Device trees

[Nathan Whitehorn]

Benno Rice wrote:
>
> On 24/07/2008, at 10:27 AM, Nathan Whitehorn wrote:
>
>> Benno Rice wrote:
>>> On 24/07/2008, at 8:47 AM, M. Warner Losh wrote:
>>>> Does FreeBSD/powerpc on the MPC85xx machine use the fake device open
>>>> firmware tree stuff from uboot (aka dts in the linux kernel sources)?
>>> I've got a machine (Mikrotik RouterBOARD 333, uses MPC8321) that has 
>>> this.  I've been working with one of the dtc developers to port 
>>> their code over to FreeBSD.  I'm trying to hide it underneath our 
>>> existing OpenFirmware stuff.  I've got a few more bugs to get out of 
>>> it but I can probably send you some patches early next week...
>>
>>
>> Can you use this to abstract real OF too? We don't support real mode 
>> Open Firmware and IBM is shipping new systems with a very basic 
>> firmware that runs only in real mode (this includes the Bimini 970MP 
>> dev board, apparently, as well as their Cell blades). The easiest 
>> course seems to be just to pretend it's a flattened device tree if we 
>> can support that.
>
> My plan (assuming some measure of available time) was to actually get 
> loader to handle turning OpenFirmware device data into an FDT which 
> could be handed to FreeBSD as part of the loader metadata.  On 
> platforms like the Mikrotik where all we get is an FDT, if we have 
> loader it can just pass it through or if we're not using loader we can 
> have a hook that shunts the data into the right place.
>
> Does that sound workable?
>
It would be nice not to require loader at all and support direct loading 
of the kernel. We already support this, and it is handy on systems with 
low-featured firmware that doesn't allow loader to easily load more files.

I think the best way forward is to support some kind of pluggable Open 
Firmware implementation. On systems with nonexistent/bad firmware, we 
can supply a fake interpreter with a very limited vocabulary to OF_init 
(I'm assuming this is what you are doing with the Microtik board) that 
either uses a passed-in FDT or generates one from the real-mode firmware 
early in the boot process.

But I don't think we should turn everything into an FDT: we gain some 
real advantages on systems with decent firmware implementations from 
being able to use them, like RTC and power management support. It would 
be silly to throw that away and I believe the existing architecture is 
flexible enough to support runtime switching between implementations.

Also, while I'm sending email to everyone, I'll give my apparently daily 
update on G5 support. We're self-hosting! This means that my patches 
(http://people.freebsd.org/~nwhitehorn/g5.diff, as before) fix building 
PPC world and that the system is now stable enough to build that world 
itself. The NFS bug has now been fixed in HEAD (thanks to dfr). You 
still can't build a kernel with INVARIANTS.
-Nathan