Pre-open container root directory

A lot of filesystem-related stuff happens inside the container root
directory, and we have used its name before. It makes sense to pre-open
it and use a *os.File handle instead.

Function names in internal/pathrs are kept as is for simplicity (and it
is an internal package), but they now accept root as *os.File.

Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>

(cherry picked from commit 28cb321887)

It was also necessary to cherry-pick a missing hunk from commit
b88635e57e ("libct: close rootFd ASAP in maskPaths") that was not
backported to 1.3 in the backported patch commit 9a2fec0a9a ("libct:
close rootFd ASAP in maskPaths").

Signed-off-by: Aleksa Sarai <aleksa@amutable.com>

internal/pathrs/mkdirall.go

@@ -31,12 +31,12 @@ import (
 // Callers need to be very careful operating on the trailing path, as trivial
 // mistakes like following symlinks can cause security bugs. Most people
 // should probably just use [MkdirAllInRoot] or [CreateInRoot].
-func MkdirAllParentInRoot(root, unsafePath string, mode os.FileMode) (*os.File, string, error) {
+func MkdirAllParentInRoot(root *os.File, unsafePath string, mode os.FileMode) (*os.File, string, error) {
 	// MkdirAllInRoot also does hallucinateUnsafePath, but we need to do it
 	// here first because when we split unsafePath into (dir, file) components
 	// we want to be doing so with the hallucinated path (so that trailing
 	// dangling symlinks are treated correctly).
-	unsafePath, err := hallucinateUnsafePath(root, unsafePath)
+	unsafePath, err := hallucinateUnsafePath(root.Name(), unsafePath)
 	if err != nil {
 		return nil, "", fmt.Errorf("failed to construct hallucinated target path: %w", err)
 	}

internal/pathrs/mkdirall_pathrslite.go

@@ -24,12 +24,11 @@ import (
 
 	"github.com/cyphar/filepath-securejoin/pathrs-lite"
 	"github.com/sirupsen/logrus"
-	"golang.org/x/sys/unix"
 )
 
 // MkdirAllInRoot attempts to make
 //
-//	path, _ := securejoin.SecureJoin(root, unsafePath)
+//	path, _ := securejoin.SecureJoin(root.Name(), unsafePath)
 //	os.MkdirAll(path, mode)
 //	os.Open(path)
 //
@@ -48,8 +47,8 @@ import (
 // handling if unsafePath has already been scoped within the rootfs (this is
 // needed for a lot of runc callers and fixing this would require reworking a
 // lot of path logic).
-func MkdirAllInRoot(root, unsafePath string, mode os.FileMode) (*os.File, error) {
-	unsafePath, err := hallucinateUnsafePath(root, unsafePath)
+func MkdirAllInRoot(root *os.File, unsafePath string, mode os.FileMode) (*os.File, error) {
+	unsafePath, err := hallucinateUnsafePath(root.Name(), unsafePath)
 	if err != nil {
 		return nil, fmt.Errorf("failed to construct hallucinated target path: %w", err)
 	}
@@ -67,13 +66,7 @@ func MkdirAllInRoot(root, unsafePath string, mode os.FileMode) (*os.File, error)
 		mode &= 0o1777
 	}
 
-	rootDir, err := os.OpenFile(root, unix.O_DIRECTORY|unix.O_CLOEXEC, 0)
-	if err != nil {
-		return nil, fmt.Errorf("open root handle: %w", err)
-	}
-	defer rootDir.Close()
-
 	return retryEAGAIN(func() (*os.File, error) {
-		return pathrs.MkdirAllHandle(rootDir, unsafePath, mode)
+		return pathrs.MkdirAllHandle(root, unsafePath, mode)
 	})
 }

internal/pathrs/root_pathrslite.go

@@ -26,11 +26,11 @@ import (
 )
 
