[vfs2] Don't take reference in Task.MountNamespaceVFS2 and MountNamespace.Root.

This fixes reference leaks related to accidentally forgetting to DecRef()
after calling one or the other.

PiperOrigin-RevId: 336918922

pkg/sentry/control/proc.go

@@ -183,9 +183,9 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI
 		if initArgs.MountNamespaceVFS2 == nil {
 			// Set initArgs so that 'ctx' returns the namespace.
 			//
-			// MountNamespaceVFS2 adds a reference to the namespace, which is
-			// transferred to the new process.
+			// Add a reference to the namespace, which is transferred to the new process.
 			initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2()
+			initArgs.MountNamespaceVFS2.IncRef()
 		}
 	} else {
 		if initArgs.MountNamespace == nil {

pkg/sentry/fs/user/path.go

@@ -121,6 +121,7 @@ func resolve(ctx context.Context, mns *fs.MountNamespace, paths []string, name s
 
 func resolveVFS2(ctx context.Context, creds *auth.Credentials, mns *vfs.MountNamespace, paths []string, name string) (string, error) {
 	root := mns.Root()
+	root.IncRef()
 	defer root.DecRef(ctx)
 	for _, p := range paths {
 		if !path.IsAbs(p) {

pkg/sentry/fs/user/user.go

@@ -105,6 +105,7 @@ func getExecUserHomeVFS2(ctx context.Context, mns *vfs.MountNamespace, uid auth.
 	const defaultHome = "/"
 
 	root := mns.Root()
+	root.IncRef()
 	defer root.DecRef(ctx)
 
 	creds := auth.CredentialsFromContext(ctx)

pkg/sentry/fsimpl/devtmpfs/devtmpfs.go

@@ -95,10 +95,13 @@ func NewAccessor(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth
 	if err != nil {
 		return nil, err
 	}
+	// Pass a reference on root to the Accessor.
+	root := mntns.Root()
+	root.IncRef()
 	return &Accessor{
 		vfsObj: vfsObj,
 		mntns:  mntns,
-		root:   mntns.Root(),
+		root:   root,
 		creds:  creds,
 	}, nil
 }

pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go

@@ -53,6 +53,7 @@ func setupDevtmpfs(t *testing.T) (context.Context, *auth.Credentials, *vfs.Virtu
 		t.Fatalf("failed to create tmpfs root mount: %v", err)
 	}
 	root := mntns.Root()
+	root.IncRef()
 	devpop := vfs.PathOperation{
 		Root:  root,
 		Start: root,

pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go

@@ -70,6 +70,7 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys
 	}
 
 	root := mntns.Root()
+	root.IncRef()
 
 	tearDown := func() {
 		root.DecRef(ctx)

pkg/sentry/fsimpl/ext/ext_test.go

@@ -82,6 +82,7 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys
 	}
 
 	root := mntns.Root()
+	root.IncRef()
 
 	tearDown := func() {
 		root.DecRef(ctx)

pkg/sentry/fsimpl/gofer/BUILD

@@ -52,6 +52,7 @@ go_library(
         "//pkg/fspath",
         "//pkg/log",
         "//pkg/p9",
+        "//pkg/refs",
         "//pkg/safemem",
         "//pkg/sentry/fs/fsutil",
         "//pkg/sentry/fs/lock",

pkg/sentry/fsimpl/proc/tasks_test.go

@@ -109,9 +109,12 @@ func setup(t *testing.T) *testutil.System {
 	if err != nil {
 		t.Fatalf("NewMountNamespace(): %v", err)
 	}
+	root := mntns.Root()
+	root.IncRef()
+	defer root.DecRef(ctx)
 	pop := &vfs.PathOperation{
-		Root:  mntns.Root(),
-		Start: mntns.Root(),
+		Root:  root,
+		Start: root,
 		Path:  fspath.Parse("/proc"),
 	}
 	if err := k.VFS().MkdirAt(ctx, creds, pop, &vfs.MkdirOptions{Mode: 0777}); err != nil {
@@ -119,8 +122,8 @@ func setup(t *testing.T) *testutil.System {
 	}
 
 	pop = &vfs.PathOperation{
-		Root:  mntns.Root(),
-		Start: mntns.Root(),
+		Root:  root,
+		Start: root,
 		Path:  fspath.Parse("/proc"),
 	}
 	mntOpts := &vfs.MountOptions{

