On 30/09/12 17:19, Stijn Devriendt wrote:
>> If I understand correctly, it's a violation (single-value should hold
>> for read and write).
>>
>> To solve it, I have the following in mind: /sys/.../gpiogroupXXX/
>> contains files "bit0" ... "bit31" which contain a gpio number each,
>> empty if "unconnected".
> 
> Unfortunately that means you can't atomically create a group.

I don't see a big advantage of having atomic create/request. Most
important is set/get, isn't it? I assume the following usage pattern:

* Create(request) - non atomic (maybe atomic but why not add GPIOs later?)
* Set - atomic
* Get - atomic
* ...

> It also creates a mess to keep ordering intact and to either
> keep the current pin state or override it at allocation-time.

Ordering should stay intact, and later add/delete operations could be
possible. I meant bit0 ... bit31 in the gpio block as such:

bit0  - "80"
bit1  - ""  (i.e. unconnected)
bit2  - "85"
bit3  - "2"
...
bit31 - ""

This scheme can support multiple gpio_chips, as discussed with Linus and
JC, which of course can't always guarantee real simultaneous I/O but
provide virtual I/O word access (32bit/64bit).

> Rules are rules, but why make the interface overly complex when
> the multi-value file is saner, cleaner and simpler?

Simply because they won't (and probably shouldn't) accept it mainline.

Roland