Persisting driver state and reload state on restart.

Upon restarts, driver has to recognize the provisioned volumes on last run. For that driver has to persist its state(provisioned volume information). We have two options: 1. Per node state: Persist the driver state on each node specific to that node. Upon driver start, it has to load the persisted information. Whenever a node registers, Controller/Master has to collect the volume information stored on that node. 2. Cluster level state: Persist the consolidated cluster level driver state, say to a configmap on Kubernetes. This is supposed to done at Controller side, but not clear how to pass node specific information to a node. This change implements the fist option, i.e, per node state. Defined a new interface, StateManager, that handles the state persistency irrespective of either per node or cluster. Also provided an implementation of StateManager that persists the state to a file. Made changes to RegistrySrerver, so that it informs the listeners upon registering of a node controller. The ControllerServer on master, listens on this and collects the volume information on node, by calling csi.ControllerServer.ListVolumes. FIXES intel#25 - Clean handling of driver restart
avalluri · May 17, 2019 · 7ab9cb7 · 7ab9cb7
1 parent 06c0d7a
commit 7ab9cb7
Show file tree

Hide file tree

Showing 18 changed files with 800 additions and 51 deletions.
diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md
@@ -142,7 +142,7 @@ argument name        | meaning                                           | type
 -mode string               | driver run mode                                | string | controller, node |
 -nodeid string             | node id                                        | string |              | nodeid
 -registryEndpoint string   | endpoint to connect/listen registry server     | string |              |
-
+-statePath                 | Directory path where to persist the state of the driver running on a node | string | absolute directory path on node | /var/lib/<drivername>
 
 Environment variables
 ---------------------

diff --git a/deploy/kubernetes-1.13/pmem-csi-direct-testing.yaml b/deploy/kubernetes-1.13/pmem-csi-direct-testing.yaml
@@ -295,6 +295,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         - -v=5
         - -coverprofile=/var/lib/pmem-csi-coverage/pmem-csi-driver-node-*.out
         env:
@@ -331,6 +332,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /sys
           name: sys-dir
         - mountPath: /var/lib/pmem-csi-coverage
@@ -371,6 +374,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.13/pmem-csi-direct.yaml b/deploy/kubernetes-1.13/pmem-csi-direct.yaml
@@ -218,6 +218,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         env:
         - name: CSI_ENDPOINT
           value: unix:///csi/csi.sock
@@ -252,6 +253,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /sys
           name: sys-dir
       - args:
@@ -289,6 +292,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.13/pmem-csi-lvm-testing.yaml b/deploy/kubernetes-1.13/pmem-csi-lvm-testing.yaml
@@ -295,6 +295,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         - -v=5
         - -coverprofile=/var/lib/pmem-csi-coverage/pmem-csi-driver-node-*.out
         env:
@@ -331,6 +332,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /var/lib/pmem-csi-coverage
           name: coverage-dir
       - args:
@@ -409,6 +412,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.13/pmem-csi-lvm.yaml b/deploy/kubernetes-1.13/pmem-csi-lvm.yaml
@@ -218,6 +218,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         env:
         - name: CSI_ENDPOINT
           value: unix:///csi/csi.sock
@@ -252,6 +253,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
       - args:
         - -v=3
         - --kubelet-registration-path=/var/lib/kubelet/plugins/pmem-csi.intel.com/csi.sock
@@ -318,6 +321,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.14/pmem-csi-direct-testing.yaml b/deploy/kubernetes-1.14/pmem-csi-direct-testing.yaml
@@ -315,6 +315,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         - -v=5
         - -coverprofile=/var/lib/pmem-csi-coverage/pmem-csi-driver-node-*.out
         env:
@@ -351,6 +352,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /sys
           name: sys-dir
         - mountPath: /var/lib/pmem-csi-coverage
@@ -391,6 +394,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.14/pmem-csi-direct.yaml b/deploy/kubernetes-1.14/pmem-csi-direct.yaml
@@ -238,6 +238,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         env:
         - name: CSI_ENDPOINT
           value: unix:///csi/csi.sock
@@ -272,6 +273,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /sys
           name: sys-dir
       - args:
@@ -309,6 +312,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.14/pmem-csi-lvm-testing.yaml b/deploy/kubernetes-1.14/pmem-csi-lvm-testing.yaml
@@ -315,6 +315,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         - -v=5
         - -coverprofile=/var/lib/pmem-csi-coverage/pmem-csi-driver-node-*.out
         env:
@@ -351,6 +352,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
         - mountPath: /var/lib/pmem-csi-coverage
           name: coverage-dir
       - args:
@@ -429,6 +432,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kubernetes-1.14/pmem-csi-lvm.yaml b/deploy/kubernetes-1.14/pmem-csi-lvm.yaml
@@ -238,6 +238,7 @@ spec:
         - -caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
         - -certFile=/certs/$(KUBE_NODE_NAME).crt
         - -keyFile=/certs/$(KUBE_NODE_NAME).key
+        - -statePath=/var/lib/pmem-csi.intel.com
         env:
         - name: CSI_ENDPOINT
           value: unix:///csi/csi.sock
@@ -272,6 +273,8 @@ spec:
           name: registry-cert
         - mountPath: /dev
           name: dev-dir
+        - mountPath: /var/lib/pmem-csi.intel.com
+          name: pmem-state-dir
       - args:
         - -v=3
         - --kubelet-registration-path=/var/lib/kubelet/plugins/pmem-csi.intel.com/csi.sock
@@ -338,6 +341,10 @@ spec:
       - name: registry-cert
         secret:
           secretName: pmem-csi-node-secrets
+      - hostPath:
+          path: /var/lib/pmem-csi.intel.com
+          type: DirectoryOrCreate
+        name: pmem-state-dir
       - hostPath:
           path: /dev
           type: DirectoryOrCreate

