[RFC] ARM: sched_clock: update epoch_cyc on resume

[Colin Cross]

On Mon, Jul 23, 2012 at 5:14 PM, Linus Walleij <linus.walleij at linaro.org> wrote:
> On Mon, Jul 23, 2012 at 9:27 PM, Colin Cross <ccross at android.com> wrote:
>> On Mon, Jul 23, 2012 at 11:55 AM, Linus Walleij
>
>> Does the clock you use for sched_clock continue to run in all suspend
>> modes? All the SoC's I've used only have a 32kHz clock in the deepest
>> suspend mode,
>
> Yes, and yes it is 32kHz.
>
>> which is not ideal for sched_clock.
>
> Not that I know why scheduling with 32kHz is so bad compared to the
> default system scheduling granularity which is HZ if you don't have
> sched_clock() implemented.
>
> Since this seems to be such an important point, what makes you
> want MHz:es for scheduling granularity? To me the biggest impact
> is actually the granularity of the timestamps in the printk:s.
>
> (It's not that I doubt your needs, more curiosity.)

There's a comment somewhere about higher resolution sched_clock
providing fairer scheduling.  With 32 kHz sched_clock, every runtime
measured by the scheduler will be wrong by up to 31.25 us.  Most
systems have a faster clock, and if it's available it seems silly not
to use it.

It's also used for tracing, where 31.25 us resolution is a little low
for function tracing or function graph tracing.

>>  For example, on
>> Tegra2 the faster 1MHz clock used for sched_clock resets in the
>> deepest suspend state (LP0) but not the shallowest suspend state
>> (LP2), and which suspend state the chip hits depends on which hardware
>> is active.  Opting out of this patch would cause Tegra's clock to
>> sometimes run in suspend, and sometimes not, which seems worse for
>> debugging than consistently not running in suspend.  I'd be surprised
>> if a similar situation didn't apply to your platform.
>
> Well being able to switch between different sched_clock() providers
> may be the ideal...
>
>>> - If it absolutely needs to be in the core code, also have a bool
>>>   field indicating whether the clock is going to die during suspend
>>>   and add new registration functions for setting that sched_clock
>>>   type, e.g. setup_sched_clock_nonsuspendable()
>>
>> Sounds reasonable if some platforms need the extra complexity.
>
> OK agreed.
>
> A connecting theme is that of being avle to flag clock sources as
> sched_clock providers. If all clocksources were tagged with
> rating, and only clocksources were used for sched_clock(), the
> kernel could select the highest-rated clock under all circumstances.
>
> But that's quite intrusive, more of an idea. :-P

sched_clock is supposed to be very low overhead compared to ktime_get,
and has some strict  requirements if CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
is not set (see kernel/sched/clock.c), but it might be possible.