Re: [RFC PATCH v2 2/6] timekeeping: Fix cross-timestamp interpolation corner case decision

[Peter Hilber]

On 15.09.23 18:10, Thomas Gleixner wrote:
> On Fri, Aug 18 2023 at 03:20, Peter Hilber wrote:
>> --- a/kernel/time/timekeeping.c
>> +++ b/kernel/time/timekeeping.c
>> @@ -1247,7 +1247,8 @@ int get_device_system_crosststamp(int (*get_time_fn)
>>  		 */
>>  		now = tk_clock_read(&tk->tkr_mono);
>>  		interval_start = tk->tkr_mono.cycle_last;
>> -		if (!cycle_between(interval_start, cycles, now)) {
>> +		if (!cycle_between(interval_start, cycles, now) &&
>> +		    cycles != interval_start) {
>>  			clock_was_set_seq = tk->clock_was_set_seq;
>>  			cs_was_changed_seq = tk->cs_was_changed_seq;
>>  			cycles = interval_start;
> 
> So the explanation in the changelog makes some sense, but this code
> without any further explanation just makes my brain explode.
> 
> This whole thing screams for a change to cycle_between() so it becomes:
> 
>      timestamp_in_interval(start, end, ts)
> 
> and make start inclusive and not exclusive, no?

I tried like this in v1 (having 'end' inclusive as well), but didn't like
the effect at the second usage site.

> 
> That's actually correct for both usage sites because for interpolation
> the logic is the same. history_begin->cycles is a valid timestamp, no?

AFAIU, with the timestamp_in_interval() change, history_begin->cycles would
become a valid timestamp. To me it looks like
adjust_historical_crosststamp() should then work unmodified for now. But
one would have to be careful with the additional corner case in the future.

So, document the current one-line change, or switch to
timestamp_in_interval()?

Thanks for the review!

Peter