FS#39141 - [nvidia] 334.21-1-x86_64 breaks CUDA completely

Attached to Project: Arch Linux
Opened by Jeremy Maitin-Shepard (jbms) - Tuesday, 04 March 2014, 08:28 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Wednesday, 05 March 2014, 02:24 GMT
Task Type Bug Report
Category Packages: Extra
Status Closed
Assigned To Ionut Biru (wonder)
Sven-Hendrik Haase (Svenstaro)
Architecture x86_64
Severity High
Priority Normal
Reported Version
Due in Version Undecided
Due Date Undecided
Percent Complete 100%
Votes 3
Private No

Details

nvidia-334.21-1-x86_64 breaks CUDA completely:

~/cuda-samples/1_Utilities/deviceQuery $ sudo ./deviceQuery
./deviceQuery Starting...

CUDA Device Query (Runtime API) version (CUDART static linking)

modprobe: FATAL: Module nvidia-uvm not found.
cudaGetDeviceCount returned 30
-> unknown error
Result = FAIL
[1] 2422 exit 1 sudo ./deviceQuery

The problem may be related to a missing nvidia-uvm kernel module which may be required to use CUDA with this version of the driver.

Downgrading to nvidia-331.49-1-x86_64 works around the problem.
This task depends upon

Closed by  Sven-Hendrik Haase (Svenstaro)
Wednesday, 05 March 2014, 02:24 GMT
Reason for closing:  Fixed
Comment by Felix Yan (felixonmars) - Tuesday, 04 March 2014, 08:31 GMT
https://devtalk.nvidia.com/default/topic/699610/334-21-driver-returns-999-on-cuinit-cuda-/

Yeah, I've opened a thread on upstream forum, should be the same problem.
Comment by Felix Yan (felixonmars) - Tuesday, 04 March 2014, 08:51 GMT
I've attached a patch for the PKGBUILD, to build and install the nvidia-uvm module respectively. Please let me know if it works since I can't test it now, thanks.
   334.21.diff (0.9 KiB)
Comment by Jeremy Maitin-Shepard (jbms) - Tuesday, 04 March 2014, 09:23 GMT
Yes, this works. It does, however, require loading nvidia_uvm.ko (running a CUDA program as root does this automatically, but running a CUDA program as a regular user does not). It would be nice if it could be automatically loaded.
Comment by Felix Yan (felixonmars) - Tuesday, 04 March 2014, 09:31 GMT
I'd rather recommend you to add nvidia_uvm into /etc/modules-load.d yourself, since not all users needs cuda on boot.
Comment by Felix Yan (felixonmars) - Tuesday, 04 March 2014, 09:45 GMT
Update: Putting the module into /etc/modules-load.d seems not enough, I still have to run a cuda test program as root before a cuda program running as a regular user will work.
Comment by Olivier Langlois (lano1106) - Tuesday, 04 March 2014, 15:49 GMT
I have experienced the same problem with package nvidia-ck

adding nvidia-uvm did fix the problem. I just wanted to add to the discussion that OpenCL is affected in the same with uvm module ommission.

Loading...