| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "github.com/carlosframework/platform/internal/edge" |
| 8 | "github.com/carlosframework/platform/internal/registry" |
| 9 | "net/http" |
| 10 | "net/http/httptest" |
| 11 | "sync/atomic" |
| 12 | "testing" |
| 13 | "time" |
| 14 | ) |
| 15 | |
| 16 | type maintenanceProc struct { |
| 17 | calls atomic.Int32 |
| 18 | exited atomic.Bool |
| 19 | gate <-chan struct{} |
| 20 | entered chan struct{} |
| 21 | err error |
| 22 | remains bool |
| 23 | } |
| 24 | |
| 25 | func (p *maintenanceProc) Stop(ctx context.Context) error { |
| 26 | if p.calls.Add(1) == 1 && p.entered != nil { |
| 27 | close(p.entered) |
| 28 | } |
| 29 | if p.gate != nil { |
| 30 | select { |
| 31 | case <-p.gate: |
| 32 | case <-ctx.Done(): |
| 33 | return ctx.Err() |
| 34 | } |
| 35 | } |
| 36 | if !p.remains { |
| 37 | p.exited.Store(true) |
| 38 | } |
| 39 | return p.err |
| 40 | } |
| 41 | func (p *maintenanceProc) Exited() bool { return p.exited.Load() } |
| 42 | |
| 43 | type maintenanceRep struct { |
| 44 | gate <-chan struct{} |
| 45 | entered chan struct{} |
| 46 | calls atomic.Int32 |
| 47 | err error |
| 48 | } |
| 49 | |
| 50 | func (r *maintenanceRep) Err() error { return nil } |
| 51 | func (r *maintenanceRep) Close(ctx context.Context) error { |
| 52 | if r.calls.Add(1) == 1 && r.entered != nil { |
| 53 | close(r.entered) |
| 54 | } |
| 55 | if r.gate != nil { |
| 56 | select { |
| 57 | case <-r.gate: |
| 58 | case <-ctx.Done(): |
| 59 | return ctx.Err() |
| 60 | } |
| 61 | } |
| 62 | return r.err |
| 63 | } |
| 64 | func maintenanceContext(t *testing.T, duration time.Duration) context.Context { |
| 65 | t.Helper() |
| 66 | ctx, cancel := context.WithTimeout(t.Context(), duration) |
| 67 | t.Cleanup(cancel) |
| 68 | return ctx |
| 69 | } |
| 70 | func maintenanceAwake(t *testing.T) (*Service, *actEntry, *maintenanceProc, *maintenanceProc, *maintenanceRep) { |
| 71 | t.Helper() |
| 72 | a, _, _, reg := testActivator(t, time.Hour) |
| 73 | t.Cleanup(func() { reg.Close() }) |
| 74 | main, side, rep := &maintenanceProc{}, &maintenanceProc{}, &maintenanceRep{} |
| 75 | e := &actEntry{host: "fence.example", route: testRoute("fence.example"), state: actAwake, proc: main, sidecar: side, rep: rep, lastActive: a.now()} |
| 76 | a.entries[e.host] = e |
| 77 | return a, e, main, side, rep |
| 78 | } |
| 79 | func TestMaintenanceFenceChecksEveryStopOutcome(t *testing.T) { |
| 80 | t.Parallel() |
| 81 | for _, scenario := range []string{"success", "main-error", "sidecar-error", "replica-error", "main-alive", "sidecar-alive", "missing-main", "missing-replica", "uncertain-sidecar"} { |
| 82 | t.Run(scenario, func(t *testing.T) { |
| 83 | a, e, main, side, rep := maintenanceAwake(t) |
| 84 | fail := errors.New("private backend detail") |
| 85 | switch scenario { |
| 86 | case "main-error": |
| 87 | main.err = fail |
| 88 | case "sidecar-error": |
| 89 | side.err = fail |
| 90 | case "replica-error": |
| 91 | rep.err = fail |
| 92 | case "main-alive": |
| 93 | main.remains = true |
| 94 | case "sidecar-alive": |
| 95 | side.remains = true |
| 96 | case "missing-main": |
| 97 | e.proc = nil |
| 98 | case "missing-replica": |
| 99 | e.rep = nil |
| 100 | case "uncertain-sidecar": |
| 101 | e.sidecarUncertain = true |
| 102 | } |
| 103 | proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1) |
| 104 | if scenario == "success" { |
| 105 | if err != nil || !proof.Matches(e.host, 1) || proof.Matches(e.host, 2) { |
| 106 | t.Fatal("checked stop missing", err) |
| 107 | } |
| 108 | } else if err != ErrStopUnproven || proof != nil { |
| 109 | t.Fatal("failed stop became evidence", err) |
| 110 | } |
| 111 | if a.fences[e.host] != 1 { |
| 112 | t.Fatal("failure removed fence") |
| 113 | } |
| 114 | if !a.Quiesced(e.host) { |
| 115 | t.Fatal("ordinary entry removal changed") |
| 116 | } |
| 117 | retry, retryErr := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1) |
| 118 | if scenario != "success" && (retryErr != ErrStopUnproven || retry != nil) { |
| 119 | t.Fatal("deleted entry erased failure") |
