Re: [Patch v2 1/2] gpio: add GPIO hogging mechanism

[Maxime Ripard]

On Sat, Dec 06, 2014 at 09:08:36PM +0900, Alexandre Courbot wrote:
> On Fri, Dec 5, 2014 at 7:24 PM, Maxime Ripard
> <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
> > On Thu, Dec 04, 2014 at 11:49:19PM +0900, Alexandre Courbot wrote:
> >> On Thu, Dec 4, 2014 at 11:27 PM, Maxime Ripard
> >> <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
> >> > Hi,
> >> >
> >> > On Thu, Dec 04, 2014 at 11:15:38PM +0900, Alexandre Courbot wrote:
> >> >> On Wed, Dec 3, 2014 at 1:12 AM, Maxime Ripard
> >> >> <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
> >> >> > On Tue, Dec 02, 2014 at 03:29:46PM +0100, Linus Walleij wrote:
> >> >> >> On Tue, Dec 2, 2014 at 3:13 PM, Alexandre Courbot <gnurou@xxxxxxxxx> wrote:
> >> >> >> > On Tue, Dec 2, 2014 at 1:36 AM, Maxime Ripard
> >> >> >> > <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote:
> >> >> >>
> >> >> >> >> The only thing I'd like to have would be that the request here would
> >> >> >> >> be non-exclusive, so that a later driver would still be allowed later
> >> >> >> >> on to request that GPIO later on and manage it itself (ideally using
> >> >> >> >> the usual gpiod_request function).
> >> >> >> >
> >> >> >> > Actually we have a plan (and I have some code too) to allow multiple
> >> >> >> > consumers per GPIO. Although like Benoit I wonder why you would want
> >> >> >> > to hog a GPIO and then request it properly later. Also, that probably
> >> >> >> > means we should abandon the hog since it actively drives the line and
> >> >> >> > would interfere with the late requested. How to do that correctly is
> >> >> >> > not really clear to me.
> >> >> >>
> >> >> >> I don't get the usecase. A hogged GPIO is per definition hogged.
> >> >> >> This sounds more like "initial settings" or something, which is another
> >> >> >> usecase altogether.
> >> >> >
> >> >> > We do have one board where we have a pin (let's say GPIO14 of the bank
> >> >> > A) that enables a regulator that will provide VCC the bank B.
> >> >> >
> >> >> > Now, both banks are handled by the same driver, but in order to have a
> >> >> > working output on the bank B, we do need to set GPIO14 as soon as
> >> >> > we're probed.
> >> >> >
> >> >> > Just relying on the usual deferred probing introduces a circular
> >> >> > dependency between the gpio-regulator that needs to grab its GPIO from
> >> >> > a driver not there yet, and the gpio driver that needs to enable its
> >> >> > gpio-regulator.
> >> >>
> >> >> I don't get it. According to what you said, the following order should
> >> >> go through IIUC:
> >> >>
> >> >> 1) bank A is probed, gpio 14 is available
> >> >> 2) gpio-regulator is probed, acquires GPIO 14, regulator for Bank B is available
> >> >> 3) bank B is probed, grabs its regulator and turn it on, probes.
> >> >>
> >> >> What am I missing?
> >> >
> >> > It would be true if bank A and B were exposed through different
> >> > drivers (or at least different instances of the same driver), which is
> >> > not the case.
> >> >
> >> > In our case, banks A and B are handled by the same instance.
> >>
> >> Ok, so both banks A and B are part of the same device/DT node. Now I
> >> think I understand the issue. You need to hog the pin so that bank B
> >> will work right after the device is probed.
> >
> > Exactly.
> >
> >> But you will still have the problem that the regulator device will
> >> *not* be available when your device is probed, so you cannot call
> >> regulator_get() for bank B anyway. I guess your only choice is to hog
> >> that pin and leave it active ad vitam eternam.
> >
> > Hmmm, indeed.
> >
> > I'll stop boring you with this then :)
> 
> Please *keep* bothering us with any doubt you may have until they are
> all cleared and you are sure this feature will be useful to you.
> Especially since we are designing DT bindings that will have to be
> carried over forever. We really want to get them right, and need input
> of potential users for that.
> 
> Having a few design arguments is a small thing compared to the hassle
> of having to work with unadapted features and bindings.

Ok. What I had in mind when I first thought about it was to set GPIO
as hogged through the GPIO flags, and then have a dumb GPIO driver
like Pantelis was suggesting.

I don't know if hogged would still be the right term, but we could
have that flag that would just allow the value to be set through
gpio_request init value, and deny/cache any subsequent change through
gpio_set_value.

gpio_request with this flag would never return EPROBE_DEFER, and just
cache the value to be set for when the driver comes in.

We could enforce driver-less GPIOs through that dumb driver, and we
would still be able to break weird dependency chains that end up in
circle.

That's just a thought though.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

Attachment: signature.asc
Description: Digital signature