cudaMalloc（）上的Cuda未知错误（ErrNo：30）

Question

I have searched for the reason but no luck. It fails on such a simple program:

#include <iostream>

using namespace std;

int main() {
  int* n;
  cout << cudaMallocManaged(&n, 4 * sizeof(int)) << endl;
  return 0;
}

The return code is 30, unknown error. cudaMalloc also fails with same code.

This is my hardware:

$ lspci | grep NV
01:00.0 3D controller: NVIDIA Corporation GF117M [GeForce 610M/710M/820M / GT 620M/625M/630M/720M] (rev a1)

$ nvidia-smi
Sat Mar  7 14:02:04 2015       
+------------------------------------------------------+                       
| NVIDIA-SMI 331.113    Driver Version: 331.113        |                       
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  NVS 5200M           Off  | 0000:01:00.0     N/A |                  N/A |
| N/A   53C  N/A     N/A /  N/A |    279MiB /  1023MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Compute processes:                                               GPU Memory |
|  GPU       PID  Process name                                     Usage      |
|=============================================================================|
|    0            Not Supported                                               |
+-----------------------------------------------------------------------------+

I am using Ubuntu 14.10, with CUDA 6.0 from official repository(hopefully, if Ubuntu does not mess it up).

It is a Lenovo T430s labtop, the card is on Optimus so that might cause some problem. I have tested on another machine and the same code works.

Update 1

OK, nvidia_uvm is not loaded...

$ lsmod |grep nv

nvidia              10744914  65 
nvram                  14362  1 thinkpad_acpi
drm                   310919  6 i915,drm_kms_helper,nvidia

$ sudo modprobe nvidia_uvm
modprobe: ERROR: ../libkmod/libkmod-module.c:816 kmod_module_insert_module() could not find module by name='nvidia_331_updates_uvm'
modprobe: ERROR: could not insert 'nvidia_331_updates_uvm': Function not implemented

Update 2

OK, I reinstalled nvidia-331-updates-uvm and the module was loaded.

$ lsmod | grep nv
nvidia_uvm             34855  0 
nvidia              10744914  66 nvidia_uvm
nvram                  14362  1 thinkpad_acpi
drm                   310919  6 i915,drm_kms_helper,nvidia

However, the code still returns error 30.

Update 3

After some more testing (mainly tried running as root), now I get error 71: operation not supported. However, if I am just using cudaMalloc it succeeded. I will also check whether my device support unified memory addressing.

Update 4

OK, my card only supports SM 2.1, so it does not support Unified Memory.

Answer 1

AFAIK nvidia_uvm kernel module is required for CUDA to work.

You need to install package with that kernel module, eg nvidia-331-uvm and enable it's autoloading by installing nvidia-modprobe package:

sudo apt-get install nvidia-modprobe nvidia-331-uvm

If you don't want to reboot after installing nvidia-modprobe , you can try to run your program as root (eg sudo ./a.out ) — module should be loaded during run as root.