Re: [PATCH v3 5/8] reset: imx8mp-audiomix: Switch to using regmap API

[Laurentiu Mihalcea <laurentiumihalcea111@gmail.com>]

On 10/30/2025 6:06 AM, Philipp Zabel wrote:
> On Mi, 2025-10-29 at 06:52 -0700, Laurentiu Mihalcea wrote:
>> From: Laurentiu Mihalcea <laurentiu.mihalcea@nxp.com>
>>
>> As far as the Linux kernel is concerned, block devices such as i.MX8MP's
>> AUDIOMIX block control or i.MX8ULP's SIM LPAV can simultaneously act as
>> clock controllers, reset controllers or mux controllers. Since these IPs
>> offer different functionalities through different subsystem APIs, it's
>> important to make sure that the register R-M-W cycles are performed under
>> the same lock across all subsystem APIs. This will ensure that registers
>> will not end up with the wrong values because of race conditions (e.g.
>> clock consumer tries to update block control register A, while, at the
>> same time, reset consumer tries to update the same block control register).
>>
>> However, the aforementioned race conditions will only impact block control
>> IPs which use the same register for multiple functionalities. For example,
>> i.MX8MP's AUDIOMIX block control IP provides clock gating functionalities
>> and reset control functionalities through different registers. This is why
>> the current approach (i.e. clock control and reset control work using
>> different locks) has worked well so far.
>>
>> Since we want to extend this driver to be usable for i.MX8ULP's SIM LPAV
>> block control IP, we need to make sure that clock control, reset control,
>> and mux control APIs use the same lock since all of these functionalities
>> are performed using the SYSCTRL0 register.
>>
>> To do so, we need to switch to the regmap API and, if possible, use the
>> parent device's regmap, which, in the case of i.MX8ULP, will be the clock
>> controller. This way, we can make sure that the clock gates and the reset
>> controller will use the same lock to perform the register R-M-W cycles.
>>
>> This change will also work fine for cases where we don't really need to
>> share the lock across multiple APIs (e.g. i.MX8MP's AUDIOMIX block
>> control) since regmap will take care of the locking we were previously
>> explicitly performing in the driver.
>>
>> The transition to the regmap API also involves some cleanup. Specifically,
>> we can make use of devres to unmap the device's memory and get rid of the
>> memory mapping-related error paths and the remove() function altogether.
>>
>> Signed-off-by: Laurentiu Mihalcea <laurentiu.mihalcea@nxp.com>
>> ---
>>  drivers/reset/reset-imx8mp-audiomix.c | 91 +++++++++++++++++----------
>>  1 file changed, 57 insertions(+), 34 deletions(-)
>>
>> diff --git a/drivers/reset/reset-imx8mp-audiomix.c b/drivers/reset/reset-imx8mp-audiomix.c
>> index e9643365a62c..3f6d11270918 100644
>> --- a/drivers/reset/reset-imx8mp-audiomix.c
>> +++ b/drivers/reset/reset-imx8mp-audiomix.c
>> @@ -11,6 +11,7 @@
>>  #include <linux/module.h>
>>  #include <linux/of.h>
>>  #include <linux/of_address.h>
>> +#include <linux/regmap.h>
>>  #include <linux/reset-controller.h>
>>  
>>  #define IMX8MP_AUDIOMIX_EARC_RESET_OFFSET	0x200
>> @@ -42,8 +43,8 @@ static const struct imx8mp_reset_map reset_map[] = {
>>  
>>  struct imx8mp_audiomix_reset {
>>  	struct reset_controller_dev rcdev;
>> -	spinlock_t lock; /* protect register read-modify-write cycle */
>>  	void __iomem *base;
> Drop base as well, better let devres handle this.
>
> [...]
>
>> +/* assumption: registered only if not using parent regmap */
>> +static void imx8mp_audiomix_reset_iounmap(void *data)
> Pass base instead of dev.


ACK. Will need some extra casts/annotations though as sparse will definitely

not be happy with this.


>
>> +{
>> +	struct imx8mp_audiomix_reset *priv = dev_get_drvdata(data);
>> +
>> +	iounmap(priv->base);
> 	void __iomem *base = data;
>
> 	iounmap(base);
>
>> +}
>> +
>> +/* assumption: dev_set_drvdata() is called before this */
> Why not just pass priv instead of dev?
>
>> +static int imx8mp_audiomix_reset_get_regmap(struct device *dev)
>> +{
>> +	struct imx8mp_audiomix_reset *priv;
>> +	int ret;
>> +
>> +	priv = dev_get_drvdata(dev);
>> +
>> +	/* try to use the parent's regmap */
>> +	priv->regmap = dev_get_regmap(dev->parent, NULL);
>> +	if (priv->regmap)
>> +		return 0;
>> +
>> +	/* ... if that's not possible then initialize the regmap right now */
>> +	priv->base = of_iomap(dev->parent->of_node, 0);
> Make base a local variable ...
>
>> +	if (!priv->base)
>> +		return dev_err_probe(dev, -ENOMEM, "failed to iomap address space\n");
>> +
>> +	ret = devm_add_action_or_reset(dev, imx8mp_audiomix_reset_iounmap, dev);
> ... and pass it as data instead of dev.
>
>> +	if (ret)
>> +		return dev_err_probe(dev, ret, "failed to register action\n");
>> +
>> +	priv->regmap = devm_regmap_init_mmio(dev, priv->base, &regmap_config);
>> +	if (IS_ERR(priv->regmap))
>> +		return dev_err_probe(dev, PTR_ERR(priv->regmap),
>> +				     "failed to initialize regmap\n");
>> +
>> +	return 0;
>> +}
>> +
>>  static int imx8mp_audiomix_reset_probe(struct auxiliary_device *adev,
>>  				       const struct auxiliary_device_id *id)
>>  {
>> @@ -105,36 +139,26 @@ static int imx8mp_audiomix_reset_probe(struct auxiliary_device *adev,
>>  	if (!priv)
>>  		return -ENOMEM;
>>  
>> -	spin_lock_init(&priv->lock);
>> -
>>  	priv->rcdev.owner     = THIS_MODULE;
>>  	priv->rcdev.nr_resets = ARRAY_SIZE(reset_map);
>>  	priv->rcdev.ops       = &imx8mp_audiomix_reset_ops;
>>  	priv->rcdev.of_node   = dev->parent->of_node;
>>  	priv->rcdev.dev	      = dev;
>>  	priv->rcdev.of_reset_n_cells = 1;
>> -	priv->base            = of_iomap(dev->parent->of_node, 0);
>> -	if (!priv->base)
>> -		return -ENOMEM;
>>  
>> +	/* keep before call to imx8mp_audiomix_reset_init_regmap() */
> Not needed if priv is passed to it directly.


right, somehow I forgot that rcdev holds a pointer to our "struct device" even though it's right there....

will make the change, thanks!


>
> regards
> Philipp