libpod: defer volume plugin init when loading volumes from DB

Do not contact volume plugins in finalizeVolume or when loading from the
DB. Initialize lazily on inspect/mount/unmount/remove; fail eagerly on create.

Fixes #28110

Signed-off-by: Yoonseo Han <yooncer00@gmail.com>

Author: yoonseo-han

libpod/boltdb_state_internal.go

@@ -298,21 +298,6 @@ func (s *BoltState) getVolumeFromDB(name []byte, volume *Volume, volBkt *bolt.Bu
 	// Need this for UsesVolumeDriver() so set it now.
 	volume.runtime = s.runtime
 
-	// Retrieve volume driver
-	if volume.UsesVolumeDriver() {
-		plugin, err := s.runtime.getVolumePlugin(volume.config)
-		if err != nil {
-			// We want to fail gracefully here, to ensure that we
-			// can still remove volumes even if their plugin is
-			// missing. Otherwise, we end up with volumes that
-			// cannot even be retrieved from the database and will
-			// cause things like `volume ls` to fail.
-			logrus.Errorf("Volume %s uses volume plugin %s, but it cannot be accessed - some functionality may not be available: %v", volume.Name(), volume.config.Driver, err)
-		} else {
-			volume.plugin = plugin
-		}
-	}
-
 	// Get the lock
 	lock, err := s.runtime.lockManager.RetrieveLock(volume.config.LockID)
 	if err != nil {

libpod/runtime_volume_common.go

@@ -70,13 +70,15 @@ func (r *Runtime) newVolume(ctx context.Context, noCreatePluginVolume bool, opti
 		}
 	}
 
-	// Plugin can be nil if driver is local, but that's OK - superfluous
-	// assignment doesn't hurt much.
-	plugin, err := r.getVolumePlugin(volume.config)
-	if err != nil {
-		return nil, fmt.Errorf("volume %s uses volume plugin %s but it could not be retrieved: %w", volume.config.Name, volume.config.Driver, err)
+	var plugin *volplugin.VolumePlugin
+	if volume.UsesVolumeDriver() {
+		var err error
+		plugin, err = r.getVolumePlugin(volume.config)
+		if err != nil {
+			return nil, fmt.Errorf("volume %s uses volume plugin %s but it could not be retrieved: %w", volume.config.Name, volume.config.Driver, err)
+		}
+		volume.cacheVolumePlugin(plugin)
 	}
-	volume.plugin = plugin
 
 	if volume.config.Driver == define.VolumeDriverLocal {
 		logrus.Debugf("Validating options for local driver")
@@ -150,13 +152,13 @@ func (r *Runtime) newVolume(ctx context.Context, noCreatePluginVolume bool, opti
 
 	// Now we get conditional: we either need to make the volume in the
 	// volume plugin, or on disk if not using a plugin.
-	if volume.plugin != nil && !noCreatePluginVolume {
+	if volume.UsesVolumeDriver() && !noCreatePluginVolume {
 		// We can't chown, or relabel, or similar the path the volume is
 		// using, because it's not managed by us.
 		// TODO: reevaluate this once we actually have volume plugins in
 		// use in production - it may be safe, but I can't tell without
 		// knowing what the actual plugin does...
-		if err := makeVolumeInPluginIfNotExist(volume.config.Name, volume.config.Options, volume.plugin); err != nil {
+		if err := makeVolumeInPluginIfNotExist(volume.config.Name, volume.config.Options, plugin); err != nil {
 			return nil, err
 		}
 	} else {
@@ -437,27 +439,22 @@ func (r *Runtime) removeVolume(ctx context.Context, v *Volume, force bool, timeo
 	if v.UsesVolumeDriver() && !ignoreVolumePlugin {
 		canRemove := true
 
-		// Do we have a volume driver?
-		if v.plugin == nil {
+		plugin, err := v.getPlugin()
+		if err != nil {
 			canRemove = false
-			removalErr = fmt.Errorf("cannot remove volume %s from plugin %s, but it has been removed from Podman: %w", v.Name(), v.Driver(), define.ErrMissingPlugin)
+			removalErr = fmt.Errorf("cannot remove volume %s from plugin %s, but it has been removed from Podman: %w", v.Name(), v.Driver(), err)
 		} else {
-			// Ping the plugin first to verify the volume still
-			// exists.
-			// We're trying to be very tolerant of missing volumes
-			// in the backend, to avoid the problems we see with
-			// sync between c/storage and the Libpod DB.
 			getReq := new(pluginapi.GetRequest)
 			getReq.Name = v.Name()
-			if _, err := v.plugin.GetVolume(getReq); err != nil {
+			if _, err := plugin.GetVolume(getReq); err != nil {
 				canRemove = false
 				removalErr = fmt.Errorf("volume %s could not be retrieved from plugin %s, but it has been removed from Podman: %w", v.Name(), v.Driver(), err)
 			}
 		}
 		if canRemove {
 			req := new(pluginapi.RemoveRequest)
 			req.Name = v.Name()
-			if err := v.plugin.RemoveVolume(req); err != nil {
+			if err := plugin.RemoveVolume(req); err != nil {
 				return fmt.Errorf("volume %s could not be removed from plugin %s: %w", v.Name(), v.Driver(), err)
 			}
 		}

libpod/sqlite_state_internal.go

@@ -350,21 +350,6 @@ func finalizeVolumeSqlite(vol *Volume) error {
 	}
 	vol.lock = lock
 
