Re: [RFT PATCH 14/21] hte: tegra194: don't access struct gpio_chip

[Bartosz Golaszewski]

On Thu, Oct 5, 2023 at 9:43 PM Dipen Patel <dipenp@xxxxxxxxxx> wrote:
>
> On 10/5/23 12:05 PM, Bartosz Golaszewski wrote:
> > On Thu, Oct 5, 2023 at 8:12 PM Dipen Patel <dipenp@xxxxxxxxxx> wrote:
> >>
> >> On 10/5/23 6:48 AM, Bartosz Golaszewski wrote:
> >>> On Thu, Oct 5, 2023 at 1:52 AM Dipen Patel <dipenp@xxxxxxxxxx> wrote:
> >>>>
> >>>> On 10/4/23 3:54 PM, Dipen Patel wrote:
> >>>>> On 10/4/23 1:33 PM, Dipen Patel wrote:
> >>>>>> On 10/4/23 1:30 PM, Dipen Patel wrote:
> >>>>>>> On 10/4/23 5:00 AM, Bartosz Golaszewski wrote:
> >>>>>>>> On Thu, Sep 7, 2023 at 9:28 AM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
> >>>>>>>>>
> >>>>>>>>> On Tue, Sep 5, 2023 at 8:53 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> >>>>>>>>>
> >>>>>>>>>> From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> >>>>>>>>>>
> >>>>>>>>>> Using struct gpio_chip is not safe as it will disappear if the
> >>>>>>>>>> underlying driver is unbound for any reason. Switch to using reference
> >>>>>>>>>> counted struct gpio_device and its dedicated accessors.
> >>>>>>>>>>
> >>>>>>>>>> Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> >>>>>>>>>
> >>>>>>>>> As Andy points out add <linux/cleanup.h>, with that fixed:
> >>>>>>>>> Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> >>>>>>>>>
> >>>>>>>>> I think this can be merged into the gpio tree after leaving some
> >>>>>>>>> slack for the HTE maintainer to look at it, things look so much
> >>>>>>>>> better after this.
> >>>>>>>>>
> >>>>>>>>> Yours,
> >>>>>>>>> Linus Walleij
> >>>>>>>>
> >>>>>>>> Dipen,
> >>>>>>>>
> >>>>>>>> if you could give this patch a test and possibly ack it for me to take
> >>>>>>>> it through the GPIO tree (or go the immutable tag from HTE route) then
> >>>>>>>> it would be great. This is the last user of gpiochip_find() treewide,
> >>>>>>>> so with it we could remove it entirely for v6.7.
> >>>>>>>
> >>>>>>> Progress so far for the RFT...
> >>>>>>>
> >>>>>>> I tried applying the patch series on 6.6-rc1 and it did not apply cleanly,
> >>>>>>> some patches I needed to manually apply and correct. With all this, it failed
> >>>>>>> compilation at some spi/spi-bcm2835 driver. I disabled that and was able to
> >>>>>>> compile. I thought I should let you know this part.
> >>>>>>>
> >>>>>>> Now, I tried to test the hte and it seems to fail finding the gpio device,
> >>>>>>> roughly around this place [1]. I thought it would be your patch series so
> >>>>>>> tried to just use 6.6rc1 without your patches and it still failed at the
> >>>>>>> same place. I have to trace back now from which kernel version it broke.
> >>>>>>
> >>>>>> [1].
> >>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/pateldipen1984/linux.git/tree/drivers/hte/hte-tegra194.c?h=for-next#n781
> >>>>>>
> >>>>>> of course with your patches it would fail for the gdev instead of the chip.
> >>>>>
> >>>>> Small update:
> >>>>>
> >>>>> I put some debugging prints in the gpio match function in the hte-tegra194.c as
> >>>>> below:
> >>>>>
> >>>>> static int tegra_gpiochip_match(struct gpio_chip *chip, void *data)
> >>>>>  {
> >>>>> +       struct device_node *node = data;
> >>>>> +       struct fwnode_handle *fw = of_node_to_fwnode(data);
