Use-after-free in _Unpickler_ReadIntoFromFile: temporary memoryview passed to readinto() can outlive its buffer
還沒有人認領這個 Issue。
- 主要語言
- Python
- 星號
- 77.2k
- 分支
- 36k
- PR 合併指標
- PR 指標待擷取
描述
Bug report
Bug description
The C implementation of pickle.Unpickler, when reading from a file-like
object that provides readinto(), hands that method a temporary memoryview
created over an internal buffer:
PyObject *buf_obj = PyMemoryView_FromMemory(buf, n, PyBUF_WRITE);
...
PyObject *read_size_obj = _Pickle_FastCall(self->readinto, buf_obj);
(Modules/_pickle.c, _Unpickler_ReadIntoFromFile.)
buf points into a short-lived buffer (e.g. the bytes object allocated in
load_counted_binbytes, which may also be reallocated by _PyBytes_Resize or
freed when unpickling ends). The memoryview is never released or invalidated
after readinto() returns. A readinto() implementation that keeps a
reference to the view can therefore use it to read or write the buffer after
it has been freed, which is a use-after-free at the C level.
This only requires a pure-Python file-like object -- no ctypes. A pure-Python
program should not be able to make the interpreter read or write freed memory.
Reproducer
import pickle, struct, gc
stashed = []
class EvilFile:
def __init__(self):
self._h = b"\x80\x05" + b"\x8e" + struct.pack("<Q", 200_000)
self._p = 0
def read(self, n=-1):
d = self._h[self._p:] if (n is None or n < 0) else self._h[self._p:self._p+n]
self._p += len(d); return d
def readline(self):
return self.read(-1)
def readinto(self, view):
stashed.append(view) # keep the view past readinto()
view[:] = b"A" * len(view); return len(view)
up = pickle.Unpickler(EvilFile())
try:
up.load() # stream ends after the payload
except EOFError:
pass
del up; gc.collect() # free the backing buffer
_ = [bytes(200_000) for _ in range(8)] # churn the allocator
stashed[0][0] # <-- use-after-free read
On a --with-address-sanitizer --with-pydebug build this reports a clean
heap-use-after-free (READ in unpack_single, the buffer freed via
Pdata_dealloc and originally allocated in load_counted_binbytes).
Root cause and fix direction
The view is a non-owning window over a raw pointer that the unpickler does not
keep alive. Other CPython sites that drive a user readinto() (e.g.
_io.RawIOBase.read()) hand out an owning object (a bytearray) whose buffer
protocol prevents it from being freed while still exported. The pickle path
should release the temporary memoryview as soon as readinto() returns, so a
surviving reference raises ValueError: operation forbidden on released memoryview object instead of dereferencing freed memory.
I have a patch with a regression test and will open a PR.
(For context: this was originally raised privately with the Python Security
Response Team, who advised opening a public issue.)
CPython versions tested on
3.16.0a0 (main, commit 5755d0f).
Operating systems tested on
macOS (arm64), --with-address-sanitizer --with-pydebug build.
Linked PRs
- gh-151048
貢獻指南
從這裡開始
- 先讀完整個 Issue,再讀專案的貢獻指南。
- 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
- Fork 儲存庫,在一個分支上完成修改。
- 送出 Pull Request,並在描述裡引用這個 Issue 編號。
研究方向
從 Modules/_pickle.c 中的 _Unpickler_ReadIntoFromFile 開始,檢查 readinto() 返回後暫時 memoryview 的處理方式。在 ASAN、pydebug 建置中使用提供的 EvilFile 重現程式,然後新增或檢查提到的回歸測試。當保留的檢視無法存取已釋放的緩衝區並引發預期的已釋放 memoryview 錯誤時,即表示完成。
由索引模型根據 Issue 內容生成。
評估
- 技術堆疊
- c, python
- 領域
- backend, security
- Issue 類型
- 缺陷
- 難度
- 4/5
- 預估耗時
- 3-5 天
- 活躍度
- 停滯
- 描述清晰度
- 描述清楚
- 新手友好度
- 20/100