On Tue, May 6, 2014 at 6:32 PM, Rob Herring <robherring2@xxxxxxxxx> wrote: > On Tue, May 6, 2014 at 7:48 PM, Frank Rowand <frowand.list@xxxxxxxxx> wrote: >> An issue with the path of SPMI nodes under /sys/bus/... was reported in >> https://lkml.org/lkml/2014/4/23/312. The symptom is that two different >> grandchild nodes of the spmi with the same node-name@unit-address will >> result in attempting to create duplicate links at >> /sys/bus/platform/devices/unit-address.node-name. It turns out that the >> specific example provided might not be an expected configuration for >> current hardware, but the reported trap remains an issue. >> >> I have been poking at the problem, trying to figure out how to cleanly >> fix the issue without breaking devicetree device creation. >> >> The first patch in the series is the one that may be a very bad idea. Or >> it may help show the way forward to deal with what I think is the major >> underlying problem. I have not finished investigating the possible negative >> side effects. And I am still thinking whether this is a conceptually good >> approach, or whether it is simply an expediant hack that hides the underlying >> problem. But I am throwing this out prematurely because I have mentioned >> it to several people, and I want to make it visible to everyone involved. >> >> The underlying architectural problem (in my opinion) is that a lot of devices >> are created by the device tree infrastructure as platform devices, when they >> truly should not be platform devices. They should not be platform devices >> because they are not physically on a platform bus, they are instead somewhere >> below some other bus. The first patch in this series is a hack which >> results in the devices still being represented by "struct platform_device" >> objects, but with a link to their parent's "struct bus_type" instead of >> to &platform_bus_type. >> >> The second patch does not require the first patch. The second patch provides >> a mechanism to allow subsystems to provide a method of naming devices to >> avoid name collisions. >> >> The third patch provides an example of a subsystem using the new feature >> provided by the second patch. >> > > I think the primary question to ask is there any added benefit to > having the additional hierarchy of devices. I don't think there is > much support to have more hierarchy from what I have seen of past > discussions. > > Another approach could be to support having multiple platform bus > instances. Then drivers can easily create a new instance for each set > of sub-devices. > > This can be solved in a much less invasive way just in the DT naming > algorithm. This is slightly different from what I had suggested of > just dropping the unit address. It keeps the unit address, but adds > the unique index on untranslate-able addresses. The diff is bigger due > to refactoring to reduce the indentation levels. It is untested and > whitespace corrupted: The unique index leads to an interesting dependency between the order of nodes in the DTB and userspace; paths to e.g. our the path to our block devices contains soc.X where X changes now and then. Fortunately soc.X won't change that often, but forcing more peripheral nodes to use the same schema would show the problem quite often. Does translatable/untranslatable refer to if this is an address translatable my the cpu or that it's just a translatable address on this specific bus? As far as I can see it's the latter and in our case (revid { reg = <0x100, 0x100>; }) seem to be translatable with the current implementation. Regards, Bjorn -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html