Re: [PATCH 1/6] gpiolib: Allow GPIO chips to request their own GPIOs

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Wednesday, March 05, 2014 02:05:50 PM Mika Westerberg wrote:
> On Wed, Mar 05, 2014 at 10:49:41AM +0800, Linus Walleij wrote:
> > On Tue, Feb 25, 2014 at 12:00 AM, Mika Westerberg
> > <mika.westerberg@linux.intel.com> wrote:
> > 
> > > Sometimes it is useful to allow GPIO chips themselves to request GPIOs they
> > > own through gpiolib API. One usecase is ACPI ASL code that should be able
> > > to toggle GPIOs through GPIO operation regions.
> > >
> > > We can't really use gpio_request() and its counterparts because it will pin
> > > the module to the kernel forever (as it calls module_get()). Instead we
> > > provide a gpiolib internal functions gpiochip_request/free_own_desc() that
> > > work the same as gpio_request() but don't manipulate module refrence count.
> > >
> > > Since it's the GPIO chip driver who requests the GPIOs in the first place
> > > we can be sure that it cannot be unloaded without the driver knowing about
> > > that. Furthermore we only limit this functionality to be available only
> > > inside gpiolib.
> > >
> > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > 
> > I fully trust you in doing the ACPI stuff in patches 2-n but on this patch
> > in particular I want Alexandre's review tag as well, as he's working
> > actively with the descriptor API and I don't want to add too many quirks
> > without his consent.
> 
> Thanks for your trust :)
> 
> > So Alexandre, what do you say about this?
> 
> I'm about to send v2 of the series with Rafael's comments taken into
> account. However, I stumbled to another locking problem:
> 
> I'm going to move taking the gpio_lock outside of __gpiod_request() and
> have __gpiod_request() to release that lock, so that we can call
> chip->request() safely.
> 
> Since we are using _irqsave()/_irqrestore() versions, it means that I need
> to pass flags as a pointer from gpiod_request() to __gpiod_request() like:
> 
> 	spin_lock_irqsave(&gpio_lock, flags);
> 	if (try_module_get(chip->owner)) {
> 		ret = __gpiod_request(desc, label, &flags);

Ouch.  Sorry for overlooking that.

> 		...
> 
> Is that acceptable or can you guys suggest some alternative? One
> alternative that I can think about is to have __gpiod_request() to take the
> lock and move try_module_get() outside to gpiod_request():
> 
> __gpiod_request()
> {
> 	unsigned long flags;
> 
> 	spin_lock_irqsave(&gpio_lock, flags);
> 	...
> }
> 
> gpiod_request():
> {
> 	...
> 	if (try_module_get(chip->owner)) {
> 		ret = __gpiod_request(desc, label);
> 		...
> }
> 
> Thoughts?

If that works, it would be better than passing the flags IMO.

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.