| 1 | //go:build linux |
| 2 | |
| 3 | package layout |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "errors" |
| 8 | "net/http" |
| 9 | "net/http/httptest" |
| 10 | "os" |
| 11 | "path/filepath" |
| 12 | "reflect" |
| 13 | "strings" |
| 14 | "sync" |
| 15 | "testing" |
| 16 | "time" |
| 17 | |
| 18 | "github.com/carlosframework/platform/internal/activator" |
| 19 | "github.com/carlosframework/platform/internal/provision" |
| 20 | "github.com/carlosframework/platform/internal/registry" |
| 21 | ) |
| 22 | |
| 23 | const ( |
| 24 | host = "h.example.com" |
| 25 | replica = "r.example.com" |
| 26 | ) |
| 27 | |
| 28 | // fakeFence is the activator's fence as the pass sees it. A fence already |
| 29 | // in fences, of any generation, refuses a new one with ErrFenceHeld, as the |
| 30 | // real exclusive install does. Otherwise every stop is proven; unproven, |
| 31 | // awake and uncertain make it answer ErrStopUnproven and shape what |
| 32 | // Quiesced and StopUncertain say after it. events records every fence, |
| 33 | // refusal and lift, and box's Reload adds "reload", so a test can read the |
| 34 | // order the pass did things in. |
| 35 | type fakeFence struct { |
| 36 | fences map[string]uint64 |
| 37 | unproven map[string]bool |
| 38 | awake map[string]bool |
| 39 | uncertain map[string]bool |
| 40 | onFence func(host string) |
| 41 | events []string |
| 42 | } |
| 43 | |
| 44 | func (f *fakeFence) FenceAndStopIfUnfenced(_ context.Context, h string, gen uint64) (*activator.StopEvidence, error) { |
| 45 | if f.fences == nil { |
| 46 | f.fences = map[string]uint64{} |
| 47 | } |
| 48 | if f.fences[h] != 0 { |
| 49 | f.events = append(f.events, "held "+h) |
| 50 | return nil, activator.ErrFenceHeld |
| 51 | } |
| 52 | f.fences[h] = gen |
| 53 | f.events = append(f.events, "fence "+h) |
| 54 | if f.onFence != nil { |
| 55 | f.onFence(h) |
| 56 | } |
| 57 | if f.unproven[h] { |
| 58 | return nil, activator.ErrStopUnproven |
| 59 | } |
| 60 | return &activator.StopEvidence{}, nil |
| 61 | } |
| 62 | |
| 63 | func (f *fakeFence) LiftFence(h string, gen uint64) bool { |
| 64 | if f.fences[h] != gen { |
| 65 | return false |
| 66 | } |
| 67 | delete(f.fences, h) |
| 68 | f.events = append(f.events, "lift "+h) |
| 69 | return true |
| 70 | } |
| 71 | |
| 72 | func (f *fakeFence) FenceGeneration(h string) uint64 { return f.fences[h] } |
| 73 | func (f *fakeFence) Quiesced(h string) bool { return !f.awake[h] } |
| 74 | func (f *fakeFence) StopUncertain(h string) bool { return f.uncertain[h] } |
| 75 | |
| 76 | func openLayoutRegistry(t *testing.T, dir string) *registry.Registry { |
| 77 | t.Helper() |
| 78 | reg, err := registry.Open(filepath.Join(dir, "registry.db")) |
| 79 | if err != nil { |
| 80 | t.Fatal(err) |
| 81 | } |
| 82 | t.Cleanup(func() { reg.Close() }) |
| 83 | return reg |
| 84 | } |
| 85 | |
| 86 | // seedFlat writes host's flat files: the database and its wal, the admin |
| 87 | // env, a litestream shadow directory with a file in it, and (for the |
| 88 | // primary host only) a quarantine. The sidecar env and the rendered bundle |
| 89 | // are absent, so the journal records them absent. |
| 90 | func seedFlat(t *testing.T, dd, h string) { |
| 91 | t.Helper() |
| 92 | db := provision.DefaultDBPath(dd, h) |
| 93 | files := map[string]string{ |
| 94 | db: "db " + h, |
| 95 | db + "-wal": "wal", |
| 96 | filepath.Join(dd, h+".env"): "CARLOS_ADMIN_TOKEN=t\n", |
| 97 | filepath.Join(dd, shadowName(h), "generation"): "g", |
| 98 | } |
| 99 | if h == host { |
| 100 | files[db+".corrupt"] = "quarantined" |
| 101 | } |
| 102 | for path, body := range files { |
| 103 | if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil { |
| 104 | t.Fatal(err) |
| 105 | } |
| 106 | if err := os.WriteFile(path, []byte(body), 0o600); err != nil { |
| 107 | t.Fatal(err) |
| 108 | } |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | // box is a temp data dir, a temp socket dir and a registry with two flat |
| 113 | // managed rows whose files exist: an ordinary instance and a replica |
| 114 | // (#162), which is its own host and migrates like any other. |
| 115 | func box(t *testing.T) (*Pass, *fakeFence) { |
| 116 | t.Helper() |
| 117 | dd := t.TempDir() |
| 118 | // The socket dir is short on purpose: t.TempDir embeds the test's name, |
| 119 | // and the nested control socket must fit a unix socket address, which |
| 120 | // Preflight enforces. |
| 121 | rd, err := os.MkdirTemp("", "ly") |
| 122 | if err != nil { |
| 123 | t.Fatal(err) |
| 124 | } |
| 125 | t.Cleanup(func() { os.RemoveAll(rd) }) |
| 126 | reg := openLayoutRegistry(t, dd) |
| 127 | for _, rt := range []registry.Route{ |
| 128 | {Host: host, Kind: "instance", Hibernate: true, Status: "ready", |
| 129 | Socket: filepath.Join(rd, host+".sock"), DBPath: provision.DefaultDBPath(dd, host)}, |
| 130 | {Host: replica, Kind: "instance", Hibernate: true, Replica: true, |
| 131 | Socket: filepath.Join(rd, replica+".sock"), DBPath: provision.DefaultDBPath(dd, replica)}, |
| 132 | } { |
| 133 | if err := reg.Upsert(rt); err != nil { |
| 134 | t.Fatal(err) |
| 135 | } |
| 136 | seedFlat(t, dd, rt.Host) |
| 137 | } |
| 138 | f := &fakeFence{} |
| 139 | p := &Pass{Registry: reg, Fence: f, DataDir: dd, SocketDir: rd, SocketsGID: -1, Now: time.Now, |
| 140 | SweptAtStart: true, |
| 141 | Reload: func() { f.events = append(f.events, "reload") }, |
| 142 | Logf: t.Logf} |
| 143 | return p, f |
| 144 | } |
| 145 | |
| 146 | func byHost(t *testing.T, rs []Result) map[string]Result { |
| 147 | t.Helper() |
| 148 | out := map[string]Result{} |
| 149 | for _, r := range rs { |
| 150 | if _, dup := out[r.Host]; dup { |
| 151 | t.Errorf("host %q reported twice: %+v", r.Host, rs) |
| 152 | } |
| 153 | out[r.Host] = r |
| 154 | } |
| 155 | return out |
| 156 | } |
| 157 | |
| 158 | func endpoints(t *testing.T, p *Pass, h string, mode Mode) Endpoint { |
| 159 | t.Helper() |
| 160 | plan, err := PlanFor(p.DataDir, p.SocketDir, h, mode) |
| 161 | if err != nil { |
| 162 | t.Fatal(err) |
| 163 | } |
| 164 | return plan.To |
| 165 | } |
| 166 | |
| 167 | func rowAt(t *testing.T, p *Pass, h string) Endpoint { |
| 168 | t.Helper() |
| 169 | rt, err := p.Registry.Get(h) |
| 170 | if err != nil || rt == nil { |
| 171 | t.Fatalf("row %s: %v %v", h, rt, err) |
| 172 | } |
| 173 | return Endpoint{DB: rt.DBPath, Socket: rt.Socket} |
| 174 | } |
| 175 | |
| 176 | func assertFlatUntouched(t *testing.T, p *Pass, h string) { |
| 177 | t.Helper() |
| 178 | if got := rowAt(t, p, h); got != endpoints(t, p, h, ModeFlat) { |
| 179 | t.Errorf("%s row moved: %+v", h, got) |
| 180 | } |
| 181 | if b, err := os.ReadFile(provision.DefaultDBPath(p.DataDir, h)); err != nil || string(b) != "db "+h { |
| 182 | t.Errorf("%s flat db: %q %v", h, b, err) |
| 183 | } |
| 184 | if JournalPresent(p.DataDir, h) { |
| 185 | t.Errorf("%s has a journal", h) |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | func TestConvergeNestedMovesFilesFlipsTheRowAndRemovesTheJournal(t *testing.T) { |
| 190 | p, f := box(t) |
| 191 | // Step 9's order: routing reloads on the committed row with the journal |
| 192 | // gone, and only then is the host released. |
| 193 | var reloads []string |
| 194 | p.Reload = func() { |
| 195 | for _, h := range []string{host, replica} { |
| 196 | if f.fences[h] != 0 { |
| 197 | if JournalPresent(p.DataDir, h) { |
| 198 | t.Errorf("reload for %s ran with its journal still present", h) |
| 199 | } |
| 200 | reloads = append(reloads, h+" "+rowAt(t, p, h).DB) |
| 201 | } |
| 202 | } |
| 203 | } |
| 204 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 205 | for _, h := range []string{host, replica} { |
| 206 | if res[h].Outcome != "migrated" { |
| 207 | t.Fatalf("%s: %+v", h, res[h]) |
| 208 | } |
| 209 | want := endpoints(t, p, h, ModeNested) |
| 210 | if got := rowAt(t, p, h); got != want { |
| 211 | t.Fatalf("%s row = %+v, want %+v", h, got, want) |
| 212 | } |
| 213 | raw := rawInstanceDir(p.DataDir, h) |
| 214 | for _, name := range []string{h + ".db", h + ".db-wal", h + ".env", filepath.Join(shadowName(h), "generation")} { |
| 215 | if _, err := os.Stat(filepath.Join(raw, name)); err != nil { |
| 216 | t.Errorf("%s not at the raw destination: %v", name, err) |
| 217 | } |
| 218 | } |
| 219 | if _, err := os.Lstat(provision.DefaultDBPath(p.DataDir, h)); !errors.Is(err, os.ErrNotExist) { |
| 220 | t.Errorf("%s flat db still present", h) |
| 221 | } |
| 222 | if JournalPresent(p.DataDir, h) { |
| 223 | t.Errorf("%s journal left behind after a clean move", h) |
| 224 | } |
| 225 | if fi, err := os.Lstat(rawRunDir(p.SocketDir, h)); err != nil || !fi.IsDir() { |
| 226 | t.Errorf("%s run directory: %v %v", h, fi, err) |
| 227 | } |
| 228 | } |
| 229 | if _, err := os.Stat(filepath.Join(rawInstanceDir(p.DataDir, host), host+".db.corrupt")); err != nil { |
| 230 | t.Errorf("quarantine not moved: %v", err) |
| 231 | } |
| 232 | if len(f.fences) != 0 { |
| 233 | t.Errorf("fences left installed after a clean move: %v", f.fences) |
| 234 | } |
| 235 | wantReloads := []string{host + " " + endpoints(t, p, host, ModeNested).DB, replica + " " + endpoints(t, p, replica, ModeNested).DB} |
| 236 | if !reflect.DeepEqual(reloads, wantReloads) { |
| 237 | t.Errorf("reloads under the fence = %v, want %v", reloads, wantReloads) |
| 238 | } |
| 239 | // Idempotent: a second converge does nothing. |
| 240 | for h, r := range byHost(t, p.Converge(context.Background(), ModeNested)) { |
| 241 | if r.Outcome != "unchanged" { |
| 242 | t.Fatalf("second converge, %s: %+v", h, r) |
| 243 | } |
| 244 | } |
| 245 | // Revert restores the flat layout byte for byte. |
| 246 | p.Reload = nil |
| 247 | res = byHost(t, p.Converge(context.Background(), ModeFlat)) |
| 248 | for _, h := range []string{host, replica} { |
| 249 | if res[h].Outcome != "reverted" { |
| 250 | t.Fatalf("revert %s = %+v", h, res[h]) |
| 251 | } |
| 252 | assertFlatUntouched(t, p, h) |
| 253 | if _, err := os.Stat(filepath.Join(p.DataDir, shadowName(h), "generation")); err != nil { |
