Re: [PATCH/RFC 2/2] gpio: of: Add DT overlay support for GPIO hogs

[Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>]

Hi Frank,

On Tue, Jan 7, 2020 at 12:34 AM Frank Rowand <frowand.list@xxxxxxxxx> wrote:
> On 12/30/19 7:38 AM, Geert Uytterhoeven wrote:
> > As GPIO hogs are configured at GPIO controller initialization time,
> > adding/removing GPIO hogs in DT overlays does not work.
> >
> > Add support for GPIO hogs described in DT overlays by registering an OF
> > reconfiguration notifier, to handle the addition and removal of GPIO hog
> > subnodes to/from a GPIO controller device node.
> >
> > Note that when a GPIO hog device node is being removed, its "gpios"
> > properties is no longer available, so we have to keep track of which
> > node a hog belongs to, which is done by adding a pointer to the hog's
> > device node to struct gpio_desc.
>
> If I have read the patches and the existing overlay source correctly,
> then some observations:
>
> - A gpio hog node added in an overlay will be properly processed.
>
> - A gpio hog node already existing in the live devicetree, but with a
>   non-active status will be properly processed if the status of the
>   gpio hog node is changed to "ok" in the overlay.
>
> - If a gpio hog node already exists in the live devicetree with an
>   active status, then any updated or added properties in that gpio
>   hog node in the overlay will have no effect.
>
>   There is a scenario where the updated property would have an effect:
>   apply a second overlay that sets the status to inactive, then apply
>   a third overlay that sets the status back to active.  This is a
>   rather contrived example and I think it should be documented as
>   not supported and the result undefined.
>
> It would be good to document this explicitly.

I didn't verify this in detail, but I believe the existing overlay
support for platform, i2c, and SPI devices behaves the same.

> > --- a/drivers/gpio/gpiolib-of.c
> > +++ b/drivers/gpio/gpiolib-of.c

> > +static int of_gpio_notify(struct notifier_block *nb, unsigned long action,
> > +                       void *arg)
> > +{
> > +     struct of_reconfig_data *rd = arg;
> > +     struct gpio_chip *chip;
> > +     int ret;
> > +
> > +     switch (of_reconfig_get_state_change(action, arg)) {
> > +     case OF_RECONFIG_CHANGE_ADD:
> > +             if (!of_property_read_bool(rd->dn, "gpio-hog"))
> > +                     return NOTIFY_OK;       /* not for us */
> > +
> > +             if (of_node_test_and_set_flag(rd->dn, OF_POPULATED))
> > +                     return NOTIFY_OK;
>
> I don't think OF_POPULATED could be already set.  It seems to be a
> bug if it is.

For a real gpio-hog it indeed is not.  But this function is called for
every change made to the device tree (add a printk() and look at the
output during boot).  So this serves as a (cheap) line of defense.
The of_find_gpiochip_by_node() call below is more expensive to call.

> > +
> > +             chip = of_find_gpiochip_by_node(rd->dn->parent);
> > +             if (chip == NULL)
> > +                     return NOTIFY_OK;       /* not for us */
>
> If I understand correctly, "not for us" is a misleading comment.
> The notification is for the node rd->dn->parent, but the device
> does not exist, so we can't do the hogging at the moment.  (If the
> device is created later, then the gpio hog child node will exist,
> and the init will "do the right thing" with the hog node -- so
> not a problem.)

This function is called for all additions to the device tree.
So rd->dn->parent may not even be a gpio controller node.
Hence unless this is a gpio controller node for this hog, this
notification is "not for us".

> > +
> > +             ret = of_gpiochip_add_hog(chip, rd->dn);
> > +             if (ret < 0) {
> > +                     pr_err("%s: failed to add hogs for %pOF\n", __func__,
> > +                            rd->dn);
> > +                     of_node_clear_flag(rd->dn, OF_POPULATED);
> > +                     return notifier_from_errno(ret);
> > +             }
> > +             break;
> > +
> > +     case OF_RECONFIG_CHANGE_REMOVE:
> > +             if (!of_node_check_flag(rd->dn, OF_POPULATED))
> > +                     return NOTIFY_OK;       /* already depopulated */
>
> I don't think OF_POPULATED could be already cleared.  It seems to be a
> bug if it is.

Same here. First line of defense.

> > +
> > +             chip = of_find_gpiochip_by_node(rd->dn->parent);
> > +             if (chip == NULL)
> > +                     return NOTIFY_OK;       /* not for us */
>
> Again, a misleading comment.

Same here. rd->dn->parent may be something else.

> > +
> > +             of_gpiochip_remove_hog(chip, rd->dn);
> > +             of_node_clear_flag(rd->dn, OF_POPULATED);
> > +             break;
> > +     }
> > +
> > +     return NOTIFY_OK;
> > +}

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds