It's important to remember that some of the possible states arise from the memory model rather than what your code intentionally does. Say a producer thread fills a buffer and queues work, and a consumer thread receives the queue entry and reads the buffer. Does the consumer read any or all of what the producer intended? It depends on whether they were scheduled in different cores, and how full their caches were, and whether there was a memory barrier that's global (which is slow and not always needed) or only applies to the queue itself. Good platforms trade away just enough performance to guarantee correctness here.