[v2,2/4] reset: Instantiate reset GPIO controller for shared reset-gpios

  Devices sharing a reset GPIO could use the reset framework for
coordinated handling of that shared GPIO line.  We have several cases of
such needs, at least for Devicetree-based platforms.

If Devicetree-based device requests a reset line, which is missing but
there is a reset-gpios property, instantiate a new "reset-gpio" platform
device which will handle such reset line.  This allows seamless handling
of such shared reset-gpios without need of changing Devicetree binding [1].

All newly registered "reset-gpio" platform devices will be stored on
their own list to avoid any duplicated devices.  The key to find each of
such platform device is the entire Devicetree GPIO specifier: phandle to
GPIO controller, GPIO number and GPIO flags.  If two devices have
conflicting "reset-gpios" property, e.g. with different ACTIVE_xxx
flags, this would spawn two separate "reset-gpio" devices, where the
second would fail probing on busy GPIO reques

Link: https://lore.kernel.org/all/YXi5CUCEi7YmNxXM@robh.at.kernel.org/ [1]
Cc: Bartosz Golaszewski <brgl@bgdev.pl>
Cc: Sean Anderson <sean.anderson@seco.com>
Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
---
 drivers/reset/core.c             | 176 ++++++++++++++++++++++++++++---
 include/linux/reset-controller.h |   4 +
 2 files changed, 167 insertions(+), 13 deletions(-)
  

On Fr, 2024-01-05 at 16:59 +0100, Krzysztof Kozlowski wrote:
> Devices sharing a reset GPIO could use the reset framework for
> coordinated handling of that shared GPIO line.  We have several cases of
> such needs, at least for Devicetree-based platforms.
> 
> If Devicetree-based device requests a reset line, which is missing but
                                                    ^^^^^^^^^^^^^^^^
Nitpick: the "resets" property is missing, not the reset line.

"If Devicetree-based device requests a reset line, but there only is a
 reset-gpios property instead of a "resets" property, ..." maybe?

> there is a reset-gpios property, instantiate a new "reset-gpio" platform
> device which will handle such reset line.  This allows seamless handling
> of such shared reset-gpios without need of changing Devicetree binding [1].
> 
> All newly registered "reset-gpio" platform devices will be stored on
> their own list to avoid any duplicated devices.

That's not strictly true. The reset_gpio_device_list only contains the
of_phandle_args for lookup.

> The key to find each of
> such platform device is the entire Devicetree GPIO specifier: phandle to
> GPIO controller, GPIO number and GPIO flags.  If two devices have
> conflicting "reset-gpios" property, e.g. with different ACTIVE_xxx
> flags, this would spawn two separate "reset-gpio" devices, where the
> second would fail probing on busy GPIO reques

request.

Is that true? The code below looks like overwrites of_phandle_args so
that only one reset-gpio device is spawned for each gpio node.

> Link: https://lore.kernel.org/all/YXi5CUCEi7YmNxXM@robh.at.kernel.org/ [1]
> Cc: Bartosz Golaszewski <brgl@bgdev.pl>
> Cc: Sean Anderson <sean.anderson@seco.com>
> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
> ---
>  drivers/reset/core.c             | 176 ++++++++++++++++++++++++++++---
>  include/linux/reset-controller.h |   4 +
>  2 files changed, 167 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/reset/core.c b/drivers/reset/core.c
> index 4d5a78d3c085..ec9b3ff419cf 100644
> --- a/drivers/reset/core.c
> +++ b/drivers/reset/core.c
> @@ -13,6 +13,7 @@
>  #include <linux/module.h>
>  #include <linux/of.h>
>  #include <linux/acpi.h>
> +#include <linux/platform_device.h>
>  #include <linux/reset.h>
>  #include <linux/reset-controller.h>
>  #include <linux/slab.h>
> @@ -23,6 +24,10 @@ static LIST_HEAD(reset_controller_list);
>  static DEFINE_MUTEX(reset_lookup_mutex);
>  static LIST_HEAD(reset_lookup_list);
>  
> +/* Protects reset_gpio_device_list */
> +static DEFINE_MUTEX(reset_gpio_device_mutex);
> +static LIST_HEAD(reset_gpio_device_list);

I would call this reset_gpio_lookup_list or
reset_gpio_phandle_args_list.

> +
>  /**
>   * struct reset_control - a reset control
>   * @rcdev: a pointer to the reset controller device
> @@ -63,6 +68,16 @@ struct reset_control_array {
>  	struct reset_control *rstc[] __counted_by(num_rstcs);
>  };
>  
> +/**
> + * struct reset_gpio_device - ad-hoc created reset-gpio device
> + * @of_args: phandle to the reset controller with all the args like GPIO number
> + * @list: list entry for the reset_lookup_list
> + */
> +struct reset_gpio_device {

Similarly, I would call this reset_gpio_lookup or
reset_gpio_phandle_args.

