Mirror/pmbootstrap
1
0
Fork 1
mirror of https://gitlab.postmarketos.org/postmarketOS/pmbootstrap.git synced 2025-07-23 04:25:10 +03:00
Code Issues Wiki Activity
pmbootstrap/pmb/init/sandbox.py
Casey Connolly 9dc294dbbf pmbootstrap: unshare PID namespaec 
Put ourselves in a new PID namespace so that daemons we might start in
the chroot like adbd will be killed on exit.

This simplifies "shutdown" since we no longer need to kill these
processes.

Signed-off-by: Casey Connolly <kcxt@postmarketos.org>
2025-06-13 21:47:42 +02:00

1146 lines
38 KiB
Python
Raw Blame History

# SPDX-License-Identifier: LGPL-2.1-or-later
# FIXME: this file is wayyy off lol
# ruff: noqa
"""
This is a standalone implementation of sandboxing which is used by mkosi. Note that this is
invoked many times while building the image and as a result, the performance of this script has a
substantial impact on the performance of mkosi itself. To keep the runtime of this script to a
minimum, please don't import any extra modules if it can be avoided.
"""
import ctypes
import os
import sys
import warnings # noqa: F401 (loaded lazily by os.execvp() which happens too late)
__version__ = "26~devel"
# The following constants are taken from the Linux kernel headers.
AT_EMPTY_PATH = 0x1000
AT_FDCWD = -100
AT_NO_AUTOMOUNT = 0x800
AT_RECURSIVE = 0x8000
AT_SYMLINK_NOFOLLOW = 0x100
BTRFS_SUPER_MAGIC = 0x9123683E
CAP_NET_ADMIN = 12
CAP_SYS_ADMIN = 21
CLONE_NEWIPC = 0x08000000
CLONE_NEWPID = 0x20000000
CLONE_NEWNET = 0x40000000
CLONE_NEWNS = 0x00020000
CLONE_NEWUSER = 0x10000000
EBADF = 9
UNSHARE_EPERM_MSGEPERM = 1
EPERM = 1
ENOENT = 2
ENOSYS = 38
F_DUPFD = 0
F_GETFD = 1
FS_IOC_GETFLAGS = 0x80086601
FS_IOC_SETFLAGS = 0x40086602
FS_NOCOW_FL = 0x00800000
LINUX_CAPABILITY_U32S_3 = 2
LINUX_CAPABILITY_VERSION_3 = 0x20080522
MNT_DETACH = 2
MOUNT_ATTR_RDONLY = 0x00000001
MOUNT_ATTR_NOSUID = 0x00000002
MOUNT_ATTR_NODEV = 0x00000004
MOUNT_ATTR_NOEXEC = 0x00000008
MOUNT_ATTR_SIZE_VER0 = 32
MOVE_MOUNT_F_EMPTY_PATH = 0x00000004
MS_BIND = 4096
MS_MOVE = 8192
MS_REC = 16384
MS_SHARED = 1 << 20
MS_SLAVE = 1 << 19
NR_mount_setattr = 442
NR_move_mount = 429
NR_open_tree = 428
OPEN_TREE_CLOEXEC = os.O_CLOEXEC
OPEN_TREE_CLONE = 1
OVERLAYFS_SUPER_MAGIC = 0x794C7630
PR_CAP_AMBIENT = 47
PR_CAP_AMBIENT_RAISE = 2
# These definitions are taken from the libseccomp headers
SCMP_ACT_ALLOW = 0x7FFF0000
SCMP_ACT_ERRNO = 0x00050000
SD_LISTEN_FDS_START = 3
SIGSTOP = 19
class mount_attr(ctypes.Structure):
_fields_ = [
("attr_set", ctypes.c_uint64),
("attr_clr", ctypes.c_uint64),
("propagation", ctypes.c_uint64),
("userns_fd", ctypes.c_uint64),
]
class cap_user_header_t(ctypes.Structure):
# __user_cap_header_struct
_fields_ = [
("version", ctypes.c_uint32),
("pid", ctypes.c_int),
]
class cap_user_data_t(ctypes.Structure):
# __user_cap_data_struct
_fields_ = [
("effective", ctypes.c_uint32),
("permitted", ctypes.c_uint32),
("inheritable", ctypes.c_uint32),
]
libc = ctypes.CDLL(None, use_errno=True)
libc.syscall.restype = ctypes.c_long
libc.unshare.argtypes = (ctypes.c_int,)
libc.statfs.argtypes = (ctypes.c_char_p, ctypes.c_void_p)
libc.eventfd.argtypes = (ctypes.c_int, ctypes.c_int)
libc.mount.argtypes = (ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_ulong, ctypes.c_char_p)
libc.pivot_root.argtypes = (ctypes.c_char_p, ctypes.c_char_p)
libc.umount2.argtypes = (ctypes.c_char_p, ctypes.c_int)
libc.capget.argtypes = (ctypes.c_void_p, ctypes.c_void_p)
libc.capset.argtypes = (ctypes.c_void_p, ctypes.c_void_p)
libc.fcntl.argtypes = (ctypes.c_int, ctypes.c_int, ctypes.c_int)
def terminal_is_dumb() -> bool:
return not sys.stdout.isatty() or not sys.stderr.isatty() or os.getenv("TERM", "") == "dumb"
class Style:
# fmt: off
bold: str = "\033[0;1;39m" if not terminal_is_dumb() else ""
blue: str = "\033[0;1;34m" if not terminal_is_dumb() else ""
