Re: [PATCH v8 1/2] mfd: bd71837: mfd driver for ROHM BD71837 PMIC

[Matti Vaittinen <matti.vaittinen@fi.rohmeurope.com>]

On Wed, Jul 04, 2018 at 11:11:02AM +0100, Lee Jones wrote:
> On Wed, 04 Jul 2018, Matti Vaittinen wrote:
> 
> > On Wed, Jul 04, 2018 at 11:39:11AM +0300, Matti Vaittinen wrote:
> > > On Tue, Jul 03, 2018 at 11:26:00AM +0200, Enric Balletbo Serra wrote:
> > > > Missatge de Matti Vaittinen <matti.vaittinen@fi.rohmeurope.com> del
> > > > dia dv., 29 de juny 2018 a les 11:47:
> > > > 
> > > > Now that you use devm calls and you don't need to unwind things I
> > > > think is better to use plain returns. So,
> > > > 
> > > > return -ENOMEM;
> > > 
> > > I have never really understood why use of gotos in error handling is
> > > discouraged.
> 
> They're not.
> 
> > > Personally I would always choose single point of exit from
> > > a function when it is simple enough to achieve (like in this case). I've
> > > written and fixed way too many functions which leak resources or
> > > accidentally keep a lock when exiting from error branches. But I know
> > > many colleagues like you who prefer not to have gotos but  in place returns
> > > instead. So I guess I'll leave the final call on this to the one who is
> > > maintainer for this code. And it is true there is no things to unwind
> > > now - which does not mean that next updater won't add such. But as I
> > > said, I know plenty of people share your view - and even though I rather
> > > maintain code with only one exit the final call is on subsystem maintainer
> > > here.
> 
> Please use gotos in the error path.
> 
> IMO, it's the nicest way to unwind (as you call it).

