fusemanager: fix container fail after ttl timeout in detach mode

cmd/containerd-stargz-grpc/main.go

@@ -181,7 +181,7 @@ func main() {
 		if err != nil {
 			log.G(ctx).WithError(err).Fatalf("failed to configure fusemanager")
 		}
-		rs, err = snbase.NewSnapshotter(ctx, filepath.Join(*rootDir, "snapshotter"), fs, snbase.AsynchronousRemove)
+		rs, err = snbase.NewSnapshotter(ctx, filepath.Join(*rootDir, "snapshotter"), fs, snbase.AsynchronousRemove, snbase.SetDetachFlag)
 		if err != nil {
 			log.G(ctx).WithError(err).Fatalf("failed to configure snapshotter")
 		}
@@ -213,10 +213,6 @@ func main() {
 		log.G(ctx).WithError(err).Fatalf("failed to serve snapshotter")
 	}
 
-	// TODO: In detach mode, rs is taken over by fusemanager,
-	// but client will send unmount request to fusemanager,
-	// and fusemanager need get mount info from local db to
-	//  determine its behavior
 	if cleanup {
 		log.G(ctx).Debug("Closing the snapshotter")
 		rs.Close()

docs/overview.md

@@ -118,6 +118,18 @@ When upgrading the fuse manager, it's recommended to follow these steps:
 
 This ensures a clean upgrade without impacting running containers.
 
+### Important Considerations
+
+Before restarting the `containerd-stargz-grpc` process, it is essential to consider the state of any running containers.
+
+1. **When to Use SIGKILL** :
+
+If there are running containers, it is crucial to terminate the `containerd-stargz-grpc` process using `SIGKILL`. This approach prevents the normal shutdown sequence from attempting to clean up the mount points of the running containers, which could disrupt their availability. By using `SIGKILL`, you ensure that the process is forcefully terminated without affecting the ongoing operations of the containers.
+
+2. **When to Use SIGTERM** :
+
+If there are no running containers, you should use `SIGTERM` to terminate the `containerd-stargz-grpc` process. This allows the process to follow its normal shutdown sequence, ensuring that it properly cleans up resources and mount points.
+
 ## Registry-related configuration
 
 You can configure stargz snapshotter for accessing registries with custom configurations.

snapshot/snapshot.go

@@ -73,6 +73,7 @@ type SnapshotterConfig struct {
 	asyncRemove                 bool
 	noRestore                   bool
 	allowInvalidMountsOnRestart bool
+	detach                      bool
 }
 
 // Opt is an option to configure the remote snapshotter
@@ -97,6 +98,11 @@ func AllowInvalidMountsOnRestart(config *SnapshotterConfig) error {
 	return nil
 }
 
+func SetDetachFlag(config *SnapshotterConfig) error {
+	config.detach = true
+	return nil
+}
+
 type snapshotter struct {
 	root        string
 	ms          *storage.MetaStore
@@ -107,6 +113,7 @@ type snapshotter struct {
 	userxattr                   bool // whether to enable "userxattr" mount option
 	noRestore                   bool
 	allowInvalidMountsOnRestart bool
+	detach                      bool
 }
 
 // NewSnapshotter returns a Snapshotter which can use unpacked remote layers
@@ -157,6 +164,7 @@ func NewSnapshotter(ctx context.Context, root string, targetFs FileSystem, opts
 		userxattr:                   userxattr,
 		noRestore:                   config.noRestore,
 		allowInvalidMountsOnRestart: config.allowInvalidMountsOnRestart,
+		detach:                      config.detach,
 	}
 
 	if err := o.restoreRemoteSnapshot(ctx); err != nil {
@@ -736,6 +744,10 @@ func (o *snapshotter) checkAvailability(ctx context.Context, key string) bool {
 }
 
 func (o *snapshotter) restoreRemoteSnapshot(ctx context.Context) error {
+	if o.detach {
+		return nil
+	}
+
 	mounts, err := mountinfo.GetMounts(nil)
 	if err != nil {
 		return err