gray: str = "\033[0;38;5;245m" if not terminal_is_dumb() else ""
red: str = "\033[31;1m" if not terminal_is_dumb() else ""
yellow: str = "\033[33;1m" if not terminal_is_dumb() else ""
reset: str = "\033[0m" if not terminal_is_dumb() else ""
# fmt: on
ENOSYS_MSG = f"""\
{Style.red}mkosi was unable to invoke the {{syscall}}() system call.{Style.reset}
This probably means either the system call is not implemented by the running kernel version ({{kver}}) or the
system call is prohibited via seccomp if mkosi is being executed inside a containerized environment.\
"""
def oserror(syscall: str, filename: str = "") -> None:
if ctypes.get_errno() == ENOSYS:
print(ENOSYS_MSG.format(syscall=syscall, kver=os.uname().version), file=sys.stderr)
raise OSError(ctypes.get_errno(), os.strerror(ctypes.get_errno()), filename or None)
def unshare(flags: int) -> None:
if libc.unshare(flags) < 0:
oserror("unshare")
def statfs(path: str) -> int:
# struct statfs is 120 bytes, which equals 15 longs. Since we only care about the first field
# and the first field is of type long, we avoid declaring the full struct by just passing an
# array of 15 longs as the output argument.
buffer = (ctypes.c_long * 15)()
if libc.statfs(path.encode(), ctypes.byref(buffer)) < 0:
oserror("statfs", path)
return int(buffer[0])
def mount(src: str, dst: str, type: str, flags: int, options: str) -> None:
srcb = src.encode() if src else None
typeb = type.encode() if type else None
optionsb = options.encode() if options else None
if libc.mount(srcb, dst.encode(), typeb, flags, optionsb) < 0:
oserror("mount", dst)
def umount2(path: str, flags: int = 0) -> None:
if libc.umount2(path.encode(), flags) < 0:
oserror("umount2", path)
def cap_permitted_to_ambient() -> None:
"""
When unsharing a user namespace and mapping the current user to itself, the user has a full
set of capabilities in the user namespace. This allows the user to do mounts after unsharing a
mount namespace for example. However, these capabilities are lost again when the user executes
a subprocess. As we also want subprocesses invoked by the user to be able to mount stuff, we
make sure the capabilities are inherited by adding all the user's capabilities to the inherited
and ambient capabilities set, which makes sure that they are passed down to subprocesses.
"""
header = cap_user_header_t(LINUX_CAPABILITY_VERSION_3, 0)
payload = (cap_user_data_t * LINUX_CAPABILITY_U32S_3)()
if libc.capget(ctypes.addressof(header), ctypes.byref(payload)) < 0:
oserror("capget")
payload[0].inheritable = payload[0].permitted
payload[1].inheritable = payload[1].permitted
if libc.capset(ctypes.addressof(header), ctypes.byref(payload)) < 0:
oserror("capset")
effective = payload[1].effective << 32 | payload[0].effective
with open("/proc/sys/kernel/cap_last_cap", "rb") as f:
last_cap = int(f.read())
libc.prctl.argtypes = (ctypes.c_int, ctypes.c_ulong, ctypes.c_ulong, ctypes.c_ulong, ctypes.c_ulong)
for cap in range(ctypes.sizeof(ctypes.c_uint64) * 8):
if cap > last_cap:
break
if effective & (1 << cap) and libc.prctl(PR_CAP_AMBIENT, PR_CAP_AMBIENT_RAISE, cap, 0, 0) < 0:
oserror("prctl")
def have_effective_cap(capability: int) -> bool:
with open("/proc/self/status", "rb") as f:
for line in f.readlines():
if line.startswith(b"CapEff:"):
return (int(line[7:], 16) & (1 << capability)) != 0
return False
def seccomp_suppress(*, chown: bool = False, sync: bool = False) -> None:
"""
There's still a few files and directories left in distributions in /usr and /etc that are
not owned by root. This causes package managers to fail to install the corresponding packages
when run from a single uid user namespace. Unfortunately, non-root users can only create files
owned by their own uid. To still allow non-root users to build images, if requested we install
a seccomp filter that makes calls to chown() and friends a noop.
