def execute_in_namespace(func, use_network_ns=True):
"""Execute a function in a child process in separate namespaces.
@param func: a parameter-less function returning an int (which will be the process' exit value)
@return: the PID of the created child process
"""
flags = (signal.SIGCHLD |
libc.CLONE_NEWNS | libc.CLONE_NEWUTS | libc.CLONE_NEWIPC | libc.CLONE_NEWUSER |
libc.CLONE_NEWPID)
if use_network_ns:
flags |= libc.CLONE_NEWNET
# We use the syscall clone() here, which is similar to fork().
# Calling it without letting Python know about it is dangerous (especially because
# we want to execute Python code in the child, too), but so far it seems to work.
# Basically we attempt to do (almost) the same that os.fork() does (cf. function os_fork_impl
# in https://github.com/python/cpython/blob/master/Modules/posixmodule.c).
# We currently do not take the import lock os.lock() does because it is only available
# via an internal API, and because the child should never import anything anyway
# (inside the container, modules might not be visible).
# It is very important, however, that we have the GIL during clone(),
# otherwise the child will often deadlock when trying to execute Python code.
# Luckily, the ctypes module allows us to hold the GIL while executing the
# function by using ctypes.PyDLL as library access instead of ctypes.CLL.
def child_func():
# This is necessary for correcting the Python interpreter state after a
# fork-like operation. For example, it resets the GIL and fixes state of
# several modules like threading and signal.
ctypes.pythonapi.PyOS_AfterFork()
return func()
with allocate_stack() as stack:
pid = libc.clone(ctypes.CFUNCTYPE(ctypes.c_int)(child_func), stack, flags, None)
return pid
def execute_in_namespace(func, use_network_ns=True):
"""Execute a function in a child process in separate namespaces.
@param func: a parameter-less function returning an int (which will be the process' exit value)
@return: the PID of the created child process
"""
flags = (signal.SIGCHLD |
libc.CLONE_NEWNS | libc.CLONE_NEWUTS | libc.CLONE_NEWIPC | libc.CLONE_NEWUSER |
libc.CLONE_NEWPID)
if use_network_ns:
flags |= libc.CLONE_NEWNET
# We use the syscall clone() here, which is similar to fork().
# Calling it without letting Python know about it is dangerous (especially because
# we want to execute Python code in the child, too), but so far it seems to work.
# Basically we attempt to do (almost) the same that os.fork() does (cf. function os_fork_impl
# in https://github.com/python/cpython/blob/master/Modules/posixmodule.c).
# We currently do not take the import lock os.lock() does because it is only available
# via an internal API, and because the child should never import anything anyway
# (inside the container, modules might not be visible).
# It is very important, however, that we have the GIL during clone(),
# otherwise the child will often deadlock when trying to execute Python code.
# Luckily, the ctypes module allows us to hold the GIL while executing the
# function by using ctypes.PyDLL as library access instead of ctypes.CLL.
def child_func():
# This is necessary for correcting the Python interpreter state after a
# fork-like operation. For example, it resets the GIL and fixes state of
# several modules like threading and signal.
ctypes.pythonapi.PyOS_AfterFork()
return func()
with allocate_stack() as stack:
pid = libc.clone(ctypes.CFUNCTYPE(ctypes.c_int)(child_func), stack, flags, None)
return pid
def execute_in_namespace(func, use_network_ns=True):
"""Execute a function in a child process in separate namespaces.
@param func: a parameter-less function returning an int
(which will be the process' exit value)
@return: the PID of the created child process
"""
flags = (
signal.SIGCHLD
| libc.CLONE_NEWNS
| libc.CLONE_NEWUTS
| libc.CLONE_NEWIPC
| libc.CLONE_NEWUSER
| libc.CLONE_NEWPID
)
if use_network_ns:
flags |= libc.CLONE_NEWNET
# We need to use the syscall clone(), which is similar to fork(), but not available
# in the Python API. We can call it directly using ctypes, but then the state of the
# Python interpreter is inconsistent, so we need to fix that. Python >= 3.7 has
# three C functions that should be called before and after fork/clone:
# https://docs.python.org/3/c-api/sys.html#c.PyOS_BeforeFork
# This is the same that os.fork() does (cf. os_fork_impl
# in https://github.com/python/cpython/blob/master/Modules/posixmodule.c).
# Furthermore, it is very important that we have the GIL during clone(),
# otherwise the child will often deadlock when trying to execute Python code.
# Luckily, the ctypes module allows us to hold the GIL while executing the
# function by using ctypes.PyDLL as library access instead of ctypes.CLL.
# Two difficulties remain:
# 1. On Python < 3.7, only PyOS_AfterFork() (to be called in the child) exists.
# Other cleanup done by os_fork_impl is not accessible to us, so we ignore it.
# For example, we do not take the import lock because it is only
# available via an internal API, and because the child should never import anything
# anyway (inside the container, modules might not be visible).
# 2. On all Python versions, the interpreter state in the child is inconsistent
# until PyOS_AfterFork_Child() is called. However, if we pass the Python function
# _python_clone_child_callback() as callback to clone and do the cleanup in
# its first line, it is too late because the Python interpreter is already used.
# This actually causes problems if benchexec is executed with a high number of
# parallel runs because of thread contention, the gil_drop_request and a deadlock
# in drop_gil (cf. https://github.com/sosy-lab/benchexec/issues/435).
# So we should avoid executing Python code at all before PyOS_AfterFork_Child().
# We do not want to take the hassle of shipping C code with BenchExec, so we use
# _generate_native_clone_child_callback() to generate machine code on the fly
# as replacement for _python_clone_child_callback(). This works for x86_64 Linux
# and we expect practically all BenchExec users to fall in this category. For others
# there is still the pure Python callback, which in practice works totally fine as
# long as there does not exist a huge number of threads.
# There is a workaround using sys.setswitchinterval(), however, it is too late to
# apply it here in this function, because gil_drop_request could already be set.
# Summary:
# - For Linux x86_64 we use native code from _generate_native_clone_child_callback()
# - Otherwise, we use sys.setswitchinterval() as workaround in localexecution.py.
# - Direct users of ContainerExecutor are fine in practice if they use few threads.
func_p = _CLONE_NESTED_CALLBACK(func) # store in variable to avoid GC
with allocate_stack() as stack:
try:
ctypes.pythonapi.PyOS_BeforeFork()
pid = libc.clone(_clone_child_callback, stack, flags, func_p)
finally:
ctypes.pythonapi.PyOS_AfterFork_Parent()
return pid