| 120 | } |
| 121 | if main.calls.Load() > 1 || side.calls.Load() > 1 || rep.calls.Load() > 1 { |
| 122 | t.Fatal("replayed spent stop") |
| 123 | } |
| 124 | }) |
| 125 | } |
| 126 | } |
| 127 | func TestMaintenanceFenceBlocksEveryAdmissionIncludingStaleRoute(t *testing.T) { |
| 128 | t.Parallel() |
| 129 | a, _, _, reg := testActivator(t, time.Hour) |
| 130 | defer reg.Close() |
| 131 | host := "absent.example" |
| 132 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), host, 3); err != ErrStopUnproven || proof != nil { |
| 133 | t.Fatal("absence became stop proof", err) |
| 134 | } |
| 135 | rt := testRoute(host) |
| 136 | if _, err := a.beginRequest(t.Context(), rt); err != ErrFenced { |
| 137 | t.Fatal("member wake passed", err) |
| 138 | } |
| 139 | if err := a.BeginSidecarRequest(t.Context(), rt); err != ErrFenced { |
| 140 | t.Fatal("sidecar passed", err) |
| 141 | } |
| 142 | for _, scheduled := range []bool{false, true} { |
| 143 | for _, stale := range []bool{false, true} { |
| 144 | for _, sidecar := range []bool{false, true} { |
| 145 | route := rt |
| 146 | if stale { |
| 147 | route.Hibernate = false |
| 148 | route.Kind = "service" |
| 149 | } |
| 150 | req := httptest.NewRequest("GET", "https://"+host+"/", nil) |
| 151 | req, answered := edge.WithAnswerFlag(req) |
| 152 | if sidecar { |
| 153 | req.Header.Set(SidecarHeader, "1") |
| 154 | } |
| 155 | out := httptest.NewRecorder() |
| 156 | proxied := false |
| 157 | proxy := http.HandlerFunc(func(http.ResponseWriter, *http.Request) { proxied = true }) |
| 158 | var handled bool |
| 159 | if scheduled { |
| 160 | handled = a.ServeScheduled(out, req, route, proxy) |
| 161 | } else { |
| 162 | handled = a.ServeHTTP(out, req, route, proxy) |
| 163 | } |
| 164 | if !handled || proxied || !*answered || out.Code != 503 || out.Header().Get("Cache-Control") != "no-store" { |
| 165 | t.Fatal("fence bypassed by stale or scheduled route") |
| 166 | } |
| 167 | } |
| 168 | } |
| 169 | } |
| 170 | if _, err := a.FenceAndStop(maintenanceContext(t, time.Second), host, 2); err != ErrFenceGeneration { |
| 171 | t.Fatal("older generation accepted") |
| 172 | } |
| 173 | // A new process has no closure evidence from the old activator's lifetime. |
| 174 | fresh := New(a.cfg, a.backend, reg) |
| 175 | if proof, err := fresh.FenceAndStop(maintenanceContext(t, time.Second), host, 3); err != ErrStopUnproven || proof != nil { |
| 176 | t.Fatal("restart invented old evidence") |
| 177 | } |
| 178 | } |
| 179 | func TestMaintenanceFenceContextAndGenerationDoNotRestoreAdmission(t *testing.T) { |
| 180 | a, e, main, _, _ := maintenanceAwake(t) |
| 181 | gate := make(chan struct{}) |
| 182 | main.gate = gate |
| 183 | main.entered = make(chan struct{}) |
| 184 | first := make(chan error, 1) |
| 185 | go func() { _, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1); first <- err }() |
| 186 | <-main.entered |
| 187 | if proof, err := a.FenceAndStop(maintenanceContext(t, 15*time.Millisecond), e.host, 2); err != ErrStopUnproven || proof != nil { |
| 188 | t.Fatal("timed-out stop claimed proof") |
| 189 | } |
| 190 | if _, err := a.beginRequest(t.Context(), e.route); err != ErrFenced { |
| 191 | t.Fatal("timeout reopened admission") |
| 192 | } |
| 193 | close(gate) |
| 194 | if err := <-first; err != ErrFenceGeneration { |
| 195 | t.Fatal("old waiter survived newer fence", err) |
| 196 | } |
| 197 | proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 2) |
| 198 | if err != nil || !proof.Matches(e.host, 2) { |
| 199 | t.Fatal("checked outcome lost after timeout", err) |
| 200 | } |
| 201 | for _, ctx := range []context.Context{nil, context.Background(), maintenanceContext(t, -time.Second)} { |
| 202 | if p, err := a.FenceAndStop(ctx, e.host, 2); err == nil || p != nil { |
| 203 | t.Fatal("unbounded or expired stop accepted") |
| 204 | } |
| 205 | } |
| 206 | a.mu.Lock() |
| 207 | started := time.Now() |
| 208 | if _, err := a.FenceAndStop(maintenanceContext(t, 15*time.Millisecond), e.host, 2); err != ErrFenceNotInstalled { |
| 209 | t.Fatal(err) |
| 210 | } |