| 254 | t.Errorf("%s shadow not back: %v", h, err) |
| 255 | } |
| 256 | } |
| 257 | if len(f.fences) != 0 { |
| 258 | t.Errorf("fences left installed after a revert: %v", f.fences) |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | // An unproven stop is accepted only for a host with no live entry, no |
| 263 | // uncertain stop on record, and an agent whose startup unit sweep |
| 264 | // succeeded. Anything else defers the host, writes no journal, and lifts |
| 265 | // the fence this pass installed. |
| 266 | func TestConvergeTrustsAnUnprovenStopOnlyWhenNothingCanBeRunning(t *testing.T) { |
| 267 | for _, tc := range []struct { |
| 268 | name string |
| 269 | swept bool |
| 270 | awake bool |
| 271 | uncertain bool |
| 272 | want string |
| 273 | }{ |
| 274 | {"swept, quiesced, certain", true, false, false, "migrated"}, |
| 275 | {"no startup sweep", false, false, false, "deferred"}, |
| 276 | {"an entry is live", true, true, false, "deferred"}, |
| 277 | {"the last stop is uncertain", true, false, true, "deferred"}, |
| 278 | } { |
| 279 | t.Run(tc.name, func(t *testing.T) { |
| 280 | p, f := box(t) |
| 281 | p.SweptAtStart = tc.swept |
| 282 | f.unproven = map[string]bool{host: true} |
| 283 | f.awake = map[string]bool{host: tc.awake} |
| 284 | f.uncertain = map[string]bool{host: tc.uncertain} |
| 285 | r := byHost(t, p.Converge(context.Background(), ModeNested))[host] |
| 286 | if r.Outcome != tc.want { |
| 287 | t.Fatalf("result = %+v, want %s", r, tc.want) |
| 288 | } |
| 289 | if tc.want == "deferred" { |
| 290 | if !strings.Contains(r.Reason, "stop unproven") { |
| 291 | t.Errorf("reason = %q", r.Reason) |
| 292 | } |
| 293 | assertFlatUntouched(t, p, host) |
| 294 | } |
| 295 | if len(f.fences) != 0 { |
| 296 | t.Errorf("a layout fence was left installed: %v", f.fences) |
| 297 | } |
| 298 | }) |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | // wakeBackend is an activator.Backend with no processes, no S3 and no |
| 303 | // replication: just enough for a real Service to wake a host. |
| 304 | type wakeBackend struct { |
| 305 | mu sync.Mutex |
| 306 | started []registry.Route |
| 307 | } |
| 308 | |
| 309 | type nopLease struct{} |
| 310 | |
| 311 | func (nopLease) Replicate(context.Context, string, time.Duration) (activator.Rep, error) { |
| 312 | return nopRep{}, nil |
| 313 | } |
| 314 | func (nopLease) Release(context.Context) error { return nil } |
| 315 | |
| 316 | type nopRep struct{} |
| 317 | |
| 318 | func (nopRep) Err() error { return nil } |
| 319 | func (nopRep) Close(context.Context) error { return nil } |
| 320 | |
| 321 | type nopProc struct{} |
| 322 | |
| 323 | func (nopProc) Stop(context.Context) error { return nil } |
| 324 | func (nopProc) Exited() bool { return false } |
| 325 | |
| 326 | func (b *wakeBackend) Acquire(context.Context, registry.Route) (activator.Lease, error) { |
| 327 | return nopLease{}, nil |
| 328 | } |
| 329 | func (b *wakeBackend) Restore(context.Context, registry.Route) (bool, error) { return false, nil } |
| 330 | func (b *wakeBackend) Seed(context.Context, registry.Route) error { return nil } |
| 331 | func (b *wakeBackend) Start(_ context.Context, rt registry.Route) (activator.Proc, string, error) { |
| 332 | b.mu.Lock() |
| 333 | defer b.mu.Unlock() |
| 334 | b.started = append(b.started, rt) |
| 335 | return nopProc{}, "", nil |
| 336 | } |
| 337 | func (b *wakeBackend) NextDue(context.Context, registry.Route) (int64, error) { return 0, nil } |
| 338 | func (b *wakeBackend) HasSidecar(registry.Route) bool { return false } |
| 339 | func (b *wakeBackend) StartSidecar(context.Context, registry.Route) (activator.Proc, error) { |
| 340 | return nil, errors.New("wakeBackend: no sidecar in this fixture") |
| 341 | } |
| 342 | func (b *wakeBackend) StartWorker(context.Context, registry.Route, map[string]string) (activator.Proc, string, error) { |
| 343 | return nil, "", errors.New("wakeBackend: no worker in this fixture") |
| 344 | } |
| 345 | |
| 346 | func realService(t *testing.T, p *Pass) (*activator.Service, *wakeBackend) { |
| 347 | t.Helper() |
| 348 | b := &wakeBackend{} |
| 349 | svc := activator.New(activator.Config{Bucket: "test-bucket", Idle: time.Hour}, b, p.Registry) |
| 350 | // The wake re-reads the committed row, so the test sees which paths it |
| 351 | // actually woke on. |
| 352 | svc.Routes = p.Registry |
| 353 | t.Cleanup(func() { |
| 354 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 355 | defer cancel() |
| 356 | svc.Shutdown(ctx) |
| 357 | }) |
| 358 | p.Fence = svc |
| 359 | return svc, b |
| 360 | } |
| 361 | |
| 362 | // After an agent restart no asleep host has a stop on record, so the real |
| 363 | // activator answers ErrStopUnproven for every one of them. The pass must |
| 364 | // still migrate such a host, and must leave it wakeable. |
| 365 | func TestConvergeAgainstARealActivatorMigratesAnEntrylessHostAndItWakes(t *testing.T) { |
| 366 | p, _ := box(t) |
| 367 | svc, b := realService(t, p) |
| 368 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 369 | if res[host].Outcome != "migrated" { |
| 370 | t.Fatalf("result = %+v", res[host]) |
| 371 | } |
| 372 | if svc.Fenced(host) { |
| 373 | t.Fatal("the host is still fenced after the pass") |
| 374 | } |
| 375 | rt, err := p.Registry.Get(host) |
| 376 | if err != nil || rt == nil { |
| 377 | t.Fatal(err) |
| 378 | } |
| 379 | rec := httptest.NewRecorder() |
| 380 | req := httptest.NewRequest(http.MethodGet, "http://"+host+"/", nil) |
| 381 | ok := http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(http.StatusOK) }) |
| 382 | if !svc.ServeHTTP(rec, req, *rt, ok) || rec.Code != http.StatusOK { |
| 383 | t.Fatalf("wake after the pass: code %d", rec.Code) |
| 384 | } |
| 385 | b.mu.Lock() |
| 386 | defer b.mu.Unlock() |
| 387 | if len(b.started) != 1 || b.started[0].DBPath != endpoints(t, p, host, ModeNested).DB { |
| 388 | t.Fatalf("started = %+v", b.started) |
| 389 | } |
| 390 | } |
| 391 | |
| 392 | func TestConvergeAgainstARealActivatorDefersWithoutAStartupSweep(t *testing.T) { |
| 393 | p, _ := box(t) |
| 394 | svc, _ := realService(t, p) |
| 395 | p.SweptAtStart = false |
| 396 | for h, r := range byHost(t, p.Converge(context.Background(), ModeNested)) { |
| 397 | if r.Outcome != "deferred" || !strings.Contains(r.Reason, "stop unproven") { |
| 398 | t.Errorf("%s: %+v", h, r) |
| 399 | } |
| 400 | if svc.Fenced(h) { |
| 401 | t.Errorf("%s left fenced after a refused quiesce", h) |
| 402 | } |
| 403 | assertFlatUntouched(t, p, h) |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | // Another operation's fence (the residence move's cut) shares the |
| 408 | // activator's one fence slot per host. The pass must neither take it over |
| 409 | // nor lift it. |
| 410 | func TestConvergeDefersAHostSomeoneElseFenced(t *testing.T) { |
| 411 | p, _ := box(t) |
| 412 | svc, _ := realService(t, p) |
| 413 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 414 | defer cancel() |
| 415 | if _, err := svc.FenceAndStop(ctx, host, 5); err != nil && !errors.Is(err, activator.ErrStopUnproven) { |
| 416 | t.Fatalf("pre-installing the other fence: %v", err) |
| 417 | } |
| 418 | p.Generation = func() uint64 { return 1000 } |
| 419 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 420 | if res[host].Outcome != "deferred" || !strings.Contains(res[host].Reason, "fenced by another operation") { |
| 421 | t.Fatalf("fenced host: %+v", res[host]) |
| 422 | } |
| 423 | if res[replica].Outcome != "migrated" { |
| 424 | t.Fatalf("the other host: %+v", res[replica]) |
| 425 | } |
| 426 | assertFlatUntouched(t, p, host) |
| 427 | if !svc.LiftFence(host, 5) { |
| 428 | t.Fatal("the original fence is not in place at its own generation") |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | // On the serial-move leg a row that becomes eligible while its host is |
| 433 | // awake is left for a pass that finds it asleep: DeferAwake defers the |
| 434 | // host before any fence is asked for, so the move never stops a running |
| 435 | // tenant. A host already at the intent is unchanged, awake or not, and the |
| 436 | // same awake host is moved when DeferAwake is off. |
| 437 | func TestConvergeDefersAnAwakeHostBeforeFencingIt(t *testing.T) { |
| 438 | p, f := box(t) |
| 439 | p.DeferAwake = true |
| 440 | f.awake = map[string]bool{host: true} |
| 441 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 442 | if r := res[host]; r.Outcome != "deferred" || !strings.HasPrefix(r.Reason, "awake") { |
| 443 | t.Fatalf("awake host: %+v", r) |
| 444 | } |
| 445 | if res[replica].Outcome != "migrated" { |
| 446 | t.Fatalf("the asleep host: %+v", res[replica]) |
| 447 | } |
| 448 | for _, e := range f.events { |
| 449 | if e == "fence "+host || e == "held "+host { |
| 450 | t.Fatalf("the awake host was offered a fence: %v", f.events) |
| 451 | } |
| 452 | } |
| 453 | assertFlatUntouched(t, p, host) |
| 454 | |
| 455 | // The next pass finds the host asleep and moves it; the replica is awake |
| 456 | // now, but it is already nested. |
| 457 | f.awake = map[string]bool{replica: true} |
| 458 | res = byHost(t, p.Converge(context.Background(), ModeNested)) |
| 459 | if res[host].Outcome != "migrated" { |
| 460 | t.Fatalf("the host once asleep: %+v", res[host]) |
| 461 | } |
| 462 | if res[replica].Outcome != "unchanged" { |
| 463 | t.Fatalf("an awake host already at the intent: %+v", res[replica]) |
| 464 | } |
| 465 | |
| 466 | p.DeferAwake = false |
| 467 | f.awake = map[string]bool{host: true} |
| 468 | if r := byHost(t, p.Converge(context.Background(), ModeFlat))[host]; r.Outcome != "reverted" { |
| 469 | t.Fatalf("without DeferAwake the awake host: %+v", r) |
| 470 | } |
| 471 | } |
| 472 | |
| 473 | func TestConvergeDefersARowThatIsNotReady(t *testing.T) { |
| 474 | p, f := box(t) |
| 475 | if err := p.Registry.SetStatus(host, "moving"); err != nil { |
| 476 | t.Fatal(err) |
| 477 | } |
| 478 | r := byHost(t, p.Converge(context.Background(), ModeNested))[host] |
