DRAAdminAccess: add example

Signed-off-by: Rita Zhang <[email protected]>

Author: ritazh

README.md

@@ -356,10 +356,72 @@ You can use the IDs of the GPUs as well as the GPU sharing settings set in
 these environment variables to verify that they were handed out in a way
 consistent with the semantics shown in the figure above.
 
+### Demo DRA Admin Access Feature
+This example driver includes support for the [DRA AdminAccess feature](https://kubernetes.io/docs/concepts/scheduling-eviction/dynamic-resource-allocation/#admin-access), which allows administrators to gain privileged access to devices already in use by other users. In this example, only workloads with admin access resource claims can view sensitive host hardware information through environment variables.
+
+When admin access is granted, the following host information is made available through environment variables:
+
+- `DRA_ADMIN_ACCESS=true` - Indicates admin access is enabled
+- `HOST_CPU_INFO` - CPU model and core count information from /proc/cpuinfo
+- `HOST_MEMORY_INFO` - Memory capacity and availability from /proc/meminfo  
+- `HOST_KERNEL_INFO` - Kernel version information from /proc/version
+- `HOST_SYSTEM_INFO` - Operating system and architecture (GOOS, GOARCH)
+- `HOST_NETWORK_INFO` - Available network interfaces from /proc/net/dev
+- `HOST_STORAGE_INFO` - Root filesystem and mount information from /proc/mounts
+
+#### CDI Integration
+
+When admin access is detected, the CDI (Container Device Interface) handler injects the host hardware environment variables into the container specification.
+
+#### Usage Example
+
+See `demo/gpu-test7.yaml` for a complete example. Key points:
+
+1. **Namespace**: Must have admin access label in order to create ResourceClaimTemplate and ResourceClaim with `adminAccess: true`. From kubernetes v1.34+, this label will be `resource.kubernetes.io/admin-access`.
+```yaml
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: gpu-test7
+  labels:
+    resource.k8s.io/admin-access: "true"
+```
+
+2. **Resource Claim Template**: Request must have `adminAccess: true`. [optional] `allocationMode: All` should be included if the workload needs to get information for all hardware devices. 
+```yaml
+spec:
+  spec:
+    devices:
+      requests:
+      - name: admin-gpu
+        exactly:
+          deviceClassName: gpu.example.com
+          allocationMode: All
+          adminAccess: true
+```
+
+3. **Container**: Will receive host hardware information via environment variables
+```bash
+echo "DRA Admin Access: $DRA_ADMIN_ACCESS"
+echo "Host CPU Info: $HOST_CPU_INFO"
+echo "Host Memory Info: $HOST_MEMORY_INFO"
+# ... additional host info variables
+```
+
+#### Testing
+
+To run this demo:
+```bash
+./demo/test-admin-access.sh
+```
+Note: These workloads can access devices already in use by other workloads. Only these workloads with admin access resource claims can view sensitive host hardware information through environment variables.
+
+### Clean Up
+
 Once you have verified everything is running correctly, delete all of the
 example apps:
 ```bash
-kubectl delete --wait=false --filename=demo/gpu-test{1,2,3,4,5}.yaml
+kubectl delete --wait=false --filename=demo/gpu-test{1,2,3,4,5,7}.yaml
 ```
 
