[New-bugs-announce] [issue38106] Race in PyThread_release_lock - can lead to MEMORY CORRUPTION and DEADLOCK

[Kirill Smelkov]

New submission from Kirill Smelkov <kirr at nexedi.com>:

Hello up there. I believe I've discovered a race in PyThread_release_lock on
Python2.7 that, on systems where POSIX semaphores are not available and Python
locks are implemented with mutexes and condition variables, can lead to MEMORY
CORRUPTION and DEADLOCK. The particular system I've discovered the bug on is
macOS Mojave 10.14.6.

The bug is already fixed on Python3 and the fix for Python2 is easy:

    git cherry-pick 187aa545165d

Thanks beforehand,
Kirill


Bug description

( Please see attached pylock_bug.pyx for the program that triggers the bug for real. )

On Darwin, even though this is considered as POSIX, Python uses
mutex+condition variable to implement its lock, and, as of 20190828, Py2.7
implementation, even though similar issue was fixed for Py3 in 2012, contains
synchronization bug: the condition is signalled after mutex unlock while the
correct protocol is to signal condition from under mutex:

  https://github.com/python/cpython/blob/v2.7.16-127-g0229b56d8c0/Python/thread_pthread.h#L486-L506
  https://github.com/python/cpython/commit/187aa545165d (py3 fix)

PyPy has the same bug for both pypy2 and pypy3:

  https://bitbucket.org/pypy/pypy/src/578667b3fef9/rpython/translator/c/src/thread_pthread.c#lines-443:465
  https://bitbucket.org/pypy/pypy/src/5b42890d48c3/rpython/translator/c/src/thread_pthread.c#lines-443:465

Signalling condition outside of corresponding mutex is considered OK by
POSIX, but in Python context it can lead to at least memory corruption if we
consider the whole lifetime of python level lock. For example the following
logical scenario:

      T1                                          T2

  sema = Lock()
  sema.acquire()

                                              sema.release()

  sema.acquire()
  free(sema)

  ...


can translate to the next C-level calls:

      T1                                          T2

  # sema = Lock()
  sema = malloc(...)
  sema.locked = 0
  pthread_mutex_init(&sema.mut)
  pthread_cond_init (&sema.lock_released)

  # sema.acquire()
  pthread_mutex_lock(&sema.mut)
  # sees sema.locked == 0
  sema.locked = 1
  pthread_mutex_unlock(&sema.mut)


                                              # sema.release()
                                              pthread_mutex_lock(&sema.mut)
                                              sema.locked = 0
                                              pthread_mutex_unlock(&sema.mut)

                      # OS scheduler gets in and relinquishes control from T2
                      # to another process
                                              ...

  # second sema.acquire()
  pthread_mutex_lock(&sema.mut)
  # sees sema.locked == 0
  sema.locked = 1
  pthread_mutex_unlock(&sema.mut)

  # free(sema)
  pthread_mutex_destroy(&sema.mut)
  pthread_cond_destroy (&sema.lock_released)
  free(sema)


  # ...
  e.g. malloc() which returns memory where sema was

                                              ...
                      # OS scheduler returns control to T2
                      # sema.release() continues
                      #
                      # BUT sema was already freed and writing to anywhere
                      # inside sema block CORRUPTS MEMORY. In particular if
                      # _another_ python-level lock was allocated where sema
                      # block was, writing into the memory can have effect on
                      # further synchronization correctness and in particular
                      # lead to deadlock on lock that was next allocated.
                                              pthread_cond_signal(&sema.lock_released)

Note that T2.pthread_cond_signal(&sema.lock_released) CORRUPTS MEMORY as it
is called when sema memory was already freed and is potentially
reallocated for another object.

The fix is to move pthread_cond_signal to be done under corresponding mutex:

  # sema.release()
  pthread_mutex_lock(&sema.mut)
  sema.locked = 0
  pthread_cond_signal(&sema.lock_released)
  pthread_mutex_unlock(&sema.mut)

by cherry-picking commit 187aa545165d ("Signal condition variables with the
mutex held. Destroy condition variables before their mutexes").


Bug history

The bug was there since 1994 - since at least [1]. It was discussed in 2001
with original code author[2], but the code was still considered to be
race-free. In 2010 the place where pthread_cond_signal should be - before or
after pthread_mutex_unlock - was discussed with the rationale to avoid
threads bouncing[3,4,5], and in 2012 pthread_cond_signal was moved to be
called from under mutex, but only for CPython3[6,7].

In 2019 the bug was (re-)discovered while testing Pygolang[8] on macOS with
CPython2 and PyPy2 and PyPy3.

[1] https://github.com/python/cpython/commit/2c8cb9f3d240
[2] https://bugs.python.org/issue433625
[3] https://bugs.python.org/issue8299#msg103224
[4] https://bugs.python.org/issue8410#msg103313
[5] https://bugs.python.org/issue8411#msg113301
[6] https://bugs.python.org/issue15038#msg163187
[7] https://github.com/python/cpython/commit/187aa545165d
[8] https://pypi.org/project/pygolang

----------
components: Interpreter Core, macOS
files: pylock_bug.pyx
messages: 351826
nosy: arigo, beazley, benjamin.peterson, brian.curtin, dabeaz, flox, gvanrossum, jyasskin, kristjan.jonsson, loewis, navytux, ned.deily, paul.moore, pitrou, python-dev, r.david.murray, rhettinger, ronaldoussoren, sbt, scoder, shihao, techtonik, terry.reedy, tim.golden, tim.peters, torsten, vstinner, ysj.ray
priority: normal
severity: normal
status: open
title: Race in PyThread_release_lock - can lead to MEMORY CORRUPTION and DEADLOCK
type: behavior
versions: Python 2.7
Added file: https://bugs.python.org/file48604/pylock_bug.pyx

_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue38106>
_______________________________________