> +	struct of_phandle_args of_args;
> +	struct list_head list;
> +};
> +
>  static const char *rcdev_name(struct reset_controller_dev *rcdev)
>  {
>  	if (rcdev->dev)
> @@ -813,13 +828,119 @@ static void __reset_control_put_internal(struct reset_control *rstc)
>  	kref_put(&rstc->refcnt, __reset_control_release);
>  }
>  
> +static bool __reset_gpios_args_match(const struct of_phandle_args *a1,
> +				     const struct of_phandle_args *a2)
> +{
> +	unsigned int i;
> +
> +	if (!a2)
> +		return false;
> +
> +	if (a1->args_count != a2->args_count)
> +		return false;
> +
> +	for (i = 0; i < a1->args_count; i++)
> +		if (a1->args[i] != a2->args[i])
> +			break;

Just return false in the loop and simplify the following to return
true.

> +
> +	/* All args matched? */
> +	if (i == a1->args_count)
> +		return true;
> +
> +	return false;
> +}
> +
> +/*
> + * @node:	node of the device requesting reset
> + * @reset_args:	phandle to the reset controller with all the args like GPIO number
> + */
> +static int __reset_add_reset_gpio_device(struct device_node *node,
> +					 struct of_phandle_args *args)
> +{
> +	struct reset_gpio_device *rgpio_dev;
> +	struct platform_device *pdev;
> +	int ret;
> +
> +	lockdep_assert_not_held(&reset_list_mutex);
> +
> +	mutex_lock(&reset_gpio_device_mutex);
> +
> +	list_for_each_entry(rgpio_dev, &reset_gpio_device_list, list) {
> +		if (args->np == rgpio_dev->of_args.np) {
> +			if (__reset_gpios_args_match(args,
> +						     &rgpio_dev->of_args)) {
> +				ret = 0;
> +				goto out_unlock;
> +			}
> +		}
> +	}
> +
> +	/* Not freed in normal path, persisent subsyst data */
> +	rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);

Since this is persistent, instead of letting the reset-gpio driver call
of_parse_phandle_with_args() again, this could be passed in via
platform data. Is there a reason not to do that instead?

> +	if (!rgpio_dev) {
> +		ret = -ENOMEM;
> +		goto out_unlock;
> +	}
> +
> +	rgpio_dev->of_args = *args;
> +	pdev = platform_device_register_data(NULL, "reset-gpio",
> +					     PLATFORM_DEVID_AUTO, &node,
> +					     sizeof(node));
> +	ret = PTR_ERR_OR_ZERO(pdev);
> +	if (!ret)
> +		list_add(&rgpio_dev->list, &reset_gpio_device_list);
> +	else
> +		kfree(rgpio_dev);
> +
> +out_unlock:
> +	mutex_unlock(&reset_gpio_device_mutex);
> +
> +	return ret;
> +}
> +
> +static struct reset_controller_dev *__reset_find_rcdev(struct of_phandle_args *args,
> +						       bool gpio_fallback,
> +						       const void *cookie)

Unused cookie.

> +{
> +	struct reset_controller_dev *r, *rcdev;
> +
> +	lockdep_assert_held(&reset_list_mutex);
> +
> +	rcdev = NULL;
> +	list_for_each_entry(r, &reset_controller_list, list) {
> +		if (args->np == r->of_node) {
> +			if (gpio_fallback) {
> +				if (__reset_gpios_args_match(args, r->of_args)) {
> +					/*
> +					 * Fake args (take first reset) and
> +					 * args_count (to matcg reset-gpio

match

> +					 * of_reset_n_cells) because reset-gpio
> +					 * has only one reset and does not care
> +					 * about reset of GPIO specifier.
> +					 */
> +					args->args[0] = 0;
> +					args->args_count = 1;

I'd expect args to be an input-only argument, but here its contents are
overwritten after a match. Why?

This has an effect in __of_reset_control_get(), that I find hard to
follow. See below.

> +					rcdev = r;
> +					break;
> +				}
> +			} else {
> +				rcdev = r;
> +				break;
> +			}
> +		}
> +	}
> +
> +	return rcdev;
> +}
> +
>  struct reset_control *
>  __of_reset_control_get(struct device_node *node, const char *id, int index,
>  		       bool shared, bool optional, bool acquired)
>  {
> +	struct of_phandle_args args = {0};
> +	bool gpio_fallback = false;
>  	struct reset_control *rstc;
> -	struct reset_controller_dev *r, *rcdev;
> -	struct of_phandle_args args;
> +	struct reset_controller_dev *rcdev;
>  	int rstc_id;
>  	int ret;
>  
> @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
>  					 index, &args);
>  	if (ret == -EINVAL)
>  		return ERR_PTR(ret);
> -	if (ret)
> -		return optional ? NULL : ERR_PTR(ret);
> +	if (ret) {
> +		/*
> +		 * There can be only one reset-gpio for regular devices, so
> +		 * don't bother with GPIO index.
> +		 */

I don't understand this comment. The GPIO index should be checked as
part of __reset_gpios_args_match(), or should it not?

> +		ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
> +						 0, &args);
> +		if (ret)
> +			return optional ? NULL : ERR_PTR(ret);
>  
> -	mutex_lock(&reset_list_mutex);
> -	rcdev = NULL;
> -	list_for_each_entry(r, &reset_controller_list, list) {
> -		if (args.np == r->of_node) {
> -			rcdev = r;
> -			break;
> -		}
> +		gpio_fallback = true;

Is there a reason not just call __reset_add_reset_gpio_device() here?
With that, there should be no need to call __reset_find_rcdev() twice.

