Gosh, I didn't know about that. My choice of poll was mostly dictated by the ALSA library API, which provides a pre-cooked set of struct pollfds for detecting when more data needs to be sent to the sound device. I suppose I could have read those and translated their contents into my poll-equivalent of choice, but it didn't seem particularly worth doing.

epoll looks like massive overkill for a job this small, anyway!