classification
Title: Prevent uses of format string based PyObject_Call* that do not produce tuple required by docs
Type: Stage:
Components: Interpreter Core Versions: Python 3.6, Python 3.5
process
Status: open Resolution:
Dependencies: Superseder:
Assigned To: Nosy List: jkloth, josh.r, vstinner
Priority: normal Keywords:

Created on 2016-04-21 18:36 by josh.r, last changed 2016-04-22 00:26 by jkloth.

Messages (3)
msg263929 - (view) Author: Josh Rosenberg (josh.r) * (Python triager) Date: 2016-04-21 18:36
PyObject_CallMethod explicitly documents that "The C arguments are described by a Py_BuildValue() format string that should produce a tuple." While PyObject_CallFunction doesn't document this requirement, it has the same behavior, and the same failures, as does the undocumented _PyObject_CallMethodId.

The issue is that, should someone pass a format string of "O", then the type of the subsequent argument determines the effect in a non-obvious way; when the argument comes from a caller, and the intent was to pass a single argument, this means that if the caller passes a non-tuple sequence, everything works, while passing a tuple tries to pass the contents of the tuple as sequential arguments. This inconsistency was the cause of both #26478 and #21209 (maybe others).

Assuming the API can't/shouldn't be changed, it should still be an error when a format string of "O" is passed and the argument is a non-tuple (because you've violated the spec; the result of BuildValue was not a tuple). Instead call_function_tail is silently rewrapping non-tuple args in a single element tuple.

I'm proposing that, in debug builds (and ideally release builds too), abstract.c's call_function_tail treat the "non-tuple" case as an error, rather than rewrapping in a single element tuple. This still allows the use case where the function is used inefficiently, but correctly (where the format string is "O" and the value is *always* a tuple that's supposed to be varargs; it should really just use PyObject_CallFunctionObjArgs/PyObject_CallMethodObjArgs/PyObject_CallObject or Id based optimized versions, but it's legal). But it will make the majority of cases where a user provided argument could be tuple or not fail fast, rather than silently behave themselves *until* they receive a tuple and misbehave.

Downside: It will require code changes for cases like PyObject_CallFunction(foo, "k", myunsigned);, where there was no risk of misbehavior, but those cases were also violating the spec, and should be fixable by changing the format string to wrap the single value in parens, e.g. "(k)".
msg263930 - (view) Author: Josh Rosenberg (josh.r) * (Python triager) Date: 2016-04-21 18:42
The motivation for this change was Mr. STINNER's comment on #26814 ( https://bugs.python.org/issue26814#msg263923 ), where he mentioned the weirdness of PyObject_CallFunction and friends, which complicates the implementation of PyObject_FastCall and alerted me to a second case ( #21209 ) in which this silent fix up has caused confusing issues in CPython (I filed #26478 so I recognized the issue).

If this fix could be made, it might be possible to eventually make the check for non-tuple arguments a debug build only check (or a check only on public APIs), allowing the implementation in release mode and/or internal APIs to avoid the work involved in constantly checking for and performing this workaround to fix doc violating code, and possible simplify PyObject_FastCall by removing the corner case.
msg263945 - (view) Author: Jeremy Kloth (jkloth) * Date: 2016-04-22 00:26
IMHO, this is a documentation bug with PyObject_CallMethod.  The change to its documentation to differ from PyObject_CallFunction was changed back in 2004.  It should have been updated then to reflect the already well-entrenched behavior of those 2 (at the time) functions, but that ship has sailed.

It would be a huge mistake to change how they handle the format strings now as they have operated they way they have since at least Python 1.5.2 (when I learned Python's C API).  There is just too much C code that would potentially break (third-party C extensions).

I think that a good change for the docs would be to separate the specification of the "building format string" away from the Py_BuildValue function so as to allow for differences in how the resulting object is created in those C functions that use that specification.  Similar, I suppose, to how the "parsing format string" is defined independently from PyArg_ParseTuple.

Now to the "attractive nuisance" that the single argument passed as varargs is handled.  I believe it may be best to introduce a new format character just for this purpose.  Possibly "T" (for tuple) or "V" (for varargs) using uppercase as that seems to be the practice for referencing objects.  And at the same time, add a check in the "call" functions (those that *use* Py_VaBuildValue to create a argument tuple, of which there are also internal ones).  The check could be as simple as:

  if (format[0] == 'O' && format[1] == '\0') {
    va_start(va, format);
    PyObject *ob = (PyObject *)va_arg(va, PyObject *);
    if (ob != NULL) {
      if (PyTuple_Check(ob)) {
        if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1,
                "use 'V' format code to support varargs")) {
          args = NULL;
        }
        else {
          args = ob;
        }
      }
      else {
        args = PyTuple_Pack(1, ob);
      }
    }
    else if (!PyErr_Occurred()) {
      PyErr_SetString(PyExc_SystemError, "argument is NULL");
      args = NULL;
    }
    va_end(va);
  }
  else {
    args = //...whatever happens now...
  }
History
Date User Action Args
2016-04-22 00:26:44jklothsetnosy: + jkloth
messages: + msg263945
2016-04-21 18:42:08josh.rsetmessages: + msg263930
2016-04-21 18:36:47josh.rcreate