>  	}
>  
> +	mutex_lock(&reset_list_mutex);
> +	rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);

This gets called with args as parsed. If there is a match, this will 
overwrite args (in the gpio_fallback case) and return NULL.

> +
>  	if (!rcdev) {
> -		rstc = ERR_PTR(-EPROBE_DEFER);
> -		goto out;
> +		if (gpio_fallback) {
> +			/*
> +			 * Registering reset-gpio device might cause immediate
> +			 * bind, thus taking reset_list_mutex lock via
> +			 * reset_controller_register().
> +			 */
> +			mutex_unlock(&reset_list_mutex);
> +			ret = __reset_add_reset_gpio_device(node, &args);

So this will also be called with args as parsed.

> +			mutex_lock(&reset_list_mutex);
> +			if (ret) {
> +				rstc = ERR_PTR(ret);
> +				goto out;
> +			}
> +			/*
> +			 * Success: reset-gpio could probe immediately, so
> +			 * re-check the lookup.
> +			 */
> +			rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);

And this will again be called with args as parsed and overwrite args
again.

> +			if (!rcdev) {
> +				rstc = ERR_PTR(-EPROBE_DEFER);
> +				goto out;
> +			}
> +			/* Success, rcdev is valid thus do not bail out */
> +		} else {
> +			rstc = ERR_PTR(-EPROBE_DEFER);
> +			goto out;
> +		}
>  	}

So at this point args is overwritten in the gpio_fallback case. I would
find it much clearer to just overwrite args here and make the first
parameter to __reset_find_rcdev() const.


regards
Philipp
  

On 08/01/2024 13:17, Philipp Zabel wrote:
> On Fr, 2024-01-05 at 16:59 +0100, Krzysztof Kozlowski wrote:
>> Devices sharing a reset GPIO could use the reset framework for
>> coordinated handling of that shared GPIO line.  We have several cases of
>> such needs, at least for Devicetree-based platforms.
>>
>> If Devicetree-based device requests a reset line, which is missing but
>                                                     ^^^^^^^^^^^^^^^^
> Nitpick: the "resets" property is missing, not the reset line.
> 
> "If Devicetree-based device requests a reset line, but there only is a
>  reset-gpios property instead of a "resets" property, ..." maybe?

Ack

> 
>> there is a reset-gpios property, instantiate a new "reset-gpio" platform
>> device which will handle such reset line.  This allows seamless handling
>> of such shared reset-gpios without need of changing Devicetree binding [1].
>>
>> All newly registered "reset-gpio" platform devices will be stored on
>> their own list to avoid any duplicated devices.
> 
> That's not strictly true. The reset_gpio_device_list only contains the
> of_phandle_args for lookup.

Ack, I will re-phrase it.

> 
>> The key to find each of
>> such platform device is the entire Devicetree GPIO specifier: phandle to
>> GPIO controller, GPIO number and GPIO flags.  If two devices have
>> conflicting "reset-gpios" property, e.g. with different ACTIVE_xxx
>> flags, this would spawn two separate "reset-gpio" devices, where the
>> second would fail probing on busy GPIO reques
> 
> request.

Ack

> 
> Is that true? 

It should be true and my tests confirmed it.

The code below looks like overwrites of_phandle_args so
> that only one reset-gpio device is spawned for each gpio node.

The code will iterate over list of of_node and of_phandle_args and
compare them with __reset_gpios_args_match(). If all match: do not
create new platform device.

If they do not match, e.g. ACTIVE_LOW -> ACTIVE_HIGH, create new
platform device. This will be the second device for the same GPIO.
Probing of that device in reset-gpio driver will fail:

[   19.198775] reset-gpio reset-gpio.2.auto: error -EBUSY: Could not get
reset gpios

because GPIO is used by reset-gpio.1.auto already.

> 
>> Link: https://lore.kernel.org/all/YXi5CUCEi7YmNxXM@robh.at.kernel.org/ [1]
>> Cc: Bartosz Golaszewski <brgl@bgdev.pl>
>> Cc: Sean Anderson <sean.anderson@seco.com>
>> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
>> ---
>>  drivers/reset/core.c             | 176 ++++++++++++++++++++++++++++---
>>  include/linux/reset-controller.h |   4 +
>>  2 files changed, 167 insertions(+), 13 deletions(-)
>>
>> diff --git a/drivers/reset/core.c b/drivers/reset/core.c
>> index 4d5a78d3c085..ec9b3ff419cf 100644
>> --- a/drivers/reset/core.c
>> +++ b/drivers/reset/core.c
>> @@ -13,6 +13,7 @@
>>  #include <linux/module.h>
>>  #include <linux/of.h>
>>  #include <linux/acpi.h>
>> +#include <linux/platform_device.h>
>>  #include <linux/reset.h>
>>  #include <linux/reset-controller.h>
>>  #include <linux/slab.h>
>> @@ -23,6 +24,10 @@ static LIST_HEAD(reset_controller_list);
>>  static DEFINE_MUTEX(reset_lookup_mutex);
>>  static LIST_HEAD(reset_lookup_list);
>>  
>> +/* Protects reset_gpio_device_list */
>> +static DEFINE_MUTEX(reset_gpio_device_mutex);
>> +static LIST_HEAD(reset_gpio_device_list);
> 
> I would call this reset_gpio_lookup_list or
> reset_gpio_phandle_args_list.

