Date:   Wed, 11 Mar 2020 11:38:00 -0700
From:   Doug Berger <opendmb@...il.com>
To:     Linus Walleij <linus.walleij@...aro.org>
Cc:     Gregory Fong <gregory.0xf0@...il.com>,
        Bartosz Golaszewski <bgolaszewski@...libre.com>,
        Florian Fainelli <f.fainelli@...il.com>,
        bcm-kernel-feedback-list <bcm-kernel-feedback-list@...adcom.com>,
        "open list:GPIO SUBSYSTEM" <linux-gpio@...r.kernel.org>,
        Linux ARM <linux-arm-kernel@...ts.infradead.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH V2] gpio: brcmstb: support gpio-line-names property

Thanks for taking the time to review this.

On 3/11/20 8:32 AM, Linus Walleij wrote:
> On Mon, Mar 9, 2020 at 8:02 PM Doug Berger <opendmb@...il.com> wrote:
> 
>> The default handling of the gpio-line-names property by the
>> gpiolib-of implementation does not work with the multiple
>> gpiochip banks per device structure used by the gpio-brcmstb
>> driver.
To expand on this description, the crux of the issue is that the
gpio-brcmstb hardware has some nicely banked registers and some
not-so-nicely-banked common registers.

This lead to the decision to implement the driver to manage multiple
banks as a single GPIO device with a single device-tree node rather than
separate devices for each bank each with its own device-tree node.

In addition, most implementations include a hardware block within an
"Always On" power island and a second block that can be powered down.
Each of these blocks is represented as a separate device with their own
device-tree node and are managed by this driver.

The gpio_chip abstraction in the gpiolib provides a lot of useful
functionality for managing the banks of GPIO for the gpio-brcmstb
driver, but unfortunately it breaks down in a couple of places because
of the common device tree node that is shared by each bank.

One area is the IRQ chip helpers which were tried but needed to be reverted.

Another is labeling, which this commit attempts to address. The
device-tree node for each device can optionally contain a single
gpio-line-names property with a list of names to be applied to the GPIO
managed by the driver.

>>
>> This commit adds driver level support for the device tree
>> property so that GPIO lines can be assigned friendly names.
>>
> 
> 
>> Signed-off-by: Doug Berger <opendmb@...il.com>
>> +static void brcmstb_gpio_set_names(struct device *dev,
>> +                                  struct brcmstb_gpio_bank *bank)
>> +{
>> +       struct device_node *np = dev->of_node;
>> +       const char **names;
>> +       int nstrings, base;
> 
> I don't understand why that thing is named "base".Since this function is applied to each bank, it is necessary to know
what the device relative index is for the first GPIO contained within
this bank. That is the purpose of this base variable. It is used to
index the device relative list of gpio labels.

GPIO0 of bank 0 would have a base of 0. GPIO0 of bank 1 would have a
base of MAX_GPIO_PER_BANK, and so on.

>> +       unsigned int i;
>> +
>> +       base = bank->id * MAX_GPIO_PER_BANK;
> 
> That would be ngpios or something.
> 
> But you alread have what you need in bank->gc.ngpio, right?
> 
> So why calculate it?
Almost. ngpios is the number of gpios in the bank which in this case is
always MAX_GPIO_PER_BANK.

bank->gc.base is almost the right value, but it is relative to the GPIO
subsystem which can include multiple devices rather than the specific
device that contains this bank.

bank->id is device relative so bank->id * MAX_GPIO_PER_BANK gives us the
desired device relative offset.

>> +       nstrings = of_property_count_strings(np, "gpio-line-names");
>> +       if (nstrings <= base)
>> +               /* Line names not present */
>> +               return;
>> +
>> +       names = devm_kcalloc(dev, MAX_GPIO_PER_BANK, sizeof(*names),
>> +                            GFP_KERNEL);
>> +       if (!names)
>> +               return;
>> +
>> +       /*
>> +        * Make sure to not index beyond the end of the number of descriptors
>> +        * of the GPIO device.
>> +        */
>> +       for (i = 0; i < bank->width; i++) {
>> +               const char *name;
>> +               int ret;
>> +
>> +               ret = of_property_read_string_index(np, "gpio-line-names",
>> +                                                   base + i, &name);
>> +               if (ret) {
>> +                       if (ret != -ENODATA)
>> +                               dev_err(dev, "unable to name line %d: %d\n",
>> +                                       base + i, ret);
>> +                       break;
>> +               }
>> +               if (*name)
>> +                       names[i] = name;
>> +       }
>> +
>> +       bank->gc.names = names;
>> +}
> 
> Why can't you just make the function
> devprop_gpiochip_set_names() public, (line in <linux/gpio/driver.h>)
> and convert your np to a fwnode and call that &bank->gc ?
This is basically the current functionality as provided by the call to
gpiochip_add_data() in probe that this commit attempts to correct.

Since the fwnode is the same for all banks of the same device each bank
repeats the first MAX_GPIO_PER_BANK label names in each bank.

This commit populates the gc.names member of each bank from the
device-tree node within the driver. This overrides the default behavior
since devprop_gpiochip_set_names() will only be called if names is NULL.

> 
> Yours,
> Linus Walleij
> 

I hope that explanation makes sense.

Thanks again,
    Doug

Powered by blists - more mailing lists