"""
if not chown and not sync:
return
libseccomp = ctypes.CDLL("libseccomp.so.2")
if libseccomp is None:
raise FileNotFoundError("libseccomp.so.2")
libseccomp.seccomp_init.argtypes = (ctypes.c_uint32,)
libseccomp.seccomp_init.restype = ctypes.c_void_p
libseccomp.seccomp_release.argtypes = (ctypes.c_void_p,)
libseccomp.seccomp_release.restype = None
libseccomp.seccomp_syscall_resolve_name.argtypes = (ctypes.c_char_p,)
libseccomp.seccomp_rule_add_exact.argtypes = (
ctypes.c_void_p,
ctypes.c_uint32,
ctypes.c_int,
ctypes.c_uint,
)
libseccomp.seccomp_load.argtypes = (ctypes.c_void_p,)
seccomp = libseccomp.seccomp_init(SCMP_ACT_ALLOW)
suppress = []
if chown:
suppress += [
b"chown",
b"chown32",
b"fchown",
b"fchown32",
b"fchownat",
b"lchown",
b"lchown32",
]
if sync:
suppress += [
b"fdatasync",
b"fsync",
b"msync",
b"sync",
b"sync_file_range",
b"sync_file_range2",
b"syncfs",
]
try:
for syscall in suppress:
id = libseccomp.seccomp_syscall_resolve_name(syscall)
libseccomp.seccomp_rule_add_exact(seccomp, SCMP_ACT_ERRNO, id, 0)
libseccomp.seccomp_load(seccomp)
finally:
libseccomp.seccomp_release(seccomp)
def lsattr(path: str) -> int:
attr = ctypes.c_int()
r = 0
fd = os.open(path, os.O_CLOEXEC | os.O_RDONLY)
libc.ioctl.argtypes = (ctypes.c_int, ctypes.c_long, ctypes.c_void_p)
if libc.ioctl(fd, FS_IOC_GETFLAGS, ctypes.byref(attr)) < 0:
r = ctypes.get_errno()
os.close(fd)
if r != 0:
raise OSError(r, os.strerror(r), path)
return attr.value
def chattr(path: str, attr: int) -> None:
cattr = ctypes.c_int(attr)
fd = os.open(path, os.O_CLOEXEC | os.O_RDONLY)
r = 0
libc.ioctl.argtypes = (ctypes.c_int, ctypes.c_long, ctypes.c_void_p)
if libc.ioctl(fd, FS_IOC_SETFLAGS, ctypes.byref(cattr)) < 0:
r = ctypes.get_errno()
os.close(fd)
if r != 0:
raise OSError(r, os.strerror(r), path)
def join_new_session_keyring() -> None:
libkeyutils = ctypes.CDLL("libkeyutils.so.1")
if libkeyutils is None:
raise FileNotFoundError("libkeyutils.so.1")
libkeyutils.keyctl_join_session_keyring.argtypes = (ctypes.c_char_p,)
libkeyutils.keyctl_join_session_keyring.restype = ctypes.c_int32
keyring = libkeyutils.keyctl_join_session_keyring(None)
if keyring == -1:
oserror("keyctl")
def mount_rbind(src: str, dst: str, attrs: int = 0) -> None:
"""
When using the old mount syscall to do a recursive bind mount, mount options are not
applied recursively. Because we want to do recursive read-only bind mounts in some cases, we
use the new mount API for that which does allow recursively changing mount options when doing
bind mounts.
"""
flags = AT_NO_AUTOMOUNT | AT_RECURSIVE | AT_SYMLINK_NOFOLLOW | OPEN_TREE_CLONE
try:
libc.open_tree.argtypes = (ctypes.c_int, ctypes.c_char_p, ctypes.c_uint)
fd = libc.open_tree(AT_FDCWD, src.encode(), flags)
except AttributeError:
libc.syscall.argtypes = (ctypes.c_long, ctypes.c_int, ctypes.c_char_p, ctypes.c_uint)
fd = libc.syscall(NR_open_tree, AT_FDCWD, src.encode(), flags)
if fd < 0:
oserror("open_tree", src)
try:
attr = mount_attr()
attr.attr_set = attrs
flags = AT_EMPTY_PATH | AT_RECURSIVE
try:
libc.mount_setattr.argtypes = (
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_uint,
ctypes.c_void_p,
ctypes.c_size_t,
)
r = libc.mount_setattr(fd, b"", flags, ctypes.addressof(attr), MOUNT_ATTR_SIZE_VER0)
except AttributeError:
libc.syscall.argtypes = (
ctypes.c_long,
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_uint,
ctypes.c_void_p,
ctypes.c_size_t,
)
r = libc.syscall(NR_mount_setattr, fd, b"", flags, ctypes.addressof(attr), MOUNT_ATTR_SIZE_VER0)
if r < 0:
oserror("mount_setattr", src)
try:
libc.move_mount.argtypes = (
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_uint,
)
r = libc.move_mount(fd, b"", AT_FDCWD, dst.encode(), MOVE_MOUNT_F_EMPTY_PATH)
except AttributeError:
libc.syscall.argtypes = (
ctypes.c_long,
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_int,
ctypes.c_char_p,
ctypes.c_uint,
)
r = libc.syscall(NR_move_mount, fd, b"", AT_FDCWD, dst.encode(), MOVE_MOUNT_F_EMPTY_PATH)
if r < 0:
oserror("move_mount", dst)
finally:
os.close(fd)
class umask:
def __init__(self, mask: int) -> None:
self.mask = mask
def __enter__(self) -> None:
self.mask = os.umask(self.mask)
def __exit__(self, *args: object, **kwargs: object) -> None:
os.umask(self.mask)
def become_user(uid: int, gid: int) -> None:
"""
This function implements the required dance to unshare a user namespace and map the current
user to itself or to root within it. The kernel only allows a process running outside of the
unshared user namespace to write the necessary uid and gid mappings, so we fork off a child
process, make it wait until the parent process has unshared a user namespace, and then writes
the necessary uid and gid mappings.