| 211 | a.mu.Unlock() |
| 212 | if time.Since(started) > 250*time.Millisecond { |
| 213 | t.Fatal("fence mutex escaped deadline") |
| 214 | } |
| 215 | } |
| 216 | func TestMaintenanceOrdinaryStopFailureRemainsObservable(t *testing.T) { |
| 217 | t.Parallel() |
| 218 | a, e, main, _, _ := maintenanceAwake(t) |
| 219 | main.err = errors.New("stop refused") |
| 220 | a.mu.Lock() |
| 221 | a.beginStopLocked(e, StopIdle) |
| 222 | done := e.stopped |
| 223 | a.mu.Unlock() |
| 224 | <-done |
| 225 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 4); err != ErrStopUnproven || proof != nil { |
| 226 | t.Fatal("ordinary stop erased failure", err) |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | func TestMaintenanceLaterOrdinaryStopCannotEraseUnresolvedCleanup(t *testing.T) { |
| 231 | t.Parallel() |
| 232 | a, e, main, _, _ := maintenanceAwake(t) |
| 233 | main.err = errors.New("earlier process may still run") |
| 234 | a.mu.Lock() |
| 235 | a.beginStopLocked(e, StopIdle) |
| 236 | done := e.stopped |
| 237 | a.mu.Unlock() |
| 238 | <-done |
| 239 | replacement := &actEntry{host: e.host, route: e.route, state: actAwake, proc: &maintenanceProc{}, sidecar: &maintenanceProc{}, rep: &maintenanceRep{}} |
| 240 | a.mu.Lock() |
| 241 | a.entries[e.host] = replacement |
| 242 | a.beginStopLocked(replacement, StopIdle) |
| 243 | done = replacement.stopped |
| 244 | a.mu.Unlock() |
| 245 | <-done |
| 246 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1); err != ErrStopUnproven || proof != nil { |
| 247 | t.Fatal("new entry erased unresolved predecessor", err) |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | type maintenanceBackend struct { |
| 252 | *fakeBackend |
| 253 | main *maintenanceProc |
| 254 | side *maintenanceProc |
| 255 | rep *maintenanceRep |
| 256 | acquireGate <-chan struct{} |
| 257 | acquireFailure error |
| 258 | acquireEntered chan struct{} |
| 259 | replicateGate <-chan struct{} |
| 260 | replicateEntered chan struct{} |
| 261 | sideGate <-chan struct{} |
| 262 | sideEntered chan struct{} |
| 263 | startFailure error |
| 264 | sideFailure error |
| 265 | replicateFailure error |
| 266 | } |
| 267 | |
| 268 | func (b *maintenanceBackend) Acquire(ctx context.Context, rt registry.Route) (Lease, error) { |
| 269 | if b.acquireEntered != nil { |
| 270 | close(b.acquireEntered) |
| 271 | } |
| 272 | if b.acquireGate != nil { |
| 273 | <-b.acquireGate |
| 274 | } |
| 275 | lease, err := b.fakeBackend.Acquire(ctx, rt) |
| 276 | if err != nil { |
| 277 | return nil, err |
| 278 | } |
| 279 | return &maintenanceLease{Lease: lease, b: b}, b.acquireFailure |
| 280 | } |
| 281 | func (b *maintenanceBackend) Start(context.Context, registry.Route) (Proc, string, error) { |
| 282 | b.mu.Lock() |
| 283 | b.starts++ |
| 284 | b.mu.Unlock() |
| 285 | if b.startFailure != nil { |
| 286 | if b.main == nil { |
| 287 | return nil, "", b.startFailure |
| 288 | } |
| 289 | return b.main, "", b.startFailure |
| 290 | } |
| 291 | return b.main, "", nil |
| 292 | } |
| 293 | func (b *maintenanceBackend) StartSidecar(context.Context, registry.Route) (Proc, error) { |
| 294 | b.mu.Lock() |
| 295 | b.sidecarStarts++ |
| 296 | b.mu.Unlock() |
| 297 | if b.sideEntered != nil { |
| 298 | close(b.sideEntered) |
| 299 | } |
| 300 | if b.sideGate != nil { |
| 301 | <-b.sideGate |
| 302 | } |
| 303 | return b.side, b.sideFailure |
| 304 | } |
| 305 | |
| 306 | type maintenanceLease struct { |
| 307 | Lease |
| 308 | b *maintenanceBackend |
| 309 | } |
| 310 | |
| 311 | func (l *maintenanceLease) Replicate(context.Context, string, time.Duration) (Rep, error) { |
| 312 | if l.b.replicateEntered != nil { |
| 313 | close(l.b.replicateEntered) |
| 314 | } |
| 315 | if l.b.replicateGate != nil { |
| 316 | <-l.b.replicateGate |
| 317 | } |
| 318 | if l.b.rep == nil { |
| 319 | return nil, l.b.replicateFailure |
| 320 | } |
| 321 | return l.b.rep, l.b.replicateFailure |
| 322 | } |
| 323 | func TestMaintenanceFenceDuringWakeOwnsItsCleanup(t *testing.T) { |
| 324 | t.Parallel() |
| 325 | for _, phase := range []string{"acquire", "replicate"} { |
