On Wed, Jun 15, 2022 at 09:24:55AM +0300, Tony Lindgren wrote: > We want to enable runtime PM for serial port device drivers in a generic > way. To do this, we want to have the serial core layer manage the > registered serial port controllers. For runtime PM, we need a way to find > the serial ports for each serial port controller device. > > The serial core manages ports. Each serial controller can have multiple > ports. As serial core has no struct device, and the serial port device > drivers have their own driver data, we cannot currently start making > use of serial core generic data easily without changing all the serial > port device drivers. Really? Why not make struct uart_port a real struct device? > We could consider adding a serial core specific struct device. It would > be a child of the serial port device, and would allow us eventually to use > device_links to add generic runtime PM calls for example. But as the serial > core layer is not a device driver, driver specific features would need to > be added, and are probably not justified for a virtual device. I think it's very justified, let's not paper over this whole thing by adding a kref stuck in in the middle and trying to hook up the PM code to it, instead of just using all of the PM logic that the driver model already provides. > Considering the above, let's improve the serial core layer so we can > manage the serial port controllers better. Let's register the controllers > with the serial core layer in addition to the serial ports. Why can't controllers be a device as well? Let's try to work with the driver model here, not work around it, if at all possible. We never did a full conversion of the serial layer to the driver core all those decades ago. Perhaps now is the time to really do that. thanks, greg k-h