| 479 | if r.Outcome != "deferred" || !strings.Contains(r.Reason, "moving") { |
| 480 | t.Fatalf("result = %+v", r) |
| 481 | } |
| 482 | for _, e := range f.events { |
| 483 | if e == "fence "+host { |
| 484 | t.Fatal("a moving row was fenced") |
| 485 | } |
| 486 | } |
| 487 | assertFlatUntouched(t, p, host) |
| 488 | } |
| 489 | |
| 490 | // The row the pass classified can change before its fence lands: a move |
| 491 | // cut starting, an operator unflagging it. The pass re-reads it under the |
| 492 | // fence and refuses, writing nothing. |
| 493 | func TestConvergeRefusesARowThatChangedUnderTheFence(t *testing.T) { |
| 494 | for _, tc := range []struct { |
| 495 | name string |
| 496 | change func(reg *registry.Registry) error |
| 497 | }{ |
| 498 | {"status", func(reg *registry.Registry) error { return reg.SetStatus(host, "moving") }}, |
| 499 | {"unflagged", func(reg *registry.Registry) error { return reg.SetHibernate(host, false) }}, |
| 500 | } { |
| 501 | t.Run(tc.name, func(t *testing.T) { |
| 502 | p, f := box(t) |
| 503 | f.onFence = func(h string) { |
| 504 | if h == host { |
| 505 | if err := tc.change(p.Registry); err != nil { |
| 506 | t.Fatal(err) |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | r := byHost(t, p.Converge(context.Background(), ModeNested))[host] |
| 511 | if r.Outcome != "refused" { |
| 512 | t.Fatalf("result = %+v", r) |
| 513 | } |
| 514 | assertFlatUntouched(t, p, host) |
| 515 | if len(f.fences) != 0 { |
| 516 | t.Errorf("fence left installed: %v", f.fences) |
| 517 | } |
| 518 | }) |
| 519 | } |
| 520 | } |
| 521 | |
| 522 | // A box whose isolation probe is not ready does not start nested tenants, |
| 523 | // so it does not make any. |
| 524 | func TestConvergeNestedIsGatedOnIsolation(t *testing.T) { |
| 525 | p, f := box(t) |
| 526 | calls := 0 |
| 527 | p.NestedOK = func() error { calls++; return errors.New("probe: not ready") } |
| 528 | res := p.Converge(context.Background(), ModeNested) |
| 529 | if len(res) != 1 || res[0].Host != "" || res[0].Outcome != "refused" || res[0].Reason != "isolation not ready: probe: not ready" { |
| 530 | t.Fatalf("results = %+v", res) |
| 531 | } |
| 532 | if calls != 1 { |
| 533 | t.Errorf("NestedOK called %d times", calls) |
| 534 | } |
| 535 | if len(f.events) != 0 { |
| 536 | t.Errorf("the pass touched hosts: %v", f.events) |
| 537 | } |
| 538 | for _, h := range []string{host, replica} { |
| 539 | assertFlatUntouched(t, p, h) |
| 540 | } |
| 541 | // Reverting and recovering are never gated: they are how a box gets |
| 542 | // back to a layout it can serve. |
| 543 | p.NestedOK = nil |
| 544 | p.Converge(context.Background(), ModeNested) |
| 545 | p.NestedOK = func() error { return errors.New("probe: not ready") } |
| 546 | for h, r := range byHost(t, p.Converge(context.Background(), ModeFlat)) { |
| 547 | if r.Outcome != "reverted" { |
| 548 | t.Errorf("revert %s under a failing probe: %+v", h, r) |
| 549 | } |
| 550 | } |
| 551 | } |
| 552 | |
| 553 | func TestConvergeSkipsForeignAndUnmanagedRows(t *testing.T) { |
| 554 | p, _ := box(t) |
| 555 | if err := p.Registry.Upsert(registry.Route{Host: "custom.example.com", Kind: "instance", Hibernate: true, |
| 556 | Socket: filepath.Join(p.SocketDir, "custom.example.com.sock"), DBPath: "/data/sub/store.db"}); err != nil { |
| 557 | t.Fatal(err) |
| 558 | } |
| 559 | if err := p.Registry.Upsert(registry.Route{Host: "ext.example.com", Kind: "instance", Hibernate: false, |
| 560 | Socket: filepath.Join(p.SocketDir, "ext.example.com.sock"), DBPath: provision.DefaultDBPath(p.DataDir, "ext.example.com")}); err != nil { |
| 561 | t.Fatal(err) |
| 562 | } |
| 563 | out := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 564 | if out["custom.example.com"].Outcome != "skipped" || !strings.Contains(out["custom.example.com"].Reason, "foreign") { |
| 565 | t.Errorf("foreign row: %+v", out["custom.example.com"]) |
| 566 | } |
| 567 | if _, listed := out["ext.example.com"]; listed { |
| 568 | t.Errorf("an unmanaged row was considered: %+v", out["ext.example.com"]) |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | // A same-uid tenant can plant <data>/.instances/<iid> as a symlink before |
| 573 | // the pass arrives. The pass moves nothing through it, and since the |
| 574 | // symlink is the tenant's own doing the host is held fenced and down |
| 575 | // (review G1) rather than released to run flat again; once the link is |
| 576 | // gone the next pass moves it from the kept fence. |
| 577 | func TestConvergeRefusesASymlinkedInstanceDirectory(t *testing.T) { |
| 578 | p, f := box(t) |
| 579 | elsewhere := t.TempDir() |
| 580 | raw := rawInstanceDir(p.DataDir, host) |
| 581 | if err := os.MkdirAll(filepath.Dir(raw), 0o700); err != nil { |
| 582 | t.Fatal(err) |
| 583 | } |
| 584 | if err := os.Symlink(elsewhere, raw); err != nil { |
| 585 | t.Fatal(err) |
| 586 | } |
| 587 | r := byHost(t, p.Converge(context.Background(), ModeNested))[host] |
| 588 | if r.Outcome != "held" || !strings.Contains(r.Reason, "symlink") { |
| 589 | t.Fatalf("result = %+v", r) |
| 590 | } |
| 591 | assertFlatUntouched(t, p, host) |
| 592 | if entries, _ := os.ReadDir(elsewhere); len(entries) != 0 { |
| 593 | t.Errorf("files landed through the symlink: %v", entries) |
| 594 | } |
| 595 | if f.fences[host] == 0 { |
| 596 | t.Error("the fence was lifted; the host would wake flat again behind its own symlink") |
| 597 | } |
| 598 | os.Remove(raw) |
| 599 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "migrated" || f.fences[host] != 0 { |
| 600 | t.Fatalf("after the symlink is gone: %+v fence=%d", r, f.fences[host]) |
| 601 | } |
| 602 | } |
| 603 | |
| 604 | func TestSerializeWrapsJournalWriteToJournalRemoval(t *testing.T) { |
| 605 | p, _ := box(t) |
| 606 | var wrapped []string |
| 607 | p.Serialize = func(fn func()) { |
| 608 | for _, h := range []string{host, replica} { |
| 609 | if JournalPresent(p.DataDir, h) { |
| 610 | t.Errorf("%s journal present before the serialized section", h) |
| 611 | } |
| 612 | } |
| 613 | fn() |
| 614 | for _, h := range []string{host, replica} { |
| 615 | if JournalPresent(p.DataDir, h) { |
| 616 | t.Errorf("%s journal present after the serialized section", h) |
| 617 | } |
| 618 | } |
| 619 | wrapped = append(wrapped, "section") |
| 620 | } |
| 621 | p.Converge(context.Background(), ModeNested) |
| 622 | if len(wrapped) != 2 { |
| 623 | t.Fatalf("serialized sections = %d, want one per host", len(wrapped)) |
| 624 | } |
| 625 | } |
| 626 | |
| 627 | func TestRecoverCompletesEveryCrashBoundary(t *testing.T) { |
| 628 | // Drive the pass through a move with a seam that aborts after each |
| 629 | // step, then Recover, and assert the end state is the nested layout |
| 630 | // with no journal, every time. |
| 631 | for _, step := range []string{"after-journal", "mid-files", "after-files", "after-cas", "before-remove"} { |
| 632 | t.Run(step, func(t *testing.T) { |
| 633 | p, f := box(t) |
| 634 | p.abortAt = step |
| 635 | p.Converge(context.Background(), ModeNested) |
| 636 | p.abortAt = "" |
| 637 | for _, h := range []string{host, replica} { |
| 638 | if !JournalPresent(p.DataDir, h) { |
| 639 | t.Fatalf("%s: no journal after aborting at %s", h, step) |
| 640 | } |
| 641 | } |
| 642 | if len(f.fences) != 0 { |
| 643 | t.Errorf("an aborted move left its fence: %v", f.fences) |
| 644 | } |
| 645 | res := byHost(t, p.Recover(context.Background(), "")) |
| 646 | for _, h := range []string{host, replica} { |
| 647 | if res[h].Outcome != "recovered" { |
| 648 | t.Fatalf("recover %s = %+v", h, res[h]) |
| 649 | } |
| 650 | if got := rowAt(t, p, h); got != endpoints(t, p, h, ModeNested) { |
| 651 | t.Errorf("%s row not nested after recovery from %s: %+v", h, step, got) |
| 652 | } |
| 653 | if b, err := os.ReadFile(filepath.Join(rawInstanceDir(p.DataDir, h), h+".db")); err != nil || string(b) != "db "+h { |
| 654 | t.Errorf("%s db not at destination after recovery from %s: %q %v", h, step, b, err) |
| 655 | } |
| 656 | if JournalPresent(p.DataDir, h) { |
| 657 | t.Errorf("%s journal remains after recovery", h) |
| 658 | } |
| 659 | } |
| 660 | if len(f.fences) != 0 { |
| 661 | t.Errorf("recovery left a fence: %v", f.fences) |
| 662 | } |
| 663 | }) |
| 664 | } |
| 665 | } |
| 666 | |
| 667 | func TestRecoverRefusesAForgedJournalAndKeepsTheHostFenced(t *testing.T) { |
| 668 | p, _ := box(t) |
| 669 | j := &Journal{Host: host, Mode: ModeNested, |
| 670 | From: Endpoint{DB: "/data/console/console.db", Socket: filepath.Join(p.SocketDir, host+".sock")}, |
| 671 | To: endpoints(t, p, host, ModeNested), |
| 672 | Phase: PhaseFiles} |
| 673 | for _, n := range artifactNames { |
| 674 | j.Artifacts = append(j.Artifacts, JournalArtifact{Name: n, Present: true}) |
| 675 | } |
| 676 | if err := WriteJournal(p.DataDir, j); err != nil { |
| 677 | t.Fatal(err) |
| 678 | } |
| 679 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 680 | if r.Outcome != "refused" || !strings.Contains(r.Reason, "not acting") { |
| 681 | t.Fatalf("forged journal: %+v", r) |
| 682 | } |
| 683 | if !JournalPresent(p.DataDir, host) { |
| 684 | t.Error("forged journal removed; the host must stay fenced for an operator") |
| 685 | } |
| 686 | if _, err := os.Stat(filepath.Join(p.DataDir, "console")); err == nil { |
| 687 | t.Error("something touched the forged source") |
| 688 | } |
| 689 | // Converge reports the host once, as the recovery refused it. |
| 690 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 691 | if res[host].Outcome != "refused" { |
| 692 | t.Errorf("converge over a refused journal: %+v", res[host]) |
| 693 | } |
| 694 | } |
| 695 | |
| 696 | // A file at both ends of a move cannot come from a crash: the destination |
| 697 | // check cleared every destination before the journal was written, and a |
| 698 | // rename leaves a file at exactly one of them. Recovery refuses to choose. |
| 699 | func TestRecoverRefusesAnArtifactAtBothEnds(t *testing.T) { |
| 700 | p, _ := box(t) |
| 701 | p.abortAt = "after-journal" |
