On 10/11/2021 11:30, Andriy Gapon wrote:
On 09/11/2021 17:56, Andriy Gapon wrote:
So, as I was saying, when the delta is large the calculations in tc_windup and
bintime_off give slightly different results and that can lead to a
discontinuity of the time when timehands are switched.
A quick follow-up.
I think that both tc_windup and bintime_off have fundamentally correct
calculations but with different precision. Both seem to produce values slightly
greater than a "true" value where the bintime fractional delta would be
calculated as tc_delta * 2^64 / tc_frequency. That's because of how th_scale is
calculated.
When the timecounter delta is greater than the frequency then the value in
tc_windup is closer to the true value because it accounts for whole seconds
precisely: a tc_frequency number of timecounter ticks is equal to one second.
bintime_off, however, converts both whole seconds and fractions using th_scale.
So, its result is consistently greater when the delta is longer than a second.
E.g., in my environment: tc_frequency = 14318180, th_scale = 1288420532460.
For a delta of 14318180 (== tc_frequency) tc_windup calculates a one second
advance, bt = { 1, 0 }.
bintime_off for the same delta will produce bt = { 1, 1093027638570944 }.
The difference is minuscule, just 59 ppm in relative terms.
But it's 59 microseconds of "jumping back in time".
I think that the precision of bintime_off is sufficient and its calculations are
faster, so I think that it's better to use the same calculations in tc_windup as
well. Especially given that they are identical for sub-second deltas and longer
deltas should be extremely rare.
I am working on patch to implement this.
The promised patch: https://people.freebsd.org/~avg/kern-tc-add-delta.diff
--
Andriy Gapon
