| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "slices" |
| 10 | "strings" |
| 11 | "sync" |
| 12 | "sync/atomic" |
| 13 | "testing" |
| 14 | "time" |
| 15 | |
| 16 | "github.com/carlosframework/platform/internal/cgroup" |
| 17 | "github.com/carlosframework/platform/internal/instancectl" |
| 18 | "github.com/carlosframework/platform/internal/registry" |
| 19 | "github.com/carlosframework/platform/internal/release" |
| 20 | ) |
| 21 | |
| 22 | // launchBackend is the swap fixture's fakeSwapBackend (lease, restore, |
| 23 | // raise, switch) with StartStaged on a REAL ExecBackend, flat or nested, so |
| 24 | // a staged spawn runs the production preparation: the admin token, the |
| 25 | // config read, the control socket's bind and, nested, the launch file and |
| 26 | // the root start call. The activator's row is mapped onto the real |
| 27 | // backend's row (rt) of the chosen layout. Its KillGen stops the |
| 28 | // generation's handle, as execStagedBackend's does. |
| 29 | type launchBackend struct { |
| 30 | *fakeSwapBackend |
| 31 | eb *ExecBackend |
| 32 | rt registry.Route |
| 33 | ctl *exeControl // nested only: what root was asked to start |
| 34 | |
| 35 | lmu sync.Mutex |
| 36 | entered []uint64 // generation numbers that reached the real backend |
| 37 | procs map[uint64]Proc |
| 38 | builds []string // the builds the real backend returned a process for |
| 39 | } |
| 40 | |
| 41 | func (b *launchBackend) StartStaged(ctx context.Context, _ registry.Route, g Gen) (Gen, Proc, error) { |
| 42 | b.lmu.Lock() |
| 43 | b.entered = append(b.entered, g.N) |
| 44 | b.lmu.Unlock() |
| 45 | got, p, err := b.eb.StartStaged(ctx, b.rt, g) |
| 46 | if p != nil { |
| 47 | b.lmu.Lock() |
| 48 | b.procs[g.N] = p |
| 49 | b.builds = append(b.builds, g.Build) |
| 50 | b.lmu.Unlock() |
| 51 | } |
| 52 | return got, p, err |
| 53 | } |
| 54 | |
| 55 | func (b *launchBackend) KillGen(ctx context.Context, _ registry.Route, n uint64) error { |
| 56 | b.lmu.Lock() |
| 57 | p := b.procs[n] |
| 58 | b.lmu.Unlock() |
| 59 | if p == nil { |
| 60 | return nil |
| 61 | } |
| 62 | _ = p.Stop(ctx) |
| 63 | if !p.Exited() { |
| 64 | return ErrStillRunning |
| 65 | } |
| 66 | return nil |
| 67 | } |
| 68 | |
| 69 | func (b *launchBackend) gens() []uint64 { |
| 70 | b.lmu.Lock() |
| 71 | defer b.lmu.Unlock() |
| 72 | return slices.Clone(b.entered) |
| 73 | } |
| 74 | |
| 75 | func (b *launchBackend) stopAll() { |
| 76 | b.lmu.Lock() |
| 77 | procs := make([]Proc, 0, len(b.procs)) |
| 78 | for _, p := range b.procs { |
| 79 | procs = append(procs, p) |
| 80 | } |
| 81 | b.lmu.Unlock() |
| 82 | for _, p := range procs { |
| 83 | ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second) |
| 84 | _ = p.Stop(ctx) |
| 85 | cancel() |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | // launched is the evidence that a build's process was started: the real |
| 90 | // backend returned a process for it; flat, the marker its script writes the |
| 91 | // moment it runs; nested, root was asked to start the unit with it. |
| 92 | func (b *launchBackend) launched(t *testing.T, id release.StagedIdentity, marker string) bool { |
| 93 | t.Helper() |
| 94 | b.lmu.Lock() |
| 95 | started := slices.Contains(b.builds, id.Version) |
| 96 | b.lmu.Unlock() |
| 97 | if started { |
| 98 | return true |
| 99 | } |
| 100 | if b.ctl != nil { |
| 101 | b.ctl.mu.Lock() |
| 102 | defer b.ctl.mu.Unlock() |
| 103 | for _, x := range b.ctl.got { |
| 104 | if x.Path == id.StagedPath { |
| 105 | return true |
| 106 | } |
| 107 | } |
| 108 | return false |
| 109 | } |
| 110 | _, err := os.Stat(marker) |
| 111 | return err == nil |
| 112 | } |
| 113 | |
| 114 | // prepGone checks that what generation n's preparation made is gone after |
| 115 | // a refusal at the launch: its control socket, and nested, its launch file. |
| 116 | func (b *launchBackend) prepGone(t *testing.T, n uint64) { |
| 117 | t.Helper() |
