Thinking In C++. Volume 2: Practical Programming

All the communication between the Entrance objects takes place through the single Count object. When the user presses , main( ) sends the pause( ) message to count. Since each Entrance::run( ) is watching the count object to see whether it is paused, this causes each Entrance to move into the waitingForCancel state, where it is no longer counting, but it is still alive. This is essential because main( ) must still be able to safely iterate over each object in the vector. Note that because there is a slight possibility that the iteration might occur before an Entrance has finished counting and moved into the waitingForCancel state, the getValue( ) function cycles through calls to sleep( ) until the object moves into waitingForCancel. (This is one form of what is called a busy wait, which is undesirable. You’ll see the preferred approach of using wait( ) later in the chapter.) Once main( ) completes its iteration through the vector, the cancel( ) message is sent to the count object, and once again all the Entrance objects are watching for this state change. At this point, they print a termination message and exit from run( ), which causes each task to be destroyed by the threading mechanism.

As this program runs, you will see the total count and the count at each entrance displayed as people walk through a turnstile. If you comment out the Guard object in Count::increment( ), you’ll notice that the total number of people is not what you expect it to be. The number of people counted by each turnstile will be different from the value in count. As long as the Mutex is there to synchronize access to the Counter, things work correctly. Keep in mind that Count::increment( ) exaggerates the potential for failure by using temp and yield( ). In real threading problems, the possibility for failure may be statistically small, so you can easily fall into the trap of believing that things are working correctly. Just as in the example above, there are likely to be hidden problems that haven’t occurred to you, so be exceptionally diligent when reviewing concurrent code.

Atomic operations

Note that Count::value( ) returns the value of count using a Guard object for synchronization. This brings up an interesting point, because this code will probably work fine with most compilers and systems without synchronization. The reason is that, in general, a simple operation such as returning an int will be an atomic operation, which means that it will probably happen in a single microprocessor instruction that will not get interrupted. (The multithreading mechanism is unable to stop a thread in the middle of a microprocessor instruction.) That is, atomic operations are not interruptible by the threading mechanism and thus do not need to be guarded.[126] In fact, if we removed the fetch of count into temp and removed the yield( ), and instead simply incremented count directly, we probably wouldn’t need a lock at all because the increment operation is usually atomic, as well.

The problem is that the C++ standard doesn’t guarantee atomicity for any of these operations. Although operations such as returning an int and incrementing an int are almost certainly atomic on most machines, there’s no guarantee. And because there’s no guarantee, you have to assume the worst. Sometimes you might investigate the atomicity behavior on a particular machine (usually by looking at assembly language) and write code based on those assumptions. That’s always dangerous and ill-advised. It’s too easy for that information to be lost or hidden, and the next person that comes along may assume that this code can be ported to another machine and then go mad tracking down the occasional glitch caused by thread collisions.