| 326 | t.Run(phase, func(t *testing.T) { |
| 327 | a, base, _, reg := testActivator(t, time.Hour) |
| 328 | defer reg.Close() |
| 329 | gate, entered := make(chan struct{}), make(chan struct{}) |
| 330 | b := &maintenanceBackend{fakeBackend: base, main: &maintenanceProc{}, rep: &maintenanceRep{}} |
| 331 | if phase == "acquire" { |
| 332 | b.acquireGate = gate |
| 333 | b.acquireEntered = entered |
| 334 | } else { |
| 335 | b.replicateGate = gate |
| 336 | b.replicateEntered = entered |
| 337 | } |
| 338 | a.backend = b |
| 339 | rt := testRoute("waking.example") |
| 340 | request := make(chan error, 1) |
| 341 | go func() { |
| 342 | release, err := a.beginRequest(t.Context(), rt) |
| 343 | if release != nil { |
| 344 | release() |
| 345 | } |
| 346 | request <- err |
| 347 | }() |
| 348 | <-entered |
| 349 | if proof, err := a.FenceAndStop(maintenanceContext(t, 15*time.Millisecond), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 350 | t.Fatal("in-flight wake yielded proof", err) |
| 351 | } |
| 352 | close(gate) |
| 353 | if err := <-request; err != ErrFenced { |
| 354 | t.Fatal("waking request admitted after fence", err) |
| 355 | } |
| 356 | proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1) |
| 357 | if phase == "acquire" { |
| 358 | if err != ErrStopUnproven || proof != nil { |
| 359 | t.Fatal("canceled pre-spawn wake invented closure", err) |
| 360 | } |
| 361 | if b.starts != 0 || b.restores != 0 { |
| 362 | t.Fatal("fenced acquisition continued startup") |
| 363 | } |
| 364 | } else { |
| 365 | if err != nil || !proof.Matches(rt.Host, 1) { |
| 366 | t.Fatal("started wake did not yield checked cleanup", err) |
| 367 | } |
| 368 | if b.main.calls.Load() != 1 || b.rep.calls.Load() != 1 { |
| 369 | t.Fatal("partial wake resources not closed") |
| 370 | } |
| 371 | } |
| 372 | }) |
| 373 | } |
| 374 | } |
| 375 | func TestMaintenanceFailedBackendStartRemainsUnproven(t *testing.T) { |
| 376 | t.Parallel() |
| 377 | a, base, _, reg := testActivator(t, time.Hour) |
| 378 | defer reg.Close() |
| 379 | a.backend = &maintenanceBackend{fakeBackend: base, startFailure: errors.New("hidden backend cleanup failed")} |
| 380 | rt := testRoute("hidden.example") |
| 381 | if _, err := a.beginRequest(t.Context(), rt); err == nil { |
| 382 | t.Fatal("failed backend started") |
| 383 | } |
| 384 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 385 | t.Fatal("hidden cleanup failure became proof", err) |
| 386 | } |
| 387 | } |
| 388 | |
| 389 | // Before design §6.7 this restart ran synchronously under a.mu, so a fence |
| 390 | // attempt that raced a stalled StartSidecar found the lock held and came |
| 391 | // back ErrFenceNotInstalled without ever touching the entry. Moving the |
| 392 | // call off a.mu means a stalled restart can't block anything else, |
| 393 | // including FenceAndStop: it now installs its fence immediately. What |
| 394 | // still bounds a short-context attempt is stop()'s own wait |
| 395 | // (sidecarStopTimeout) for the stalled restart to land before it can act |
| 396 | // on e.sidecar, so the outcome changes (ErrStopUnproven, fence installed) |
| 397 | // but a stalled sidecar still keeps a checked stop from completing |
| 398 | // instantly, which is the property this test exists to cover. |
| 399 | func TestMaintenanceFenceSerializesSidecarSweepReplacement(t *testing.T) { |
| 400 | t.Parallel() |
| 401 | a, e, _, side, rep := maintenanceAwake(t) |
| 402 | gate, entered := make(chan struct{}), make(chan struct{}) |
| 403 | base := a.backend.(*fakeBackend) |
| 404 | base.sidecarStaged = true |
| 405 | next := &maintenanceProc{} |
| 406 | a.backend = &maintenanceBackend{fakeBackend: base, side: next, sideGate: gate, sideEntered: entered, rep: rep} |
| 407 | side.exited.Store(true) |
| 408 | e.sidecarWanted = true |
| 409 | // sweepOnce itself now returns immediately. The restart it launches |
| 410 | // runs in its own goroutine, which is what's actually gated below. |
| 411 | a.sweepOnce() |
| 412 | <-entered |
| 413 | |
