Set: `BUILD_SET` opcode can be failed with segfault

[Eclips4]

cpython/Python/generated_cases.c.h

Lines 1648 to 1667 in 36b139a

	TARGET(BUILD_SET){
	PyObject **values = &PEEK(oparg);
	PyObject *set;
	set = PySet_New(NULL);
	int err = 0;
	for (int i = 0; i < oparg; i++){
	PyObject *item = values[i];
	if (err == 0)
	err = PySet_Add(set, item);
	Py_DECREF(item);
	}
	if (err != 0){
	Py_DECREF(set);
	if (true){STACK_SHRINK(oparg); goto error}
	}
	STACK_SHRINK(oparg);
	STACK_GROW(1);
	POKE(1, set);
	DISPATCH();
	}

&

cpython/Python/bytecodes.c

Lines 1303 to 1316 in 36b139a

	inst(BUILD_SET, (values[oparg] -- set)){
	set = PySet_New(NULL);
	int err = 0;
	for (int i = 0; i < oparg; i++){
	PyObject *item = values[i];
	if (err == 0)
	err = PySet_Add(set, item);
	Py_DECREF(item);
	}
	if (err != 0){
	Py_DECREF(set);
	ERROR_IF(true, error);
	}
	}

Doesn't take in account case, when PySet_New(NULL) returns NULL.
We are checking that PySet_Add doesn't return a non-zero(-1) value.
But, PySet_Add has a check, that first argument is a subclass of set. Which fails, if we will pass (PyObject *) NULL as first argument. Why?

#define PySet_Check(ob) \
 (Py_IS_TYPE((ob), &PySet_Type) || \
 PyType_IsSubtype(Py_TYPE(ob), &PySet_Type))

PySet_Add uses this macross. But, Py_TYPE will be failed with segfault when try to access ob_type of (PyObject *) NULL.

Implementation of Py_TYPE:

static inline PyTypeObject* Py_TYPE(PyObject *ob){
 return ob->ob_type;
}

(gdb) call (PyObject *) NULL
$1 = (PyObject *) 0x0
(gdb) call $1->ob_type
Cannot access memory at address 0x8

So, we should add check, that value of PySet_New is not-null.

Linked PRs

• gh-101952: Fix possible segfault in BUILD_SET opcode #101958

[kumaraditya303]

cc @gvanrossum Appears to be introduced in #100912

[gvanrossum]

Yeah, the original code had

if (set == NULL)
 goto error;

which somehow got lost.

How can I repro this?

[gvanrossum]

If someone could come up with a fix I'll review it, that's the quickest way.

[Eclips4]

I can fix it, but i don't know yet how to reproduce it. I was just looking through the bytecodes.c file =)

[gvanrossum]

It's probably hard to repro, because the only way PySet_New() can error is if it runs out of memory. I think restoring those two lines will do it (nothing has been popped off the stack at that point yet).

[Eclips4]

Okay, i'll soon send a PR