On Mon, Oct 05, 2020 at 02:59:04PM -0500, Rob Herring wrote: > On Mon, Oct 5, 2020 at 2:36 PM Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote: > > > > On Mon, Oct 05, 2020 at 12:18:12PM -0700, Matthias Kaehlcke wrote: > > > On Mon, Oct 05, 2020 at 12:15:27PM -0400, Alan Stern wrote: > > > > The conclusion is that we need to have code that is aware of some > > > > detailed needs of a specific device but is not part of the device's > > > > driver. I'm not sure what the best way to implement this would be. > > > > > > Wouldn't it be possible to load the module when the DT specifies that > > > the device exists? For USB the kernel would need the VID/PID to identify > > > the module, these could be extracted from the compatible string. > > > > Loading a driver module whenever DT says a device exists? Not a bad > > idea. I don't know what would be involved, but no doubt it is possible. > > MODULE_DEVICE_TABLE mostly as I mentioned in my other reply. > > > Note that, except for a few special cases, the kernel identifies the > > appropriate driver for USB hubs not by the VID/PID but instead by the > > device class or interface class. I suppose the compatible string could > > include that information too? > > We can go back to 1998 OpenFirmware and it's already there[1]. > 'usb,class9' for a hub. There's a few other variations defined. That should work if the initialization is simple enough that the info in the device tree is sufficient (e.g. switching a single regulator on), otherwise a device specific compatible string would be needed. > > > Having the initialization code outside of the driver could lead to code > > > duplication, since the driver might want to power the device down in > > > certain situations (e.g. system suspend). > > > > True. On the other hand, how common do you think it would be for > > drivers not to want to mess with the power settings? > > I think in that case you'd generally want firmware to enable things > and the kernel then does no power control. > > We have ~1500 boards using DT and maybe ~10 with USB devices described > in DT. So the whole thing is not common to begin with. It's probably not very common, but might be more common than the DT suggests. Many devices probably don't specify their hub(s) or other USB devices explicitly when the initialization is done in firmware. In case a generic solution for all types of devices and busses is not required we would still need a driver to address at least the conditional power down of a hub during system suspend. Doug summarized the state of the discussion about the bindings for hubs (https://lore.kernel.org/patchwork/patch/1313000/#1511757) maybe we should continue from there?
