CSI: retry claims from client when max claims are reached

When the alloc runner claims a volume, an allocation for a previous
version of the job may still have the volume claimed because it's
still shutting down. In this case we'll receive an error from the
server. Retry this error until we succeed or until a very long timeout
expires, to give operators a chance to recover broken plugins.

.changelog/12113.txt

@@ -0,0 +1,3 @@
+```release-note:bug
+csi: Fixed a bug where allocations with volume claims would fail their first placement after a reschedule
+```

client/allocrunner/csi_hook.go

@@ -3,6 +3,7 @@ package allocrunner
 import (
 	"context"
 	"fmt"
+	"strings"
 	"sync"
 	"time"
 
@@ -28,6 +29,7 @@ type csiHook struct {
 	nodeSecret           string
 
 	volumeRequests     map[string]*volumeAndRequest
+	minBackoffInterval time.Duration
 	maxBackoffInterval time.Duration
 	maxBackoffDuration time.Duration
 }
@@ -47,6 +49,7 @@ func newCSIHook(alloc *structs.Allocation, logger hclog.Logger, csi csimanager.M
 		updater:              updater,
 		nodeSecret:           nodeSecret,
 		volumeRequests:       map[string]*volumeAndRequest{},
+		minBackoffInterval:   time.Second,
 		maxBackoffInterval:   time.Minute,
 		maxBackoffDuration:   time.Hour * 24,
 	}
@@ -213,11 +216,10 @@ func (c *csiHook) claimVolumesFromAlloc() (map[string]*volumeAndRequest, error)
 			},
 		}
 
-		var resp structs.CSIVolumeClaimResponse
-		if err := c.rpcClient.RPC("CSIVolume.Claim", req, &resp); err != nil {
+		resp, err := c.claimWithRetry(req)
+		if err != nil {
 			return nil, fmt.Errorf("could not claim volume %s: %w", req.VolumeID, err)
 		}
-
 		if resp.Volume == nil {
 			return nil, fmt.Errorf("Unexpected nil volume returned for ID: %v", pair.request.Source)
 		}
@@ -230,6 +232,51 @@ func (c *csiHook) claimVolumesFromAlloc() (map[string]*volumeAndRequest, error)
 	return result, nil
 }
 
+// claimWithRetry tries to claim the volume on the server, retrying
+// with exponential backoff capped to a maximum interval
+func (c *csiHook) claimWithRetry(req *structs.CSIVolumeClaimRequest) (*structs.CSIVolumeClaimResponse, error) {
+
+	// note: allocrunner hooks don't have access to the client's
+	// shutdown context, just the allocrunner's shutdown; if we make
+	// it available in the future we should thread it through here so
+	// that retry can exit gracefully instead of dropping the
+	// in-flight goroutine
+	ctx, cancel := context.WithTimeout(context.TODO(), c.maxBackoffDuration)
+	defer cancel()
+
+	var resp structs.CSIVolumeClaimResponse
+	var err error
+	backoff := c.minBackoffInterval
+	t, stop := helper.NewSafeTimer(0)
+	defer stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, err
+		case <-t.C:
+		}
+
+		err = c.rpcClient.RPC("CSIVolume.Claim", req, &resp)
+		if err == nil {
+			break
+		}
+		if !strings.Contains(err.Error(), structs.ErrCSIVolumeMaxClaims.Error()) {
+			break
+		}
+
+		if backoff < c.maxBackoffInterval {
+			backoff = backoff * 2
+			if backoff > c.maxBackoffInterval {
+				backoff = c.maxBackoffInterval
+			}
+		}
+		c.logger.Debug(
+			"volume could not be claimed because it is in use, retrying in %v", backoff)
+		t.Reset(backoff)
+	}
+	return &resp, err
+}
+
 func (c *csiHook) shouldRun() bool {
 	tg := c.alloc.Job.LookupTaskGroup(c.alloc.TaskGroup)
 	for _, vol := range tg.Volumes {
@@ -286,7 +333,7 @@ func (c *csiHook) unmountWithRetry(pair *volumeAndRequest) error {
 	ctx, cancel := context.WithTimeout(context.TODO(), c.maxBackoffDuration)
 	defer cancel()
 	var err error
-	backoff := time.Second
+	backoff := c.minBackoffInterval
 	t, stop := helper.NewSafeTimer(0)
 	defer stop()
 	for {
@@ -307,6 +354,8 @@ func (c *csiHook) unmountWithRetry(pair *volumeAndRequest) error {
 				backoff = c.maxBackoffInterval
 			}
 		}
+		c.logger.Debug(
+			"volume could not be unmounted, retrying in %v", backoff)
 		t.Reset(backoff)
 	}
 	return nil

client/allocrunner/csi_hook_test.go

@@ -7,12 +7,14 @@ import (
 	"testing"
 	"time"
 
