On Wed, May 7, 2014 at 2:32 AM, 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. > I fear that it will be more invasive that you think it will be. Right now the bus_type abstraction is a mechanism for matching drivers and devices. The bus type contains a bag of device drivers, and it tries to match one of them to a device when a device gets registered to that device (or when a driver gets registered, try to match it to one of the devices it already knows about). You can see this in the /sys/bus/<type>/drivers and /sys/bus/<type>/devices directories. Splitting the platform bus type into multiple instances is not trivial because the drivers will only be available to one instance. You'd need to figure out how to make a device driver available to multiple bus_type instances (ideally without having to manually add the driver to each bus_type at module load time). > 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: > > diff --git a/drivers/of/platform.c b/drivers/of/platform.c > index 404d1da..c77dd7a 100644 > --- a/drivers/of/platform.c > +++ b/drivers/of/platform.c > @@ -105,23 +105,33 @@ void of_device_make_bus_id(struct device *dev) > * For MMIO, get the physical address > */ > reg = of_get_property(node, "reg", NULL); > - if (reg) { > - if (of_can_translate_address(node)) { > - addr = of_translate_address(node, reg); > - } else { > - addrp = of_get_address(node, 0, NULL, NULL); > - if (addrp) > - addr = of_read_number(addrp, 1); > - else > - addr = OF_BAD_ADDR; > - } > - if (addr != OF_BAD_ADDR) { > - dev_set_name(dev, "%llx.%s", > - (unsigned long long)addr, node->name); > - return; > - } > + if (!reg) > + goto no_bus_id; > + > + if (of_can_translate_address(node)) { > + addr = of_translate_address(node, reg); > + if (addr == OF_BAD_ADDR) > + goto no_bus_id; > + > + dev_set_name(dev, "%llx.%s", > + (unsigned long long)addr, node->name); > + return; > } > > + addrp = of_get_address(node, 0, NULL, NULL); > + if (!addrp) > + goto no_bus_id; > + > + addr = of_read_number(addrp, 1); > + if (addr == OF_BAD_ADDR) > + goto no_bus_id; > + > + magic = atomic_add_return(1, &bus_no_reg_magic); > + dev_set_name(dev, "%llx.%s.%d", (unsigned long long)addr, > + node->name, magic - 1); > + return; > + > +no_bus_id: Looks like a reasonable change to me. g. > /* > * No BusID, use the node name and add a globally incremented > * counter (and pray...) -- 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
