2).Infof("Registered endpoint %v", e)
	m.mutex.Unlock()

	if old != nil {
		old.stop()
	}

	go func() {
		e.run()
		e.stop()
		m.mutex.Lock()
		if old, ok := m.endpoints[r.ResourceName]; ok && old == e {
			m.markResourceUnhealthy(r.ResourceName)
		}
		glog.V(2).Infof("Unregistered endpoint %v", e)
		m.mutex.Unlock()
	}()
}

• 首先检查注册的这个device plugin是否已经注册过，如果注册过，则获取已经缓存的devices。

• 再检查device plugin的socket是否能dial成功，如果dial失败，则说明device plugin没正常启动。如果dial成功，就根据已经缓存的devices重新初始化Endpoint，EndpointImpl的定义如下：

  	type endpointImpl struct {
  		client     pluginapi.DevicePluginClient
  		clientConn *grpc.ClientConn
  
  		socketPath   string
  		resourceName string
  		stopTime     time.Time
  
  		devices map[string]pluginapi.Device
  		mutex   sync.Mutex
  
  		cb monitorCallback
  	}
  

• 为了防止在EndpointImpl重新初始化的过程中device plugin进行re-register，初始化完成后再次获取缓存中该device plugin的Endpoint，并与初始化之前的Endpoint对象进行比对:

  • 如果不是同一个对象，则说明在初始化过程中发生了re-register，那么就invoke Endpoint的stop接口，关闭gRPC连接，并设置Endpoint的stopTime为当前时间，Register流程以失败结束。
  • 否则继续后面流程。

• 如果该device plugin之前注册过，那么再重新调用Endpoint的run()启动之前，先调用Endpoint的stop关闭gRPC连接，并设置Endpoint的stopTime为当前时间。

• 然后启动golang协程执行Endpoint的run()，在run方法中:

  • 调用device plugin的ListAndWatch gRPC接口，通过长连接持续获取ListAndWatch gRPC stream，
  • 从stream流中获取的devices与Endpoint中缓存的devices进行比对，得到需要add/delete/update的devices，
  • 然后调用Endpoint的callback（也就是ManagerImpl注册的callback方法genericDeviceUpdateCallback）进行Device Manager的缓存更新并写到checkpoint文件中。

• 直到与device plugin的gRPC连接发生errListAndWatch错误，跳出持续获取stream的死循环，然后调用Endpoint的stop关闭gRPC连接，并设置Endpoint的stopTime为当前时间。

• invoke stop后，再标记该device plugin对应的所有devices为unhealthy，即设置healthyDevices为空， 所有原来healthy的devices都加到unhealthyDevices中，此时表示注册失败。

调用Device Plugin的Allocate接口

注册UpdatePluginResources为Pod Admit Handler

kubelet在NewMainKubelet中会注册一系列的Pod Admit Handler，当有Pod需要创建的时，都会先调用这些Pod Admit Handler进行处理，其中klet.containerManager.UpdatePluginResources就是kubelet Device Manager为Pod分配devices的。

pkg/kubelet/kubelet.go:893

func NewMainKubelet( ... ) (*Kubelet, error) {
	...
	
	klet.admitHandlers.AddPodAdmitHandler(lifecycle.NewPredicateAdmitHandler(klet.getNodeAnyWay, criticalPodAdmissionHandler, klet.containerManager.UpdatePluginResources))
	
	...
}
	
pkg/kubelet/cm/container_manager_linux.go:618

func (cm *containerManagerImpl) UpdatePluginResources(node *schedulercache.NodeInfo, attrs *lifecycle.PodAdmitAttributes) error {
	return cm.deviceManager.Allocate(node, attrs)
}

