Add vfio mode to generate CDI specs for NVIDIA passthrough GPUs

[elezar]

This change adds a vfio mode to the nvcdi API (and nvidia-ctk generate command).

---
cdiVersion: 0.5.0
kind: nvidia.com/pgpu
devices:
    - name: "0"
      containerEdits:
        deviceNodes:
            - path: /dev/vfio/5
              hostPath: /dev/vfio/5
containerEdits:
    env:
        - NVIDIA_VISIBLE_DEVICES=void
    deviceNodes:
        - path: /dev/vfio/vfio
          hostPath: /dev/vfio/vfio

[elezar]

@cdesiniotis if this is still valid, please rebase or close.

[elezar]

@varunrsekar would this be interesting for what you're looking at in NVIDIA/k8s-dra-driver-gpu#183?

[varunrsekar]

This is missing a device node for /dev/vfio/vfio:

&cdi.ContainerEdits{
    ContainerEdits: &specs.ContainerEdits{
        DeviceNodes: []*specs.DeviceNode{
	    &specs.DeviceNode{
	        Path: "/dev/vfio/vfio",
	    },
	},
    },
}

[varunrsekar]

@varunrsekar would this be interesting for what you're looking at in NVIDIA/k8s-dra-driver#183?

yeah this will help!

[cdesiniotis]

@varunrsekar I am no longer working on this -- feel free to take ownership of this PR if you need it.

[coveralls]

Pull Request Test Coverage Report for Build 16119484166

Details

• 69 of 81 (85.19%) changed or added relevant lines in 4 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.3%) to 34.398%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
pkg/nvcdi/mode.go	0	1	0.0%
pkg/nvcdi/lib.go	6	8	75.0%
pkg/nvcdi/lib-vfio.go	59	68	86.76%

Totals
Change from base Build 16115644541:	0.3%
Covered Lines:	4565
Relevant Lines:	13271

---

💛 - Coveralls

[ArangoGutierrez]

2 non blocking nits

[Copilot]

Pull Request Overview

This PR adds a new VFIO mode to the nvcdi API and the nvidia-ctk generate command to support NVIDIA GPU passthrough via VFIO.

• Introduces a ModeVfio constant and integrates it into the mode resolver.
• Adds WithPCILib option and dependency on nvpci for querying PCI devices.
• Implements vfiolib to generate VFIO-based CDI specs, including unit tests.

Reviewed Changes

Copilot reviewed 6 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
pkg/nvcdi/options.go	Import nvpci and add WithPCILib option
pkg/nvcdi/mode.go	Register ModeVfio in the list of supported modes
pkg/nvcdi/lib.go	Wire up nvpcilib in New for ModeVfio
pkg/nvcdi/lib-vfio.go	Implement vfiolib and VFIO device spec generation
pkg/nvcdi/lib-vfio_test.go	Add unit tests for the new VFIO mode
go.mod	Bump github.com/NVIDIA/go-nvlib for PCI library support

Comments suppressed due to low confidence (3)

pkg/nvcdi/lib-vfio.go:93

• [nitpick] Consider using %q instead of %v to quote the invalid ID in the error message for clarity ("invalid channel ID %q: %w").

			return nil, fmt.Errorf("invalid channel ID %v: %w", id, err)

pkg/nvcdi/lib-vfio_test.go:28

• Tests cover positive VFIO scenarios but don’t verify that non-vfio-pci devices are filtered out or that error paths (e.g., invalid ID parsing, PCI library failures) behave as expected. Consider adding cases for those.

func TestModeVfio(t *testing.T) {

pkg/nvcdi/lib-vfio.go:33

• [nitpick] The field group in vfioDevice is ambiguous — consider renaming it to iommuGroup to make its purpose clearer.

	group   int

[Copilot]

The comment should use the uppercase acronym PCI for consistency with the function name (WithPCILib).

Suggested change

	// WithPCILib sets the pci library to be used for CDI spec generation.
	// WithPCILib sets the PCI library to be used for CDI spec generation.

[Copilot]

When path starts with a slash, filepath.Join(l.devRoot, path) will ignore l.devRoot. You may need to strip the leading slash or construct the host path as filepath.Join(l.devRoot, path[1:]).

Suggested change

	HostPath: filepath.Join(l.devRoot, path),
	HostPath: filepath.Join(l.devRoot, path[1:]),

[Copilot]

Same issue here: joining an absolute path to l.devRoot will ignore the root override. Consider constructing this path without the leading slash or using path.Join on relative segments.

Suggested change

	HostPath: filepath.Join(l.devRoot, "/dev/vfio/vfio"),
	HostPath: filepath.Join(l.devRoot, "dev/vfio/vfio"),