Re: [RFC][PATCH 1/3] PM /devfreq: add user frequency limits into devfreq struct

From: Lukasz Luba
Date: Tue Feb 16 2021 - 05:42:15 EST


Hi Chanwoo,

On 2/15/21 3:00 PM, Chanwoo Choi wrote:
Hi Lukasz,

On Fri, Feb 12, 2021 at 7:28 AM Lukasz Luba <lukasz.luba@xxxxxxx> wrote:



On 2/11/21 11:07 AM, Lukasz Luba wrote:
Hi Chanwoo,

On 2/3/21 10:21 AM, Lukasz Luba wrote:
Hi Chanwoo,

Thank you for looking at this.

On 2/3/21 10:11 AM, Chanwoo Choi wrote:
Hi Lukasz,

When accessing the max_freq and min_freq at devfreq-cooling.c,
even if can access 'user_max_freq' and 'lock' by using the 'devfreq'
instance,
I think that the direct access of variables
(lock/user_max_freq/user_min_freq)
of struct devfreq are not good.

Instead, how about using the 'DEVFREQ_TRANSITION_NOTIFIER'
notification with following changes of 'struct devfreq_freq'?

I like the idea with devfreq notification. I will have to go through the
code to check that possibility.

Also, need to add codes into devfreq_set_target() for initializing
'new_max_freq' and 'new_min_freq' before sending the DEVFREQ_POSTCHANGE
notification.

diff --git a/include/linux/devfreq.h b/include/linux/devfreq.h
index 147a229056d2..d5726592d362 100644
--- a/include/linux/devfreq.h
+++ b/include/linux/devfreq.h
@@ -207,6 +207,8 @@ struct devfreq {
struct devfreq_freqs {
unsigned long old;
unsigned long new;
+ unsigned long new_max_freq;
+ unsigned long new_min_freq;
};


And I think that new 'user_min_freq'/'user_max_freq' are not necessary.
You can get the current max_freq/min_freq by using the following steps:

get_freq_range(devfreq, &min_freq, &max_freq);
dev_pm_opp_find_freq_floor(pdev, &min_freq);
dev_pm_opp_find_freq_floor(pdev, &max_freq);

So that you can get the 'max_freq/min_freq' and then
initialize the 'freqs.new_max_freq and freqs.new_min_freq'
with them as following:

in devfreq_set_target()
get_freq_range(devfreq, &min_freq, &max_freq);
dev_pm_opp_find_freq_floor(pdev, &min_freq);
dev_pm_opp_find_freq_floor(pdev, &max_freq);
freqs.new_max_freq = min_freq;
freqs.new_max_freq = max_freq;
devfreq_notify_transition(devfreq, &freqs, DEVFREQ_POSTCHANGE);

I will plumb it in and check that option. My concern is that function
get_freq_range() would give me the max_freq value from PM QoS, which
might be my thermal limit - lower that user_max_freq. Then I still
need

I've been playing with PM QoS notifications because I thought it would
be possible to be notified in thermal for all new set values - even from
devfreq sysfs user max_freq write, which has value higher that the
current limit set by thermal governor. Unfortunately PM QoS doesn't
send that information by design. PM QoS also by desing won't allow
me to check first two limits in the plist - which would be thermal
and user sysfs max_freq.

I will experiment with this notifications and share the results.
That you for your comments.

I have experimented with your proposal. Unfortunately, the value stored
in the pm_qos which is read by get_freq_range() is not the user max
freq. It's the value from thermal devfreq cooling when that one is
lower. Which is OK in the overall design, but not for my IPA use case.

What comes to my mind is two options:
1) this patch proposal, with simple solution of two new variables
protected by mutex, which would maintain user stored values
2) add a new notification chain in devfreq to notify about new
user written value, to which devfreq cooling would register; that
would allow devfreq cooling to get that value instantly and store
locally

3) How about new define for existing notification chain:
#define DEVFREQ_USER_CHANGE (2)

I think that if we add the notification with specific actor like user change
or OPP change or others, it is not proper. But, we can add the notification
for min or max frequency change timing. Because the devfreq already has
the notification for current frequency like DEVFREQ_PRECHANGE,
DEVFREQ_POSTCHANGE.