Allocate

kubelet在创建Pod前，会invoke Device Manager的Allocate方法，为Pod中的每个Container请求分配对应的devices，kubelet会将请求转发到对应的Endpoint的Allocate方法， 然后请求会到对应的device plugin进行处理。

pkg/kubelet/cm/devicemanager/manager.go:259

func (m *ManagerImpl) Allocate(node *schedulercache.NodeInfo, attrs *lifecycle.PodAdmitAttributes) error {
	pod := attrs.Pod
	devicesToReuse := make(map[string]sets.String)
	// TODO: Reuse devices between init containers and regular containers.
	for _, container := range pod.Spec.InitContainers {
		if err := m.allocateContainerResources(pod, &container, devicesToReuse); err != nil {
			return err
		}
		m.podDevices.addContainerAllocatedResources(string(pod.UID), container.Name, devicesToReuse)
	}
	for _, container := range pod.Spec.Containers {
		if err := m.allocateContainerResources(pod, &container, devicesToReuse); err != nil {
			return err
		}
		m.podDevices.removeContainerAllocatedResources(string(pod.UID), container.Name, devicesToReuse)
	}

	m.mutex.Lock()
	defer m.mutex.Unlock()

	// quick return if no pluginResources requested
	if _, podRequireDevicePluginResource := m.podDevices[string(pod.UID)]; !podRequireDevicePluginResource {
		return nil
	}

	m.sanitizeNodeAllocatable(node)
	return nil
}

• 调用allocateContainerResources为Pod中的init container分配devices，并更新ManagerImpl中PodDevices缓存；
• 调用allocateContainerResources为Pod中的regular container分配devices，并更新ManagerImpl中PodDevices缓存；
• 调用sanitizeNodeAllocatable更新scheduler cache中Node对应Resource Name的Allocatable Resource；

allocateContainerResources

pkg/kubelet/cm/devicemanager/manager.go:608

func (m *ManagerImpl) allocateContainerResources(pod *v1.Pod, container *v1.Container, devicesToReuse map[string]sets.String) error {
	podUID := string(pod.UID)
	contName := container.Name
	allocatedDevicesUpdated := false
	// Extended resources are not allowed to be overcommitted.
	// Since device plugin advertises extended resources,
	// therefore Requests must be equal to Limits and iterating
	// over the Limits should be sufficient.
	for k, v := range container.Resources.Limits {
		resource := string(k)
		needed := int(v.Value())
		glog.V(3).Infof("needs %d %s", needed, resource)
		if !m.isDevicePluginResource(resource) {
			continue
		}
		// Updates allocatedDevices to garbage collect any stranded resources
		// before doing the device plugin allocation.
		if !allocatedDevicesUpdated {
			m.updateAllocatedDevices(m.activePods())
			allocatedDevicesUpdated = true
		}
		allocDevices, err := m.devicesToAllocate(podUID, contName, resource, needed, devicesToReuse[resource])
		if err != nil {
			return err
		}
		if allocDevices == nil || len(allocDevices) <= 0 {
			continue
		}

		startRPCTime := time.Now()
		
		m.mutex.Lock()
		e, ok := m.endpoints[resource]
		m.mutex.Unlock()
		if !ok {
			m.mutex.Lock()
			m.allocatedDevices = m.podDevices.devices()
			m.mutex.Unlock()
			return fmt.Errorf("Unknown Device Plugin %s", resource)
		}

		devs := allocDevices.UnsortedList()
		
		glog.V(3).Infof("Making allocation request for devices %v for device plugin %s", devs, resource)
		resp, err := e.allocate(devs)
		metrics.DevicePluginAllocationLatency.WithLabelValues(resource).Observe(metrics.SinceInMicroseconds(startRPCTime))
		if err != nil {
			m.mutex.Lock()
			m.allocatedDevices = m.podDevices.devices()
			m.mutex.Unlock()
			return err
		}

		// Update internal cached podDevices state.
		m.mutex.Lock()
		m.podDevices.insert(podUID, contName, resource, allocDevices, resp.ContainerResponses[0])
		m.mutex.Unlock()
	}

	// Checkpoints device to container allocation information.
	return m.writeCheckpoint()
}