+	"github.com/pkg/errors"
 	"github.com/stretchr/testify/require"
 
 	"github.com/hashicorp/nomad/client/allocrunner/interfaces"
 	"github.com/hashicorp/nomad/client/pluginmanager"
 	"github.com/hashicorp/nomad/client/pluginmanager/csimanager"
 	cstructs "github.com/hashicorp/nomad/client/structs"
+	"github.com/hashicorp/nomad/helper"
 	"github.com/hashicorp/nomad/helper/testlog"
 	"github.com/hashicorp/nomad/nomad/mock"
 	"github.com/hashicorp/nomad/nomad/structs"
@@ -34,6 +36,9 @@ func TestCSIHook(t *testing.T) {
 	testcases := []struct {
 		name                   string
 		volumeRequests         map[string]*structs.VolumeRequest
+		startsUnschedulable    bool
+		startsWithClaims       bool
+		expectedClaimErr       error
 		expectedMounts         map[string]*csimanager.MountInfo
 		expectedMountCalls     int
 		expectedUnmountCalls   int
@@ -89,6 +94,58 @@ func TestCSIHook(t *testing.T) {
 			expectedUnpublishCalls: 1,
 		},
 
+		{
+			name: "fatal error on claim",
+			volumeRequests: map[string]*structs.VolumeRequest{
+				"vol0": {
+					Name:           "vol0",
+					Type:           structs.VolumeTypeCSI,
+					Source:         "testvolume0",
+					ReadOnly:       true,
+					AccessMode:     structs.CSIVolumeAccessModeSingleNodeReader,
+					AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
+					MountOptions:   &structs.CSIMountOptions{},
+					PerAlloc:       false,
+				},
+			},
+			startsUnschedulable: true,
+			expectedMounts: map[string]*csimanager.MountInfo{
+				"vol0": &csimanager.MountInfo{Source: fmt.Sprintf(
+					"test-alloc-dir/%s/testvolume0/ro-file-system-single-node-reader-only", alloc.ID)},
+			},
+			expectedMountCalls:     0,
+			expectedUnmountCalls:   0,
+			expectedClaimCalls:     1,
+			expectedUnpublishCalls: 0,
+			expectedClaimErr: errors.New(
+				"claim volumes: could not claim volume testvolume0: volume is currently unschedulable"),
+		},
+
+		{
+			name: "retryable error on claim",
+			volumeRequests: map[string]*structs.VolumeRequest{
+				"vol0": {
+					Name:           "vol0",
+					Type:           structs.VolumeTypeCSI,
+					Source:         "testvolume0",
+					ReadOnly:       true,
+					AccessMode:     structs.CSIVolumeAccessModeSingleNodeReader,
+					AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
+					MountOptions:   &structs.CSIMountOptions{},
+					PerAlloc:       false,
+				},
+			},
+			startsWithClaims: true,
+			expectedMounts: map[string]*csimanager.MountInfo{
+				"vol0": &csimanager.MountInfo{Source: fmt.Sprintf(
+					"test-alloc-dir/%s/testvolume0/ro-file-system-single-node-reader-only", alloc.ID)},
+			},
+			expectedMountCalls:     1,
+			expectedUnmountCalls:   1,
+			expectedClaimCalls:     2,
+			expectedUnpublishCalls: 1,
+		},
+
 		// TODO: this won't actually work on the client.
 		// https://github.com/hashicorp/nomad/issues/11798
 		//
@@ -136,7 +193,12 @@ func TestCSIHook(t *testing.T) {
 
 			callCounts := map[string]int{}
 			mgr := mockPluginManager{mounter: mockVolumeMounter{callCounts: callCounts}}
-			rpcer := mockRPCer{alloc: alloc, callCounts: callCounts}
+			rpcer := mockRPCer{
+				alloc:            alloc,
+				callCounts:       callCounts,
+				hasExistingClaim: helper.BoolToPtr(tc.startsWithClaims),
+				schedulable:      helper.BoolToPtr(!tc.startsUnschedulable),
+			}
 			ar := mockAllocRunner{
 				res: &cstructs.AllocHookResources{},
 				caps: &drivers.Capabilities{
@@ -145,17 +207,24 @@ func TestCSIHook(t *testing.T) {
 				},
 			}
 			hook := newCSIHook(alloc, logger, mgr, rpcer, ar, ar, "secret")
-			hook.maxBackoffInterval = 100 * time.Millisecond
-			hook.maxBackoffDuration = 2 * time.Second
+			hook.minBackoffInterval = 1 * time.Millisecond
+			hook.maxBackoffInterval = 10 * time.Millisecond
+			hook.maxBackoffDuration = 500 * time.Millisecond
 
 			require.NotNil(t, hook)
 