Maybe, we can add the following notification for min/max_freq.
The following min_freq and max_freq values will be calculated by
get_freq_range().
DEVFREQ_MIN_FREQ_PRECHANGE
DEVFREQ_MIN_FREQ_POSTCHANGE
DEVFREQ_MAX_FREQ_PRECHANGE
DEVFREQ_MAX_FREQ_POSTCHANGE

Would it be then possible to pass the user max freq value written via
sysfs? Something like in the example below, when writing into max sysfs:

1) starting in max_freq_store()
freqs.new_max_freq = max_freq;
devfreq_notify_transition(devfreq, &freqs, DEVFREQ_MAX_FREQ_PRECHANGE);
dev_pm_qos_update_request()

2)then after a while in devfreq_set_target()
get_freq_range(devfreq, &min_freq, &max_freq);
dev_pm_opp_find_freq_floor(pdev, &min_freq);
dev_pm_opp_find_freq_floor(pdev, &max_freq);
freqs.new_min_freq = min_freq;
freqs.new_max_freq = max_freq;
devfreq_notify_transition(devfreq, &freqs, DEVFREQ_MAX_FREQ_POSTCHANGE);

This 2nd part is called after the PM QoS has changed that limit,
so might be missing (in case value was higher that current),
but thermal would know about that, so no worries.




Then a modified devfreq_notify_transition() would get:
@@ -339,6 +339,10 @@ static int devfreq_notify_transition(struct devfreq
*devfreq,

srcu_notifier_call_chain(&devfreq->transition_notifier_list,
DEVFREQ_POSTCHANGE, freqs);
break;
+ case DEVFREQ_USER_CHANGE:
+ srcu_notifier_call_chain(&devfreq->transition_notifier_list,
+ DEVFREQ_USER_CHANGE, freqs);
+ break;
default:
return -EINVAL;
}

If that is present, I can plumb your suggestion with:
struct devfreq_freq {
+ unsigned long new_max_freq;
+ unsigned long new_min_freq;

and populate them with values in the max_freq_store() by adding at the
end:

freqs.new_max_freq = max_freq;
mutex_lock();
devfreq_notify_transition(devfreq, &freqs, DEVFREQ_USER_CHANGE);
mutex_unlock();

I would handle this notification in devfreq cooling and keep the
value there, for future IPA checks.

If you agree, I can send next version of the patch set.


What do you think Chanwoo?

I thought that your suggestion to expose the user input for min/max_freq.
But, these values are not valid for the public user. Actually, the devfreq core
handles these values only internally without any explicit access from outside.

I'm not sure that it is right or not to expose the internal value of
devfreq struct.
Until now, I think that it is not proper to show the interval value outside.

Because the devfreq subsystem only provides the min_freq and max_freq
which reflect the all requirement of user input/cooling policy/OPP
instead of user_min_freq, user_max_freq.

If we provide the user_min_freq, user_max_freq via DEVFREQ notification,
we have to make the new sysfs attributes for user_min_freq and user_max_freq
to show the value to the user. But, it seems that it is not nice.

I would say we don't have to expose it. Let's take a closer look into
an example. The main problem is with GPUs. The middleware is aware of
the OPPs in the GPU. If the middleware wants to switch into different
power-performance mode e.g. power-saving, it writes into this sysfs
the max allowed freq. IPA does not know about it and makes wrong
decisions. As you said, the sysfs read operation combines all:
user input/cooling policy/OPP, but that's not a problem for this aware
middleware. So it can stay as is.
The only addition would be this 'notification about user attempt of
reducing max device speed' internally inside the kernel, for those
subsystems which are interested in it.


Actually, I have no other idea how to support your feature.
We try to find the more proper method.


Thank you for coming back with your comments. I know it's not
an easy feature but I hope we can find a solution.

Regards,
Lukasz