5 files changed, 42 insertions, 1 deletions
diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
index 5209a17af..3cc63e732 100644
--- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
+++ b/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,
diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go
index be29a2363..2f856ce36 100644
--- a/pkg/sentry/fsimpl/tmpfs/pipe_test.go
+++ b/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,
diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go
index a199eb33d..ce4e3eda7 100644
--- a/pkg/sentry/fsimpl/tmpfs/regular_file.go
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go
@@ -293,7 +293,7 @@ func (rf *regularFile) Translate(ctx context.Context, required, optional memmap.
 		optional.End = pgend
 	}
 
-	cerr := rf.data.Fill(ctx, required, optional, rf.memFile, rf.memoryUsageKind, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) {
+	cerr := rf.data.Fill(ctx, required, optional, rf.size, rf.memFile, rf.memoryUsageKind, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) {
 		// Newly-allocated pages are zeroed, so we don't need to do anything.
 		return dsts.NumBytes(), nil
 	})
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
index cefec8fde..e2a0aac69 100644
--- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
@@ -74,6 +74,8 @@ type filesystem struct {
 	mu sync.RWMutex `state:"nosave"`
 
 	nextInoMinusOne uint64 // accessed using atomic memory operations
+
+	root *dentry
 }
 
 // Name implements vfs.FilesystemType.Name.
@@ -81,6 +83,9 @@ func (FilesystemType) Name() string {
 	return Name
 }
 
+// Release implements vfs.FilesystemType.Release.
+func (FilesystemType) Release(ctx context.Context) {}
+
 // FilesystemOpts is used to pass configuration data to tmpfs.
 //
 // +stateify savable
@@ -194,6 +199,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 		fs.vfsfs.DecRef(ctx)
 		return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType)
 	}
+	fs.root = root
 	return &fs.vfsfs, &root.vfsd, nil
 }
 
@@ -205,6 +211,37 @@ func NewFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *au
 // Release implements vfs.FilesystemImpl.Release.
 func (fs *filesystem) Release(ctx context.Context) {
 	fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor)
+	fs.mu.Lock()
+	if fs.root.inode.isDir() {
+		fs.root.releaseChildrenLocked(ctx)
+	}
+	fs.mu.Unlock()
+}
+
+// releaseChildrenLocked is called on the mount point by filesystem.Release() to
+// destroy all objects in the mount. It performs a depth-first walk of the
+// filesystem and "unlinks" everything by decrementing link counts
+// appropriately. There should be no open file descriptors when this is called,
+// so each inode should only have one outstanding reference that is removed once
+// its link count hits zero.
+//
+// Note that we do not update filesystem state precisely while tearing down (for
+// instance, the child maps are ignored)--we only care to remove all remaining
+// references so that every filesystem object gets destroyed. Also note that we
+// do not need to trigger DecRef on the mount point itself or any child mount;
+// these are taken care of by the destructor of the enclosing MountNamespace.
+//
+// Precondition: filesystem.mu is held.
+func (d *dentry) releaseChildrenLocked(ctx context.Context) {
+	dir := d.inode.impl.(*directory)
+	for _, child := range dir.childMap {
+		if child.inode.isDir() {
+			child.releaseChildrenLocked(ctx)
+			child.inode.decLinksLocked(ctx) // link for child/.
+			dir.inode.decLinksLocked(ctx)   // link for child/..
+		}
+		child.inode.decLinksLocked(ctx) // link for child
+	}
 }
 
 // immutable
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go
index 99c8e3c0f..fc5323abc 100644
--- a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go
+++ b/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)