RFC: PEP 490 - Chain exceptions at C level
Hi, Here is the first draft of my PEP to chain automatically exceptions at C level in Python 3.5. Plain text version of the PEP below. HTML version of the PEP: https://www.python.org/dev/peps/pep-0490/ It has already an implementation, the pyerr_chain.patch file attached to http://bugs.python.org/issue23763 The patch is very short. We may modify more functions to hide the previous exception when the new exception contains as much or more information. Keeping the previous exception waste memory, make reference cycles more likely and is useless when displaying the full exception chain. I wrote a huge patch modifying a lot of functions to explicitly *not* chain exceptions: pyerr_match_clear-2.patch of the issue #23763. I wrote the patch to show where exceptions will be chained with pyerr_chain.patch applied. But we might apply it if we want to keep exactly the same behaviour between Python 3.4 and 3.5 for C modules the stdlib. pyerr_match_clear-2.patch adds also checks on the type of the previous exception to not replace an "unexpected" exception. For example, we can replace an expected TypeError with a ValueError (without chaining them), but we should not raise a ValueError if we got a KeyboardInterrupt, a MemoryError or other unexpected exceptions. But such changes are unrelated to the PEP and may be done in a separated issue. The starting point of my PEP is my work on CPython done with the pyfailmalloc module. This module injects MemoryError. I fixed a ton of issues in the code handling errors in CPython (see the issue #18408). I noticed that Python didn't complain when exceptions are lost. I added many assertions to detect these mistakes. Recently, I modified Python 3.5 to raise an exception even in release mode (issue #23571) if a C function returns NULL without an exception set, or if a C function returns a result with an exception set. Victor PEP: 490 Title: Chain exceptions at C level Version: $Revision$ Last-Modified: $Date$ Author: Victor Stinner <victor.stinner@gmail.com> Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 25-March-2015 Python-Version: 3.5 Abstract ======== Chain exceptions at C level, as already done at Python level. Rationale ========= Python 3 introduced a new killer feature: exceptions are chained by default, PEP 3134. Example:: try: raise TypeError("err1") except TypeError: raise ValueError("err2") Output:: Traceback (most recent call last): File "test.py", line 2, in <module> raise TypeError("err1") TypeError: err1 During handling of the above exception, another exception occurred: Traceback (most recent call last): File "test.py", line 4, in <module> raise ValueError("err2") ValueError: err2 Exceptions are chained by default in Python code, but not in extensions written in C. A new private ``_PyErr_ChainExceptions()`` function was introduced in Python 3.4.3 and 3.5 to chain exceptions. Currently, it must be called explicitly to chain exceptions and its usage is not trivial. Example of ``_PyErr_ChainExceptions()`` usage from the ``zipimport`` module to chain the previous ``OSError`` to a new ``ZipImportError`` exception:: PyObject *exc, *val, *tb; PyErr_Fetch(&exc, &val, &tb); PyErr_Format(ZipImportError, "can't open Zip file: %R", archive); _PyErr_ChainExceptions(exc, val, tb); This PEP proposes to also chain exceptions automatically at C level to stay consistent and give more information on failures to help debugging. The previous example becomes simply:: PyErr_Format(ZipImportError, "can't open Zip file: %R", archive); Proposal ======== Modify PyErr_*() functions to chain exceptions ---------------------------------------------- Modify C functions raising exceptions of the Python C API to automatically chain exceptions: modify ``PyErr_SetString()``, ``PyErr_Format()``, ``PyErr_SetNone()``, etc. Modify functions to not chain exceptions ---------------------------------------- Keeping the previous exception is not always interesting when the new exception contains information of the previous exception or even more information, especially when the two exceptions have the same type. Example of an useless exception chain with ``int(str)``:: TypeError: a bytes-like object is required, not 'type' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: int() argument must be a string, a bytes-like object or a number, not 'type' The new ``TypeError`` exception contains more information than the previous exception. The previous exception should be hidden. The ``PyErr_Clear()`` function