Re: soc_button_array: ignored -EPROBE_DEFER loses the buttons on 68% of boots (Surface Pro 11)

[Hans de Goede]

Hi,

On 30-Aug-26 16:14, Sergey Lebedev wrote:
> Hello,
> 
> On a Microsoft Surface Pro 11 for Business (Intel, Lunar Lake) the power and
> volume buttons fail to appear on most boots. The probe runs, gives up
> silently, and the device is lost until something rebinds it by hand.
> 
> Measured across the 40 boots in this machine's journal, using the monotonic
> offset of the first "input: gpio-keys" message:
> 
>   driver attached on its own   13 boots   at 1.28 - 1.78 s
>   did not attach               27 boots   68 %
> 
> The timing is binary. Either the inputs appear at about 1.3 s, or they never
> appear at all - no intermediate values in 40 boots.
> 
> Where it fails
> --------------
> 
> Nothing at all is logged on a failing boot, so I built with
> soc_button_array.dyndbg=+p on the kernel command line. On four consecutive
> failing boots the probe printed exactly one line:
> 
>   [    1.286567] soc_button_array MSHW0040:00: OEM Platform Revision 129
> 
> That eliminates most of the possibilities:
> 
>   - the probe does run
>   - it passes soc_device_check_MSHW0040() - oem_platform_rev is 129, not 0
>   - it passes gpiod_count() - the dev_dbg("no GPIO attached, ignoring...")
>     would have printed with dyndbg on, and did not
> 
> By elimination the buttons are lost inside soc_button_device_create(), at
> 
>         error = soc_button_lookup_gpio(&pdev->dev, info->acpi_index,
>                                        &gpio, &irq);
>         if (error || irq < 0) {
>                 /*
>                  * Skip GPIO if not present. Note we deliberately
>                  * ignore -EPROBE_DEFER errors here. ...
>                  */
>                 continue;
>         }
> 
> Every button is skipped, n_buttons stays 0, and the two returns that follow
> are both silent:
> 
>         if (n_buttons == 0) { error = -ENODEV; goto err_free_mem; }
>   and
>         if (!priv->children[0] && !priv->children[1]) return -ENODEV;
> 
> which is why nothing reaches the log.
> 
> The GPIO controller here is INTC105D, five gpiochips driven by
> pinctrl-intel-platform. Both that and soc_button_array are modules in this
> config:
> 
>   CONFIG_INPUT_SOC_BUTTON_ARRAY=m
>   CONFIG_PINCTRL_INTEL_PLATFORM=m
> 
> A manual bind always works: 25 unbind/bind cycles at runtime gave 25
> successes and 0 failures. The failure exists only in the early-boot window.
> 
> One hypothesis, tested and wrong
> ---------------------------------
> 
> Since both are modules, load order looked like the cause, so I tried
> 
>   softdep soc_button_array pre: pinctrl-intel-platform
> 
> verified applied (modprobe --show-depends inserts pinctrl first), with the
> rescue that normally binds it by hand disabled so nothing masked the result.
> Five boots: 2 successes, 3 failures - the same rate as before.
> 
> pinctrl_intel_platform is a dependency of fifteen other modules and is
> loaded long before either way. softdep orders modules; the race is against
> the asynchronous probe of the individual INTC105D:0N devices, which register
> their gpiochips at their own pace. Reporting this so nobody else spends the
> evening on it.
> 
> Why no patch
> ------------
> 
> The comment says -EPROBE_DEFER is discarded because Intel uses "virtual
> GPIOs" that never materialise, so deferring on them would hang the probe
> forever. A fix has to tell that case apart from a controller that simply has
> not probed yet, and I do not know which signal you would want to use for
> that. Guessing would risk reintroducing the problem the workaround exists
> for.
> > Happy to test whatever shape you would prefer - the machine is here and I
> can build and boot kernels on it.

Thank you for your detailed bug report. I've just submitted a patch which
I believe should fix this:

https://lore.kernel.org/platform-driver-x86/20260908214410.70422-1-johannes.goede@xxxxxxxxxxxxxxxx/

Please give this a try and confirm if it fixes things.

Regards,

Hans