Re: [PATCH] libfdt: place new nodes & properties after the parent's ones

[Rob Herring]

On Wed, Feb 12, 2020 at 12:57 AM Marek Szyprowski
<m.szyprowski@xxxxxxxxxxx> wrote:
>
> Hi Rob,
>
> On 11.02.2020 21:29, Rob Herring wrote:
> > On Mon, Feb 10, 2020 at 5:44 PM David Gibson
> > <david@xxxxxxxxxxxxxxxxxxxxx> wrote:
> >> On Mon, Feb 10, 2020 at 12:40:19PM +0100, Marek Szyprowski wrote:
> >>> Hi David,
> >>>
> >>> On 05.02.2020 06:45, David Gibson wrote:
> >>>> On Tue, Feb 04, 2020 at 01:58:44PM +0100, Marek Szyprowski wrote:
> >>>>> While applying dt-overlays using libfdt code, the order of the applied
> >>>>> properties and sub-nodes is reversed. This should not be a problem in
> >>>>> ideal world (mainline), but this matters for some vendor specific/custom
> >>>>> dtb files. This can be easily fixed by the little change to libfdt code:
> >>>>> any new properties and sub-nodes should be added after the parent's node
> >>>>> properties and subnodes.
> >>>>>
> >>>>> Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
> >>>> I'm not convinced this is a good idea.
> >>>>
> >>>> First, anything that relies on the order of properties or subnodes in
> >>>> a dtb is deeply, fundamentally broken.  That can't even really be a
> >>>> problem with a dtb file itself, only with the code processing it.
> >>> I agree about the properties, but generally the order of nodes usually
> >>> implies the order of creation of some devices or objects.
> >> Huh?  From the device tree client's point of view the devices just
> >> exist - the order of creation should not be visible to it.
> > I'm not sure if downstream is different, but upstream this stems from
> > Linux initcalls being processed in link order within a given level.
> > It's much better than it used to be, but short of randomizing the
> > ordering, I'm not sure we'll ever find and fix all these hidden
> > dependencies.
>
> Downstream is probably much worse, because I've seen a lots of custom
> code iterating over the nodes and doing its initialization.
>
> >>> This sometimes
> >>> has some side-effects.
> >> If those side effects matter, your code is broken.  If you need to
> >> apply an order to nodes, you should be looking at 'reg' or other
> >> properties.
> > The general preference is to sort by 'reg'. And we try to catch and
> > reject any node re-ordering patches.
>
> If one applies an dt-overlay with current libfdt, the order of the all
> nodes from that overlay is reversed.

Sure, but my caring about the ordering ends at the source level.

Rob