NixOS: Add support for CDI

[ereslibre]

Description of changes

This builds on top of #278969. It provides two main features to NixOS related to the Container Device Interface (CDI):

• virtualisation.containers.cdi.static attrset of JSONs. This can be used to provide static valid JSON files for CDI that will be placed under /etc/cdi.

• virtualisation.containers.cdi.dynamic.nvidia.enable boolean. When enabled, it will use the nvidia-container-toolkit to generate a CDI spec based on the current machine, as a systemd unit file. The generated JSON file will be patched to accomodate the NixOS host system, and will be placed under /run/cdi/nvidia-container-toolkit.json.

As per the CDI spec, CDI-compliant runtimes need to inspect /etc/cdi (static configuration) and /var/run/cdi (dynamic configuration); prioritizing what is under the dynamic configuration over the static configuration in case of clash.

So, this implementation is compliant with this specification.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandboxing enabled in nix.conf? (See Nix manual)
  • sandbox = relaxed
  • sandbox = true
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 24.05 Release Notes (or backporting 23.05 and 23.11 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

---

Add a 👍 reaction to pull requests you find important.

[ereslibre]

The last commit contains more cleanup than is needed, since we support Docker < 25, and thus, we'll still need the container runtime approach, given the CDI support was included in Docker 25.

When the final PR is submitted I'll get rid of that cleanup commit.

[aaronmondal]

I like how this cleanup makes it easier to keep track of all the nvidia dependencies.

One thing I'd like to add is that some sort of integration test would probably be a good idea. I'm not sure whether this is possible as they'd probably have to run in a NixOS VM though.

@SomeoneSerge Might have some comments on the new layout of the package.nix/default.nix files.

[SomeoneSerge]

Overall looks much simpler and cleaner than what we had

[ereslibre]

I'd not tackle pkgs/by-name in this instance and request an exception in this case.

[SomeoneSerge]

Thank you for the latest round of changes! I think the warnings look good. Ofborg is happy. I'm attaching a few more questions of which only the one about cdi.{dynamic,static} makes me hesitate about merging.

This is part of the to-be public interface so my assessment is that it's important to resolve this concern now. I don't know what the actual cost of introducing an interface in nixpkgs-unstable and potentially changing it before the release is.

You've done a lot of work on this PR, and I think we're almost there. Sorry to be annoying.

[SomeoneSerge]

I feel a bit disoriented, but I think all of my remaining comments are mere nitpicks. We also outlined a few TODOs for the immediate future (correct handling of closures, decisions about /usr/bin, libc priorities, deciding about whether to delete nvidia-container-toolkit-cdi-generate as a separate nixos module).

I think I'm convinced enough by the public-facing side of the PR. I also think this should just work for most of the basic use-cases (the overlay experience may need more looking into; and yes, I haven't yet tested this branch myself). In the (hopefully unlikely) case this PR does break somebody's workflow, I am online and respond to github pings, and I expect @ereslibre to be available as well, so we can always discuss and fix stuff.

...Is it time?

[ereslibre]

I am online and respond to github pings, and I expect @ereslibre to be available as well, so we can always discuss and fix stuff.

Likewise here, I'm happy to take ownership of this bit.

[aaronmondal]

So I just tested this on an instance with a T4 GPU on GCP. Seems to "just work" out of the box.

One thing I'm noticing is that the generated cdi file is not symlinked to /etc/cdi/nvidia.yaml as described in the nvidia docs. This doesn't seem to be an issue but higher level tooling (like the nvidia-operator for k8s) might have issues with this. I'd expect this to be solvable on the side of said tooling though, so this seems fine to me (and we can always change this later).

Fantastic work!

[ereslibre]

I just want to double check a couple things and answer all remaining conversations on this PR and mark it as ready to review, sorry for this last wait. Final stretch :)

Thanks a ton for your confirmation @aaronmondal :)

[ereslibre]

One thing I'm noticing is that the generated cdi file is not symlinked to /etc/cdi/nvidia.yaml as described in the nvidia docs. This doesn't seem to be an issue but higher level tooling (like the nvidia-operator for k8s) might have issues with this. I'd expect this to be solvable on the side of said tooling though, so this seems fine to me (and we can always change this later).

We'll have to check with the nvidia-operator and other components, but in principle both /etc/cdi and /var/run/cdi should be inspected by the runtimes (according to the CDI spec, prioritizing /var/run/cdi over /etc/cdi).

With this implementation, users can provide their static configuration, that is placed under /etc/cdi, and the "dynamic configuration", that is only implemented for nvidia (as of today) will place the dynamic configuration under /run/cdi (to which /var/run is linked to in NixOS.)

[ereslibre]

Everything seems to be fine, this is the only thing I cannot explain:

❯ podman run --rm --device=nvidia.com/gpu=all -it ubuntu:latest ldd /usr/bin/nvidia-smi
	linux-vdso.so.1 (0x00007fff0c181000)
	libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f94ec666000)
	libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f94ec57f000)
	libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f94ec578000)
	libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f94ec34f000)
	librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007f94ec34a000)
	/nix/store/cyrrf49i2hm1w7vn2j945ic3rrzgxbqs-glibc-2.38-44/lib/ld-linux-x86-64.so.2 => /lib64/ld-linux-x86-64.so.2 (0x00007f94ec671000)

What is exactly the meaning of => on this instance, and why is /lib64 there?:

/nix/store/cyrrf49i2hm1w7vn2j945ic3rrzgxbqs-glibc-2.38-44/lib/ld-linux-x86-64.so.2 => /lib64/ld-linux-x86-64.so.2 (0x00007f94ec671000)

Anyhow, I think we can move on, and fix any issues that might appear due to this change. As I have stated before, I take ownership of this logic and I'm happy to address those issues.

[nixos-discourse]

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/docker-rootless-with-nvidia-support/37069/21

[MathieuMoalic]

I just want to say thanks for the work. I was watching this issue closely as I would benefit a lot from this. Thanks!

[stuzenz]

Likewise, I have been looking into this issue and just found this pull request. Thank you for all the work on this issue - greatly appreciated.

I have linked this pull request to some reddit questions about GPU use with docker/podman on NixOS where others are trying to find their way through it.

Many thanks!

[ereslibre]

I have linked this pull request to some reddit questions about GPU use with docker/podman on NixOS where others are trying to find their way through it.

Thanks a lot for this @stuzenz. Just a reminder that this feature was updated in subsequent PR's and that virtualisation.containers.cdi.static and virtualisation.containers.cdi.dynamic are no longer a thing. hardware.nvidia-container-toolkit.enable should be used instead.