[PATCH v2 2/2] ARM: delay: allow timer-based delay implementation to be selected

[santosh.shilimkar@ti.com (Shilimkar, Santosh)]

On Tue, Jul 17, 2012 at 8:40 AM, Shinya Kuribayashi
<shinya.kuribayashi.px@renesas.com> wrote:
> Will, Stephen and Santosh,
>
> On 7/13/2012 8:13 PM, Will Deacon wrote:
>> I was anticipating that the platform would set the initial loops_per_jiffy
>> value if it requires udelays before loop calibration and the default of 4k
>> is wildly off.
>
> I overlooked two different lpj setups were involved at hand.
>
> First one was, the initial loops_per_jiffy value of 4k was too small for
> almost all processors running Linux today, so I set up loops_per_jiffy
> _early_, calculated from the CPU clock speed.  I didn't mentioned this
> before, sorry for confusion.
>
> So my initial loops_per_jiffy is not 4k at this point.  It's optimized
> for loop-based delay with the CPU running at 1.2GHz (much bigger than
> default 4k).
>
> And later, init_current_timer_delay() got processed.  Actual udelay()
> behavior switched from loop-based delay to timer-based one immediately,
> while my loops_per_jiffy was not changed/updated to appropriate value.
>
> This is why my udelay()s, used after init_current_timer_delay(), were
> taking considerable long time to expire.   Note that my initial tests
> for Will's patchset was done using a loadable module dedicated for
> udelay tests, that was prepared for 2.6.35/3.0 kernels beforehand.
>
> And this time, I confirmed that updating loops_per_jiffy at the same
> time as lpj_fine, works perfectly as expected for me.
>
>> If people need loops_per_jiffy to be updated at the same time as lpj_fine,
>> I can post that as a separate patch (below) as Russell has merged v2 of these
>> patches into his delay branch. That said, I'd certainly like to know if this
>> is actually a real problem (and can't be solved by choosing a compromise value
>> as the initial loops_per_jiffy). I think Shinya was doing some tests so
>> I'll wait to see how those went.
>
> From my observations:
>
> (1) loops_per_jiffy can easily be calculated from the CPU clock speed.
> If your platform is capable of detecting CPU frequency at run-time,
> settingi up loops_per_jiffy _early_ can allow you early use of udelay()s.
>
> Or even if you don't need udelay() early, setting up lpj_fine (or having
> calibrate_delay_is_known()) allows you to skip calibrate_delay() later.
> This is useful and can be applied to both UP and SMP systems.
>
> (2) For SMP platforms, if you need ealy use of udelay(), you have to
> update loops_per_jiffy at the same time as init_current_timer_delay().
> It could be done in init_current_timer_delay(), or platforms can take
> care of that, that need udelay() available early.  Either one should be
> fine with me.

Thanks for the detailed explanation. CPU clock detection is indeed the
nit way to skip the calibration overhead and this was one of the comment
when I tried to push the skipping of calibration for secondary CPUs.

Looks like you have a working patch for the clock detection. Will
you able to post that patch so that this long pending calibration
for secondary CPUs gets optimized.

Regards
Santosh