On Mon, Nov 13, 2017 at 10:20:28AM +0000, Lee Jones wrote: > On Mon, 13 Nov 2017, Johan Hovold wrote: > > > On Mon, Nov 13, 2017 at 09:11:44AM +0000, Lee Jones wrote: > > > On Sun, 12 Nov 2017, Johan Hovold wrote: > > > > > > > [ +CC: Lee, Rob and device-tree list ] > > > > > > > > On Sat, Nov 11, 2017 at 09:50:59AM -0800, Dmitry Torokhov wrote: > > > > > On Sat, Nov 11, 2017 at 04:43:37PM +0100, Johan Hovold wrote: > > > > > > A helper purported to look up a child node based on its name was using > > > > > > the wrong of-helper and ended up prematurely freeing the parent of-node > > > > > > while searching the whole device tree depth-first starting at the parent > > > > > > node. > > > > > > > > > > Ugh, this all is pretty ugly business. Can we teach MFD to allow > > > > > specifying firmware node to be attached to the platform devices it > > > > > creates in mfd_add_device() so that the leaf drivers simply call > > > > > device_property_read_XXX() on their own device and not be bothered with > > > > > weird OF refcount issues or what node they need to locate and parse? > > > > > > If a child compatible is provided, we already set the child's > > > of_node. It's then up to the driver (set) author(s) to use it in the > > > correct manner. > > > > > > > Yeah, that may have helped. You can actually specify a compatible string > > > > in struct mfd_cell today which does make mfd_add_device() associate a > > > > matching child node. > > > > > > > > Some best practice regarding how to deal with MFD and device tree would > > > > be good to determine and document too. For example, when should > > > > of_platform_populate() be used in favour of mfd_add_device()? > > > > > > When the device supports DT and its entire hierarchical layout, along > > > with all of its attributes can be expressed in DT. > > > > Ok, a follow up: When there are different variants of an MFD and that > > affects the child drivers, then that should be expressed in in the child > > node compatibles rather than having the child match on the parent node? > > > > I'm asking because this came up recently during review and their seems > > to be no precedent for matching on the parent compatible in child > > drivers: > > > > https://lkml.kernel.org/r/20171105154725.GA11226@localhost > > Accessing the parent's of_device_id .data directly doesn't sit well > with me. The parent driver should pass this type of configuration > though pdata IMHO. The child driver is only matching on the parent-node compatible string IIRC, and therefore keeps its own table of all parent compatibles with its own set of (child) private match data (i.e. the parent compatible property is matched first by the parent driver, and then again by the child). Passing through pdata here is not possible since mfd_add_device() isn't used, right? It could of course be described using properties of the child node (e.g. by using different child compatible strings). > > > > And how best to deal with sibling nodes, which is part of the problem > > > > here (I think the mfd should have provided a flag rather than having > > > > subdrivers deal with sibling nodes, for example). > > > > > > I disagree. The only properties the MFD (parent) driver is interested > > > in is ones which are shared across multiple child devices. > > > *Everything* which pertains to only a single child device should be > > > handled by its accompanying driver. > > > > Even if that means leaking details of one child driver into a sibling? > > Not sure what you mean here. Could you please elaborate or provide an > example? I mean that the sibling node needs to be aware of the name, compatible string, or other node properties of its sibling node to be able to parse sibling nodes itself (rather than the sibling or parent doing so on its behalf). But it seems you answer this below. > > Isn't it then cleaner to use the parent MFD to coordinate between the > > cells, just as we do for IO? > > > > In this case a child driver looked up a sibling node based on name, but > > This should not be allowed. If >1 sibling requires access to a > particular property, this is normally evidence enough that this > property should be shared and handled by the parent. > > > that doesn't mean the node is active, that there's a driver bound, or > > that the sibling node has some other random property for example. The > > parent could be used for such coordination, if only to pass information > > from one sibling to another. > > Right. Ok, it seems we're in agreement here. Given that MFD has evolved over time and device-tree support has been added retroactively to some drivers, we've ended up with a multitude of different ways of dealing with such issues. I think it may still be a good idea to jot down some best practices for future driver developers. Thanks, Johan