NVIDIA GPUs: add documentation on how to use (for users), and pointers to ebuilds (for devs)

[t-lo]

No description provided.

[pothos]

Sayan has a draft ready, will be filed as PR soon.

One possible improvement I noticed is that we could switch the nvidia.service unit to be Type=oneshot and RemainAfterExit=true and set Before=containerd.service to prevent race conditions with containers that want to use the GPU.

[sayanchowdhury]

PR flatcar-archive/flatcar-docs#307

[shsamkit]

Is there any work going on related to chaning nvidia.service unit to oneshot type? It would really help with sequencing the service containers.
ATM waitFor/required doesn't really work with nvidia.service because it doesn't wait for the service to be actually completed.

[pothos]

I suggested it but didn't start the work, feel free to file a PR with the above suggestions.

[shsamkit]

@pothos Looks like there may not be a need for changing the service-type afterall. I was able to change the service-type using the drop-in-units

If you want to try it out, this is the butane config I used

---
variant: flatcar
version: 1.0.0
storage:
  files:
    - path: /etc/systemd/system/nvidia.service.d/service-type-simple.conf
      contents:
        inline: |
          [Service]
          Type=oneshot

[pothos]

I think you should also have RemainAfterExit=true or it will run multiple times, and I think we should just do this by default, so if you want, you can file a PR where you set this and Before=containerd.service.
The file is here: https://github.com/flatcar/scripts/blob/main/sdk_container/src/third_party/coreos-overlay/x11-drivers/nvidia-drivers/files/units/nvidia.service

[shsamkit]

I have opened an issue and working on a PR for it.
About Before=containerd.service, a quick look at the installation script I assume the installation runs installation with containers. Not sure if that would work? I may be wrong tho

[jepio]

The installation uses systemd-nspawn and the idea was that containers may want to use GPUs so that's why ordering the nvidia.service before containerd.service.

[sayanchowdhury]

We have added documentation for the users, and with the migration to the base image - the ebuild is simplified so closing the issue.