| 414 | if proof, err := a.FenceAndStop(maintenanceContext(t, 15*time.Millisecond), e.host, 1); err != ErrStopUnproven || proof != nil { |
| 415 | t.Fatal("blocked replacement yielded proof", err) |
| 416 | } |
| 417 | if a.wakeAllowed(e.host) { |
| 418 | t.Fatal("fenced admission attempt failed to install its fence") |
| 419 | } |
| 420 | close(gate) |
| 421 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1); err != nil || !proof.Matches(e.host, 1) { |
| 422 | t.Fatal("replacement escaped checked stop", err) |
| 423 | } |
| 424 | waitFor(t, "the discarded replacement to be stopped", func() bool { return next.calls.Load() == 1 }) |
| 425 | before := base.sidecarStartCount() |
| 426 | a.sweepOnce() |
| 427 | if base.sidecarStartCount() != before { |
| 428 | t.Fatal("fenced sweep restarted sidecar") |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | type maintenanceDue struct { |
| 433 | DueStore |
| 434 | route registry.Route |
| 435 | entered chan struct{} |
| 436 | gate <-chan struct{} |
| 437 | } |
| 438 | |
| 439 | func (d *maintenanceDue) DueInstances(int64) ([]registry.Route, error) { |
| 440 | close(d.entered) |
| 441 | <-d.gate |
| 442 | return []registry.Route{d.route}, nil |
| 443 | } |
| 444 | func TestMaintenanceFenceStopsStaleScheduledWake(t *testing.T) { |
| 445 | t.Parallel() |
| 446 | a, b, _, reg := testActivator(t, time.Hour) |
| 447 | defer reg.Close() |
| 448 | rt := testRoute("scheduled.example") |
| 449 | rt.Status = "ready" |
| 450 | gate, entered := make(chan struct{}), make(chan struct{}) |
| 451 | a.due = &maintenanceDue{DueStore: reg, route: rt, entered: entered, gate: gate} |
| 452 | swept := make(chan struct{}) |
| 453 | go func() { a.sweepOnce(); close(swept) }() |
| 454 | <-entered |
| 455 | if _, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven { |
| 456 | t.Fatal(err) |
| 457 | } |
| 458 | close(gate) |
| 459 | <-swept |
| 460 | b.mu.Lock() |
| 461 | starts, acquires := b.starts, b.acquires |
| 462 | b.mu.Unlock() |
| 463 | if starts != 0 || acquires != 0 || a.entryCount() != 0 { |
| 464 | t.Fatal("stale due query crossed installed fence") |
| 465 | } |
| 466 | } |
| 467 | func TestMaintenanceAwakeFenceSuppressesSidecarAdoption(t *testing.T) { |
| 468 | t.Parallel() |
| 469 | a, e, _, side, _ := maintenanceAwake(t) |
| 470 | base := a.backend.(*fakeBackend) |
| 471 | base.sidecarStaged = true |
| 472 | side.exited.Store(true) |
| 473 | a.fences = map[string]uint64{e.host: 1} |
| 474 | a.sweepOnce() |
| 475 | if base.sidecarStartCount() != 0 { |
| 476 | t.Fatal("awake fence allowed sidecar adoption") |
| 477 | } |
| 478 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1); err != nil || proof == nil { |
| 479 | t.Fatal(err) |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | func TestMaintenanceFailedStartCleansReturnedHandleBeforeRelease(t *testing.T) { |
| 484 | t.Parallel() |
| 485 | a, base, _, reg := testActivator(t, time.Hour) |
| 486 | defer reg.Close() |
| 487 | gate, entered := make(chan struct{}), make(chan struct{}) |
| 488 | proc := &maintenanceProc{gate: gate, entered: entered} |
| 489 | a.backend = &maintenanceBackend{fakeBackend: base, main: proc, startFailure: errors.New("started then failed")} |
| 490 | rt := testRoute("partial.example") |
| 491 | done := make(chan error, 1) |
| 492 | go func() { _, err := a.beginRequest(t.Context(), rt); done <- err }() |
| 493 | select { |
| 494 | case <-entered: |
| 495 | case <-time.After(3 * time.Second): |
| 496 | t.Fatal("returned process cleanup was never attempted") |
| 497 | } |
| 498 | base.mu.Lock() |
| 499 | released := base.releases |
| 500 | base.mu.Unlock() |
| 501 | if released != 0 { |
| 502 | t.Fatal("lease released before returned process cleanup") |
| 503 | } |
| 504 | close(gate) |
| 505 | if err := <-done; err == nil { |
| 506 | t.Fatal("partial failed start admitted") |
| 507 | } |
| 508 | if proc.calls.Load() != 1 || !proc.Exited() { |
| 509 | t.Fatal("returned process handle was abandoned") |
| 510 | } |
| 511 | base.mu.Lock() |
| 512 | released = base.releases |
| 513 | base.mu.Unlock() |
| 514 | if released != 1 { |
