Re: [PATCH] PM / devfreq: Add sysfs attributes to simple_ondemand governor

[Lukasz Luba <lukasz.luba@arm.com>]

On 1/18/21 6:14 PM, Greg KH wrote:
> On Mon, Jan 18, 2021 at 05:56:06PM +0000, Lukasz Luba wrote:
>>
>>
>> On 1/18/21 5:17 PM, Greg KH wrote:
>>> On Fri, Jan 15, 2021 at 05:05:30PM +0000, Lukasz Luba wrote:
>>>> The simple_ondemand devfreq governor is used by quite a few devices, like
>>>> GPUs, DSPs, memory controllers, etc. It implements algorithm which tries
>>>> to predict the device frequency based on past statistics. There are two
>>>> tunables for the algorithm: 'upthreshold' and 'downdifferential'. These
>>>> tunables change the behavior of the decision, e.g. how fast to increase
>>>> the frequency or how rapidly limit the frequency. These values might be
>>>> different based on the application which is currently running, e.g.
>>>> different behavior is needed for a game than for web browsing or clean
>>>> desktop. The patch exports these two tunables so they can be adjusted
>>>> based on current need. There is also a check with the allowed ranges
>>>> to make sure the values are correct and safe.
>>>>
>>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>>>> ---
>>>>    drivers/devfreq/governor_simpleondemand.c | 135 ++++++++++++++++++++++
>>>>    1 file changed, 135 insertions(+)
>>>>
>>>> diff --git a/drivers/devfreq/governor_simpleondemand.c b/drivers/devfreq/governor_simpleondemand.c
>>>> index d57b82a2b570..4b3c182e0a49 100644
>>>> --- a/drivers/devfreq/governor_simpleondemand.c
>>>> +++ b/drivers/devfreq/governor_simpleondemand.c
>>>> @@ -15,6 +15,7 @@
>>>>    /* Default constants for DevFreq-Simple-Ondemand (DFSO) */
>>>>    #define DFSO_UPTHRESHOLD	(90)
>>>>    #define DFSO_DOWNDIFFERENCTIAL	(5)
>>>> +#define DFSO_MAX_VALUE		(100)
>>>>    static int devfreq_simple_ondemand_func(struct devfreq *df,
>>>>    					unsigned long *freq)
>>>>    {
>>>> @@ -84,15 +85,149 @@ static int devfreq_simple_ondemand_func(struct devfreq *df,
>>>>    	return 0;
>>>>    }
>>>> +static ssize_t upthreshold_show(struct device *dev,
>>>> +				struct device_attribute *attr, char *buf)
>>>> +{
>>>> +	struct devfreq_simple_ondemand_data *data;
>>>> +	struct devfreq *df = to_devfreq(dev);
>>>> +
>>>> +	if (!df->data)
>>>> +		return -EINVAL;
>>>> +
>>>> +	data = df->data;
>>>> +
>>>> +	return sprintf(buf, "%d\n", data->upthreshold);
>>>
>>> sysfs_emit()?
>>>
>>> Also, you forgot the Documentation/ABI/ updates for new sysfs files :(
>>
>> True, I will remember next time.
>>
>>>
>>>
>>>> +}
>>>> +
>>>> +static ssize_t upthreshold_store(struct device *dev,
>>>> +				 struct device_attribute *attr,
>>>> +				 const char *buf, size_t count)
>>>> +{
>>>> +	struct devfreq_simple_ondemand_data *data;
>>>> +	struct devfreq *df = to_devfreq(dev);
>>>> +	unsigned int value;
>>>> +	int ret;
>>>> +
>>>> +	if (!df->data)
>>>> +		return -EINVAL;
>>>> +
>>>> +	data = df->data;
>>>> +
>>>> +	ret = kstrtouint(buf, 10, &value);
>>>> +	if (ret < 0)
>>>> +		return -EINVAL;
>>>> +
>>>> +	mutex_lock(&df->lock);
>>>> +
>>>> +	if (value > DFSO_MAX_VALUE || value <= data->downdifferential) {