| 702 | p.Converge(context.Background(), ModeNested) |
| 703 | p.abortAt = "" |
| 704 | dst := filepath.Join(rawInstanceDir(p.DataDir, host), host+".db") |
| 705 | if err := os.WriteFile(dst, []byte("planted"), 0o600); err != nil { |
| 706 | t.Fatal(err) |
| 707 | } |
| 708 | r := byHost(t, p.Recover(context.Background(), ""))[host] |
| 709 | if r.Outcome != "refused" || !strings.Contains(r.Reason, "both") { |
| 710 | t.Fatalf("result = %+v", r) |
| 711 | } |
| 712 | if !JournalPresent(p.DataDir, host) { |
| 713 | t.Error("journal removed over an ambiguous artifact") |
| 714 | } |
| 715 | if b, _ := os.ReadFile(provision.DefaultDBPath(p.DataDir, host)); string(b) != "db "+host { |
| 716 | t.Error("the source database was touched") |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | // A journal at row says every file it records present is at its |
| 721 | // destination. Recovery checks that before it flips the row, because a row |
| 722 | // pointed at a missing database would have the next wake restore a second |
| 723 | // one there. |
| 724 | func TestRecoverAtRowRefusesAMissingDestinationBeforeTheCAS(t *testing.T) { |
| 725 | p, f := box(t) |
| 726 | p.abortAt = "after-files" |
| 727 | p.Converge(context.Background(), ModeNested) |
| 728 | p.abortAt = "" |
| 729 | j, found, err := ReadJournal(p.DataDir, host) |
| 730 | if err != nil || !found { |
| 731 | t.Fatalf("setup: %v %v", found, err) |
| 732 | } |
| 733 | if err := UpdatePhase(p.DataDir, j, PhaseRow); err != nil { |
| 734 | t.Fatal(err) |
| 735 | } |
| 736 | wal := filepath.Join(rawInstanceDir(p.DataDir, host), host+".db-wal") |
| 737 | if err := os.Remove(wal); err != nil { |
| 738 | t.Fatal(err) |
| 739 | } |
| 740 | f.events = nil |
| 741 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 742 | if r.Outcome != "refused" || !strings.Contains(r.Reason, wal) { |
| 743 | t.Fatalf("result = %+v, want a refusal naming %s", r, wal) |
| 744 | } |
| 745 | if got, found, _ := ReadJournal(p.DataDir, host); !found || got.Phase != PhaseRow { |
| 746 | t.Errorf("journal not kept at row: %+v", got) |
| 747 | } |
| 748 | if rowAt(t, p, host) != endpoints(t, p, host, ModeFlat) { |
| 749 | t.Errorf("the row flipped over a missing file: %+v", rowAt(t, p, host)) |
| 750 | } |
| 751 | } |
| 752 | |
| 753 | // Design §3.3: a journal is completed to its own endpoint, and the host is |
| 754 | // then converged to the fleet's intent before it is released. The fence |
| 755 | // is held from the recovery through the convergence: nothing lifts it in |
| 756 | // between, so the host never wakes at the journal's endpoint. |
| 757 | func TestConvergeRecoversAJournalBeforeConvergingToANewIntent(t *testing.T) { |
| 758 | p, f := box(t) |
| 759 | p.abortAt = "after-files" |
| 760 | p.Converge(context.Background(), ModeNested) |
| 761 | p.abortAt = "" |
| 762 | f.events = nil |
| 763 | // The intent changed to flat while the move was in flight: the journal |
| 764 | // completes to ITS endpoint (nested), then the host converges to flat. |
| 765 | res := byHost(t, p.Converge(context.Background(), ModeFlat)) |
| 766 | for _, h := range []string{host, replica} { |
| 767 | if res[h].Outcome != "recovered" { |
| 768 | t.Errorf("%s: %+v", h, res[h]) |
| 769 | } |
| 770 | assertFlatUntouched(t, p, h) |
| 771 | } |
| 772 | want := []string{ |
| 773 | "fence " + host, "reload", "reload", "lift " + host, |
| 774 | "fence " + replica, "reload", "reload", "lift " + replica, |
| 775 | } |
| 776 | if !reflect.DeepEqual(f.events, want) { |
| 777 | t.Fatalf("events = %v\nwant %v", f.events, want) |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | // With no intent recorded, a journal is completed and the host released |
| 782 | // where the journal ends. |
| 783 | func TestRecoverWithNoIntentReleasesAtTheJournalsEndpoint(t *testing.T) { |
| 784 | p, f := box(t) |
| 785 | p.abortAt = "after-files" |
| 786 | p.Converge(context.Background(), ModeNested) |
| 787 | p.abortAt = "" |
| 788 | for h, r := range byHost(t, p.Recover(context.Background(), "")) { |
| 789 | if r.Outcome != "recovered" || rowAt(t, p, h) != endpoints(t, p, h, ModeNested) { |
| 790 | t.Errorf("%s: %+v at %+v", h, r, rowAt(t, p, h)) |
| 791 | } |
| 792 | } |
| 793 | if len(f.fences) != 0 { |
| 794 | t.Errorf("fences: %v", f.fences) |
| 795 | } |
| 796 | } |
| 797 | |
| 798 | // When the agent starts and cannot read the fleet's intent, a revert in |
| 799 | // flight is completed but kept fenced at phase done: releasing it flat |
| 800 | // could let a forged revert run a nested host unisolated until an intent |
| 801 | // is read. A move to nested is completed and released as usual. |
| 802 | func TestRecoverHoldsACompletedRevertWhileTheIntentIsUnreadable(t *testing.T) { |
| 803 | p, f := box(t) |
| 804 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 805 | t.Fatalf("setup: %+v", res) |
| 806 | } |
| 807 | p.abortAt = "after-cas" |
| 808 | p.Converge(context.Background(), ModeFlat) |
| 809 | p.abortAt = "" |
| 810 | res := byHost(t, p.Recover(context.Background(), IntentUnreadable)) |
| 811 | for _, h := range []string{host, replica} { |
| 812 | if res[h].Outcome != "deferred" || !strings.Contains(res[h].Reason, "intent") { |
| 813 | t.Errorf("%s: %+v", h, res[h]) |
| 814 | } |
| 815 | j, found, err := ReadJournal(p.DataDir, h) |
| 816 | if err != nil || !found || j.Phase != PhaseDone { |
| 817 | t.Fatalf("%s journal: %+v %v %v", h, j, found, err) |
| 818 | } |
| 819 | if rowAt(t, p, h) != endpoints(t, p, h, ModeFlat) { |
| 820 | t.Errorf("%s row not completed to flat", h) |
| 821 | } |
| 822 | } |
| 823 | if len(f.fences) != 0 { |
| 824 | t.Errorf("the journal is the fence; the activator's must not stay: %v", f.fences) |
| 825 | } |
| 826 | // Once the intent is read, the held host converges to it. |
| 827 | res = byHost(t, p.Recover(context.Background(), ModeNested)) |
| 828 | for _, h := range []string{host, replica} { |
| 829 | if res[h].Outcome != "recovered" || rowAt(t, p, h) != endpoints(t, p, h, ModeNested) || JournalPresent(p.DataDir, h) { |
| 830 | t.Errorf("%s after the intent: %+v at %+v", h, res[h], rowAt(t, p, h)) |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | // A move toward nested is released even while the intent is unreadable. |
| 835 | q, _ := box(t) |
| 836 | q.abortAt = "after-files" |
| 837 | q.Converge(context.Background(), ModeNested) |
| 838 | q.abortAt = "" |
| 839 | for h, r := range byHost(t, q.Recover(context.Background(), IntentUnreadable)) { |
| 840 | if r.Outcome != "recovered" || JournalPresent(q.DataDir, h) { |
| 841 | t.Errorf("%s nested journal under an unreadable intent: %+v", h, r) |
| 842 | } |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | // abortedNestedMove leaves both hosts with a nested journal at phase files |
| 847 | // and their files already at the nested destination. |
| 848 | func abortedNestedMove(t *testing.T, p *Pass) { |
| 849 | t.Helper() |
| 850 | p.abortAt = "after-files" |
| 851 | p.Converge(context.Background(), ModeNested) |
| 852 | p.abortAt = "" |
| 853 | for _, h := range []string{host, replica} { |
| 854 | if j, found, err := ReadJournal(p.DataDir, h); err != nil || !found || j.Phase != PhaseFiles { |
| 855 | t.Fatalf("setup %s: %+v %v %v", h, j, found, err) |
| 856 | } |
| 857 | } |
| 858 | } |
| 859 | |
| 860 | func installForeignFence(t *testing.T, svc *activator.Service, h string, gen uint64) { |
| 861 | t.Helper() |
| 862 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 863 | defer cancel() |
| 864 | if _, err := svc.FenceAndStop(ctx, h, gen); err != nil && !errors.Is(err, activator.ErrStopUnproven) { |
| 865 | t.Fatalf("installing the foreign fence: %v", err) |
| 866 | } |
| 867 | } |
| 868 | |
| 869 | // A journal on a host another operation holds fenced is completed under |
| 870 | // that operation's fence: the pass installs and lifts nothing. When the |
| 871 | // completed endpoint is the fleet's intent there is nothing left to do, |
| 872 | // and the journal goes. |
| 873 | func TestRecoverUnderAForeignFenceCompletesAndRemovesAtTheIntent(t *testing.T) { |
| 874 | p, _ := box(t) |
| 875 | svc, _ := realService(t, p) |
| 876 | abortedNestedMove(t, p) |
| 877 | installForeignFence(t, svc, host, 5) |
| 878 | res := byHost(t, p.Recover(context.Background(), ModeNested)) |
| 879 | for _, h := range []string{host, replica} { |
| 880 | if res[h].Outcome != "recovered" || JournalPresent(p.DataDir, h) || rowAt(t, p, h) != endpoints(t, p, h, ModeNested) { |
| 881 | t.Errorf("%s: %+v at %+v", h, res[h], rowAt(t, p, h)) |
| 882 | } |
| 883 | } |
| 884 | if svc.Fenced(replica) { |
| 885 | t.Error("the unfenced host was left fenced") |
| 886 | } |
| 887 | if !svc.LiftFence(host, 5) { |
| 888 | t.Fatal("the foreign fence is not in place at its own generation") |
| 889 | } |
| 890 | } |
| 891 | |
| 892 | // When the completed endpoint is not the intent, the pass cannot converge |
| 893 | // the host without a fence of its own. It keeps the journal at done, which |
| 894 | // keeps the host fenced if the other operation lifts its fence, and the |
| 895 | // next recovery after that lift converges it. |
| 896 | func TestRecoverUnderAForeignFenceKeepsTheJournalWhenNotAtTheIntent(t *testing.T) { |
| 897 | p, _ := box(t) |
| 898 | svc, _ := realService(t, p) |
| 899 | abortedNestedMove(t, p) |
| 900 | installForeignFence(t, svc, host, 5) |
| 901 | res := byHost(t, p.Recover(context.Background(), ModeFlat)) |
| 902 | if r := res[host]; r.Outcome != "deferred" || !strings.Contains(r.Reason, "fence") { |
| 903 | t.Fatalf("held host: %+v", r) |
| 904 | } |
| 905 | // The move waits for the journal and the journal waits for the move, |
| 906 | // so the reason has to tell an operator both ways out. |
| 907 | for _, want := range []string{"instance move", "abort the move", provision.LayoutJournalPath(p.DataDir, host), "stands at its destination"} { |
| 908 | if r := res[host]; !strings.Contains(r.Reason, want) { |
| 909 | t.Errorf("reason %q does not say %q", r.Reason, want) |
