Re: [PATCH] ath9k: fix build error with LEDS_CLASS=m

On Mon, 25 Jan 2021 at 14:09, Arnd Bergmann [off-list ref] wrote:

On Mon, Jan 25, 2021 at 12:40 PM Krzysztof Kozlowski [off-list ref] wrote:

quoted

On Mon, 25 Jan 2021 at 12:36, Arnd Bergmann [off-list ref] wrote:

quoted

From: Arnd Bergmann <arnd@arndb.de>

When CONFIG_ATH9K is built-in but LED support is in a loadable
module, both ath9k drivers fails to link:

x86_64-linux-ld: drivers/net/wireless/ath/ath9k/gpio.o: in function `ath_deinit_leds':
gpio.c:(.text+0x36): undefined reference to `led_classdev_unregister'
x86_64-linux-ld: drivers/net/wireless/ath/ath9k/gpio.o: in function `ath_init_leds':
gpio.c:(.text+0x179): undefined reference to `led_classdev_register_ext'

The problem is that the 'imply' keyword does not enforce any dependency
but is only a weak hint to Kconfig to enable another symbol from a
defconfig file.

Change imply to a 'depends on LEDS_CLASS' that prevents the incorrect
configuration but still allows building the driver without LED support.

The 'select MAC80211_LEDS' is now ensures that the LED support is
actually used if it is present, and the added Kconfig dependency
on MAC80211_LEDS ensures that it cannot be enabled manually when it
has no effect.

But we do not want to have this dependency (selecting MAC80211_LEDS).
I fixed this problem here:
https://lore.kernel.org/lkml/20201227143034.1134829-1-krzk@kernel.org/ (local)
Maybe let's take this approach?

Generally speaking, I don't like to have a device driver specific Kconfig
setting 'select' a subsystem', for two reasons:

- you suddenly get asked for tons of new LED specific options when
  enabling seemingly benign options

- Mixing 'depends on' and 'select' leads to bugs with circular
  dependencies that usually require turning some other 'select'
  into 'depends on'.

The problem with LEDS_CLASS in particular is that there is a mix of drivers
using one vs the other roughly 50:50.

Yes, you are right, I also don't like it. However it was like this
before my commit so I am not introducing a new issue. The point is
that in your choice the MAC80211_LEDS will be selected if LEDS_CLASS
is present, which is exactly what I was trying to fix/remove. My WiFi
dongle does not have a LED and it causes a periodic (every second)
event. However I still have LEDS_CLASS for other LEDS in the system.

Best regards,
Krzysztof

On Mon, 25 Jan 2021 at 14:09, Arnd Bergmann [off-list ref] wrote:

On Mon, Jan 25, 2021 at 12:40 PM Krzysztof Kozlowski [off-list ref] wrote:

quoted

On Mon, 25 Jan 2021 at 12:36, Arnd Bergmann [off-list ref] wrote:

quoted

From: Arnd Bergmann <arnd@arndb.de>

When CONFIG_ATH9K is built-in but LED support is in a loadable
module, both ath9k drivers fails to link:

x86_64-linux-ld: drivers/net/wireless/ath/ath9k/gpio.o: in function `ath_deinit_leds':
gpio.c:(.text+0x36): undefined reference to `led_classdev_unregister'
x86_64-linux-ld: drivers/net/wireless/ath/ath9k/gpio.o: in function `ath_init_leds':
gpio.c:(.text+0x179): undefined reference to `led_classdev_register_ext'

The problem is that the 'imply' keyword does not enforce any dependency
but is only a weak hint to Kconfig to enable another symbol from a
defconfig file.

Change imply to a 'depends on LEDS_CLASS' that prevents the incorrect
configuration but still allows building the driver without LED support.

The 'select MAC80211_LEDS' is now ensures that the LED support is
actually used if it is present, and the added Kconfig dependency
on MAC80211_LEDS ensures that it cannot be enabled manually when it
has no effect.

But we do not want to have this dependency (selecting MAC80211_LEDS).
I fixed this problem here:
https://lore.kernel.org/lkml/20201227143034.1134829-1-krzk@kernel.org/ (local)
Maybe let's take this approach?

Generally speaking, I don't like to have a device driver specific Kconfig
setting 'select' a subsystem', for two reasons:

- you suddenly get asked for tons of new LED specific options when
  enabling seemingly benign options

- Mixing 'depends on' and 'select' leads to bugs with circular
  dependencies that usually require turning some other 'select'
  into 'depends on'.

The problem with LEDS_CLASS in particular is that there is a mix of drivers
using one vs the other roughly 50:50.

      Arnd