Fix runsc association of gofers and FDs on VFS2.

Updates #1487

PiperOrigin-RevId: 311443628

2 files changed, 95 insertions, 17 deletions

diff --git a/runsc/boot/fs.go b/runsc/boot/fs.go
index 8df5cc989..e1181271a 100644
--- a/runsc/boot/fs.go
+++ b/runsc/boot/fs.go
@@ -770,14 +770,8 @@ func (c *containerMounter) getMountNameAndOptions(conf *Config, m specs.Mount) (
 		useOverlay bool
 	)
 
-	for _, opt := range m.Options {
-		// When options include either "bind" or "rbind", this behaves as
-		// bind mount even if the mount type is equal to a filesystem supported
-		// on runsc.
-		if opt == "bind" || opt == "rbind" {
-			m.Type = bind
-			break
-		}
+	if isBindMount(m) {
+		m.Type = bind
 	}
 
 	switch m.Type {
@@ -807,6 +801,18 @@ func (c *containerMounter) getMountNameAndOptions(conf *Config, m specs.Mount) (
 	return fsName, opts, useOverlay, nil
 }
 
+func isBindMount(m specs.Mount) bool {
+	for _, opt := range m.Options {
+		// When options include either "bind" or "rbind", this behaves as
+		// bind mount even if the mount type is equal to a filesystem supported
+		// on runsc.
+		if opt == "bind" || opt == "rbind" {
+			return true
+		}
+	}
+	return false
+}
+
 func (c *containerMounter) getMountAccessType(mount specs.Mount) FileAccessType {
 	if hint := c.hints.findMount(mount); hint != nil {
 		return hint.fileAccessType()
diff --git a/runsc/boot/vfs.go b/runsc/boot/vfs.go
index 7378fbc95..147c901c4 100644
--- a/runsc/boot/vfs.go
+++ b/runsc/boot/vfs.go
@@ -203,28 +203,61 @@ func (c *containerMounter) createMountNamespaceVFS2(ctx context.Context, conf *C
 }
 
 func (c *containerMounter) mountSubmountsVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials) error {
-	c.prepareMountsVFS2()
+	mounts, err := c.prepareMountsVFS2()
+	if err != nil {
+		return err
+	}
 
-	for _, submount := range c.mounts {
+	for i := range mounts {
+		submount := &mounts[i]
 		log.Debugf("Mounting %q to %q, type: %s, options: %s", submount.Source, submount.Destination, submount.Type, submount.Options)
-		if err := c.mountSubmountVFS2(ctx, conf, mns, creds, &submount); err != nil {
+		if err := c.mountSubmountVFS2(ctx, conf, mns, creds, submount); err != nil {
 			return err
 		}
 	}
 
 	// TODO(gvisor.dev/issue/1487): implement mountTmp from fs.go.
 
-	return c.checkDispenser()
+	return nil
+}
+
+type mountAndFD struct {
+	specs.Mount
+	fd int
 }
 
-func (c *containerMounter) prepareMountsVFS2() {
+func (c *containerMounter) prepareMountsVFS2() ([]mountAndFD, error) {
+	// Associate bind mounts with their FDs before sorting since there is an
+	// undocumented assumption that FDs are dispensed in the order in which
+	// they are required by mounts.
+	var mounts []mountAndFD
+	for _, m := range c.mounts {
+		fd := -1
+		// Only bind mounts use host FDs; see
+		// containerMounter.getMountNameAndOptionsVFS2.
+		if m.Type == bind || isBindMount(m) {
+			fd = c.fds.remove()
+		}
+		mounts = append(mounts, mountAndFD{
+			Mount: m,
+			fd:    fd,
+		})
+	}
+	if err := c.checkDispenser(); err != nil {
+		return nil, err
+	}
+
 	// Sort the mounts so that we don't place children before parents.
-	sort.Slice(c.mounts, func(i, j int) bool { return len(c.mounts[i].Destination) < len(c.mounts[j].Destination) })
+	sort.Slice(mounts, func(i, j int) bool {
+		return len(mounts[i].Destination) < len(mounts[j].Destination)
+	})
+
+	return mounts, nil
 }
 
 // TODO(gvisor.dev/issue/1487): Implement submount options similar to the VFS1
-//  version.
-func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials, submount *specs.Mount) error {
+// version.
+func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials, submount *mountAndFD) error {
 	root := mns.Root()
 	defer root.DecRef()
 	target := &vfs.PathOperation{
@@ -233,7 +266,7 @@ func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config,
 		Path:  fspath.Parse(submount.Destination),
 	}
 
-	fsName, options, useOverlay, err := c.getMountNameAndOptions(conf, *submount)
+	fsName, options, useOverlay, err := c.getMountNameAndOptionsVFS2(conf, submount)
 	if err != nil {
 		return fmt.Errorf("mountOptions failed: %w", err)
 	}
@@ -263,6 +296,45 @@ func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config,
 	return nil
 }
 
+// getMountNameAndOptionsVFS2 retrieves the fsName, opts, and useOverlay values
+// used for mounts.
+func (c *containerMounter) getMountNameAndOptionsVFS2(conf *Config, m *mountAndFD) (string, []string, bool, error) {
+	var (
+		fsName     string
+		opts       []string
+		useOverlay bool
+	)
+
+	if isBindMount(m.Mount) {
+		m.Type = bind
+	}
+
+	switch m.Type {
+	case devpts.Name, devtmpfs.Name, proc.Name, sys.Name:
+		fsName = m.Type
+	case nonefs:
+		fsName = sys.Name
+	case tmpfs.Name:
+		fsName = m.Type
+
+		var err error
+		opts, err = parseAndFilterOptions(m.Options, tmpfsAllowedOptions...)
+		if err != nil {
+			return "", nil, false, err
+		}
+
+	case bind:
+		fsName = gofer.Name
+		opts = p9MountOptions(m.fd, c.getMountAccessType(m.Mount), true /* vfs2 */)
+		// If configured, add overlay to all writable mounts.
+		useOverlay = conf.Overlay && !mountFlags(m.Options).ReadOnly
+
+	default:
+		log.Warningf("ignoring unknown filesystem type %q", m.Type)
+	}
+	return fsName, opts, useOverlay, nil
+}
+
 func (c *containerMounter) makeSyntheticMount(ctx context.Context, currentPath string, root vfs.VirtualDentry, creds *auth.Credentials) error {
 	target := &vfs.PathOperation{
 		Root:  root,