Missed libraries in container compared to what GPU driver installs

[qiutongs]

I tried to install NVIDIA DRA 25.3.0-rc.4" on GKE. The CDI spec mounts individual library files (real file or symlink) into container while GKE mounts the entire GPU driver directory into container. I compared the library list between them and identified the missed libraries.

Real files (non-symlinks): libEGL.so.1.1.0, libGL.so.1.7.0, libGLESv1_CM.so.1.2.0, libGLESv2.so.2.1.0, libGLX.so.0, libGLdispatch.so.0, libOpenCL.so.1.0.0, libOpenGL.so.0, libnvidia-api.so.1, libnvidia-egl-xcb.so.1.0.0, libnvidia-egl-xlib.so.1.0.0

Symlinks for compile time (.so): libnvcuvid.so, libnvidia-allocator.so, libnvidia-cfg.so, libnvidia-encode.so, libnvidia-fbc.so, libnvidia-ml.so, libnvidia-nvvm.so, libnvidia-ptxjitcompiler.so, libnvidia-sandboxutils.so, libvdpau_nvidia.so

Symlinks for runtime (.so.X): None

From .so files, the plan should be enabling create-soname-symlinks hook for everyone.

Please look into the real file libraries.

[qiutongs]

@klueska

[klueska]

The second bullet should be addressed in the upcoming rc.5 release. Please see: #432

@elezar can you comment to the first bullet?

[elezar]

Real files (non-symlinks): libEGL.so.1.1.0, libGL.so.1.7.0, libGLESv1_CM.so.1.2.0, libGLESv2.so.2.1.0, libGLX.so.0, libGLdispatch.so.0, libOpenCL.so.1.0.0, libOpenGL.so.0, libnvidia-api.so.1, libnvidia-egl-xcb.so.1.0.0, libnvidia-egl-xlib.so.1.0.0

The logic to find these graphics / rendering libraries in the driver installation may need to be updated to allow them to be detected when using the GKE driver installation. Would you be able to provide information on where these are located in the driver root? For some of these they may be missed because these do not have the same suffix as the "core" driver libraries (e.g. libcuda.so.<DRIVER_VERSION>). I have created NVIDIA/nvidia-container-toolkit#1215 to track following up on this in the CDI spec generation logic.

Symlinks for runtime (.so.X): None

SONAME symlinks are not mounted into the container, but are created when ldconfig is run. With the changes merged in #432 we add explicit creation of the symlinks for driver libraries.