Re: [PATCH v3 4/4] gpio: Hide valid_mask from direct assignments

[Matti Vaittinen]

Hi Doug,

On 13/04/2025 02:00, Doug Anderson wrote:
> Hi,
>
> On Wed, Mar 5, 2025 at 5:23 AM Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote:
> > The valid_mask member of the struct gpio_chip is unconditionally written
> > by the GPIO core at driver registration. Current documentation does not
> > mention this but just says the valid_mask is used if it's not NULL. This
> > lured me to try populating it directly in the GPIO driver probe instead
> > of using the init_valid_mask() callback. It took some retries with
> > different bitmaps and eventually a bit of code-reading to understand why
> > the valid_mask was not obeyed. I could've avoided this trial and error if
> > the valid_mask was hidden in the struct gpio_device instead of being a
> > visible member of the struct gpio_chip.
> >
> > Help the next developer who decides to directly populate the valid_mask
> > in struct gpio_chip by hiding the valid_mask in struct gpio_device and
> > keep it internal to the GPIO core.
> >
> > Suggested-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> > Signed-off-by: Matti Vaittinen <mazziesaccount@xxxxxxxxx>
> > Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> > ---
> > Revision history:
> > v2 => v3:
> >   - Rebase to gpio/for-next
> > v1 => v2:
> >   - Hide the valid_mask instead of documenting it as internal to GPIO
> >     core as suggested by Linus W.
> > https://lore.kernel.org/all/Z71qphikHPGB0Yuv@mva-rohm/
> > ---
> >   drivers/gpio/gpiolib.c      | 16 ++++++++--------
> >   drivers/gpio/gpiolib.h      |  3 +++
> >   include/linux/gpio/driver.h |  8 --------
> >   3 files changed, 11 insertions(+), 16 deletions(-)
>
> FWIW, I've found that this patch is crashing me at bootup on my
> sc7180-trogdor board. The problem is pretty obvious in gdb.
> "gc->gpiodev" is NULL in gpiochip_line_is_valid().

Thanks for debugging this! I find this odd. It seems to me the 
pinctrl-msm.c is calling the gpiochip_add_data() for the chip, in the 
msm_gpio_init() - which is called from the msm_pinctrl_probe().

The gpiochip_add_data() should go to the gpiochip_add_data_with_key() - 
where the gpiodev should be allocated and set.

I don't spot any successful code path where the gpiodev was not allocated.

> 0xffff80008066c760 in gpiochip_line_is_valid (gc=0xffff000083223890,
> offset=offset@entry=66) at drivers/gpio/gpiolib.c:746
> 746             if (likely(!gc->gpiodev->valid_mask))
> (gdb) bt
> #0  0xffff80008066c760 in gpiochip_line_is_valid
> (gc=0xffff000083223890, offset=offset@entry=66) at
> drivers/gpio/gpiolib.c:746
> #1  0xffff800080666338 in msm_pinmux_request (pctldev=<optimized out>,

Ah, but now I see the call comes from the pinmux. Looking at the 
msm_pinctrl_probe() - the pincontroller is registered before the gpio. 
Maybe, with unlucky timing, the request happens right after registering 
the pinctrl - but before registering the gpios.

This, I think, can be a bug even before this change (because the 
valid_mask is not initialized prior the gpio registration) - but this 
change now made it obvious.

I see the probe is actually an exported function, and there are mentions 
about ACPI support etc. I don't really know if there are valid cases 
where the pincontroller should be usable without the gpiochip. If this 
is the case, the unconditional call to the gpiochip_line_is_valid() from 
the msm_pinmux_request() smells wrong.

I am not sure about the right fix. One could try:

@@ -1568,6 +1568,10 @@ int msm_pinctrl_probe(struct platform_device *pdev,
        if (pctrl->irq < 0)
                return pctrl->irq;

+       ret = msm_gpio_init(pctrl);
+       if (ret)
+               return ret;
+
        pctrl->desc.owner = THIS_MODULE;
        pctrl->desc.pctlops = &msm_pinctrl_ops;
        pctrl->desc.pmxops = &msm_pinmux_ops;
@@ -1582,10 +1586,6 @@ int msm_pinctrl_probe(struct platform_device *pdev,
                return PTR_ERR(pctrl->pctrl);
        }

-       ret = msm_gpio_init(pctrl);
-       if (ret)
-               return ret;
-
        platform_set_drvdata(pdev, pctrl);

        dev_dbg(&pdev->dev, "Probed Qualcomm pinctrl driver\n")

but I am not at all this is the fix we're looking after. I wonder if 
Krzysztof has any suggestions? (Seeing he has been authoring some 
changes here :] )

Yours,
	-- Matti


> offset=66) at drivers/pinctrl/qcom/pinctrl-msm.c:152
> #2  0xffff800080662314 in pin_request (pctldev=0xffff000082686ac0,
> pin=66, owner=0xffff000082c02790 "3500000.pinctrl", gpio_range=0x0)
>      at drivers/pinctrl/pinmux.c:176
> #3  0xffff800080662900 in pinmux_enable_setting
> (setting=0xffff000082684b40) at drivers/pinctrl/pinmux.c:445
> #4  0xffff80008065fd54 in pinctrl_commit_state (p=0xffff000083a07520,
> state=0xffff000082684a40) at drivers/pinctrl/core.c:1300
> #5  0xffff8000806605bc in pinctrl_select_state (p=0xffff000083223890,
> p@entry=0xffff000082686ac0, state=0x42) at drivers/pinctrl/core.c:1381
> #6  pinctrl_claim_hogs (pctldev=0xffff000082686ac0) at
> drivers/pinctrl/core.c:2136
> #7  pinctrl_enable (pctldev=0xffff000082686ac0) at drivers/pinctrl/core.c:2156
> #8  0xffff800080660814 in pinctrl_register
> (pctldesc=0xffff000083223a90, dev=0xffff000081406410,
> driver_data=0xffff000083223880) at drivers/pinctrl/core.c:2193
> #9  0xffff800080660df4 in devm_pinctrl_register
> (dev=0xffff000081406410, pctldesc=0xffff000083223a90,
> driver_data=0xffff000083223880) at drivers/pinctrl/core.c:2313
> #10 0xffff8000806657b4 in msm_pinctrl_probe (pdev=0xffff000081406400,
> soc_data=<optimized out>) at drivers/pinctrl/qcom/pinctrl-msm.c:1579
> #11 0xffff80008066afcc in sc7180_pinctrl_probe
> (pdev=0xffff000083223890) at
> drivers/pinctrl/qcom/pinctrl-sc7180.c:1147
> #12 0xffff80008089583c in platform_probe (_dev=0xffff000081406410) at
> drivers/base/platform.c:1404
>
> (gdb) print gc->gpiodev
> $1 = (struct gpio_device *) 0x0
>
> -Doug