| 1 | package hostops |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "crypto/ed25519" |
| 6 | "errors" |
| 7 | "strings" |
| 8 | "sync" |
| 9 | "testing" |
| 10 | "time" |
| 11 | |
| 12 | "github.com/carlosframework/platform/internal/hostclaim" |
| 13 | "github.com/carlosframework/platform/internal/release" |
| 14 | ) |
| 15 | |
| 16 | // racingStore loses the next armed admission CAS: just before it, it runs |
| 17 | // race (a concurrent renewal, settle or phase write through the same raw |
| 18 | // store), so the armed Put's ETag is stale and fails with ErrCASConflict. |
| 19 | type racingStore struct { |
| 20 | *release.DirStore |
| 21 | mu sync.Mutex |
| 22 | race func() |
| 23 | fired int |
| 24 | } |
| 25 | |
| 26 | func (s *racingStore) arm(race func()) { |
| 27 | s.mu.Lock() |
| 28 | s.race = race |
| 29 | s.mu.Unlock() |
| 30 | } |
| 31 | |
| 32 | func (s *racingStore) Put(ctx context.Context, key string, body []byte, tag string) (string, error) { |
| 33 | s.mu.Lock() |
| 34 | race := s.race |
| 35 | if strings.HasSuffix(key, "/admission.json") && race != nil { |
| 36 | s.race = nil |
| 37 | s.fired++ |
| 38 | } else { |
| 39 | race = nil |
| 40 | } |
| 41 | s.mu.Unlock() |
| 42 | if race != nil { |
| 43 | race() |
| 44 | } |
| 45 | return s.DirStore.Put(ctx, key, body, tag) |
| 46 | } |
| 47 | |
| 48 | func racingWriter(t *testing.T) (*ControlWriter, *racingStore, func() time.Time) { |
| 49 | t.Helper() |
| 50 | raw := &racingStore{DirStore: release.NewDirStore(t.TempDir())} |
| 51 | pub, _, err := ed25519.GenerateKey(nil) |
| 52 | if err != nil { |
| 53 | t.Fatal(err) |
| 54 | } |
| 55 | now := time.Date(2026, 10, 7, 12, 0, 0, 0, time.UTC) |
| 56 | clock := func() time.Time { return now } |
| 57 | w, err := NewControlWriter(raw, pub, clock) |
| 58 | if err != nil { |
| 59 | t.Fatal(err) |
| 60 | } |
| 61 | return w, raw, clock |
| 62 | } |
| 63 | |
| 64 | func startRegistration(t *testing.T, w *ControlWriter, clock func() time.Time, id string) *StartHandle { |
| 65 | t.Helper() |
| 66 | r := Registration{Kind: "create", Host: "x.example", Owner: testOwner, AppID: "app", Epoch: 1, Claim: SavedClaim{Absent: true}} |
| 67 | h, err := w.RegisterStart(t.Context(), "bab", "demo", "app", id, r, clock()) |
| 68 | if err != nil { |
| 69 | t.Fatal(err) |
| 70 | } |
| 71 | return h |
| 72 | } |
| 73 | |
| 74 | // Review P1-3: a routine renewal racing the settle CAS must not leave a known |
| 75 | // outcome "issued" forever. |
| 76 | func TestSettleSurvivesRenewalCAS(t *testing.T) { |
| 77 | ctx := t.Context() |
| 78 | w, raw, clock := racingWriter(t) |
| 79 | seedAdmission(t, raw.DirStore, "bab", "demo", openAdmission("app")) |
| 80 | h := startRegistration(t, w, clock, "reg") |
| 81 | _, settle, err := h.Runner(clock).Issue(ctx, Effect{ID: "eff", Resource: "store:x", Action: "grant"}) |
| 82 | if err != nil { |
| 83 | t.Fatal(err) |
| 84 | } |
| 85 | raw.arm(func() { |
| 86 | if err := h.Renew(ctx, clock()); err != nil { |
| 87 | t.Errorf("renew: %v", err) |
| 88 | } |
| 89 | }) |
| 90 | if err = settle("ok", nil); err != nil { |
| 91 | t.Fatalf("settle lost to a renewal: %v", err) |
| 92 | } |