| 910 | } |
| 911 | } |
| 912 | if r := res[host]; strings.Contains(r.Reason, "\u2014") { |
| 913 | t.Errorf("reason has an em dash: %q", r.Reason) |
| 914 | } |
| 915 | if j, found, err := ReadJournal(p.DataDir, host); err != nil || !found || j.Phase != PhaseDone { |
| 916 | t.Fatalf("journal: %+v %v %v", j, found, err) |
| 917 | } |
| 918 | if rowAt(t, p, host) != endpoints(t, p, host, ModeNested) { |
| 919 | t.Errorf("row not completed to the journal's endpoint: %+v", rowAt(t, p, host)) |
| 920 | } |
| 921 | if res[replica].Outcome != "recovered" { |
| 922 | t.Errorf("unfenced host: %+v", res[replica]) |
| 923 | } |
| 924 | assertFlatUntouched(t, p, replica) |
| 925 | if !svc.Fenced(host) { |
| 926 | t.Fatal("the foreign fence was lifted") |
| 927 | } |
| 928 | if !svc.LiftFence(host, 5) { |
| 929 | t.Fatal("the foreign fence is not at its own generation") |
| 930 | } |
| 931 | // The other operation is gone: the next recovery converges the host. |
| 932 | if r := byHost(t, p.Recover(context.Background(), ModeFlat))[host]; r.Outcome != "recovered" { |
| 933 | t.Fatalf("after the foreign fence lifted: %+v", r) |
| 934 | } |
| 935 | assertFlatUntouched(t, p, host) |
| 936 | if svc.Fenced(host) { |
| 937 | t.Error("left fenced") |
| 938 | } |
| 939 | } |
| 940 | |
| 941 | // Under a foreign fence the pass has no stop proof of its own, so it acts |
| 942 | // only on a host with no live entry. Otherwise nothing changes. |
| 943 | func TestRecoverUnderAForeignFenceDefersAnAwakeHost(t *testing.T) { |
| 944 | p, f := box(t) |
| 945 | abortedNestedMove(t, p) |
| 946 | f.fences = map[string]uint64{host: 5} |
| 947 | f.awake = map[string]bool{host: true} |
| 948 | f.events = nil |
| 949 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 950 | if r.Outcome != "deferred" { |
| 951 | t.Fatalf("result = %+v", r) |
| 952 | } |
| 953 | if j, found, _ := ReadJournal(p.DataDir, host); !found || j.Phase != PhaseFiles { |
| 954 | t.Fatalf("journal changed: %+v", j) |
| 955 | } |
| 956 | if rowAt(t, p, host) != endpoints(t, p, host, ModeFlat) { |
| 957 | t.Errorf("row changed: %+v", rowAt(t, p, host)) |
| 958 | } |
| 959 | if f.fences[host] != 5 { |
| 960 | t.Errorf("foreign fence = %d", f.fences[host]) |
| 961 | } |
| 962 | for _, e := range f.events { |
| 963 | if e == "lift "+host || e == "fence "+host { |
| 964 | t.Errorf("the pass touched the host's fence: %v", f.events) |
| 965 | } |
| 966 | } |
| 967 | } |
| 968 | |
| 969 | // Completing a journal is never gated, but taking the host on toward a |
| 970 | // nested intent is, as Converge(ModeNested) is. With the probe failing, a |
| 971 | // revert completes and the host is held flat, not released: there is no |
| 972 | // degraded mode (design §8.5), and a released flat host is exactly what a |
| 973 | // forged revert wants. The journal at done is the fence, so the pass |
| 974 | // lifts its in-memory one. |
| 975 | func TestRecoverGatesTheLegTowardNestedAndHoldsFlat(t *testing.T) { |
| 976 | p, f := box(t) |
| 977 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 978 | t.Fatalf("setup: %+v", res) |
| 979 | } |
| 980 | p.abortAt = "after-files" |
| 981 | p.Converge(context.Background(), ModeFlat) |
| 982 | p.abortAt = "" |
| 983 | calls := 0 |
| 984 | p.NestedOK = func() error { calls++; return errors.New("probe: not ready") } |
| 985 | f.events = nil |
| 986 | res := byHost(t, p.Recover(context.Background(), ModeNested)) |
| 987 | for _, h := range []string{host, replica} { |
| 988 | if r := res[h]; r.Outcome != "deferred" || !strings.Contains(r.Reason, "isolation not ready: probe: not ready") { |
| 989 | t.Errorf("%s: %+v", h, r) |
| 990 | } |
| 991 | assertHeldFlat(t, p, h) |
| 992 | } |
| 993 | if calls != 1 { |
| 994 | t.Errorf("NestedOK called %d times, want once per recovery", calls) |
| 995 | } |
| 996 | want := []string{"fence " + host, "reload", "lift " + host, "fence " + replica, "reload", "lift " + replica} |
| 997 | if !reflect.DeepEqual(f.events, want) { |
| 998 | t.Fatalf("events = %v\nwant %v", f.events, want) |
| 999 | } |
| 1000 | // Once the probe passes, the next recovery takes the held host on to |
| 1001 | // nested. |
| 1002 | p.NestedOK = nil |
| 1003 | for h, r := range byHost(t, p.Recover(context.Background(), ModeNested)) { |
| 1004 | if r.Outcome != "recovered" || JournalPresent(p.DataDir, h) || rowAt(t, p, h) != endpoints(t, p, h, ModeNested) { |
| 1005 | t.Errorf("%s after the probe passed: %+v", h, r) |
| 1006 | } |
| 1007 | } |
| 1008 | } |
| 1009 | |
| 1010 | // assertHeldFlat is a host whose revert completed and is held: the row |
| 1011 | // and files at flat, and the journal still there at done. |
| 1012 | func assertHeldFlat(t *testing.T, p *Pass, h string) { |
| 1013 | t.Helper() |
| 1014 | if got := rowAt(t, p, h); got != endpoints(t, p, h, ModeFlat) { |
| 1015 | t.Errorf("%s row not at flat: %+v", h, got) |
| 1016 | } |
| 1017 | if b, err := os.ReadFile(provision.DefaultDBPath(p.DataDir, h)); err != nil || string(b) != "db "+h { |
| 1018 | t.Errorf("%s flat db: %q %v", h, b, err) |
| 1019 | } |
| 1020 | if j, found, err := ReadJournal(p.DataDir, h); err != nil || !found || j.Mode != ModeFlat || j.Phase != PhaseDone { |
| 1021 | t.Errorf("%s journal: %+v %v %v, want a flat journal held at done", h, j, found, err) |
| 1022 | } |
| 1023 | } |
| 1024 | |
| 1025 | // addFlatRow adds a fresh flat managed row with its files, after any setup |
| 1026 | // that would otherwise have moved it. |
| 1027 | func addFlatRow(t *testing.T, p *Pass, h string) { |
| 1028 | t.Helper() |
| 1029 | if err := p.Registry.Upsert(registry.Route{Host: h, Kind: "instance", Hibernate: true, |
| 1030 | Socket: filepath.Join(p.SocketDir, h+".sock"), DBPath: provision.DefaultDBPath(p.DataDir, h)}); err != nil { |
| 1031 | t.Fatal(err) |
| 1032 | } |
| 1033 | seedFlat(t, p.DataDir, h) |
| 1034 | } |
| 1035 | |
| 1036 | // forgeFlatJournal writes a well-formed revert journal for a nested host, |
| 1037 | // recording each artifact present exactly when it is at its nested source, |
| 1038 | // except the names in absent, which it records absent regardless. It is |
| 1039 | // what a flat tenant sharing the agent's uid can write into <data>/layout. |
| 1040 | func forgeFlatJournal(t *testing.T, p *Pass, h string, absent ...string) { |
| 1041 | t.Helper() |
| 1042 | plan, err := PlanFor(p.DataDir, p.SocketDir, h, ModeFlat) |
| 1043 | if err != nil { |
| 1044 | t.Fatal(err) |
| 1045 | } |
| 1046 | present := make([]bool, len(plan.Artifacts)) |
| 1047 | for i, a := range plan.Artifacts { |
| 1048 | ok, _ := lstatPresent(a.Src) |
| 1049 | present[i] = ok && a.Name != "socket" && !slicesContains(absent, a.Name) |
| 1050 | } |
| 1051 | if err := WriteJournal(p.DataDir, plan.Journal(present)); err != nil { |
| 1052 | t.Fatal(err) |
| 1053 | } |
| 1054 | } |
| 1055 | |
| 1056 | func slicesContains(xs []string, x string) bool { |
| 1057 | for _, v := range xs { |
| 1058 | if v == x { |
| 1059 | return true |
| 1060 | } |
| 1061 | } |
| 1062 | return false |
| 1063 | } |
| 1064 | |
| 1065 | // A journal completed to an endpoint that is not the fleet's intent keeps |
| 1066 | // the host fenced until the converge leg succeeds. If the leg refuses, the |
| 1067 | // journal stays at done: a forged revert must not leave a nested host |
| 1068 | // released flat. |
| 1069 | func TestRecoverHoldsTheJournalWhenTheLegTowardTheIntentRefuses(t *testing.T) { |
| 1070 | p, f := box(t) |
| 1071 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 1072 | t.Fatalf("setup: %+v", res) |
| 1073 | } |
| 1074 | forgeFlatJournal(t, p, host) |
| 1075 | // A route that owns <host>.config.env under the flat rules makes the |
| 1076 | // way back to nested refuse at the ownership preflight. |
| 1077 | if err := p.Registry.Upsert(registry.Route{Host: host + ".config", Kind: "instance", |
| 1078 | Socket: filepath.Join(p.SocketDir, host+".config.sock"), DBPath: provision.DefaultDBPath(p.DataDir, host+".config")}); err != nil { |
| 1079 | t.Fatal(err) |
| 1080 | } |
| 1081 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 1082 | if r.Outcome != "deferred" || !strings.Contains(r.Reason, "config") { |
| 1083 | t.Fatalf("result = %+v", r) |
| 1084 | } |
| 1085 | j, found, err := ReadJournal(p.DataDir, host) |
| 1086 | if err != nil || !found || j.Phase != PhaseDone || j.Mode != ModeFlat { |
| 1087 | t.Fatalf("journal not held at done: %+v %v %v", j, found, err) |
| 1088 | } |
| 1089 | if len(f.fences) != 0 { |
| 1090 | t.Errorf("fences: %v", f.fences) |
| 1091 | } |
| 1092 | // Every later pass holds it the same way; it is never released flat. |
| 1093 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "deferred" || !JournalPresent(p.DataDir, host) { |
| 1094 | t.Fatalf("second pass: %+v", r) |
| 1095 | } |
| 1096 | } |
| 1097 | |
| 1098 | // The attack the hold closes, end to end: a forged revert that records a |
| 1099 | // real nested artifact absent. Recovery refuses it outright, and the host |
| 1100 | // stays nested and fenced. |
| 1101 | func TestRecoverDoesNotReleaseANestedHostFlatOverAForgedRevert(t *testing.T) { |
| 1102 | p, _ := box(t) |
| 1103 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 1104 | t.Fatalf("setup: %+v", res) |
| 1105 | } |
| 1106 | shadow := filepath.Join(rawInstanceDir(p.DataDir, host), shadowName(host)) |
| 1107 | if _, err := os.Stat(shadow); err != nil { |
| 1108 | t.Fatalf("setup: %v", err) |
| 1109 | } |
| 1110 | forgeFlatJournal(t, p, host, "litestream-shadow") |
| 1111 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 1112 | if r.Outcome == "recovered" { |
| 1113 | t.Fatalf("result = %+v", r) |
| 1114 | } |
| 1115 | if rowAt(t, p, host) != endpoints(t, p, host, ModeNested) { |
| 1116 | t.Errorf("row released flat: %+v", rowAt(t, p, host)) |
| 1117 | } |
| 1118 | if !JournalPresent(p.DataDir, host) { |
| 1119 | t.Error("journal removed; the host is no longer fenced") |
| 1120 | } |
| 1121 | } |
| 1122 | |
| 1123 | // A journal's "absent" is a claim a crash cannot contradict: the host has |
