|Title:||Robust Stack Overflow Handling|
|Author:||Mark Shannon <mark at hotpy.org>|
- Rejection Notice
- Backwards Compatibility
- Security Implications
- Performance Impact
- Rejected Ideas
- Open Issues
This PEP proposes that Python should treat machine stack overflow differently from runaway recursion.
This would allow programs to set the maximum recursion depth to fit their needs and provide additional safety guarantees.
If this PEP is accepted, then the following program will run safely to completion:
sys.setrecursionlimit(1_000_000) def f(n): if n: f(n-1) f(500_000)
and the following program will raise a StackOverflow, without causing a VM crash:
sys.setrecursionlimit(1_000_000) class X: def __add__(self, other): return self + other X() + 1
CPython uses a single recursion depth counter to prevent both runaway recursion and C stack overflow. However, runaway recursion and machine stack overflow are two different things. Allowing machine stack overflow is a potential security vulnerability, but limiting recursion depth can prevent the use of some algorithms in Python.
Currently, if a program needs to deeply recurse it must manage the maximum recursion depth allowed, hopefully managing to set it in the region between the minimum needed to run correctly and the maximum that is safe to avoid a memory protection error.
By separating the checks for C stack overflow from checks for recursion depth, pure Python programs can run safely, using whatever level of recursion they require.
CPython currently relies on a single limit to guard against potentially dangerous stack overflow in the virtual machine and to guard against run away recursion in the Python program.
This is a consequence of the implementation which couples the C and Python call stacks. By breaking this coupling, we can improve both the usability of CPython and its safety.
The recursion limit exists to protect against runaway recursion, the integrity of the virtual machine should not depend on it. Similarly, recursion should not be limited by implementation details.
Two new exception classes will be added, StackOverflow and RecursionOverflow, both of which will be sub-classes of RecursionError
A StackOverflow exception will be raised whenever the interpreter or builtin module code determines that the C stack is at or nearing a limit of safety. StackOverflow is a sub-class of RecursionError, so any code that handles RecursionError will handle StackOverflow
A RecursionOverflow exception will be raised when a call to a Python function causes the recursion limit to be exceeded. This is a slight change from current behavior which raises a RecursionError. RecursionOverflow is a sub-class of RecursionError, so any code that handles RecursionError will continue to work as before.
In order to provide the above guarantees and ensure that any program that worked previously continues to do so, the Python and C stack will need to be separated. That is, calls to Python functions from Python functions, should not consume space on the C stack. Calls to and from builtin functions will continue to consume space on the C stack.
The size of the C stack will be implementation defined, and may vary from machine to machine. It may even differ between threads. However, there is an expectation that any code that could run with the recursion limit set to the previous default value, will continue to run.
Many operations in Python perform some sort of call at the C level. Most of these will continue to consume C stack, and will result in a StackOverflow exception if uncontrolled recursion occurs.
Other implementations are required to fail safely regardless of what value the recursion limit is set to.
If the implementation couples the Python stack to the underlying VM or hardware stack, then it should raise a RecursionOverflow exception when the recursion limit is exceeded, but the underlying stack does not overflow. If the underlying stack overflows, or is near to overflow, then a StackOverflow exception should be raised.
A new function, Py_CheckStackDepth() will be added, and the behavior of Py_EnterRecursiveCall() will be modified slightly.
int Py_CheckStackDepth(const char *where) will return 0 if there is no immediate danger of C stack overflow. It will return -1 and set an exception, if the C stack is near to overflowing. The where parameter is used in the exception message, in the same fashion as the where parameter of Py_EnterRecursiveCall().
Py_EnterRecursiveCall() will be modified to call Py_CheckStackDepth() before performing its current function.
Py_LeaveRecursiveCall() will remain unchanged.
This feature is fully backwards compatible at the Python level. Some low-level tools, such as machine-code debuggers, will need to be modified. For example, the gdb scripts for Python will need to be aware that there may be more than one Python frame per C frame.
C code that uses the Py_EnterRecursiveCall(), PyLeaveRecursiveCall() pair of functions will continue to work correctly. In addition, Py_EnterRecursiveCall() may raise a StackOverflow exception.
New code should use the Py_CheckStackDepth() function, unless the code wants to count as a Python function call with regard to the recursion limit.
We recommend that "python-like" code, such as Cython-generated functions, use Py_EnterRecursiveCall(), but other code use Py_CheckStackDepth().
It will no longer be possible to crash the CPython virtual machine through recursion.
It is unlikely that the performance impact will be significant.
The additional logic required will probably have a very small negative impact on performance. The improved locality of reference from reduced C stack use should have some small positive impact.
It is hard to predict whether the overall effect will be positive or negative, but it is quite likely that the net effect will be too small to be measured.
Gauging whether a C stack overflow is imminent is difficult. So we need to be conservative. We need to determine a safe bounds for the stack, which is not something possible in portable C code.
For major platforms, the platform specific API will be used to provide an accurate stack bounds. However, for minor platforms some amount of guessing may be required. While this might sound bad, it is no worse than the current situation, where we guess that the size of the C stack is at least 1000 times the stack space required for the chain of calls from _PyEval_EvalFrameDefault to _PyEval_EvalFrameDefault.
This means that in some cases the amount of recursion possible may be reduced. In general, however, the amount of recursion possible should be increased, as many calls will use no C stack.
Our general approach to determining a limit for the C stack is to get an address within the current C frame, as early as possible in the call chain. The limit can then be guessed by adding some constant to that.
Calls in the interpreter are handled by the CALL_FUNCTION, CALL_FUNCTION_KW, CALL_FUNCTION_EX and CALL_METHOD instructions. The code for those instructions will be modified so that when a Python function or method is called, instead of making a call in C, the interpreter will setup the callee's frame and continue interpretation as normal.
The RETURN_VALUE instruction will perform the reverse operation, except when the current frame is the entry frame of the interpreter when it will return as normal.
None, as yet.
None, as yet.
This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive.