Add support for MutableCSINodeAllocatableCount

The CSI list's all PCIe devices that are not of
type VIRTIO_BLOCK_DEVICE and subtracts them from
the theoretically maximum, so kubernetes can report
a correct dynamic max volume count that can be attached
for each node.

Signed-off-by: Niclas Schad <niclas.schad@stackit.cloud>

Author: nschad

pkg/csi/blockstorage/controllerserver.go

@@ -370,6 +370,10 @@ func (cs *controllerServer) ControllerPublishVolume(ctx context.Context, req *cs
 
 	_, err = cloud.AttachVolume(ctx, instanceID, volumeID)
 	if err != nil {
+		// Trigger's an immediate `NodeGetInfo` RPC call when MutableCSINodeAllocatableCount is enabled
+		if stackiterrors.IsTooManyDevicesError(err) {
+			return nil, status.Errorf(codes.ResourceExhausted, "[ControllerPublishVolume] Node can't accept any more volumes %v. All PCIe lanes are exhausted!", err)
+		}
 		klog.Errorf("Failed to AttachVolume: %v", err)
 		return nil, status.Errorf(codes.Internal, "[ControllerPublishVolume] Attach Volume failed with error %v", err)
 	}

pkg/csi/blockstorage/nodeserver.go

@@ -308,8 +308,16 @@ func (ns *nodeServer) NodeGetInfo(ctx context.Context, _ *csi.NodeGetInfoRequest
 	}
 
 	maxVolumesPerNode := DetermineMaxVolumesByFlavor(flavor)
-	// Subtract 1 for root disk and another for configDrive/spare
-	maxVolumesPerNode -= 2
+
+	// Subtract already mounted Volumes
+	emptyPCIeRootPorts, err := mount.CountNonVirtioBlockDevices()
+	if err != nil {
+		klog.Errorf("[NodeGetInfo] unable to retrieve PCIe root ports %v", err)
+		emptyPCIeRootPorts = 0
+	}
+
+	maxVolumesPerNode -= emptyPCIeRootPorts
+	klog.V(4).Infof("Determined %d PCIe ports occupied by non virtio block devices", emptyPCIeRootPorts)
 	klog.V(4).Infof("Determined node to support %d volumes", maxVolumesPerNode)
 
 	nodeInfo := &csi.NodeGetInfoResponse{

pkg/csi/blockstorage/utils.go

@@ -97,7 +97,7 @@ func DetermineMaxVolumesByFlavor(flavor string) int64 {
 		return 159
 	default:
 		// All other flavors can mount 28 volumes
-		return 25
+		return 28
 	}
 }
 

pkg/csi/blockstorage/utils_test.go

@@ -12,14 +12,14 @@ var _ = Describe("Util Test", func() {
 			maxVolumes := DetermineMaxVolumesByFlavor(flavor)
 			Expect(maxVolumes).To(Equal(int64(expectedMaxVolumes)))
 		},
-			Entry("Intel 3rd Gen", "c3i.2", 25),
-			Entry("Intel 2rd Gen", "c2i.2", 25),
-			Entry("Intel 1st Gen", "c1.2", 25),
-			Entry("AMD 1st Gen without overprovisioning", "s1a.8d", 25),
+			Entry("Intel 3rd Gen", "c3i.2", 28),
+			Entry("Intel 2rd Gen", "c2i.2", 28),
+			Entry("Intel 1st Gen", "c1.2", 28),
+			Entry("AMD 1st Gen without overprovisioning", "s1a.8d", 28),
 			Entry("AMD 2nd Gen without overprovisioning", "s2a.8d", 159),
 			Entry("Nvidia GPU", "n2.14d.g1", 10),
 			Entry("Nvidia GPU", "n2.56d.g4", 10),
-			Entry("ARM Gen1Link without CPU-overprovisioning ARM Gen1", "g1r.4d", 25),
+			Entry("ARM Gen1Link without CPU-overprovisioning ARM Gen1", "g1r.4d", 28),
 		)
 	})
 })

pkg/csi/util/mount/mount_darwin.go

@@ -17,3 +17,8 @@ func newDeviceStats(statfs *unix.Statfs_t) *DeviceStats {
 		UsedInodes:      int64(statfs.Files) - int64(statfs.Ffree),
 	}
 }
+
+func CountNonVirtioBlockDevices() (int64, error) {
+	// not implemented
+	return 0, nil
+}

pkg/csi/util/mount/mount_linux.go

@@ -4,6 +4,15 @@ package mount
 
 import "golang.org/x/sys/unix"
 
