On Thu, Sep 30, 2010 at 5:02 PM, Andreas Färber andreas.faerber@web.de wrote:
Hello,
I'm investigating the idea of a new QEMU chrp machine derived from mac99. The background is that there are ofmem issues when running Haiku on the g3beige machine, whereas running AIX on mac99 fails with the fix in place that makes Haiku go further. I assume this is due to differences in memory layout so that I need to move OpenBIOS to another location in RAM and anything else that clobbers the memory locations AIX assumes to be able to claim for itself.
Not having (low-level) access to an IBM Power System / System p / pSeries, the best source I could find was a JS20 device tree [1]. (Some older RS/6000 ones for comparison are available at [2].)
Interestingly the nvram is at 0x00000100ff800000 and pci at 0x08000000f8000000, i.e. above 32-bit address space. With sizeof(void*) == 4, drivers/macio.c:nvram is unable to save and access such an address.
ofmem cannot deal with 64-bit virtual/physical addresses either, since it uses 32-bit ucell in translation_t and range_t. The MMU translations on mac99 don't match Apple's OpenFirmware on my G5 size-wise.
It looks like the cell size is 32 bits, but both #address-cells and #size-cells is 2.
What's the plan for fixing this? Will we need an openbios-ppc64 after all, or is there a way to fix these issues in openbios-ppc using uint64_t and some assembler magic? The 970 init (arch/ppc/qemu/init.c:ppc64_patch_handlers) apparently does something clever for the interrupt handlers and seems able to use 64-bit ld/std instructions in start.S. That still wouldn't help with ucell based structs and C functions of course...
If you don't have to implement JS20, are there any other machines where all devices are located inside the first 4G?
There are more trees here: http://penguinppc.org/historical/dev-trees-html/trees-index.html