Question

I am writing a bit of code that implements an (unsigned) integer counter.

1. It is used from an arbitrary number of threads.
2. A thread should get a unique value every time until overflow.
3. If integer type range is overflown, zero should be returned.
4. I have at my disposal atomic increment function (returns old value), and atomic compare-and-swap function.

All scenarios I have come up with so far suffer from race conditions on overflow. Is it possible to implement the counter with these constraints?

Answer 1

You can build everything from compare-and-swap. The general algorithm is

int DoSomeAtomicCalculation()
{
    static int value;
    int capture, newValue;
    do {
        capture = value;
        newValue = some_function_based_on(capture);
    } while (!compare_and_swap(value, capture, newValue));
    return newValue;
}

(I’m assuming that compare_and_swap takes a variable, a comparison value, and a swap value, and it returns true if the comparison succeeded (and the swap occurred).