| 515 | t.Fatal("lease release missing or repeated", released) |
| 516 | } |
| 517 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 518 | t.Fatal("partial failure erased uncertainty", err) |
| 519 | } |
| 520 | } |
| 521 | func TestMaintenanceFailedSidecarHandlesRemainOwned(t *testing.T) { |
| 522 | t.Parallel() |
| 523 | for _, phase := range []string{"wake", "sweep"} { |
| 524 | t.Run(phase, func(t *testing.T) { |
| 525 | a, base, _, reg := testActivator(t, time.Hour) |
| 526 | defer reg.Close() |
| 527 | base.sidecarStaged = true |
| 528 | main, partial, rep := &maintenanceProc{}, &maintenanceProc{}, &maintenanceRep{} |
| 529 | b := &maintenanceBackend{fakeBackend: base, main: main, side: partial, rep: rep, sideFailure: errors.New("sidecar started then failed")} |
| 530 | a.backend = b |
| 531 | rt := testRoute("partial-sidecar.example") |
| 532 | if phase == "wake" { |
| 533 | release, err := a.beginRequest(t.Context(), rt) |
| 534 | if err != nil { |
| 535 | t.Fatal(err) |
| 536 | } |
| 537 | release() |
| 538 | } else { |
| 539 | old := &maintenanceProc{} |
| 540 | old.exited.Store(true) |
| 541 | a.entries[rt.Host] = &actEntry{host: rt.Host, route: rt, state: actAwake, proc: main, rep: rep, sidecar: old, sidecarWanted: true, lastActive: a.now()} |
| 542 | a.sweepOnce() |
| 543 | // The restart call runs off a.mu (design §6.7); give it a |
| 544 | // moment to land before inspecting what it installed. |
| 545 | waitFor(t, "sidecar restart to land", func() bool { return base.sidecarStartCount() == 1 && a.sidecarSettled(rt.Host) }) |
| 546 | } |
| 547 | a.mu.Lock() |
| 548 | entry := a.entries[rt.Host] |
| 549 | owned := entry.sidecar == partial && entry.sidecarUncertain |
| 550 | a.mu.Unlock() |
| 551 | if !owned { |
| 552 | t.Fatal("failed sidecar handle was discarded") |
| 553 | } |
| 554 | before := base.sidecarStartCount() |
| 555 | a.sweepOnce() |
| 556 | if base.sidecarStartCount() != before { |
| 557 | t.Fatal("live partial sidecar was replaced") |
| 558 | } |
| 559 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 560 | t.Fatal("failed sidecar yielded trusted evidence", err) |
| 561 | } |
| 562 | if partial.calls.Load() != 1 || !partial.Exited() { |
| 563 | t.Fatal("checked stop did not own partial sidecar") |
| 564 | } |
| 565 | }) |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | func TestMaintenanceFailedReplicationClosesReturnedReplicaWithoutDoubleRelease(t *testing.T) { |
| 570 | t.Parallel() |
| 571 | for _, closeFails := range []bool{false, true} { |
| 572 | t.Run(fmt.Sprint(closeFails), func(t *testing.T) { |
| 573 | a, base, _, reg := testActivator(t, time.Hour) |
| 574 | defer reg.Close() |
| 575 | mainGate, mainEntered := make(chan struct{}), make(chan struct{}) |
| 576 | repGate, repEntered := make(chan struct{}), make(chan struct{}) |
| 577 | main := &maintenanceProc{gate: mainGate, entered: mainEntered} |
| 578 | rep := &maintenanceRep{gate: repGate, entered: repEntered} |
| 579 | if closeFails { |
| 580 | rep.err = errors.New("replica close failed") |
| 581 | } |
| 582 | a.backend = &maintenanceBackend{fakeBackend: base, main: main, rep: rep, replicateFailure: errors.New("replica opened then failed")} |
| 583 | rt := testRoute("partial-replica.example") |
| 584 | done := make(chan error, 1) |
| 585 | go func() { _, err := a.beginRequest(t.Context(), rt); done <- err }() |
| 586 | select { |
| 587 | case <-mainEntered: |
| 588 | case <-time.After(3 * time.Second): |
| 589 | t.Fatal("returned main cleanup missing") |
| 590 | } |
| 591 | if rep.calls.Load() != 0 { |
| 592 | t.Fatal("replica closed before main stopped") |
| 593 | } |
| 594 | close(mainGate) |
| 595 | select { |
| 596 | case <-repEntered: |
| 597 | case <-time.After(3 * time.Second): |
| 598 | t.Fatal("returned replica cleanup missing") |
| 599 | } |
| 600 | if !main.Exited() { |
| 601 | t.Fatal("replica closed before main exited") |
| 602 | } |
| 603 | base.mu.Lock() |
| 604 | released := base.releases |
| 605 | base.mu.Unlock() |
| 606 | if released != 0 { |
| 607 | t.Fatal("lease released while replica close pending") |
