Re: [PATCH 1/7] pwm: Add pwm core driver

[Lars-Peter Clausen]

Arun MURTHY wrote:
>>>> Shouldn't PWM_DEVICES select HAVE_PWM?
>>>
>>> No not required, the entire concept is to remove HAVE_PWM and use
>> PWM_CORE.
>>
>> Well in patch 4 you say that PWM_CORE is currently limited to ARM.
>> Furthermore you
>> change the pwm-backlight and pwm-led Kconfig entries to depend on
>> HAVE_PWM ||
>> PWM_CORE. Adding a select HAVE_PWM here would make those changes
>> unnecessary.
> HAVE_PWM is retained just because the mips pwm driver is not aligned with the pwm core driver.
> On mips pwm driver aligning to the pwm core driver HAVE_PWM will be replaced by PWM_CORE.
> 
>> HAVE_PWM should be set, when pwm_* functions are available. When your
>> pwm-core driver
>> is selected they are available.
> On applying this patch set pwm_* function will be exported in pwm_core driver and in mips pwm driver.
> Since mips pwm driver is not aligned with the pwm core, HAVE_PWM is retained and removed in places where pwm drivers register to pwm core driver.

pwm_{enable,disable,request,free} are the interface of the pwm api. Your pwm-core is
one implementation of that interface. A somewhat special though, because it tries to
be a generic implementation.
There are still other implementations though. For example right now the mips jz4740 one.
In my opinion HAVE_PWM should be defined if there is a implementation for the pwm
interface is available.
I know that your plan is that in the end pwm-core is the only implementation of the
pwm interface.

But right now it is not and on the other hand some SoC implementors might choose that
they want to provide their own minimal pwm interface implementation.
Furthermore this would allow you to start with pwm-core for one SoC which you have on
your desk and where you can properly test things and keep the patches clean from
clutter changing all the different archs.
Once pwm-core is in a proper shape you or other people can start porting all the
different SoC support code to pwm-core.

Similar behavior is for example true for the gpio api. There is gpiolib which is the
generic implementation which allows having gpio chips outside of the chip. On the
other hand there are still archs which choose to have their own gpio api implementation.

> 
>>> pwm_device will be passed to each and every pwm driver that are
>> registered as client with pwm core.
>>> The list consists of the registered pwm drivers and is to be handled
>> by pwm core.
>>> Why should each and every pwm driver get to know about the entire pwm
>> driver list?
>> Declare the list field to be private, by saying that it should only be
>> touched by the
>> core. Right now you allocate a rather small additional structure for
>> each registered
>> device. This could be easily be avoided by embedding the list field
>> into the
>> pwm_device struct.
> 
> The one that is being allocated in register is the pwm_device and this has to. Because each pwm driver will have their own data related to ops, pwm_id.
> Also note that there exists an element "data" that points to the pwm device specific information. Hence this allocation is required.
> 
>>>>> +	}
>>>>> +	pwm->pwm_dev = pwm_dev;
>>>>> +	list_add_tail(&pwm->list, &di->list);
>>>>> +	up_write(&pwm_list_lock);
>>>>> +
>>>> I guess you only need to lock the list when accessing the list and
>>>> adding the new
>>>> pwm_dev.
>>> Oops, thanks for pointing out, will implement this in the v2 patch.
> Coming back to this, I guess the locking has to be done while traversing the list also, as my present pointer in the list my get over written by the time I add an element to list. Please let me know if I am wrong.
> 
>>>>> +struct pwm_ops {
>>>>> +	int (*pwm_config)(struct pwm_device *pwm, int duty_ns, int
>>>> period_ns);
>>>>> +	int (*pwm_enable)(struct pwm_device *pwm);
>>>>> +	int (*pwm_disable)(struct pwm_device *pwm);
>>>>> +	char *name;
>>>>> +};
>>>>> +
>>>> Shouldn't name be part of the pwm_device? That would allow the ops
>> to
>>>> be shared
>>>> between different devices.
>>> Good catch, the reason being that 2 or more devices can share the
>> same ops and get registered to pwm core.
>>> But the catch lies while identifying the pwm device while the clients
>> are requesting for.
>>> The pwm backlight will request the pwm driver by name. This is
>> parameter that distinguishes among different pwm devices irrespective
>> of same ops or not.
>> Yes. And thats why it should go into the pwm_device struct itself.
>>
>> If an additional ops struct is allocated for each device anyway we
>> would be better of
>> embedding it directly into the device struct instead of just holding a
>> pointer to it.
> Yes ops structure will be allocated. But how can we get access to the ops structure of another driver?
> And moreover two pwm driver sharing same ops ideally means a single pwm module. If not everything atleast the pwm registers of two different modules changes. So this scenario can never occur.
> 
>>>>>  #endif /* __LINUX_PWM_H */
>>>> It might be also a good idea to add a device class for pwm devices.
>>> Sure, but can you please explain with an example the use case.
>>>
>> Well, for one it helps to keep data structured.
>> And there would be functions to traverse all devices of a class, so you
>> could get rid
>> of your "di" list.
> Sorry, I didn't get you can you please elaborate more?
> 
Sure. You would create a "struct class" device class for pwm devices. For each
registered pwm device there would then be a "struct device" representing the pwm
device whithin the linux device tree.
This has serveral advantages:
For one you can use this for keeping track of the all the pwm devices instead of
having your custom list. You can use class_for_each_device and class_find_device
instead of traversing your list. This would make the core much simpler and more readable.
Also you can use the device structure for refcounting of modules and devices. Right
now if a pwm-core driver, like the twl6040, is build as a module you can remove the
module while another driver, for example pwm-backlight, is using a pwm device from
the pwm-core driver. Then upon accessing the pwm device from the pwm-backlight driver
the code would crash, because it would access already freed memory.

- Lars


> Thanks and Regards,
> Arun R Murthy
> -------------

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/