| 1124 | // been fenced since the plan was made, so nothing recorded absent can have |
| 1125 | // appeared at either end. Recovery refuses one that has, naming it. |
| 1126 | func TestRecoverRefusesAnArtifactRecordedAbsentThatExists(t *testing.T) { |
| 1127 | for _, end := range []string{"source", "destination"} { |
| 1128 | t.Run(end, func(t *testing.T) { |
| 1129 | p, _ := box(t) |
| 1130 | abortedNestedMove(t, p) |
| 1131 | plan, _ := PlanFor(p.DataDir, p.SocketDir, host, ModeNested) |
| 1132 | sidecar := artifactNamed(t, plan, "sidecar-env") |
| 1133 | path := sidecar.Src |
| 1134 | if end == "destination" { |
| 1135 | path = sidecar.Dst |
| 1136 | } |
| 1137 | if err := os.WriteFile(path, []byte("X=1\n"), 0o600); err != nil { |
| 1138 | t.Fatal(err) |
| 1139 | } |
| 1140 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 1141 | if r.Outcome != "refused" || !strings.Contains(r.Reason, path) || !strings.Contains(r.Reason, "absent") { |
| 1142 | t.Fatalf("result = %+v", r) |
| 1143 | } |
| 1144 | if j, found, _ := ReadJournal(p.DataDir, host); !found || j.Phase != PhaseFiles { |
| 1145 | t.Fatalf("journal changed: %+v", j) |
| 1146 | } |
| 1147 | }) |
| 1148 | } |
| 1149 | } |
| 1150 | |
| 1151 | // Completing a journal is never gated, so a failing probe does not leave a |
| 1152 | // crashed move fenced. Only fresh moves wait for the probe. |
| 1153 | func TestConvergeNestedCompletesAJournalWhileTheProbeFails(t *testing.T) { |
| 1154 | p, _ := box(t) |
| 1155 | abortedNestedMove(t, p) |
| 1156 | addFlatRow(t, p, "f.example.com") |
| 1157 | calls := 0 |
| 1158 | p.NestedOK = func() error { calls++; return errors.New("probe: not ready") } |
| 1159 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1160 | for _, h := range []string{host, replica} { |
| 1161 | if res[h].Outcome != "recovered" || JournalPresent(p.DataDir, h) || rowAt(t, p, h) != endpoints(t, p, h, ModeNested) { |
| 1162 | t.Errorf("%s: %+v", h, res[h]) |
| 1163 | } |
| 1164 | } |
| 1165 | if r := res[""]; r.Outcome != "refused" || r.Reason != "isolation not ready: probe: not ready" { |
| 1166 | t.Errorf("pass result: %+v", r) |
| 1167 | } |
| 1168 | if _, listed := res["f.example.com"]; listed { |
| 1169 | t.Errorf("a fresh move ran under a failing probe: %+v", res["f.example.com"]) |
| 1170 | } |
| 1171 | assertFlatUntouched(t, p, "f.example.com") |
| 1172 | if calls != 1 { |
| 1173 | t.Errorf("NestedOK called %d times", calls) |
| 1174 | } |
| 1175 | } |
| 1176 | |
| 1177 | // One pass asks the probe once, whether the question comes from a |
| 1178 | // recovery's leg toward nested or from the fresh moves. |
| 1179 | func TestConvergeAsksTheIsolationGateOncePerPass(t *testing.T) { |
| 1180 | for _, tc := range []struct { |
| 1181 | name string |
| 1182 | gate error |
| 1183 | }{{"failing", errors.New("probe: not ready")}, {"ready", nil}} { |
| 1184 | t.Run(tc.name, func(t *testing.T) { |
| 1185 | p, _ := box(t) |
| 1186 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 1187 | t.Fatalf("setup: %+v", res) |
| 1188 | } |
| 1189 | p.abortAt = "after-files" |
| 1190 | p.Converge(context.Background(), ModeFlat) |
| 1191 | p.abortAt = "" |
| 1192 | addFlatRow(t, p, "f.example.com") |
| 1193 | calls := 0 |
| 1194 | p.NestedOK = func() error { calls++; return tc.gate } |
| 1195 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1196 | if calls != 1 { |
| 1197 | t.Fatalf("NestedOK called %d times in one pass", calls) |
| 1198 | } |
| 1199 | want := "recovered" |
| 1200 | if tc.gate != nil { |
| 1201 | want = "deferred" |
| 1202 | } |
| 1203 | for _, h := range []string{host, replica} { |
| 1204 | if res[h].Outcome != want { |
| 1205 | t.Errorf("%s: %+v, want %s", h, res[h], want) |
| 1206 | } |
| 1207 | } |
| 1208 | if tc.gate != nil { |
| 1209 | assertHeldFlat(t, p, host) |
| 1210 | assertFlatUntouched(t, p, "f.example.com") |
| 1211 | } else if res["f.example.com"].Outcome != "migrated" || rowAt(t, p, host) != endpoints(t, p, host, ModeNested) { |
| 1212 | t.Errorf("ready probe: %+v", res) |
| 1213 | } |
| 1214 | }) |
| 1215 | } |
| 1216 | } |
| 1217 | |
| 1218 | // A host whose database is not on this box cannot move; finding that out |
| 1219 | // must not cost it a stop, or directories and a destination check, on |
| 1220 | // every pass. |
| 1221 | func TestConvergeRefusesAHostWithNoLocalDatabaseWithoutFencingIt(t *testing.T) { |
| 1222 | p, f := box(t) |
| 1223 | db := provision.DefaultDBPath(p.DataDir, host) |
| 1224 | if err := os.Remove(db); err != nil { |
| 1225 | t.Fatal(err) |
| 1226 | } |
| 1227 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1228 | if r := res[host]; r.Outcome != "refused" || !strings.Contains(r.Reason, "db is missing at "+db) { |
| 1229 | t.Fatalf("result = %+v", r) |
| 1230 | } |
| 1231 | for _, e := range f.events { |
| 1232 | if strings.HasSuffix(e, " "+host) { |
| 1233 | t.Errorf("the host's fence was touched: %v", f.events) |
| 1234 | } |
| 1235 | } |
| 1236 | if _, err := os.Lstat(rawInstanceDir(p.DataDir, host)); !errors.Is(err, os.ErrNotExist) { |
| 1237 | t.Errorf("directories were made for a host that cannot move: %v", err) |
| 1238 | } |
| 1239 | if res[replica].Outcome != "migrated" { |
| 1240 | t.Errorf("the other host: %+v", res[replica]) |
| 1241 | } |
| 1242 | } |
| 1243 | |
| 1244 | // crashedRevert leaves both hosts nested-then-mid-revert: each has a flat |
| 1245 | // journal at phase files with its files already at the flat destination. |
| 1246 | func crashedRevert(t *testing.T, p *Pass) { |
| 1247 | t.Helper() |
| 1248 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "migrated" { |
| 1249 | t.Fatalf("setup: %+v", res) |
| 1250 | } |
| 1251 | p.abortAt = "after-files" |
| 1252 | p.Converge(context.Background(), ModeFlat) |
| 1253 | p.abortAt = "" |
| 1254 | } |
| 1255 | |
| 1256 | // The handover from the completed journal to the converge leg's journal |
| 1257 | // is one rename, and anything going wrong on it leaves the host down: the |
| 1258 | // held journal stays whole, the pass's fence stays installed, and nothing |
| 1259 | // is released. A same-uid tenant can make <data>/layout unwritable at the |
| 1260 | // right moment, so this is reachable, not only a disk fault. |
| 1261 | func TestRecoverKeepsTheFenceWhenTheHandoverFails(t *testing.T) { |
| 1262 | if os.Geteuid() == 0 { |
| 1263 | t.Skip("root ignores the directory mode this test relies on") |
| 1264 | } |
| 1265 | p, f := box(t) |
| 1266 | crashedRevert(t, p) |
| 1267 | dir := journalDir(p.DataDir) |
| 1268 | calls := 0 |
| 1269 | p.Serialize = func(fn func()) { |
| 1270 | calls++ |
| 1271 | // The second section on the first host is the handover: make the |
| 1272 | // temp file impossible to create there, as a racing tenant could. |
| 1273 | if calls == 2 { |
| 1274 | if err := os.Chmod(dir, 0o500); err != nil { |
| 1275 | t.Fatal(err) |
| 1276 | } |
| 1277 | defer os.Chmod(dir, 0o700) |
| 1278 | } |
| 1279 | fn() |
| 1280 | } |
| 1281 | f.events = nil |
| 1282 | r := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 1283 | if r.Outcome != "refused" || !strings.Contains(r.Reason, "held by this pass after a failed handover") { |
| 1284 | t.Fatalf("result = %+v", r) |
| 1285 | } |
| 1286 | j, found, err := ReadJournal(p.DataDir, host) |
| 1287 | if err != nil || !found || j.Mode != ModeFlat || j.Phase != PhaseDone { |
| 1288 | t.Fatalf("held journal not intact: %+v %v %v", j, found, err) |
| 1289 | } |
| 1290 | if f.fences[host] == 0 { |
| 1291 | t.Error("the pass's fence was lifted after a failed handover") |
| 1292 | } |
| 1293 | for _, e := range f.events { |
| 1294 | if e == "lift "+host { |
| 1295 | t.Errorf("released after a failed handover: %v", f.events) |
| 1296 | } |
| 1297 | } |
| 1298 | if left := tempFilesIn(t, dir); len(left) != 0 { |
| 1299 | t.Errorf("temp files left: %v", left) |
| 1300 | } |
| 1301 | } |
| 1302 | |
| 1303 | // A fence the pass kept after a failed handover is its own, not another |
| 1304 | // operation's. The next pass recognises it by its generation, re-runs the |
| 1305 | // leg once <data>/layout is writable again, and lifts it when the host |
| 1306 | // stands at the intent. |
| 1307 | func TestRecoverRetriesAHandoverUnderTheFenceItKept(t *testing.T) { |
| 1308 | if os.Geteuid() == 0 { |
| 1309 | t.Skip("root ignores the directory mode this test relies on") |
| 1310 | } |
| 1311 | p, f := box(t) |
| 1312 | crashedRevert(t, p) |
| 1313 | dir := journalDir(p.DataDir) |
| 1314 | calls := 0 |
| 1315 | p.Serialize = func(fn func()) { |
| 1316 | calls++ |
| 1317 | if calls == 2 { |
| 1318 | if err := os.Chmod(dir, 0o500); err != nil { |
| 1319 | t.Fatal(err) |
| 1320 | } |
| 1321 | defer os.Chmod(dir, 0o700) |
| 1322 | } |
| 1323 | fn() |
| 1324 | } |
| 1325 | first := byHost(t, p.Recover(context.Background(), ModeNested))[host] |
| 1326 | if first.Outcome != "refused" || !strings.Contains(first.Reason, "held by this pass after a failed handover") || !strings.Contains(first.Reason, "the next pass retries it") { |
| 1327 | t.Fatalf("first pass: %+v", first) |
| 1328 | } |
| 1329 | kept := f.fences[host] |
| 1330 | if kept == 0 { |
| 1331 | t.Fatal("the fence was not kept") |
| 1332 | } |
| 1333 | p.Serialize = nil |
| 1334 | f.events = nil |
| 1335 | res := p.Recover(context.Background(), ModeNested) |
| 1336 | second := byHost(t, res)[host] |
| 1337 | if second.Outcome != "recovered" || !strings.Contains(second.Reason, "converged to nested") { |
| 1338 | t.Fatalf("second pass: %+v", second) |
| 1339 | } |
| 1340 | for _, r := range append(res, first) { |
| 1341 | if strings.Contains(r.Reason, "another operation") { |
| 1342 | t.Errorf("the pass's own fence reported as another operation's: %+v", r) |
| 1343 | } |
| 1344 | } |
| 1345 | if JournalPresent(p.DataDir, host) || rowAt(t, p, host) != endpoints(t, p, host, ModeNested) { |
| 1346 | t.Fatalf("host not converged: journal %v, row %+v", JournalPresent(p.DataDir, host), rowAt(t, p, host)) |
| 1347 | } |
| 1348 | if f.fences[host] != 0 { |