Ack

> 
>> +
>>  /**
>>   * struct reset_control - a reset control
>>   * @rcdev: a pointer to the reset controller device
>> @@ -63,6 +68,16 @@ struct reset_control_array {
>>  	struct reset_control *rstc[] __counted_by(num_rstcs);
>>  };
>>  
>> +/**
>> + * struct reset_gpio_device - ad-hoc created reset-gpio device
>> + * @of_args: phandle to the reset controller with all the args like GPIO number
>> + * @list: list entry for the reset_lookup_list
>> + */
>> +struct reset_gpio_device {
> 
> Similarly, I would call this reset_gpio_lookup or
> reset_gpio_phandle_args.

Ack

> 
>> +	struct of_phandle_args of_args;
>> +	struct list_head list;
>> +};
>> +
>>  static const char *rcdev_name(struct reset_controller_dev *rcdev)
>>  {
>>  	if (rcdev->dev)
>> @@ -813,13 +828,119 @@ static void __reset_control_put_internal(struct reset_control *rstc)
>>  	kref_put(&rstc->refcnt, __reset_control_release);
>>  }
>>  
>> +static bool __reset_gpios_args_match(const struct of_phandle_args *a1,
>> +				     const struct of_phandle_args *a2)
>> +{
>> +	unsigned int i;
>> +
>> +	if (!a2)
>> +		return false;
>> +
>> +	if (a1->args_count != a2->args_count)
>> +		return false;
>> +
>> +	for (i = 0; i < a1->args_count; i++)
>> +		if (a1->args[i] != a2->args[i])
>> +			break;
> 
> Just return false in the loop and simplify the following to return
> true.

Ack
> 
>> +
>> +	/* All args matched? */
>> +	if (i == a1->args_count)
>> +		return true;
>> +
>> +	return false;
>> +}
>> +
>> +/*
>> + * @node:	node of the device requesting reset
>> + * @reset_args:	phandle to the reset controller with all the args like GPIO number
>> + */
>> +static int __reset_add_reset_gpio_device(struct device_node *node,
>> +					 struct of_phandle_args *args)
>> +{
>> +	struct reset_gpio_device *rgpio_dev;
>> +	struct platform_device *pdev;
>> +	int ret;
>> +
>> +	lockdep_assert_not_held(&reset_list_mutex);
>> +
>> +	mutex_lock(&reset_gpio_device_mutex);
>> +
>> +	list_for_each_entry(rgpio_dev, &reset_gpio_device_list, list) {
>> +		if (args->np == rgpio_dev->of_args.np) {
>> +			if (__reset_gpios_args_match(args,
>> +						     &rgpio_dev->of_args)) {
>> +				ret = 0;
>> +				goto out_unlock;
>> +			}
>> +		}
>> +	}
>> +
>> +	/* Not freed in normal path, persisent subsyst data */
>> +	rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);
> 
> Since this is persistent, instead of letting the reset-gpio driver call
> of_parse_phandle_with_args() again, this could be passed in via
> platform data. Is there a reason not to do that instead?

We can pass it as read only platform data, but we cannot pass the
ownership. This is associated with registered platform device, not with
bound one device->driver.

Imagine case:
1. modprobe reset-gpio,
2. Driver is bound to the device,
3. unbind  (echo > unbind)
4. rmmod
5. goto 1

> 
>> +	if (!rgpio_dev) {
>> +		ret = -ENOMEM;
>> +		goto out_unlock;
>> +	}
>> +
>> +	rgpio_dev->of_args = *args;
>> +	pdev = platform_device_register_data(NULL, "reset-gpio",
>> +					     PLATFORM_DEVID_AUTO, &node,
>> +					     sizeof(node));
>> +	ret = PTR_ERR_OR_ZERO(pdev);
>> +	if (!ret)
>> +		list_add(&rgpio_dev->list, &reset_gpio_device_list);
>> +	else
>> +		kfree(rgpio_dev);
>> +
>> +out_unlock:
>> +	mutex_unlock(&reset_gpio_device_mutex);
>> +
>> +	return ret;
>> +}
>> +
>> +static struct reset_controller_dev *__reset_find_rcdev(struct of_phandle_args *args,
>> +						       bool gpio_fallback,
>> +						       const void *cookie)
> 
> Unused cookie.