I'll keep the gotos but clean other stuff for patch v9 then.
> 
> > Actually, If it was completely my call the probe would look something
> > like this:
> > 
> > +static int bd71837_i2c_probe(struct i2c_client *i2c,
> > +                           const struct i2c_device_id *id)
> > +{
> > +       struct bd71837 *bd71837;
> > +       struct bd71837_board *board_info;
> > +       int gpio_intr = 0;
> > +
> > +       const char *errstr = "No IRQ configured";
> > +       int ret = -EINVAL;
> > +
> > +       bd71837 = devm_kzalloc(&i2c->dev, sizeof(struct bd71837), GFP_KERNEL);
> > +
> > +       if (bd71837 == NULL)
> > +               return -ENOMEM;
> > +
> > +       board_info = dev_get_platdata(&i2c->dev);
> > +
> > +       if (!board_info)
> > +               gpio_intr = i2c->irq;
> > +       else
> > +               gpio_intr = board_info->gpio_intr;
> > +
> > +       if (!gpio_intr)
> > +               goto err_out;
> > +
> > +       i2c_set_clientdata(i2c, bd71837);
> > +       bd71837->dev = &i2c->dev;
> > +       bd71837->i2c_client = i2c;
> > +       bd71837->chip_irq = gpio_intr;
> > +
> > +       errstr = "regmap initialization failed";
> > +
> > +       bd71837->regmap = devm_regmap_init_i2c(i2c, &bd71837_regmap_config);
> > +       ret = PTR_ERR(bd71837->regmap);
> > +       if (IS_ERR(bd71837->regmap))
> > +               goto err_out;
> > +
> > +       errstr = "Read BD71837_REG_DEVICE failed";
> > +       ret = bd71837_reg_read(bd71837, BD71837_REG_REV);
> > +       if (ret < 0)
> > +               goto err_out;
> > +
> > +       errstr = "Failed to add irq_chip";
> > +       ret = devm_regmap_add_irq_chip(&i2c->dev, bd71837->regmap,
> > +                                      bd71837->chip_irq, IRQF_ONESHOT, 0,
> > +                                      &bd71837_irq_chip, &bd71837->irq_data);
> > +       if (ret < 0)
> > +               goto err_out;
> > +
> > +       errstr = "Failed to configure button short press timeout";
> > +       ret = regmap_update_bits(bd71837->regmap,
> > +                                BD71837_REG_PWRONCONFIG0,
> > +                                BD718XX_PWRBTN_PRESS_DURATION_MASK,
> > +                                BD718XX_PWRBTN_SHORT_PRESS_10MS);
> > +       if (ret < 0)
> > +               goto err_out;
> > +
> > +       /* According to BD71847 datasheet the HW default for long press
> > +        * detection is 10ms. So lets change it to 10 sec so we can actually
> > +        * get the short push and allow gracefull shut down
> > +        */
> > +       ret = regmap_update_bits(bd71837->regmap,
> > +                                BD71837_REG_PWRONCONFIG1,
> > +                                BD718XX_PWRBTN_PRESS_DURATION_MASK,
> > +                                BD718XX_PWRBTN_LONG_PRESS_10S);
> > +
> > +       errstr = "Failed to configure button long press timeout";
> > +       if (ret < 0)
> > +               goto err_out;
> > +
> > +       btns[0].irq = regmap_irq_get_virq(bd71837->irq_data,
> > +                                         BD71837_INT_PWRBTN_S);
> > +
> > +       errstr = "Failed to get the IRQ";
> > +       ret = btns[0].irq;
> > +       if (btns[0].irq < 0)
> > +               goto err_out;
> > +
> > +       errstr = "Failed to create subdevices";
> > +       ret = devm_mfd_add_devices(bd71837->dev, PLATFORM_DEVID_AUTO,
> > +                                  bd71837_mfd_cells,
> > +                                  ARRAY_SIZE(bd71837_mfd_cells), NULL, 0,
> > +                                  regmap_irq_get_domain(bd71837->irq_data));
> > +       if (ret) {
> > +err_out:
> > +               if (errstr)
> > +                       dev_err(&i2c->dev, "%s (%d)\n", errstr, ret);
> > +       }
> > +
> > +       return ret;
> > +}
> > 
> > What do you think of this? To my eye it is nice. It keeps single point of
> > exit and introduces only simple if-statements without the need of curly
> > brackets. And finally the error prints string works as a comment too.
> > I've seen bunch of constructs like this on the networking side but I
> > have no idea if this is frowned on this subsystem =) Oh, and probe abowe
> > is just to illustrate the idea, I did not even try compiling it yet.
> 
> That is horrible.  I nearly vomited on my keyboard. 

Note to self: Never to buy second hand keyboard from Lee =)

> It doesn't flow
> anywhere nearly as nicely has sorting out all of the error handling
> *after* it has been detected.  You're sacrificing readability to save
> a single line and do not save any *actual* lines of code, only a brace.

I was expecting something like this comment =) But the truth is that one
gets used to reading this quickly. Well, this still sounds like I should
not try convincing you - so you can stay heretic ;)
> 
> Landing a goto in the middle of a statement is messy and unsightly.

This is another thing one gets used to. I've actually seen plenty of
code using

	if (0) {
error_label:
	....
	}

for error handling. But again - you can keep your view and I'll adopt to
it here =)

> What happens when you have some resources to free?  The last few lines
> will become very messy, very quickly.

One can just build the usual clean-up sequence inside the last if (ret)
using different goto lables. Eg:

	if (ret) {
err_unwind_X:
		undo_x();
err_unwind_Y:
		undo_y();
err_out:
		dev_err(...);
	}
> 
> Nit: "something == NULL" is better written as "!something".

Oh, I personally liked the !foo more as Enric - but I will write the
NULL in case it won't make line too long. This is not a big deal to me.

> Nit: Please use proper multi-line comments as per the Coding Style.

Will do.

Thanks for quick reply! I will send new version today or tomorrow.

Best Regards
    Matti Vaittinen