 // OpenInRoot opens the given path inside the root with the provided flags. It
-// is effectively shorthand for [securejoin.OpenInRoot] followed by
+// is effectively shorthand for [securejoin.OpenatInRoot] followed by
 // [securejoin.Reopen].
-func OpenInRoot(root, subpath string, flags int) (*os.File, error) {
+func OpenInRoot(root *os.File, subpath string, flags int) (*os.File, error) {
 	handle, err := retryEAGAIN(func() (*os.File, error) {
-		return pathrs.OpenInRoot(root, subpath)
+		return pathrs.OpenatInRoot(root, subpath)
 	})
 	if err != nil {
 		return nil, err
@@ -45,7 +45,7 @@ func OpenInRoot(root, subpath string, flags int) (*os.File, error) {
 // open(O_CREAT|O_NOFOLLOW) semantics. If you want the creation to use O_EXCL,
 // include it in the passed flags. The fileMode argument uses unix.* mode bits,
 // *not* os.FileMode.
-func CreateInRoot(root, subpath string, flags int, fileMode uint32) (*os.File, error) {
+func CreateInRoot(root *os.File, subpath string, flags int, fileMode uint32) (*os.File, error) {
 	dirFd, filename, err := MkdirAllParentInRoot(root, subpath, 0o755)
 	if err != nil {
 		return nil, err
@@ -61,5 +61,5 @@ func CreateInRoot(root, subpath string, flags int, fileMode uint32) (*os.File, e
 	if err != nil {
 		return nil, err
 	}
-	return os.NewFile(uintptr(fd), root+"/"+subpath), nil
+	return os.NewFile(uintptr(fd), root.Name()+"/"+subpath), nil
 }

libcontainer/criu_linux.go

@@ -540,10 +540,16 @@ func isOnTmpfs(path string, mounts []*configs.Mount) bool {
 // This function also creates missing mountpoints as long as they
 // are not on top of a tmpfs, as CRIU will restore tmpfs content anyway.
 func (c *Container) prepareCriuRestoreMounts(mounts []*configs.Mount) error {
+	rootFd, err := os.OpenFile(c.config.Rootfs, unix.O_DIRECTORY|unix.O_CLOEXEC|unix.O_PATH, 0)
+	if err != nil {
+		return fmt.Errorf("open rootfs handle: %w", err)
+	}
+	defer rootFd.Close()
+
 	umounts := []string{}
 	defer func() {
 		for _, u := range umounts {
-			mntFile, err := pathrs.OpenInRoot(c.config.Rootfs, u, unix.O_PATH)
+			mntFile, err := pathrs.OpenInRoot(rootFd, u, unix.O_PATH)
 			if err != nil {
 				logrus.Warnf("Error during cleanup unmounting %s: open handle: %v", u, err)
 				continue
@@ -577,7 +583,7 @@ func (c *Container) prepareCriuRestoreMounts(mounts []*configs.Mount) error {
 			continue
 		}
 		me := mountEntry{Mount: m}
-		if err := me.createOpenMountpoint(c.config.Rootfs); err != nil {
+		if err := me.createOpenMountpoint(rootFd); err != nil {
 			return fmt.Errorf("create criu restore mountpoint for %s mount: %w", me.Destination, err)
 		}
 		if me.dstFile != nil {

libcontainer/rootfs_linux.go

@@ -34,7 +34,7 @@ const defaultMountFlags = unix.MS_NOEXEC | unix.MS_NOSUID | unix.MS_NODEV
 
 // mountConfig contains mount data not specific to a mount point.
 type mountConfig struct {
-	root            string
+	root            *os.File
 	label           string
 	cgroup2Path     string
 	rootlessCgroups bool
@@ -105,8 +105,17 @@ func prepareRootfs(pipe *syncSocket, iConfig *initConfig) (err error) {
 		return fmt.Errorf("error preparing rootfs: %w", err)
 	}
 
+	// Pre-open rootfs. NOTE that if we need to re-enable support for mounting
+	// on top of container root (see issue 5070), we will need to reopen rootFd
+	// after such mounts.
+	rootFd, err := os.OpenFile(config.Rootfs, unix.O_DIRECTORY|unix.O_CLOEXEC|unix.O_PATH, 0)
+	if err != nil {
+		return fmt.Errorf("open rootfs handle: %w", err)
+	}
+	defer rootFd.Close()
+
 	mountConfig := &mountConfig{
-		root:            config.Rootfs,
+		root:            rootFd,
 		label:           config.MountLabel,
 		cgroup2Path:     iConfig.Cgroup2Path,
 		rootlessCgroups: config.RootlessCgroups,
@@ -166,7 +175,7 @@ func prepareRootfs(pipe *syncSocket, iConfig *initConfig) (err error) {
 
 	setupDev := needsSetupDev(config)
 	if setupDev {
-		if err := createDevices(config); err != nil {
+		if err := createDevices(config, rootFd); err != nil {
 			return fmt.Errorf("error creating device nodes: %w", err)
 		}
 		if err := setupPtmx(config); err != nil {
@@ -216,6 +225,7 @@ func prepareRootfs(pipe *syncSocket, iConfig *initConfig) (err error) {
 	if err != nil {
 		return fmt.Errorf("error jailing process inside rootfs: %w", err)
 	}
+	rootFd.Close()
 