Ack
> 
>> +{
>> +	struct reset_controller_dev *r, *rcdev;
>> +
>> +	lockdep_assert_held(&reset_list_mutex);
>> +
>> +	rcdev = NULL;
>> +	list_for_each_entry(r, &reset_controller_list, list) {
>> +		if (args->np == r->of_node) {
>> +			if (gpio_fallback) {
>> +				if (__reset_gpios_args_match(args, r->of_args)) {
>> +					/*
>> +					 * Fake args (take first reset) and
>> +					 * args_count (to matcg reset-gpio
> 
> match

Ack
> 
>> +					 * of_reset_n_cells) because reset-gpio
>> +					 * has only one reset and does not care
>> +					 * about reset of GPIO specifier.
>> +					 */
>> +					args->args[0] = 0;
>> +					args->args_count = 1;
> 
> I'd expect args to be an input-only argument, but here its contents are
> overwritten after a match. Why?
> 
> This has an effect in __of_reset_control_get(), that I find hard to
> follow. See below.
> 
>> +					rcdev = r;
>> +					break;
>> +				}
>> +			} else {
>> +				rcdev = r;
>> +				break;
>> +			}
>> +		}
>> +	}
>> +
>> +	return rcdev;
>> +}
>> +
>>  struct reset_control *
>>  __of_reset_control_get(struct device_node *node, const char *id, int index,
>>  		       bool shared, bool optional, bool acquired)
>>  {
>> +	struct of_phandle_args args = {0};
>> +	bool gpio_fallback = false;
>>  	struct reset_control *rstc;
>> -	struct reset_controller_dev *r, *rcdev;
>> -	struct of_phandle_args args;
>> +	struct reset_controller_dev *rcdev;
>>  	int rstc_id;
>>  	int ret;
>>  
>> @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
>>  					 index, &args);
>>  	if (ret == -EINVAL)
>>  		return ERR_PTR(ret);
>> -	if (ret)
>> -		return optional ? NULL : ERR_PTR(ret);
>> +	if (ret) {
>> +		/*
>> +		 * There can be only one reset-gpio for regular devices, so
>> +		 * don't bother with GPIO index.
>> +		 */
> 
> I don't understand this comment. The GPIO index should be checked as
> part of __reset_gpios_args_match(), or should it not?

This and earlier comment are result of a bit hacky approach to the
problem: how to find reset controllers for that GPIO?

The point is that our reset gpio controller has only 1 reset, thus
of_reset_n_cells=1. However args_count from of_parse_handle is >0, which
later is compared in reset core:

https://elixir.bootlin.com/linux/latest/source/drivers/reset/core.c#L859

That part we need to match.

I could make the reset-gpio driver to have of_reset_n_cells=2, but what
would be the point? The rest of the cells are not really relevant,
because you cannot refer to this reset gpio controller with any other
arguments.

To remind: my solution spawns one reset-gpio controller for one GPIO.

> 
>> +		ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
>> +						 0, &args);
>> +		if (ret)
>> +			return optional ? NULL : ERR_PTR(ret);
>>  
>> -	mutex_lock(&reset_list_mutex);
>> -	rcdev = NULL;
>> -	list_for_each_entry(r, &reset_controller_list, list) {
>> -		if (args.np == r->of_node) {
>> -			rcdev = r;
>> -			break;
>> -		}
>> +		gpio_fallback = true;
> 
> Is there a reason not just call __reset_add_reset_gpio_device() here?
> With that, there should be no need to call __reset_find_rcdev() twice.

Hm, could be, although not sure if code would be simpler.

This entire function handles two cases:
1. Get normal reset controller ("resets" OF property),
2. If above fails, get reset-gpio controller ("reset-gpios" OF property)

Therefore the entire solution is following approach:
1. of_parse_phandle(resets)
1b. error? Then of_parse_phandle(reset-gpios)
2. Find reset-controller based on any of above phandles.
3. error? Check if we created reset-gpios platform device. If not:
create new reset-gpios platform device/
3b. Platform device could probe, so lookup again for reset controller or
defer probe.

What type of flow do you propose?

> 
>>  	}
>>  
>> +	mutex_lock(&reset_list_mutex);
>> +	rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> 
> This gets called with args as parsed. If there is a match, this will 
> overwrite args (in the gpio_fallback case) and return NULL.

Overwrite not complete. It will only overwrite args_count and return a
valid rcdev.
I do not see overwriting in case of returning NULL.

> 
>> +
>>  	if (!rcdev) {
>> -		rstc = ERR_PTR(-EPROBE_DEFER);
>> -		goto out;
>> +		if (gpio_fallback) {
>> +			/*
>> +			 * Registering reset-gpio device might cause immediate
>> +			 * bind, thus taking reset_list_mutex lock via
>> +			 * reset_controller_register().
>> +			 */
>> +			mutex_unlock(&reset_list_mutex);
>> +			ret = __reset_add_reset_gpio_device(node, &args);
> 
> So this will also be called with args as parsed.
> 
>> +			mutex_lock(&reset_list_mutex);
>> +			if (ret) {
>> +				rstc = ERR_PTR(ret);
>> +				goto out;
>> +			}
>> +			/*
>> +			 * Success: reset-gpio could probe immediately, so
>> +			 * re-check the lookup.
>> +			 */
>> +			rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> 
> And this will again be called with args as parsed and overwrite args
> again.>
>> +			if (!rcdev) {
>> +				rstc = ERR_PTR(-EPROBE_DEFER);
>> +				goto out;
>> +			}
>> +			/* Success, rcdev is valid thus do not bail out */
>> +		} else {
>> +			rstc = ERR_PTR(-EPROBE_DEFER);
>> +			goto out;
>> +		}
>>  	}
> 
> So at this point args is overwritten in the gpio_fallback case. I would
> find it much clearer to just overwrite args here and make the first
> parameter to __reset_find_rcdev() const.

