Question

I’ve been asked the following question and I’m not sure what the correct answer is to it:

If monitors are implemented by replacing condition variables with semaphores 
(counters set to 0) with down() and up() as wait and signal, respectively,
would the monitors work correctly?

I’d be tempted to say it is a correct implementation because semaphores and condition variables can replace each other, correct? Is there a better explanation?

Answer 1

You are asking about a semaphore initialized to 1, which is also called a binary semaphore.

The answer depends on a particular implementation (or definition) of these primitives, however a typical difference is that monitors have thread ownership and semaphores do not. This impacts a variety of scenarios.

• For a semaphore, it is completely normal that one thread does all the ups and another one does all the downs. This should not be allowed by monitors.
• Suppose you have a class with private state and multiple public methods, all of which lock a common monitor upon entry and unlock upon exit in order to protect that state. Suppose also that public method A internally calls public method B. Then a recursive monitor will correctly allow calling method A (which involves a sequence lock-lock-unlock-unlock) whereas the same with a semaphore-implemented monitor would result in a deadlock upon the second attempt to lock, using the same thread.
• Suppose that a thread acquires a named system wide monitor and crashes. Because of thread ownership, it is possible to automatically unlock the monitor and allow any waiting threads to acquire it. With a binary semaphore this is not possible, and such a situation will typically never resolve satisfactorily until a reboot. [Edit: Note that even in case of a mutex, the described recovery mechanism may not be desirable or enabled by default, as the state of the protected resources or data structures is basically undefined after a crash midway. The decision whether to re-acquire or to recover differently is typically left to application code.]

So, a binary semaphore is similar to a monitor, but do not expect it to behave identically.