| 93 | if raw.fired != 1 { |
| 94 | t.Fatalf("race not exercised: %d", raw.fired) |
| 95 | } |
| 96 | a, err := ReadAdmission(ctx, raw.DirStore, "bab", "demo", "app") |
| 97 | if err != nil { |
| 98 | t.Fatal(err) |
| 99 | } |
| 100 | if e := a.Value.Effects["eff"]; e.State != "complete" || e.ResultReference != "ok" { |
| 101 | t.Fatalf("effect after settle: %+v", e) |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | // Review P1-3: Renew and Bind themselves retry a lost CAS. |
| 106 | func TestRenewAndBindRetryLostCAS(t *testing.T) { |
| 107 | ctx := t.Context() |
| 108 | w, raw, clock := racingWriter(t) |
| 109 | seedAdmission(t, raw.DirStore, "bab", "demo", openAdmission("app")) |
| 110 | one := startRegistration(t, w, clock, "one") |
| 111 | two := startRegistration(t, w, clock, "two") |
| 112 | raw.arm(func() { _ = two.Renew(ctx, clock()) }) |
| 113 | if err := one.Renew(ctx, clock()); err != nil { |
| 114 | t.Fatalf("renew lost to a sibling renewal: %v", err) |
| 115 | } |
| 116 | raw.arm(func() { _ = two.Renew(ctx, clock()) }) |
| 117 | if err := one.Bind(ctx, clock(), "inc-1", clock()); err != nil { |
| 118 | t.Fatalf("bind lost to a sibling renewal: %v", err) |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | // Review P1-3: the prepare and issue CASes run before the call, so they |
| 123 | // replay after a lost race; the effect ends issued exactly once. |
| 124 | func TestIssueReplaysPreparedAcrossCAS(t *testing.T) { |
| 125 | ctx := t.Context() |
| 126 | w, raw, clock := racingWriter(t) |
| 127 | seedAdmission(t, raw.DirStore, "bab", "demo", openAdmission("app")) |
| 128 | other := startRegistration(t, w, clock, "other") |
| 129 | h := startRegistration(t, w, clock, "reg") |
| 130 | raw.arm(func() { _ = other.Renew(ctx, clock()) }) |
| 131 | _, settle, err := h.Runner(clock).Issue(ctx, Effect{ID: "eff", Resource: "store:x", Action: "grant"}) |
| 132 | if err != nil { |
| 133 | t.Fatalf("issue lost to a renewal: %v", err) |
| 134 | } |
| 135 | a, _ := ReadAdmission(ctx, raw.DirStore, "bab", "demo", "app") |
| 136 | if a.Value.Effects["eff"].State != "issued" { |
| 137 | t.Fatalf("effect: %+v", a.Value.Effects["eff"]) |
| 138 | } |
| 139 | if err = settle("ok", nil); err != nil { |
| 140 | t.Fatal(err) |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | // Review P1-3: SetPhase and the reopen retry; a retried reopen bumps the |
| 145 | // epoch exactly once. |
| 146 | func TestSetPhaseAndFinishRetryLostCAS(t *testing.T) { |
| 147 | ctx := t.Context() |
| 148 | w, raw, clock := racingWriter(t) |
| 149 | seedAdmission(t, raw.DirStore, "bab", "demo", openAdmission("app")) |
| 150 | h, err := w.BeginExclusive(ctx, "bab", "demo", "app", AppOperation{ID: "op", Kind: OpAppMutation, AppID: "app"}, testOwner) |
| 151 | if err != nil { |
| 152 | t.Fatal(err) |
| 153 | } |
| 154 | _, settle, err := h.Runner(clock).Issue(ctx, Effect{ID: "op", Resource: "app-mutation", Action: "test"}) |
| 155 | if err != nil { |
| 156 | t.Fatal(err) |
| 157 | } |
| 158 | if err = settle("joined", nil); err != nil { |
| 159 | t.Fatal(err) |
| 160 | } |
| 161 | bump := func() { |
| 162 | before, err := ReadAdmission(ctx, raw.DirStore, "bab", "demo", "app") |