"""
ppid = os.getpid()
event = libc.eventfd(0, 0)
if event < 0:
oserror("eventfd")
pid = os.fork()
if pid == 0:
try:
os.read(event, ctypes.sizeof(ctypes.c_uint64))
os.close(event)
# Fork again for UID map, bweh this is suuuper gross
newpid = os.fork()
if newpid == 0:
os.execl(
"/usr/bin/newuidmap",
"newuidmap",
str(ppid),
"0",
str(uid),
"1",
"1",
"100000",
"9999",
# 12345 is the UID of the "pmos" or "build" user in our containers
"12345",
"110000",
"1",
"10000",
"120000",
"1",
)
else:
os.waitpid(newpid, 0)
os.execl(
"/usr/bin/newgidmap",
"newgidmap",
str(ppid),
"0",
str(uid),
"1",
"1",
"100000",
"9999",
"12345",
"110000",
"1",
"10000",
"120000",
"1",
)
except OSError as e:
print(e.strerror, flush=True)
os._exit(e.errno or 1)
except BaseException as e:
print(e, flush=True)
os._exit(1)
else:
print("", flush=True)
os._exit(0)
try:
unshare(CLONE_NEWUSER)
except OSError as e:
if e.errno == EPERM:
print(UNSHARE_EPERM_MSG, file=sys.stderr)
raise
finally:
os.write(event, ctypes.c_uint64(1))
os.close(event)
_, status = os.waitpid(pid, 0)
rc = os.waitstatus_to_exitcode(status)
if rc != 0:
raise OSError(rc, os.strerror(rc))
def acquire_privileges(*, become_root: bool = False) -> bool:
if have_effective_cap(CAP_SYS_ADMIN) and (os.getuid() == 0 or not become_root):
return False
if become_root:
become_user(0, 0)
else:
become_user(os.getuid(), os.getgid())
cap_permitted_to_ambient()
return True
def userns_has_single_user() -> bool:
try:
with open("/proc/self/uid_map", "rb") as f:
lines = f.readlines()
except FileNotFoundError:
return False
return len(lines) == 1 and int(lines[0].split()[-1]) == 1
def chase(root: str, path: str) -> str:
if root == "/":
return os.path.realpath(path)
cwd = os.getcwd()
fd = os.open("/", os.O_CLOEXEC | os.O_PATH | os.O_DIRECTORY)
try:
os.chroot(root)
os.chdir("/")
return joinpath(root, os.path.realpath(path))
finally:
os.fchdir(fd)
os.close(fd)
os.chroot(".")
os.chdir(cwd)
def splitpath(path: str) -> tuple[str, ...]:
return tuple(p for p in path.split("/") if p)
def joinpath(path: str, *paths: str) -> str:
return os.path.join(path, *[p.lstrip("/") for p in paths])
def is_relative_to(one: str, two: str) -> bool:
return os.path.commonpath((one, two)) == two
def pack_file_descriptors() -> int:
fds = []
with os.scandir("/proc/self/fd") as it:
for e in it:
if not e.is_symlink() and (e.is_file() or e.is_dir()):
continue
try:
fd = int(e.name)
except ValueError:
continue
if fd < SD_LISTEN_FDS_START:
continue
fds.append(fd)
# os.scandir() either opens a file descriptor to the given path or dups the given file descriptor. Either
# way, there will be an extra file descriptor in the fds array that's not valid anymore now, so find out
# which one and drop it.
fds = sorted(fd for fd in fds if libc.fcntl(fd, F_GETFD, 0) >= 0)