I think I get your point. Overwriting happens after we store the
original of_args, but the code is indeed not intuitive. I think I can
move it further, as you suggested.


Best regards,
Krzysztof
  

On Di, 2024-01-09 at 11:59 +0100, Krzysztof Kozlowski wrote:
> On 08/01/2024 13:17, Philipp Zabel wrote:
> > On Fr, 2024-01-05 at 16:59 +0100, Krzysztof Kozlowski wrote:
> > Is that true? 
> 
> It should be true and my tests confirmed it.
> 
> The code below looks like overwrites of_phandle_args so
> > that only one reset-gpio device is spawned for each gpio node.
> 
> The code will iterate over list of of_node and of_phandle_args and
> compare them with __reset_gpios_args_match(). If all match: do not
> create new platform device.
> 
> If they do not match, e.g. ACTIVE_LOW -> ACTIVE_HIGH, create new
> platform device. This will be the second device for the same GPIO.
> Probing of that device in reset-gpio driver will fail:
>
> [   19.198775] reset-gpio reset-gpio.2.auto: error -EBUSY: Could not get
> reset gpios
> 
> because GPIO is used by reset-gpio.1.auto already.

Thank you for the clarification.
I only understood later in the mail and didn't update this properly.

> > > +	/* Not freed in normal path, persisent subsyst data */
> > > +	rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);
> > 
> > Since this is persistent, instead of letting the reset-gpio driver call
> > of_parse_phandle_with_args() again, this could be passed in via
> > platform data. Is there a reason not to do that instead?
> 
> We can pass it as read only platform data, but we cannot pass the
> ownership. This is associated with registered platform device, not with
> bound one device->driver.
> 
> Imagine case:
> 1. modprobe reset-gpio,
> 2. Driver is bound to the device,
> 3. unbind  (echo > unbind)
> 4. rmmod
> 5. goto 1

Keeping ownership on the list is fine, the reset-gpio driver makes its
own copy of of_phandle_args anyway. I was just wondering whether it
could make this copy from platform data instead of from the
of_parse_phandle_with_args() return value.

[...]
> > 
> > > @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
> > >  					 index, &args);
> > >  	if (ret == -EINVAL)
> > >  		return ERR_PTR(ret);
> > > -	if (ret)
> > > -		return optional ? NULL : ERR_PTR(ret);
> > > +	if (ret) {
> > > +		/*
> > > +		 * There can be only one reset-gpio for regular devices, so
> > > +		 * don't bother with GPIO index.
> > > +		 */
> > 
> > I don't understand this comment. The GPIO index should be checked as
> > part of __reset_gpios_args_match(), or should it not?
> 
> This and earlier comment are result of a bit hacky approach to the
> problem: how to find reset controllers for that GPIO?
> 
> The point is that our reset gpio controller has only 1 reset, thus
> of_reset_n_cells=1. However args_count from of_parse_handle is >0, which
> later is compared in reset core:
> 
> https://elixir.bootlin.com/linux/latest/source/drivers/reset/core.c#L859
> 
> That part we need to match.
> 
> I could make the reset-gpio driver to have of_reset_n_cells=2, but what
> would be the point? The rest of the cells are not really relevant,
> because you cannot refer to this reset gpio controller with any other
> arguments.
> 
> To remind: my solution spawns one reset-gpio controller for one GPIO.

Thank you. I think we could also just make that check

	if (WARN_ON(!rcdev->of_args && ...))

instead and skip the of_xlate call in that case (or implement of_xlate
in the reset-gpio driver to make this more explicit).

> > 
> > > +		ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
> > > +						 0, &args);
> > > +		if (ret)
> > > +			return optional ? NULL : ERR_PTR(ret);
> > >  
> > > -	mutex_lock(&reset_list_mutex);
> > > -	rcdev = NULL;
> > > -	list_for_each_entry(r, &reset_controller_list, list) {
> > > -		if (args.np == r->of_node) {
> > > -			rcdev = r;
> > > -			break;
> > > -		}
> > > +		gpio_fallback = true;
> > 
> > Is there a reason not just call __reset_add_reset_gpio_device() here?
> > With that, there should be no need to call __reset_find_rcdev() twice.
> 
> Hm, could be, although not sure if code would be simpler.
> 
> This entire function handles two cases:
> 1. Get normal reset controller ("resets" OF property),
> 2. If above fails, get reset-gpio controller ("reset-gpios" OF property)
> 
> Therefore the entire solution is following approach:
> 1. of_parse_phandle(resets)
> 1b. error? Then of_parse_phandle(reset-gpios)
> 2. Find reset-controller based on any of above phandles.
> 3. error? Check if we created reset-gpios platform device. If not:
> create new reset-gpios platform device/
> 3b. Platform device could probe, so lookup again for reset controller or
> defer probe.
> 
> What type of flow do you propose?

