pá 3. 7. 2026 v 14:11 odesílatel Tomas Glozar <[email protected]> napsal:
>
> pá 12. 6. 2026 v 19:55 Crystal Wood <[email protected]> napsala:
>
> > > OSNOISE_LL_OPTIONS takes three options: name - struct osnoise_context
> > > field name (written "<opt>" above), path - filename inside
> > > /sys/kernel/tracing/osnoise passed to libtracefs, and init_val - initial
> > > value of struct fields, corresponding to an otherwise invalid option
> > > (some options use OSNOISE_OPTION_INIT_VAL = -1, some use
> > > OSNOISE_TIME_INIT_VAL = 0).
> >
> > Can we simplify by always using -1? Especially since that's already
> > treated as the universal "invalid" by osnoise_read_ll_config().
> >
> > FWIW using "init val" to mean "invalid" rather than "default" is a bit
> > unintuitive.
> >
>
> The idea behind *_INIT_VAL is to re-use a value that is invalid on the
> osnoise tracer side to mean "not (read from tracer and) set yet (on
> the RTLA side)". OSNOISE_TIME_INIT_VAL is used for values where 0 is
> invalid (e.g. period, runtime), OSNOISE_OPTION_INIT_VAL is used for
> those where 0 is valid. I believe the distinction comes from the fact
> that some of the osnoise options are unsigned on the kernel side, and
> -1 (= 18446744073709551615) is actually a valid value there. E.g.:
>
> [root@cs9 osnoise]# cat period_us
> 1000000
> [root@cs9 osnoise]# echo 18446744073709551615 > period_us
> [root@cs9 osnoise]# cat period_us
> 18446744073709551615
>
> There are values where both 0 and -1 are valid (stop_tracing_us,
> stop_tracing_total_us, timerlat_align_us), and they use
> OSNOISE_OPTION_INIT_VAL. Those cannot use OSNOISE_TIME_INIT_VAL as 0
> is a meaningful and common value for them (disabled for
> stop_us/stop_total_us, zero alignment for timerlat_align_us). Those
> will break if pre-set to 18446744073709551615.
>
> Merging OSNOISE_OPTION_INIT_VAL and OSNOISE_TIME_INIT_VAL would
> introduce this breakage to options that are using
> OSNOISE_TIME_INIT_VAL now. IMHO the cleanest solution is dropping the
> entire complex logic around checking if the value was read from the
> kernel already, as there is no user: RTLA always saves the value once,
> then restores it at the end.
>
Ah sorry, I missed that we use get_llong_from_str() for everything,
which already treats all values as (signed) long long, despite period
and runtime ("OSNOISE_OPTION_INIT_VAL" options) being declared
unsigned. We can just drop the INIT_VALs then and just use -1.
Tomas