 And wait for them to terminate:
@@ -374,6 +436,8 @@ gpu-test3   pod0   1/1     Terminating   0          31m
 gpu-test3   pod1   1/1     Terminating   0          31m
 gpu-test4   pod0   1/1     Terminating   0          31m
 gpu-test5   pod0   4/4     Terminating   0          31m
+gpu-test7   pod0   1/1     Terminating   0          31m
+gpu-test7   pod1   1/1     Terminating   0          31m
 ...
 ```
 

cmd/dra-example-kubeletplugin/cdi.go

@@ -36,18 +36,20 @@ const (
 )
 
 type CDIHandler struct {
-	cache *cdiapi.Cache
+	cache            *cdiapi.Cache
+	hostHardwareInfo *HostHardwareInfo
 }
 
-func NewCDIHandler(config *Config) (*CDIHandler, error) {
+func NewCDIHandler(config *Config, hostHardwareInfo *HostHardwareInfo) (*CDIHandler, error) {
 	cache, err := cdiapi.NewCache(
 		cdiapi.WithSpecDirs(config.flags.cdiRoot),
 	)
 	if err != nil {
 		return nil, fmt.Errorf("unable to create a new CDI cache: %w", err)
 	}
 	handler := &CDIHandler{
-		cache: cache,
+		cache:            cache,
+		hostHardwareInfo: hostHardwareInfo,
 	}
 
 	return handler, nil
@@ -96,9 +98,24 @@ func (cdi *CDIHandler) CreateClaimSpecFile(claimUID string, devices PreparedDevi
 			ContainerEdits: &cdispec.ContainerEdits{
 				Env: []string{
 					fmt.Sprintf("GPU_DEVICE_%s_RESOURCE_CLAIM=%s", device.DeviceName[4:], claimUID),
+					fmt.Sprintf("DRA_ADMIN_ACCESS=%t", device.AdminAccess),
 				},
 			},
 		}
+
+		// If this device has admin access, inject host hardware information
+		if device.AdminAccess {
+			hostEnvVars := []string{
+				fmt.Sprintf("HOST_CPU_INFO=%s", cdi.hostHardwareInfo.CPUInfo),
+				fmt.Sprintf("HOST_MEMORY_INFO=%s", cdi.hostHardwareInfo.MemInfo),
+				fmt.Sprintf("HOST_KERNEL_INFO=%s", cdi.hostHardwareInfo.KernelInfo),
+				fmt.Sprintf("HOST_SYSTEM_INFO=%s", cdi.hostHardwareInfo.SystemInfo),
+				fmt.Sprintf("HOST_NETWORK_INFO=%s", cdi.hostHardwareInfo.NetworkInfo),
+				fmt.Sprintf("HOST_STORAGE_INFO=%s", cdi.hostHardwareInfo.StorageInfo),
+			}
+			claimEdits.ContainerEdits.Env = append(claimEdits.ContainerEdits.Env, hostEnvVars...)
+		}
+
 		claimEdits.Append(device.ContainerEdits)
 
 		cdiDevice := cdispec.Device{

cmd/dra-example-kubeletplugin/discovery.go

@@ -17,9 +17,12 @@
 package main
 
 import (
+	"bufio"
 	"fmt"
 	"math/rand"
 	"os"
+	"runtime"
+	"strings"
 
 	resourceapi "k8s.io/api/resource/v1beta1"
 	"k8s.io/apimachinery/pkg/api/resource"
@@ -84,3 +87,181 @@ func hash(s string) int64 {
 	}
 	return h
 }
+
+// HostHardwareInfo represents hardware information available to admin access.
+type HostHardwareInfo struct {
+	CPUInfo     string
+	MemInfo     string
+	KernelInfo  string
+	SystemInfo  string
+	NetworkInfo string
+	StorageInfo string
+}
+
+// GetHostHardwareInfo gathers hardware information from the host system.
+func GetHostHardwareInfo() (*HostHardwareInfo, error) {
+	info := &HostHardwareInfo{}
+
+	// Get CPU information
+	if cpuInfo, err := readFile("/proc/cpuinfo"); err == nil {
+		info.CPUInfo = extractCPUModel(cpuInfo)
+	} else {
+		info.CPUInfo = fmt.Sprintf("CPU Info unavailable: %v", err)
+	}
+
+	// Get memory information
+	if memInfo, err := readFile("/proc/meminfo"); err == nil {
+		info.MemInfo = extractMemInfo(memInfo)
+	} else {
+		info.MemInfo = fmt.Sprintf("Memory Info unavailable: %v", err)
+	}
+
+	// Get kernel information
+	if kernelInfo, err := readFile("/proc/version"); err == nil {
+		info.KernelInfo = strings.TrimSpace(kernelInfo)
+	} else {
+		info.KernelInfo = fmt.Sprintf("Kernel Info unavailable: %v", err)
+	}
+
+	// Get system information (architecture, OS)
+	info.SystemInfo = fmt.Sprintf("GOOS: %s, GOARCH: %s", runtime.GOOS, runtime.GOARCH)
+
+	// Get network information
+	if netInfo, err := getNetworkInfo(); err == nil {
+		info.NetworkInfo = netInfo
+	} else {
+		info.NetworkInfo = fmt.Sprintf("Network Info unavailable: %v", err)
+	}
+
+	// Get storage information
+	if storageInfo, err := getStorageInfo(); err == nil {
+		info.StorageInfo = storageInfo
+	} else {
+		info.StorageInfo = fmt.Sprintf("Storage Info unavailable: %v", err)
+	}
+
+	return info, nil
+}
+