| 163 | if err != nil { |
| 164 | t.Error(err) |
| 165 | return |
| 166 | } |
| 167 | // An unrelated same-op intent write (another worker's phase journal). |
| 168 | if _, err = w.updateAdmission(ctx, before, func(v *AppAdmission) error { v.Op.Intent["note"] = "x"; return nil }); err != nil { |
| 169 | t.Error(err) |
| 170 | } |
| 171 | } |
| 172 | raw.arm(bump) |
| 173 | if err = h.SetPhase(ctx, "", "done", map[string]string{"k": "v"}); err != nil { |
| 174 | t.Fatalf("SetPhase lost: %v", err) |
| 175 | } |
| 176 | raw.arm(bump) |
| 177 | if err = h.FinishMutation(ctx); err != nil { |
| 178 | t.Fatalf("finish lost: %v", err) |
| 179 | } |
| 180 | a, _ := ReadAdmission(ctx, raw.DirStore, "bab", "demo", "app") |
| 181 | if a.Value.State != "open" || a.Value.Epoch != 2 { |
| 182 | t.Fatalf("reopen: state %s epoch %d", a.Value.State, a.Value.Epoch) |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | func restoreFixture(t *testing.T) (*ControlWriter, *release.DirStore, func() time.Time, *OperationHandle, string) { |
| 187 | t.Helper() |
| 188 | ctx := t.Context() |
| 189 | w, raw, clock := testWriter(t) |
| 190 | seedAdmission(t, raw, "bab", "demo", openAdmission("app")) |
| 191 | if err := release.TrashApp(ctx, raw, "bab", "demo", "owner", clock()); err != nil { |
| 192 | t.Fatal(err) |
| 193 | } |
| 194 | _, tag, err := raw.Get(ctx, release.AppTombstoneKey("bab", "demo")) |
| 195 | if err != nil { |
| 196 | t.Fatal(err) |
| 197 | } |
| 198 | h, err := w.BeginExclusive(ctx, "bab", "demo", "app", AppOperation{ID: "op", Kind: OpAppMutation, AppID: "app", Intent: map[string]string{"action": "restore"}}, testOwner) |
| 199 | if err != nil { |
| 200 | t.Fatal(err) |
| 201 | } |
| 202 | if err = h.PrepareRestore(ctx, tag, clock().Unix()); err != nil { |
| 203 | t.Fatal(err) |
| 204 | } |
| 205 | return w, raw, clock, h, tag |
| 206 | } |
| 207 | |
| 208 | // SQL-fix order BeginExclusive -> PrepareRestore -> Issue: a crash after |
| 209 | // the receipt and before Issue is recovered from the receipt alone, the |
| 210 | // epoch advances once, and a late Issue gets the recorded result, never a |
| 211 | // permit. |
| 212 | func TestRecoverPreparedReceiptWithNoEffect(t *testing.T) { |
| 213 | ctx := t.Context() |
| 214 | w, raw, clock, h, _ := restoreFixture(t) |
| 215 | if err := w.RecoverMutation(ctx, "bab", "demo", "app", "op"); err != nil { |
| 216 | t.Fatal(err) |
| 217 | } |
| 218 | a, err := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 219 | if err != nil { |
| 220 | t.Fatal(err) |
| 221 | } |
| 222 | e := a.Value.Effects["op"] |
| 223 | if a.Value.State != "open" || a.Value.Op != nil || a.Value.Epoch != 2 || a.Value.LastRestore.State != "complete" || e.State != "complete" || e.ResultReference != "restore-reconciled" { |
| 224 | t.Fatalf("not recovered: %+v", a.Value) |
| 225 | } |
| 226 | if _, _, err = raw.Get(ctx, release.AppTombstoneKey("bab", "demo")); err == nil { |
| 227 | t.Fatal("tombstone survived") |
| 228 | } |
| 229 | _, _, err = h.Runner(clock).Issue(ctx, Effect{ID: "op", Resource: OpAppMutation, Action: "restore", Inputs: map[string]string{"tombstone_etag": "x"}}) |
