fix(hostagent): bind mlx5_core to DPU PFs before configuring VFs

internal/provisioning/hostagent/networkmanager/network_manager.go

@@ -185,6 +185,28 @@ func (nm *NetworkManager) processNetworkRequest(nr NetworkRequest) error {
 		return nil
 	}
 	operations := []networkOperation{
+		{
+			name: "EnsureDriverBoundP0",
+			f: func(nr NetworkRequest) error {
+				return hostutil.NewPCIHelper(nr.PCIAddress).PF(0).BindDriver("mlx5_core")
+			},
+		},
+		{
+			name: "EnsureDriverBoundP1",
+			f: func(nr NetworkRequest) error {
+				pf := hostutil.NewPCIHelper(nr.PCIAddress).PF(1)
+				isDPU, err := pf.IsDPU()
+				if err != nil {
+					if os.IsNotExist(err) {
+						return nil
+					}
+					return fmt.Errorf("failed to check if device is DPU: %w", err)
+				} else if !isDPU {
+					return nil
+				}
+				return pf.BindDriver("mlx5_core")
+			},
+		},
 		{
 			name: "CreateP0VF",
 			f: func(nr NetworkRequest) error {

internal/provisioning/hostagent/util/pci.go

@@ -144,6 +144,38 @@ func (h *PCIHelper) SetNumOfVFs(num int) error {
 	return os.WriteFile(numvfsPath, []byte(fmt.Sprintf("%d", num)), 0644)
 }
 
+// IsDriverBound returns true if a driver is currently bound to the device.
+func (h *PCIHelper) IsDriverBound() (bool, error) {
+	if _, err := os.Lstat(filepath.Join(h.Path(), "driver")); err != nil {
+		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, fmt.Errorf("failed to stat driver symlink: %w", err)
+	}
+	return true, nil
+}
+
+// BindDriver writes the device's BDF to /sys/bus/pci/drivers/<driverName>/bind,
+// causing the kernel to attempt to bind the named driver to the device.
+// No-op if a driver is already bound. May block while the kernel runs the
+// driver's probe routine; if the device's firmware is in pre-init, the bind
+// can return -ETIMEDOUT after the kernel's wait_fw_init timeout.
+func (h *PCIHelper) BindDriver(driverName string) error {
+	bound, err := h.IsDriverBound()
+	if err != nil {
+		return err
+	}
+	if bound {
+		return nil
+	}
+	bdf := filepath.Base(h.Path())
+	bindPath := filepath.Join(h.sysFSRoot, "bus/pci/drivers", driverName, "bind")
+	if err := os.WriteFile(bindPath, []byte(bdf), 0644); err != nil {
+		return fmt.Errorf("failed to bind %s to %s: %w", bdf, driverName, err)
+	}
+	return nil
+}
+
 // GetMTU returns the current MTU of the PF interface
 func (h *PCIHelper) GetMTU() (int, error) {
 	interfaceName, err := h.InterfaceName()

internal/provisioning/hostagent/util/pci_test.go

@@ -522,6 +522,86 @@ var _ = Describe("PCI", func() {
 		})
 	})
 
