[PATCH 1/3] clocksource: exynos_mct: Fix ftrace

[Doug Anderson]

Daniel,

On Sun, Jun 15, 2014 at 2:18 PM, Daniel Lezcano
<daniel.lezcano at linaro.org> wrote:
> On 06/04/2014 07:30 PM, Doug Anderson wrote:
>>
>> In (93bfb76 clocksource: exynos_mct: register sched_clock callback) we
>> supported using the MCT as a scheduler clock.  We properly marked
>> exynos4_read_sched_clock() as notrace.  However, we then went and
>> called another function that _wasn't_ notrace.  That means if you do:
>>
>>    cd /sys/kernel/debug/tracing/
>>    echo function_graph > current_tracer
>>
>> You'll get a crash.
>>
>> Fix this (but still let other readers of the MCT be trace-enabled) by
>> adding an extra function.  It's important to keep other users of MCT
>> traceable because the MCT is actually quite slow.
>
>
>
> Hi Doug,
>
> could you elaborate ? I don't get the 'because the MCT ... slow'

Sorry, I was trying to avoid duplication in the series and it's more
obvious when you look at parts 2 and 3 of the series.  ;)

Doing the math (please correct any miscalculations) using the numbers
from the other patches: You can see that the existing code takes
1323852 us for 1000000 gettimeofday in userspace.  The fastest
implementation (just shaving to a 32-bit timer) gets us as fast as
~1000000 us for 1000000 gettimeofday in userspace.

>From profiling, I believe that gettimeofday from userspace is about
50% overhead (system call, multiplication, copies, etc) and about 50%
MCT read.  That means that the fastest you can possibly do an MCT read
is in .5us or 500ns.

I believe an A15 has something like 1 or 2 cycles per instruction.  If
it were 2 cycles per instruction, it can execute a normal instruction
on a 2GHz machine in .5ns.  That means we can execute 1000 normal
instructions in the time it takes to do a since MCT access.

...so I guess that's what I'd call slow.  ;)  What do you think?  I
know that the MCT read shows up in whole system profiles of
gettimeofday.

-Doug