ondemand governor and passive cooling broken?

ondemand governor and passive cooling broken?

[Thomas Renninger]

Machine is a Thinkpad with a Pentium M ~1700MHz running with the speedstep-centrino module.

Something seems to be wrong with the cooling system and
the CPU temp goes up from 50 C to 90 C in about a minute.

With the userspace governor the machine stays at 1200 MHz on full load
and bounces around 80C (passive trip point), that's how it should be.

However, passive cooling seems to be broken with current kernels
and ondemand governor?
Machine stays at 1700 MHz until critical shutdown (94 C).

Is passive cooling supposed to work with the ondemand governor?
If yes, can someone confirm seeing it working with old kernels and
the speedstep-centrino module?

Thanks,

    Thomas

Re: ondemand governor and passive cooling broken?

[Dominik Brodowski]

Hi,

With which kernel do you see this behaviour? Can you re-test with 2.6.12.x,
please?

Thanks,
	Dominik

Re: ondemand governor and passive cooling broken?

[Thomas Renninger]

Hi,

This makes ondemand governor aware of thermal limits.
The policy was buffered and new policy values never reached the governor.

I just used the cpufreq_get_policy call every time the load is checked,
maybe the new policy should be made available via notifier as it is done
for the cur_freq in the userspace governor?
If this sounds reasonable I can come up with a bigger patch on Monday.

However, thermal cooling still does not always work right:

  - Sometimes the limit is not decreased correctly in polling frequency time
    (I saw it at max for more than 20 secs, polling freq 5 secs, passive tp exceeded).
    Strange enough, that seems to happen only in rare cases.
  - I also saw the limit stuck at lowest frequency even after getting back in
    normal temperature zones.
  - And once you entered passive cooling you always see the "cooling: passive"
    flag set in /proc/acpi/thermal_zone/*/cooling_mode


    Thomas


--- drivers/cpufreq/cpufreq_ondemand.c.orig     2005-08-20 19:26:18.000000000 -0600
+++ drivers/cpufreq/cpufreq_ondemand.c  2005-08-20 19:28:04.000000000 -0600
@@ -254,6 +254,10 @@
        if (!this_dbs_info->enable)
                return;

+       int ret = cpufreq_get_policy(this_dbs_info->cur_policy, cpu);
+       if (ret)
+               return;
+
        policy = this_dbs_info->cur_policy;
        /*
         * Every sampling_rate, we check, if current idle time is less

RE: ondemand governor and passive cooling broken?

[Pallipadi, Venkatesh]

 

>-----Original Message-----
>From: Thomas Renninger [mailto:mail@renninger.de] 
>Sent: Saturday, August 20, 2005 9:49 PM
>To: Dominik Brodowski
>Cc: cpufreq@lists.linux.org.uk; Pallipadi, Venkatesh
>Subject: Re: ondemand governor and passive cooling broken?
>
>Hi,
>
>This makes ondemand governor aware of thermal limits.
>The policy was buffered and new policy values never reached 
>the governor.

I don't think there is an issue with buffering here.
This_dbs_info->cur_policy is a pointer to the cpufreq policy (and not
allocated by ondemand governor). So, any change to max/min in cpufreq
should be automatically getting reflected. Atleast that was the intent
with having policy pointer. If it is not happening then that should be a
bug. I don't think that this memcpy is required here.

>I just used the cpufreq_get_policy call every time the load is checked,
>maybe the new policy should be made available via notifier as 
>it is done
>for the cur_freq in the userspace governor?
>If this sounds reasonable I can come up with a bigger patch on Monday.
>
>However, thermal cooling still does not always work right:
>
>  - Sometimes the limit is not decreased correctly in polling 
>frequency time
>    (I saw it at max for more than 20 secs, polling freq 5 
>secs, passive tp exceeded).
>    Strange enough, that seems to happen only in rare cases.
>  - I also saw the limit stuck at lowest frequency even after 
>getting back in
>    normal temperature zones.

I think we need to do some more things to intergrate cpufreq and thermal
cooling. Especially in relation to polling frequency. Probably, ondemand
governor should slow down its polling interval to something higher than
thermal polling frequency as soon as it notices the passive cooling is
on.


Thanks,
Venki

Re: ondemand governor and passive cooling broken?

[Thomas Renninger]

Pallipadi, Venkatesh wrote:
>  
> 
>>-----Original Message-----
>>From: Thomas Renninger [mailto:mail@renninger.de] 
>>Sent: Saturday, August 20, 2005 9:49 PM
>>To: Dominik Brodowski
>>Cc: cpufreq@lists.linux.org.uk; Pallipadi, Venkatesh
>>Subject: Re: ondemand governor and passive cooling broken?
>>
>>Hi,
>>
>>This makes ondemand governor aware of thermal limits.
>>The policy was buffered and new policy values never reached 
>>the governor.
> 
> I don't think there is an issue with buffering here.
> This_dbs_info->cur_policy is a pointer to the cpufreq policy (and not
> allocated by ondemand governor). So, any change to max/min in cpufreq
> should be automatically getting reflected. Atleast that was the intent
> with having policy pointer. If it is not happening then that should be a
> bug. I don't think that this memcpy is required here.
> 
You are right. I tried to reproduce this problem with a powernow-k8 machine
and thermal cooling worked just fine.

>>I just used the cpufreq_get_policy call every time the load is checked,
>>maybe the new policy should be made available via notifier as 
>>it is done
>>for the cur_freq in the userspace governor?
>>If this sounds reasonable I can come up with a bigger patch on Monday.
>>
>>However, thermal cooling still does not always work right:
>>
>> - Sometimes the limit is not decreased correctly in polling 
>>frequency time
>>   (I saw it at max for more than 20 secs, polling freq 5 
>>secs, passive tp exceeded).
>>   Strange enough, that seems to happen only in rare cases.
>> - I also saw the limit stuck at lowest frequency even after 
>>getting back in
>>   normal temperature zones.
> 
> I think we need to do some more things to intergrate cpufreq and thermal
> cooling. Especially in relation to polling frequency. Probably, ondemand
> governor should slow down its polling interval to something higher than
> thermal polling frequency as soon as it notices the passive cooling is
> on.

Don't know. I will try to get a speedstep-centrino machine and have a look
whether this makes a difference and also play around a bit with thermal polling freq
and ondemand configs. Do you think this could be a driver specific problem?

Thanks,

      Thomas