# The following is a reimplementation of pack_fds() in systemd.
if len(fds) == 0:
return 0
start = 0
while True:
restart_from = -1
for i in range(start, len(fds)):
if fds[i] == SD_LISTEN_FDS_START + i:
continue
nfd = libc.fcntl(fds[i], F_DUPFD, SD_LISTEN_FDS_START + i)
if nfd < 0:
oserror("fnctl")
try:
os.close(fds[i])
except OSError as e:
if e.errno != EBADF:
raise
fds[i] = nfd
if nfd != (SD_LISTEN_FDS_START + i) and restart_from < 0:
restart_from = i
if restart_from < 0:
break
start = restart_from
assert fds[0] == SD_LISTEN_FDS_START
return len(fds)
class FSOperation:
def __init__(self, dst: str, *, relative: bool = False) -> None:
self.dst = dst
self.relative = relative
def execute(self, oldroot: str, newroot: str) -> None:
raise NotImplementedError()
@classmethod
def optimize(cls, fsops: list["FSOperation"]) -> list["FSOperation"]:
binds: dict[BindOperation, None] = {}
rest = []
for fsop in fsops:
if isinstance(fsop, BindOperation):
binds[fsop] = None
else:
rest.append(fsop)
# Drop all bind mounts that are mounted from beneath another bind mount to the same
# location within the new rootfs.
optimized = [
m
for m in binds
if not any(
m != n
and m.readonly == n.readonly
and m.required == n.required
and m.relative == n.relative
and is_relative_to(m.src, n.src)
and is_relative_to(m.dst, n.dst)
and os.path.relpath(m.src, n.src) == os.path.relpath(m.dst, n.dst)
for n in binds
)
]
# Make sure bind mounts override other operations on the same destination by appending them
# to the rest and depending on python's stable sort behavior. Additionally, relative operations
# always go last.
return sorted([*rest, *optimized], key=lambda fsop: (fsop.relative, splitpath(fsop.dst)))
class BindOperation(FSOperation):
def __init__(self, src: str, dst: str, *, readonly: bool, required: bool, relative: bool) -> None:
self.src = src
self.readonly = readonly
self.required = required
super().__init__(dst, relative=relative)
def __hash__(self) -> int:
return hash((splitpath(self.src), splitpath(self.dst), self.readonly, self.required))
def __eq__(self, other: object) -> bool:
return isinstance(other, BindOperation) and self.__hash__() == other.__hash__()
def execute(self, oldroot: str, newroot: str) -> None:
src = chase(newroot if self.relative else oldroot, self.src)
if not os.path.exists(src) and not self.required:
return
# If we're mounting a file on top of a symlink, mount directly on top of the symlink instead of
# resolving it.
dst = joinpath(newroot, self.dst)
if not os.path.isdir(src) and os.path.islink(dst):
return mount_rbind(src, dst, attrs=MOUNT_ATTR_RDONLY if self.readonly else 0)
dst = chase(newroot, self.dst)
if not os.path.exists(dst):
isfile = os.path.isfile(src)
with umask(~0o755):
os.makedirs(os.path.dirname(dst), exist_ok=True)
with umask(~0o644 if isfile else ~0o755):
if isfile:
os.close(os.open(dst, os.O_CREAT | os.O_CLOEXEC | os.O_EXCL))
else:
os.mkdir(dst)
mount_rbind(src, dst, attrs=MOUNT_ATTR_RDONLY if self.readonly else 0)
class ProcOperation(FSOperation):
def execute(self, oldroot: str, newroot: str) -> None:
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(dst, exist_ok=True)
mount_rbind(joinpath(oldroot, "proc"), dst)
class BinfmtOperation(FSOperation):
def execute(self, oldroot: str, newroot: str) -> None:
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(dst, exist_ok=True)
mount("binfmt_misc", dst, "binfmt_misc", 0, "")
class DevOperation(FSOperation):
def __init__(self, ttyname: str, dst: str) -> None:
self.ttyname = ttyname
super().__init__(dst)
def execute(self, oldroot: str, newroot: str) -> None:
