Re: [PATCH 3/3] of/gpio: Introduce of_put_gpio(), add ref counting for OF GPIO chips

[Grant Likely <grant.likely@secretlab.ca>]

On Tue, Feb 9, 2010 at 12:14 PM, Anton Vorontsov
<avorontsov@ru.mvista.com> wrote:
> On Tue, Feb 09, 2010 at 10:28:15AM -0700, Grant Likely wrote:
> [...]
>> Rather than having a lock at the device tree data pointer level which
>> mixes usage with potentially many other drivers; wouldn't it make more
>> sense to use a mutex at the of_gc subsystem context?
>
> I don't think so.
>
> of_gc = np->data;
> lock(of_gc); (or lock(devtree))
> <do something with of_gc>
>
> doesn't provide us what we need, i.e. it doesn't guarantee that
> np->data (of_gc) is still alive.
>
> And here:
>
> lock(np->data); (or lock(devtree))
> of_gc = np->data;
> lock(of_gc);
> <do something with of_gc>
>
> The second lock becomes useless (unless you also refcount np->data
> usage and can drop the devtree/np->data lock, and grab some other
> kind of lock, e.g. mutex, but this is silly).

Okay, I'm convinced now.  The model is wrong.  struct of_gc does need
to be a member of struct gpio_chip and conditionally compiled against
CONFIG_OF_GPIO.  This locking requirement is just too plain ugly, and
dereferencing the np->data pointer in this way is dangerous (what if
something that is not struct of_gc is stored there).

Put of_gc into struct gpio_chip, and I'll completely support that approach.

g.

-- 
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.