netwatch/internal/reportbuf/datadir.go

package reportbuf

import (
	"errors"
	"io/fs"
	"os"
	"os/user"
	"path/filepath"
	"strconv"
	"syscall"
)

// ErrDataDirOutsideVolume is returned by PrepareDataDir for a DATA_DIR
// that is not the volume or a path below it, written in full.
var ErrDataDirOutsideVolume = errors.New(
	"must be /data or a path below it, with no '.', '..' or extra '/'")

// PrepareDataDir gets dir, the server's DATA_DIR, ready for owner, the
// user the server runs as, so that a host directory mounted at volume,
// /data in the image, needs no preparing: it creates dir, gives volume
// and the directories, regular files and symbolic links in it to
// owner, and gives volume and dir the mode the server gives a directory
// it creates. It leaves a file with a second name (hard link) as it
// is, as that name could be outside volume, and logs its path; it
// leaves a device, named pipe or socket as it is too (see chownFile).
// dir must be volume or a path below it, with no '.', '..', empty part
// or '/' at the end. On any system but Linux, the system of the image, it
// fails at the first file in volume that is not a directory (see
// chownFile).
//
// bin/entrypoint.sh runs this as root, which would follow a symbolic
// link anywhere, so every step goes through an os.Root opened on
// volume: it follows a link only when it is written as a relative
// path that stays inside volume, and refuses any other. A process on
// the host can swap a link onto a path in volume at any moment while
// this runs. Even then, the os.Root checks each link as it reaches
// it. MkdirAll creates each directory inside a parent it already has
// open, never following a link at the name it creates, and follows a
// link on the path only as the os.Root allows, so a relative link
// inside volume can lead it to create directories elsewhere inside
// volume. The owners and modes are set on files already opened (see
// chownAll and chmodDir), so the most that process can do is make a
// step fail or wait, or act on something else inside volume.
func PrepareDataDir(volume, dir string, owner *user.User) error {
	// rel is dir as a path from volume; IsLocal is false for one that
	// leads out of it.
	rel, err := filepath.Rel(volume, dir)
	if err != nil || dir != filepath.Clean(dir) || !filepath.IsLocal(rel) {
		return ErrDataDirOutsideVolume
	}

	uid, err := strconv.Atoi(owner.Uid)
	if err != nil {
		return err
	}

	gid, err := strconv.Atoi(owner.Gid)
	if err != nil {
		return err
	}

	root, err := os.OpenRoot(volume)
	if err != nil {
		return err
	}

	defer func() { _ = root.Close() }()

	err = root.MkdirAll(rel, dirPerms)
	if err != nil {
		return err
	}

	err = chownAll(root, ".", uid, gid)
	if err != nil {
		return err
	}

	err = chmodDir(root, ".")
	if err != nil {
		return err
	}

	return chmodDir(root, rel)
}

// chownAll gives name, a directory inside root, and the directories,
// regular files and symbolic links in it to uid and gid, but not the
// files chownFile leaves. It opens each through root and gives the
// open file to them, as chmodDir sets a mode on the open directory.
// Root.Lchown would not hold: it acts on whatever is at the name by
// then, and a hard link swapped in there shares its owner with a file
// that could be outside root. A symbolic link is given to them itself,
// never what it points to, because in a directory that anyone can
// write to and that has the sticky bit set, as /tmp has, the kernel
// follows a link only for the link's owner or when the link and the
// directory have the same owner. chownAll goes into an entry only when
// the read found a directory there, so it follows no link it finds;
// one swapped in afterwards for a directory is followed only as the
// os.Root allows, and one swapped in for any other file is not
// followed at all (see chownFile).
func chownAll(root *os.Root, name string, uid, gid int) error {
	entries, err := chownDir(root, name, uid, gid)
	if err != nil {
		return err
	}

	for _, entry := range entries {
		entryName := filepath.Join(name, entry.Name())
		if entry.IsDir() {
			err = chownAll(root, entryName, uid, gid)
		} else {
			err = chownFile(root, entryName, uid, gid)
		}

		if err != nil {
			return err
		}
	}

	return nil
}

// chownDir gives name, a directory inside root, to uid and gid, and
// returns what it holds. Like chmodDir, it opens name through root and
// refuses anything but a directory, which has no second name (hard
// link). It reads the open directory, not root.FS(), which refuses a
// name that is not valid UTF-8.
func chownDir(root *os.Root, name string, uid, gid int) ([]fs.DirEntry, error) {
	dir, err := root.Open(name)
	if err != nil {
		return nil, err
	}

	defer func() { _ = dir.Close() }()

	info, err := dir.Stat()
	if err != nil {
		return nil, err
	}

	if !info.IsDir() {
		return nil, &fs.PathError{Op: "chown", Path: name, Err: syscall.ENOTDIR}
	}

	err = dir.Chown(uid, gid)
	if err != nil {
		return nil, err
	}

	return dir.ReadDir(-1)
}

// chmodDir gives name, a directory inside root, the mode the server
// gives a directory it creates. Root.Chmod would not hold: on Linux it
// checks that name is not a symbolic link, then sets the mode by name,
// following a link swapped in between. So chmodDir opens name through
// root and sets the mode on the open directory. It refuses anything
// but a directory: a directory has no second name (hard link), so the
// one opened is inside root, where any other file could be a hard link
// to one outside.
func chmodDir(root *os.Root, name string) error {
	dir, err := root.Open(name)
	if err != nil {
		return err
	}

	defer func() { _ = dir.Close() }()

	info, err := dir.Stat()
	if err != nil {
		return err
	}

	if !info.IsDir() {
		return &fs.PathError{Op: "chmod", Path: name, Err: syscall.ENOTDIR}
	}

	return dir.Chmod(dirPerms)
}