Re: [BISECTED] "conservative" cpufreq governor broken

[Steven Noonan <steven@uplinklabs.net>]

On Wed, Oct 7, 2009 at 1:05 AM, Steven Noonan <steven@uplinklabs.net> wrote:
> On Wed, Oct 7, 2009 at 12:49 AM, Eero Nurkkala
> <ext-eero.nurkkala@nokia.com> wrote:
>> On Wed, 2009-10-07 at 09:30 +0200, ext Steven Noonan wrote:
>>>
>>> Okay, wow, I'm a moron. I misread what cpu_idle() was intended to be
>>> for. I thought that cpu_idle() was a function that was periodically
>>> called whenever the CPU had nothing to do, but now I see that it's
>>> actually the main loop. I should really read the code next time.
>>>
>>> I've moved the statistics printout code to the _inside_ of that
>>> infinite loop and retested. I had it print every several hundred
>>> iterations. Here's the results (note the machine was idle the whole
>>> time, except for about the first 10-20 seconds while the machine
>>> booted):
>>>
>>> [    3.627716] timings[0]: 2250511125 / 3627716116
>>> [    6.946216] timings[0]: 4780901366 / 6946213531
>>> [   13.355182] timings[0]: 9385417604 / 13355183525
>>> [   18.551304] timings[1]: 16300853077 / 18551301189
>>> [   21.589039] timings[0]: 15984495433 / 21589037480
>>> [   47.152733] timings[1]: 44386121538 / 47152731476
>>> [   51.682630] timings[0]: 45713834076 / 51682628295
>>> [   79.587359] timings[0]: 73524821916 / 79587356820
>>> [   88.630110] timings[1]: 85324277596 / 88630109605
>>> [   96.082386] timings[0]: 89691306072 / 96082384539
>>>
>>
>> Those look good.
>>
>> Well, might as well then go for:
>> /drivers/cpufreq/cpufreq_conservative.c
>> dbs_check_cpu() ->
>> load = 100 * (wall_time - idle_time) / wall_time; <- What is your load?
>
> That's probably the problem...
>
> [   40.632277] cpufreq load = 100 * (66667 - 3310) / 66667 = 95
> [   40.698947] cpufreq load = 100 * (66661 - 3238) / 66661 = 95
> [   73.965425] cpufreq load = 100 * (66667 - 12820) / 66667 = 80
> [   74.032095] cpufreq load = 100 * (66661 - 1124) / 66661 = 98
> [  107.298571] cpufreq load = 100 * (66666 - 13092) / 66666 = 80
> [  107.365301] cpufreq load = 100 * (66722 - 3317) / 66722 = 95
> [  140.631717] cpufreq load = 100 * (66666 - 3311) / 66666 = 95
> [  140.698387] cpufreq load = 100 * (66662 - 3237) / 66662 = 95
>
> idle_time is wrong.

Actually, it's more likely that the idle_time there is correct and
there's something else that's going runaway. My system's fans are
running at about 4000 RPM, when they'd normally be running at 2000 RPM
for this load average. I suspect there's something actually going
wild.

>
>> Let assume load is sane, look for (in dbs_check_cpu())
>>        if (load < (dbs_tuners_ins.down_threshold - 10)) {
>>
>> whether it is taken ever...if not, what is your
>> (dbs_tuners_ins.down_threshold - 10) ?
>>
>> - Eero
>>
>>
>