[PATCH v7 0/5] pinctrl: mediatek: Enable module build support

[Linus Walleij]

On Thu, Jul 23, 2026 at 5:58 AM Justin Yeh <justin.yeh at mediatek.com> wrote:

> This series lets the MediaTek pinctrl drivers be built as loadable
> kernel modules. This is required for Android GKI + vendor_dlkm
> deployments, where vendor-specific drivers must be kept separate from
> the GKI vmlinux and loaded as modules from the vendor partition.
>
> Enabling the individual SoC drivers as modules is a single logical step,
> but it has a few prerequisites, so the series is:
>
>   1. Fix a pre-existing GPIO chip lifecycle bug. The gpio_chip lives in
>      device-managed memory but is registered with the non-managed
>      gpiochip_add_data(). While the drivers were built-in only this was
>      harmless, but once they can be unbound/rmmod'd, devm frees the
>      backing memory while the chip is still registered (use-after-free).
>      Switch to devm_gpiochip_add_data() in the shared probe/init paths.
>
>   2. Fix a pre-existing EINT resource leak on unbind. mtk_eint_do_init()
>      creates an IRQ domain, an interrupt mapping per EINT line and a
>      chained handler on the parent interrupt, none of which were ever
>      released. Once the drivers can be unbound/rmmod'd this leaves a
>      dangling IRQ domain, mappings whose chip data points at freed
>      memory, and a chained handler still firing into that freed data.
>      Tear them down with a device-managed action, matching the devm
>      lifecycle used for the gpio_chip.
>
>   3. Make the shared common code modular. The SoC drivers link against
>      pinctrl-mtk-common.c (v1), pinctrl-moore.c and pinctrl-mtmips.c,
>      whose Kconfig symbols were bool and which exported nothing. Without
>      this, selecting a SoC driver as =m forces the common symbol to =y
>      and the module fails to link against the unexported entry points.
>      Convert PINCTRL_MTK, PINCTRL_MTK_MOORE and PINCTRL_MTK_MTMIPS to
>      tristate, export the entry points, and add MODULE_LICENSE()/
>      MODULE_DESCRIPTION(). The already-modular v2 common code
>      additionally needs mtk_rmw() exported, since it is called directly
>      by SoC drivers such as mt7623. The shared symbols are exported in
>      the "MTK_PINCTRL" symbol namespace (EXPORT_SYMBOL_NS_GPL()), and
>      each SoC driver that uses them declares MODULE_IMPORT_NS().
>
>   4. Convert the MT7986 driver to a single initcall. It registers two
>      platform drivers (mt7986a/mt7986b) and used to call arch_initcall()
>      twice. A module has only one module_init(), so two arch_initcall()s
>      break the module build with a redefinition of
>      init_module()/__inittest(). Fold both drivers into one
>      platform_register_drivers() call from a single initcall. No
>      functional change when built-in.
>
>   5. Flip every SoC driver's Kconfig from bool to tristate and add the
>      module metadata. This also adds the MODULE_DEVICE_TABLE(of, ...)
>      entries that several drivers were missing, so the resulting modules
>      carry the modalias information udev needs to autoload them on a
>      matching device.
>
> The series builds cleanly under allmodconfig on arm64 (all SoC drivers
> as =m).
>
> Changes in v7:
>  - patch 2/5 (free EINT resources on unbind): add synchronize_irq() in
>    the teardown, after detaching the chained handler and before disposing
>    of the domain/mappings, so any in-flight handler has finished before
>    the data it references is freed, as suggested by AngeloGioacchino Del
>    Regno. Also add the missing #include <linux/interrupt.h> that
>    synchronize_irq() needs. Added his Reviewed-by.
>  - patch 5/5 (enable module build support): add MODULE_DEVICE_TABLE(of,
>    ...) to the SoC drivers that were missing it. AngeloGioacchino and
>    Chen-Yu Tsai both pointed this out; without it the module still
>    builds but is not autoloaded by udev on a matching device.
>
> Review status: patches 1/5 and 4/5 carry Chen-Yu Tsai's Reviewed-by;
> patch 2/5 carries AngeloGioacchino Del Regno's Reviewed-by; patch 5/5
> carries AngeloGioacchino Del Regno's Reviewed-by (from the per-driver
> patches reviewed in v4, preserved across the squash). Patch 3/5 has not
> been reviewed yet.

All patches applied, I had to rebase patch 5/5, please check the
result!

Yours,
Linus Walleij