| 1349 | t.Fatalf("the kept fence (generation %d) is still installed at %d", kept, f.fences[host]) |
| 1350 | } |
| 1351 | if !slicesContains(f.events, "lift "+host) || slicesContains(f.events, "fence "+host) { |
| 1352 | t.Errorf("events = %v, want the kept fence lifted and no new one installed", f.events) |
| 1353 | } |
| 1354 | // The record went with the fence: a later foreign fence at the same |
| 1355 | // generation is not mistaken for the pass's own. |
| 1356 | p.keptMu.Lock() |
| 1357 | _, still := p.kept[host] |
| 1358 | p.keptMu.Unlock() |
| 1359 | if still { |
| 1360 | t.Error("the kept generation is still recorded after the lift") |
| 1361 | } |
| 1362 | } |
| 1363 | |
| 1364 | // A successful handover leaves exactly one journal, the converge leg's |
| 1365 | // own, and no temp file. |
| 1366 | func TestRecoverHandsOverToTheNewJournalInOneStep(t *testing.T) { |
| 1367 | p, _ := box(t) |
| 1368 | crashedRevert(t, p) |
| 1369 | p.abortAt = "after-journal" |
| 1370 | p.Recover(context.Background(), ModeNested) |
| 1371 | p.abortAt = "" |
| 1372 | j, found, err := ReadJournal(p.DataDir, host) |
| 1373 | if err != nil || !found || j.Mode != ModeNested || j.Phase != PhaseFiles { |
| 1374 | t.Fatalf("journal after the handover: %+v %v %v", j, found, err) |
| 1375 | } |
| 1376 | entries, err := os.ReadDir(journalDir(p.DataDir)) |
| 1377 | if err != nil { |
| 1378 | t.Fatal(err) |
| 1379 | } |
| 1380 | var names []string |
| 1381 | for _, e := range entries { |
| 1382 | names = append(names, e.Name()) |
| 1383 | } |
| 1384 | want := []string{host + ".json", replica + ".json"} |
| 1385 | if !reflect.DeepEqual(names, want) { |
| 1386 | t.Fatalf("journal dir = %v, want %v", names, want) |
| 1387 | } |
| 1388 | } |
| 1389 | |
| 1390 | // The presence scan runs inside the section config sync is locked out of, |
| 1391 | // so a bundle rendered just before the pass took the lock is recorded and |
| 1392 | // moved, never stranded at the flat path behind a journal that calls it |
| 1393 | // absent. |
| 1394 | func TestThePresenceScanRunsInsideTheSerializedSection(t *testing.T) { |
| 1395 | p, _ := box(t) |
| 1396 | rendered := false |
| 1397 | p.Serialize = func(fn func()) { |
| 1398 | if !rendered { |
| 1399 | // Config sync's last render, landing while it still held the |
| 1400 | // lock the pass is now taking. |
| 1401 | for _, h := range []string{host, replica} { |
| 1402 | if err := os.WriteFile(filepath.Join(p.DataDir, h+".config.env"), []byte("A=1\n"), 0o600); err != nil { |
| 1403 | t.Fatal(err) |
| 1404 | } |
| 1405 | } |
| 1406 | rendered = true |
| 1407 | } |
| 1408 | fn() |
| 1409 | } |
| 1410 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1411 | for _, h := range []string{host, replica} { |
| 1412 | if res[h].Outcome != "migrated" { |
| 1413 | t.Fatalf("%s: %+v", h, res[h]) |
| 1414 | } |
| 1415 | if _, err := os.Lstat(filepath.Join(p.DataDir, h+".config.env")); !errors.Is(err, os.ErrNotExist) { |
| 1416 | t.Errorf("%s bundle stranded at the flat path", h) |
| 1417 | } |
| 1418 | if _, err := os.Stat(filepath.Join(rawInstanceDir(p.DataDir, h), h+".config.env")); err != nil { |
| 1419 | t.Errorf("%s bundle not moved: %v", h, err) |
| 1420 | } |
| 1421 | } |
| 1422 | } |
| 1423 | |
| 1424 | // Review F1 end to end, as amended for landing: a planted hard link in |
| 1425 | // the host's future directory is quarantined and the host migrates; the |
| 1426 | // aside copy is outside the instance tree, status gets the event as a |
| 1427 | // note, and the clean sibling is unaffected. |
| 1428 | func TestConvergeQuarantinesPlantedEntriesAndMigrates(t *testing.T) { |
| 1429 | p, f := box(t) |
| 1430 | secret := filepath.Join(p.DataDir, "litestream.env") |
| 1431 | if err := os.WriteFile(secret, []byte("AWS_SECRET_ACCESS_KEY=x\n"), 0o600); err != nil { |
| 1432 | t.Fatal(err) |
| 1433 | } |
| 1434 | dir := rawInstanceDir(p.DataDir, host) |
| 1435 | if err := os.MkdirAll(dir, 0o700); err != nil { |
| 1436 | t.Fatal(err) |
| 1437 | } |
| 1438 | if err := os.Link(secret, filepath.Join(dir, "creds")); err != nil { |
| 1439 | t.Skipf("hard links unavailable here: %v", err) |
| 1440 | } |
| 1441 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1442 | for _, h := range []string{host, replica} { |
| 1443 | if res[h].Outcome != "migrated" { |
| 1444 | t.Fatalf("%s: %+v", h, res[h]) |
| 1445 | } |
| 1446 | } |
| 1447 | if f.fences[host] != 0 { |
| 1448 | t.Fatal("fence still installed after the move") |
| 1449 | } |
| 1450 | if _, err := os.Lstat(filepath.Join(dir, "creds")); !errors.Is(err, os.ErrNotExist) { |
| 1451 | t.Errorf("the planted link is still in the instance directory: %v", err) |
| 1452 | } |
| 1453 | q, err := os.ReadDir(asideDir(p.DataDir, host)) |
| 1454 | if err != nil || len(q) != 1 || !strings.HasPrefix(q[0].Name(), "creds.q-") { |
| 1455 | t.Fatalf("quarantine = %v %v", q, err) |
| 1456 | } |
| 1457 | if n := p.Note(host); !strings.Contains(n, "kept aside") || !strings.Contains(n, "creds.q-") { |
| 1458 | t.Errorf("note = %q", n) |
| 1459 | } |
| 1460 | if _, held := p.Held(host); held { |
| 1461 | t.Error("a quarantined entry became a hold") |
| 1462 | } |
| 1463 | } |
| 1464 | |
| 1465 | // Review G1: a database with a second name is refused before the journal |
| 1466 | // is written, and the refusal does NOT let the host run flat again. One |
| 1467 | // `ln /data/B.db /data/.x` would otherwise keep a tenant's own move |
| 1468 | // refused for ever while it woke flat on every request. The host stays |
| 1469 | // fenced and down, status reads the hold, the refusal names the inode and |
| 1470 | // the find that lists the other name, a flat intent lifts the fence, and |
| 1471 | // once the second name is gone the move goes through from the kept fence. |
| 1472 | func TestConvergeHoldsAFlatHostWhoseMoveWasRefusedAfterItsStop(t *testing.T) { |
| 1473 | p, f := box(t) |
| 1474 | other := filepath.Join(p.DataDir, ".x") |
| 1475 | if err := os.Link(provision.DefaultDBPath(p.DataDir, host), other); err != nil { |
| 1476 | t.Skipf("hard links unavailable here: %v", err) |
| 1477 | } |
| 1478 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1479 | r := res[host] |
| 1480 | if r.Outcome != "held" || !strings.Contains(r.Reason, "2 names") || !strings.Contains(r.Reason, "-samefile") || !strings.Contains(r.Reason, "inode") { |
| 1481 | t.Fatalf("linked database: %+v", r) |
| 1482 | } |
| 1483 | assertFlatUntouched(t, p, host) |
| 1484 | if f.fences[host] == 0 { |
| 1485 | t.Fatal("the fence was lifted; the host would wake flat on its next request") |
| 1486 | } |
| 1487 | if why, held := p.Held(host); !held || !strings.Contains(why, "2 names") { |
| 1488 | t.Fatalf("Held = %q %v", why, held) |
| 1489 | } |
| 1490 | // The tenant keeps the link alive: every pass meets the same refusal, |
| 1491 | // and the host stays down each time, without a second fence or lift. |
| 1492 | res = byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1493 | if res[host].Outcome != "held" || f.fences[host] == 0 { |
| 1494 | t.Fatalf("second pass: %+v fence=%d", res[host], f.fences[host]) |
| 1495 | } |
| 1496 | lifts := 0 |
| 1497 | for _, e := range f.events { |
| 1498 | if e == "lift "+host { |
| 1499 | lifts++ |
| 1500 | } |
| 1501 | } |
| 1502 | if lifts != 0 { |
| 1503 | t.Fatalf("the held fence was lifted and re-installed: %v", f.events) |
| 1504 | } |
| 1505 | // The fleet decides it wants flat after all: the hold ends and the |
| 1506 | // host runs flat, as asked. |
| 1507 | res = byHost(t, p.Converge(context.Background(), ModeFlat)) |
| 1508 | if res[host].Outcome != "unchanged" || f.fences[host] != 0 { |
| 1509 | t.Fatalf("flat intent: %+v fence=%d", res[host], f.fences[host]) |
| 1510 | } |
| 1511 | if _, held := p.Held(host); held { |
| 1512 | t.Fatal("still held under a flat intent") |
| 1513 | } |
| 1514 | // Back to nested, still linked: held again. Then the operator removes |
| 1515 | // the second name, and the next pass moves the host and releases it. |
| 1516 | if res := byHost(t, p.Converge(context.Background(), ModeNested)); res[host].Outcome != "held" { |
| 1517 | t.Fatalf("nested again: %+v", res[host]) |
| 1518 | } |
| 1519 | os.Remove(other) |
| 1520 | res = byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1521 | if res[host].Outcome != "migrated" || f.fences[host] != 0 { |
| 1522 | t.Fatalf("after the link is gone: %+v fence=%d", res[host], f.fences[host]) |
| 1523 | } |
| 1524 | if _, held := p.Held(host); held { |
| 1525 | t.Fatal("still held after the move went through") |
| 1526 | } |
| 1527 | } |
| 1528 | |
| 1529 | // A nested host refused its revert is not held: it is already isolated, |
| 1530 | // and running is the safe side of that refusal. |
| 1531 | func TestConvergeDoesNotHoldANestedHostRefusedItsRevert(t *testing.T) { |
| 1532 | p, f := box(t) |
| 1533 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "migrated" { |
| 1534 | t.Fatalf("migrate: %+v", r) |
| 1535 | } |
| 1536 | // A database already at the flat path makes the revert refuse. |
| 1537 | if err := os.WriteFile(provision.DefaultDBPath(p.DataDir, host), []byte("stray"), 0o600); err != nil { |
| 1538 | t.Fatal(err) |
| 1539 | } |
| 1540 | r := byHost(t, p.Converge(context.Background(), ModeFlat))[host] |
| 1541 | if r.Outcome != "refused" || !strings.Contains(r.Reason, "two databases") { |
| 1542 | t.Fatalf("revert: %+v", r) |
| 1543 | } |
| 1544 | if f.fences[host] != 0 { |
| 1545 | t.Fatal("a refused revert left the nested host fenced") |
| 1546 | } |
| 1547 | } |
| 1548 | |
| 1549 | // Review F2: an awake host is deferred, but not for ever. Once it has been |
| 1550 | // seen awake for the grace, the pass stops it as it stops any host and |
| 1551 | // moves it; a host that was deferred and then found at the intent forgets |
| 1552 | // its clock. |
| 1553 | func TestConvergeStopsAnAwakeHostOnceItsGraceIsSpent(t *testing.T) { |
| 1554 | p, f := box(t) |
| 1555 | p.DeferAwake = true |
| 1556 | p.AwakeGrace = time.Hour |