| 118 | var sock string |
| 119 | p := b.eb.pathsFor(b.rt) |
| 120 | switch p.layout { |
| 121 | case LayoutNested: |
| 122 | sock = nestedGenSocketIn(p.runDir, n) |
| 123 | launch := filepath.Join(p.dataDir, instancectl.GenLaunchFileName(n)) |
| 124 | if _, err := os.Stat(launch); !os.IsNotExist(err) { |
| 125 | t.Fatalf("generation %d's launch file %s after a refusal at the launch: %v, want it removed", n, launch, err) |
| 126 | } |
| 127 | default: |
| 128 | sock = b.eb.flatGenSocket(b.rt.Host, n) |
| 129 | } |
| 130 | if _, err := os.Stat(sock + ".control"); !os.IsNotExist(err) { |
| 131 | t.Fatalf("generation %d's control socket after a refusal at the launch: %v, want it unbound", n, err) |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | // newLaunchBackend builds the real backend for layout and wraps fb. |
| 136 | func newLaunchBackend(t *testing.T, layout string, fb *fakeSwapBackend) *launchBackend { |
| 137 | t.Helper() |
| 138 | b := &launchBackend{fakeSwapBackend: fb, procs: map[uint64]Proc{}} |
| 139 | switch layout { |
| 140 | case "flat": |
| 141 | eb := testExecBackend(t) |
| 142 | eb.RunDirOverride = t.TempDir() |
| 143 | host := "launch.example.com" |
| 144 | stageLive(t, eb, host, "v0") |
| 145 | b.eb, b.rt = eb, registry.Route{Host: host, Kind: "instance", Hibernate: true, |
| 146 | DBPath: filepath.Join(eb.DataDir, host+".db"), Socket: filepath.Join(eb.RunDirOverride, host+".sock")} |
| 147 | case "nested": |
| 148 | eb, rt := nestedBackend(t, &fakeControl{}, &fakeUnit{}) |
| 149 | eb.GenContainment = true |
| 150 | b.ctl = &exeControl{} |
| 151 | eb.Control = b.ctl |
| 152 | b.eb, b.rt = eb, rt |
| 153 | default: |
| 154 | t.Fatalf("layout %q", layout) |
| 155 | } |
| 156 | if got, _ := b.eb.Layout(b.rt); (layout == "flat") != (got == LayoutFlat) || (layout == "nested") != (got == LayoutNested) { |
| 157 | t.Fatalf("fixture layout %v, want %s", got, layout) |
| 158 | } |
| 159 | t.Cleanup(b.stopAll) |
| 160 | return b |
| 161 | } |
| 162 | |
| 163 | // markerScript is a staged build that writes marker the moment it runs. |
| 164 | func markerScript(t *testing.T, dir, ver, marker string) release.StagedIdentity { |
| 165 | t.Helper() |
| 166 | return stagedScript(t, dir, ver, "#!/bin/sh\necho ran > '"+marker+"'\nexec sleep 30\n") |
| 167 | } |
| 168 | |
| 169 | // onCall runs f on the call numbered want (counting from 1) and on no |
| 170 | // other. |
| 171 | func onCall(want int32, f func()) func(string) { |
| 172 | var n atomic.Int32 |
| 173 | return func(string) { |
| 174 | if n.Add(1) == want { |
| 175 | f() |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | // Final review F1: every staged instance spawn's composed gate is asked at |
| 181 | // the backend's final launch point, AFTER the backend's own preparation: |
| 182 | // flat, after the token, the config read and the control socket's bind, |
| 183 | // right before the process starts; nested, after the launch file, right |
| 184 | // before root is asked to start the unit. A lease lost, a fence set or the |
| 185 | // attempt superseded during that preparation starts nothing, removes what |
| 186 | // the preparation made, records no execution evidence and is abandonment, |
| 187 | // never a failure of the build. The four callers: a planned wake's B, its |
| 188 | // fallback, a hold's pinned build and an overlap successor, on both |
| 189 | // layouts. Before the fix the gate ran only before StartStaged, and each of |
| 190 | // these launched. |
| 191 | func TestTheExecGateIsAskedAtTheBackendsLaunchPoint(t *testing.T) { |
| 192 | for _, layout := range []string{"flat", "nested"} { |
| 193 | t.Run(layout+"/planned B, lease lost", func(t *testing.T) { |
| 194 | b := newLaunchBackend(t, layout, &fakeSwapBackend{fakeBackend: &fakeBackend{}}) |
| 195 | dir := t.TempDir() |
| 196 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 197 | next := markerScript(t, dir, "v2", mark) |
