Re: [BISECTED] "conservative" cpufreq governor broken

[Steven Noonan]

On Tue, Oct 6, 2009 at 5:54 PM, Steven Noonan <steven@xxxxxxxxxxxxxx> wrote:
> On Tue, Oct 6, 2009 at 5:36 AM, Eero Nurkkala
> <ext-eero.nurkkala@xxxxxxxxx> wrote:
>> On Tue, 2009-10-06 at 13:22 +0200, ext Steven Noonan wrote:
>>> On Tue, Oct 6, 2009 at 3:43 AM, Eero Nurkkala
>>> <ext-eero.nurkkala@xxxxxxxxx> wrote:
>>> > On Tue, 2009-10-06 at 12:22 +0200, ext Steven Noonan wrote:
>>> >>
>>> >> I would suspect you have to have CONFIG_NO_HZ enabled to be able to
>>> >> reproduce the issue (considering the title of the bisected commit and
>>> >> my own config). Do you have it enabled?
>>> >>
>>> >
>>> > Yes, it's enabled.
>>> >
>>> >> > And another round:
>>> >> >
>>> >> > cpufreq stats: OP1:16,78%, OP2:0,24%, OP3:5,14%, OP4:77,83%  (72)
>>> >> >
>>> >> > Just once more after doing nothing:
>>> >> > OP1:7,41%, OP2:0,11%, OP3:2,38%, OP4:90,10%  (82)
>>> >> >
>>> >> > So I can't agree it's broken. The patch you bisected, actually filtered
>>> >> > out such phenomenon, in which an IRQ made the cpufreq framework
>>> >> > occasionally think we were idling, although we were not. So you got
>>> >> > "bonus" idle time that shouldn't been there in the first place. Now that
>>> >> > the "bonus" idle time is not there, your system load may indeed be so
>>> >> > high that the system never spends 80% or more time in idle? Could that
>>> >> > be the case? Of course, even though I can't agree it's broken, doesn't
>>> >> > mean it isn't somehow broken ;) It'd be nice to get info on other
>>> >> > systems as well...
>>> >>
>>> >> Interestingly, "ondemand" (the governor fixed by the bisected commit)
>>> >> works fine. "conservative" is the only broken one.
>>> >>
>>> >
>>> > If you took timestamps in /arch/x86/kernel/process_**.c:
>>> > (let's assume process_64.c) in cpu_idle()
>>> > around enter_idle(); and __exit_idle(), took the diff,
>>> > added the diffs up, and compared it to system uptime, you could see how
>>> > much time you spend in idle()? I think it's possible that
>>> > even if the cpu load is near 0%, the system may idle only for a bare
>>> > moment (that translates to a buggy pm_idle()), and time is spent
>>> > elsewhere (less than 80% in idle).
>>>
>>> This makes logical sense, but how should I test this? Is there a way
>>> to do this with existing tracers?
>>
>> Tracers may by themselves add some load into the system.
>>
>> If I were you, I'd add something like: (I have only one CPU BTW)
>>
>> static ktime_t time_prior_idle;
>> static int64_t idle_total;
>>
>> time_prior_idle = ktime_get();
>> <idle stuff>
>> idle_total += ktime_to_ns(ktime_sub(ktime_get(), time_prior_idle));
>>
>> and have a sysfs hook (something already present, so you can just cat
>> it) with a trace:
>>
>> printk("Times: %lld, %lld \n", idle_total, ktime_to_ns(ktime_get()));
>>
>> Sample output:
>> 374758812519, 386768249832
>>
>> So I have 386768249832 / 386768249832 = 96.9 % time spent in idle in
>> this case.
>>
>> (Right, this should provide somewhat descent info, hopefully ;) )
>>
>
> Well, I tried adding the code to cpu_idle() as suggested, but it never
> printed anything. Apparently cpu_idle() isn't ever being called here.
> Even added a 'BUG();' at the beginning of the function and it never
> hit it. Of course, I'm probably missing something obvious. Is there a
> separate cpu_idle()-esque function for SMP?
>

Oh crap. Perhaps it's more insidious. I reverted the bisected commit
and it _DID_ hit the line I added. So cpu_idle is never entered with
the bisected commit. Bizarre.

- Steven
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/