CSI: unique volume per allocation

CHANGELOG.md

@@ -14,6 +14,7 @@ BUG FIXES:
 IMPROVEMENTS:
  * cli: Update defaults for `nomad operator debug` flags `-interval` and `-server-id` to match common usage. [[GH-10121](https://github.com/hashicorp/nomad/issues/10121)]
  * consul/connect: Enable setting `local_bind_address` field on connect upstreams [[GH-6248](https://github.com/hashicorp/nomad/issues/6248)]
+ * csi: Added support for jobs to request a unique volume ID per allocation. [[GH-10136](https://github.com/hashicorp/nomad/issues/10136)]
  * driver/docker: Added support for optional extra container labels. [[GH-9885](https://github.com/hashicorp/nomad/issues/9885)]
  * driver/docker: Added support for configuring default logger behavior in the client configuration. [[GH-10156](https://github.com/hashicorp/nomad/issues/10156)]
 

api/tasks.go

@@ -382,6 +382,7 @@ type VolumeRequest struct {
 	Source       string           `hcl:"source,optional"`
 	ReadOnly     bool             `hcl:"read_only,optional"`
 	MountOptions *CSIMountOptions `hcl:"mount_options,block"`
+	PerAlloc     bool             `hcl:"per_alloc,optional"`
 	ExtraKeysHCL []string         `hcl1:",unusedKeys,optional" json:"-"`
 }
 

client/allocrunner/csi_hook.go

@@ -92,8 +92,14 @@ func (c *csiHook) Postrun() error {
 			mode = structs.CSIVolumeClaimWrite
 		}
 
+		source := pair.request.Source
+		if pair.request.PerAlloc {
+			// NOTE: PerAlloc can't be set if we have canaries
+			source = source + structs.AllocSuffix(c.alloc.Name)
+		}
+
 		req := &structs.CSIVolumeUnpublishRequest{
-			VolumeID: pair.request.Source,
+			VolumeID: source,
 			Claim: &structs.CSIVolumeClaim{
 				AllocationID: c.alloc.ID,
 				NodeID:       c.alloc.NodeID,
@@ -159,8 +165,13 @@ func (c *csiHook) claimVolumesFromAlloc() (map[string]*volumeAndRequest, error)
 			claimType = structs.CSIVolumeClaimRead
 		}
 
+		source := pair.request.Source
+		if pair.request.PerAlloc {
+			source = source + structs.AllocSuffix(c.alloc.Name)
+		}
+
 		req := &structs.CSIVolumeClaimRequest{
-			VolumeID:     pair.request.Source,
+			VolumeID:     source,
 			AllocationID: c.alloc.ID,
 			NodeID:       c.alloc.NodeID,
 			Claim:        claimType,

command/agent/job_endpoint.go

@@ -944,6 +944,7 @@ func ApiTgToStructsTG(job *structs.Job, taskGroup *api.TaskGroup, tg *structs.Ta
 				Type:     v.Type,
 				ReadOnly: v.ReadOnly,
 				Source:   v.Source,
+				PerAlloc: v.PerAlloc,
 			}
 
 			if v.MountOptions != nil {

e2e/csi/input/use-ebs-volume.nomad

@@ -9,8 +9,9 @@ job "use-ebs-volume" {
 
   group "group" {
     volume "test" {
-      type   = "csi"
-      source = "ebs-vol0"
+      type      = "csi"
+      source    = "ebs-vol"
+      per_alloc = true
     }
 
     task "task" {

e2e/terraform/compute.tf

@@ -63,7 +63,7 @@ data "external" "packer_sha" {
 sha=$(git log -n 1 --pretty=format:%H packer)
 echo "{\"sha\":\"$${sha}\"}"
 EOT
-]
+  ]
 
 }
 

e2e/terraform/terraform.tfvars

@@ -15,5 +15,5 @@ nomad_local_binary = ""      # overrides nomad_sha and nomad_version if set
 
