Re: [PATCH] gpio: create empty /sys/class/gpio with SYSFS disabled

[Bartosz Golaszewski <brgl@xxxxxxxx>]

On Mon, Oct 14, 2024 at 2:10 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
>
> From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
>
> User-space may want to use some kind of a compatibility layer for the
> deprecated GPIO sysfs ABI. This would typically involve mounting
> a fuse-based filesystem using the GPIO character device to emulate the
> sysfs behavior and layout.
>
> With GPIO_SYSFS disabled, the /sys/class/gpio directory doesn't exist
> and user-space cannot create it. In order to facilitate moving away from
> the sysfs, add a new Kconfig option that indicates to GPIOLIB that is
> should create an empty directory where the GPIO class interface would
> exist and enable it by default.
>
> Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> ---
> As I've mentioned under a different patch, I'm working on a user-space
> compatibility layer for the sysfs GPIO interface. FUSE allows us to
> emulate almost all its functionalities using libgpiod except for
> mounting it under /sys/class/gpio if GPIO_SYSFS is disabled. User-space
> cannot create directories in sysfs so if we want to allow users to mount
> the FUSE emulator under the old location, we need to create an empty
> /sys/class/gpio directory from the kernel and this is what is patch is
> for.
>
>  drivers/gpio/Kconfig   | 18 ++++++++++++++++++
>  drivers/gpio/gpiolib.c | 12 ++++++++++++
>  2 files changed, 30 insertions(+)
>
> diff --git a/drivers/gpio/Kconfig b/drivers/gpio/Kconfig
> index bfa6b5a2c537..f2e7163cb46c 100644
> --- a/drivers/gpio/Kconfig
> +++ b/drivers/gpio/Kconfig
> @@ -69,6 +69,24 @@ config GPIO_SYSFS
>           use the character device /dev/gpiochipN with the appropriate
>           ioctl() operations instead.
>
> +config GPIO_SYSFS_CLASS_DIR_STUB
> +       bool "Create empty /sys/class/gpio directory" if EXPERT
> +       depends on !GPIO_SYSFS
> +       default y
> +       help
> +         Say Y here to create an empty /sys/class/gpio directory.
> +
> +         User-space may want to use some kind of a compatibility layer for the
> +         deprecated GPIO sysfs ABI. This would typically involve mounting
> +         a fuse-based filesystem using the GPIO character device to emulate
> +         the sysfs behavior and layout.
> +
> +         This option makes GPIOLIB create an empty directory at /sys/class/gpio
> +         where user-space can mount the sysfs replacement and avoid having to
> +         change existing programs to adjust to different filesystem paths.
> +
> +         If unsure, say Y.
> +
>  config GPIO_CDEV
>         bool
>         prompt "Character device (/dev/gpiochipN) support" if EXPERT
> diff --git a/drivers/gpio/gpiolib.c b/drivers/gpio/gpiolib.c
> index 97346b746ef5..31efb580beb8 100644
> --- a/drivers/gpio/gpiolib.c
> +++ b/drivers/gpio/gpiolib.c
> @@ -4873,6 +4873,12 @@ static struct device_driver gpio_stub_drv = {
>         .probe = gpio_stub_drv_probe,
>  };
>
> +#if IS_ENABLED(CONFIG_GPIO_SYSFS_CLASS_DIR_STUB)
> +static const struct class gpio_class_stub = {
> +       .name = "gpio",
> +};
> +#endif /* CONFIG_GPIO_SYSFS_CLASS_DIR_STUB */
> +
>  static int __init gpiolib_dev_init(void)
>  {
>         int ret;
> @@ -4899,6 +4905,12 @@ static int __init gpiolib_dev_init(void)
>                 return ret;
>         }
>
> +#if IS_ENABLED(CONFIG_GPIO_SYSFS_CLASS_DIR_STUB)
> +       ret = class_register(&gpio_class_stub);

Gah I forgot to unregister the class. Also now that I think about it
we should maybe expose the class kobject in the kobject.h header and
use sysfs_create_mount_point() like what we do for debugfs, configfs
and other mountpoints at /sys/kernel?

In any case: please let me know what you think about this change in general.

Bart

> +       if (ret)
> +               pr_err("gpiolib: failed to create the empty gpio class directory\n");
> +#endif /* CONFIG_GPIO_SYSFS_CLASS_DIR_STUB */
> +
>         gpiolib_initialized = true;
>         gpiochip_setup_devs();
>
> --
> 2.43.0
>