nextstation stuff

nextstation stuff

[John]

I finally got a very old next cube (rom 2.2) to net boot a kernel. =
  It has the video register bug, so it hung immediately after =
startup unless I either booted NEXTSTEP first, or entered

e 2004184
ea
^c

at the monitor by hand.  Adding the=20
> #define VIDEOOUTLIMIT	IO(0x02004184)
> *(ulong*)VIDEOOUTLIMIT =3D 0xea;
patch to the kernel fixed it for automatic startup.

My rom 3.3 systems still choke, and manually setting the pc still =
won't get them going.
It looks like the problem is due to the lack of a unixthread load =
command in the mach header.  I am going to hack l2 and see if I =
can build a kernel that loads properly on all NeXT roms.

Unfortunately, my cube has 16 1meg simms, only 8 megs visible to =
plan9.  8 megs is not enough to relink the kernel in...

Before I decided to wipe the scsi drive for swapping, I tried =
swapping to a file over u9fs.  This caused system traps during =
linking.  I can certainly see how it would fail (part of the u9fs =
client gets swapped out), but the docs say "swap to raw devices or =
remote servers".  Is that supposed to work?

After changing over to 9nextstationdisk and made a real swap =
partitiion, things worked fine, but linking was =
sssslllllooooowwwww.  I ordered 64 megs of ram to fix that =
problem...

I spent a while figring out how to get the proper plan9 bootp =
extension information to the cube.  NeXT's bootpd does not support =
the vendor specific extensions, so I had to compile something =
else.

The bootpd source included with the plan9 cdrom compiled with some =
warnings, but died shortly after starting, so I grabbed the latest =
2.4 bootpd source from the net.  After adding -posix, it compiled =
and linked with no warnings.   I couldn't see a way to enter raw =
text vendor data in the 2.4 bootptab (the vm=3D"text..." format is =
gone), so I wound up just making a gross hack in the source to =
create exactly the packets I want.  I will probably go back and =
make the 2.1 code included on the cd run properly, because that =
can at least be configured properly from the bootptab file.

The initial bootp packet from a net booting NeXT machine will have =
"NeXT" as the vendor id, and the return packet with booting info =
must also have that.  NeXTs also seem to require that the bootfile =
be tftpd from the same system sending the bootp packets.

The bootp info packet from the plan 9 kernel after choosing a tcp =
root will have "p9  " as the vendor id, and you want to return 'p9 =
 255.255.255.0 fileserver_ip 0.0.0.0 gateway_ip" as the vendor =
extension (assuming a type C network and no auth server).

Unfortunately, I had to remove the old cube from the NEXTSTEP =
netinfo database, or the bootpd daemons would both send confusing =
packets.  I don't think the NeXT bootp can't just be replaced, =
because it is intertwined with netinfo.


So, finally, I have things running smoothly the way I want. :-)

I was bummed to find that alef (and hence acme) doesn't exist far =
the '020.

Does cfs work with u9fs?  It wouldn't automatically start up, and =
when I started it manually, it failed as soon as I went into the =
cached directory.

Has anyone made a plan9 system with a >8bit color display?  I =
might try and bring the system up on a color nextstation.


John Carmack
Id Software

nextstation stuff

[dhog]

>Before I decided to wipe the scsi drive for swapping, I tried =
>swapping to a file over u9fs.  This caused system traps during =
>linking.  I can certainly see how it would fail (part of the u9fs =
>client gets swapped out), but the docs say "swap to raw devices or =
>remote servers".  Is that supposed to work?

In theory it should work fine.  The "u9fs client" is the mount driver, which
lives in the kernel and can't be swapped out.  Most of tcp/ip is in the kernel
also.  The thing to watch out for is that arpd is a user process, and thus it
can be swapped out.  This can be a problem if your server likes to time out
arp table entries too frequently.  I noticed that Solaris flushes its arp table
unconditionally (except for permanent entries) every few seconds (though
I think this is reconfigurable).

>Does cfs work with u9fs?  It wouldn't automatically start up, and =
>when I started it manually, it failed as soon as I went into the =
>cached directory.

The last I looked, u9fs didn't implement the "clwalk" operation (only
used by cfs), so unless this has been fixed, it won't work.