• device plugin提供的Resource属于Kubernetes Extended Resources，所以其Resource QoS只能是Guaranted。
• 每次在为Pod分配devices之前，都去检查一下此时的active pods，并与podDevices缓存中的pods进行比对，将已经terminated的Pods的devices从podDevices中删除，即进行了devices的GC操作。
• 从healthyDevices中随机分配对应数量的devices给该Pod，并注意更新allocatedDevices，否则会导致一个device被分配给多个Pod。
• 拿到devices后，就调用Endpoint的Allocate方法（进而调用对应device plugin的Allocate gRPC Service），device plugin返回ContainerAllocateResponse(包括注入的环境变量、挂载信息、Annotations)。
• 更新podDevices缓存信息，并将ManagerImpl中缓存数据更新到checkpoint文件中。

思考：当init container结束后，对应分配的devices会被释放吗？ 目前还不会释放devices，在Allocate前只会回收Terminated Pods的devices，并没有回收init container的devices。要优化这个也是比较简单的，只要修改上面代码中updateAllocatedDevices方法内的逻辑就行了，增加init container的devices回收逻辑。
所以当前版本最好不会要在init container中使用devices，虽然这种场景几乎不存在。

维护NodeStatus中Device Plugin管理的Resource Capacity

当kubelet更新node status时会调用GetCapacity更新device plugins对应的Resource信息。

pkg/kubelet/kubelet_node_status.go:599

func (kl *Kubelet) setNodeStatusMachineInfo(node *v1.Node) {
	...
	devicePluginCapacity, devicePluginAllocatable, removedDevicePlugins = kl.containerManager.GetDevicePluginResourceCapacity()
	...
}	


pkg/kubelet/cm/container_manager_linux.go:881

func (cm *containerManagerImpl) GetDevicePluginResourceCapacity() (v1.ResourceList, v1.ResourceList, []string) {
	return cm.deviceManager.GetCapacity()
}

下面是GetCapacity的具体代码实现，逻辑很简单：

• 检测healthyDevices对应的device plugin是否已经从缓存中删除或者已经停止超过5min，如果满足以上条件之一，则从endpoints和healthyDevices缓存中删除这些devices。
• 检测unhealthyDevices对应的device plugin是否已经从缓存中删除或者已经停止超过5min，如果满足以上条件之一，则从endpoints和unhealthyDevices缓存中删除这些devices。
• 如果缓存发生变化，则更新到checkpoint文件中。

pkg/kubelet/cm/devicemanager/manager.go:414

func (m *ManagerImpl) GetCapacity() (v1.ResourceList, v1.ResourceList, []string) {
	needsUpdateCheckpoint := false
	var capacity = v1.ResourceList{}
	var allocatable = v1.ResourceList{}
	deletedResources := sets.NewString()
	m.mutex.Lock()
	for resourceName, devices := range m.healthyDevices {
		e, ok := m.endpoints[resourceName]
		if (ok && e.stopGracePeriodExpired()) || !ok {
		
			if !ok {
				glog.Errorf("unexpected: healthyDevices and endpoints are out of sync")
			}
			delete(m.endpoints, resourceName)
			delete(m.healthyDevices, resourceName)
			deletedResources.Insert(resourceName)
			needsUpdateCheckpoint = true
		} else {
			capacity[v1.ResourceName(resourceName)] = *resource.NewQuantity(int64(devices.Len()), resource.DecimalSI)
			allocatable[v1.ResourceName(resourceName)] = *resource.NewQuantity(int64(devices.Len()), resource.DecimalSI)
		}
	}
	for resourceName, devices := range m.unhealthyDevices {
		e, ok := m.endpoints[resourceName]
		if (ok && e.stopGracePeriodExpired()) || !ok {
			if !ok {
				glog.Errorf("unexpected: unhealthyDevices and endpoints are out of sync")
			}
			delete(m.endpoints, resourceName)
			delete(m.unhealthyDevices, resourceName)
			deletedResources.Insert(resourceName)
			needsUpdateCheckpoint = true
		} else {
			capacityCount := capacity[v1.ResourceName(resourceName)]
			unhealthyCount := *resource.NewQuantity(int64(devices.Len()), resource.DecimalSI)
			capacityCount.Add(unhealthyCount)
			capacity[v1.ResourceName(resourceName)] = capacityCount
		}
	}
	m.mutex.Unlock()
	if needsUpdateCheckpoint {
		m.writeCheckpoint()
	}
	return capacity, allocatable, deletedResources.UnsortedList()
}

