Wait until all futexes are available?

> An alternative approach is to have the consumer wait on a single futex.

Waking up precisely one waiter if always the fastest choice.

> This implies that the producers must run within the same process or share some state in shared memory.

Yes, but futex's only work with shared memory. Communicating via shared memory is orders of magnitude faster than another other method. That's where futex's get their speed from. So if you want speed shared memory unavoidable.

If speed is not such a big issue there are as you say lots of choices, including epoll(), pipes, semop()'s ...

> My task-graph framework would have been simpler if the futex API let my consumer wait on a set of futexes and wake only when all inputs had been satisfied.

Of course it would be. Things are always simpler when there is a pre-existing library that does exactly what you want. What triggered my interest here is the claim that library should be in the kernel.