| 230 | if result, ok := EffectResult(err); !ok || result != "restore-reconciled" { |
| 231 | t.Fatalf("late issue got a permit or the wrong refusal: %v", err) |
| 232 | } |
| 233 | // Idempotent: the issuer's own recovery call now succeeds. |
| 234 | if err = w.RecoverMutation(ctx, "bab", "demo", "app", "op"); err != nil { |
| 235 | t.Fatalf("second recovery: %v", err) |
| 236 | } |
| 237 | a, _ = ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 238 | if a.Value.Epoch != 2 { |
| 239 | t.Fatalf("epoch bumped twice: %d", a.Value.Epoch) |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | // Recovery's decision on an issued restore effect stands: the issuer's late |
| 244 | // settle neither overwrites it nor fails. |
| 245 | func TestSettleNeverOverwritesRecoveredOutcome(t *testing.T) { |
| 246 | ctx := t.Context() |
| 247 | w, raw, clock, h, _ := restoreFixture(t) |
| 248 | _, settle, err := h.Runner(clock).Issue(ctx, Effect{ID: "op", Resource: OpAppMutation, Action: "restore"}) |
| 249 | if err != nil { |
| 250 | t.Fatal(err) |
| 251 | } |
| 252 | if err = w.RecoverMutation(ctx, "bab", "demo", "app", "op"); err != nil { |
| 253 | t.Fatal(err) |
| 254 | } |
| 255 | callErr := errors.New("lost response") |
| 256 | if err = settle("", callErr); !errors.Is(err, callErr) { |
| 257 | t.Fatalf("late settle: %v", err) |
| 258 | } |
| 259 | a, _ := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 260 | if e := a.Value.Effects["op"]; e.State != "complete" || e.ResultReference != "restore-reconciled" { |
| 261 | t.Fatalf("recovered outcome overwritten: %+v", e) |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | func mailMember(t *testing.T, w *ControlWriter, clock func() time.Time) (*MailGateLease, *OperationHandle, func(string, error) error) { |
| 266 | t.Helper() |
| 267 | ctx := t.Context() |
| 268 | l, err := w.AcquireMailGate(ctx, MailGateOwner{Console: "c", Enrollment: "e", Controller: "k"}, MailGateScope{}) |
| 269 | if err != nil { |
| 270 | t.Fatal(err) |
| 271 | } |
| 272 | if err = w.RecordMailGateMember(ctx, l, MailGateMember{Account: "bab", App: "demo", AppID: "app", Operation: "op", Effect: "op"}); err != nil { |
| 273 | t.Fatal(err) |
| 274 | } |
| 275 | h, err := w.BeginExclusive(ctx, "bab", "demo", "app", AppOperation{ID: "op", Kind: OpAppMutation, AppID: "app", Intent: map[string]string{"action": "mail", "gate": l.ID}}, testOwner) |
| 276 | if err != nil { |
| 277 | t.Fatal(err) |
| 278 | } |
| 279 | _, settle, err := h.Runner(clock).Issue(ctx, Effect{ID: "op", Resource: "mail-gate:" + l.ID, Action: "mail", Inputs: map[string]string{"gate": l.ID}}) |
| 280 | if err != nil { |
| 281 | t.Fatal(err) |
| 282 | } |
| 283 | return l, h, settle |
| 284 | } |
| 285 | |
| 286 | // Review P1-2: a mail member whose effect completed is recovered after a |
| 287 | // crash, but only once the gate no longer holds its gate ID. |
| 288 | func TestMailMutationRecoversOnlyAfterGateRelease(t *testing.T) { |
| 289 | ctx := t.Context() |
| 290 | w, raw, clock := testWriter(t) |
| 291 | seedAdmission(t, raw, "bab", "demo", openAdmission("app")) |
| 292 | l, _, settle := mailMember(t, w, clock) |
