What about a simplified version of io_uring_peek_batch_cqe?

[Ebenezer-group]

I have this function:

auto submit (){
    static int seen=0;
    ::io_uring_cq_advance(&rng,seen);
    int rc;
    while((rc=::io_uring_submit_and_wait(&rng,1))<0){
      if(-EINTR!=rc)raise("waitCqe",rc);
    }
    static ::io_uring_cqe *cqes[10];
    seen=::io_uring_peek_batch_cqe(&rng,&cqes[0],10);
    return ::std::span<::io_uring_cqe*>(&cqes[0],seen);
}

in this file.

This version of io_uring_peek_batch seems more complicated than I want:

unsigned io_uring_peek_batch_cqe(struct io_uring *ring,
				 struct io_uring_cqe **cqes, unsigned count)
{
	unsigned ready;
	bool overflow_checked = false;
	int shift = 0;

	if (ring->flags & IORING_SETUP_CQE32)
		shift = 1;

again:
	ready = io_uring_cq_ready(ring);
	if (ready) {
		unsigned head = *ring->cq.khead;
		unsigned mask = ring->cq.ring_mask;
		unsigned last;
		int i = 0;

		count = count > ready ? ready : count;
		last = head + count;
		for (;head != last; head++, i++)
			cqes[i] = &ring->cq.cqes[(head & mask) << shift];

		return count;
	}

	if (overflow_checked)
		return 0;

	if (cq_ring_needs_flush(ring)) {
		io_uring_get_events(ring);
		overflow_checked = true;
		goto again;
	}

	return 0;
}

I would like a version that trusts me that something is ready. There could be an assert added that 'ready' is a positive number after the call to io_uring_cq_ready. Something like this:

unsigned io_uring_peek_batch_cqe(struct io_uring *ring,
				 struct io_uring_cqe **cqes, unsigned count)
{
	unsigned ready;
	int shift = 0;

	if (ring->flags & IORING_SETUP_CQE32)
		shift = 1;

	ready = io_uring_cq_ready(ring);
	unsigned head = *ring->cq.khead;
	unsigned mask = ring->cq.ring_mask;
	unsigned last;

	count = count > ready ? ready : count;
	last = head + count;
	for (int i=0;head != last; head++, i++)
		cqes[i] = &ring->cq.cqes[(head & mask) << shift];

	return count;
}

Building with this simplified version reduces the size of my executable by several hundred bytes.

[axboe]

You could surely add that, might make sense to just abstract out that part into a __io_uring_peek_batch_cqe() helper.