In a blog post (Linux LED Subsystem), Fabio Baltieri details using the leds-gpio driver to quickly add GPIO-connected LEDs to a SoC. His method references nslu2-setup.c as an example implementation which utilizes static structures and the platform_add_devices function to add the LED devices. This is fine for an unchanging platform device where the driver will be compiled into the kernel - using static structures makes perfect sense but it does have a major drawback for developers.

If you find yourself developing a board and you want to load and unload a module rather than constantly rebooting while you work on writing parts of the driver you will have a big problem. leds-gpio does not get the opportunity to release (during platform_device_unregister) these statically initialized devices. This means they will be left out in the cold and you cannot reload the module successfully. It makes sense: the underling platform_device_register that platform_add_devices calls is meant for statically initialized devices. Since this memory cannot be freed there is no sense in the release hook being assigned - doing so will only lead to problems.

Rather than statically initializing the devices, they can be created with a combination of platform_device_alloc, platform_device_add_data, and platform_device_add. Within the call to platform_device_alloc, an unregister hook to free the memory is added. This means a later call to platform_device_unregister will actually remove the devices. Thus you can do work in a module that can be cleanly unloaded and loaded again.