diff --git a/deploy/kustomize/driver/pmem-csi.yaml b/deploy/kustomize/driver/pmem-csi.yaml
@@ -113,7 +113,8 @@ spec:
                   "-registryEndpoint=$(PMEM_CSI_CONTROLLER_PORT_10000_TCP)",
                   "-caFile=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt",
                   "-certFile=/certs/$(KUBE_NODE_NAME).crt",
-                  "-keyFile=/certs/$(KUBE_NODE_NAME).key"
+                  "-keyFile=/certs/$(KUBE_NODE_NAME).key",
+                  "-statePath=/var/lib/pmem-csi.intel.com"
               ]
         # Passing /dev to container may cause container creation error because
         # termination-log is located on /dev/ by default, re-locate to /tmp
@@ -148,6 +149,8 @@ spec:
           mountPath: /certs/
         - name : dev-dir
           mountPath: /dev
+        - name: pmem-state-dir
+          mountPath: /var/lib/pmem-csi.intel.com
       - name: driver-registrar
         imagePullPolicy: Always
         image: quay.io/k8scsi/csi-node-driver-registrar:v1.X.Y
@@ -179,6 +182,10 @@ spec:
         - name: registry-cert
           secret:
             secretName: pmem-csi-node-secrets
+        - name: pmem-state-dir
+          hostPath:
+            path: /var/lib/pmem-csi.intel.com
+            type: DirectoryOrCreate
         - name: dev-dir
           hostPath:
             path: /dev

diff --git a/pkg/pmem-csi-driver/controllerserver-master.go b/pkg/pmem-csi-driver/controllerserver-master.go
@@ -53,8 +53,6 @@ type pmemVolume struct {
 	// ID of nodes where the volume provisioned/attached
 	// It would be one if simple volume, else would be more than one for "cached" volume
 	nodeIDs map[string]VolumeStatus
-	// VolumeType
-	volumeType PmemPersistencyModel
 }
 
 type masterController struct {
@@ -65,6 +63,7 @@ type masterController struct {
 
 var _ csi.ControllerServer = &masterController{}
 var _ PmemService = &masterController{}
+var _ RegistryListener = &masterController{}
 var volumeMutex = keymutex.NewHashed(-1)
 
