Re: [PATCH v5 1/6] platform/x86: firmware_attributes_class: Add device initialization methods

["Kurt Borja" <kuurtb@gmail.com>]

Hi Thomas,

On Sun Jul 6, 2025 at 3:42 AM -03, Thomas Weißschuh wrote:
> On 2025-07-05 00:33:56-0300, Kurt Borja wrote:
>> From: Thomas Weißschuh <linux@weissschuh.net>
>> 
>> Currently each user of firmware_attributes_class has to manually set up
>> kobjects, devices, etc.
>> 
>> Provide this infrastructure out-of-the-box through the newly introduced
>> fwat_device_register().
>> 
>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>> Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
>> Co-developed-by: Kurt Borja <kuurtb@gmail.com>
>> Signed-off-by: Kurt Borja <kuurtb@gmail.com>
>> ---
>>  drivers/platform/x86/firmware_attributes_class.c | 125 +++++++++++++++++++++++
>>  drivers/platform/x86/firmware_attributes_class.h |  28 +++++
>>  2 files changed, 153 insertions(+)
>> 
>> diff --git a/drivers/platform/x86/firmware_attributes_class.c b/drivers/platform/x86/firmware_attributes_class.c
>> index 736e96c186d9dc6d945517f090e9af903e93bbf4..290364202cce64bb0e9046e0b2bbb8d85e2cbc6f 100644
>> --- a/drivers/platform/x86/firmware_attributes_class.c
>> +++ b/drivers/platform/x86/firmware_attributes_class.c
>> @@ -2,7 +2,14 @@
>>  
>>  /* Firmware attributes class helper module */
>>  
>> +#include <linux/cleanup.h>
>> +#include <linux/device.h>
>> +#include <linux/device/class.h>
>> +#include <linux/kdev_t.h>
>> +#include <linux/kobject.h>
>>  #include <linux/module.h>
>> +#include <linux/slab.h>
>> +#include <linux/types.h>
>>  #include "firmware_attributes_class.h"
>>  
>>  const struct class firmware_attributes_class = {
>> @@ -10,6 +17,122 @@ const struct class firmware_attributes_class = {
>>  };
>>  EXPORT_SYMBOL_GPL(firmware_attributes_class);
>>  
>> +static void fwat_device_release(struct device *dev)
>> +{
>> +	struct fwat_device *fadev = to_fwat_device(dev);
>> +
>> +	kfree(fadev);
>> +}
>> +
>> +/**
>> + * fwat_device_register - Create and register a firmware-attributes class
>> + *			  device
>> + * @parent: Parent device
>> + * @name: Name of the class device
>> + * @drvdata: Drvdata of the class device
>> + * @groups: Extra groups for the "attributes" directory
>> + *
>> + * Return: pointer to the new fwat_device on success, ERR_PTR on failure
>> + */
>> +struct fwat_device *
>> +fwat_device_register(struct device *parent, const char *name, void *drvdata,
>> +		     const struct attribute_group **groups)
>> +{
>> +	struct fwat_device *fadev;
>> +	int ret;
>> +
>> +	if (!parent || !name)
>> +		return ERR_PTR(-EINVAL);
>> +
>> +	fadev = kzalloc(sizeof(*fadev), GFP_KERNEL);
>> +	if (!fadev)
>> +		return ERR_PTR(-ENOMEM);
>> +
>> +	fadev->groups = groups;
>> +	fadev->dev.class = &firmware_attributes_class;
>> +	fadev->dev.parent = parent;
>> +	fadev->dev.release = fwat_device_release;
>> +	dev_set_drvdata(&fadev->dev, drvdata);
>> +	ret = dev_set_name(&fadev->dev, "%s", name);
>> +	if (ret) {
>> +		kfree(fadev);
>> +		return ERR_PTR(ret);
>> +	}
>> +	ret = device_register(&fadev->dev);
>> +	if (ret)
>> +		return ERR_PTR(ret);
>
> I think we need a put_device() here.

Indeed. I'll fix it.

