Possibly wrong readers count in boost::thread_v2::upgrade_mutex

[daniel-boehme]

Because of issues (#265, #361) in the default shared mutex implementation, I reviewed the code of the V2 mutex classes.

It seems like upgrade_mutex::try_unlock_shared_and_lock_until(const boost::chrono::time_point<Clock, Duration>& abs_time) and upgrade_mutex::try_unlock_upgrade_and_lock_until(const boost::chrono::time_point<Clock, Duration>& abs_time) can set the readers count to a wrong value.

Both functions contain the same piece of code:

      count_t num_readers = (state_ & n_readers_) - 1;
      state_ &= ~n_readers_;
      state_ |= (write_entered_ | num_readers);
      if (!gate2_.wait_until(lk, abs_time, boost::bind(
            &upgrade_mutex::no_readers, boost::ref(*this))))
      {
        ++num_readers;
        state_ &= ~(write_entered_ | n_readers_);
        state_ |= num_readers;

The problem is with ++num_readers. Between reading it out of the state and then setting it again, the mutex is unlocked, so readers can unlock_shared(). This is even necessary, as the code actually waits until the readers count reaches 0.

So IMHO, the line in question should read:

num_readers = (state_ & n_readers_) +1;