Re: [PATCH net-next v4 1/2] dpll: zl3073x: add channel ToD, phase step and TIE operations

[Ivan Vecera]

Sashiko findings with replies...

> zl3073x_poll_zero_u8() hard-codes the interval:
>
> #define ZL_POLL_SLEEP_US 10
>
> so the new timeouts [...] all wait at a 10 us cadence. [...] giving on
> the order of 10^5 bus reads while waiting out a WR_NEXT_1HZ command
> [...]
> Would a millisecond-scale interval variant be preferable for these
> three cases, so the shared bus and the driver's monitor kworker are not
> starved for up to a second?

Good catch. I did some testing and found that even the existing
registers typically clear within low hundreds of iterations, so the
fixed 10 us sleep is already unnecessarily aggressive for all callers.

I will add a prep patch that replaces the fixed ZL_POLL_SLEEP_US with
timeout_us / 50 in zl3073x_poll_zero_u8(). This scales the sleep
interval proportionally to the timeout, targeting ~50 iterations
regardless of timeout magnitude:
  - existing 30-50 ms timeouts: 600-1000 us sleep
  - new 1000-3000 ms PTP timeouts: 20-60 ms sleep

> Can the tick-crossed branch skip the 20 ms rollover guard and shift the
> applied time by one second?
>
> The margin test is gated on !diff.tv_sec, so when the 1 Hz tick crossed
> between the two zl3073x_chan_tod_read() calls only the seconds get 
fixed up
> and the remaining margin is never compared against threshold_ns [...]

Good point. Although this requires ~1 s delay between two consecutive
register reads (practically unreachable), the fix is clean and costs
nothing. I will restructure the guard to separate the tick-crossing
fixup from the margin check so both apply independently:

    diff = timespec64_sub(ts_next, ts_cur);

    if (diff.tv_sec < 0)
        ts_next.tv_sec -= diff.tv_sec;

    if (diff.tv_sec <= 0 && diff.tv_nsec < threshold_ns) {
        fsleep((unsigned long)diff.tv_nsec / NSEC_PER_USEC + 1);
        ts_next.tv_sec++;
    }

Thanks,
Ivan