Re: [PATCH v2 3/3] Input: pwm-beeper: add optional amplifier regulator

[David Lechner <david@lechnology.com>]

On 01/15/2017 06:34 PM, Dmitry Torokhov wrote:
> On Sun, Jan 15, 2017 at 06:12:29PM -0600, David Lechner wrote:
>> On 01/14/2017 01:19 PM, Dmitry Torokhov wrote:
>>> On Wed, Jan 11, 2017 at 02:02:01PM -0600, David Lechner wrote:
>>>> This adds an optional regulator to the pwm-beeper device. This regulator
>>>> acts as an amplifier. The amplifier is only enabled while beeping in order
>>>> to reduce power consumption.
>>>>
>>>> Tested on LEGO MINDSTORMS EV3, which has a speaker connected to PWM through
>>>> an amplifier.
>>>>
>>>> Signed-off-by: David Lechner <david@lechnology.com>
>>>> ---
>>>> drivers/input/misc/pwm-beeper.c | 29 ++++++++++++++++++++++++++++-
>>>> 1 file changed, 28 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/input/misc/pwm-beeper.c b/drivers/input/misc/pwm-beeper.c
>>>> index 30ac227..708e88e 100644
>>>> --- a/drivers/input/misc/pwm-beeper.c
>>>> +++ b/drivers/input/misc/pwm-beeper.c
>>>> @@ -14,6 +14,7 @@
>>>>  */
>>>>
>>>> #include <linux/input.h>
>>>> +#include <linux/regulator/consumer.h>
>>>> #include <linux/module.h>
>>>> #include <linux/kernel.h>
>>>> #include <linux/of.h>
>>>> @@ -25,8 +26,10 @@
>>>> struct pwm_beeper {
>>>> 	struct input_dev *input;
>>>> 	struct pwm_device *pwm;
>>>> +	struct regulator *reg;
>>>> 	struct work_struct work;
>>>> 	unsigned long period;
>>>> +	bool reg_enabled;
>>>> };
>>>>
>>>> #define HZ_TO_NANOSECONDS(x) (1000000000UL/(x))
>>>> @@ -38,8 +41,20 @@ static void __pwm_beeper_set(struct pwm_beeper *beeper)
>>>> 	if (period) {
>>>> 		pwm_config(beeper->pwm, period / 2, period);
>>>> 		pwm_enable(beeper->pwm);
>>>> -	} else
>>>> +		if (beeper->reg) {
>>>> +			int error;
>>>> +
>>>> +			error = regulator_enable(beeper->reg);
>>>> +			if (!error)
>>>> +				beeper->reg_enabled = true;
>>>> +		}
>>>> +	} else {
>>>> +		if (beeper->reg_enabled) {
>>>> +			regulator_disable(beeper->reg);
>>>> +			beeper->reg_enabled = false;
>>>> +		}
>>>> 		pwm_disable(beeper->pwm);
>>>> +	}
>>>> }
>>>>
>>>> static void pwm_beeper_work(struct work_struct *work)
>>>> @@ -82,6 +97,10 @@ static void pwm_beeper_stop(struct pwm_beeper *beeper)
>>>> {
>>>> 	cancel_work_sync(&beeper->work);
>>>>
>>>> +	if (beeper->reg_enabled) {
>>>> +		regulator_disable(beeper->reg);
>>>> +		beeper->reg_enabled = false;
>>>> +	}
>>>> 	if (beeper->period)
>>>> 		pwm_disable(beeper->pwm);
>>>> }
>>>> @@ -111,6 +130,14 @@ static int pwm_beeper_probe(struct platform_device *pdev)
>>>> 		return error;
>>>> 	}
>>>>
>>>> +	beeper->reg = devm_regulator_get_optional(&pdev->dev, "amp");
>>>
>>> If you do not use optional regulator then you will not have to check if
>>> you have it or not everywhere: regulator core will give you a dummy that
>>> you can toggle to your heart's content.
>>
>> Some months ago, I learned that if you are not using device tree and
>> you do not call regulator_has_full_constraints(), then you do not
>> get a dummy regulator. And here, we are only checking if the
>> regulator exists in one place. We will still need the checks for
>> beeper->reg_enabled to keep calls to regulator_enable() and
>> regulator_disable() balanced.
>
> Why? You do not have checks for calls to pwm_enable() and pwm_disable(),
> (or rather beeper->period is used as such flag) why regulator would be
> any different?

regulator_enable() has a __must_check attribute on it, so we get 
compiler warnings if we do not check the return value. Also, if enabling 
the regulator fails and returns an error, then calling 
regulator_disable() later would cause an imbalance.

pwm_enable() and pwm_disable() work differently because they don't count 
how many times they have been called. regulator_enable() and 
regulator_disable(), on the other hand, work like reference counting.

>
>>
>> On the other hand, it is recommended that you always call
>> regulator_has_full_constraints(), so I don't mind changing it if
>> that is what you think we should do. But, I don't really see much of
>> an advantage to changing it compared to the current implementation.
>
> It greatly simplifies control flow in the driver (since I believe you
> can get rid of the flags you introduced).
>
> As far as arch not having full constraints - I am not sure if this makes
> sense anymore. I am not quite sure what the original intent here was, we
> should probably ask Mark Brown. But a lot of drivers do expect the dummy
> substitution to imply work.

I am OK with using the dummy regulator, but I don't see how I can get 
rid of the beeper->reg_enabled flag.