On Wed, Oct 26, 2016 at 4:23 PM, Jerome Brunet [off-list ref] wrote:

[Me]

quoted

We usually refer to the local numberspace on the GPIO controller
as "offsets", so line offsets 0...31 on a gpiochip with 31 lines.

The ngpio in struct gpio_chip is the number of lines on that
controller,
and should nominally map 1:1 to hwirq sources.

Indeed it should be the the case, and for meson, it is pretty close.
The irqchip controller provide a number of hwirq. Each hwirq maps to
one, and only one, pin. But since not every pins are connected to the
irqchip controller, the opposite is not true.

Taking an example with 16 gpios, here is what it could look like with
the exception we have on meson :

gpio offset [ 0  1  2  3  4  5  6  7  8  9  10 11 12 13 14 15 ]
hwirq num   [ 0  1  2  3] NC NC[4  5  6  7  8  9  10]NC NC NC

Like gpio offset are used (internally) in the driver to find
appropriate gpio registers and bit, the hwirq has a meaning too.
It is the setting you put in the channel multiplexer of the controller
to select the proper pin to spy on.

In the end, these gpio offset and hwirq number are different. I would
prefer to have hwirq == gpio and go your way, it would make my life
easier, but I don't see how it would work.

The irqchip controller cares only about the hwirq number. You can
actually request an interrupt directly to the controller by asking the
proper hwirq number (in DT for example), without involving the gpio
driver (tested).

The relation between the pins and the interrupt number is provided by
the manufacturer in the Datasheet [1], in the section GPIO Interrupt.

I think I kind of get it.

This reminds me of recent patches to the generic GPIOLIB_IRQCHIP
where we made it possible to "mask" set of IRQ lines, saying
"these are in the irqdomain, but you cannot map them".

See
commit 79b804cb6af4f128b2c53f0887c02537a7eb5824
"gpiolib: Make it possible to exclude GPIOs from IRQ domain"
commit 96b2cca64fa3e1a31b573bb308b2944c802aacf3
"gpio: stmpe: forbid unused lines to be mapped as IRQs"

So what we do in the generic case is put a linear map over all
the lines/offsets, then punch holes in that map and say
"this and this and this can not be mapped as IRQ".

As you can see in _gpiochip_irqchip_add() we pre-map all
except those with irq_create_mapping().

Does this scheme fit with your usecase? I would guess so,
just that your domain is hierarchic, not simple/linear.

Maybe the takeaway is that your map does not need to
be "dense", i.e. every hwirq is in use. It can be sparse.
It is stored in a radix tree anyways.

Looking at other gpio drivers, it is not uncommon to have some simple
calculation to get from gpio offset to the hwirq number. I don't get
what is the specific problem here ?

It's OK to use the offset vs hwirq.

I just misunderstand it as the global GPIO number, that is
not OK.

Yours,
Linus Walleij
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html