Use After Free when assigning into a memoryview

[chilaxan]

Bug report

within memoryview.c, I have found two Use After Frees, both based around memory_ass_sub.
The first is if a class with a malicious __index__ method is used as the index for the assignment, its index method is called after the memoryview is checked if it is released. This allows the index method to release the memory view and backing buffer, leading to a write to freed memory when the write completes. The same vuln exists if the class with a malicious index method is used as the assigned value, as its __index__ method is called inside of pack_single

# memoryview Use After Free (memory_ass_sub)
uaf_backing = bytearray(bytearray.__basicsize__)
uaf_view = memoryview(uaf_backing).cast('n') # ssize_t format

class weird_index:
    def __index__(self):
        global memory_backing
        uaf_view.release() # release memoryview (UAF)
        # free `uaf_backing` memory and allocate a new bytearray into it
        memory_backing = uaf_backing.clear() or bytearray()
        return 2 # `ob_size` idx

# by the time this line finishes executing, it writes the max ptr size
# into the `ob_size` slot of `memory_backing`
uaf_view[weird_index()] = (2 ** (tuple.__itemsize__ * 8) - 1) // 2
memory = memoryview(memory_backing)
memory[id(250) + int.__basicsize__] = 100
print(250) # prints 100

Your environment

• CPython versions tested on: Python 3.10.2 (main, Feb 2 2022, 07:36:01) [Clang 12.0.0 (clang-1200.0.32.29)] on darwin
• Operating system and architecture: MacOS, 64bit

[chilaxan]

Tested on 3.12.0 built with ./configure --enable-optimizations

[Fidget-Spinner]

@vstinner would you class this as a security vuln or just a user bug? IMO it's definitely a bug that we should fix on our end considering memorview promises to raise ValueError after .release. But there exists many ways to corrupt memory in CPython and write to arbitrary addresses (e.g. ctypes or compiling your own bytecode with malicious LOAD_FAST instructions). So I'm not sure if you would class this as a vuln?

[Fidget-Spinner]

CC @serhiy-storchaka too for your opinion please.

[serhiy-storchaka]

Nice. Many thanks to you @chilaxan for the reproducer. It only works on non-debug build, but I am sure that it is possible to get the same result on a debug build if slightly change the code.

This is definitely a serious bug, but I am not sure that it can be classified as a practical security vulnerability. Of course the bug can manifest not only with malicious __index__ (it is just a convenient way to reproduce it consistently), but with any __index__ implemented in Python if you use memoryview and multithreading. You only need to be exceptionally (un)lucky to get it. In theory the attacker can attack the program which have all three components (threads, writing to memoryview and objects with Python implemented __index__), but it is pure hypothetical scenario.

[kumaraditya303]

See also #91153 which is a similar bug but with bytearray.

[gvanrossum]

We should fix both but once an attacker can execute this code they have full control already, so not a vulnerability.

[serhiy-storchaka]

Here are simpler examples (they work on debug build):

size = 100
class MyIndex:
    def __index__(self):
        m.release()
        global ba
        ba = bytearray(size)
        return 4

ba = None
m = memoryview(bytearray(b'\xff'*size))
m[MyIndex()] = 42
print(ba[4])  # prints 42, should be 0

ba = None
m = memoryview(bytearray(b'\xff'*size))
m[:MyIndex()] = b'spam'
print(ba[:4])  # prints bytearray(b'spam'), should be bytearray(b'\x00\x00\x00\x00')

ba = None
m = memoryview(bytearray(b'\xff'*size))
m[1] = MyIndex()
print(ba[1])  # prints 4, should be 0

[vstinner]

@vstinner would you class this as a security vuln or just a user bug?

If you ask me, the exploit makes the assumption than a attacker can execute arbitrary Python code. It doesn't fit into Python Security Model: https://python-security.readthedocs.io/security.html#python-security-model

I consider this issue as a regular bug.

Python doesn't provide any warranty that it's impossible to crash Python with malicious code. Using ctypes you can read/write arbitrary memory.