Marc Gonzalez <[email protected]> writes: > On 20/01/2016 17:10, Måns Rullgård wrote: > >> Marc Zyngier wrote: >> >>>> + if (of_property_read_u32(node, "reg", &ctl)) >>>> + panic("%s: failed to get reg base", node->name); >>>> + >>>> + chip = kzalloc(sizeof(*chip), GFP_KERNEL); >>>> + chip->ctl = ctl; >>>> + chip->base = base; >> >> As I said before, this assumes the outer DT node uses a ranges >> property. Normally reg properties work the same whether they specify an >> offset within an outer "ranges" or have a full address directly. It >> would be easy enough to make this work with either, so I don't see any >> reason not to. > > IIRC, I was told very early in the review process that the ranges prop > was mandatory. Lemme look for it... It was Arnd: > > http://thread.gmane.org/gmane.linux.ports.arm.kernel/444131/focus=444207 > >> You are missing a ranges property that describes what address >> space these addresses are in. >> >> 'ranges;' would be wrong here, as the interrupt controller is >> not a bus. If you have no ranges property, the bus is interpreted >> as having its own address space with no relation to the parent bus >> (e.g. an I2C bus uses addresses that are not memory mapped). >> >> Just list the addresses that are actually decoded by child >> devices here. > > Did I misunderstand?
It's still possible to create such a device tree, and that will fail in very hard to debug ways. Better to be a bit robust. -- Måns Rullgård