+	Context("PCIHelper.IsDriverBound", Label("PCIHelper", "IsDriverBound"), func() {
+		It("should return false when device has no driver symlink", func() {
+			mock := createMockSysfs("0000:b1:00.0", "0xa2dc\n", "", nil, "")
+			defer mock.Cleanup()
+
+			helper := NewPCIHelper("0000:b1:00.0").SetSysFS(mock.TempSysfsDir())
+			bound, err := helper.IsDriverBound()
+			Expect(err).NotTo(HaveOccurred())
+			Expect(bound).To(BeFalse())
+		})
+
+		It("should return true when driver symlink exists", func() {
+			mock := createMockSysfs("0000:b1:00.0", "0xa2dc\n", "", nil, "")
+			defer mock.Cleanup()
+
+			// Create a fake driver symlink to mimic the kernel binding the device
+			driverTarget := filepath.Join(mock.TempSysfsDir(), "bus/pci/drivers/mlx5_core")
+			Expect(os.MkdirAll(driverTarget, 0755)).To(Succeed())
+			driverSymlink := filepath.Join(mock.PCIDevicesDir(), "0000:b1:00.0", "driver")
+			Expect(os.Symlink(driverTarget, driverSymlink)).To(Succeed())
+
+			helper := NewPCIHelper("0000:b1:00.0").SetSysFS(mock.TempSysfsDir())
+			bound, err := helper.IsDriverBound()
+			Expect(err).NotTo(HaveOccurred())
+			Expect(bound).To(BeTrue())
+		})
+	})
+
+	Context("PCIHelper.BindDriver", Label("PCIHelper", "BindDriver"), func() {
+		It("should write the BDF to the driver's bind file when no driver is bound", func() {
+			mock := createMockSysfs("0000:b1:00.0", "0xa2dc\n", "", nil, "")
+			defer mock.Cleanup()
+
+			// Create the driver's bind file (kernel-managed in real sysfs)
+			bindDir := filepath.Join(mock.TempSysfsDir(), "bus/pci/drivers/mlx5_core")
+			Expect(os.MkdirAll(bindDir, 0755)).To(Succeed())
+			bindPath := filepath.Join(bindDir, "bind")
+			Expect(os.WriteFile(bindPath, []byte{}, 0644)).To(Succeed())
+
+			helper := NewPCIHelper("0000:b1:00.0").SetSysFS(mock.TempSysfsDir())
+			Expect(helper.BindDriver("mlx5_core")).To(Succeed())
+
+			content, err := os.ReadFile(bindPath)
+			Expect(err).NotTo(HaveOccurred())
+			Expect(string(content)).To(Equal("0000:b1:00.0"))
+		})
+
+		It("should be a no-op when a driver is already bound", func() {
+			mock := createMockSysfs("0000:b1:00.0", "0xa2dc\n", "", nil, "")
+			defer mock.Cleanup()
+
+			// Pre-bind by creating the driver symlink
+			driverTarget := filepath.Join(mock.TempSysfsDir(), "bus/pci/drivers/mlx5_core")
+			Expect(os.MkdirAll(driverTarget, 0755)).To(Succeed())
+			driverSymlink := filepath.Join(mock.PCIDevicesDir(), "0000:b1:00.0", "driver")
+			Expect(os.Symlink(driverTarget, driverSymlink)).To(Succeed())
+
+			// Create the bind file too, but pre-populated to detect any unwanted writes
+			bindPath := filepath.Join(driverTarget, "bind")
+			Expect(os.WriteFile(bindPath, []byte("untouched"), 0644)).To(Succeed())
+
+			helper := NewPCIHelper("0000:b1:00.0").SetSysFS(mock.TempSysfsDir())
+			Expect(helper.BindDriver("mlx5_core")).To(Succeed())
+
+			content, err := os.ReadFile(bindPath)
+			Expect(err).NotTo(HaveOccurred())
+			Expect(string(content)).To(Equal("untouched"))
+		})
+
+		It("should return error when bind file does not exist", func() {
+			mock := createMockSysfs("0000:b1:00.0", "0xa2dc\n", "", nil, "")
+			defer mock.Cleanup()
+
+			helper := NewPCIHelper("0000:b1:00.0").SetSysFS(mock.TempSysfsDir())
+			err := helper.BindDriver("mlx5_core")
+			Expect(err).To(HaveOccurred())
+			Expect(err.Error()).To(ContainSubstring("failed to bind 0000:b1:00.0 to mlx5_core"))
+		})
+	})
+
 	Context("DiscoverDPUs", Label("DiscoverDPUs"), func() {
 		It("should return error when PCI devices directory does not exist", func() {
 			// Create mock to get cleanup() for path restoration, then point to non-existent path