GetCapacity更新NodeStatus如下数据：

• registered device plugin resource Capacity
• registered device plugin resource Allocatable
• previously registered resources that are no longer active

调用Device Plugin的PreStartContainer接口

在kubelet的GetResource中，会调用DeviceManager的GetDeviceRunContainerOptions，并将这些options添加到kubecontainer.RunContainerOptions中。RunContainerOptions包括Envs、Mounts、Devices、PortMappings、Annotations等信息。

pkg/kubelet/cm/container_manager_linux.go:601

// TODO: move the GetResources logic to PodContainerManager.
func (cm *containerManagerImpl) GetResources(pod *v1.Pod, container *v1.Container) (*kubecontainer.RunContainerOptions, error) {
	opts := &kubecontainer.RunContainerOptions{}
	// Allocate should already be called during predicateAdmitHandler.Admit(),
	// just try to fetch device runtime information from cached state here
	devOpts, err := cm.deviceManager.GetDeviceRunContainerOptions(pod, container)
	if err != nil {
		return nil, err
	} else if devOpts == nil {
		return opts, nil
	}
	opts.Devices = append(opts.Devices, devOpts.Devices...)
	opts.Mounts = append(opts.Mounts, devOpts.Mounts...)
	opts.Envs = append(opts.Envs, devOpts.Envs...)
	opts.Annotations = append(opts.Annotations, devOpts.Annotations...)
	return opts, nil
}

• Device Manager的GetDeviceRunContainerOptions会根据pluginOpts的PreStartRequired是否为true，决定是否调用device plugin的PreStartContainer gRPC Service。

注意：如果某个device plugin的PreStartRequired为true，那么需要注册kubelet Device Manager调用device plugin的PreStartContainer接口的超时时间是30s，即30s内必须完成PreStartContainer的逻辑并返回。

pkg/kubelet/cm/devicemanager/manager.go:688

// GetDeviceRunContainerOptions checks whether we have cached containerDevices
// for the passed-in  and returns its DeviceRunContainerOptions
// for the found one. An empty struct is returned in case no cached state is found.
func (m *ManagerImpl) GetDeviceRunContainerOptions(pod *v1.Pod, container *v1.Container) (*DeviceRunContainerOptions, error) {
	podUID := string(pod.UID)
	contName := container.Name
	for k := range container.Resources.Limits {
		resource := string(k)
		if !m.isDevicePluginResource(resource) {
			continue
		}
		err := m.callPreStartContainerIfNeeded(podUID, contName, resource)
		if err != nil {
			return nil, err
		}
	}
	m.mutex.Lock()
	defer m.mutex.Unlock()
	return m.podDevices.deviceRunContainerOptions(string(pod.UID), container.Name), nil
}

• 然后deviceRunContainerOptions负责封装Container的Envs、Mount points、Device files、Annotations。

总结

本文对Kubelet Device Manager的核心代码进行了走读分析，对其整个工作流有了较深的理解。另外，分别对kubelet的Register服务、kubelet调用device plugin的Allocate接口等做了分析，尤其要注意kubelet device plugins的checkpoint机制（/var/lib/kubelet/device-plugins/kubelet_internal_checkpoint）的重要性。

转载于:https://my.oschina.net/jxcdwangtao/blog/1803036