Message-ID: <wa6ajhkrceg3h24c7jmniws2lkkrb3zhhyqpeo3y7d74hbh25x@brt3etvm7dh3>
Date: Sat, 3 May 2025 08:30:04 +0300
From: Dmitry Baryshkov <dmitry.baryshkov@....qualcomm.com>
To: Matti Vaittinen <mazziesaccount@...il.com>
Cc: Doug Anderson <dianders@...omium.org>,
        Matti Vaittinen <matti.vaittinen@...rohmeurope.com>,
        Linus Walleij <linus.walleij@...aro.org>,
        Bartosz Golaszewski <brgl@...ev.pl>, linux-gpio@...r.kernel.org,
        linux-kernel@...r.kernel.org, Biju Das <biju.das.jz@...renesas.com>,
        Geert Uytterhoeven <geert@...ux-m68k.org>,
        linux-arm-msm <linux-arm-msm@...r.kernel.org>,
        Krzysztof Kozlowski <krzk+dt@...nel.org>,
        Bjorn Andersson <andersson@...nel.org>
Subject: Re: [PATCH v3 4/4] gpio: Hide valid_mask from direct assignments

On Sat, May 03, 2025 at 12:41:30AM +0300, Dmitry Baryshkov wrote:
> On Sun, Apr 13, 2025 at 11:51:29AM +0300, Matti Vaittinen wrote:
> > On 13/04/2025 11:08, Matti Vaittinen wrote:
> > > Hi Doug,
> > > 
> > > On 13/04/2025 02:00, Doug Anderson wrote:
> > > > Hi,
> > > > 
> > > > On Wed, Mar 5, 2025 at 5:23 AM Matti Vaittinen
> > > > <mazziesaccount@...il.com> 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@...aro.org>
> > > > > Signed-off-by: Matti Vaittinen <mazziesaccount@...il.com>
> > > > > Reviewed-by: Linus Walleij <linus.walleij@...aro.org>
> > > > > ---
> > > > > 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@...ry=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@...ry=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 :] )
> 
> I think a correct fix for the pinctrl-msm driver would to use
> devm_pinctrl_register_and_init() and then pinctrl_enable() after
> registering GPIO chip, I'm going to submit a relevant patch. However I
> can't stop but notice that pinctrl-msm is not unique in the pattern of
> simply calling [devm_]pinctrl_register() and then registering a GPIO
> chip. This patch makes this pattern much more fragile.

Unfortunately... my idea to split pinctrl_register() into _and_init()
and pinctrl_enable() calls didn't work. Internally GPIOchip reigstration
calls of_pinctrl_get(), gets NULL and fails gpiochip registration with
-EPROBE_DEFER.

> 
> > 
> > +Björn
> > 
> > > 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@...ry=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
> > > 
> > 
> 
> -- 
> With best wishes
> Dmitry

-- 
With best wishes
Dmitry

Powered by blists - more mailing lists