Re: [PATCH v9 3/4] dtc: Plugin and fixup support

[Phil Elwell <phil-FnsA7b+Nu9XbIbC87yuRow@public.gmane.org>]

On 24/11/2016 14:23, Pantelis Antoniou wrote:
> Hi Phil,
>> I would have thought that all DTBs already contain enough dependency
>> information in the form of the phandles themselves. One of the first
>> things the kernel does is to unflatten the DTB, and that is the obvious
>> point to resolve the phandles and generate the necessary dependencies.
>> Can you explain how both __fixups__ and __local_fixups__ aid this
>> process? Ideally they wouldn't duplicate any information already in the
>> tree, since then you have to cope with the possibility of malformed DTBs
>> where the two don't actually match.
> No, the DTBs by themselves do not contain enough information to build the
> probe dependency graph, because phandles are simply converted to 32 bit
> cell values on compile.
>
> For instance take a case of one node using the other:
>
> foo_label: foo { };
> bar { use = <&foo_label>; };
>
> A standard compile would generate a bar node as bar { use = <1>; };
>
> While using the -@ switch you’d get a local ref that says that there is
> a phandle cell value at offset 0 of bar/use property. Looking up the
> phandle value you can see that this property references the foo node.
>
> So when building the probe order graph the foo node should be probed
> before bar.
Yes, you are quite right. During compilation the phandleness of a
phandle reference gets lost - it's just an integer. All you really need
is one bit per word, and you could dream up clever coding schemes for
that sparse map, but extra information is needed.

I still think that __fixups__ is unnecessary for base DTBs, but since it
will be empty anyway it just looks a bit awkward rather than taking up
much space.

Seemingly random load order of device nodes has been a cause of
confusion and annoyance, so I'm happy that there is a plan to eliminate
them. If the DTBs have to grow to make that possible then so be it.

Phil