Large latency 'k_usleep() / k_wakeup()`

[TorstenRobitzki]

I try to have a LED blinking while doing some asynchronous background peripheral interactions. This runs on a nrf53 eval board.

In the main() thread, I create the necessary timing, using k_usleep() in a loop, leaving the loop, once the function returns 0.

    for ( int32_t sleep = k_usleep( sleeptime ); sleep; sleep = k_usleep( sleep ) )
    {
        ... handle peripheral interaction
    }

This works quite good and the LED is blinking with the desired frequency. I have in parallel a state machine, that prepaires and starts SPI transactions. An Interrupt service routine is installed at the end of the SPI transaction, that simply calls k_wakeup() with the thread id of the main thread.

I've toggle a debug pin right after the call to k_wakeup() and right after the call to k_usleep(). And now, I observe, that the latency between the wakeup call and the actual wakeup is round about 5ms! Which means, that I have a delay of 5ms between my SPI transactions.

I can not find anything in the documentation of k_usleep() nor k_wakeup() that would justify this large latency. If I replace the sleep loop by k_yield() I end up with at latency of ~20µs and when I replace it with _WFI(), I end up with 400ns.

Is this intended behavior?

[TorstenRobitzki]

My mistake was, to use IRQ_DIRECT_CONNECT to register the IRQs (and not IRQ_CONNECT). After changing this and using a semaphore to wake the main thread, I measure a context switch time of ~2µs.