Hi Bartosz,
On Mo, 2024-02-19 at 16:33 +0100, Bartosz Golaszewski wrote:
> On Thu, Feb 15, 2024 at 11:03 AM Philipp Zabel <p.zabel@xxxxxxxxxxxxxx> wrote:
> >
> > On Mi, 2024-02-14 at 17:23 +0100, Théo Lebrun wrote:
> > [...]
> > > diff --git a/drivers/gpio/gpio-nomadik.c b/drivers/gpio/gpio-nomadik.c
> > > new file mode 100644
> > > index 000000000000..e39477e1a58f
> > > --- /dev/null
> > > +++ b/drivers/gpio/gpio-nomadik.c
> > > @@ -0,0 +1,660 @@
> > [...]
> > > +static int nmk_gpio_probe(struct platform_device *dev)
> > > +{
> > [...]
> > > + ret = gpiochip_add_data(chip, nmk_chip);
> >
> > Use devm_gpiochip_add_data() to cleanup on unbind, before nmk_chip goes
> > away. Or make the driver un-unbindable via suppress_bind_attrs. In that
> > case you could drop devm_ prefixes everywhere for consistency.
> >
>
> No! Why? What about error paths in probe() where you want to undo everything?
Brain fog moment. I was triggered by the mixture of devm_ and non-devm_
calls and jumped to the wrong conclusion.
Yes, keeping devm_ for error cleanup is of course correct, and with
suppress_bind_attrs it'd even be ok to use non-devm_
gpiochip_add_data(), as long as there can be no error return
afterwards.
regards
Philipp
