Device: Improvements to GPU CDI mode & Ubuntu Core validation (#14348)

#14347 (comment)

lxd/device/cdi/spec.go

@@ -77,6 +77,12 @@ func generateNvidiaSpec(s *state.State, cdiID ID, inst instance.Instance) (*spec
 		//
 		rootPath = os.Getenv("SNAP") + "/gpu-2404-2"
 		devRootPath = "/"
+
+		// Let's ensure that user did:
+		// snap connect mesa-2404:kernel-gpu-2404 pc-kernel
+		if !shared.PathExists("/snap/lxd/current/gpu-2404-2/usr/bin/nvidia-smi") {
+			return nil, fmt.Errorf("Failed to find nvidia-smi tool. Please ensure that pc-kernel snap is connected to mesa-2404.")
+		}
 	} else if shared.InSnap() {
 		rootPath = "/var/lib/snapd/hostfs"
 		devRootPath = rootPath

lxd/device/gpu_physical.go

@@ -75,6 +75,14 @@ func (d *gpuPhysical) validateConfig(instConf instance.ConfigReader) error {
 				return fmt.Errorf(`Cannot use %q when "id" is set`, field)
 			}
 		}
+
+		// If user requests CDI in conjunction with nvidia.runtime=true we should forbid that.
+		if shared.IsTrue(instConf.ExpandedConfig()["nvidia.runtime"]) {
+			_, err := cdi.ToCDI(d.config["id"])
+			if err == nil {
+				return fmt.Errorf("CDI mode is incompatible with nvidia.runtime=true")
+			}
+		}
 	}
 
 	return nil