pkg/sentry/fsimpl/testutil/testutil.go

@@ -46,16 +46,18 @@ type System struct {
 
 // NewSystem constructs a System.
 //
-// Precondition: Caller must hold a reference on MntNs, whose ownership
+// Precondition: Caller must hold a reference on mns, whose ownership
 // is transferred to the new System.
 func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns *vfs.MountNamespace) *System {
+	root := mns.Root()
+	root.IncRef()
 	s := &System{
 		t:     t,
 		Ctx:   ctx,
 		Creds: auth.CredentialsFromContext(ctx),
 		VFS:   v,
 		MntNs: mns,
-		Root:  mns.Root(),
+		Root:  root,
 	}
 	return s
 }
@@ -254,10 +256,10 @@ func (d *DirentCollector) Contains(name string, typ uint8) error {
 	defer d.mu.Unlock()
 	dirent, ok := d.dirents[name]
 	if !ok {
-		return fmt.Errorf("No dirent named %q found", name)
+		return fmt.Errorf("no dirent named %q found", name)
 	}
 	if dirent.Type != typ {
-		return fmt.Errorf("Dirent named %q found, but was expecting type %s, got: %+v", name, linux.DirentType.Parse(uint64(typ)), dirent)
+		return fmt.Errorf("dirent named %q found, but was expecting type %s, got: %+v", name, linux.DirentType.Parse(uint64(typ)), dirent)
 	}
 	return nil
 }

pkg/sentry/fsimpl/tmpfs/benchmark_test.go

@@ -193,6 +193,7 @@ func BenchmarkVFS2TmpfsStat(b *testing.B) {
 
 			// Create nested directories with given depth.
 			root := mntns.Root()
+			root.IncRef()
 			defer root.DecRef(ctx)
 			vd := root
 			vd.IncRef()
@@ -387,6 +388,7 @@ func BenchmarkVFS2TmpfsMountStat(b *testing.B) {
 
 			// Create the mount point.
 			root := mntns.Root()
+			root.IncRef()
 			defer root.DecRef(ctx)
 			pop := vfs.PathOperation{
 				Root:  root,

pkg/sentry/fsimpl/tmpfs/pipe_test.go

@@ -165,6 +165,7 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy
 
 	// Create the pipe.
 	root := mntns.Root()
+	root.IncRef()
 	pop := vfs.PathOperation{
 		Root:  root,
 		Start: root,

pkg/sentry/fsimpl/tmpfs/tmpfs_test.go

@@ -46,6 +46,7 @@ func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentr
 		return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err)
 	}
 	root := mntns.Root()
+	root.IncRef()
 	return vfsObj, root, func() {
 		root.DecRef(ctx)
 		mntns.DecRef(ctx)

pkg/sentry/fsimpl/verity/verity_test.go

@@ -70,6 +70,7 @@ func newVerityRoot(ctx context.Context, t *testing.T) (*vfs.VirtualFilesystem, v
 		return nil, vfs.VirtualDentry{}, fmt.Errorf("NewMountNamespace: %v", err)
 	}
 	root := mntns.Root()
+	root.IncRef()
 	t.Helper()
 	t.Cleanup(func() {
 		root.DecRef(ctx)

pkg/sentry/kernel/kernel.go

@@ -841,14 +841,16 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} {
 		if ctx.args.MountNamespaceVFS2 == nil {
 			return nil
 		}
-		// MountNamespaceVFS2.Root() takes a reference on the root dirent for us.
-		return ctx.args.MountNamespaceVFS2.Root()
+		root := ctx.args.MountNamespaceVFS2.Root()
+		root.IncRef()
+		return root
 	case vfs.CtxMountNamespace:
 		if ctx.k.globalInit == nil {
 			return nil
 		}
-		// MountNamespaceVFS2 takes a reference for us.
-		return ctx.k.GlobalInit().Leader().MountNamespaceVFS2()
+		mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2()
+		mntns.IncRef()
+		return mntns
 	case fs.CtxDirentCacheLimiter:
 		return ctx.k.DirentCacheLimiter
 	case inet.CtxStack:
@@ -904,14 +906,13 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID,
 	if VFS2Enabled {
 		mntnsVFS2 = args.MountNamespaceVFS2
 		if mntnsVFS2 == nil {
-			// MountNamespaceVFS2 adds a reference to the namespace, which is
-			// transferred to the new process.
+			// Add a reference to the namespace, which is transferred to the new process.
 			mntnsVFS2 = k.globalInit.Leader().MountNamespaceVFS2()
+			mntnsVFS2.IncRef()
 		}
 		// Get the root directory from the MountNamespace.
 		root := mntnsVFS2.Root()
-		// The call to newFSContext below will take a reference on root, so we
-		// don't need to hold this one.
+		root.IncRef()
 		defer root.DecRef(ctx)
 