-	// Retrieve volume driver
-	if vol.UsesVolumeDriver() {
-		plugin, err := vol.runtime.getVolumePlugin(vol.config)
-		if err != nil {
-			// We want to fail gracefully here, to ensure that we
-			// can still remove volumes even if their plugin is
-			// missing. Otherwise, we end up with volumes that
-			// cannot even be retrieved from the database and will
-			// cause things like `volume ls` to fail.
-			logrus.Errorf("Volume %s uses volume plugin %s, but it cannot be accessed - some functionality may not be available: %v", vol.Name(), vol.config.Driver, err)
-		} else {
-			vol.plugin = plugin
-		}
-	}
-
 	vol.valid = true
 
 	return nil

libpod/volume.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"maps"
+	"sync"
 	"time"
 
 	"github.com/sirupsen/logrus"
@@ -27,7 +28,7 @@ type Volume struct {
 
 	ignoreIfExists bool
 	valid          bool
-	plugin         *plugin.VolumePlugin
+	volumePlugin   func() (*plugin.VolumePlugin, error)
 	runtime        *Runtime
 	lock           lock.Locker
 }
@@ -130,6 +131,26 @@ func (v *Volume) Driver() string {
 	return v.config.Driver
 }
 
+// getPlugin returns the volume plugin driver, initializing on first use.
+func (v *Volume) getPlugin() (*plugin.VolumePlugin, error) {
+	if !v.UsesVolumeDriver() {
+		return nil, nil
+	}
+	if v.volumePlugin == nil {
+		v.volumePlugin = sync.OnceValues(func() (*plugin.VolumePlugin, error) {
+			return v.runtime.getVolumePlugin(v.config)
+		})
+	}
+	return v.volumePlugin()
+}
+
+// cacheVolumePlugin records a plugin that was already validated (e.g. volume create).
+func (v *Volume) cacheVolumePlugin(p *plugin.VolumePlugin) {
+	v.volumePlugin = sync.OnceValues(func() (*plugin.VolumePlugin, error) {
+		return p, nil
+	})
+}
+
 // Scope retrieves the volume's scope.
 // Libpod does not implement volume scoping, and this is provided solely for
 // Docker compatibility. It returns only "local".

libpod/volume_inspect.go

@@ -32,15 +32,16 @@ func (v *Volume) Inspect() (*define.InspectVolumeData, error) {
 		logrus.Debugf("Querying volume plugin %s for status", v.config.Driver)
 		data.Mountpoint = v.state.MountPoint
 
-		if v.plugin == nil {
-			return nil, fmt.Errorf("volume %s uses volume plugin %s but it is not available, cannot inspect: %w", v.Name(), v.config.Driver, define.ErrMissingPlugin)
+		plugin, err := v.getPlugin()
+		if err != nil {
+			return nil, fmt.Errorf("volume %s uses volume plugin %s but it is not available, cannot inspect: %w", v.Name(), v.config.Driver, err)
 		}
 
 		// Retrieve status for the volume.
 		// Need to query the volume driver.
 		req := new(pluginapi.GetRequest)
 		req.Name = v.Name()
-		resp, err := v.plugin.GetVolume(req)
+		resp, err := plugin.GetVolume(req)
 		if err != nil {
 			return nil, fmt.Errorf("retrieving volume %s information from plugin %s: %w", v.Name(), v.Driver(), err)
 		}

libpod/volume_internal_common.go

@@ -50,14 +50,15 @@ func (v *Volume) mount() error {
 	// ours more. So hardcode container ID to something reasonable, and use
 	// the same one for everything.
 	if v.UsesVolumeDriver() {
-		if v.plugin == nil {
-			return fmt.Errorf("volume plugin %s (needed by volume %s) missing: %w", v.Driver(), v.Name(), define.ErrMissingPlugin)
+		plugin, err := v.getPlugin()
+		if err != nil {
+			return fmt.Errorf("volume plugin %s (needed by volume %s) missing: %w", v.Driver(), v.Name(), err)
 		}
 
 		req := new(pluginapi.MountRequest)
 		req.Name = v.Name()
 		req.ID = pseudoCtrID
-		mountPoint, err := v.plugin.MountVolume(req)
+		mountPoint, err := plugin.MountVolume(req)
 		if err != nil {
 			return err
 		}
@@ -156,14 +157,15 @@ func (v *Volume) unmount(force bool) error {
 
 	if v.state.MountCount == 0 {
 		if v.UsesVolumeDriver() {
-			if v.plugin == nil {
-				return fmt.Errorf("volume plugin %s (needed by volume %s) missing: %w", v.Driver(), v.Name(), define.ErrMissingPlugin)
+			plugin, err := v.getPlugin()
+			if err != nil {
+				return fmt.Errorf("volume plugin %s (needed by volume %s) missing: %w", v.Driver(), v.Name(), err)
 			}
 
 			req := new(pluginapi.UnmountRequest)
 			req.Name = v.Name()
 			req.ID = pseudoCtrID
-			if err := v.plugin.UnmountVolume(req); err != nil {
+			if err := plugin.UnmountVolume(req); err != nil {
 				return err
 			}