| 608 | } |
| 609 | close(repGate) |
| 610 | if err := <-done; err == nil { |
| 611 | t.Fatal("partial replica failure admitted") |
| 612 | } |
| 613 | base.mu.Lock() |
| 614 | released = base.releases |
| 615 | base.mu.Unlock() |
| 616 | if released != 0 || rep.calls.Load() != 1 { |
| 617 | t.Fatal("returned replica lease released twice", released, rep.calls.Load()) |
| 618 | } |
| 619 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 620 | t.Fatal("partial replica uncertainty erased", err) |
| 621 | } |
| 622 | }) |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | func TestMaintenanceFailedAcquireReleasesReturnedLease(t *testing.T) { |
| 627 | t.Parallel() |
| 628 | a, base, _, reg := testActivator(t, time.Hour) |
| 629 | defer reg.Close() |
| 630 | a.backend = &maintenanceBackend{fakeBackend: base, acquireFailure: errors.New("lease returned with acquisition failure")} |
| 631 | rt := testRoute("partial-lease.example") |
| 632 | if _, err := a.beginRequest(t.Context(), rt); err == nil { |
| 633 | t.Fatal("failed acquisition admitted") |
| 634 | } |
| 635 | base.mu.Lock() |
| 636 | released, starts := base.releases, base.starts |
| 637 | base.mu.Unlock() |
| 638 | if released != 1 || starts != 0 { |
| 639 | t.Fatal("returned lease abandoned or wake continued", released, starts) |
| 640 | } |
| 641 | if proof, err := a.FenceAndStop(maintenanceContext(t, time.Second), rt.Host, 1); err != ErrStopUnproven || proof != nil { |
| 642 | t.Fatal("failed acquisition became stop proof", err) |
| 643 | } |
| 644 | } |
| 645 | |
| 646 | // TestFencedReportsAnInstalledFenceUntilItIsLifted: Fenced is the layout |
| 647 | // pass's own affirmative read of what FenceAndStop and LiftFence leave |
| 648 | // behind, so it must agree with them at every step rather than invert |
| 649 | // wakeAllowed's private sense of the same map. |
| 650 | func TestFencedReportsAnInstalledFenceUntilItIsLifted(t *testing.T) { |
| 651 | a, e, _, _, _ := maintenanceAwake(t) |
| 652 | if a.Fenced(e.host) { |
| 653 | t.Fatal("fence reported before one was installed") |
| 654 | } |
| 655 | if _, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 1); err != nil { |
| 656 | t.Fatal(err) |
| 657 | } |
| 658 | if !a.Fenced(e.host) { |
| 659 | t.Fatal("fence not reported as installed") |
| 660 | } |
| 661 | if !a.LiftFence(e.host, 1) { |
| 662 | t.Fatal("LiftFence reported nothing to release") |
| 663 | } |
| 664 | if a.Fenced(e.host) { |
| 665 | t.Fatal("fence still reported after it was released") |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | // TestFenceGenerationReportsTheHeldGeneration: the layout pass tells its |
| 670 | // own kept fence from another operation's by the generation, so the read |
| 671 | // must say exactly which generation holds, and 0 once none does. |
| 672 | func TestFenceGenerationReportsTheHeldGeneration(t *testing.T) { |
| 673 | a, e, _, _, _ := maintenanceAwake(t) |
| 674 | if g := a.FenceGeneration(e.host); g != 0 { |
| 675 | t.Fatalf("generation %d before any fence", g) |
| 676 | } |
| 677 | if _, err := a.FenceAndStop(maintenanceContext(t, time.Second), e.host, 9); err != nil { |
| 678 | t.Fatal(err) |
| 679 | } |
| 680 | if g := a.FenceGeneration(e.host); g != 9 { |
| 681 | t.Fatalf("generation = %d, want 9", g) |
| 682 | } |
| 683 | a.LiftFence(e.host, 9) |
| 684 | if g := a.FenceGeneration(e.host); g != 0 { |
| 685 | t.Fatalf("generation %d after the lift", g) |
| 686 | } |
| 687 | } |
| 688 | |
| 689 | // TestStopUncertainCoversLiveEntriesAndIncompleteHistory: a pass deciding |
| 690 | // whether ErrStopUnproven can be treated as a completed stop needs both |
| 691 | // halves of this: an entry that is still live right now (waking, awake or |
| 692 | // stopping), and a past stop observation that never finished every check. |
| 693 | // It must answer false for a host it has never seen at all, since that is |
| 694 | // not evidence of anything. |
| 695 | func TestStopUncertainCoversLiveEntriesAndIncompleteHistory(t *testing.T) { |
| 696 | t.Parallel() |
| 697 | a, e, main, _, _ := maintenanceAwake(t) |
