Re: [PATCH 2/6] PowerPC 440EPx: Sequoia DTS

[Segher Boessenkool <segher@kernel.crashing.org>]

> Aha!  Ok, now I understand the sorts of situations you're talking
> about.  By "not direct mapped", I thought you were talking about some
> kind of access via address/data registers on some indirect bus
> controller, rather than weird variations on endianness and
> bit-swizzling.
>
> Hrm.. this is a property of how the flash is wired onto the bus,
> rather than of the flash chips themselves, so I'm not entirely sure
> where description of it belongs.
>
> Simplest option seems to me to add a property "endianness" or
> "bit-swizzling" or something which can be defined to describe some odd
> connections.  If absent we'd default to direct mapping.  Segher, is
> that idea going to cause you to scream?

No, that's fine with me.  I would recommend either using a
_good_ _descriptive_ name for such a property describing the
swizzling, if this swizzling is common; or just put the whole
bloody weirdo address permutation into some nice big array,
something like

address-permutation = <0 1 3 2 4 5 7 6 e f d c a b 9 8>;

>>     Well, "device-width" is not the thing that we care about either. 
>> ;-)
>
> Well, yes but we need to encode it somehow.  Segher preferred
> device-width to interleave, because it's closer to a description of
> the actual hardware, rather than an abstration decribing its wiring.
>
> In other words I *still* don't see any reason to prefer giving the
> interleave.

You *need* to know the device width.  You *need* to know
the bank width.  And for 99% of the cases you don't need
to know anything more since most hardware designs aren't
insane.  Case closed I'd say.


>> Didn't know that the spec allows commas after @...
>
> Well, now you do.  I believe USB usually uses that format, too.

PCI too, and it even has an official binding :-)

>>     Don't we need "ranges" here, at least from the formal PoV -- as 
>> the parent
>> and child address spaces differ? I know the physmap_of parser doesn't 
>> care but
>> still...
>
> That's one I've wondered about.  To describe the partitions address
> space as lying (ultimately) in the physical address space, which in a
> sense it does, yes we'd need a ranges property here.  But we also have
> a 'reg' at the top level which would overlap with that hypothetical
> ranges which would be weird.  Or we could exclude the top-level reg,
> but then that's a pain if we do want to map the flash as a whole.
>
> So I left out ranges, on the grounds that there isn't actually
> anything at present which will attempt to access flash partitions
> "generically" as a device tree device.

It looks good to me like this.

In a real OF, the "register" access for the flash partition
node would be handled by its parent node, which would know
to do the direct-mapping thing (at least mapping it to _its_
parent, which typically asks the nodes further up, etc.)

For the kernel world, we should just document it in the binding.

> I'm not sold on this approach, but I haven't heard you give a better
> argument yet.

I haven't heard or thought of anything better either.  Using "ranges"
is conceptually wrong, even ignoring the technical problems that come
with it.


Segher