rtio: completion queue event count overflow

[JordanYates]

Describe the bug

The internal counter used to keep track of completion queue accounts is only ever incremented, never decreasing.

zephyr/include/zephyr/rtio/rtio.h

Line 1180 in 3d6dde4

atomic_inc(&r->cq_count);

This counter is used by rtio_submit to determine when the submission has completed:

zephyr/include/zephyr/rtio/rtio.h

Line 1453 in 3d6dde4

uintptr_t cq_count = (uintptr_t)atomic_get(&r->cq_count) + wait_count;

zephyr/include/zephyr/rtio/rtio.h

Lines 1473 to 1476 in 3d6dde4

	while ((uintptr_t)atomic_get(&r->cq_count) < cq_count) {
	Z_SPIN_DELAY(10);
	k_yield();
	}

A typical RTIO I2C transaction might like the following:

1. rtio_sqe_prep_tiny_write
2. rtio_sqe_prep_read
3. rtio_sqe_prep_callback_no_cqe`

So 2 CQE events per I2C transaction. If there were 50 I2C transactions per second, this counter would overflow in ~250 days (2**31 / 50 / 2 / 86400).

Increasing or decreasing the I2C transaction rate would scale the overflow duration linearly.
At overflow, the second check will pass immediately:

cq_count = INT32_MAX + 1; // INT32_MIN
while(INT32_MAX < INT32_MIN) {
   // Never enters
}

Expected behavior

Internal counter overflow should either be detected and handled, or reset in some way through atomic_dec.

Environment (please complete the following information):

• Zephyr v4.0

[teburd]

We need the counter to know when to return, so detecting the rollover is expected and should already work. Agreed a test should show this working