nvidia-drivers do not create uvm device, even when uvm flag is set. So we must execute the script of http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#runfile-verifications every time we want to use cuda. Instead we could create a udev rule to create the device when uvm module is loaded. This udev rule could create the device itself or run the proposed script by nvidia. My proposed udev rule: ACTION=="add", SUBSYSTEM=="module", DEVPATH=="/module/nvidia_uvm", PROGRAM="/bin/grep nvidia-uvm /proc/devices", ENV{dev_num}="%c{1}", RUN+="/bin/mknod -m 666 /dev/nvidia-uvm c $env{dev_num} 0" Please include that udev rule when uvm flag is set and if possible propose this udev rule upstream. Reproducible: Always
Works for me and you didn't include any information about your system so I guess you need to read the bug reporting guidelines before you continue.
Can you please elaborate more? Does your system creates a uvm device? I removed 'sometimes' from the title, it's misleading.
(In reply to Plero H from comment #0) > My proposed udev rule: > > ACTION=="add", SUBSYSTEM=="module", DEVPATH=="/module/nvidia_uvm", > PROGRAM="/bin/grep nvidia-uvm /proc/devices", ENV{dev_num}="%c{1}", > RUN+="/bin/mknod -m 666 /dev/nvidia-uvm c $env{dev_num} 0" Are you sure that should be read/write for all users?
*** This bug has been marked as a duplicate of bug 506696 ***