Re: Dual core Power Mac G5

[Michael <macallan%netbsd.org@localhost>]

Hello,

On Nov 11, 2008, at 6:01 PM, Yevgeny Binder wrote:

> On Nov 11, 2008, at 9:57 AM, Michael wrote:
>
> > ...
> > I kinda-sorta succeeded getting my G5 booted over the network  
> > ( hint: the firmware wants your DHCP server to have the  
> > 'authoritative' flag set and you can feed it the kernel directly,  
> > without ofwboot.xcf ) but the kernel dies very early in the  
> > process. The early rasops console appears to work though, I got a  
> > few lines of output with wrong colours ( meaning it wasn't  
> > OF_write() ). Didn't make it to CPU attachment though.
> > I've been busy getting native Xorg going on all kinds of different  
> > hardware, hopefully I'll have time for the G5 very soon. Of course,  
> > any help is more than welcome.
>
> I've been out of the loop for a while since my Google SoC project,  
> so I don't know how relevant this is, but I did manage to netboot a  
> dual-processor single-core G5 into single user mode under its own  
> console after modifying Sanjay Lal's original code; I also had a  
> friend successfully boot a dual-core Power Mac G5 with the same  
> kernel. The patches (though now two years old) are available at http://netbsd-soc.cvs.sourceforge.net/viewvc/netbsd-soc/ppcg5 
> . I doubt that this code has been integrated into the official  
> sources, so it might be of some use to you now.

Thanks, this will definitely help.
I've been playing with my G5 a bit today and ran into a bunch of issues:
- the 'available' property in /memory doesn't obey #address-cells on  
my G5. Wrote a workaround, not sure how to reliably detect it, for now  
I check if (size >> 2) % (#address-cells + #size-cells) == 0 but that  
may not catch everything.
- the early rasops console works out of the box, OF maps all the BARs  
we need where we need them
- the whole thing hangs when I enable address translation. If I only  
enable address translation for instructions I get a little further so  
I guess I'm still missing hw mappings somewhere.
- detecting memory and so on seems to do The Right Thing(tm)
- the openpic node is no longer under mac-io - it's /u4/mpic now and  
the address is no longer relative to the parent node

have fun
Michael