 	// Apply root mount propagation flags.
 	// This must be done after pivot_root/chroot because the mount propagation flag is applied
@@ -361,7 +371,7 @@ func mountCgroupV1(m mountEntry, c *mountConfig) error {
 			// symlink(2) is very dumb, it will just shove the path into
 			// the link and doesn't do any checks or relative path
 			// conversion. Also, don't error out if the cgroup already exists.
-			if err := os.Symlink(mc, filepath.Join(c.root, m.Destination, ss)); err != nil && !os.IsExist(err) {
+			if err := os.Symlink(mc, filepath.Join(c.root.Name(), m.Destination, ss)); err != nil && !errors.Is(err, os.ErrExist) {
 				return err
 			}
 		}
@@ -425,15 +435,22 @@ func doTmpfsCopyUp(m mountEntry, mountLabel string) (Err error) {
 	}
 	defer tmpDirFile.Close()
 
+	hostRootFd, err := os.OpenFile("/", unix.O_DIRECTORY|unix.O_CLOEXEC|unix.O_PATH, 0)
+	if err != nil {
+		return fmt.Errorf("tmpcopyup: open host root: %w", err)
+	}
+	defer hostRootFd.Close()
+
 	// Configure the *host* tmpdir as if it's the container mount. We change
 	// m.dstFile since we are going to mount *on the host*.
 	hostMount := mountEntry{
 		Mount:   m.Mount,
 		dstFile: tmpDirFile,
 	}
-	if err := hostMount.mountPropagate("/", mountLabel); err != nil {
+	if err := hostMount.mountPropagate(hostRootFd, mountLabel); err != nil {
 		return err
 	}
+	hostRootFd.Close()
 	defer func() {
 		if Err != nil {
 			if err := unmount(tmpDir, unix.MNT_DETACH); err != nil {
@@ -502,9 +519,10 @@ func statfsToMountFlags(st unix.Statfs_t) int {
 	return flags
 }
 
-func (m *mountEntry) createOpenMountpoint(rootfs string) (Err error) {
+func (m *mountEntry) createOpenMountpoint(root *os.File) (Err error) {
+	rootfs := root.Name()
 	unsafePath := pathrs.LexicallyStripRoot(rootfs, m.Destination)
-	dstFile, err := pathrs.OpenInRoot(rootfs, unsafePath, unix.O_PATH)
+	dstFile, err := pathrs.OpenInRoot(root, unsafePath, unix.O_PATH)
 	defer func() {
 		if dstFile != nil && Err != nil {
 			_ = dstFile.Close()
@@ -541,9 +559,9 @@ func (m *mountEntry) createOpenMountpoint(rootfs string) (Err error) {
 			dstIsFile = !fi.IsDir()
 		}
 		if dstIsFile {
-			dstFile, err = pathrs.CreateInRoot(rootfs, unsafePath, unix.O_CREAT|unix.O_EXCL|unix.O_NOFOLLOW, 0o644)
+			dstFile, err = pathrs.CreateInRoot(root, unsafePath, unix.O_CREAT|unix.O_EXCL|unix.O_NOFOLLOW, 0o644)
 		} else {
-			dstFile, err = pathrs.MkdirAllInRoot(rootfs, unsafePath, 0o755)
+			dstFile, err = pathrs.MkdirAllInRoot(root, unsafePath, 0o755)
 		}
 		if err != nil {
 			return fmt.Errorf("make mountpoint %q: %w", m.Destination, err)
@@ -589,7 +607,7 @@ func mountToRootfs(c *mountConfig, m mountEntry) error {
 		// has been a "fun" attack scenario in the past.
 		// TODO: This won't be necessary once we switch to libpathrs and we can
 		//       stop all of these symlink-exchange attacks.
-		rootfs := c.root
+		rootfs := c.root.Name()
 		dest := filepath.Clean(m.Destination)
 		if !pathrs.IsLexicallyInRoot(rootfs, dest) {
 			// Do not use securejoin as it resolves symlinks.
@@ -936,7 +954,7 @@ func reOpenDevNull() error {
 }
 
 // Create the device nodes in the container.
-func createDevices(config *configs.Config) error {
+func createDevices(config *configs.Config, rootFd *os.File) error {
 	useBindMount := userns.RunningInUserNS() || config.Namespaces.Contains(configs.NEWUSER)
 	for _, node := range config.Devices {
 