can be called to clear the current exception before raising a new exception, to not chain the current exception with a new exception. Modify functions to chain exceptions ------------------------------------ Some functions save and then restore the current exception. If a new exception is raised, the exception is currently displayed into sys.stderr or ignored depending on the function. Some of these functions should be modified to chain exceptions instead. Examples of function ignoring the new exception(s): * ``ptrace_enter_call()``: ignore exception * ``subprocess_fork_exec()``: ignore exception raised by enable_gc() * ``t_bootstrap()`` of the ``_thread`` module: ignore exception raised by trying to display the bootstrap function to ``sys.stderr`` * ``PyDict_GetItem()``, ``_PyDict_GetItem_KnownHash()``: ignore exception raised by looking for a key in the dictionary * ``_PyErr_TrySetFromCause()``: ignore exception * ``PyFrame_LocalsToFast()``: ignore exception raised by ``dict_to_map()`` * ``_PyObject_Dump()``: ignore exception. ``_PyObject_Dump()`` is used to debug, to inspect a running process, it should not modify the Python state. * ``Py_ReprLeave()``: ignore exception "because there is no way to report them" * ``type_dealloc()``: ignore exception raised by ``remove_all_subclasses()`` * ``PyObject_ClearWeakRefs()``: ignore exception? * ``call_exc_trace()``, ``call_trace_protected()``: ignore exception * ``remove_importlib_frames()``: ignore exception * ``do_mktuple()``, helper used by ``Py_BuildValue()`` for example: ignore exception? * ``flush_io()``: ignore exception * ``sys_write()``, ``sys_format()``: ignore exception * ``_PyTraceback_Add()``: ignore exception * ``PyTraceBack_Print()``: ignore exception Examples of function displaying the new exception to ``sys.stderr``: * ``atexit_callfuncs()``: display exceptions with ``PyErr_Display()`` and return the latest exception, the function calls multiple callbacks and only returns the latest exception * ``sock_dealloc()``: log the ``ResourceWarning`` exception with ``PyErr_WriteUnraisable()`` * ``slot_tp_del()``: display exception with ``PyErr_WriteUnraisable()`` * ``_PyGen_Finalize()``: display ``gen_close()`` exception with ``PyErr_WriteUnraisable()`` * ``slot_tp_finalize()``: display exception raised by the ``__del__()`` method with ``PyErr_WriteUnraisable()`` * ``PyErr_GivenExceptionMatches()``: display exception raised by ``PyType_IsSubtype()`` with ``PyErr_WriteUnraisable()`` Backward compatibility ====================== A side effect of chaining exceptions is that exceptions store traceback objects which store frame objects which store local variables. Local variables are kept alive by exceptions. A common issue is a reference cycle between local variables and exceptions: an exception is stored in a local variable and the frame indirectly stored in the exception. The cycle only impacts applications storing exceptions. The reference cycle can now be fixed with the new ``traceback.TracebackException`` object introduced in Python 3.5. It stores informations required to format a full textual traceback without storing local variables. The ``asyncio`` is impacted by the reference cycle issue. This module is also maintained outside Python standard library to release a version for Python 3.3. ``traceback.TracebackException`` will maybe be backported in a private ``asyncio`` module to fix reference cycle issues. Alternatives ============ No change --------- A new private ``_PyErr_ChainExceptions()`` function is enough to chain manually exceptions. Exceptions will only be chained explicitly where it makes sense. New helpers to chain exceptions ------------------------------- Functions like ``PyErr_SetString()`` don't chain automatically exceptions. To make the usage of ``_PyErr_ChainExceptions()`` easier, new private functions are added: * ``_PyErr_SetStringChain(exc_type, message)`` * ``_PyErr_FormatChain(exc_type, format, ...)