>>>> +		mutex_unlock(&df->lock);
>>>> +		return -EINVAL;
>>>> +	}
>>>> +
>>>> +	data->upthreshold = value;
>>>> +	mutex_unlock(&df->lock);
>>>> +
>>>> +	return count;
>>>> +}
>>>> +static DEVICE_ATTR_RW(upthreshold);
>>>> +
>>>> +static ssize_t downdifferential_show(struct device *dev,
>>>> +				     struct device_attribute *attr, char *buf)
>>>> +{
>>>> +	struct devfreq_simple_ondemand_data *data;
>>>> +	struct devfreq *df = to_devfreq(dev);
>>>> +
>>>> +	if (!df->data)
>>>> +		return -EINVAL;
>>>> +
>>>> +	data = df->data;
>>>> +
>>>> +	return sprintf(buf, "%d\n", data->downdifferential);
>>>> +}
>>>> +
>>>> +static ssize_t downdifferential_store(struct device *dev,
>>>> +				      struct device_attribute *attr,
>>>> +				      const char *buf, size_t count)
>>>> +{
>>>> +	struct devfreq_simple_ondemand_data *data;
>>>> +	struct devfreq *df = to_devfreq(dev);
>>>> +	unsigned int value;
>>>> +	int ret;
>>>> +
>>>> +	if (!df->data)
>>>> +		return -EINVAL;
>>>> +
>>>> +	data = df->data;
>>>> +
>>>> +	ret = kstrtouint(buf, 10, &value);
>>>> +	if (ret < 0)
>>>> +		return -EINVAL;
>>>> +
>>>> +	mutex_lock(&df->lock);
>>>> +
>>>> +	if (value > DFSO_MAX_VALUE || value >= data->upthreshold) {
>>>> +		mutex_unlock(&df->lock);
>>>> +		return -EINVAL;
>>>> +	}
>>>> +
>>>> +	data->downdifferential = value;
>>>> +	mutex_unlock(&df->lock);
>>>> +
>>>> +	return count;
>>>> +}
>>>> +static DEVICE_ATTR_RW(downdifferential);
>>>> +
>>>> +static void devfreq_simple_ondemand_sysfs_setup(struct devfreq *df)
>>>> +{
>>>> +	struct devfreq_simple_ondemand_data *data;
>>>> +	int ret;
>>>> +
>>>> +	if (!df->data) {
>>>> +		/* The memory will be freed automatically */
>>>> +		df->data = devm_kzalloc(&df->dev,
>>>> +				sizeof(struct devfreq_simple_ondemand_data),
>>>> +				GFP_KERNEL);
>>>> +		if (!df->data) {
>>>> +			dev_warn(&df->dev, "Unable to allocate memory");
>>>> +			return;
>>>> +		}
>>>> +	}
>>>> +
>>>> +	data = df->data;
>>>> +
>>>> +	/* After new allocation setup default values, since they are used */
>>>> +	if (!data->upthreshold)
>>>> +		data->upthreshold = DFSO_UPTHRESHOLD;
>>>> +
>>>> +	if (!data->downdifferential)
>>>> +		data->downdifferential = DFSO_DOWNDIFFERENCTIAL;
>>>> +
>>>> +	ret = sysfs_create_file(&df->dev.kobj, &dev_attr_upthreshold.attr);
>>>> +	if (ret < 0)
>>>> +		dev_warn(&df->dev, "Unable to create 'upthreshold' attr\n");
>>>> +
>>>> +	ret = sysfs_create_file(&df->dev.kobj, &dev_attr_downdifferential.attr);
>>>> +	if (ret < 0)
>>>> +		dev_warn(&df->dev, "Unable to create 'downdifferential' attr\n");
>>>
>>> You just raced with userspace and lost :(
>>>
>>> Please use the default sysfs groups so that it all works properly.
>>>
>>> Huge hint, when calling sysfs_* from a driver, that usually means you
>>> are doing something wrong.
>>
>> I should have used kobject_init_and_add() and the default devfreq
>> sysfs group as a parent.
> 
> No, never use "raw" kobjects in the sys/devices/ tree, that is a sure
> way to ensure that userspace will never be notified or see your
> attributes.

Interesting, I've seen it in the schedutil governor, but might in a
different context.

Regards,
Lukasz