Re: [patch/rfc 2.6.19-rc5] arch-neutral GPIO calls

[David Brownell <david-b@pacbell.net>]

On Monday 13 November 2006 9:56 am, Thiago Galesi wrote:
> I guess that a kind of "name addressing" would be the way to go, we
> need to get to it by "location" (I'm thinking PortA 20, PortB 5 rather
> than pin number or some other arbitrary convention; we need a way to
> not need to look up what 'Port X pin B' should be called.

That's highly platform-specific.  AT91 and AVR32 use that style of
addressing ... most others just talk about GPIO numbers.  In all cases,
the platform can assign _some_ number to each GPIO signal.


> Another thing that may be considered is the ability to get 'pointers'
> for GPIOs.

The GPIO identifiers are unsigned integers.  If you really crave pointers,
you can cast those integers to pointers, and back.  But why bother?  :)


> And, of course, protecting GPIOs from concurrent accesses 

Did you read the proposal?  There are *already* two mechanisms for that:

 - gpio_request()/gpio_free() ... protecting from concurrent access
   between drivers, also known as bugs.

 - gpio_set_value() ... implementations can lock internally, as needed.
   ditto gpio_get_value(), but only bizarre hardware would need it.

Re that last, consider two different types of GPIO controller:

  * One of them has just one register with the output values, and a
    RISC CPU without bit set/clear instructions to use on it.  So a
    spinlock is needed to cover reading the register, modifying that
    value, then rewriting the value.

  * Better GPIO controller designs (normal ones) have separate registers
    for "write mask to set" and "write mask to clear" ... no spinlock
    needed, the hardware sorts everything out.

I did update the docs to make clear that gpio get/set calls are atomic;
hard to imagine single bit operations that aren't, but it's worth
removing potential confusion.

- Dave