-			require.NoError(t, hook.Prerun())
-			mounts := ar.GetAllocHookResources().GetCSIMounts()
-			require.NotNil(t, mounts)
-			require.Equal(t, tc.expectedMounts, mounts)
+			if tc.expectedClaimErr != nil {
+				require.EqualError(t, hook.Prerun(), tc.expectedClaimErr.Error())
+				mounts := ar.GetAllocHookResources().GetCSIMounts()
+				require.Nil(t, mounts)
+			} else {
+				require.NoError(t, hook.Prerun())
+				mounts := ar.GetAllocHookResources().GetCSIMounts()
+				require.NotNil(t, mounts)
+				require.Equal(t, tc.expectedMounts, mounts)
+				require.NoError(t, hook.Postrun())
+			}
 
-			require.NoError(t, hook.Postrun())
 			require.Equal(t, tc.expectedMountCalls, callCounts["mount"])
 			require.Equal(t, tc.expectedUnmountCalls, callCounts["unmount"])
 			require.Equal(t, tc.expectedClaimCalls, callCounts["claim"])
@@ -168,25 +237,11 @@ func TestCSIHook(t *testing.T) {
 
 // HELPERS AND MOCKS
 
-func testVolume(id string) *structs.CSIVolume {
-	vol := structs.NewCSIVolume(id, 0)
-	vol.Schedulable = true
-	vol.RequestedCapabilities = []*structs.CSIVolumeCapability{
-		{
-			AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
-			AccessMode:     structs.CSIVolumeAccessModeSingleNodeReader,
-		},
-		{
-			AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
-			AccessMode:     structs.CSIVolumeAccessModeSingleNodeWriter,
-		},
-	}
-	return vol
-}
-
 type mockRPCer struct {
-	alloc      *structs.Allocation
-	callCounts map[string]int
+	alloc            *structs.Allocation
+	callCounts       map[string]int
+	hasExistingClaim *bool
+	schedulable      *bool
 }
 
 // RPC mocks the server RPCs, acting as though any request succeeds
@@ -195,7 +250,7 @@ func (r mockRPCer) RPC(method string, args interface{}, reply interface{}) error
 	case "CSIVolume.Claim":
 		r.callCounts["claim"]++
 		req := args.(*structs.CSIVolumeClaimRequest)
-		vol := testVolume(req.VolumeID)
+		vol := r.testVolume(req.VolumeID)
 		err := vol.Claim(req.ToClaim(), r.alloc)
 		if err != nil {
 			return err
@@ -215,6 +270,44 @@ func (r mockRPCer) RPC(method string, args interface{}, reply interface{}) error
 	return nil
 }
 
+// testVolume is a helper that optionally starts as unschedulable /
+// claimed until after the first claim RPC is made, so that we can
+// test retryable vs non-retryable failures
+func (r mockRPCer) testVolume(id string) *structs.CSIVolume {
+	vol := structs.NewCSIVolume(id, 0)
+	vol.Schedulable = *r.schedulable
+	vol.RequestedCapabilities = []*structs.CSIVolumeCapability{
+		{
+			AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
+			AccessMode:     structs.CSIVolumeAccessModeSingleNodeReader,
+		},
+		{
+			AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
+			AccessMode:     structs.CSIVolumeAccessModeSingleNodeWriter,
+		},
+	}
+
+	if *r.hasExistingClaim {
+		vol.AccessMode = structs.CSIVolumeAccessModeSingleNodeReader
+		vol.AttachmentMode = structs.CSIVolumeAttachmentModeFilesystem
+		vol.ReadClaims["another-alloc-id"] = &structs.CSIVolumeClaim{
+			AllocationID:   "another-alloc-id",
+			NodeID:         "another-node-id",
+			Mode:           structs.CSIVolumeClaimRead,
+			AccessMode:     structs.CSIVolumeAccessModeSingleNodeReader,
+			AttachmentMode: structs.CSIVolumeAttachmentModeFilesystem,
+			State:          structs.CSIVolumeClaimStateTaken,
+		}
+	}
+
+	if r.callCounts["claim"] >= 0 {
+		*r.hasExistingClaim = false
+		*r.schedulable = true
+	}
+
+	return vol
+}
+
 type mockVolumeMounter struct {
 	callCounts map[string]int
 }