ext: wrap file-descriptor-based mount facilities

This only implements what is required by package container for now.

Signed-off-by: Ophestra <cat@gensokyo.uk>

1 files changed, 267 insertions, 0 deletions

diff --git a/ext/fs.go b/ext/fs.go
new file mode 100644
index 00000000..f2bce87c
--- /dev/null
+++ b/ext/fs.go
@@ -0,0 +1,267 @@
+package ext
+
+import (
+	"os"
+	"runtime"
+	"syscall"
+	"unsafe"
+)
+
+// include/uapi/linux/mount.h
+
+/*
+ * move_mount() flags.
+ */
+const (
+	MOVE_MOUNT_F_SYMLINKS   = 1 << iota /* Follow symlinks on from path */
+	MOVE_MOUNT_F_AUTOMOUNTS             /* Follow automounts on from path */
+	MOVE_MOUNT_F_EMPTY_PATH             /* Empty from path permitted */
+	_
+	MOVE_MOUNT_T_SYMLINKS   /* Follow symlinks on to path */
+	MOVE_MOUNT_T_AUTOMOUNTS /* Follow automounts on to path */
+	MOVE_MOUNT_T_EMPTY_PATH /* Empty to path permitted */
+	_
+	MOVE_MOUNT_SET_GROUP /* Set sharing group instead */
+	MOVE_MOUNT_BENEATH   /* Mount beneath top mount */
+)
+
+/*
+ * fsopen() flags.
+ */
+const (
+	FSOPEN_CLOEXEC = 1 << iota
+)
+
+/*
+ * fspick() flags.
+ */
+const (
+	FSPICK_CLOEXEC = 1 << iota
+	FSPICK_SYMLINK_NOFOLLOW
+	FSPICK_NO_AUTOMOUNT
+	FSPICK_EMPTY_PATH
+)
+
+/*
+ * The type of fsconfig() call made.
+ */
+const (
+	FSCONFIG_SET_FLAG        = iota /* Set parameter, supplying no value */
+	FSCONFIG_SET_STRING             /* Set parameter, supplying a string value */
+	FSCONFIG_SET_BINARY             /* Set parameter, supplying a binary blob value */
+	FSCONFIG_SET_PATH               /* Set parameter, supplying an object by path */
+	FSCONFIG_SET_PATH_EMPTY         /* Set parameter, supplying an object by (empty) path */
+	FSCONFIG_SET_FD                 /* Set parameter, supplying an object by fd */
+	FSCONFIG_CMD_CREATE             /* Create new or reuse existing superblock */
+	FSCONFIG_CMD_RECONFIGURE        /* Invoke superblock reconfiguration */
+	FSCONFIG_CMD_CREATE_EXCL        /* Create new superblock, fail if reusing existing superblock */
+)
+
+/*
+ * fsmount() flags.
+ */
+const (
+	FSMOUNT_CLOEXEC = 1 << iota
+)
+
+/*
+ * Mount attributes.
+ */
+const (
+	MOUNT_ATTR_RDONLY      = 0x00000001 /* Mount read-only */
+	MOUNT_ATTR_NOSUID      = 0x00000002 /* Ignore suid and sgid bits */
+	MOUNT_ATTR_NODEV       = 0x00000004 /* Disallow access to device special files */
+	MOUNT_ATTR_NOEXEC      = 0x00000008 /* Disallow program execution */
+	MOUNT_ATTR__ATIME      = 0x00000070 /* Setting on how atime should be updated */
+	MOUNT_ATTR_RELATIME    = 0x00000000 /* - Update atime relative to mtime/ctime. */
+	MOUNT_ATTR_NOATIME     = 0x00000010 /* - Do not update access times. */
+	MOUNT_ATTR_STRICTATIME = 0x00000020 /* - Always perform atime updates */
+	MOUNT_ATTR_NODIRATIME  = 0x00000080 /* Do not update directory access times */
+	MOUNT_ATTR_IDMAP       = 0x00100000 /* Idmap mount to @userns_fd in struct mount_attr. */
+	MOUNT_ATTR_NOSYMFOLLOW = 0x00200000 /* Do not follow symlinks */
+)
+
+// FS provides low-level wrappers around the suite of file-descriptor-based
+// mount facilities in Linux.
+type FS struct {
+	fd uintptr
+	c  runtime.Cleanup
+}
+
+// newFS allocates a new [FS] for the specified fd.
+func newFS(fd uintptr) *FS {
+	fs := FS{fd: fd}
+	fs.c = runtime.AddCleanup(&fs, func(fd uintptr) {
+		_ = syscall.Close(int(fd))
+	}, fd)
+	return &fs
+}
+
+// Close closes the underlying filesystem context.
+func (fs *FS) Close() error {
+	if fs == nil {
+		return syscall.EINVAL
+	}
+	err := syscall.Close(int(fs.fd))