| 698 | |
| 699 | if !a.StopUncertain(e.host) { |
| 700 | t.Fatal("a live entry must read as stop-uncertain") |
| 701 | } |
| 702 | if a.StopUncertain("never-seen.example") { |
| 703 | t.Fatal("a host with no entry and no history must not read as uncertain") |
| 704 | } |
| 705 | |
| 706 | // Drive the entry through an ordinary stop that fails, which removes the |
| 707 | // entry but leaves an incomplete observation in stopHistory. |
| 708 | main.err = errors.New("stop refused") |
| 709 | a.mu.Lock() |
| 710 | a.beginStopLocked(e, StopIdle) |
| 711 | done := e.stopped |
| 712 | a.mu.Unlock() |
| 713 | <-done |
| 714 | if a.Owns(e.host) { |
| 715 | t.Fatal("entry should have been forgotten once the stop finished") |
| 716 | } |
| 717 | if !a.StopUncertain(e.host) { |
| 718 | t.Fatal("an incomplete stop observation must read as uncertain") |
| 719 | } |
| 720 | |
| 721 | // A different host's clean stop completes its observation and leaves no |
| 722 | // live entry either. StopUncertain must read false for it, proving the |
| 723 | // predicate is per-host rather than sticky once anything has failed. |
| 724 | clean, ce, _, _, _ := maintenanceAwake(t) |
| 725 | if proof, err := clean.FenceAndStop(maintenanceContext(t, time.Second), ce.host, 1); err != nil || !proof.Matches(ce.host, 1) { |
| 726 | t.Fatalf("checked clean stop: %v", err) |
| 727 | } |
| 728 | if clean.StopUncertain(ce.host) { |
| 729 | t.Fatal("a complete stop observation with no live entry must not read as uncertain") |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | // TestFenceAndStopIfUnfencedRefusesAnyHeldFence: the layout pass shares the |
| 734 | // one fence slot per host with the residence move, and FenceAndStop would |
| 735 | // replace a lower generation. The exclusive variant decides under the same |
| 736 | // lock that installs, so a fence of any generation, lower, equal or |
| 737 | // higher, refuses it and stays exactly as it was, and nothing is stopped. |
| 738 | func TestFenceAndStopIfUnfencedRefusesAnyHeldFence(t *testing.T) { |
| 739 | t.Parallel() |
| 740 | for _, held := range []uint64{5, 10, 20} { |
| 741 | t.Run(fmt.Sprint(held), func(t *testing.T) { |
| 742 | a, e, main, _, _ := maintenanceAwake(t) |
| 743 | a.mu.Lock() |
| 744 | a.fences = map[string]uint64{e.host: held} |
| 745 | a.mu.Unlock() |
| 746 | _, err := a.FenceAndStopIfUnfenced(maintenanceContext(t, time.Second), e.host, 10) |
| 747 | if !errors.Is(err, ErrFenceHeld) { |
| 748 | t.Fatalf("err = %v, want ErrFenceHeld", err) |
| 749 | } |
| 750 | a.mu.Lock() |
| 751 | got := a.fences[e.host] |
| 752 | a.mu.Unlock() |
| 753 | if got != held { |
| 754 | t.Fatalf("held fence is %d, want %d", got, held) |
| 755 | } |
| 756 | if main.calls.Load() != 0 || !a.Owns(e.host) { |
| 757 | t.Fatal("a refused exclusive fence stopped the host") |
| 758 | } |
| 759 | }) |
| 760 | } |
| 761 | } |
| 762 | |
| 763 | // TestFenceAndStopIfUnfencedIsFenceAndStopOnAnUnfencedHost: with no fence |
| 764 | // held it installs, stops and proves exactly as FenceAndStop does. |
| 765 | func TestFenceAndStopIfUnfencedIsFenceAndStopOnAnUnfencedHost(t *testing.T) { |
| 766 | t.Parallel() |
| 767 | a, e, _, _, _ := maintenanceAwake(t) |
| 768 | proof, err := a.FenceAndStopIfUnfenced(maintenanceContext(t, time.Second), e.host, 7) |
| 769 | if err != nil || !proof.Matches(e.host, 7) { |
| 770 | t.Fatalf("exclusive fence on an unfenced awake host: %v", err) |
| 771 | } |
| 772 | if !a.Fenced(e.host) || a.Owns(e.host) { |
| 773 | t.Fatal("not fenced and stopped") |
| 774 | } |
| 775 | // An entry-less host with no stop on record answers ErrStopUnproven |
| 776 | // with its fence installed, as FenceAndStop does. |
| 777 | if _, err := a.FenceAndStopIfUnfenced(maintenanceContext(t, time.Second), "idle.example", 3); !errors.Is(err, ErrStopUnproven) { |
| 778 | t.Fatalf("entry-less host: %v", err) |
| 779 | } |
| 780 | if !a.LiftFence("idle.example", 3) { |
| 781 | t.Fatal("the unproven exclusive fence was not installed at its generation") |
| 782 | } |
| 783 | } |
| 784 | |
amadan