Re: [RFC PATCH 3/5] gpio: gpiolib: Add chardev support for maintaining GPIO values on reset

From: Charles Keepax
Date: Thu Oct 26 2017 - 05:11:32 EST


On Thu, Oct 26, 2017 at 10:35:39AM +1030, Andrew Jeffery wrote:
> On Wed, 2017-10-25 at 09:14 +0100, Charles Keepax wrote:
> > On Fri, Oct 20, 2017 at 07:32:53PM +1030, Andrew Jeffery wrote:
> > > On Fri, 2017-10-20 at 09:27 +0200, Linus Walleij wrote:
> > > > I don't see it as helpful to give userspace control over whether the line
> > > > is persistent or not. It is more reasonable to assume persistance for
> > > > userspace use cases, don't you think? Whether the system goes to sleep
> > > > or the gpiochip resets should not make a door suddenly close or the
> > > > lights in the christmas tree go out, right? I think if the gpiochip supports
> > > > persistance of any kind, we should try to use it and not have userspace
> > > > provide flags for that.
> > >
> > > Right. I guess the counter argument to your examples is if the gpio is
> > > controlling any active process that we don't want to continue if we've
> > > lost the capacity to monitor some other inputs (some kind of dead-man's 
> > > switch). But maybe the argument is that should be implemented in the
> > > kernel anyway?
> > >
> >
> > To me it certainly feels like decisions like this should live in
> > the kernel, your talking about things that could cause very weird
> > hardware behaviour if set wrong, so it makes sense to me to have
> > that responsibility guarded in the kernel.
>
> I feel that taking this argument to its logical conclusion leads to
> never exporting any GPIOs to userspace and doing everything in the
> kernel. If userspace has exported the GPIO and is managing its state,
> then it can *already* cause very weird hardware behaviour if set wrong.
> The fact that userspace is controlling the GPIO state and not the
> kernel already says that the kernel doesn't know how to manage it, so
> why not expose the option for userspace to set the persistence, given
> that it should know what it's doing?

Admittedly yes, I guess it really comes down to use-cases. There
are fairly strong use-cases to control GPIOs from user-space
that justify the risks. The use-cases for being able to set
non-persistent GPIOs from user-space seem less clear to me, but
if they exist I certainly don't have any objection.

Thanks,
Charles