bug in pthreads implementation of recursive_mutex

From: Feldman, Yulik (yulik.feldman_at_[hidden])
Date: 2003-05-03 11:10:06

Hi,

 

It looks that there is a bug in
.../boost_1_30_0/libs/thread/src/recursive_mutex.cpp. The bug shows in the
pthreads implementation of the recursive_mutex when the mutex is used with
the "condition" synchronization primitive. Here is the description of the
bug:

 

When the client code calls wait() function on a condition sending as a
parameter a lock object associated with a recursive_mutex
("m_Condition.wait(Lock)"), the following functions get called:

 

boost::condition::wait(L& Lock) at condition.hpp:91

boost::condition::do_wait(M& mutex) at condition.hpp:137

boost::detail::thread::lock_ops<boost::recursive_mutex>::unlock(Mutex& m,
lock_state& state) at lock.hpp:54

boost::recursive_mutex::do_unlock(cv_state& state)

 

The implementation of the last function on "winthreads" and "mptasks" resets
the 'm_count' data member to zero. The "pthreads" implementation at
recursive_mutex.cpp:399 doesn't reset the 'm_count' to zero. As a result,
when the calling thread enters pthread_cond_wait() and the mutex is unlocked
by the pthreads the 'm_count' still remains non-zero. Starting from this
point, the things may go bad. Consider, for example, that another thread
tries to acquire the mutex. The thread calls:

 

boost::recursive_mutex::scoped_lock Lock(m_Mutex)

boost::detail::thread::scoped_lock<boost::recursive_mutex>::scoped_lock<boos
t::recursive_mutex>(Mutex& mx, initially_locked = true) at lock.hpp:66

boost::detail::thread::scoped_lock<boost::recursive_mutex>::lock() at
lock.hpp:76

boost::detail::thread::lock_ops<boost::recursive_mutex>::lock(Mutex& m) at
lock.hpp:34

boost::recursive_mutex::do_lock() at recursive_mutex.cpp:307

 

The last function locks the mutex again, and, if
BOOST_HAS_PTHREAD_MUTEXATTR_SETTYPE is defined, increments the counter to 2.
Now, since the count is more than 1, the function unlocks the mutex. As a
result, the mutex remains unlocked, after the lock function is called. 

 

The same problem exists in recursive_try_mutex and recursive_timed_mutex.

 

I've tried to fix the code as follows, which seemed to help in my particular
case:

 

----------------------------------------------------------------------------
------------------------------------

*** recursive_mutex.cpp.orig    Sun Feb 23 19:38:11 2003

--- recursive_mutex.cpp Sat May  3 17:32:44 2003

***************

*** 412,417 ****

--- 412,418 ----

  

      state.pmutex = &m_mutex;

      state.count = m_count;

+     m_count = 0;

  }

  

  recursive_try_mutex::recursive_try_mutex()

***************

*** 608,613 ****

--- 609,615 ----

  

      state.pmutex = &m_mutex;

      state.count = m_count;

+     m_count = 0;

  }

  

  recursive_timed_mutex::recursive_timed_mutex()

***************

*** 787,792 ****

--- 789,795 ----

  

      state.pmutex = &m_mutex;

      state.count = m_count;

+     m_count = 0;

  }

  #elif defined(BOOST_HAS_MPTASKS)

----------------------------------------------------------------------------
------------------------------------

 

Any comments? Is this indeed a bug? Is the fix correct?

 

Thanks,

            Yulik.

 

[Non-text portions of this message have been removed]