On Tue, May 16, 2023 at 10:03:45PM +0300, Andy Shevchenko wrote:
> On Mon, May 15, 2023 at 1:13 PM Charles Keepax
> <ckeepax@xxxxxxxxxxxxxxxxxxxxx> wrote:
> > On Fri, May 12, 2023 at 10:19:14PM +0300, andy.shevchenko@xxxxxxxxx wrote:
> > > Fri, May 12, 2023 at 01:28:36PM +0100, Charles Keepax kirjoitti:
> > > > + if (!of_property_read_bool(dev_of_node(cs42l43->dev), "gpio-ranges")) {
> > > > + ret = gpiochip_add_pin_range(&priv->gpio_chip, priv->gpio_chip.label,
> > > > + 0, 0, CS42L43_NUM_GPIOS);
> > > > + if (ret) {
> > > > + dev_err(priv->dev, "Failed to add GPIO pin range: %d\n", ret);
> > > > + goto err_pm;
> > > > + }
> > > > + }
> > >
> > > Besides the fact that we have a callback for this, why GPIO library can't
> > > handle this for you already?
> >
> > Apologies but I am not quite sure I follow you, in the device
> > tree case this will be handled by the GPIO library. But for ACPI
> > this information does not exist so has to be called manually, the
> > library does not necessarily know which values to call with,
> > although admittedly our case is trivial but not all are.
>
> Why can't the firmware provide this information? _DSD() is a part of
> ACPI v5.1 IIRC.
>
I am very very far from confident we can guarantee that will be
present in the ACPI. The ACPI is typically made for and by the
Windows side.
> Although it might require moving some code from gpiolib-of.c to
> gpiolib.c with replacing OF APIs with agnostic ones.
>
I really think if we want to start doing things that way on ACPI
platforms someone with a little more clout than us needs to start
doing it first. If Intel or someone was doing it that way it
might give us a little more levelage to push it as being the
"correct" way to do it.
I will switch to the callback, but really don't think we can rely
on this being in DSD yet.
>
> > > > +static int cs42l43_pin_remove(struct platform_device *pdev)
> > > > +{
> > > > + pm_runtime_disable(&pdev->dev);
> > >
> > > This is simply wrong order because it's a mix of non-devm_*() followed by
> > > devm_*() calls in the probe.
> > >
> >
> > I had missed there are now devm_pm_runtime calls, I will switch
> > to that. But I would like to understand the wrong order, remove
> > will be called before the devm bits are destroyed and it seems
> > reasonable to disable the pm_runtime before destroying the
> > pinctrl device. What exactly would run in the wrong order here?
>
> At the ->remove() stage after this call an IRQ can be fired (or on SMP
> systems any other APIs can be called), for example. So, would it be a
> problem to service it with PM disabled?
>
> But in any case the shuffling ordering like this is prone to subtle
> bugs. I prefer to have strict ordering if there is nothing preventing
> from doing that way.
Yeah happy enough to use devm_ here, just didn't know it existed
and wanted to better understand your concerns as I was having
difficulty seeing the issue.
Thanks,
Charles
