Re: [PATCH] gpio-mvebu: no hardcoded timer assignment for pwm

[Finn Behrens]

Hi,

Sorry for taking so so long, life happened.
Currently working on a v2 but maybe reasonable to discuss some things already here.

(And sorry for first sending html, to long not worked on kernel and trust my mail client to much, so second resend without html)

> On 8. Feb 2024, at 09:05, Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx> wrote:
>
> Hello,
>
> On Tue, Jan 30, 2024 at 11:55:13AM +0100, Finn Behrens wrote:
>> Removes the hardcoded timer assignment of timers to pwm controllers.
>> This allows to use more than one pwm per gpio bank.
>>
>> Original patch with chip_data interface by Heisath <jannis@xxxxxxxxxx <mailto:jannis@xxxxxxxxxx>>
>>
>> Link: https://wiki.kobol.io/helios4/pwm/#patch-requirement
>> Co-developed-by: Yureka Lilian <yuka@xxxxxxxx <mailto:yuka@xxxxxxxx>>
>> Signed-off-by: Yureka Lilian <yuka@xxxxxxxx <mailto:yuka@xxxxxxxx>>
>> Signed-off-by: Finn Behrens <me@xxxxxxxxx <mailto:me@xxxxxxxxx>>
>
> I find this patch hard to understand and I hope it's more complicated
> than it could be. I wonder if it would be beneficial to split this patch
> in two. In the first patch just introduce the new structures with all
> the necessary renaming and only in the second patch implement the added
> flexibility.
I will try for the v2, currently not sure how easy that will be as most things are used directly and will not work when not used.
But might have to rewrite quite a bit as think I found a possible race condition in this patch. (further down below with the static variable)
>
> Some more details below.
>
>> drivers/gpio/gpio-mvebu.c | 223 ++++++++++++++++++++++++--------------
>> 1 file changed, 139 insertions(+), 84 deletions(-)
>>
>> diff --git a/drivers/gpio/gpio-mvebu.c b/drivers/gpio/gpio-mvebu.c
>> index a13f3c18ccd4..303ea3be0b69 100644
>> --- a/drivers/gpio/gpio-mvebu.c
>> +++ b/drivers/gpio/gpio-mvebu.c
>> @@ -94,21 +94,43 @@
>>
>> #define MVEBU_MAX_GPIO_PER_BANK		32
>>
>> -struct mvebu_pwm {
>> +enum mvebu_pwm_ctrl {
>> +	MVEBU_PWM_CTRL_SET_A = 0,
>> +	MVEBU_PWM_CTRL_SET_B,
>> +	MVEBU_PWM_CTRL_MAX
>> +};
>> +
>> +struct mvebu_pwmchip {
>> 	struct regmap		*regs;
>> 	u32			 offset;
>> 	unsigned long		 clk_rate;
>> -	struct gpio_desc	*gpiod;
>> -	struct pwm_chip		 chip;
>> 	spinlock_t		 lock;
>> -	struct mvebu_gpio_chip	*mvchip;
>> +	bool			 in_use;
>>
>> 	/* Used to preserve GPIO/PWM registers across suspend/resume */
>> -	u32			 blink_select;
>> 	u32			 blink_on_duration;
>> 	u32			 blink_off_duration;
>> };
>>
>> +struct mvebu_pwm_chip_drv {
>> +	enum mvebu_pwm_ctrl	 ctrl;
>> +	struct gpio_desc	*gpiod;
>> +	bool			 master;
>> +};
>> +
>> +struct mvebu_pwm {
>> +	struct pwm_chip		 chip;
>> +	struct mvebu_gpio_chip	*mvchip;
>> +	struct mvebu_pwmchip	 controller;
>> +	enum mvebu_pwm_ctrl	 default_controller;
>> +
>> +	/* Used to preserve GPIO/PWM registers across suspend/resume */
>> +	u32				 blink_select;
>> +	struct mvebu_pwm_chip_drv	 drv[];
>> +};
>
> So we have three different structures related to pwm. Some highlevel
> description (in a comment or at least the commit log) about how the
> hardware works and which struct describes what would be helpful. I gave
> up after 15 min of reading this patch and trying to understand it.
>
>> +static struct mvebu_pwmchip  *mvebu_pwm_list[MVEBU_PWM_CTRL_MAX];
>
> Huh, a static variable. Does that mean we can only have one mvebu_gpio
> device?
This is the global over the gpio aliases gpio0 and gpio1.
As far as I understand this we take the id from the alias in the probe function and there we can only ever have the id 0 and 1.

While looking over this again I noticed that we therefore don’t have a working lock with this, and now gpio0 can take the pwm from gpio1 which is the main idea behind this patch.
My idea would be convert the lock from the pwm chip to a global lock in this module and use that for all pwm locks and this global list. I don’t know if there is a less global variable that we can access from both gpio0 and gpio1?
>
>> +
>> struct mvebu_gpio_chip {
>> 	struct gpio_chip   chip;
>> 	struct regmap     *regs;
>> @@ -285,12 +307,12 @@ mvebu_gpio_write_level_mask(struct mvebu_gpio_chip *mvchip, u32 val)
>>  * Functions returning offsets of individual registers for a given
>>  * PWM controller.
>>  */
>> -static unsigned int mvebu_pwmreg_blink_on_duration(struct mvebu_pwm *mvpwm)
>> +static unsigned int mvebu_pwmreg_blink_on_duration(struct mvebu_pwmchip *mvpwm)
>
> I'm a fan of picking always the same variable name for the same thing
> and different names for different things. "mvpwm" is used for variables
> of type struct mvebu_pwmchip and struct mvebu_pwm.
>
>> {
>> 	return mvpwm->offset + PWM_BLINK_ON_DURATION_OFF;
>> }
>
> Best regards
> Uwe
>
> -- 
> Pengutronix e.K.                           | Uwe Kleine-König            |
> Industrial Linux Solutions                 | https://www.pengutronix.de/ |