Re: [PATCH v5 3/5] gpiolib: use kref in gpio_desc

[Bartosz Golaszewski]

pt., 28 lut 2020 o 23:33 Linus Walleij <linus.walleij@xxxxxxxxxx> napisaÅ(a):
>
> On Fri, Feb 21, 2020 at 4:48 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
>
> > From: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
> >
> > GPIO descriptors are freed by consumers using gpiod_put(). The name of
> > this function suggests some reference counting is going on but it's not
> > true.
> >
> > Use kref to actually introduce reference counting for gpio_desc objects.
> > Add a corresponding gpiod_get() helper for increasing the reference count.
> >
> > This doesn't change anything for already existing (correct) drivers but
> > allows us to keep track of GPIO descs used by multiple users.
> >
> > Signed-off-by: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
>
> I'm having some trouble figuring out if we might be reinventing
> a wheel here.
>
> A while back there was a proposed patch to add device links
> between GPIO producers and consumers, so that a GPIO
> chip won't be dropped while there are active consumers.
>
> (I don't remember who sent the patch.)
>
> We have a similar functionality in pin control if the
> .link_consumers property is set on the pincontrol device.
> I was thinking about making that compulsory at one point.
>
> The device links use a kref already existing in struct
> device and would in this case be the kref in the struct
> device for the struct gpio_device.
>
> So if that existed, gpiod_ref could just grab another
> device_link_add().
>

I was always under the impression that device links are aimed mostly
at runtime PM.

> Maybe we should just add device links between all
> GPIO consumers (devices) and struct gpio_device:s
> struct device and implement it like this so we don't
> have to back out of this later?
>
> C.f. commit
> commit 036f394dd77f8117346874151793ec38967d843f
> pinctrl: Enable device link creation for pin control
>

Yes, definitely looks like it's done with PM in mind. Maybe we should
do what nvmem does? Define a struct device_type for GPIO chips with an
appropriate release() callback and use get_device() and put_device()?
Although nvmem seems to use kref for cells and device reference
counting somewhat separately - maybe that's something to address too.

> (...)
> > @@ -81,6 +81,7 @@ struct gpio_descs *__must_check gpiod_get_array(struct device *dev,
> >  struct gpio_descs *__must_check gpiod_get_array_optional(struct device *dev,
> >                                                         const char *con_id,
> >                                                         enum gpiod_flags flags);
> > +struct gpio_desc *gpiod_ref(struct gpio_desc *desc);
> >  void gpiod_put(struct gpio_desc *desc);
> >  void gpiod_put_array(struct gpio_descs *descs);
>
> You forgot to add a stub for the case where GPIOLIB is not
> compiled in I think? (Lower in the same file.)
>
> Yours,
> Linus Walleij

Yeah this is fixed in the next version (with a different subject since
it no longer concerns nvmem that much).

Bart