Q: sched_clock() vs. clocksource, how to implement correctly

[js@sig21.net (Johannes Stezenbach)]

On Sat, Apr 24, 2010 at 10:50:11PM +0200, Johannes Stezenbach wrote:
> /*
>  * Orion's sched_clock implementation. It has a resolution of
>  * at least 7.5ns (133MHz TCLK) and a maximum value of 834 days.
>  *
>  * Because the hardware timer period is quite short (21 secs if
>  * 200MHz TCLK) and because cnt32_to_63() needs to be called at
>  * least once per half period to work properly, a kernel timer is
>  * set up to ensure this requirement is always met.
>  */
> #define TCLK2NS_SCALE_FACTOR 8

I found the following discussion of the sched_clock()
implementation trade-offs very informative:
http://lkml.org/lkml/2010/4/15/299

It mentions clocksource_calc_mult_shift() which was added in 2.6.33,
however I had some difficulties understanding the meaning of the minsec
parameter, especially since all existing callers use a value of 4.
But when using minsec = 365*24*60*60 (1 year) it results in the shift value of 8.

So finally the pieces connect together :-)

> BTW, even though this uses TCLK2NS_SCALE_FACTOR of 8, the same file
> uses a shift vaue of 20 for the orion_clksrc...

It seems clocksource_cyc2ns() is used in kernel/time/timekeeping.c
and kernel/time/clocksource.c only on relatively
small delta values, so there's no need to worry
about overflow and a large clocksource.shift and .mult is OK.

(Apparently the minsec value of 4 mentioned above is suitable
for timekeeping?  Where does the 4 come from?)


Thanks,
Johannes