I propose to reorder after parsing the phandles: check/create the gpio
platform device right after parsing the gpio handle. Only then lock
reset_list_mutex look for the rcdev.

1. of_parse_phandle(resets)
1b. error? Then of_parse_phandle(reset-gpios)
2b. gpio? Then check if we created reset-gpios platform device. If not:
create new reset-gpios platform device/, defer if probe failed
3. Lock reset_list_mutex, find reset-controller based on any of above
phandles.

> 
> > 
> > >  	}
> > >  
> > > +	mutex_lock(&reset_list_mutex);
> > > +	rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> > 
> > This gets called with args as parsed. If there is a match, this will 
> > overwrite args (in the gpio_fallback case) and return NULL.
> 
> Overwrite not complete. It will only overwrite args_count and return a
> valid rcdev.
> I do not see overwriting in case of returning NULL.

Sorry, I meant to write

"This gets called with args as parsed. If there is a match, this will 
 overwrite args (in the gpio_fallback case) _or_ return NULL."

at least at the end, when I understood the following.

> > 
> > > +
> > >  	if (!rcdev) {

So in this non-NULL branch there was no overwriting.

> > > -		rstc = ERR_PTR(-EPROBE_DEFER);
> > > -		goto out;
> > > +		if (gpio_fallback) {
> > > +			/*
> > > +			 * Registering reset-gpio device might cause immediate
> > > +			 * bind, thus taking reset_list_mutex lock via
> > > +			 * reset_controller_register().
> > > +			 */
> > > +			mutex_unlock(&reset_list_mutex);
> > > +			ret = __reset_add_reset_gpio_device(node, &args);
> > 
> > So this will also be called with args as parsed.
> > 
> > > +			mutex_lock(&reset_list_mutex);
> > > +			if (ret) {
> > > +				rstc = ERR_PTR(ret);
> > > +				goto out;
> > > +			}
> > > +			/*
> > > +			 * Success: reset-gpio could probe immediately, so
> > > +			 * re-check the lookup.
> > > +			 */
> > > +			rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> > 
> > And this will again be called with args as parsed and overwrite args
> > again.>
> > > +			if (!rcdev) {
> > > +				rstc = ERR_PTR(-EPROBE_DEFER);
> > > +				goto out;
> > > +			}
> > > +			/* Success, rcdev is valid thus do not bail out */
> > > +		} else {
> > > +			rstc = ERR_PTR(-EPROBE_DEFER);
> > > +			goto out;
> > > +		}
> > >  	}
> > 
> > So at this point args is overwritten in the gpio_fallback case. I would
> > find it much clearer to just overwrite args here and make the first
> > parameter to __reset_find_rcdev() const.
> 
> I think I get your point. Overwriting happens after we store the
> original of_args, but the code is indeed not intuitive. I think I can
> move it further, as you suggested.

Now I think we can skip the overwriting altogether and just adapt the
following of_reset_n_cells check ad of_xlate call as mentioned above.

regards
Philipp
  

On 09/01/2024 12:58, Philipp Zabel wrote:
>>>> +	/* Not freed in normal path, persisent subsyst data */
>>>> +	rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);
>>>
>>> Since this is persistent, instead of letting the reset-gpio driver call
>>> of_parse_phandle_with_args() again, this could be passed in via
>>> platform data. Is there a reason not to do that instead?
>>
>> We can pass it as read only platform data, but we cannot pass the
>> ownership. This is associated with registered platform device, not with
>> bound one device->driver.
>>
>> Imagine case:
>> 1. modprobe reset-gpio,
>> 2. Driver is bound to the device,
>> 3. unbind  (echo > unbind)
>> 4. rmmod
>> 5. goto 1
> 
> Keeping ownership on the list is fine, the reset-gpio driver makes its
> own copy of of_phandle_args anyway. I was just wondering whether it
> could make this copy from platform data instead of from the
> of_parse_phandle_with_args() return value.

Looks like it could. This could save us few lines of code in
reset-gpio.c. I'll try it.

> 
> [...]
>>>
>>>> @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
>>>>  					 index, &args);
>>>>  	if (ret == -EINVAL)
>>>>  		return ERR_PTR(ret);
>>>> -	if (ret)
>>>> -		return optional ? NULL : ERR_PTR(ret);
>>>> +	if (ret) {
>>>> +		/*
>>>> +		 * There can be only one reset-gpio for regular devices, so
>>>> +		 * don't bother with GPIO index.
>>>> +		 */
>>>
>>> I don't understand this comment. The GPIO index should be checked as
>>> part of __reset_gpios_args_match(), or should it not?
>>
>> This and earlier comment are result of a bit hacky approach to the
>> problem: how to find reset controllers for that GPIO?
>>
>> The point is that our reset gpio controller has only 1 reset, thus
>> of_reset_n_cells=1. However args_count from of_parse_handle is >0, which
>> later is compared in reset core:
>>
>> https://elixir.bootlin.com/linux/latest/source/drivers/reset/core.c#L859
>>
>> That part we need to match.
>>
>> I could make the reset-gpio driver to have of_reset_n_cells=2, but what
>> would be the point? The rest of the cells are not really relevant,
>> because you cannot refer to this reset gpio controller with any other
>> arguments.
>>
>> To remind: my solution spawns one reset-gpio controller for one GPIO.
> 
> Thank you. I think we could also just make that check
> 
> 	if (WARN_ON(!rcdev->of_args && ...))
> 
> instead and skip the of_xlate call in that case (or implement of_xlate
> in the reset-gpio driver to make this more explicit).

