Re: [PATCH 0/1] drivers: devfreq: use DELAYED_WORK in DEVFREQ monitoring subsystem

[Chanwoo Choi]

On 1/31/20 9:42 AM, Chanwoo Choi wrote:
> Hi Lukasz,
> 
> On 1/30/20 8:47 PM, Lukasz Luba wrote:
>> Hi Chanwoo, MyungJoo,
>>
>> Gentle ping. The issue is not only in the devfreq itself,
>> but also it affects thermal. The devfreq cooling rely on
>> busy_time and total_time updated by the devfreq monitoring
>> (in simple_ondemand).
>> Thermal uses DELAYED_WORK and is more reliable, but uses stale
>> data from devfreq_dev_stats. It is especially visible when
>> you have cgroup spanning one cluster. Android uses cgroups
>> heavily. You can make easily this setup using 'taskset',
>> run some benchmarks and observe 'devfreq_monitor' traces and
>> timestamps, i.e. for your exynos-bus.
>>
>> The patch is really non-invasive and simple. It can be a good starting
>> point for testing and proposing other solutions.
> 
> Sorry for late reply. I'm preparing the RFC patch about my approach
> to support this requirement as following:
> 
> As you knew, DEFERRABLE_WORK with CONFIG_NO_HZ focuses on removing
> the redundant of power-consumption by preventing the unneeded wakeup
> from idle state if there are no any interrupts and runnable threads.
> 
> Finally, I agree the requirement of delaywd_work for devfreq subsystem.
> But, I would like to support both deferrable_work and delayed_work
> on devfreq subsystem. It is better to select either deferrable_work
> or delayed_work by user like Kamil's suggestion[1].
> [1] https://lore.kernel.org/patchwork/patch/1164317/
> - [2/4] PM / devfreq: add possibility for delayed work
> 
> But, I want to change the timer type for devfreq device
> using simple_ondemand governor via sysfs as following:
> 
> Example:
> 
> 1.
> enum work_timer_type {
> 	DEVFREQ_WORK_TIMER_DEFERRABLE = 0,
> 	DEVFREQ_WORK_TIMER_DELAYED = 0,
> };
> 
> struct devfreq_simple_ondemand_data {
> 	unsigned int upthreshold;
> 	unsigned int downdifferential;
> 	enum work_timer_type timer_type;
> };
> 
> The developer of devfreq device driver can choose
> the default work time type by initializing the 'timer_type of 
> struct devfreq_simple_ondemand_data'.
> 
> 2. Change the work timer type at the runtime
> - Change the work timer type from 'deferrable' to 'delayed'
> $ echo delayed > /sys/class/devfreq/devfreq0/work_timer_type
> $ cat /sys/class/devfreq/devfreq0/work_timer_type
> delayed
> 
> - Change the work timer type from 'delayed' to 'deferrable'
> $ echo deferrable > /sys/class/devfreq/devfreq0/work_timer_type
> $ cat /sys/class/devfreq/devfreq0/work_timer_type
> deferrable
> 

And
Only show '/sys/class/devfreq/devfreq0/work_timer_type' sysfs attribute,
if devfreq device uses the simple_ondemand. Because this 'work_timer_type'
sysfs attribute only depends on simple_ondemand governor and are useful.

So, 'work_timer_type' sysfs attribute will be handled
at drivers/devfreq/governor_simpleondemand.c.

After posting my suggestion, we can discuss it.


> I'm developing the RFC patch and then I'll send it as soon as possible.
> 
>>
>> Regards,
>> Lukasz
>>
>> On 1/27/20 3:17 PM, lukasz.luba@xxxxxxx wrote:
>>> From: Lukasz Luba <lukasz.luba@xxxxxxx>
>>>
>>> Hi all,
>>>
>>> This patch is a continuation of my previous work for fixing DEVFREQ monitoring
>>> subsystem [1]. The issue is around DEFERRABLE_WORK, which uses TIMER_DEFERRABLE
>>> under the hood which will work normally when the system is busy, but will not
>>> cause a CPU to come out of idle and serve the DEVFREQ monitoring requests.
>>>
>>> This is especially important in the SMP systems with many CPUs, when the load
>>> balance tries to keep some CPUs idle. The next service request could not be
>>> triggered when the CPU went idle in the meantime.
>>>
>>> The DELAYED_WORK is going to be triggered even on an idle CPU. This will allow
>>> to call the DEVFREQ monitoring in reliable intervals. Some of the drivers might
>>> use internal counters to monitor their load, when the DEVFREQ work is not
>>> triggered in a predictable way, these counters might overflow leaving the
>>> device in undefined state.
>>>
>>> To observe the difference, the trace output might be used, i.e.
>>>
>>> echo 1 > /sys/kernel/debug/tracing/events/devfreq/enable
>>> #your test starts here, i.e. 'sleep 5' or 'dd ' or 'gfxbench'
>>> echo 0 > /sys/kernel/debug/tracing/events/devfreq/enable
>>> cat /sys/kernel/debug/tracing/trace
>>>
>>> When there are some registered devfreq drivers, you should see the traces
>>> 'devfreq_moniotor' triggered in reliable intervals.
>>>
>>> The patch set is based on Chanwoo's devfreq repository and branch
>>> 'devfreq-next' [2].
>>>
>>> Regards,
>>> Lukasz Luba
>>>
>>> [1] https://protect2.fireeye.com/url?k=d26154c0-8fb20fd4-d260df8f-0cc47a31ce4e-ba68a61e16ee1965&u=https://lkml.org/lkml/2019/2/12/1179
>>> [2] https://git.kernel.org/pub/scm/linux/kernel/git/chanwoo/linux.git/log/?h=devfreq-next
>>>
>>>
>>> Lukasz Luba (1):
>>> ÂÂ drivers: devfreq: add DELAYED_WORK to monitoring subsystem
>>>
>>> Â drivers/devfreq/KconfigÂÂ | 19 +++++++++++++++++++
>>> Â drivers/devfreq/devfreq.c |Â 6 +++++-
>>> Â 2 files changed, 24 insertions(+), 1 deletion(-)
>>>
>>
>>
> 
> 


-- 
Best Regards,
Chanwoo Choi
Samsung Electronics