+	fs.c.Stop()
+	return err
+}
+
+// OpenFS creates a new filesystem context.
+func OpenFS(fsname string, flags int) (fs *FS, err error) {
+	var s *byte
+	s, err = syscall.BytePtrFromString(fsname)
+	if err != nil {
+		return
+	}
+	fd, _, errno := syscall.Syscall(
+		SYS_FSOPEN,
+		uintptr(unsafe.Pointer(s)),
+		uintptr(flags|FSOPEN_CLOEXEC),
+		0,
+	)
+	if errno != 0 {
+		err = os.NewSyscallError("fsopen", errno)
+	} else {
+		fs = newFS(fd)
+	}
+	return
+}
+
+// PickFS selects filesystem for reconfiguration.
+func PickFS(dirfd int, pathname string, flags int) (fs *FS, err error) {
+	var s *byte
+	s, err = syscall.BytePtrFromString(pathname)
+	if err != nil {
+		return
+	}
+	fd, _, errno := syscall.Syscall(
+		SYS_FSPICK,
+		uintptr(dirfd),
+		uintptr(unsafe.Pointer(s)),
+		uintptr(flags|FSPICK_CLOEXEC),
+	)
+	if errno != 0 {
+		err = os.NewSyscallError("fspick", errno)
+	} else {
+		fs = newFS(fd)
+	}
+	return
+}
+
+// config configures new or existing filesystem context.
+func (fs *FS) config(cmd uint, key *byte, value unsafe.Pointer, aux int) (err error) {
+	_, _, errno := syscall.Syscall6(
+		SYS_FSCONFIG,
+		fs.fd,
+		uintptr(cmd),
+		uintptr(unsafe.Pointer(key)),
+		uintptr(value),
+		uintptr(aux),
+		0,
+	)
+	if errno != 0 {
+		err = os.NewSyscallError("fsconfig", errno)
+	}
+	return
+}
+
+// SetFlag sets the flag parameter named by key. ([FSCONFIG_SET_FLAG])
+func (fs *FS) SetFlag(key string) (err error) {
+	var s *byte
+	s, err = syscall.BytePtrFromString(key)
+	if err != nil {
+		return
+	}
+
+	return fs.config(FSCONFIG_SET_FLAG, s, nil, 0)
+}
+
+// SetString sets the string parameter named by key to the value specified by
+// value. ([FSCONFIG_SET_STRING])
+func (fs *FS) SetString(key, value string) (err error) {
+	var s0 *byte
+	s0, err = syscall.BytePtrFromString(key)
+	if err != nil {
+		return
+	}
+
+	var s1 *byte
+	s1, err = syscall.BytePtrFromString(value)
+	if err != nil {
+		return
+	}
+
+	return fs.config(FSCONFIG_SET_STRING, s0, unsafe.Pointer(s1), 0)
+}
+
+// mount instantiates mount object from filesystem context.
+func (fs *FS) mount(flags, attrFlags int) (fsfd int, err error) {
+	r, _, errno := syscall.Syscall(
+		SYS_FSMOUNT,
+		fs.fd,
+		uintptr(flags|FSMOUNT_CLOEXEC),
+		uintptr(attrFlags),
+	)
+	fsfd = int(r)
+	if errno != 0 {
+		err = os.NewSyscallError("fsmount", errno)
+	}
+	return
+}
+
+// MoveMount moves or attaches mount object to filesystem.
+func MoveMount(
+	fromDirfd int,
+	fromPathname string,
+	toDirfd int,
+	toPathname string,
+	flags int,
+) (err error) {
+	var s0 *byte
+	s0, err = syscall.BytePtrFromString(fromPathname)
+	if err != nil {
+		return
+	}
+
+	var s1 *byte
+	s1, err = syscall.BytePtrFromString(toPathname)
+	if err != nil {
+		return
+	}
+
+	_, _, errno := syscall.Syscall6(
+		SYS_MOVE_MOUNT,
+		uintptr(fromDirfd),
+		uintptr(unsafe.Pointer(s0)),
+		uintptr(toDirfd),
+		uintptr(unsafe.Pointer(s1)),
+		uintptr(flags),
+		0,
+	)
+	if errno != 0 {
+		err = os.NewSyscallError("move_mount", errno)
+	}
+	return
+}
+
+// Mount attaches the underlying filesystem context to the specified pathname.
+func (fs *FS) Mount(pathname string, attrFlags int) error {
+	if err := fs.config(FSCONFIG_CMD_CREATE_EXCL, nil, nil, 0); err != nil {
+		return err
+	}
+	fd, err := fs.mount(0, attrFlags)
+	if err != nil {
+		return err
+	}
+	err = MoveMount(
+		fd, "",
+		-1, pathname,
+		MOVE_MOUNT_F_EMPTY_PATH,
+	)
+	closeErr := syscall.Close(fd)
+	if err == nil {
+		err = closeErr
+	}
+	return err
+}