[SOLVED] Nividia-driver Slackbuild CUDA initializaton error/glitch RFC

kingbeowulf · 07-13-2014, 01:25 PM

I sent this to the SBo mailing list, but figured that I would get a bit better coverage here.
----------------------------
I just realized a fix for a glitch with the new nvidia-driver and CUDA initialization. Since the slackbuild is already in "approved" and its not fatal unless you run CUDA/OpenCL programs, I'll fix this up on the next update. This issue popped up somewhere in the 331.xx drivers.

For those of you, as I do, who run BOINC, or another CUDA program, you may have wondered about "no usable GPUs found" message in the BOINC log. Turns out, some weird process in the way Nvidia does things now, since the introduction of nvidia-uvm.ko, prevents the creation of a required device node unless you first run your CUDA program as "root." This then creates the device node and all is then well as a regular user. You should see this in order to run CUDA:

Code:

$ ls -l /dev/nvidia-uvm 
crw-rw-rw- 1 root root 248, 0 Jul 13 01:45 /dev/nvidia-uvm

A quick test program, as well as possible fixes, are in the References below. The question now is, how best to handle this in the slackbuild?

1. put this in doinst.sh

Code:

mknod -m 666 /dev/nvidia-uvm c 249 0
chgrp video /dev/nvidia-uvm

(or maybe 660?)

2. udev rule?

Code:

KERNEL=="nvidia_uvm", RUN+="/usr/bin/bash -c '/usr/bin/mknod -m 666 /dev/nvidia-uvm c $(grep nvidia-uvm /proc/devices | cut -d \ -f 1) 0; /usr/bin/chgrp video /dev/nvidia-uvm'"

3. Compile nvidia-modprobe by default (optional now) so that X userland can make the node, or invoke via doinst.sh if not.

Any comments and advice is appreciated.

References:

irgunII · 07-13-2014, 06:09 PM

I hope you get this set up, as I've been running seti@home (BOINC) since 2000. I'd sure like to be able to really kick the RAC's up by using gpu's as they're a billion times faster (exagerating, but you know what I mean, heh).

Are you saying though that I *can* use the gpu(s) if I start my BOINC as root first, then close it and restart it as user?

kingbeowulf · 07-13-2014, 08:08 PM

Yes, running any CUDA program as root first will create the needed device nodes. Alternatively, run

Code:

OPTAPPS=yes ./nvidia-driver.SlackBuild

to build the optional software. Than, as root (or root terminal) run

Code:

nvidia-modprobe -c 0 -u

Alternatively, just run the "mknod" example above. You will need to do one of those after every reboot until I can figure out how make it automagical (and I have no clue how to properly write udev rules...).

kingbeowulf · 07-16-2014, 01:52 PM

Just an update for those that reguire CUDA/OpenCL to work for the new nvidia-340.24 scripts at Slackbuilds.org. Thanks to willysr for pushing it through his SBo branch.

Soon, the updated nvidia-driver SlackBuild script will post to compile and install the previously "optional applications." Please refer to the README to correctly set up 'nvidia-modprobe' in your 'rc.local'. This will make sure that /dev/nvidia-uvm is created after every reboot. This seems a bit cleaner than using udev rules and 'mknod'.

Enjoy.