| 198 | good := markerScript(t, dir, "v1", filepath.Join(t.TempDir(), "v1.ran")) |
| 199 | a, evs := launchActivator(t, b, WakePlan{Boot: next, Mode: ModeOverlap, FirstWake: true, Fallback: &good}) |
| 200 | b.eb.launchHook = onCall(1, func() { b.lastRep().setErr(errors.New("lease renewal failed")) }) |
| 201 | _, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 202 | if !errors.Is(err, errSwapLeaseLost) || classifyStep(nil, err) == stepFailed { |
| 203 | t.Fatalf("wake = %v, want the lost lease, abandonment", err) |
| 204 | } |
| 205 | gens := b.gens() |
| 206 | if len(gens) != 1 { |
| 207 | t.Fatalf("the backend was entered for generations %v, want B's alone (the refusal is inside it)", gens) |
| 208 | } |
| 209 | if b.launched(t, next, mark) { |
| 210 | t.Fatal("B was launched after the lease was lost during its preparation") |
| 211 | } |
| 212 | b.prepGone(t, gens[0]) |
| 213 | time.Sleep(50 * time.Millisecond) |
| 214 | if k := evs.kinds(); len(k) != 0 { |
| 215 | t.Fatalf("events %v: a refusal at the launch was recorded", k) |
| 216 | } |
| 217 | }) |
| 218 | t.Run(layout+"/planned fallback, fenced", func(t *testing.T) { |
| 219 | defer func(k time.Duration) { firstWakeKillBound = k }(firstWakeKillBound) |
| 220 | firstWakeKillBound = 500 * time.Millisecond |
| 221 | b := newLaunchBackend(t, layout, &fakeSwapBackend{fakeBackend: &fakeBackend{}}) |
| 222 | dir := t.TempDir() |
| 223 | markB, markGood := filepath.Join(t.TempDir(), "v2.ran"), filepath.Join(t.TempDir(), "v1.ran") |
| 224 | next, good := markerScript(t, dir, "v2", markB), markerScript(t, dir, "v1", markGood) |
| 225 | a, evs := launchActivator(t, b, WakePlan{Boot: next, Mode: ModeOverlap, FirstWake: true, Fallback: &good}) |
| 226 | const host = "one.example.com" |
| 227 | a.genHealth = func(context.Context, Gen) error { return errors.New("GET / answered 500") } |
| 228 | b.eb.launchHook = onCall(2, func() { |
| 229 | a.mu.Lock() |
| 230 | a.fences = map[string]uint64{host: 1} |
| 231 | a.mu.Unlock() |
| 232 | }) |
| 233 | if _, err := a.beginRequest(context.Background(), testRoute(host)); err == nil { |
| 234 | t.Fatal("served after the fallback was refused") |
| 235 | } |
| 236 | if gens := b.gens(); len(gens) != 2 { |
| 237 | t.Fatalf("the backend was entered for generations %v, want B's and the fallback's", gens) |
| 238 | } |
| 239 | waitFor(t, "B's launch evidence", func() bool { return b.launched(t, next, markB) }) |
| 240 | time.Sleep(100 * time.Millisecond) |
| 241 | if b.launched(t, good, markGood) { |
| 242 | t.Fatal("the good build was launched on a host fenced during its preparation") |
| 243 | } |
| 244 | b.prepGone(t, b.gens()[1]) |
| 245 | // B ran and failed: that is recorded. The refused fallback is not |
| 246 | // a fallback that served. |
| 247 | evs.wait(t, GenFailed) |
| 248 | if k := evs.kinds(); slices.Contains(k, GenFellBack) { |
| 249 | t.Fatalf("events %v: a fallback refused at its launch was reported as served", k) |
| 250 | } |
| 251 | }) |
| 252 | t.Run(layout+"/hold's pinned build, lease lost", func(t *testing.T) { |
| 253 | f := newSwapFixture(t) |
| 254 | b := newLaunchBackend(t, layout, f.b) |
| 255 | f.a.backend = b |
| 256 | sink := &lockedSink{} |
| 257 | f.a.Logs = sink |
| 258 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 259 | id := markerScript(t, t.TempDir(), "v2", mark) |
| 260 | at := f.attempt(ModeHold) |
| 261 | at.StagedPath, at.Digest = id.StagedPath, id.Digest |
| 262 | b.eb.launchHook = onCall(1, func() { f.b.lastRep().setErr(errors.New("lease renewal failed")) }) |
| 263 | res := f.a.Swap(context.Background(), at, f.newest) |
| 264 | if res.Outcome != SwapAbandoned || !strings.Contains(res.Reason, errSwapLeaseLost.Error()) { |
| 265 | t.Fatalf("Swap = %+v, want abandoned: the lost lease", res) |
| 266 | } |
| 267 | gens := b.gens() |
| 268 | if len(gens) != 1 { |
| 269 | t.Fatalf("the backend was entered for generations %v, want the pin's alone", gens) |
