[PATCH] nomadik: prevent sched_clock() wraparound v3

[Linus Walleij]

Hm I wonder why this mail doesn't show up on linux-arm-kernel... Probably
too many CC.

2010/11/17 Rabin Vincent <rabin at rab.in>:
> On Wed, Nov 17, 2010 at 7:15 PM, Linus Walleij
>> +       /* Find the apropriate mult and shift factors */
>> +       clocks_calc_mult_shift(&sched_mult, &sched_shift,
>> +                              rate, NSEC_PER_SEC, SCHED_CLOCK_MIN_WRAP);
>> +       /* We need to multiply by atleast 2 to get rid of bit 63 */
>> +       BUG_ON(sched_mult < 2);
>
> No, you want to multiply by an even number, not "atleast 2".  Please
> also drop the "cycles &= 0x7FFFFFFFFFFFFFFFLLU" in sched_clock().

Ayeah, the Orion trick, that's better what was I thinking of..

>> +       /* Let's see what we get, take max counter and scale it */
>> +       days = 0xFFFFFFFFFFFFFFFFLLU >> sched_shift;
>> +       do_div(days, rate);
>> +       do_div(days, (3600*24));
>> +
>> +       pr_debug("sched_clock: using %d bits @ %lu Hz wrap in %lu days\n",
>> +               (64 - sched_shift), rate, (unsigned long) days);
>
> This days calculation is wrong because you need to divide by
> NSECS_PER_SEC, not rate.

Yes thanks, I was trying to be clever leaving out sched_mult but it doesn't
loose bits properly. I replaced it with:

days = (0xFFFFFFFFFFFFFFFFLLU * sched_mult) >> sched_shift;
do_div(days, NSEC_PER_SEC);
do_div(days, (3600*24));

And it yields:
sched_clock: using 55 bits @ 8333125 Hz wrap in 416 days

Which is what is to be expected. (I requested a minimum of 365 days)

Thanks a lot Rabin I'll fire off v4...

Yours,
Linus Walleij