| 1557 | clock := time.Unix(1700000000, 0) |
| 1558 | p.Now = func() time.Time { return clock } |
| 1559 | f.awake = map[string]bool{host: true} |
| 1560 | |
| 1561 | res := byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1562 | if r := res[host]; r.Outcome != "deferred" || !strings.Contains(r.Reason, "1h0m0s") { |
| 1563 | t.Fatalf("first pass: %+v", r) |
| 1564 | } |
| 1565 | clock = clock.Add(30 * time.Minute) |
| 1566 | res = byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1567 | if r := res[host]; r.Outcome != "deferred" || !strings.Contains(r.Reason, "30m0s") { |
| 1568 | t.Fatalf("half way: %+v", r) |
| 1569 | } |
| 1570 | for _, e := range f.events { |
| 1571 | if e == "fence "+host { |
| 1572 | t.Fatalf("the awake host was fenced inside its grace: %v", f.events) |
| 1573 | } |
| 1574 | } |
| 1575 | clock = clock.Add(30 * time.Minute) |
| 1576 | res = byHost(t, p.Converge(context.Background(), ModeNested)) |
| 1577 | if r := res[host]; r.Outcome != "migrated" { |
| 1578 | t.Fatalf("grace spent: %+v", r) |
| 1579 | } |
| 1580 | if got := rowAt(t, p, host); got != endpoints(t, p, host, ModeNested) { |
| 1581 | t.Fatalf("row = %+v", got) |
| 1582 | } |
| 1583 | // The move went through the fence, so the running tenant was stopped |
| 1584 | // for it, and released after. |
| 1585 | fenced, lifted := false, false |
| 1586 | for _, e := range f.events { |
| 1587 | fenced = fenced || e == "fence "+host |
| 1588 | lifted = lifted || e == "lift "+host |
| 1589 | } |
| 1590 | if !fenced || !lifted { |
| 1591 | t.Fatalf("events = %v, want a fence and a lift for %s", f.events, host) |
| 1592 | } |
| 1593 | if len(p.awakeSince) != 0 { |
| 1594 | t.Errorf("awake clocks still held: %v", p.awakeSince) |
| 1595 | } |
| 1596 | } |
| 1597 | |
| 1598 | // Landing follow-up 2: the hold survives an agent restart. While a host is |
| 1599 | // held its marker sits in the layout directory; a fresh pass (a new agent) |
| 1600 | // re-derives the hold from the files before anything wakes, re-holds while |
| 1601 | // the planted link persists, and drops the marker once the link is gone |
| 1602 | // or the intent is not nested. A marker the tenant deleted costs only one |
| 1603 | // grace, since the next pass meets the same refusal. |
| 1604 | func TestRestoreHoldsReholdsFromTheMarkerWhileTheReasonPersists(t *testing.T) { |
| 1605 | p, f := box(t) |
| 1606 | other := filepath.Join(p.DataDir, ".x") |
| 1607 | if err := os.Link(provision.DefaultDBPath(p.DataDir, host), other); err != nil { |
| 1608 | t.Skipf("hard links unavailable here: %v", err) |
| 1609 | } |
| 1610 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "held" { |
| 1611 | t.Fatalf("first pass: %+v", r) |
| 1612 | } |
| 1613 | marker := filepath.Join(p.DataDir, "layout", host+".hold") |
| 1614 | if b, err := os.ReadFile(marker); err != nil || !strings.Contains(string(b), "2 names") { |
| 1615 | t.Fatalf("hold marker: %q %v", b, err) |
| 1616 | } |
| 1617 | if hosts, _ := JournalHosts(p.DataDir); len(hosts) != 0 { |
| 1618 | t.Fatalf("a hold marker read as a journal: %v", hosts) |
| 1619 | } |
| 1620 | |
| 1621 | // The agent restarts: a new pass over the same registry and files, a |
| 1622 | // new fence (every tenant is stopped at start). |
| 1623 | restart := func() (*Pass, *fakeFence) { |
| 1624 | f2 := &fakeFence{} |
| 1625 | p2 := &Pass{Registry: p.Registry, Fence: f2, DataDir: p.DataDir, SocketDir: p.SocketDir, SocketsGID: -1, |
| 1626 | Now: time.Now, SweptAtStart: true, Logf: t.Logf} |
| 1627 | return p2, f2 |
| 1628 | } |
| 1629 | p2, f2 := restart() |
| 1630 | res := byHost(t, p2.RestoreHolds(context.Background(), ModeNested)) |
| 1631 | if r := res[host]; r.Outcome != "held" || !strings.Contains(r.Reason, "restored at start") { |
| 1632 | t.Fatalf("restore: %+v", r) |
| 1633 | } |
| 1634 | if f2.fences[host] == 0 { |
| 1635 | t.Fatal("the restored hold installed no fence; the host would wake flat") |
| 1636 | } |
| 1637 | if why, held := p2.Held(host); !held || !strings.Contains(why, "2 names") { |
| 1638 | t.Fatalf("Held after restore = %q %v", why, held) |
| 1639 | } |
| 1640 | // Unreadable intent is the fail-closed side: held too. |
| 1641 | p3, f3 := restart() |
| 1642 | if r := byHost(t, p3.RestoreHolds(context.Background(), IntentUnreadable))[host]; r.Outcome != "held" || f3.fences[host] == 0 { |
| 1643 | t.Fatalf("unreadable intent: %+v fence=%d", r, f3.fences[host]) |
| 1644 | } |
| 1645 | // A flat intent drops the marker and holds nothing. |
| 1646 | p4, f4 := restart() |
| 1647 | if r := byHost(t, p4.RestoreHolds(context.Background(), ModeFlat))[host]; r.Outcome != "released" || f4.fences[host] != 0 { |
| 1648 | t.Fatalf("flat intent: %+v fence=%d", r, f4.fences[host]) |
| 1649 | } |
| 1650 | if _, err := os.Lstat(marker); !errors.Is(err, os.ErrNotExist) { |
| 1651 | t.Fatalf("marker survived a flat intent: %v", err) |
| 1652 | } |
| 1653 | // Held again, then the operator removes the link: at the next start |
| 1654 | // the marker's reason is gone, so it is dropped and nothing is held; |
| 1655 | // the first pass then moves the host. |
| 1656 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "held" { |
| 1657 | t.Fatalf("re-hold: %+v", r) |
| 1658 | } |
| 1659 | os.Remove(other) |
| 1660 | p5, f5 := restart() |
| 1661 | if r := byHost(t, p5.RestoreHolds(context.Background(), ModeNested))[host]; r.Outcome != "released" || f5.fences[host] != 0 { |
| 1662 | t.Fatalf("reason gone: %+v fence=%d", r, f5.fences[host]) |
| 1663 | } |
| 1664 | if _, err := os.Lstat(marker); !errors.Is(err, os.ErrNotExist) { |
| 1665 | t.Fatalf("marker survived its reason: %v", err) |
| 1666 | } |
| 1667 | if r := byHost(t, p5.Converge(context.Background(), ModeNested))[host]; r.Outcome != "migrated" { |
| 1668 | t.Fatalf("after restore: %+v", r) |
| 1669 | } |
| 1670 | _ = f |
| 1671 | } |
| 1672 | |
| 1673 | // A hold released by the move going through takes its marker with it. |
| 1674 | func TestHoldMarkerIsRemovedWhenTheHostMoves(t *testing.T) { |
| 1675 | p, _ := box(t) |
| 1676 | other := filepath.Join(p.DataDir, ".x") |
| 1677 | if err := os.Link(provision.DefaultDBPath(p.DataDir, host), other); err != nil { |
| 1678 | t.Skipf("hard links unavailable here: %v", err) |
| 1679 | } |
| 1680 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "held" { |
| 1681 | t.Fatalf("hold: %+v", r) |
| 1682 | } |
| 1683 | marker := filepath.Join(p.DataDir, "layout", host+".hold") |
| 1684 | if _, err := os.Lstat(marker); err != nil { |
| 1685 | t.Fatal(err) |
| 1686 | } |
| 1687 | os.Remove(other) |
| 1688 | if r := byHost(t, p.Converge(context.Background(), ModeNested))[host]; r.Outcome != "migrated" { |
| 1689 | t.Fatalf("after cleanup: %+v", r) |
| 1690 | } |
| 1691 | if _, err := os.Lstat(marker); !errors.Is(err, os.ErrNotExist) { |
| 1692 | t.Fatalf("marker survived the move: %v", err) |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | // The aside directory is carlos 0700 like the rest of the layout |
| 1697 | // directory, so a flat tenant can plant a symlink at it, or a file at a |
| 1698 | // sibling's directory under it, and a refusal there would hold the |
| 1699 | // sibling (or every host) down with one command. The blocker is renamed |
| 1700 | // to a random name under /data/layout and the move goes on. |
| 1701 | func TestConvergeDisplacesABlockerAtTheAsideDirectory(t *testing.T) { |
| 1702 | for _, tc := range []struct { |
| 1703 | name string |
| 1704 | plant func(t *testing.T, dd string) |
| 1705 | }{ |
| 1706 | {"symlink at aside/", func(t *testing.T, dd string) { |
| 1707 | if err := os.MkdirAll(filepath.Join(dd, "layout"), 0o700); err != nil { |
| 1708 | t.Fatal(err) |
| 1709 | } |
| 1710 | if err := os.Symlink(t.TempDir(), filepath.Join(dd, "layout", "aside")); err != nil { |
| 1711 | t.Fatal(err) |
| 1712 | } |
| 1713 | }}, |
| 1714 | {"file at aside/<host>", func(t *testing.T, dd string) { |
| 1715 | if err := os.MkdirAll(filepath.Join(dd, "layout", "aside"), 0o700); err != nil { |
| 1716 | t.Fatal(err) |
| 1717 | } |
| 1718 | if err := os.WriteFile(filepath.Join(dd, "layout", "aside", host), []byte("planted"), 0o600); err != nil { |
| 1719 | t.Fatal(err) |
| 1720 | } |
| 1721 | }}, |
| 1722 | } { |
| 1723 | t.Run(tc.name, func(t *testing.T) { |
| 1724 | p, f := box(t) |
| 1725 | tc.plant(t, p.DataDir) |
| 1726 | // Something for the move to set aside, so the aside directory |
| 1727 | // is needed: a planted entry in the host's future directory. |
| 1728 | dir := rawInstanceDir(p.DataDir, host) |
| 1729 | if err := os.MkdirAll(dir, 0o700); err != nil { |
| 1730 | t.Fatal(err) |
| 1731 | } |
| 1732 | if err := os.WriteFile(filepath.Join(dir, "x"), []byte("x"), 0o600); err != nil { |
| 1733 | t.Fatal(err) |
| 1734 | } |
| 1735 | r := byHost(t, p.Converge(context.Background(), ModeNested))[host] |
| 1736 | if r.Outcome != "migrated" { |
| 1737 | t.Fatalf("result = %+v", r) |
| 1738 | } |
| 1739 | if f.fences[host] != 0 { |
| 1740 | t.Fatal("fence left installed") |
| 1741 | } |
| 1742 | fi, err := os.Lstat(asideDir(p.DataDir, host)) |
| 1743 | if err != nil || !fi.IsDir() { |
| 1744 | t.Fatalf("aside/<host> = %v %v, want a real directory", fi, err) |
| 1745 | } |
| 1746 | q, err := os.ReadDir(asideDir(p.DataDir, host)) |
| 1747 | if err != nil || len(q) != 1 || !strings.HasPrefix(q[0].Name(), "x.q-") { |
| 1748 | t.Fatalf("quarantine = %v %v", q, err) |
| 1749 | } |
| 1750 | entries, err := os.ReadDir(filepath.Join(p.DataDir, "layout")) |
| 1751 | if err != nil { |
| 1752 | t.Fatal(err) |
| 1753 | } |
| 1754 | blockers := 0 |
| 1755 | for _, e := range entries { |
| 1756 | if strings.Contains(e.Name(), ".blocker-") { |
| 1757 | blockers++ |
| 1758 | } |
| 1759 | } |
| 1760 | if blockers != 1 { |
| 1761 | t.Fatalf("layout dir = %v, want the one displaced blocker", entries) |
| 1762 | } |
| 1763 | if _, held := p.Held(host); held { |
| 1764 | t.Fatal("the blocker became a hold") |
| 1765 | } |
| 1766 | }) |
| 1767 | } |
| 1768 | } |
| 1769 | |
amadan