Re: [PATCH 1/2] thermal: devfreq_cooling: Use PM QoS to set frequency limits

[Leonard Crestez]

On 10.01.2020 19:49, Matthias Kaehlcke wrote:
> Now that devfreq supports limiting the frequency range of a device
> through PM QoS make use of it instead of disabling OPPs that should
> not be used.
> 
> Signed-off-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
> ---

It is not obvious but this changes behavior when min max requests 
conflict (min > max): with PM QoS a MIN_FREQUENCY request takes 
precedence but if higher OPPs are disabled then this will override 
MIN_FREQUENCY.

There are very few users of this functionality so I don't think there 
are any systems that depend on this behaving one way or the other but 
perhaps it should be mentioned in commit message?

As far as I can tell the only user of devfreq_cooling in upstream is 
drivers/gpu/drm/panfrost?

>   drivers/thermal/devfreq_cooling.c | 66 ++++++++++---------------------
>   1 file changed, 20 insertions(+), 46 deletions(-)
> 
> diff --git a/drivers/thermal/devfreq_cooling.c b/drivers/thermal/devfreq_cooling.c
> index ef59256887ff..3a63603afcf2 100644
> --- a/drivers/thermal/devfreq_cooling.c
> +++ b/drivers/thermal/devfreq_cooling.c
> @@ -24,11 +24,13 @@
>   #include <linux/idr.h>
>   #include <linux/slab.h>
>   #include <linux/pm_opp.h>
> +#include <linux/pm_qos.h>
>   #include <linux/thermal.h>
>   
>   #include <trace/events/thermal.h>
>   
> -#define SCALE_ERROR_MITIGATION 100
> +#define HZ_PER_KHZ		1000
> +#define SCALE_ERROR_MITIGATION	100
>   
>   static DEFINE_IDA(devfreq_ida);
>   
> @@ -65,49 +67,9 @@ struct devfreq_cooling_device {
>   	struct devfreq_cooling_power *power_ops;
>   	u32 res_util;
>   	int capped_state;
> +	struct dev_pm_qos_request req_max_freq;
>   };
>   
> -/**
> - * partition_enable_opps() - disable all opps above a given state
> - * @dfc:	Pointer to devfreq we are operating on
> - * @cdev_state:	cooling device state we're setting
> - *
> - * Go through the OPPs of the device, enabling all OPPs until
> - * @cdev_state and disabling those frequencies above it.
> - */
> -static int partition_enable_opps(struct devfreq_cooling_device *dfc,
> -				 unsigned long cdev_state)
> -{
> -	int i;
> -	struct device *dev = dfc->devfreq->dev.parent;
> -
> -	for (i = 0; i < dfc->freq_table_size; i++) {
> -		struct dev_pm_opp *opp;
> -		int ret = 0;
> -		unsigned int freq = dfc->freq_table[i];
> -		bool want_enable = i >= cdev_state ? true : false;
> -
> -		opp = dev_pm_opp_find_freq_exact(dev, freq, !want_enable);
> -
> -		if (PTR_ERR(opp) == -ERANGE)
> -			continue;
> -		else if (IS_ERR(opp))
> -			return PTR_ERR(opp);
> -
> -		dev_pm_opp_put(opp);
> -
> -		if (want_enable)
> -			ret = dev_pm_opp_enable(dev, freq);
> -		else
> -			ret = dev_pm_opp_disable(dev, freq);
> -
> -		if (ret)
> -			return ret;
> -	}
> -
> -	return 0;
> -}
> -
>   static int devfreq_cooling_get_max_state(struct thermal_cooling_device *cdev,
>   					 unsigned long *state)
>   {
> @@ -134,7 +96,7 @@ static int devfreq_cooling_set_cur_state(struct thermal_cooling_device *cdev,
>   	struct devfreq_cooling_device *dfc = cdev->devdata;
>   	struct devfreq *df = dfc->devfreq;
>   	struct device *dev = df->dev.parent;
> -	int ret;
> +	unsigned long freq;
>   
>   	if (state == dfc->cooling_state)
>   		return 0;
> @@ -144,9 +106,10 @@ static int devfreq_cooling_set_cur_state(struct thermal_cooling_device *cdev,
>   	if (state >= dfc->freq_table_size)
>   		return -EINVAL;
>   
> -	ret = partition_enable_opps(dfc, state);
> -	if (ret)
> -		return ret;
> +	freq = dfc->freq_table[state];
> +
> +	dev_pm_qos_update_request(&dfc->req_max_freq,
> +				  DIV_ROUND_UP(freq, HZ_PER_KHZ));
>   
>   	dfc->cooling_state = state;
>   
> @@ -529,6 +492,12 @@ of_devfreq_cooling_register_power(struct device_node *np, struct devfreq *df,
>   	if (err)
>   		goto free_dfc;
>   
> +	err = dev_pm_qos_add_request(df->dev.parent, &dfc->req_max_freq,
> +				     DEV_PM_QOS_MAX_FREQUENCY,
> +				     PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
> +	if (err < 0)
> +		goto remove_qos_req;
> +
>   	err = ida_simple_get(&devfreq_ida, 0, 0, GFP_KERNEL);
>   	if (err < 0)
>   		goto free_tables;
> @@ -552,6 +521,10 @@ of_devfreq_cooling_register_power(struct device_node *np, struct devfreq *df,
>   
>   release_ida:
>   	ida_simple_remove(&devfreq_ida, dfc->id);
> +
> +remove_qos_req:
> +	dev_pm_qos_remove_request(&dfc->req_max_freq); > +

A quirk of the dev_pm_qos API is that dev_pm_qos_remove_request prints a 
WARN splat if !dev_pm_qos_request_active and this can true on 
dev_pm_qos_add_request error.

I dealt with this by checking dev_pm_qos_request_active explicitly but 
perhaps dev_pm_qos API could be changed? In general "free/release" 
functions shouldn't complain if there's nothing to do.

>   free_tables:
>   	kfree(dfc->power_table);
>   	kfree(dfc->freq_table);
> @@ -600,6 +573,7 @@ void devfreq_cooling_unregister(struct thermal_cooling_device *cdev)
>   
>   	thermal_cooling_device_unregister(dfc->cdev);
>   	ida_simple_remove(&devfreq_ida, dfc->id);
> +	dev_pm_qos_remove_request(&dfc->req_max_freq);
>   	kfree(dfc->power_table);
>   	kfree(dfc->freq_table);