linux: let the kernel resolve the tun.dev %d template atomically

Resolving the template in userspace (LinkList then TUNSETIFF) races with
concurrent instances: both can pick nebula0, and with multiqueue the loser
silently attaches as a second queue to the winner's device instead of
failing. TUNSETIFF natively substitutes a single %d via dev_alloc_name,
atomically and always allocating a fresh device, so keep the fail-fast
validation but pass the template through and read the resolved name back.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fg7oD3uZ4jZjjQ4gRsPAwB

examples/config.yml

@@ -242,6 +242,10 @@ tun:
   # When tun is disabled, a lighthouse can be started without a local tun interface (and therefore without root)
   disabled: false
   # Name of the device. If not set, a default will be chosen by the OS.
+  # For Linux: a single `%d` anywhere in the name is treated as a template and replaced with the
+  #   lowest number that yields an unused device name (e.g. `nebula%d` becomes `nebula0`, then `nebula1`, and so on, `neb%dprod` becomes `neb0prod`).
+  #   Only on Linux: `nebula%d` is the default if tun.dev is unset.
+  #   The name, both before and after %d substitution, must be shorter than the kernel limit of 16 characters.
   # For macOS: if set, must be in the form `utun[0-9]+`.
   # For NetBSD: Required to be set, must be in the form `tun[0-9]+`
   dev: nebula1

overlay/tun_linux.go

@@ -5,6 +5,7 @@ package overlay
 
 import (
 	"encoding/binary"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -250,6 +251,17 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 }
 
 func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+	// Validate the device name up front so a bad tun.dev fails fast, before we
+	// open /dev/net/tun or leak a file descriptor. A single %d in the name is
+	// substituted by the kernel during TUNSETIFF (dev_alloc_name) with the
+	// lowest number that yields an unused device name. Resolving the template
+	// in the kernel keeps the pick-a-name/create-the-device pair atomic, so
+	// concurrent callers can never race each other to the same name.
+	tunName := c.GetString("tun.dev", "nebula%d")
+	if err := validateTunName(tunName); err != nil {
+		return nil, err
+	}
+
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
 		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
@@ -277,12 +289,11 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	if multiqueue {
 		req.Flags |= unix.IFF_MULTI_QUEUE
 	}
-	nameStr := c.GetString("tun.dev", "")
+	copy(req.Name[:], tunName)
-	copy(req.Name[:], nameStr)
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
 		_ = unix.Close(fd)
 		return nil, &NameError{
-			Name:       nameStr,
+			Name:       tunName,
 			Underlying: err,
 		}
 	}
@@ -298,6 +309,27 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	return t, nil
 }
 
+func validateTunName(tunName string) error {
+	if !strings.Contains(tunName, "%d") {
+		if len(tunName) >= unix.IFNAMSIZ {
+			return fmt.Errorf("tun.dev %q is not shorter than the maximum device name length of %d", tunName, unix.IFNAMSIZ)
+		}
+		return nil
+	}
+	if strings.Count(tunName, "%d") > 1 {
+		return fmt.Errorf("tun.dev template %q may only contain a single %%d", tunName)
+	}
+	if tunName == "%d" {
+		return errors.New("please don't name your tun device '%d'")
+	}
+	// The kernel substitutes the %d itself and requires the template, like a
+	// literal name, to be NUL-terminated within IFNAMSIZ bytes.
+	if len(tunName) >= unix.IFNAMSIZ {
+		return fmt.Errorf("tun.dev template %q is not shorter than the maximum device name length of %d", tunName, unix.IFNAMSIZ)
+	}
+	return nil
+}
+
 // newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
 func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	tfd, err := newTunFd(fd)
@@ -768,7 +800,6 @@ func (t *tun) isGatewayInVpnNetworks(gwAddr netip.Addr) bool {
 
 func (t *tun) getGatewaysFromRoute(r *netlink.Route) routing.Gateways {
 	var gateways routing.Gateways
-
 	link, err := netlink.LinkByName(t.Device)
 	if err != nil {
 		t.l.Error("Ignoring route update: failed to get link by name", "deviceName", t.Device)

overlay/tun_linux_test.go

@@ -3,7 +3,12 @@
 
 package overlay
 
-import "testing"
+import (
+	"strings"
+	"testing"
+
+	"golang.org/x/sys/unix"
+)
 
 var runAdvMSSTests = []struct {
 	name     string
@@ -32,3 +37,39 @@ func TestTunAdvMSS(t *testing.T) {
 		})
 	}
 }
+
+func TestValidateTunName(t *testing.T) {
+	// A device name must be shorter than IFNAMSIZ (i.e. IFNAMSIZ-1 chars max).
+	maxLenName := strings.Repeat("a", unix.IFNAMSIZ-1)
+
+	tests := []struct {
+		name    string
+		tmpl    string
+		wantErr bool
+	}{
+		{"short literal name is fine", "nebula1", false},
+		{"literal name at the max length is fine", maxLenName, false},
+		{"literal name at IFNAMSIZ is rejected", strings.Repeat("a", unix.IFNAMSIZ), true},
+		{"trailing template is fine", "nebula%d", false},
+		{"mid-string template is fine", "neb%dprod", false},
+		{"leading template is fine", "%dnebula", false},
+		{"template at the max length is fine", strings.Repeat("a", unix.IFNAMSIZ-3) + "%d", false},
+		{"template at IFNAMSIZ is rejected", strings.Repeat("a", unix.IFNAMSIZ-2) + "%d", true},
+		{"bare %d is rejected", "%d", true},
+		{"multiple %d is rejected", "neb%d%dprod", true},
+		{"over-long template is rejected", strings.Repeat("a", unix.IFNAMSIZ-1) + "%d", true},
+		{"over-long mid-string template is rejected", "neb%d" + strings.Repeat("a", unix.IFNAMSIZ-3), true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			err := validateTunName(tt.tmpl)
+			if tt.wantErr && err == nil {
+				t.Fatalf("expected an error for %q, got none", tt.tmpl)
+			}
+			if !tt.wantErr && err != nil {
+				t.Fatalf("unexpected error for %q: %v", tt.tmpl, err)
+			}
+		})
+	}
+}