@@ -947,7 +965,7 @@ func createDevices(config *configs.Config) error {
 
 		// containers running in a user namespace are not allowed to mknod
 		// devices so we can just bind mount it from the host.
-		if err := createDeviceNode(config.Rootfs, node, useBindMount); err != nil {
+		if err := createDeviceNode(rootFd, node, useBindMount); err != nil {
 			return err
 		}
 	}
@@ -967,12 +985,12 @@ func bindMountDeviceNode(destDir *os.File, destName string, node *devices.Device
 }
 
 // Creates the device node in the rootfs of the container.
-func createDeviceNode(rootfs string, node *devices.Device, bind bool) error {
+func createDeviceNode(rootFd *os.File, node *devices.Device, bind bool) error {
 	if node.Path == "" {
 		// The node only exists for cgroup reasons, ignore it here.
 		return nil
 	}
-	destDir, destName, err := pathrs.MkdirAllParentInRoot(rootfs, node.Path, 0o755)
+	destDir, destName, err := pathrs.MkdirAllParentInRoot(rootFd, node.Path, 0o755)
 	if err != nil {
 		return fmt.Errorf("mkdir parent of device inode %q: %w", node.Path, err)
 	}
@@ -1384,7 +1402,12 @@ func maskPaths(rootFs string, paths []string, mountLabel string) error {
 					// resolves the underlying inode via procfs and re-opens it through
 					// rootFd, so the resulting fd is anchored to the real path inside the
 					// container rootfs even if path was a /proc/self/fd/N alias.
-					reopened, err := reopenAfterMount(rootFs, dstFh, unix.O_PATH|unix.O_CLOEXEC)
+					rootFd, err := os.OpenFile(rootFs, unix.O_DIRECTORY|unix.O_CLOEXEC|unix.O_PATH, 0)
+					if err != nil {
+						return fmt.Errorf("open rootfs handle for masked paths: %w", err)
+					}
+					reopened, err := reopenAfterMount(rootFd, dstFh, unix.O_PATH|unix.O_CLOEXEC)
+					rootFd.Close()
 					if err != nil {
 						return fmt.Errorf("can't reopen shared directory mask: %w", err)
 					}
@@ -1407,16 +1430,17 @@ func maskPaths(rootFs string, paths []string, mountLabel string) error {
 	return nil
 }
 
-func reopenAfterMount(rootfs string, f *os.File, flags int) (_ *os.File, Err error) {
+func reopenAfterMount(rootFd, f *os.File, flags int) (_ *os.File, Err error) {
 	fullPath, err := procfs.ProcSelfFdReadlink(f)
 	if err != nil {
 		return nil, fmt.Errorf("get full path: %w", err)
 	}
+	rootfs := rootFd.Name()
 	if !pathrs.IsLexicallyInRoot(rootfs, fullPath) {
 		return nil, fmt.Errorf("mountpoint %q is outside of rootfs %q", fullPath, rootfs)
 	}
 	unsafePath := pathrs.LexicallyStripRoot(rootfs, fullPath)
-	reopened, err := pathrs.OpenInRoot(rootfs, unsafePath, flags)
+	reopened, err := pathrs.OpenInRoot(rootFd, unsafePath, flags)
 	if err != nil {
 		return nil, fmt.Errorf("re-open mountpoint %q: %w", unsafePath, err)
 	}
@@ -1446,7 +1470,7 @@ func reopenAfterMount(rootfs string, f *os.File, flags int) (_ *os.File, Err err
 
 // Do the mount operation followed by additional mounts required to take care
 // of propagation flags. This will always be scoped inside the container rootfs.
-func (m *mountEntry) mountPropagate(rootfs string, mountLabel string) error {
+func (m *mountEntry) mountPropagate(rootFd *os.File, mountLabel string) error {
 	var (
 		data  = label.FormatMountLabel(m.Data, mountLabel)
 		flags = m.Flags
@@ -1472,7 +1496,7 @@ func (m *mountEntry) mountPropagate(rootfs string, mountLabel string) error {
 	//
 	// TODO: Use move_mount(2) on newer kernels so that this is no longer
 	// necessary on modern systems.
-	newDstFile, err := reopenAfterMount(rootfs, m.dstFile, unix.O_PATH)
+	newDstFile, err := reopenAfterMount(rootFd, m.dstFile, unix.O_PATH)
 	if err != nil {
 		return fmt.Errorf("reopen mountpoint after mount: %w", err)
 	}