Hidden edges in the Python object graph
2019-03-08 · view article source
It’s well known that Python doesn’t really have a concept of private data. Underscore-named attributes are nothing more than a convention; anyone can access them, and they can leak implementation details across class boundaries in a way that (say) private fields in Java can’t.
When actual encapsulation is desired—say, to prevent leaking new attributes onto an object—a common practice is to use a closure to capture the desired the value rather than assigning it to an attribute. For instance, in the following code, it might seem unlikely that we could implement
callback such that the program always completes successfully:
def make_greeter(expected_name): """Return a function that expects to greet a very specific guest.""" def greet(actual_name): if expected_name != actual_name: raise ValueError("Who are you?") else: print("Hello, %s!" % actual_name) return greet def greet_guest(greet, choose_guest): guest = choose_guest() greet(guest) def main(): greet_guest(make_greeter(input()), callback) def callback(): # Wherever to begin? raise NotImplementedError() if __name__ == "__main__": main()
But in fact we can, with the help of two “features” of Python dynamism. The following definition suffices:
def callback(): import inspect parent_frame = inspect.stack().frame greeter = parent_frame.f_locals["greet"] expected_name = greeter.__closure__.cell_contents return expected_name
In case it’s not clear, here’s what’s going on. We first introspect the runtime environment to traverse up the call stack, and obtain a reference to our caller’s activation record. This includes our caller’s local variables, which are conveniently keyed by name, so we can easily grab whatever
greet function was passed into
greet_guest. Finally, we can directly access that function’s closure environment, pulling the contents of the first (and only) value over which it closes.
In Python 2, the syntax is a bit different—
parent_frame is just a tuple, so replace
__closure__ is instead called
func_closure. The semantics, however, are the same.
In case you were wondering, Python at least has the good sense to make local variable dictionaries and closure cells read-only, so you can’t completely wreak havoc.
While this code probably isn’t quite ready for production, it can be useful in a pinch when you want to inspect a running Python process. For instance, I used the following incantation at a REPL to get the details of a particularly hard-to-access value that was being mutated by the Python import system:
import inspect print(inspect.getsource( type(tensorboard.program) # a dynamically generated class… .__getattr__ # …whose member function closes over… .__closure__.cell_contents # …a decorated function… .__closure__.cell_contents # …whose underlying function… .__closure__.cell_contents # …contains a value of interest! ))
Sure, it requires knowing the memory layout pretty well ahead of time—but hey, it works!
Is nothing secret?
So, closure-captured values and locals in ancestor function calls are accessible. This works even across threads, because
sys._current_frames() gives references to stack frame objects for all running threads, and a frame reference is all we need to perform the traversals above. Are there any places that we can stash a value such that later code can’t access it? At the very least, values that are never bound to a name look pretty safe: in
print(input() + callback()), the body of
callback will be hard-pressed to discover the result of
Python, of course, thwarts any such attempt at sanity by directly providing
gc.get_objects(), which returns a list of all objects tracked by the garbage collector. This almost suffices, except that some values aren’t tracked by the garbage collector (e.g.,
dicts containing only atomic keys and values), so this isn’t quite good enough. While a definitive answer proves elusive, one conclusion seems clear: if for some reason you find yourself wanting to make the assumption that you can effectively isolate a piece of secret memory from untrusted Python code running in the same process… don’t!