Multi-seat X crashes

[Erik W. <om...@vc...>]

I've got a system I'm building for a school in a small isolated Andean
town in Bolivia, and I'm leaving on Sept 4th to go install it.  Problem
is, it's currently crashing ;-(

The machine is a basic Athlon 3000+ box with 6 (six) Radeon 7000 PCI
video cards, and a whole mess of USB keyboards and mice.  It's running
Ubuntu 5.04 with xorg 6.8.2, and I've just tried xfree 4.3.0 with even
worse results so far.


Here are the tricks I've got in place currently:

1) At bootup, I have a script that starts each X server, one at a time,
just long enough to fully initialize the video card.  The xorg.conf.N
file is pretty standard except for the evdev (USB) setup, and
VGAAccess=false.  SingleCard is explicitly set to *false*.  GLX/DRI are
removed because they don't work with multiple cards as far as I can tell
so far (would love to solve that though).

2) Once this is done, gdm.conf has 6 lines as follows:

/usr/X11R6/bin/X -br -audit 0 -xf86config /etc/X11/xorg.conf.0s :0 vt7
-sharevts

The xorg.conf.Ns for this final configuration is identical except
SingleCard is set to *true*.


The initial startup without SingleCard seems to be necessary in order
for the VGA BIOS of the non-primary cards to initialize properly.  If I
neglect to perform that step and try to start the SingleCard servers
right after bootup, any server on any card *except* the primary (bootup
console) crashes somewhere in the middle of executing the BIOS (strace
shows vm86old() or whatever it is several thousand times, then it seizes).

Now, once this is all up and running, it seems to operate perfectly.  I
can log into each head separately and each looks as if it were its own
dedicated computer.  That's the goal of the setup, obviously, as a
single machine is more efficient in many respects (power, heat, cost,
and most importantly in this case: shipping volume).

The problem is that the machine will then seize up with no warning
anywhere from ~5 to ~30 minutes after starting up GDM.  This happens
whether I am using one of the heads or not, and whether I've even
*touched* the system or not.  It just as regularly happens when I've
just gone through the entire Ubuntu package list in Synaptic and am
about to start downloading (sigh!) as when I'm not even in the same room
and have started GDM remotely.

There are *no* messages of any kind, *anywhere*.


I'm attempting to set up a KGDB environment, but need to hunt down
either a null modem cable or the kgdb-over-ethernet patches for gdb
before I can begin.  Even then I have had no experience actually
debugging the kernel (have only done a bit of driver development without
gdb) so I don't know what kind of luck I will have.

Given the fact that this machine is basically worthless if it crashes
once it's installed in Bolivia and I've come back home, I *really* need
to try to solve this in the next 2 weeks.  If not we'll be taking the
gear but attempting to scrounge together as many single-headed machines
as we can to make use of all the LCDs we're bringing with us.  Lame, but
better than nothing. ;-(

I've tried messing with combinations of SingleCard, VGAAccess, NoInt10,
DPMS=false, -novtswitch, etc. to no avail.  I haven't exhaustively tried
NoAccel, and need to start taking more thorough notes of the
combinations and results, but otherwise I'm almost totally out of ideas.
 Re: DPMS it doesn't seem to have any relationship at all to shutdown or
wakeup of the displays.

The only time I got any hint of having made progress was when I added
-novtswitch to gdm.conf.  Running two heads the machine made it through
several near-complete deb-builds of xfree86 without crashing.  However,
once the two displays went to sleep, I tied to wake them.  The primary
head came back, but the secondary showed no signs of life at all.  Over
ssh I killed the server (-TERM), and the machine promptly crashed as
soon as gdm tried to start the replacement server.

As mentioned above I tried xfree86 4.3.0-dfsg.1, after rebuilding it
with the -sharevts and -novtswitch patches ported over from the xorg
packages.  With the same config files and commandlines, it won't even
bring up two displays without crashing during initialization.  I'll do a
few more tests tomorrow to see if I've screwed a config somewhere
compared to xorg, but I'm not very optimistic.

ATI's proprietary driver isn't my favorite thing, but I'd put up with it
if it solved the problem.  Only thing is, it doesn't support anything
older than the 9xxx series anyway...

I'm in #xorg on freenode, but haven't gotten any significant suggestions
so far, and it takes way too long to explain the issue ;-(  If there are
any other (active!) mailing lists I can send this to that might be
relevant, I'd appreciate any suggestions.

TIA,
   Omega
   aka Erik Walthinsen
   om...@vc...