 # Example overrides:
 # nomad_sha = "38e23b62a7700c96f4898be777543869499fea0a"
-# nomad_local_binary = "../../pkg/linux_amd/nomad"
+# nomad_local_binary = "../../pkg/linux_amd64/nomad"
 # nomad_local_binary_client_windows_2016_amd64 = ["../../pkg/windows_amd64/nomad.exe"]

e2e/terraform/volumes.tf

@@ -30,8 +30,8 @@ data "template_file" "ebs_volume_hcl" {
   count    = var.volumes ? 1 : 0
   template = <<EOT
 type = "csi"
-id = "ebs-vol0"
-name = "ebs-vol0"
+id = "ebs-vol[0]"
+name = "ebs-vol"
 external_id = "${aws_ebs_volume.csi[0].id}"
 access_mode = "single-node-writer"
 attachment_mode = "file-system"

nomad/state/state_store.go

@@ -2135,10 +2135,7 @@ func (s *StateStore) CSIVolumeByID(ws memdb.WatchSet, namespace, id string) (*st
 	if obj == nil {
 		return nil, nil
 	}
-	vol, ok := obj.(*structs.CSIVolume)
-	if !ok {
-		return nil, fmt.Errorf("volume row conversion error")
-	}
+	vol := obj.(*structs.CSIVolume)
 
 	// we return the volume with the plugins denormalized by default,
 	// because the scheduler needs them for feasibility checking

nomad/structs/diff_test.go

@@ -3639,6 +3639,7 @@ func TestTaskGroupDiff(t *testing.T) {
 						Type:     "host",
 						Source:   "foo-src",
 						ReadOnly: true,
+						PerAlloc: true,
 					},
 				},
 			},
@@ -3656,6 +3657,12 @@ func TestTaskGroupDiff(t *testing.T) {
 								Old:  "",
 								New:  "foo",
 							},
+							{
+								Type: DiffTypeAdded,
+								Name: "PerAlloc",
+								Old:  "",
+								New:  "true",
+							},
 							{
 								Type: DiffTypeAdded,
 								Name: "ReadOnly",

nomad/structs/funcs.go

@@ -329,6 +329,17 @@ func AllocName(job, group string, idx uint) string {
 	return fmt.Sprintf("%s.%s[%d]", job, group, idx)
 }
 
+// AllocSuffix returns the alloc index suffix that was added by the AllocName
+// function above.
+func AllocSuffix(name string) string {
+	idx := strings.LastIndex(name, "[")
+	if idx == -1 {
+		return ""
+	}
+	suffix := name[idx:]
+	return suffix
+}
+
 // ACLPolicyListHash returns a consistent hash for a set of policies.
 func ACLPolicyListHash(policies []*ACLPolicy) string {
 	cacheKeyHash, err := blake2b.New256(nil)

nomad/structs/structs.go

@@ -6100,14 +6100,19 @@ func (tg *TaskGroup) Validate(j *Job) error {
 		mErr.Errors = append(mErr.Errors, fmt.Errorf("Only one task may be marked as leader"))
 	}
 
-	// Validate the Host Volumes
+	// Validate the volume requests
 	for name, decl := range tg.Volumes {
 		if !(decl.Type == VolumeTypeHost ||
 			decl.Type == VolumeTypeCSI) {
 			mErr.Errors = append(mErr.Errors, fmt.Errorf("Volume %s has unrecognised type %s", name, decl.Type))
 			continue
 		}
 
+		if decl.PerAlloc && tg.Update != nil && tg.Update.Canary > 0 {
+			mErr.Errors = append(mErr.Errors,
+				fmt.Errorf("Volume %s cannot be per_alloc when canaries are in use", name))
+		}
+
 		if decl.Source == "" {
 			mErr.Errors = append(mErr.Errors, fmt.Errorf("Volume %s has an empty source", name))
 		}

nomad/structs/structs_test.go

@@ -1106,6 +1106,28 @@ func TestTaskGroup_Validate(t *testing.T) {
 	err = tg.Validate(&Job{})
 	require.Contains(t, err.Error(), `Volume foo has an empty source`)
 
+	tg = &TaskGroup{
+		Name: "group-a",
+		Update: &UpdateStrategy{
+			Canary: 1,
+		},
+		Volumes: map[string]*VolumeRequest{
+			"foo": {
+				Type:     "csi",
+				PerAlloc: true,
+			},
+		},
+		Tasks: []*Task{
+			{
+				Name:      "task-a",
+				Resources: &Resources{},
+			},
+		},
+	}
+	err = tg.Validate(&Job{})
+	require.Contains(t, err.Error(), `Volume foo has an empty source`)
+	require.Contains(t, err.Error(), `Volume foo cannot be per_alloc when canaries are in use`)
+
 	tg = &TaskGroup{
 		Volumes: map[string]*VolumeRequest{
 			"foo": {

nomad/structs/volumes.go

@@ -91,6 +91,7 @@ type VolumeRequest struct {
 	Source       string
 	ReadOnly     bool
 	MountOptions *CSIMountOptions
+	PerAlloc     bool
 }
 
 func (v *VolumeRequest) Copy() *VolumeRequest {

scheduler/feasible.go

@@ -227,31 +227,38 @@ func (c *CSIVolumeChecker) SetNamespace(namespace string) {
 	c.namespace = namespace
 }
 
