Re: [PATCH v2 5/6] gpiolib: consolidate GPIO lookups

[Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>]

On Thu, Nov 10, 2022 at 03:42:40PM +0200, Andy Shevchenko wrote:
> On Wed, Nov 09, 2022 at 11:00:29AM -0800, Dmitry Torokhov wrote:
> > On Wed, Nov 09, 2022 at 01:25:06PM +0200, Andy Shevchenko wrote:
> > > On Tue, Nov 08, 2022 at 04:26:50PM -0800, Dmitry Torokhov wrote:
> 
> ...
> 
> > > > +static struct gpio_desc *gpiod_find_by_fwnode(struct fwnode_handle *fwnode,
> > > > +					      struct device *consumer,
> > > > +					      const char *con_id,
> > > > +					      unsigned int idx,
> > > > +					      enum gpiod_flags *flags,
> > > > +					      unsigned long *lookupflags)
> > > >  {
> > > > -	unsigned long lflags = GPIO_LOOKUP_FLAGS_DEFAULT;
> > > 
> > > > -	struct gpio_desc *desc = ERR_PTR(-ENODEV);
> > > 
> > > Not sure why this is needed. Now I see that else branch has been changed,
> > > but looking closer to it, we can drop it completely, while leaving this
> > > line untouched, correct?
> > 
> > Yes. I believe removing an initializer and doing a series of if/else
> > if/else was discussed and [soft] agreed-on in the previous review cycle,
> > but I can change it back.
> > 
> > I think we still need to have it return -ENOENT and not -ENODEV/-EINVAL
> > so that we can fall back to GPIO lookup tables when dealing with an
> > unsupported node type.
> 
> Right, okay, let's go with whatever variant you find better.
> 
> ...
> 
> > > > +	if (!IS_ERR_OR_NULL(fwnode))
> > > 
> > > I think this is superfluous check.
> > > 
> > > Now in the form of this series, you have only a single dev_dbg() that tries to
> > > dereference it. Do we really need to have it there, since every branch has its
> > > own dev_dbg() anyway?
> > 
> > As I mentioned, I like to keep this check to show the reader that we
> > should only descend into gpiod_find_by_fwnode() if we have a valid
> > fwnode. It is less about code generation and more about the intent.
> 
> Yes, but if fwnode is not found, we have a next check for that.

No, the check you are talking about is for the GPIO not being located.
It does not have anything to do with fwnode validity. You are relying on
intimate knowledge of gpiod_find_by_fwnode() implementation and the fact
that in the current form it will withstand ERR_PTR-encoded or NULL
fwnode.

I want to have the source code so clear in its intent so that I can be
woken up in the middle of the night with a huge hangover and still be
able to tell how it is supposed to behave.

> I really don't
> think we lose anything by dropping the check and gaining the code generation as
> a side effect.

This is cold path, happening only on startup. I am not saying that we
want to make it slow unnecessarily, but a condition branch that might
even get optimized out is not something we should be concerned here.

Thanks.

-- 
Dmitry