Re: [RFC PATCHv2 0/7] devfreq: Add devfreq-event class to provide raw data for devfreq device

From: Chanwoo Choi
Date: Wed Dec 10 2014 - 08:22:00 EST


Hi Krzysztof,

On Wed, Dec 10, 2014 at 7:07 PM, Krzysztof Kozlowski
<k.kozlowski@xxxxxxxxxxx> wrote:
> On wto, 2014-12-09 at 23:12 +0900, Chanwoo Choi wrote:
>> This patchset add new devfreq_event class to provide raw data to determine
>> current utilization of device which is used for devfreq governor.
>>
>> [Description of devfreq-event class]
>> This patchset add new devfreq_event class for devfreq_event device which provide
>> raw data (e.g., memory bus utilization/GPU utilization). This raw data from
>> devfreq_event data would be used for the governor of devfreq subsystem.
>> - devfreq_event device : Provide raw data for governor of existing devfreq device
>> - devfreq device : Monitor device state and change frequency/voltage of device
>> using the raw data from devfreq_event device
>>
>> The devfreq subsystem support generic DVFS(Dynamic Voltage/Frequency Scaling)
>> for Non-CPU Devices. The devfreq device would dertermine current device state
>> using various governor (e.g., ondemand, performance, powersave). After completed
>> determination of system state, devfreq device would change the frequency/voltage
>> of devfreq device according to the result of governor.
>>
>> But, devfreq governor must need basic data which indicates current device state.
>> Existing devfreq subsystem only consider devfreq device which check current system
>> state and determine proper system state using basic data. There is no subsystem
>> for device providing basic data to devfreq device.
>>
>> The devfreq subsystem must need devfreq_event device(data-provider device) for
>> existing devfreq device. So, this patch add new devfreq_event class for
>> devfreq_event device which read various basic data(e.g, memory bus utilization,
>> GPU utilization) and provide measured data to existing devfreq device through
>> standard APIs of devfreq_event class.
>>
>> The following description explains the feature of two kind of devfreq class:
>> - devfreq class (existing)
>> : devfreq consumer device use raw data from devfreq_event device for
>> determining proper current system state and change voltage/frequency
>> dynamically using various governors.
>> - devfreq_event class (new)
>> : Provide measured raw data to devfreq device for governor
>>
>> Also, the devfreq-event device would support various type event as following:
>> : DEVFREQ_EVENT_TYPE_RAW_DATA
>> : DEVFREQ_EVENT_TYPE_UTILIZATION
>> : DEVFREQ_EVENT_TYPE_BANDWIDTH
>> : DEVFREQ_EVENT_TYPE_LATENCY
>>
>> [For example]
>> If board dts includes PPMU_DMC0/DMC1/CPU event node,
>> would show following sysfs entry. Also devfreq driver(e.g., exynos4_bus.c)
>> can get the instance of devfreq-event device by using provided API and then
>> get raw data which reflect the current state of device.
>>
>
> Hi,
>
> Overall I like the idea. I understand that now devfreq devices (like
> exynos devfreq) should bind themselves to buses, not to PPMU. It makes
> sense to me because bus clock and bus voltage are properties of bus, not
> monitoring unit.
>
> I see that this is still a RFC so it would be hard to base current
> devfreq work on top of it.

The goal of this patch set is not for devfreq device. This patch set
just is used for
devfreq-event device according to patch description.

The devfreq must need devfreq-event device (e.g., exynos-ppmu.c)
but current devfreq subsystem have not been supported any reason for
exyons-ppmu.c as devfreq-event device.

I discussed this same issue on following patch[1] which was posted to support
exynos4 busfreq by me.
[1] https://lkml.org/lkml/2014/3/14/132
- [PATCHv2 8/8] devfreq: exynos4: Add busfreq driver for exynos4210/exynos4x12
After discussion, I thought that devfreq-event class must be needed
before supporting
the dt of devfreq device.

So, I think that devfreq have to support the devfreq-event subsystem
for exynos-ppmu. And then I'll work to support the device-tree of devfreq device
as next job.

>
> One more general comment: you're adding a some API which is not used by
> current devfreq_event user (exynos). For example the exclusive flag or
> event types. I think it will be simpler to stick to the basic approach
> (reduced API). If some other user emerges then new API will be added.

I think that devfreq-event must support both event_flag(e.g., exclusive flag)
and event_type(e.g., raw_data, utilization, bandwidth, latency) because
devfreq-event device (e.g., exynos-ppmu) would provide various event data.
For example, exynos-ppmu driver could support the utilization/bandwidth/latency
of each IP of Exynos SoC. I checked PPMU IP on Exynos TRM.

Also, If specific devfreq-event device should be used on only one device driver,
devfrewq-event class have to support 'exclusive' flag. If
devfreq-event class don't
support the exclusive flag, devfreq-event could not guarantee the
integrity of event
from devfreq-event device.

Thanks for your comment.

Best Regards,
Chanwoo Choi
--
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/