 		// Grab the working directory.
@@ -1648,16 +1649,16 @@ func (ctx supervisorContext) Value(key interface{}) interface{} {
 		if ctx.k.globalInit == nil {
 			return vfs.VirtualDentry{}
 		}
-		mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2()
-		defer mntns.DecRef(ctx)
-		// Root() takes a reference on the root dirent for us.
-		return mntns.Root()
+		root := ctx.k.GlobalInit().Leader().MountNamespaceVFS2().Root()
+		root.IncRef()
+		return root
 	case vfs.CtxMountNamespace:
 		if ctx.k.globalInit == nil {
 			return nil
 		}
-		// MountNamespaceVFS2() takes a reference for us.
-		return ctx.k.GlobalInit().Leader().MountNamespaceVFS2()
+		mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2()
+		mntns.IncRef()
+		return mntns
 	case fs.CtxDirentCacheLimiter:
 		return ctx.k.DirentCacheLimiter
 	case inet.CtxStack:

pkg/sentry/kernel/task.go

@@ -735,7 +735,6 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock {
 func (t *Task) IsChrooted() bool {
 	if VFS2Enabled {
 		realRoot := t.mountNamespaceVFS2.Root()
-		defer realRoot.DecRef(t)
 		root := t.fsContext.RootDirectoryVFS2()
 		defer root.DecRef(t)
 		return root != realRoot
@@ -868,7 +867,6 @@ func (t *Task) MountNamespace() *fs.MountNamespace {
 func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace {
 	t.mu.Lock()
 	defer t.mu.Unlock()
-	t.mountNamespaceVFS2.IncRef()
 	return t.mountNamespaceVFS2
 }
 

pkg/sentry/syscalls/linux/vfs2/execve.go

@@ -110,8 +110,7 @@ func execveat(t *kernel.Task, dirfd int32, pathnameAddr, argvAddr, envvAddr user
 	}
 
 	// Load the new TaskContext.
-	mntns := t.MountNamespaceVFS2() // FIXME(jamieliu): useless refcount change
-	defer mntns.DecRef(t)
+	mntns := t.MountNamespaceVFS2()
 	wd := t.FSContext().WorkingDirectoryVFS2()
 	defer wd.DecRef(t)
 	remainingTraversals := uint(linux.MaxSymlinkTraversals)

pkg/sentry/vfs/mount.go

@@ -727,14 +727,12 @@ func (mnt *Mount) Root() *Dentry {
 	return mnt.root
 }
 
-// Root returns mntns' root. A reference is taken on the returned
-// VirtualDentry.
+// Root returns mntns' root. It does not take a reference on the returned Dentry.
 func (mntns *MountNamespace) Root() VirtualDentry {
 	vd := VirtualDentry{
 		mount:  mntns.root,
 		dentry: mntns.root.root,
 	}
-	vd.IncRef()
 	return vd
 }
 

runsc/boot/loader.go

@@ -903,7 +903,7 @@ func (l *Loader) executeAsync(args *control.ExecArgs) (kernel.ThreadID, error) {
 	// Get the container MountNamespace from the Task. Try to acquire ref may fail
 	// in case it raced with task exit.
 	if kernel.VFS2Enabled {
-		// task.MountNamespace() does not take a ref, so we must do so ourselves.
+		// task.MountNamespaceVFS2() does not take a ref, so we must do so ourselves.
 		args.MountNamespaceVFS2 = tg.Leader().MountNamespaceVFS2()
 		if !args.MountNamespaceVFS2.TryIncRef() {
 			return 0, fmt.Errorf("container %q has stopped", args.ContainerID)
@@ -925,7 +925,6 @@ func (l *Loader) executeAsync(args *control.ExecArgs) (kernel.ThreadID, error) {
 		root := args.MountNamespaceVFS2.Root()
 		ctx := vfs.WithRoot(l.k.SupervisorContext(), root)
 		defer args.MountNamespaceVFS2.DecRef(ctx)
-		defer root.DecRef(ctx)
 		envv, err := user.MaybeAddExecUserHomeVFS2(ctx, args.MountNamespaceVFS2, args.KUID, args.Envv)
 		if err != nil {
 			return 0, err

runsc/boot/loader_test.go

@@ -491,6 +491,7 @@ func TestCreateMountNamespaceVFS2(t *testing.T) {
 			}
 
 			root := mns.Root()
+			root.IncRef()
 			defer root.DecRef(ctx)
 			for _, p := range tc.expectedPaths {
 				target := &vfs.PathOperation{