Ack

> 
>>>
>>>> +		ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
>>>> +						 0, &args);
>>>> +		if (ret)
>>>> +			return optional ? NULL : ERR_PTR(ret);
>>>>  
>>>> -	mutex_lock(&reset_list_mutex);
>>>> -	rcdev = NULL;
>>>> -	list_for_each_entry(r, &reset_controller_list, list) {
>>>> -		if (args.np == r->of_node) {
>>>> -			rcdev = r;
>>>> -			break;
>>>> -		}
>>>> +		gpio_fallback = true;
>>>
>>> Is there a reason not just call __reset_add_reset_gpio_device() here?
>>> With that, there should be no need to call __reset_find_rcdev() twice.
>>
>> Hm, could be, although not sure if code would be simpler.
>>
>> This entire function handles two cases:
>> 1. Get normal reset controller ("resets" OF property),
>> 2. If above fails, get reset-gpio controller ("reset-gpios" OF property)
>>
>> Therefore the entire solution is following approach:
>> 1. of_parse_phandle(resets)
>> 1b. error? Then of_parse_phandle(reset-gpios)
>> 2. Find reset-controller based on any of above phandles.
>> 3. error? Check if we created reset-gpios platform device. If not:
>> create new reset-gpios platform device/
>> 3b. Platform device could probe, so lookup again for reset controller or
>> defer probe.
>>
>> What type of flow do you propose?
> 
> I propose to reorder after parsing the phandles: check/create the gpio
> platform device right after parsing the gpio handle. Only then lock
> reset_list_mutex look for the rcdev.
> 
> 1. of_parse_phandle(resets)
> 1b. error? Then of_parse_phandle(reset-gpios)
> 2b. gpio? Then check if we created reset-gpios platform device. If not:
> create new reset-gpios platform device/, defer if probe failed
> 3. Lock reset_list_mutex, find reset-controller based on any of above
> phandles.

Could work, let me try. I have impression this was my first approach
which resulted in a bit more complicated code, but I don't remember the
details now.

> 
>>
>>>
>>>>  	}
>>>>  
>>>> +	mutex_lock(&reset_list_mutex);
>>>> +	rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
>>>
>>> This gets called with args as parsed. If there is a match, this will 
>>> overwrite args (in the gpio_fallback case) and return NULL.
>>
>> Overwrite not complete. It will only overwrite args_count and return a
>> valid rcdev.
>> I do not see overwriting in case of returning NULL.
> 
> Sorry, I meant to write
> 
> "This gets called with args as parsed. If there is a match, this will 
>  overwrite args (in the gpio_fallback case) _or_ return NULL."
> 
> at least at the end, when I understood the following.
> 
>>>
>>>> +
>>>>  	if (!rcdev) {
> 
> So in this non-NULL branch there was no overwriting.
> 
>>>> -		rstc = ERR_PTR(-EPROBE_DEFER);
>>>> -		goto out;
>>>> +		if (gpio_fallback) {
>>>> +			/*
>>>> +			 * Registering reset-gpio device might cause immediate
>>>> +			 * bind, thus taking reset_list_mutex lock via
>>>> +			 * reset_controller_register().
>>>> +			 */
>>>> +			mutex_unlock(&reset_list_mutex);
>>>> +			ret = __reset_add_reset_gpio_device(node, &args);
>>>
>>> So this will also be called with args as parsed.
>>>
>>>> +			mutex_lock(&reset_list_mutex);
>>>> +			if (ret) {
>>>> +				rstc = ERR_PTR(ret);
>>>> +				goto out;
>>>> +			}
>>>> +			/*
>>>> +			 * Success: reset-gpio could probe immediately, so
>>>> +			 * re-check the lookup.
>>>> +			 */
>>>> +			rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
>>>
>>> And this will again be called with args as parsed and overwrite args
>>> again.>
>>>> +			if (!rcdev) {
>>>> +				rstc = ERR_PTR(-EPROBE_DEFER);
>>>> +				goto out;
>>>> +			}
>>>> +			/* Success, rcdev is valid thus do not bail out */
>>>> +		} else {
>>>> +			rstc = ERR_PTR(-EPROBE_DEFER);
>>>> +			goto out;
>>>> +		}
>>>>  	}
>>>
>>> So at this point args is overwritten in the gpio_fallback case. I would
>>> find it much clearer to just overwrite args here and make the first
>>> parameter to __reset_find_rcdev() const.
>>
>> I think I get your point. Overwriting happens after we store the
>> original of_args, but the code is indeed not intuitive. I think I can
>> move it further, as you suggested.
> 
> Now I think we can skip the overwriting altogether and just adapt the
> following of_reset_n_cells check ad of_xlate call as mentioned above.


Yep!

Best regards,
Krzysztof