Heap buffer overflow in `Packer(buf_size=…)` with an unstable `__index__`

[marinelay]

I found a heap buffer overflow fuzzing the C extension. The input is contrived (it needs an object whose __index__ answers differently each call) but sharing it in case it's useful.

Versions

msgpack 1.2.1 (pip install msgpack), CPython 3.12, Linux x86_64. The same
code is on main.

Reproducer

import msgpack

class Unstable:
    """__index__ answers with a small size first and a large one afterwards."""
    def __init__(self, first, rest):
        self._v = [first, rest]
        self._n = 0
    def __index__(self):
        v = self._v[min(self._n, 1)]
        self._n += 1
        return v
    __int__ = __index__

msgpack.Packer(buf_size=Unstable(600, 1 << 20)).pack(b"A" * 100000) # SIGSEGV

Output

[1]    123498 segmentation fault (core dumped)  python

Sanitizer build

AddressSanitizer: heap-buffer-overflow
    #1 msgpack_pack_write      msgpack/pack.h:53:5
    #2 msgpack_pack_raw_body   msgpack/pack_template.h:488:16
    #3 Packer._pack_inner      _cmsgpack.c:7137:18
    #4 Packer._pack            _cmsgpack.c:8391:15
  allocated by:
    #1 Packer.__cinit__        _cmsgpack.c:5931

What happens

__cinit__ uses buf_size at two lines, and Cython converts it separately at each, so the object is asked for its value once per line:

def __cinit__(self, buf_size=256*1024, **_kwargs):
    self.pk.buf = <char*> PyMem_Malloc(buf_size)   # :114  -> 600
    if self.pk.buf == NULL:
        raise MemoryError("Unable to allocate internal buffer.")
    self.pk.buf_size = buf_size                    # :117  -> 1048576

(_packer.pyx:114, :117)

With Unstable(600, 1 << 20) the packer holds 600 bytes and records 1048576 as its capacity, and everything afterwards works from the recorded number:

if (len + l > bs) {          /* pack.h:45 — bs = pk.buf_size = 1048576 */
    ...grow...
}
memcpy(buf + len, data, l);

100000 is under 1048576, so the buffer never grows and the memcpy writes 100000 bytes into a 600-byte block.
(pack.h:45)

---

It reads like a bug to me, but the input is not a natural one, so I'm leaving it here.
Entirely reasonable if this is out of scope.