Mr.Coffee is a code name for JavaStation-1, Sun's first attempt at a Java Station. It's a sun4m architecture machine based on SPARCStation-4 design with some modifications. It uses microSPARC-II processor.
Mr.Coffee uses PC keyboard and mouse. NetBSD/sparc does not support them currently, though the work is in progress to write necessary glue code.
TCX framebuffer is supported, so you can currently run
Xsun(1) in "output only" mode ;-).
Mr.Coffee machines come with two two different OpenBoot versions. Older machines use OpenBoot 2.30 and are supported by NetBSD/sparc since summer 2000. However my understanding is that most machines come with OpenFirmware-style OpenBoot 3.x and NetBSD/sparc had no support for those.
The rest of this page is a collection of technical notes
about these OFW machines and what is necessary to get NetBSD/sparc
running on them. This is a work in progress. I currently have NetBSD
running on an OFW machine, but there are still few issues to sort out
and code is not yet integrated into -current.
at this point type any key and you'll be greated with ok prompt.ttya initialized Release 3.11.1 Version 1 created 1998/03/18 20:10 Probing Memory Bank #0 32 Megabytes Probing Memory Bank #1 32 Megabytes Ethernet address 8:0:20:87:b4:c4 JavaStation OpenBoot 3.11.1, 64 MB memory installed, Serial #8893636. Ethernet address 8:0:20:87:b4:c4, Host ID: 8087b4c4. Type any key to interrupt automatic startup 6
OpenBoot 3.x is an implementation of IEEE 1275 OpenFirmware.
Hint for those familiar with OBP2: use dev instead of cd and use .properties instead of .attributes to navigate device tree.
OpenFirmware does not provide "romvec" interface of older boot proms, instead client is passed an OF entry point named client interface handler (in %o3). Fortunately, OpenFirmware support was already imported into NetBSD/sparc from sparc64 port, and though it was not really tested before it turned out to be fairly complete and well integrated.
Another difference is that client program (boot.net, secondary boot in our case) must be a real a.out or ELF executable with accurate segment sizes! Bogus header we snap on in stand/boot will cause a crash (see port-sparc/12548).
The problem with OFW in these machines is that it's screwed in few important places. I'd rather not share with anyone those few strong words I have for its engineers.
To quote SPARC supplement:
When a client program begins execution, an Open Firmware implementations use of any virtual address space outside of the ranges 0xffd0.0000-0xffef.ffff and 0xfe00.0000-0xfeff.ffff shall have ceased, except for the virtual address space and associated memory that is allocated for the client programs code and data, as specified in the client program header.
The problem with Mr.Coffee PROM is that it occupies address space f000.0000-f007.ffff outright intruding into the address space we want to use for kernel!
NetBSD/sparc KERNBASE is 0xf0000000 so we have an address conflict. We can't move OFW, so we have to move kernel. In principle it should be possible to move kernel to f008.0000 and stay within F0 region, but it will also require a re-layout of the kernel and some more work to get initial mapping of the kernel into high virtual addresses. Since that's too radical for me, I moved kernel to region EF, but it turned out that some code in kernel had alignment assumptions about NUREG and NKREG, so I moved kernel to region E8. [NB: boot.net has to be recompiled accordingly with include/loadfile_machdep.h matching the new KERNBASE]
I'm not sure if this is a good thing to propagate to the main tree, so for -current integration it might be neccessary to redo this part. Alternatively JavaStation will require a custom kernel. We'll see.
NetBSD gets a lot of interesting info from CPU node.
On sun4m the sbus is sitting under iommu. Unfortunately device tree does not have a node describing iommu and the sbus node is right under the root. Thus we have to insinuate the iommu node if we see sbus node under root on sun4m.
The other place where we have to work around this omission is in
bootpath_build to make bootpath match the attachment
hierarchy.
NetBSD uses model name to detect mk48t08 chip.
le node (ehternet) has correct old style
intr property of 0x26 0x0, but new
style interrupts property contains zero. I changed
sbus_get_intr to fall back to intr property
if interrupts exists but is zero (Sbus interrupt levels
are 1 to 7).
At first I tried to sprinkle a lot of workarounds in the kernel code, but then I read FCode manuals and came up with a better solution. Now boot.net patches the PROM to correct most of the glitches so that kernel can see what it expects to see. The only problem that can't be corrected this way is the lack of iommu node, but the workaround for this problem is quite clean and non-intrusive.
wscons is an undocumented mess (see kern/9770 :).
It might be easier to write an emulation layer on top of i8042 to make it look like sun keyboard/mouse. But in the long run (i.e. for Krups) wscons seems the way to go.
more to follow...