`` * ``_PyErr_SetNoneChain(exc_type)`` * ``_PyErr_SetObjectChain(exc_type, exc_value)`` Helper functions to raise specific exceptions like ``_PyErr_SetKeyError(key)`` or ``PyErr_SetImportError(message, name, path)`` don't chain exceptions. The generic ``_PyErr_ChainExceptions(exc_type, exc_value, exc_tb)`` should be used to chain exceptions with these helper functions. Appendix ======== PEPs ---- * `PEP 3134 -- Exception Chaining and Embedded Tracebacks <https://www.python.org/dev/peps/pep-3134/>`_ (Python 3.0): new ``__context__`` and ``__cause__`` attributes for exceptions * `PEP 415 - Implement context suppression with exception attributes <https://www.python.org/dev/peps/pep-0415/>`_ (Python 3.3): ``raise exc from None`` * `PEP 409 - Suppressing exception context <https://www.python.org/dev/peps/pep-0409/>`_ (superseded by the PEP 415) Python C API ------------ The header file ``Include/pyerror.h`` declares functions related to exceptions. Functions raising exceptions: * ``PyErr_SetNone(exc_type)`` * ``PyErr_SetObject(exc_type, exc_value)`` * ``PyErr_SetString(exc_type, message)`` * ``PyErr_Format(exc, format, ...)`` Helpers to raise specific exceptions: * ``PyErr_BadArgument()`` * ``PyErr_BadInternalCall()`` * ``PyErr_NoMemory()`` * ``PyErr_SetFromErrno(exc)`` * ``PyErr_SetFromWindowsErr(err)`` * ``PyErr_SetImportError(message, name, path)`` * ``_PyErr_SetKeyError(key)`` * ``_PyErr_TrySetFromCause(prefix_format, ...)`` Manage the current exception: * ``PyErr_Clear()``: clear the current exception, like ``except: pass`` * ``PyErr_Fetch(exc_type, exc_value, exc_tb)`` * ``PyErr_Restore(exc_type, exc_value, exc_tb)`` * ``PyErr_GetExcInfo(exc_type, exc_value, exc_tb)`` * ``PyErr_SetExcInfo(exc_type, exc_value, exc_tb)`` Others function to handle exceptions: * ``PyErr_ExceptionMatches(exc)``: check to implement ``except exc: ...`` * ``PyErr_GivenExceptionMatches(exc1, exc2)`` * ``PyErr_NormalizeException(exc_type, exc_value, exc_tb)`` * ``_PyErr_ChainExceptions(exc_type, exc_value, exc_tb)`` Python Issues ------------- Chain exceptions: * `Issue #23763: Chain exceptions in C <http://bugs.python.org/issue23763>`_ * `Issue #23696: zipimport: chain ImportError to OSError <http://bugs.python.org/issue23696>`_ * `Issue #21715: Chaining exceptions at C level <http://bugs.python.org/issue21715>`_: added ``_PyErr_ChainExceptions()`` * `Issue #18488: sqlite: finalize() method of user function may be called with an exception set if a call to step() method failed <http://bugs.python.org/issue18488>`_ * `Issue #23781: Add private _PyErr_ReplaceException() in 2.7 <http://bugs.python.org/issue23781>`_ * `Issue #23782: Leak in _PyTraceback_Add <http://bugs.python.org/issue23782>`_ Changes preventing to loose exceptions: * `Issue #23571: Raise SystemError if a function returns a result with an exception set <http://bugs.python.org/issue23571>`_ * `Issue #18408: Fixes crashes found by pyfailmalloc <http://bugs.python.org/issue18408>`_ Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End:
On 26 March 2015 at 22:30, Victor Stinner <victor.stinner@gmail.com> wrote:
Hi,
Here is the first draft of my PEP to chain automatically exceptions at C level in Python 3.5. Plain text version of the PEP below. HTML version of the PEP:
https://www.python.org/dev/peps/pep-0490/
It has already an implementation, the pyerr_chain.patch file attached to http://bugs.python.org/issue23763
The patch is very short.
I like the idea, but I think it would raise the priority of the proposed change to the display of chained exceptions in http://bugs.python.org/issue12535 That proposal suggests adding the line "<truncated: another exception occurred>" to the start of earlier tracebacks in the chain so you get an up front hint that a chained exception will be shown after the initial traceback.
Backward compatibility ======================
A side effect of chaining exceptions is that exceptions store traceback objects which store frame objects which store local variables. Local variables are kept alive by exceptions. A common issue is a reference cycle between local variables and exceptions: an exception is stored in a local variable and the frame indirectly stored in the exception. The cycle only impacts applications storing exceptions.
The reference cycle can now be fixed with the new ``traceback.TracebackException`` object introduced in Python 3.5. It stores informations required to format a full textual traceback without storing local variables.
Also potentially worth noting here is "traceback.clear_frames()", introduced in 3.4 to explicitly clear local variables from a traceback: https://docs.python.org/3/library/traceback.html#traceback.clear_frames PEP 442 also makes it more likely that cycles will be cleared properly, even if some of the objects involved have __del__ methods: https://www.python.org/dev/peps/pep-0442/ However, I'll grant that neither of those can be backported the way that traceback.TracebackException can be. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
participants (2)
-
Nick Coghlan
-
Victor Stinner