-func (c *CSIVolumeChecker) SetVolumes(volumes map[string]*structs.VolumeRequest) {
+func (c *CSIVolumeChecker) SetVolumes(allocName string, volumes map[string]*structs.VolumeRequest) {
+
 	xs := make(map[string]*structs.VolumeRequest)
+
 	// Filter to only CSI Volumes
 	for alias, req := range volumes {
 		if req.Type != structs.VolumeTypeCSI {
 			continue
 		}
-
-		xs[alias] = req
+		if req.PerAlloc {
+			// provide a unique volume source per allocation
+			copied := req.Copy()
+			copied.Source = copied.Source + structs.AllocSuffix(allocName)
+			xs[alias] = copied
+		} else {
+			xs[alias] = req
+		}
 	}
 	c.volumes = xs
 }
 
 func (c *CSIVolumeChecker) Feasible(n *structs.Node) bool {
-	hasPlugins, failReason := c.hasPlugins(n)
-
-	if hasPlugins {
+	ok, failReason := c.isFeasible(n)
+	if ok {
 		return true
 	}
 
 	c.ctx.Metrics().FilterNode(n, failReason)
 	return false
 }
 
-func (c *CSIVolumeChecker) hasPlugins(n *structs.Node) (bool, string) {
+func (c *CSIVolumeChecker) isFeasible(n *structs.Node) (bool, string) {
 	// We can mount the volume if
 	// - if required, a healthy controller plugin is running the driver
 	// - the volume has free claims, or this job owns the claims

scheduler/feasible_test.go

@@ -309,6 +309,13 @@ func TestCSIVolumeChecker(t *testing.T) {
 	require.NoError(t, err)
 	index++
 
+	vid3 := "volume-id[0]"
+	vol3 := vol.Copy()
+	vol3.ID = vid3
+	err = state.CSIVolumeRegister(index, []*structs.CSIVolume{vol3})
+	require.NoError(t, err)
+	index++
+
 	alloc := mock.Alloc()
 	alloc.NodeID = nodes[4].ID
 	alloc.Job.TaskGroups[0].Volumes = map[string]*structs.VolumeRequest{
@@ -332,11 +339,17 @@ func TestCSIVolumeChecker(t *testing.T) {
 	noVolumes := map[string]*structs.VolumeRequest{}
 
 	volumes := map[string]*structs.VolumeRequest{
-		"baz": {
+		"shared": {
 			Type:   "csi",
 			Name:   "baz",
 			Source: "volume-id",
 		},
+		"unique": {
+			Type:     "csi",
+			Name:     "baz",
+			Source:   "volume-id[0]",
+			PerAlloc: true,
+		},
 		"nonsense": {
 			Type:   "host",
 			Name:   "nonsense",
@@ -390,7 +403,7 @@ func TestCSIVolumeChecker(t *testing.T) {
 	}
 
 	for i, c := range cases {
-		checker.SetVolumes(c.RequestedVolumes)
+		checker.SetVolumes(alloc.Name, c.RequestedVolumes)
 		if act := checker.Feasible(c.Node); act != c.Result {
 			t.Fatalf("case(%d) failed: got %v; want %v", i, act, c.Result)
 		}
@@ -407,7 +420,7 @@ func TestCSIVolumeChecker(t *testing.T) {
 	checker.SetNamespace(structs.DefaultNamespace)
 
 	for _, node := range nodes {
-		checker.SetVolumes(volumes)
+		checker.SetVolumes(alloc.Name, volumes)
 		act := checker.Feasible(node)
 		require.False(t, act, "request with missing volume should never be feasible")
 	}

scheduler/generic_sched.go

@@ -547,6 +547,7 @@ func (s *GenericScheduler) computePlacements(destructive, place []placementResul
 
 			// Compute penalty nodes for rescheduled allocs
 			selectOptions := getSelectOptions(prevAllocation, preferredNode)
+			selectOptions.AllocName = missing.Name()
 			option := s.selectNextOption(tg, selectOptions)
 
 			// Store the available nodes by datacenter