Re: [PATCH 2/4] PM / devfreq: add possibility for delayed work

[Matthias Kaehlcke]

Hi,

On Mon, Dec 09, 2019 at 03:44:23PM +0100, Kamil Konieczny wrote:
> Current devfreq workqueue uses deferred timer. Introduce sysfs
> file delayed_timer and use it for change from deferred to
> delayed work. The default is to use old deferred one, which
> saves power, but can miss increased demand for higher bus
> frequency if timer was assigned to idle cpu.
> 
> Signed-off-by: Kamil Konieczny <k.konieczny@xxxxxxxxxxx>
> ---
>  Documentation/ABI/testing/sysfs-class-devfreq | 10 ++++
>  drivers/devfreq/devfreq.c                     | 46 ++++++++++++++++++-
>  include/linux/devfreq.h                       |  2 +
>  3 files changed, 57 insertions(+), 1 deletion(-)
> 
> diff --git a/Documentation/ABI/testing/sysfs-class-devfreq b/Documentation/ABI/testing/sysfs-class-devfreq
> index 9758eb85ade3..07bfd0df6a4a 100644
> --- a/Documentation/ABI/testing/sysfs-class-devfreq
> +++ b/Documentation/ABI/testing/sysfs-class-devfreq
> @@ -30,6 +30,16 @@ Description:
>  		target_freq when get_cur_freq() is not implemented by
>  		devfreq driver.
>  
> +What:		/sys/class/devfreq/.../delayed_timer
> +Date:		December 2019
> +Contact:	Kamil Konieczny <k.konieczny@xxxxxxxxxxx>
> +Description:
> +		This ABI shows or clears timer type used by devfreq
> +		workqueue. When 0, it uses default deferred timer.
> +		When set to 1 devfreq will use delayed timer. Example
> +		useage:
> +			echo 1 > /sys/class/devfreq/.../delayed_timer
> +
>  What:		/sys/class/devfreq/.../target_freq
>  Date:		September 2012
>  Contact:	Rajagopal Venkat <rajagopal.venkat@xxxxxxxxxx>
> diff --git a/drivers/devfreq/devfreq.c b/drivers/devfreq/devfreq.c
> index 955949c6fc1f..c277d1770fef 100644
> --- a/drivers/devfreq/devfreq.c
> +++ b/drivers/devfreq/devfreq.c
> @@ -445,7 +445,11 @@ void devfreq_monitor_start(struct devfreq *devfreq)
>  	if (devfreq->governor->interrupt_driven)
>  		return;
>  
> -	INIT_DEFERRABLE_WORK(&devfreq->work, devfreq_monitor);
> +	if (devfreq->delayed_timer)
> +		INIT_DELAYED_WORK(&devfreq->work, devfreq_monitor);
> +	else
> +		INIT_DEFERRABLE_WORK(&devfreq->work, devfreq_monitor);
> +
>  	if (devfreq->profile->polling_ms)
>  		queue_delayed_work(devfreq_wq, &devfreq->work,
>  			msecs_to_jiffies(devfreq->profile->polling_ms));
> @@ -698,6 +702,7 @@ struct devfreq *devfreq_add_device(struct device *dev,
>  	devfreq->last_status.current_frequency = profile->initial_freq;
>  	devfreq->data = data;
>  	devfreq->nb.notifier_call = devfreq_notifier_call;
> +	devfreq->delayed_timer = false;

devfreq is zero initialized (allocated with kzalloc()), hence this is
unnecessary.

>  
>  	if (!devfreq->profile->max_state && !devfreq->profile->freq_table) {
>  		mutex_unlock(&devfreq->lock);
> @@ -1288,6 +1293,44 @@ static ssize_t available_governors_show(struct device *d,
>  }
>  static DEVICE_ATTR_RO(available_governors);
>  
> +static ssize_t delayed_timer_show(struct device *dev,
> +				  struct device_attribute *attr, char *buf)
> +{
> +	int i;
> +
> +	i = to_devfreq(dev)->delayed_timer ? 1 : 0;
> +	return sprintf(buf, "%d\n", i);

get rid of 'i' and just use df->delayed_timer in sprintf().

> +}
> +
> +static ssize_t delayed_timer_store(struct device *dev,
> +				   struct device_attribute *attr,
> +				   const char *buf, size_t count)
> +{
> +	struct devfreq *df = to_devfreq(dev);
> +	bool old_timer;

Not a great name, the variable doesn't hold a timer. I'd suggest something
like 'prev_val'.

> +	int value, ret;
> +
> +	if (!df->governor)
> +		return -EINVAL;
> +
> +	ret = kstrtoint(buf, 10, &value);
> +	if (ret || (value != 1 && value != 0))
> +		return -EINVAL;

use kstrtobool() instead of partially re-implementing it.

> +
> +	mutex_lock(&df->lock);
> +	old_timer = df->delayed_timer;
> +	df->delayed_timer = value == 0 ? false : true;

What's wrong with:

df->delayed_timer = value;

?

> +	mutex_unlock(&df->lock);

Does the locking as is actually provide any value? The use case seems to
be concurrent setting of the sysfs attribute. The lock is released after
the assignment, hence the value of 'df->delayed_timer' could be overwritten
before the condition below is evaluated.

If you want to protect against this case you need something like this:

// don't release the lock before evaluating the condition

> +	if (old_timer != df->delayed_timer) {
  	   	mutex_unlock(&df->lock);
> +		devfreq_monitor_stop(df);
> +		devfreq_monitor_start(df);
> +	}
  	else {
		mutex_unlock(&df->lock);
	}

I don't pretend it's pretty ;-)