# We don't put actual devices in /dev, just the API stuff in there that all manner of
# things depend on, like /dev/null.
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(dst, exist_ok=True)
# Note that the mode is crucial here. If the default mode (1777) is used, trying to access
# /dev/null fails with EACCESS for unknown reasons.
mount("tmpfs", dst, "tmpfs", 0, "mode=0755")
for node in ("null", "zero", "full", "random", "urandom", "tty", "fuse"):
nsrc = joinpath(oldroot, "dev", node)
if node == "fuse" and not os.path.exists(nsrc):
continue
ndst = joinpath(dst, node)
os.close(os.open(ndst, os.O_CREAT | os.O_CLOEXEC | os.O_EXCL))
mount(nsrc, ndst, "", MS_BIND, "")
for i, node in enumerate(("stdin", "stdout", "stderr")):
os.symlink(f"/proc/self/fd/{i}", joinpath(dst, node))
os.symlink("/proc/self/fd", joinpath(dst, "fd"))
os.symlink("/proc/kcore", joinpath(dst, "core"))
with umask(~0o1777):
os.mkdir(joinpath(dst, "shm"), mode=0o1777)
with umask(~0o755):
os.mkdir(joinpath(dst, "pts"))
mount("devpts", joinpath(dst, "pts"), "devpts", 0, "newinstance,ptmxmode=0666,mode=620")
os.symlink("pts/ptmx", joinpath(dst, "ptmx"))
if self.ttyname:
os.close(os.open(joinpath(dst, "console"), os.O_CREAT | os.O_CLOEXEC | os.O_EXCL))
mount(joinpath(oldroot, self.ttyname), joinpath(dst, "console"), "", MS_BIND, "")
class TmpfsOperation(FSOperation):
def execute(self, oldroot: str, newroot: str) -> None:
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(dst, exist_ok=True)
options = "" if any(dst.endswith(suffix) for suffix in ("/tmp", "/var/tmp")) else "mode=0755"
mount("tmpfs", dst, "tmpfs", 0, options)
class DirOperation(FSOperation):
def execute(self, oldroot: str, newroot: str) -> None:
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(os.path.dirname(dst), exist_ok=True)
mode = 0o1777 if any(dst.endswith(suffix) for suffix in ("/tmp", "/var/tmp")) else 0o755
if not os.path.exists(dst):
with umask(~mode):
os.mkdir(dst, mode=mode)
class SymlinkOperation(FSOperation):
def __init__(self, src: str, dst: str) -> None:
self.src = src
super().__init__(dst)
def execute(self, oldroot: str, newroot: str) -> None:
dst = joinpath(newroot, self.dst)
try:
return os.symlink(self.src, dst)
except FileExistsError:
if os.path.islink(dst) and os.readlink(dst) == self.src:
return
if os.path.isdir(dst):
raise
# If the target already exists and is not a directory, create the symlink somewhere else and mount
# it over the existing file or symlink.
os.symlink(self.src, "/symlink")
mount_rbind("/symlink", dst)
os.unlink("/symlink")
class WriteOperation(FSOperation):
def __init__(self, data: str, dst: str) -> None:
self.data = data
super().__init__(dst)
def execute(self, oldroot: str, newroot: str) -> None:
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(os.path.dirname(dst), exist_ok=True)
with open(dst, "wb") as f:
f.write(self.data.encode())
class OverlayOperation(FSOperation):
def __init__(self, lowerdirs: tuple[str, ...], upperdir: str, workdir: str, dst: str) -> None:
self.lowerdirs = lowerdirs
self.upperdir = upperdir
self.workdir = workdir
super().__init__(dst)
# This supports being used as a context manager so we can reuse the logic for mount_overlay()
# in mounts.py.
def __enter__(self) -> None:
self.execute("/", "/")
def __exit__(self, *args: object, **kwargs: object) -> None:
umount2(self.dst)
def execute(self, oldroot: str, newroot: str) -> None:
lowerdirs = tuple(chase(oldroot, p) for p in self.lowerdirs)
upperdir = (
chase(oldroot, self.upperdir) if self.upperdir and self.upperdir != "tmpfs" else self.upperdir
)
workdir = chase(oldroot, self.workdir) if self.workdir else None
dst = chase(newroot, self.dst)
with umask(~0o755):
os.makedirs(os.path.dirname(dst), exist_ok=True)
mode = 0o1777 if any(dst.endswith(suffix) for suffix in ("/tmp", "/var/tmp")) else 0o755
if not os.path.exists(dst):
with umask(~mode):
os.mkdir(dst, mode=mode)
options = [
f"lowerdir={':'.join(lowerdirs)}",
"userxattr",
# Disable the inodes index and metacopy (only copy metadata upwards if possible)
# options. If these are enabled (e.g., if the kernel enables them by default),
# the mount will fail if the upper directory has been earlier used with a different
# lower directory, such as with a build overlay that was generated on top of a
# different temporary root.
# See https://www.kernel.org/doc/html/latest/filesystems/overlayfs.html#sharing-and-copying-layers
# and https://github.com/systemd/mkosi/issues/1841.
"index=off",
"metacopy=off",
]
if upperdir and upperdir == "tmpfs":
mount("tmpfs", dst, "tmpfs", 0, "mode=0755")
with umask(~mode):
os.mkdir(f"{dst}/upper", mode=mode)
with umask(~0o755):
os.mkdir(f"{dst}/work")
options += [f"upperdir={dst}/upper", f"workdir={dst}/work"]
else:
if upperdir:
options += [f"upperdir={upperdir}"]
if workdir:
options += [f"workdir={workdir}"]
mount("overlayfs", dst, "overlay", 0, ",".join(options))
ANSI_HIGHLIGHT = "\x1b[0;1;39m" if os.isatty(2) else ""
ANSI_NORMAL = "\x1b[0m" if os.isatty(2) else ""
HELP = f"""\
mkosi-sandbox [OPTIONS...] COMMAND [ARGUMENTS...]
