Re: [PATCH 3/3] [POWERPC] QE: implement GPIO LIB API

[Anton Vorontsov <cbouatmailru@gmail.com>]

On Sun, Jan 27, 2008 at 01:59:51PM -0700, Grant Likely wrote:
a> On 1/27/08, Anton Vorontsov <cbouatmailru@gmail.com> wrote:
> > On Sun, Jan 27, 2008 at 02:42:12PM +0100, Jochen Friedrich wrote:
> > > Hi Anton,
> > >
> > > > +static void qe_gpio_set(struct gpio_chip *gc, unsigned int gpio, int val)
> > > > +{
> > > > +   struct of_mm_gpio_chip *mm_gc = to_of_mm_gpio_chip(gc);
> > > > +   struct port_regs *regs = mm_gc->regs;
> > > > +   u32 pin_mask;
> > > > +   u32 tmp_val;
> > > > +
> > > > +   /* calculate pin location */
> > > > +   pin_mask = (u32) (1 << (NUM_OF_PINS - 1 - gpio));
> > > > +
> > > > +   tmp_val = in_be32(&regs->cpdata);
> > > > +
> > > > +   if (val == 0)
> > > > +           out_be32(&regs->cpdata, ~pin_mask & tmp_val);
> > > > +   else
> > > > +           out_be32(&regs->cpdata, pin_mask | tmp_val);
> > > > +}
> > >
> > > I see a possible problem with this (and in the corresponding call in CPM1, as well):
> > >
> > > if there is a pin configured as open drain, you might accidently switch this pin to 0
> > > while switching a different pin, if an external device is pulling the pin to 0.
> >
> > Unfortunately I can't think out any workaround for this, except
> > implementing generic gpio_bank_{,un}lock(gpio_pin_on_the_bank), and
> > start using it in the drivers that might care about this issue. Though,
> > looking into i2c-gpio.c I don't clearly see were we can insert these
> > locks, there should be "start/end transaction" handlers or something,
> > but it seems that it's in the bitbanging code, not in the i2c-gpio
> > driver..
> >
> > Actually, I see this as a hardware limitation. For example, on ARMs
> > PXA2xx, there are separate, per bank, read/set/clear GPIO registers,
> > not all-in-one data register.
> 
> I've run into this exact issue on other GPIO hardware too.  It's not
> uncommon behaviour in GPIO hardware.
> 
> The solution is to not depend on the hardware to remember what the
> output pin values should be.  Add a shadow register in the driver
> private data.  Set the pin state for each output pin in the shadow
> register and then write that value to the hardware.  That way input
> state doesn't interfere with the output values.

Great idea, much thanks. Would be easy to implement also.

> Also, you do still need spinlocks around the manipulation of the
> shared registers; otherwise you'll have very hard to debug race
> conditions.  Probably one spin lock per bank.

With GPIO LIB we already have per bank spinlock, so it isn't
a problem.


Thanks,

-- 
Anton Vorontsov
email: cbou@mail.ru
backup email: ya-cbou@yandex.ru
irc://irc.freenode.net/bd2