gitea: forge-account reads the forge's own host key file before scanning its sshd (issue 1939)

modules/gitea/cmd/forge-account-provider/main.go

@@ -65,6 +65,7 @@ func configFromEnv(getenv func(string) string) (Config, error) {
 		HostFile:    getenv("MESH_FORGE_HOST_FILE"),
 		OriginsFile: getenv("MESH_GITEA_ORIGINS_FILE"),
 		Port:        port,
+		HostKeyFile: getenv("MESH_GITEA_HOST_KEY_FILE"), // optional
 		Forge: forgehttp.Client{URL: strings.TrimRight(getenv("MESH_GITEA_URL"), "/"), User: getenv("MESH_GITEA_ADMIN_USER"),
 			PasswordFile: getenv("MESH_GITEA_ADMIN_PASSWORD_FILE"), Timeout: 20 * time.Second},
 		Runner: execRunner{},

modules/gitea/cmd/forge-account-provider/service.go

@@ -44,10 +44,13 @@ type Config struct {
 	// origins, one per line.
 	HostFile, OriginsFile string
 	// Port is where the forge's sshd is published on the mesh.
-	Port   int
-	Forge  forgehttp.Client
-	Runner Runner
-	Log    func(format string, a ...any)
+	Port int
+	// HostKeyFile is the forge's own ed25519 host key (the .pub file in its data directory). It is read first; when it
+	// is unset or unreadable the forge's sshd is asked with ssh-keyscan instead (issue 1939).
+	HostKeyFile string
+	Forge       forgehttp.Client
+	Runner      Runner
+	Log         func(format string, a ...any)
 }
 
 // Service answers the consumers of the provision `forge-account` (novox/hq ADR 0399, ADR 0400; issue 955): for each
@@ -448,14 +451,30 @@ func (s *Service) httpsBase() (string, error) {
 	return "", fmt.Errorf("%s names no https origin", s.cfg.OriginsFile)
 }
 
-// scan reads the forge's ed25519 host key from its own sshd and writes it as the machines' known_hosts line.
+// scan gives the forge's ed25519 host key as the machines' known_hosts line. The key the forge itself holds is read
+// from its file first; only when that cannot be read is the forge's sshd asked over the network, and when that fails
+// too (the forge's container is restarting) the key read before is kept, as a host key does not change (issue 1939).
 func (s *Service) scan(ctx context.Context) (string, error) {
 	host, err := s.hostName()
 	if err != nil {
 		return "", err
 	}
+	if s.cfg.HostKeyFile != "" {
+		raw, ferr := os.ReadFile(s.cfg.HostKeyFile)
+		if ferr == nil {
+			if line, perr := knownHostLine("127.0.0.1 "+strings.TrimSpace(string(raw)), host, s.cfg.Port); perr == nil {
+				return line, nil
+			}
+		}
+	}
 	out, err := s.cfg.Runner.Run(ctx, "ssh-keyscan", "-t", "ed25519", "-p", fmt.Sprint(s.cfg.Port), "127.0.0.1")
 	if err != nil {
+		s.mu.Lock()
+		kept := s.known
+		s.mu.Unlock()
+		if kept != "" {
+			return kept, nil
+		}
 		return "", fmt.Errorf("the forge's ssh host key cannot be read: %w", err)
 	}
 	return knownHostLine(string(out), host, s.cfg.Port)

modules/gitea/cmd/forge-account-provider/service_test.go

@@ -696,3 +696,42 @@ func TestNothingIsSaidForAPerNodeUserWithoutMeshKeys(t *testing.T) {
 		}
 	}
 }
+
+// The forge's own host key file is read first: no scan over the network is needed (issue 1939).
+func TestTheHostKeyIsReadFromTheForgesOwnFile(t *testing.T) {
+	r := newRig(t)
+	f := filepath.Join(t.TempDir(), "ssh_host_ed25519_key.pub")
+	if err := os.WriteFile(f, []byte("ssh-ed25519 AAAAC3FILEKEY forge@host\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	r.svc.cfg.HostKeyFile = f
+	r.runner.err = errors.New("sshd down")
+	got, err := r.svc.scan(context.Background())
+	if err != nil || got != "[novox.internal]:222 ssh-ed25519 AAAAC3FILEKEY" {
+		t.Fatalf("scan = %q, %v", got, err)
+	}
+	if r.runner.args != nil {
+		t.Errorf("the forge's sshd was asked: %v", r.runner.args)
+	}
+}
+
+// A file that cannot be read falls back to the forge's sshd.
+func TestAnUnreadableHostKeyFileFallsBackToTheScan(t *testing.T) {
+	r := newRig(t)
+	r.svc.cfg.HostKeyFile = filepath.Join(t.TempDir(), "missing.pub")
+	got, err := r.svc.scan(context.Background())
+	if err != nil || got != "[novox.internal]:222 ssh-ed25519 AAAAC3HOSTKEY" {
+		t.Fatalf("scan = %q, %v", got, err)
+	}
+}
+
+// While the forge restarts, a key read before is kept: the sshd is not up, the key has not changed.
+func TestARestartingForgeKeepsTheKeyReadBefore(t *testing.T) {
+	r := newRig(t)
+	r.svc.known = "[novox.internal]:222 ssh-ed25519 AAAAC3HOSTKEY"
+	r.runner.err = errors.New("connection refused")
+	got, err := r.svc.scan(context.Background())
+	if err != nil || got != r.svc.known {
+		t.Fatalf("scan = %q, %v", got, err)
+	}
+}

modules/gitea/module.json

@@ -537,6 +537,7 @@
           "MESH_GITEA_ADMIN_PASSWORD_FILE": "${dir:state}/admin.secret",
           "MESH_GITEA_ORIGINS_FILE": "${dir:mesh-state}/origins",
           "MESH_GITEA_SSH_PORT": "222",
+          "MESH_GITEA_HOST_KEY_FILE": "${dir:data}/ssh/ssh_host_ed25519_key.pub",
           "MESH_FORGE_HOST_FILE": "/etc/gitea/forge-host"
         }
       }