Re: [PATCH 8/8] gpio: sim: new testing module

[Bartosz Golaszewski]

On Mon, Feb 1, 2021 at 1:49 PM Andy Shevchenko
<andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
>
> On Mon, Feb 01, 2021 at 11:59:31AM +0100, Bartosz Golaszewski wrote:
> > On Mon, Feb 1, 2021 at 11:28 AM Andy Shevchenko
> > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > On Sat, Jan 30, 2021 at 09:37:55PM +0100, Bartosz Golaszewski wrote:
> > > > On Fri, Jan 29, 2021 at 4:57 PM Andy Shevchenko
> > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > > > On Fri, Jan 29, 2021 at 02:46:24PM +0100, Bartosz Golaszewski wrote:
>
> ...
>
> > > > > > +static int gpio_sim_set_config(struct gpio_chip *gc,
> > > > > > +                               unsigned int offset, unsigned long config)
> > > > > > +{
> > > > > > +     struct gpio_sim_chip *chip = gpiochip_get_data(gc);
> > > > > > +
> > > > > > +     switch (pinconf_to_config_param(config)) {
> > > > >
> > > > > > +     case PIN_CONFIG_BIAS_PULL_UP:
> > > > > > +             return gpio_sim_apply_pull(chip, offset, 1);
> > > > > > +     case PIN_CONFIG_BIAS_PULL_DOWN:
> > > > > > +             return gpio_sim_apply_pull(chip, offset, 0);
> > > > >
> > > > > But aren't we got a parameter (1 or 0) from config? And hence
> > > > >
> > > > >         case PIN_CONFIG_BIAS_PULL_UP:
> > > > >         case PIN_CONFIG_BIAS_PULL_DOWN:
> > > > >                 return gpio_sim_apply_pull(chip, offset, <param>);
> > > > >
> > > > > ?
> > > >
> > > > I believe this is more explicit and so easier to read if you don't
> > > > know the GPIO and pinctrl internals.
> > >
> > > If we ever go to change meanings of the values in param, it will require to fix
> > > this occurrence which seems to me suboptimal.
> > >
> >
> > Why would we do it? This is internal to this driver.
> >
> > > > > > +     default:
> > > > > > +             break;
> > > > > > +     }
> > > > > > +
> > > > > > +     return -ENOTSUPP;
> > > > > > +}
>
> Up to you.
> My personal vote for using the embedded param, because it makes code consistent
> and if anybody takes this driver as an example for something, it will be better
> example in such case..
>
> ...
>
> > > > > > +static ssize_t gpio_sim_sysfs_line_store(struct device *dev,
> > > > > > +                                      struct device_attribute *attr,
> > > > > > +                                      const char *buf, size_t len)
> > > > > > +{
> > > > > > +     struct gpio_sim_attribute *line_attr = to_gpio_sim_attr(attr);
> > > > > > +     struct gpio_sim_chip *chip = dev_get_drvdata(dev);
> > > > > > +     int ret, val;
> > > > >
> > > > > > +     ret = kstrtoint(buf, 0, &val);
> > > > > > +     if (ret)
> > > > > > +             return ret;
> > > > > > +     if (val != 0 && val != 1)
> > > > > > +             return -EINVAL;
> > > > >
> > > > > kstrtobool() ?
> > > > >
> > > >
> > > > No, we really only want 0 or 1, no yes, Y etc.
> > >
> > > Side note: But you allow 0x00001, for example...
> >
> > Good point. In that case we should check if len > 2 and if buf[0] ==
> > '1' or '0' and that's all we allow.
>
> Up to you also. I don't like such a strictness. OTOH we can relax afterwards if
> needed.
>
> > > Then why not to use unsigned type from the first place and add a comment?
> > >
> > > > > > +     ret = gpio_sim_apply_pull(chip, line_attr->offset, val);
> > > > > > +     if (ret)
> > > > > > +             return ret;
> > > > > > +
> > > > > > +     return len;
> > > > > > +}
>
> ...
>
> > > > > > +struct gpio_sim_chip_config {
> > > > > > +     struct config_item item;
> > > > > > +
> > > > > > +     /*
> > > > > > +      * If pdev is NULL, the item is 'pending' (waiting for configuration).
> > > > > > +      * Once the pointer is assigned, the device has been created and the
> > > > > > +      * item is 'live'.
> > > > > > +      */
> > > > > > +     struct platform_device *pdev;
> > > > >
> > > > > Are you sure
> > > > >
> > > > >         struct device *dev;
> > > > >
> > > > > is not sufficient?
> > > > >
> > > >
> > > > It may be but I really prefer those simulated devices to be on the platform bus.
> > >
> > > My point here is that there is no need to keep specific bus devices type,
> > > because you may easily derive it from the struct device pointer. Basically if
> > > you are almost using struct device in your code (seems to me the case), you
> > > won't need to carry bus specific one and dereference it each time.
> >
> > But don't we need a bus to even register a device? I haven't checked
> > in a long time but IIRC it's mandatory.
> >
> > Let me give you a different argument - the platform device offers a
> > very simple API for registering devices with properties being
> > duplicated behind the scenes etc. It seems to me that registering a
> > bare struct device * would take more boiler-plate code for not much
> > gain.
>
> Yes, I'm not objecting the platform bus choice. I'm objecting the keeping of
> the pointer to the bus specific structure.
>
> There are helpers like to_platform_device() which make the bus specific
> pointers go away from the structures and easier code when you use exactly
> pointer to struct device rather than bus specific one.
>

Ok I get it. We almost never dereference it though. We do it in probe,
but there's no way around it. In sysfs callbacks we already get a
pointer to struct device. And when unregistering the platform device,
we need to pass it as struct platform_device anyway. I don't see any
gain from that and would prefer to keep it as is.

Bart