What is a counting semaphore good for?

Two main uses. First, managing a pool of identical resources: initialize the count to N (the number of available units, say N DMA buffers or N connection slots), and a task takes the semaphore (decrementing) to claim a unit and gives it back (incrementing) when done; when the count hits zero, further takers block until something is returned, naturally throttling usage to the available units without manual bookkeeping. Second, counting events so none are lost: if an ISR or producer can generate events faster than a consumer task drains them, a binary semaphore would coalesce multiple gives into one (the consumer sees "at least one" but loses the count), whereas a counting semaphore increments per event, so the consumer can take repeatedly and handle exactly as many events as occurred. So the mental model is "a binary semaphore says something happened; a counting semaphore says how many happened" or "how many units are available." You size the maximum count to the resource pool size or the worst-case event burst. It still has no ownership (it's a semaphore), so it's for counting/signaling, not for resource protection where you'd want a mutex's inheritance.