>
>> +
>> +	fadev->attrs_kset = kset_create_and_add("attributes", NULL, &fadev->dev.kobj);
>> +	if (!fadev->attrs_kset) {
>> +		ret = -ENOMEM;
>> +		goto out_device_unregister;
>> +	}
>> +
>> +	ret = sysfs_create_groups(&fadev->attrs_kset->kobj, groups);
>> +	if (ret)
>> +		goto out_kset_unregister;
>
> It would be nicer for userspace to add the device to the hierarchy
> only when it is set up fully.
> Replacing device_register() with a device_initialize() above and
> device_add() down here.

Sure!

In the end however, children kobjects corresponding to "firmware
attributes" will still be added after device_add(), with
fwat_create_group().

Obviously, that only applies if we can all agree on the approach of
Patch 2. If you could take a look I would be very grateful!

>
>> +
>> +	return fadev;
>> +
>> +out_kset_unregister:
>> +	kset_unregister(fadev->attrs_kset);
>
> I *think* the driver core should clean up any child objects
> automatically, so this is unnecessary.

Hmm - I think filesystem is automatically cleaned up. But not putting
down the kset ref would leak it's allocated memory.

I'll verify this.

>
>> +
>> +out_device_unregister:
>> +	device_unregister(&fadev->dev);
>> +
>> +	return ERR_PTR(ret);
>> +}
>> +EXPORT_SYMBOL_GPL(fwat_device_register);
>> +
>> +void fwat_device_unregister(struct fwat_device *fadev)
>> +{
>> +	if (!fadev)
>> +		return;
>> +
>> +	sysfs_remove_groups(&fadev->attrs_kset->kobj, fadev->groups);
>> +	kset_unregister(fadev->attrs_kset);
>
> The also the two lines above would be unnecessary.

For the reason above I think kset_unregister() is still required.

Removing groups manually, on the other hand, is not required (I think).
I can remove that call.

>
>> +	device_unregister(&fadev->dev);
>> +}
>> +EXPORT_SYMBOL_GPL(fwat_device_unregister);
>> +
>> +static void devm_fwat_device_release(void *data)
>> +{
>> +	struct fwat_device *fadev = data;
>> +
>> +	fwat_device_unregister(fadev);
>> +}
>> +
>> +/**
>> + * devm_fwat_device_register - Create and register a firmware-attributes class
>> + *			       device
>> + * @parent: Parent device
>> + * @name: Name of the class device
>> + * @data: Drvdata of the class device
>> + * @groups: Extra groups for the class device (Optional)
>> + *
>> + * Device managed version of fwat_device_register().
>> + *
>> + * Return: pointer to the new fwat_device on success, ERR_PTR on failure
>> + */
>> +struct fwat_device *
>> +devm_fwat_device_register(struct device *parent, const char *name, void *data,
>> +			  const struct attribute_group **groups)
>> +{
>> +	struct fwat_device *fadev;
>> +	int ret;
>> +
>> +	fadev = fwat_device_register(parent, name, data, groups);
>> +	if (IS_ERR(fadev))
>> +		return fadev;
>> +
>> +	ret = devm_add_action_or_reset(parent, devm_fwat_device_release, fadev);
>> +	if (ret)
>> +		return ERR_PTR(ret);
>> +
>> +	return fadev;
>> +}
>> +EXPORT_SYMBOL_GPL(devm_fwat_device_register);
>
> ... and also all of the devm stuff.

Could you elaborate on this?

>
>> +
>>  static __init int fw_attributes_class_init(void)
>>  {
>>  	return class_register(&firmware_attributes_class);
>> @@ -23,5 +146,7 @@ static __exit void fw_attributes_class_exit(void)
>>  module_exit(fw_attributes_class_exit);
>>  
>>  MODULE_AUTHOR("Mark Pearson <markpearson@lenovo.com>");
>> +MODULE_AUTHOR("Thomas Weißschuh <linux@weissschuh.net>");
>> +MODULE_AUTHOR("Kurt Borja <kuurtb@gmail.com>");
>>  MODULE_DESCRIPTION("Firmware attributes class helper module");
>>  MODULE_LICENSE("GPL");
>
> <snip>

Thanks for reviewing this!


-- 
 ~ Kurt