On Fri, Dec 1, 2023 at 4:24 PM Nuno Sá <noname.nuno@xxxxxxxxx> wrote: > > If a pins .direction_output() fails, .set_value() will not be called > > on it either. > > This is where I lost you :( devm_gpiod_get() (and similar interfaces) will set up the default mode for the line, as input or output (with value, calling .direction_output) so most likely it will fail already there, and the driver will not probe or userspace client will fail. > So, I'm might be overcomplicating things but... Again, > the case where someone wired up HW so that we can actually use the pin to drive the > line high (having an external pull up). In that case, If I return error, then I won't > be able to effectively set the line high (as you said, set_value will not be called > on it either). > > Now, I do understand that if we have the line flagged as GPIO_OPEN_DRAIN, then > gpiolib will switch the line to input which means we will set the line in high-z > which means that if we have a pull up, then the line will be high. I mean, it works > but it would be strange if someone wants to have the line as output high and after > trying to set the it high, it sees the pin moving to input. But if this is how it > should be, fine by me. What do you mean by "sees the pin moving to input". If you mean electrically then yes, it goes to high-Z. If you mean logically, as seen by software and GPIO and debugfs, not really. I think a good exercise to see how it works is to just walk through the code in drivers/gpio/gpiolib.c for e.g. gpiod_set_value() gpiod_set_value_nocheck() gpio_set_open_drain_value_commit() >direction_input() is indeed called, but that is just a way of using the hardware, logically, inside gpiolib, the line is handled as high. > Yes, that is the only thing we have. Meaning that there is no hw setting to set the > pins to open drain. Open drain is what they are. That is why I'm not seeing the point > in having PIN_CONFIG_DRIVE_OPEN_DRAIN implemented. For satisfying the logic. We have several cases where callbacks are just returning a 0 error code for such corner cases, and as Andy points out push-pull requests should return an error. Yours, Linus Walleij