| 293 | if err := settle("joined", nil); err != nil { |
| 294 | t.Fatal(err) |
| 295 | } |
| 296 | // Crash before Release: the gate still holds this gate ID. |
| 297 | if err := w.RecoverMutation(ctx, "bab", "demo", "app", "op"); !errors.Is(err, ErrPending) { |
| 298 | t.Fatalf("reopened while the gate is held: %v", err) |
| 299 | } |
| 300 | if err := w.ReleaseMailGate(ctx, l); err != nil { |
| 301 | t.Fatal(err) |
| 302 | } |
| 303 | if err := w.RecoverMutation(ctx, "bab", "demo", "app", "op"); err != nil { |
| 304 | t.Fatalf("complete effect + free gate not recovered: %v", err) |
| 305 | } |
| 306 | a, _ := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 307 | if a.Value.State != "open" || a.Value.Epoch != 2 { |
| 308 | t.Fatalf("not reopened: %+v", a.Value) |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | // Review P1-2: an uncertain mail effect keeps gate and op until the fenced |
| 313 | // reconciler settles it from readback; the old lease is fenced; the op then |
| 314 | // recovers. |
| 315 | func TestReconcileMailGateReadbackDoesNotProveQuiescence(t *testing.T) { |
| 316 | ctx := t.Context() |
| 317 | w, raw, clock := testWriter(t) |
| 318 | seedAdmission(t, raw, "bab", "demo", openAdmission("app")) |
| 319 | l, _, settle := mailMember(t, w, clock) |
| 320 | callErr := errors.New("timeout") |
| 321 | _ = settle(MailAttemptsResult([]MailAttempt{{Surface: "store", Method: "Put", Key: "accounts/bab/email/domains/example.com.json"}}), callErr) |
| 322 | if _, err := w.ReconcileMailGate(ctx, l.ID, 0, nil); !errors.Is(err, ErrMailUnproven) { |
| 323 | t.Fatalf("uncertain effect reopened: %v", err) |
| 324 | } |
| 325 | if err := w.RecoverMutation(ctx, "bab", "demo", "app", "op"); !errors.Is(err, ErrPending) { |
| 326 | t.Fatalf("op reopened: %v", err) |
| 327 | } |
| 328 | a, _ := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 329 | if a.Value.Effects["op"].State != "uncertain" || a.Value.Op == nil { |
| 330 | t.Fatal("durable ownership was lost") |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | // Review P1-2: no readback authority (DNS) or a still-issued effect refuses; |
| 335 | // gate, op and effect all stay held. |
| 336 | func TestReconcileMailGateRefusesUnproven(t *testing.T) { |
| 337 | for _, tc := range []string{"dns", "issued"} { |
| 338 | t.Run(tc, func(t *testing.T) { |
| 339 | ctx := t.Context() |
| 340 | w, raw, clock := testWriter(t) |
| 341 | seedAdmission(t, raw, "bab", "demo", openAdmission("app")) |
| 342 | l, _, settle := mailMember(t, w, clock) |
| 343 | want := "issued" |
| 344 | if tc == "dns" { |
| 345 | _ = settle(MailAttemptsResult([]MailAttempt{{Surface: "dns", Method: "EnsureRecord", Name: "x.example.com|TXT"}}), errors.New("timeout")) |
| 346 | want = "uncertain" |
| 347 | } |
| 348 | if _, err := w.ReconcileMailGate(ctx, l.ID, 0, nil); !errors.Is(err, ErrMailUnproven) { |
| 349 | t.Fatalf("unproven reconciled: %v", err) |
| 350 | } |
| 351 | g, _, _ := w.readMailGate(ctx) |
| 352 | a, _ := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 353 | if g.State != "held" || g.ID != l.ID || a.Value.Effects["op"].State != want || a.Value.Op == nil { |
