nvidia-graphics-drivers-* should recommend nvidia-modprobe

[Michael Schaller]

nvidia-modprobe allows non-root users to load NVIDIA kernel modules and it creates the respective device entries. This is important for CUDA users as they would otherwise have to run a workaround script to load the kernel modules and create the device entries. This is documented here:
http://developer.download.nvidia.com/compute/cuda/6_5/rel/docs/CUDA_Getting_Started_Linux.pdf -> 4.8 Verification

Because of this newer nvidia-graphics-drivers packages (331 and newer) should at least recommend nvidia-modprobe. Furthermore the nvidia-modprobe is not available from xorg-edgers. Utopic includes an nvidia-modprobe package though.

[Joakim Fogelström Hardingz (birdstream)]

So, just to be clear, is this why i can't get CUDA to work with the 343.22 drivers from xorg-edgers PPA?

[Michael Schaller (misch-9)]

The issue I'm referring to here is present since 331.67. The default NVIDIA driver in Trusty is 331.38.
This issue is characterized by CUDA applications failing during initialization with an error similar to this one:
CUDA Runtime API error 30: unknown error

For newer NVIDIA driver you need to install the nvidia-###-uvm package (### == major NVIDIA driver version).
After that you need to ensure after every reboot that the nvidia-uvm kernel module is loaded and /dev/nvidia-uvm is present.
If you have nvidia-modprobe installed then it takes care of these two conditions if you launch a CUDA application.
A CUDA application would also react similar to nvidia-modprobe if you would run it with root rights.
As a last resort you can run the script I've mentioned in the bug description to load the nvidia-uvm kernel module and create the /dev/nvidia-uvm character device after a reboot.

Of all solutions nvidia-modprobe is the most convenient one as everything works out of the box if needed but it might be considered a security concern as it is a setuid binary. You can install the nvidia-modprobe package from Utopic if you want to give it a try.

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[YAFU (yafu)]

I don't know is this is the same problem, but this is happening in Ubuntu 14.10 Utopic beta, nVidia 331.89:
https://bugs.launchpad.net/ubuntu/+source/nvidia-graphics-drivers-331-updates/+bug/1375841

I hope this will be solved before 14.10 be released.

[YAFU (yafu)]

Well, I installed the "nvidia-modprobe" package and that seems to solve the problem.
$ lsmod | grep nvi
nvidia_uvm 34855 0
nvidia 10728519 41 nvidia_uvm
drm 310919 3 nvidia

Definitely when the nvidia drive is installed then "nvidia-modprobe" should be automatically installed as a dependency.

[Graham Inggs (ginggs)]

nvidia-modprobe is now available in trusty-backports

[Romaine Rose-Cameron (r0ma1ne)]

I'm using Linux MInt 17.1 which uses Trusty and 'nvidia-modprobe' isnt a package i can install, it doesnt come with nvidia-331 or nvidia-340 (343 and 346 dont support my card so i havent bothered to test) as i'm sure it did along with a few other programs that havent been availible in the last few days like nvidia-smi and nvidia-xconfig, even that dodgy persistenced thing. I'll make another, more detailed bug rebort on that as it's pretty much off topic. I'm just letting you know i cant get nvidia-modprobe using any of the methods suggested.

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Graham Inggs (ginggs)]

@r0ma1ne: Check the Mint documentation on how to enable the backports repository.
I do not know if they keep in sync with the Ubuntu backports though. If they don't, you should be able to download and install the .deb from here:
https://launchpad.net/ubuntu/+source/nvidia-modprobe/

[YAFU (yafu)]

Putting the package in backports is just a workaround. nvidia-modprobe should be automatically installed when you install the nvidia driver. Does it happening currently with new versions of Ubuntu?
Otherwise, this is still a big problem for users:
http://blenderartists.org/forum/showthread.php?355719-No-Cuda-option

@r0ma1ne, you read that thread in latest posts (user have mint 17)

[Graham Inggs (ginggs)]

@yafu:
> Does it happening currently with new versions of Ubuntu?

No it doesn't. That is why this bug is still open.

[Graham Inggs (ginggs)]

MIR (main inclusion request) filed: LP: #1421209

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Márcio Jorge PimentaParola (parola-marcio)]

i hope that is resololved, but the help from you are good...
thanks

[YAFU (yafu)]

@Are you Ubuntu package maintainer? Why you are assigned to this issue?

