[LinuxPPS] testing PPS with ktimer

[Ken MacLeod]

On Tue, Nov 9, 2010 at 11:09 AM, Ken MacLeod <ken at bitsko.slc.ut.us> wrote:
> I'm tracing through the ntpd code and it currently is discarding the
> PPS time source around line 1310 of ntp_refclock.c in refclock_pps()
> in this test:
>
>        /*
>         * There can be zero, one or two PPS pulses between polls,
>         * depending on the poll interval relative to the PPS interval.
>         * The pulse must be newer and within the range gate relative
>         * to the last pulse.
>         */
>        if (ap->ts.tv_sec <= timeout.tv_sec || abs(ap->ts.tv_nsec -
>            timeout.tv_nsec) > RANGEGATE)
>                return (0);
>
> I'm digging around here now to find out what the pattern is.

Problem found:  the issue is in pps-ktimer.c, pps_ktimer_event().

pps_ktimer_event reschedules itself with mod_timer(&ktimer, jiffies +
HZ) which _should_ reschedule in one second, but due to delays (most
specifically the pr_info in that routine) it ends up being more than a
second each time.  Enough that it is greater than RANGEGATE.  You can
see this with ppstest by watching the time skew.

I made a quick fix to calculate the delay from getnstimeofday and ntpd
is now using the pps-ktimer source.