Some containers worked well with legacy mode but no longer work since the switch to CDI mode in version 1.18.0 #1456
Description
As the title says, some Docker containers that previously worked in legacy mode have stopped working in cdi mode since the default switch to version 1.18.0. I've done a lot of testing and still can't figure out why it's not working. So, I'd like to get an expert's opinion on the subject to try to guide me a little in solving this problem.
Firstly, here are some issues demonstrating that it no longer works since the upgrade to version 1.18.0:
- [BUG] Nvidia HW Transcoding Issues linuxserver/docker-plex#432 (comment)
- [BUG] Nvidia container toolkit 1.18 breaks Zink support linuxserver/docker-baseimage-selkies#112
It may be a configuration issue, something that needs to be changed in the Docker container, something I haven't understood. For now, I'm trying to figure out why it's not working like it used to.
For example, in the case of the plex container, plex can no longer find the cuda library. However, the ldd cache (ld.so.cache) does contain a reference to the library.
$ ldconfig -p | grep "cuda"
libcuda.so.1 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/nvidia/current/libcuda.so.1
libcuda.so (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/nvidia/current/libcuda.so
After doing some research, I get the impression that Plex performs a dlopen of the library. So I tried doing a dlopen myself in the same container, and my binary was able to find the cuda library. The only way I can get Plex to find the library is to add the path to the cuda library to the LD_LIBRARY_PATH environment variable.
And there appear to be errors on other containers as the second example provides.
Is it possible to know if this behavior is normal since the update? What needs to be done to make everything work properly?
I've spent quite a bit of time searching, but any help would be welcome, thank you.