Is atomicity really needed when implementing a thread lock?

Suppose I have a piece of code that must not be executed concurrently. My (presumed naive) approach at attempting a thread lock would look something like this:

int lock = 0;

DWORD WINAPI ThreadProc(LPVOID lpParam)
{
    if (lock)
        return 1;

    lock = 1;

    /* code goes here */

    lock = 0;

    return 0;
}

When testing with the following:

for (i = 0; i < 2; i++)
    thandle[i] = CreateThread(NULL, 0, ThreadProc, NULL, 0, &tid[i]);

WaitForMultipleObjects(2, thandle, TRUE, INFINITE);

I always get the desired effect that only the first thread to be created actually gets to the code and returns 0. However, I keep coming across the suggestion that this method may fail because the lock is not implemented using atomic operations.

I cannot see a way of creating the threads simultaneously, so, in practical terms, is atomicity really needed here? Can someone provide an example that would cause the above to fail?