On 15:05 Thu 30 May , Arnd Bergmann wrote: > On Thursday 30 May 2013, Tomi Valkeinen wrote: > > On 30/05/13 14:12, Jean-Christophe PLAGNIOL-VILLARD wrote: > > > On 12:34 Thu 30 May , Tomi Valkeinen wrote: > > >> On some platforms DPI requires a regulator to be enabled to power up the > > >> output pins. This regulator is, for some reason, currently attached to > > >> the virtual omapdss device, instead of the DPI device. This does not > > >> work for DT, as the regulator mappings need to be described in the DT > > >> data, and the virtual omapdss device is not present there. > > >> > > >> Fix the issue by acquiring the regulator in the DPI device. To retain > > >> compatibility with the current board files, the old method of getting > > >> the regulator is kept. The old method can be removed when the board > > >> files have been changed to pass the regulator to DPI. > > > > > > as discuss with Arnd we should handle regular enable and disable at device > > > probe for every device as we do for pinctrl > > > > I'm not sure what you mean. Enable of what? The regulator? Why would we > > enable it in the device's probe, as the device may never even be used? > > It's an idea I had a while ago, but not yet discussed in the open. > > Jean-Christophe just posted patches to move the mapping of interrupt numbers > into platform_drv_probe(), just before calling the driver ->probe() callback, > and we already have similar code to set up the default pinctrl state of > a device before calling probe(). > > This can be extended to further subsystems, but that has to be done > carefully to avoid regressions. Ideally we would move a lot of boilerplate > code out of the driver specific ->probe() function into common code. > Possible candidates for this include: > > * calling devm_request_mem_region for the "reg" property > * calling devm_ioremap on the "reg" property" > * calling devm_gpio_request for all gpio lines > * calling devm_regulator_get on all regulators > * calling devm_reset_control_get on all reset lines > * calling devm_dma_request_slave_channel on all dma channels > * calling devm_of_pwm_get for all pwm channels > * ... > > For most of these (maybe all), I think we need some form of opt-in > model on the driver side because there are cases where aquiring some > of these resources is not mandatory, and it only works if the driver > is using DT probing. > > IF we want to do this, it also needs a lot of thought, and we shouldn't > do it carelessly. We might also need some extra infrastructure in revres > to simplify access to the resources we got from the OF node. > > The irq resources are particularly trivial because we already claim > them in of_platform_populate, so moving that to platform_drv_probe() > is straightforward and solves existing bugs without creating a huge > regression potential, but it's harder for the others.
Yeah I agree with Arnd we need to start to move this way but for DT only first and carefully Best Regards, J. > > Arnd > -- > To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in > the body of a message to majord...@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html