nvidia crashing info

[Open S. S. <in...@op...>]

Here's some more info on the nvidia crashing problem so many people
have experienced in one form or another:
Our old Debian software with XFree86 used the nvidia driver and was
very stable.  Now with Ubuntu 5.10 and Xorg we are having a lot more
problems.  Using the nv open source driver instead of the binary
nvidia, things are very stable.  Restarting the primary x server with
ctrl-alt-bs repeatedly can crash things, but we have gdm set to not
restart the x servers.  However, the nvidia binary driver is crashing
about 1/20 times during logout.  A gdm restart gets things back, but
the crashing is unacceptable.  We have noticed interrupt conflicts
that made things even worse, but even with the nvidia cards on a
different interrupt from the ethernet card we still see crashes.  It
would be interesting to see if we could get each nvidia card to have
its own interrupt if that would fix things.  Of course interrupt
conflicts mean the nvidia driver is not pci spec compliant, but what
will it take to convince nvidia to take a serious look at this issue?

We just turned off a third station on one of our boxes to see if it
was more stable, and it still crashed using the nvidia driver, but
even more interestingly, the gdm.log had this message:

NVIDIA: could not open the device file /dev/nvidia2 (Input/output error).
(EE) NVIDIA(0): Failed to initialize the NVIDIA graphics device!

and /dev/nvidia2 shouldn't even be used!  Why is the primary card
trying to open /dev/nvidia2???  We were only running 2 xservers which
should be using /dev/nvidia0 and /dev/nvidia1.  /dev/nvidia2 was
probably created during our initial probe (we probe all cards even if
we don't have enough keyboards and mice in case we want to start
additional servers later) but it should not be referenced at all.

This kind of error tells us something about how the cards are
conflicting with other cards in the system, but only nvidia can make
sense of it since the drivers are closed.  We are fed up, and have 5
ATI cards getting delivered tomorrow for experimentation: two 9250
pci, 1 9250 agp, and an x300 pci-e.  Supposedly you can run two
independent x servers with one dual head card using the binary ATI
driver, which would be a cost savings.  The multiXnest approach gets
you that even with nvidia, but it is slow. Everyone says ATI is not
linux friendly, but maybe once we figure out all the quirks it will
not crash like nvidia.  If Matrox was a little more Linux friendly we
might try them, and their new pci-e 1x cards would be very useful in
modern systems to allow more cards, but Matrox prices are just insane.

Thanks,
Mike Pardee
Open Sense Solutions LLC
http://opensensesolutions.com