+func readFile(path string) (string, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return "", err
+	}
+	return string(data), nil
+}
+
+func extractCPUModel(cpuInfo string) string {
+	scanner := bufio.NewScanner(strings.NewReader(cpuInfo))
+	cpuCount := 0
+	var modelName string
+
+	for scanner.Scan() {
+		line := scanner.Text()
+		if strings.HasPrefix(line, "processor") {
+			cpuCount++
+		}
+		if strings.HasPrefix(line, "model name") {
+			parts := strings.Split(line, ":")
+			if len(parts) > 1 {
+				modelName = strings.TrimSpace(parts[1])
+			}
+		}
+	}
+
+	if modelName != "" {
+		return fmt.Sprintf("%d x %s", cpuCount, modelName)
+	}
+	return fmt.Sprintf("%d CPU cores", cpuCount)
+}
+
+func extractMemInfo(memInfo string) string {
+	scanner := bufio.NewScanner(strings.NewReader(memInfo))
+	var totalMem, availableMem string
+
+	for scanner.Scan() {
+		line := scanner.Text()
+		if strings.HasPrefix(line, "MemTotal:") {
+			parts := strings.Fields(line)
+			if len(parts) >= 2 {
+				totalMem = parts[1] + "kB"
+			}
+		}
+		if strings.HasPrefix(line, "MemAvailable:") {
+			parts := strings.Fields(line)
+			if len(parts) >= 2 {
+				availableMem = parts[1] + "kB"
+			}
+		}
+	}
+
+	if totalMem != "" && availableMem != "" {
+		return fmt.Sprintf("Total: %s, Available: %s", totalMem, availableMem)
+	}
+	return "Memory information parsing failed"
+}
+
+func getNetworkInfo() (string, error) {
+	// Read network interfaces
+	interfaces, err := readFile("/proc/net/dev")
+	if err != nil {
+		return "", err
+	}
+
+	scanner := bufio.NewScanner(strings.NewReader(interfaces))
+	var interfaceNames []string
+
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if strings.Contains(line, ":") && !strings.Contains(line, "Inter-|") && !strings.Contains(line, "face |") {
+			parts := strings.Split(line, ":")
+			if len(parts) > 0 {
+				ifName := strings.TrimSpace(parts[0])
+				if ifName != "lo" { // Skip loopback
+					interfaceNames = append(interfaceNames, ifName)
+				}
+			}
+		}
+	}
+
+	return fmt.Sprintf("Network Interfaces: %s", strings.Join(interfaceNames, ", ")), nil
+}
+
+func getStorageInfo() (string, error) {
+	// Read mounted filesystems
+	mounts, err := readFile("/proc/mounts")
+	if err != nil {
+		return "", err
+	}
+
+	scanner := bufio.NewScanner(strings.NewReader(mounts))
+	var rootFS string
+	var mountCount int
+
+	for scanner.Scan() {
+		line := scanner.Text()
+		fields := strings.Fields(line)
+		if len(fields) >= 3 {
+			mountPoint := fields[1]
+			fsType := fields[2]
+
+			if mountPoint == "/" {
+				rootFS = fmt.Sprintf("Root FS: %s (%s)", fields[0], fsType)
+			}
+
+			// Count real filesystems (not virtual ones)
+			if !strings.HasPrefix(fields[0], "/proc") &&
+				!strings.HasPrefix(fields[0], "/sys") &&
+				!strings.HasPrefix(fields[0], "/dev/pts") &&
+				fsType != "tmpfs" && fsType != "devtmpfs" &&
+				fsType != "cgroup" && fsType != "cgroup2" {
+				mountCount++
+			}
+		}
+	}
+
+	if rootFS != "" {
+		return fmt.Sprintf("%s, Total mounts: %d", rootFS, mountCount), nil
+	}
+	return fmt.Sprintf("Storage mounts: %d", mountCount), nil
+}

cmd/dra-example-kubeletplugin/driver.go

@@ -101,6 +101,9 @@ func (d *driver) Shutdown(logger klog.Logger) error {
 
 func (d *driver) PrepareResourceClaims(ctx context.Context, claims []*resourceapi.ResourceClaim) (map[types.UID]kubeletplugin.PrepareResult, error) {
 	klog.Infof("PrepareResourceClaims is called: number of claims: %d", len(claims))
+	for i, claim := range claims {
+		klog.Infof("Claim %d: UID=%s, Namespace=%s, Name=%s", i, claim.UID, claim.Namespace, claim.Name)
+	}
 	result := make(map[types.UID]kubeletplugin.PrepareResult)
 
 	for _, claim := range claims {
@@ -113,6 +116,7 @@ func (d *driver) PrepareResourceClaims(ctx context.Context, claims []*resourceap
 func (d *driver) prepareResourceClaim(_ context.Context, claim *resourceapi.ResourceClaim) kubeletplugin.PrepareResult {
 	preparedPBs, err := d.state.Prepare(claim)
 	if err != nil {
+		klog.Errorf("Error preparing devices for claim %v: %v", claim.UID, err)
 		return kubeletplugin.PrepareResult{
 			Err: fmt.Errorf("error preparing devices for claim %v: %w", claim.UID, err),
 		}