{ANSI_HIGHLIGHT}Run the specified command in a custom sandbox.{ANSI_NORMAL}
-h --help Show this help
--version Show package version
--tmpfs DST Mount a new tmpfs on DST
--dev DST Mount dev on DST
--proc DST Mount procfs on DST
--dir DST Create a new directory at DST
--bind SRC DST Bind mount the host path SRC to DST
--bind-try SRC DST Bind mount the host path SRC to DST if it exists
--ro-bind SRC DST Bind mount the host path SRC to DST read-only
--ro-bind-try SRC DST Bind mount the host path SRC to DST read-only if it exists
--symlink SRC DST Create a symlink at DST pointing to SRC
--write DATA DST Write DATA to DST
--overlay-lowerdir DIR Add a lower directory for the next overlayfs mount
--overlay-upperdir DIR Set the upper directory for the next overlayfs mount
--overlay-workdir DIR Set the working directory for the next overlayfs mount
--overlay DST Mount an overlay filesystem at DST
--unsetenv NAME Unset the environment variable with name NAME
--setenv NAME VALUE Set the environment variable with name NAME to VALUE
--chdir DIR Change the working directory in the sandbox to DIR
--same-dir Change the working directory in the sandbox to $PWD
--become-root Map the current user/group to root:root in the sandbox
--suppress-chown Make chown() syscalls in the sandbox a noop
--suppress-sync Make sync() syscalls in the sandbox a noop
--unshare-net Unshare the network namespace if possible
--unshare-ipc Unshare the IPC namespace if possible
--suspend Stop process before execve()
See the mkosi-sandbox(1) man page for details.\
"""
UNSHARE_EPERM_MSG = f"""\
{Style.red}mkosi was forbidden to unshare namespaces{Style.reset}.
This probably means your distribution has restricted unprivileged user namespaces.
Please consult the REQUIREMENTS section of the mkosi man page, e.g. via "mkosi
documentation", for workarounds.\
"""
def setup_mounts(fsops: list[FSOperation]) -> None:
# We need a workspace to setup the sandbox, the easiest way to do this in a tmpfs, since it's
# automatically cleaned up. We need a mountpoint to put the workspace on and it can't be root,
# so let's use /tmp which is almost guaranteed to exist.
mount("tmpfs", "/tmp", "tmpfs", 0, "")
os.chdir("/tmp")
with umask(~0o755):
# This is where we set up the sandbox rootfs
os.mkdir("newroot")
# This is the old rootfs which is used as the source for mounts in the new rootfs.
os.mkdir("oldroot")
# Make sure that newroot is a mountpoint.
mount("newroot", "newroot", "", MS_BIND | MS_REC, "")
# Make the workspace in /tmp / and put the old rootfs in oldroot.
if libc.pivot_root(b".", b"oldroot") < 0:
# pivot_root() can fail in the initramfs since / isn't a mountpoint there, so let's fall
# back to MS_MOVE if that's the case.
# First we move the old rootfs to oldroot.
mount("/", "oldroot", "", MS_BIND | MS_REC, "")
# Then we move the workspace (/tmp) to /.
mount(".", "/", "", MS_MOVE, "")
# chroot and chdir to fully make the workspace the new root.
os.chroot(".")
os.chdir(".")
# When we use MS_MOVE we have to unmount oldroot/tmp manually to reveal the original /tmp
# again as it might contain stuff that we want to mount into the sandbox.
umount2("oldroot/tmp", MNT_DETACH)
for fsop in fsops:
fsop.execute("oldroot", "newroot")
# Now that we're done setting up the sandbox let's pivot root into newroot to make it the new
# root. We use the pivot_root(".", ".") process described in the pivot_root() man page.
os.chdir("newroot")
# We're guaranteed to have / be a mount when we get here, so pivot_root() won't fail anymore,
# even if we're in the initramfs.
if libc.pivot_root(b".", b".") < 0:
oserror("pivot_root")
# As documented in the pivot_root() man page, this will unmount the old rootfs.
umount2(".", MNT_DETACH)
# Avoid surprises by making sure the sandbox's mount propagation is shared. This doesn't
# actually mean mounts get propagated into the host. Instead, a new mount propagation peer
# group is set up.
mount("", ".", "", MS_SHARED | MS_REC, "")
def main() -> None:
