Remove incorrect reference count decrement for dim.

[wr-web]

Related issue

#1021

Description

A reference to item is incorrectly decremented after a failed PyTuple_SetItem call, leading to a potential use-after-free vulnerability or double-decrement crash.

Affected Code

DeepRec/tensorflow/python/eager/pywrap_tensor.cc

Lines 593 to 598 in d1c5a6e

	if (MaybeRaiseExceptionFromTFStatus(self->status, PyExc_ValueError) ||
	dim == nullptr || PyTuple_SetItem(shape, i, dim) != 0) {
	// Cleanup self->status before returning.
	TF_SetStatus(self->status, TF_OK, "");
	Py_DECREF(shape);
	if (dim != nullptr) Py_DECREF(dim);

Root Cause

PyTuple_SetItem Source Code

PyTuple_SetItem Document

PyList_SetItem Python Forum Discussion

Therefore, whether the function succeeds or fails, it will steal a reference count of the third argument. It must not call Py_XDECREF/Py_DECREF in case of failure.

Evidence：Reference Counting and Ownership in CPython Native API

Borrowed Reference

A borrowed reference is a reference obtained from an object that you don't own. You don't need to decrement its reference count when you're done with it, but you must ensure the object stays alive while you're using it (e.g., by creating an owned reference with Py_INCREF if necessary). Borrowed references are typically returned by functions like PyList_GetItem(), which returns an item from a list without incrementing its reference count.

New Reference

A new reference (also called an "owned reference") is a reference that you have ownership of. When you receive a new reference from a function (such as PyObject_New() or Py_BuildValue()), you are responsible for calling Py_DECREF() on it when you no longer need it to properly decrement its reference count. Failure to do so causes memory leaks.

Stolen Reference (Stealing)

A stolen reference is when a function takes ownership of a reference you pass to it. When you pass an object reference to a function that "steals" it, you no longer own that reference, and you should not call Py_DECREF() on it afterward. The function assumes full responsibility for managing the reference count of that object.

Example: PyTuple_SetItem

PyTuple_SetItem is a classic example of a reference-stealing function. According to the documentation:

Note This function “steals” a reference to o and discards a reference to an item already in the tuple at the affected position.

However, a critical ambiguity in the documentation is that it does not clearly state whether the reference is stolen in the case of function failure. This is fundamentally an all-or-nothing problem: when you pass an object to a stealing function, you must understand whether ownership is unconditionally transferred or only transferred on success.

Looking at the CPython source code clarifies this behavior:

int
PyTuple_SetItem(PyObject *op, Py_ssize_t i, PyObject *newitem)
{
    PyObject **p;
    if (!PyTuple_Check(op) || !_PyObject_IsUniquelyReferenced(op)) {
        Py_XDECREF(newitem);
        PyErr_BadInternalCall();
        return -1;
    }
    if (i < 0 || i >= Py_SIZE(op)) {
        Py_XDECREF(newitem);
        PyErr_SetString(PyExc_IndexError,
                        "tuple assignment index out of range");
        return -1;
    }
    p = ((PyTupleObject *)op) -> ob_item + i;
    Py_XSETREF(*p, newitem);
    return 0;
}

As demonstrated in the source code above, PyTuple_SetItem unconditionally calls Py_XDECREF(newitem) when the type check fails—meaning it always steals the reference, even on failure. The function takes ownership of newitem regardless of whether it successfully inserts the item into the tuple.

This behavior has serious implications for correct API usage. Consider the following incorrect code:

if (PyTuple_SetItem(xxx, yyy, something) < 0) {
    Py_DECREF(something);  // DANGER: Use-After-Free!
}

This code is defective because it leads to a use-after-free vulnerability. Since PyTuple_SetItem already stole the reference (and decremented it on failure via Py_XDECREF), the additional Py_DECREF(something) in the error-handling block causes a double decrement, potentially leading to an assertion failure in debug builds or memory corruption and crashes in release builds.

The correct pattern is simply:

if (PyTuple_SetItem(a, b, something) < 0) {
    // Do NOT call Py_DECREF on 'something' - the reference was already stolen
    return NULL;  // or handle error appropriately
}

In summary, when dealing with stealing functions in the CPython API, you must relinquish all ownership responsibility for the passed reference and never decrement it after the call, regardless of the return value. Always consult the source code or thoroughly documented behavior to confirm whether a function truly provides an all-or-nothing stealing guarantee.