Re: [PATCH v2 3/7] timecounter: allow for non-power of two overflow
From: Thomas Gleixner
Date: Wed Nov 24 2021 - 15:55:50 EST
Martin,
On Tue, Nov 09 2021 at 10:50, Martin Kaistra wrote:
> * see CYCLECOUNTER_MASK() helper macro
> * @mult: cycle to nanosecond multiplier
> * @shift: cycle to nanosecond divisor (power of two)
> + * @overflow_point: non-power of two overflow point (optional),
> + * smaller than mask
> */
> struct cyclecounter {
> u64 (*read)(const struct cyclecounter *cc);
> u64 mask;
> u32 mult;
> u32 shift;
> + u64 overflow_point;
> };
>
> /**
> diff --git a/kernel/time/timecounter.c b/kernel/time/timecounter.c
> index e6285288d765..afd2910a9724 100644
> --- a/kernel/time/timecounter.c
> +++ b/kernel/time/timecounter.c
> @@ -39,6 +39,9 @@ static u64 timecounter_read_delta(struct timecounter *tc)
> /* calculate the delta since the last timecounter_read_delta(): */
> cycle_delta = (cycle_now - tc->cycle_last) & tc->cc->mask;
>
> + if (tc->cc->overflow_point && (cycle_now - tc->cycle_last) > tc->cc->mask)
> + cycle_delta -= tc->cc->mask - tc->cc->overflow_point;
TBH, this took me more than one twisted braincell to grok.
With support for clocks which do not wrap at power of 2 boundaries we
already lose the unconditional fast path no matter what. So what's the
point of having two conditions and doing this convoluted math here?
In timecounter_init():
tc->ovfl = cc->ovfl ? cc->ovfl : cc->mask + 1;
which makes it a common path in timecounter_read_delta():
cycle_delta = cycle_now - tc->cycle_last;
if ((s64)cycle_delta) < 0)
cycle_delta += tc->ovfl;
which produces way better binary code.
The conditional does not really matter for the timecounter use cases as
that calculation is noise compared to the actual cc->read() access.
Aside of that the same problem exists in timecounter_cyc2time()...
After that we probably should do a treewide sweep to get rid of cc->mask
to avoid confusion and subtle to understand errors when some code uses
cc->mask instead of cc->ovfl.
Thanks,
tglx