On Wed, Oct 26, 2016 at 4:23 PM, Jerome Brunet <jbrunet@xxxxxxxxxxxx> wrote: > [Me] >> 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@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html