+var (
+	pciAddressRegex = regexp.MustCompile(`^[0-9a-fA-F]{4}:[0-9a-fA-F]{2}:[0-9a-fA-F]{2}\.[0-9a-fA-F]$`)
+)
+
+const (
+	RedhatVendor      = "0x1af4"
+	VirtioBlockDevice = "0x1042"
+)
+
 func newDeviceStats(statfs *unix.Statfs_t) *DeviceStats {
 	return &DeviceStats{
 		Block: false,
@@ -17,3 +26,74 @@ func newDeviceStats(statfs *unix.Statfs_t) *DeviceStats {
 		UsedInodes:      int64(statfs.Files) - int64(statfs.Ffree),
 	}
 }
+
+// CountNonVirtioBlockDevices returns the number of PCIe Root ports who
+// are currently occupied by anything else than an VIRTIO 1.0 Block Device
+// returns zero when something went wrong
+func CountNonVirtioBlockDevices() (int64, error) {
+	const pciPath = "/sys/bus/pci/devices"
+
+	// Get all PCI devices
+	devices, err := os.ReadDir(pciPath)
+	if err != nil {
+		return 0, fmt.Errorf("failed to read PCI bus: %w", err)
+	}
+
+	pcieSlotsOccupiedByNonBlockDevice := 0
+
+	for _, dev := range devices {
+		devPath := filepath.Join(pciPath, dev.Name())
+
+		// 1. Identify if it's a Root Port / Bridge
+		// We check the 'class' file. PCI Bridge class code starts with 0x0604
+		classBuf, err := os.ReadFile(filepath.Join(devPath, "class"))
+		if err != nil {
+			klog.Errorf("failed to read PCI device class %s : %v", devPath, err)
+			continue
+		}
+		class := strings.TrimSpace(string(classBuf))
+
+		// Class 0x060400 is a PCI-to-PCI bridge (standard for Root Ports)
+		if strings.HasPrefix(class, "0x0604") {
+			// 2. Check if the port has downstream devices
+			// If the bridge has children, they appear as subdirectories
+			// matching the PCI address format (e.g., 0000:01:00.0)
+			files, err2 := os.ReadDir(devPath)
+			if err2 != nil {
+				klog.Errorf("failed to read dir %s : %v", devPath, err2)
+			}
+			for _, file := range files {
+				// Ignore PCI bus directories such as pci001 pci002 and pci010
+				// Devices must follow <domain:bus:device.function> format
+				if pciAddressRegex.MatchString(file.Name()) {
+					isNonBlockDevice := IsNonBlockDevice(devPath, file)
+					if isNonBlockDevice {
+						pcieSlotsOccupiedByNonBlockDevice++
+					}
+					break
+				}
+			}
+		} else {
+			klog.V(4).Infof("skipping class %s: path: %s", class, devPath)
+		}
+	}
+
+	return int64(pcieSlotsOccupiedByNonBlockDevice), nil
+}
+
+func IsNonBlockDevice(devPath string, file os.DirEntry) bool {
+	var isNonBlockDevice bool
+	pciDevicePath := filepath.Join(devPath, file.Name())
+	vendorBuf, err := os.ReadFile(filepath.Join(pciDevicePath, "vendor"))
+	if err != nil {
+		klog.Errorf("failed to read PCI device vendor %s : %v", pciDevicePath, err)
+	}
+	deviceBuf, err := os.ReadFile(filepath.Join(pciDevicePath, "device"))
+	if err != nil {
+		klog.Errorf("failed to read PCI device file %s : %v", pciDevicePath, err)
+	}
+	if strings.TrimSpace(string(vendorBuf)) == RedhatVendor && strings.TrimSpace(string(deviceBuf)) != VirtioBlockDevice {
+		isNonBlockDevice = true
+	}
+	return isNonBlockDevice
+}

pkg/stackit/stackiterrors/errors.go

@@ -4,9 +4,10 @@ import (
 	"errors"
 	"fmt"
 	"net/http"
+	"strings"
 
 	oapiError "github.com/stackitcloud/stackit-sdk-go/core/oapierror"
-	wait "github.com/stackitcloud/stackit-sdk-go/services/iaas/v2api/wait"
+	"github.com/stackitcloud/stackit-sdk-go/services/iaas/v2api/wait"
 )
 
 var ErrNotFound = errors.New("failed to find object")
@@ -20,6 +21,17 @@ func IsNotFound(err error) bool {
 	return oAPIError.StatusCode == http.StatusNotFound
 }
 
+func IsTooManyDevicesError(err error) bool {
+	var oAPIError *oapiError.GenericOpenAPIError
+	if ok := errors.As(err, &oAPIError); !ok {
+		return false
+	}
+
+	// TODO: Improve this if possible
+	return oAPIError.StatusCode == http.StatusForbidden &&
+		strings.Contains(oAPIError.ErrorMessage, "maximum allowed number of disk devices")
+}
+
 func IgnoreNotFound(err error) error {
 	if IsNotFound(err) {
 		return nil