| 354 | t.Fatalf("released something: gate %+v effect %+v", g, a.Value.Effects["op"]) |
| 355 | } |
| 356 | if err := w.RecoverMutation(ctx, "bab", "demo", "app", "op"); !errors.Is(err, ErrPending) { |
| 357 | t.Fatalf("reopened: %v", err) |
| 358 | } |
| 359 | }) |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | // Review P1-1 (v2 residue): an app grant can neither replace a foreign |
| 364 | // holder by naming its own op in the new claim, nor overwrite a foreign op |
| 365 | // on a free claim. |
| 366 | func TestAppGrantCannotTakeForeignClaim(t *testing.T) { |
| 367 | g := lifecycleGrant{kind: "app", account: "bab", app: "demo", appID: "app", operation: "mine"} |
| 368 | foreign := &hostclaim.Claim{Host: "x.example", Gen: 4, Holder: &hostclaim.Holder{Account: "zed", App: "web", Incarnation: "z1"}, Op: hostclaim.Op{Kind: hostclaim.KindServing}} |
| 369 | steal := &hostclaim.Claim{Host: "x.example", Gen: 5, Holder: &hostclaim.Holder{Account: "bab", App: "demo", Incarnation: "b1"}, Op: hostclaim.Op{ID: "mine"}} |
| 370 | if appGrantAllows(g, foreign, steal) { |
| 371 | t.Fatal("foreign holder replaced by naming the grant's op") |
| 372 | } |
| 373 | evict := &hostclaim.Claim{Host: "x.example", Gen: 5, Op: hostclaim.Op{ID: "mine", Kind: hostclaim.KindPurging}} |
| 374 | if appGrantAllows(g, foreign, evict) { |
| 375 | t.Fatal("foreign holder removed") |
| 376 | } |
| 377 | fenced := &hostclaim.Claim{Host: "x.example", Gen: 4, Op: hostclaim.Op{ID: "theirs", Kind: hostclaim.KindPurging}} |
| 378 | if appGrantAllows(g, fenced, evict) { |
| 379 | t.Fatal("foreign purge fence overwritten") |
| 380 | } |
| 381 | free := &hostclaim.Claim{Host: "x.example", Gen: 4} |
| 382 | if !appGrantAllows(g, free, evict) { |
| 383 | t.Fatal("own fence on a free claim refused") |
| 384 | } |
| 385 | if claimTransitionShape("x.example", free, &hostclaim.Claim{Host: "x.example", Gen: 9}) || claimTransitionShape("y.example", nil, &hostclaim.Claim{Host: "x.example", Gen: 1}) { |
| 386 | t.Fatal("gen skip or host mismatch admitted") |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | // Review P1-1 (v2 residue): a lifecycle grant dies with its registration. |
| 391 | func TestLifecycleGrantDiesWithRegistration(t *testing.T) { |
| 392 | ctx := t.Context() |
| 393 | w, raw, clock := testWriter(t) |
| 394 | seedAdmission(t, raw, "bab", "demo", openAdmission("app")) |
| 395 | h := startRegistration(t, w, clock, "reg") |
| 396 | grants, _ := h.Lifecycle(ctx).Value(lifecycleGrantKey{}).([]lifecycleGrant) |
| 397 | if len(grants) != 1 { |
| 398 | t.Fatal("no grant") |
| 399 | } |
| 400 | if live, err := grantLive(ctx, raw, grants[0]); err != nil || !live { |
| 401 | t.Fatalf("live registration's grant refused: %v", err) |
| 402 | } |
| 403 | before, _ := ReadAdmission(ctx, raw, "bab", "demo", "app") |
| 404 | if _, err := w.updateAdmission(ctx, before, func(v *AppAdmission) error { delete(v.Registrations, "reg"); return nil }); err != nil { |
| 405 | t.Fatal(err) |
| 406 | } |
| 407 | if live, err := grantLive(ctx, raw, grants[0]); err != nil || live { |
| 408 | t.Fatalf("drained registration's grant still live: %v", err) |
| 409 | } |
| 410 | } |
| 411 | |
amadan