> >>>>> +       if (!fw || !chip->fwnode)
> >>>>> +               pr_err("dipen patel: fw is null\n");
> >>>>>
> >>>>> -       pr_err("%s:%d\n", __func__, __LINE__);
> >>>>> +       pr_err("dipen patel, %s:%d: %s, %s, %s, match?:%d, fwnode name:%s\n",
> >>>>> __func__, __LINE__, chip->label, node->name, node->full_name, (chip->fwnode ==
> >>>>> fw), fw->dev->init_name);
> >>>>>         return chip->fwnode == of_node_to_fwnode(data);
> >>>>>  }
> >>>>>
> >>>>> The output of the printfs looks like below:
> >>>>> [    3.955194] dipen patel: fw is null -----> this message started appearing
> >>>>> when I added !chip->fwnode test in the if condition line.
> >>>>>
> >>>>> [    3.958864] dipen patel, tegra_gpiochip_match:689: tegra234-gpio, gpio,
> >>>>> gpio@c2f0000, match?:0, fwnode name:(null)
> >>>>>
> >>>>> I conclude that chip->fwnode is empty. Any idea in which conditions that node
> >>>>> would be empty?
> >>>>
> >>>> sorry for spamming, one last message before I sign off for the day....
> >>>>
> >>>> Seems, adding below in the tegra gpio driver resolved the issue I am facing, I
> >>>> was able to verify your patch series.
> >>>>
> >>>> diff --git a/drivers/gpio/gpio-tegra186.c b/drivers/gpio/gpio-tegra186.c
> >>>> index d87dd06db40d..a56c159d7136 100644
> >>>> --- a/drivers/gpio/gpio-tegra186.c
> >>>> +++ b/drivers/gpio/gpio-tegra186.c
> >>>> @@ -989,6 +989,8 @@ static int tegra186_gpio_probe(struct platform_device *pdev)
> >>>>                 offset += port->pins;
> >>>>         }
> >>>>
> >>>> +       gpio->gpio.fwnode = of_node_to_fwnode(pdev->dev.of_node);
> >>>> +
> >>>>         return devm_gpiochip_add_data(&pdev->dev, &gpio->gpio, gpio);
> >>>>  }
> >>>>
> >>>> Now, few follow up questions:
> >>>> 1) is this the correct way of setting the chip fwnode in the gpio driver?
> >>>
> >>> You shouldn't need this. This driver already does:
> >>>
> >>>     gpio->gpio.parent = &pdev->dev;
> >>>
> >>> so fwnode should be assigned in gpiochip_add_data_with_key(). Can you
> >>> check why this doesn't happen?
> >>
> >> I do not see anywhere chip->fwnode being set in the gpiochip_add_* function.
> >> The only reference I see is here [1]. Does it mean I need to change my match
> >> function from:
> >>
> >> chip->fwnode == of_node_to_fwnode(data)
> >>
> >> to:
> >> dev_fwnode(chip->parent) == of_node_to_fwnode(data)?
> >
> > No! chip->fwnode is only used to let GPIOLIB know which fwnode to
> > assign to the GPIO device (struct gpio_device).
> What do you suggest I should use for the match as I do not see chip->fwnode
> being set?
>

This is most likely going to be a longer discussion. I suggest that in
the meantime you just assign the gc->fwnode pointer explicitly from
the platform device in the tegra GPIO driver and use it in the lookup
function. Note that this is NOT wrong or a hack. It's just that most
devices don't need to be looked up using gpio_device_find().

Bart

> >
> > Bart
> >
> >>
> >> [1]:
> >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpio/gpiolib.c?h=v6.6-rc1#n767
> >>
> >>>
> >>> Bart
> >>>
> >>>> 2) Or should I use something else in hte matching function instead of fwnode so
> >>>> to avoid adding above line in the gpio driver?
> >>>>
> >>>>>
> >>>>>>>
> >>>>>>>>
> >>>>>>>> Bart
> >>>>>>>
> >>>>>>
> >>>>>
> >>>>
> >>
>