# We don't use argparse as it takes +- 10ms to import and since this is primarily for internal
# use, it's not necessary to have amazing UX for this CLI interface so it's trivial to write
# ourselves.
argv = list(reversed(sys.argv[1:]))
fsops: list[FSOperation] = []
setenv = []
unsetenv = []
lowerdirs = []
upperdir = ""
workdir = ""
chdir = None
become_root = suppress_chown = suppress_sync = unshare_net = unshare_ipc = suspend = pack_fds = False
ttyname = os.ttyname(2) if os.isatty(2) else ""
while argv:
arg = argv.pop()
if arg == "--":
break
if arg in ("-h", "--help"):
print(HELP, file=sys.stderr)
sys.exit(0)
elif arg == "--version":
print(__version__, file=sys.stderr)
sys.exit(0)
if arg == "--tmpfs":
fsops.append(TmpfsOperation(argv.pop()))
elif arg == "--dev":
fsops.append(DevOperation(ttyname, argv.pop()))
elif arg == "--proc":
fsops.append(ProcOperation(argv.pop()))
elif arg == "--dir":
fsops.append(DirOperation(argv.pop()))
elif arg in ("--bind", "--ro-bind", "--bind-try", "--ro-bind-try"):
readonly = arg.startswith("--ro")
required = not arg.endswith("-try")
src = argv.pop()
fsops.append(
BindOperation(
src.removeprefix("+"),
argv.pop(),
readonly=readonly,
required=required,
relative=src.startswith("+"),
)
)
elif arg == "--symlink":
fsops.append(SymlinkOperation(argv.pop(), argv.pop()))
elif arg == "--write":
fsops.append(WriteOperation(argv.pop(), argv.pop()))
elif arg == "--overlay-lowerdir":
lowerdirs.append(argv.pop())
elif arg == "--overlay-upperdir":
upperdir = argv.pop()
elif arg == "--overlay-workdir":
workdir = argv.pop()
elif arg == "--overlay":
fsops.append(OverlayOperation(tuple(reversed(lowerdirs)), upperdir, workdir, argv.pop()))
upperdir = ""
workdir = ""
lowerdirs = []
elif arg == "--unsetenv":
unsetenv.append(argv.pop())
elif arg == "--setenv":
setenv.append((argv.pop(), argv.pop()))
elif arg == "--chdir":
chdir = argv.pop()
elif arg == "--same-dir":
chdir = os.getcwd()
elif arg == "--become-root":
become_root = True
elif arg == "--suppress-chown":
suppress_chown = True
elif arg == "--suppress-sync":
suppress_sync = True
elif arg == "--unshare-net":
unshare_net = True
elif arg == "--unshare-ipc":
unshare_ipc = True
elif arg == "--suspend":
suspend = True
elif arg == "--pack-fds":
pack_fds = True
elif arg.startswith("-"):
raise ValueError(f"Unrecognized option {arg}")
else:
argv.append(arg)
break
argv.reverse()
argv = argv or ["bash"]
# Make sure all destination paths are absolute.
for fsop in fsops:
if fsop.dst[0] != "/":
raise ValueError(f"{fsop.dst} is not an absolute path")
fsops = FSOperation.optimize(fsops)
for k, v in setenv:
os.environ[k] = v
for e in unsetenv:
if e in os.environ:
del os.environ[e]
if pack_fds:
nfds = pack_file_descriptors()
if nfds > 0:
os.environ["LISTEN_FDS"] = str(nfds)
os.environ["LISTEN_PID"] = str(os.getpid())
namespaces = CLONE_NEWNS
if unshare_net and have_effective_cap(CAP_NET_ADMIN):
namespaces |= CLONE_NEWNET
if unshare_ipc:
namespaces |= CLONE_NEWIPC
userns = acquire_privileges(become_root=become_root)
seccomp_suppress(
# If we're root in a user namespace with a single user, we're still not going to be able to
# chown() stuff, so check for that and apply the seccomp filter as well in that case.
chown=suppress_chown and (userns or userns_has_single_user()),
sync=suppress_sync,
)
try:
unshare(namespaces)
except OSError as e:
# This can happen here as well as in become_user, it depends on exactly
# how the userns restrictions are implemented.
if e.errno == EPERM:
print(UNSHARE_EPERM_MSG, file=sys.stderr)
raise
# If we unshared the user namespace the mount propagation of root is changed to slave automatically.
if not userns:
mount("", "/", "", MS_SLAVE | MS_REC, "")
setup_mounts(fsops)
if chdir:
os.chdir(chdir)
if suspend:
os.kill(os.getpid(), SIGSTOP)
try:
os.execvp(argv[0], argv)
except OSError as e:
# Let's return a recognizable error when the binary we're going to execute is not found.
# We use 127 as that's the exit code used by shells when a program to execute is not found.
if e.errno == ENOENT:
sys.exit(127)
raise
if __name__ == "__main__":
main()
Reference in a new issue View git blame Copy permalink