Re: [PATCH v3 3/9] gpio: 104-dio-48e: Migrate to the regmap-irq API

[Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>]

On Mon, Nov 28, 2022 at 11:56:15AM +0200, Andy Shevchenko wrote:
> On Mon, Nov 28, 2022 at 11:51:10AM +0200, Andy Shevchenko wrote:
> > On Sun, Nov 27, 2022 at 05:00:40PM -0500, William Breathitt Gray wrote:
> > > On Sun, Nov 27, 2022 at 07:31:48PM +0100, Michael Walle wrote:
> > > > Am 2022-11-22 11:29, schrieb William Breathitt Gray:

...

> > > > gpiochip doesn't seem to be the correct place, gpiochip_add_irqchip()
> > > > is a noop for gpio-regmap, right? So using gpiochip_irqchip_init_hw()
> > > > seems wrong.
> > > > 
> > > > Maybe make gpio-regmap call it on its own? But really we just connect
> > > > the regmap-irq to the gpiochip irqdomain.
> > > 
> > > I think you're right, it feels strange to handle IRQ initialization via
> > > the GPIO framework. Maybe somewhere in regmap_irq might be more
> > > appropriate?
> > 
> > The problem that that callback solves is possible interrupt storm, spurious
> > interrupts, and Use Before Initialized.
> > 
> > If you can guarantee that in your case it never happens, add a comment
> > and go on.
> > 
> > (It might be useful to tweak code a bit and try CONFIG_DEBUG_SHIRQ=y)
> > 
> > > > What is the purpose of the
> > > > .init_hw callback? I've looked at other drivers which use regmap-irq
> > > > and they all seem to just initialize the hardware in their _probe().
> > > > 
> > > > -michael
> > > 
> > > I'm not opposed to initializing the hardware in _probe(), although I can
> > > see merit in pushing that operation instead closer to the framework
> > > where the initialization is actually relevant.
> > > 
> > > Andy, maybe you can shed some light about .init_hw; I think you
> > > introduced it to gpiolib in commit 9411e3aaa6342.
> > 
> > It seems that commit message doesn't fully  explain the situation behind
> > that change. But it was observed in real life, see above.
> 
> FWIW, real life example:
> e986f0e602f1 ("pinctrl: intel: fix unexpected interrupt")

And another one (seems found from SW perspective):
a33912061607 ("gpio: lynxpoint: Move hardware initialization to callback")

-- 
With Best Regards,
Andy Shevchenko