[YAFU (yafu)]

Sorry, my previous post was for Márcio Jorge PimentaParola

[Alberto Milone (albertomilone)]

I'm the maintainer. Reassigning the bug back to me

[VF (fviktor)]

I faced this problem on Ubuntu 15.04. The missing device nodes / kernel module can manifest in all kind of failures and error messages, depending on which application you try to use. Due to the misleading error messages it took hours to finally track down the issue and find this ticket with the correct explanation and solution. Installing the nvidia-modprobe package solved the problem instantly. Having this package recommended with the nvidia driver and/or a dependency of the nvidia-cuda-toolkit package could have saved this wasted effort.

[libuntu (libuntu)]

the bug is solved installing nvidia-modprobe!

[Olli Niemi (olliniem)]

Ubuntu 15.10 status: Installing only nvidia-352 and nvidia-cuda-toolkit packages forces Blender 2.76 to be started as root for the CUDA option to be visible in Blender menus (no errors reported in Blender). Installing also nvidia-modprobe fixes this problem: Blender can be started as normal user.

Starting Blender as root naturally saves the project files as root:root as well, possibly also other files, including user preferences. The nvidia-modprobe package installation seems more like a requirement in this use case.

[Alberto Milone (albertomilone)]

Just an update: while we are not going to use nvidia-modprobe, we are going to use a different tool I am working on, that will rely on udev and will accomplish the same as nvidia-modprobe.

[ragtag (ragtag-ragtag)]

With the nvidia-352 driver in 14.04, there no longer seems to be an uvm package, and the nvidia-modprobe package is at version 340.21.

In the past I've added "nvidia-uvm" to /etc/modules, and the following from here [ http://developer.download.nvidia.com/compute/cuda/6_5/rel/docs/CUDA_Getting_Started_Linux.pdf ] to /etc/rc.local to get CUDA working in Blender:

/sbin/modprobe nvidia-uvm
if [ "$?" -eq 0 ]; then
 # Find out the major device number used by the nvidia-uvm driver
 D=`grep nvidia-uvm /proc/devices | awk '{print $1}'`
 mknod -m 666 /dev/nvidia-uvm c $D 0
else
 exit 1
fi

With nvidia-modprobe installed, "lsmod | grep nvi" does list nvidia-uvm, but there is not /dev/nvidia-uvm device created. Running the above script snippet creates /dev/nvidia-uvm, but CUDA is still not working in Blender.

It would be great if CUDA worked out of the box, either when you install the nvidia-352 driver, or at least when you install libcuda1-352.

[ragtag (ragtag-ragtag)]

It seems I was using a faulty test to see if CUDA was working. Apparently the version of Blender that comes with 14.04, is not compiled with CUDA support. Running the latest version of Blender from blender.org got working CUDA support without doing anything beyond installing nvidia-modprob, nvidia-352 and libcuda1-352.

I second the original ticket, that nvidia-modprobe should probably be added to the dependencies, but it seems like you're working on a different solution there, so this probably wont be a problem in the near future.

[Michael Schaller (misch-9)]

Alberto, can you give an update on the udev approach to fix this issue?
Furthermore if there is a specific bug we could follow please link it here.

[Alberto Milone (albertomilone)]

@Michael: the idea is to have all the modules loaded and all the devices available on boot, even when starting a non graphical environment. This way we do not need nvidia-modprobe at all.

[Michael Schaller (misch-9)]

@Alberto: I know the plan and I fully agree with it. I was asking for a status update and if there is a bug or something else that we could follow/monitor instead of asking you for an update. ;-)

[Alberto Milone (albertomilone)]

@Michael: no, there is no bug report for it yet. I am finally working on it, so I expect to deliver a fix for 16.04 soon.

[Michael Schaller (misch-9)]

@Alberto: Thanks. Will this also be backported to the long-term supported 340 and 352 driver series?

[Alberto Milone (albertomilone)]

Yes, a backport is certainly possible. I'll create a new bug report for it.

[Alberto Milone (albertomilone)]

here is the new bug report LP: #1540896

[Michael Schaller (misch-9)]

Awesome. Thanks, Alberto.