 func NewMasterControllerServer(rs *registryServer) *masterController {
@@ -73,17 +72,62 @@ func NewMasterControllerServer(rs *registryServer) *masterController {
 		csi.ControllerServiceCapability_RPC_LIST_VOLUMES,
 		csi.ControllerServiceCapability_RPC_GET_CAPACITY,
 	}
-	return &masterController{
+	cs := &masterController{
 		DefaultControllerServer: NewDefaultControllerServer(serverCaps),
 		rs:                      rs,
 		pmemVolumes:             map[string]*pmemVolume{},
 	}
+
+	rs.AddListener(cs)
+
+	return cs
 }
 
 func (cs *masterController) RegisterService(rpcServer *grpc.Server) {
 	csi.RegisterControllerServer(rpcServer, cs)
 }
 
+// OnNodeAdded retrieves the existing volumes at recently added Node.
+// It uses ControllerServer.ListVolume() CSI call to retrieve volumes.
+func (cs *masterController) OnNodeAdded(ctx context.Context, node NodeInfo) {
+	conn, err := cs.rs.ConnectToNodeController(node.NodeID)
+	if err != nil {
+		glog.Warningf("Failed to connect to node controller at : %s on node %s: %s", node.Endpoint, node.NodeID, err.Error())
+		return
+	}
+
+	csiClient := csi.NewControllerClient(conn)
+	resp, err := csiClient.ListVolumes(ctx, &csi.ListVolumesRequest{})
+	if err != nil {
+		glog.Warningf("Failed to get volumes on node %s: %s", node.NodeID, err.Error())
+	}
+
+	glog.V(5).Infof("Found Volumes at %s: %v", node.NodeID, resp.Entries)
+
+	for _, entry := range resp.Entries {
+		v := entry.GetVolume()
+		if v == nil { /* this shouldn't happen */
+			continue
+		}
+		if vol, ok := cs.pmemVolumes[v.VolumeId]; ok && vol != nil {
+			// This is possibly Cache volume, so just add this node id.
+			vol.nodeIDs[node.NodeID] = Created
+		} else {
+			cs.pmemVolumes[v.VolumeId] = &pmemVolume{
+				id:   v.VolumeId,
+				size: v.CapacityBytes,
+				name: v.VolumeContext["Name"],
+				nodeIDs: map[string]VolumeStatus{
+					node.NodeID: Created,
+				},
+			}
+		}
+	}
+}
+
+func (cs *masterController) OnNodeDeleted(ctx context.Context, node NodeInfo) {
+}
+
 func (cs *masterController) CreateVolume(ctx context.Context, req *csi.CreateVolumeRequest) (*csi.CreateVolumeResponse, error) {
 	var vol *pmemVolume
 	chosenNodes := map[string]VolumeStatus{}
@@ -117,14 +161,13 @@ func (cs *masterController) CreateVolume(ctx context.Context, req *csi.CreateVol
 		id, _ := uuid.NewUUID() //nolint: gosec
 		volumeID := id.String()
 		inTopology := []*csi.Topology{}
-		volumeType := pmemPersistencyModelNone
 		cacheCount := uint64(1)
 
 		if req.Parameters == nil {
 			req.Parameters = map[string]string{}
 		} else {
 			if val, ok := req.Parameters[pmemParameterKeyPersistencyModel]; ok {
-				volumeType = PmemPersistencyModel(val)
+				volumeType := PmemPersistencyModel(val)
 				if volumeType == pmemPersistencyModelCache {
 					if val, ok := req.Parameters[pmemParameterKeyCacheSize]; ok {
 						c, err := strconv.ParseUint(val, 10, 64)
@@ -191,11 +234,10 @@ func (cs *masterController) CreateVolume(ctx context.Context, req *csi.CreateVol
 		glog.V(3).Infof("Chosen nodes: %v", chosenNodes)
 
 		vol = &pmemVolume{
-			id:         volumeID,
-			name:       req.Name,
-			size:       asked,
-			nodeIDs:    chosenNodes,
-			volumeType: volumeType,
+			id:      volumeID,
+			name:    req.Name,
+			size:    asked,
+			nodeIDs: chosenNodes,
 		}
 		cs.pmemVolumes[volumeID] = vol
 		glog.V(3).Infof("CreateVolume: Record new volume as %v", *vol)