def grandchild():
"""Setup everything inside the process that finally exec()s the tool."""
try:
# We know that this process has PID 2 in the inner namespace,
# but we actually need to know its PID in the outer namespace
# such that parent can put us into the correct cgroups. According to
# http://man7.org/linux/man-pages/man7/pid_namespaces.7.html,
# there are two ways to achieve this: sending a message with the PID
# via a socket (but Python 2 lacks a convenient API for sendmsg),
# and reading /proc/self in the outer procfs instance
# (that's what we do).
my_outer_pid = container.get_my_pid_from_procfs()
container.mount_proc(self._container_system_config)
container.drop_capabilities()
container.reset_signal_handling()
child_setup_fn() # Do some other setup the caller wants.
# Signal readiness to parent by sending our PID
# and wait until parent is also ready
os.write(to_parent, str(my_outer_pid).encode())
received = os.read(from_parent, 1)
assert received == MARKER_PARENT_COMPLETED, received
finally:
# close remaining ends of pipe
os.close(from_parent)
os.close(to_parent)
def grandchild():
"""Setup everything inside the process that finally exec()s the tool."""
try:
# We know that this process has PID 2 in the inner namespace,
# but we actually need to know its PID in the outer namespace
# such that parent can put us into the correct cgroups.
# According to http://man7.org/linux/man-pages/man7/pid_namespaces.7.html,
# there are two ways to achieve this: sending a message with the PID
# via a socket (but Python < 3.3 lacks a convenient API for sendmsg),
# and reading /proc/self in the outer procfs instance (that's what we do).
my_outer_pid = container.get_my_pid_from_procfs()
container.mount_proc()
container.drop_capabilities()
container.reset_signal_handling()
child_setup_fn() # Do some other setup the caller wants.
# Signal readiness to parent by sending our PID and wait until parent is also ready
os.write(to_parent, str(my_outer_pid).encode())
received = os.read(from_parent, 1)
assert received == b"\0", received
finally:
# close remaining ends of pipe
os.close(from_parent)
os.close(to_parent)
def _init_container_and_load_tool(
tool_module,
temp_dir,
network_access,
dir_modes,
container_system_config,
container_tmpfs, # ignored, tmpfs is always used
):
"""Initialize container for the current process and load given tool-info module."""
# Prepare for private home directory, some tools write there
if container_system_config:
dir_modes.setdefault(container.CONTAINER_HOME, container.DIR_HIDDEN)
os.environ["HOME"] = container.CONTAINER_HOME
# Preparations
temp_dir = temp_dir.encode()
dir_modes = collections.OrderedDict(
sorted(
((path.encode(), kind) for (path, kind) in dir_modes.items()),
key=lambda tupl: len(tupl[0]),
))
uid = container.CONTAINER_UID if container_system_config else os.getuid()
gid = container.CONTAINER_GID if container_system_config else os.getgid()
# Create container.
# Contrary to ContainerExecutor, which uses clone to start a new process in new
# namespaces, we use unshare, which puts the current process (the multiprocessing
# worker process) into new namespaces.
# The exception is the PID namespace, which will only apply to children processes.
flags = (libc.CLONE_NEWNS
| libc.CLONE_NEWUTS
| libc.CLONE_NEWIPC
| libc.CLONE_NEWUSER
| libc.CLONE_NEWPID)
if not network_access:
flags |= libc.CLONE_NEWNET
libc.unshare(flags)
# Container config
container.setup_user_mapping(os.getpid(), uid, gid)
_setup_container_filesystem(temp_dir, dir_modes, container_system_config)
if container_system_config:
libc.sethostname(container.CONTAINER_HOSTNAME)
if not network_access:
container.activate_network_interface("lo")
# Because this process is not actually in the new PID namespace, we fork.
# The child will be in the new PID namespace and will assume the role of the acting
# multiprocessing worker (which it can do because it inherits the file descriptors
# that multiprocessing uses for communication).
# The original multiprocessing worker (the parent of the fork) must do nothing in
# order to not confuse multiprocessing.
pid = os.fork()
if pid:
container.drop_capabilities()
# block parent such that it does nothing
os.waitpid(pid, 0)
os._exit(0)
# Finalize container setup in child
container.mount_proc(container_system_config) # only possible in child
container.drop_capabilities()
libc.prctl(libc.PR_SET_DUMPABLE, libc.SUID_DUMP_DISABLE, 0, 0, 0)
logging.debug("Loading tool-info module %s in container", tool_module)
global tool
try:
tool = __import__(tool_module, fromlist=["Tool"]).Tool()
except BaseException as e:
tool = None
return e
return tool.__doc__
def _init_container(
temp_dir,
network_access,
dir_modes,
container_system_config,
container_tmpfs, # ignored, tmpfs is always used
):
"""
Create a fork of this process in a container. This method only returns in the fork,
so calling it seems like moving the current process into a container.
"""
# Prepare for private home directory, some tools write there
if container_system_config:
dir_modes.setdefault(container.CONTAINER_HOME, container.DIR_HIDDEN)
os.environ["HOME"] = container.CONTAINER_HOME
# Preparations
temp_dir = temp_dir.encode()
dir_modes = collections.OrderedDict(
sorted(
((path.encode(), kind) for (path, kind) in dir_modes.items()),
key=lambda tupl: len(tupl[0]),
)
)
uid = container.CONTAINER_UID if container_system_config else os.getuid()
gid = container.CONTAINER_GID if container_system_config else os.getgid()
# Create container.
# Contrary to ContainerExecutor, which uses clone to start a new process in new
# namespaces, we use unshare, which puts the current process (the multiprocessing
# worker process) into new namespaces.
# The exception is the PID namespace, which will only apply to children processes.
flags = (
libc.CLONE_NEWNS
| libc.CLONE_NEWUTS
| libc.CLONE_NEWIPC
| libc.CLONE_NEWUSER
| libc.CLONE_NEWPID
)
if not network_access:
flags |= libc.CLONE_NEWNET
try:
libc.unshare(flags)
except OSError as e:
if (
e.errno == errno.EPERM
and util.try_read_file("/proc/sys/kernel/unprivileged_userns_clone") == "0"
):
raise BenchExecException(
"Unprivileged user namespaces forbidden on this system, please "
"enable them with 'sysctl kernel.unprivileged_userns_clone=1' "
"or disable container mode"
)
else:
raise BenchExecException(
"Creating namespace for container mode failed: " + os.strerror(e.errno)
)
# Container config
container.setup_user_mapping(os.getpid(), uid, gid)
_setup_container_filesystem(temp_dir, dir_modes, container_system_config)
if container_system_config:
libc.sethostname(container.CONTAINER_HOSTNAME)
if not network_access:
container.activate_network_interface("lo")
# Because this process is not actually in the new PID namespace, we fork.
# The child will be in the new PID namespace and will assume the role of the acting
# multiprocessing worker (which it can do because it inherits the file descriptors
# that multiprocessing uses for communication).
# The original multiprocessing worker (the parent of the fork) must do nothing in
# order to not confuse multiprocessing.
pid = os.fork()
if pid:
container.drop_capabilities()
# block parent such that it does nothing
os.waitpid(pid, 0)
os._exit(0)
# Finalize container setup in child
container.mount_proc(container_system_config) # only possible in child
container.drop_capabilities()
libc.prctl(libc.PR_SET_DUMPABLE, libc.SUID_DUMP_DISABLE, 0, 0, 0)
container.setup_seccomp_filter()