| 270 | } |
| 271 | if b.launched(t, id, mark) { |
| 272 | t.Fatal("the hold's build was launched after the lease was lost during its preparation") |
| 273 | } |
| 274 | b.prepGone(t, gens[0]) |
| 275 | if got := sink.errors(); len(got) != 0 { |
| 276 | t.Fatalf("member error events %q for an abandoned hold", got) |
| 277 | } |
| 278 | }) |
| 279 | t.Run(layout+"/overlap successor, superseded", func(t *testing.T) { |
| 280 | f := newSwapFixture(t) |
| 281 | b := newLaunchBackend(t, layout, f.b) |
| 282 | f.a.backend = b |
| 283 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 284 | id := markerScript(t, t.TempDir(), "v2", mark) |
| 285 | at := f.attempt(ModeOverlap) |
| 286 | at.StagedPath, at.Digest = id.StagedPath, id.Digest |
| 287 | b.eb.launchHook = onCall(1, func() { f.rev.Add(1) }) |
| 288 | res := f.a.Swap(context.Background(), at, f.newest) |
| 289 | if res.Outcome != SwapAbandoned || !strings.Contains(res.Reason, errSwapSuperseded.Error()) { |
| 290 | t.Fatalf("Swap = %+v, want abandoned: superseded", res) |
| 291 | } |
| 292 | gens := b.gens() |
| 293 | if len(gens) != 1 { |
| 294 | t.Fatalf("the backend was entered for generations %v, want the successor's alone", gens) |
| 295 | } |
| 296 | if b.launched(t, id, mark) { |
| 297 | t.Fatal("the successor was launched after its attempt was superseded during its preparation") |
| 298 | } |
| 299 | b.prepGone(t, gens[0]) |
| 300 | if g, members := f.a.genStateForTest(f.host); g.N != f.incumbent.N || len(members) != 0 { |
| 301 | t.Fatalf("current %+v, members %v: want the incumbent serving and an empty operation set", g, members) |
| 302 | } |
| 303 | }) |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | // launchActivator is planActivator over a launchBackend. |
| 308 | func launchActivator(t *testing.T, b *launchBackend, plan WakePlan) (*Service, *genLog) { |
| 309 | t.Helper() |
| 310 | reg, err := registry.Open(filepath.Join(t.TempDir(), "registry.db")) |
| 311 | if err != nil { |
| 312 | t.Fatal(err) |
| 313 | } |
| 314 | a := New(Config{Bucket: "test-bucket", Idle: time.Hour}, b, reg) |
| 315 | a.gens = true |
| 316 | evs := &genLog{} |
| 317 | a.OnGen = evs.add |
| 318 | a.Planner = func(context.Context, registry.Route) WakePlan { return plan } |
| 319 | a.genHealth = func(context.Context, Gen) error { return nil } |
| 320 | t.Cleanup(func() { a.Shutdown(context.Background()) }) |
| 321 | return a, evs |
| 322 | } |
| 323 | |
| 324 | // Merge triage T9: on a first wake whose health window is short (under the |
| 325 | // verdict floor, so B would not be tried and the good build would boot in |
| 326 | // its place), a refusal by B's own gate is abandonment: nothing boots, B is |
| 327 | // not reported untried, and nothing is recorded. verdictFloorGate asks the |
| 328 | // gate first, so a build held since the plan (or a lost lease, a fence) is |
| 329 | // never mistaken for a lack of time. Here B's selection check passes at |
| 330 | // checkPlan and refuses at B's exec. With the floor asked first, the good |
| 331 | // build booted untried and the wake reported GenHealthy. |
| 332 | func TestAGateRefusalOnAShortWindowAbandonsRatherThanGoingUntried(t *testing.T) { |
| 333 | shrinkVerdictFloor(t, time.Hour) // every window is under the floor |
| 334 | dir := t.TempDir() |
| 335 | good, next := stagedID(t, dir, "v1"), stagedID(t, dir, "v2") |
| 336 | a, b, evs := planActivator(t, WakePlan{}) |
| 337 | var asked atomic.Int32 |
| 338 | a.Planner = func(context.Context, registry.Route) WakePlan { |
| 339 | return WakePlan{Boot: next, Mode: ModeOverlap, FirstWake: true, Fallback: &good, |
| 340 | Check: func(build string) error { |
| 341 | if build == "v2" && asked.Add(1) >= 2 { |
| 342 | return errors.New("v2 failed on a sibling") |
| 343 | } |
| 344 | return nil |
| 345 | }} |
| 346 | } |
| 347 | _, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 348 | if !errors.Is(err, errPlanStale) || errors.Is(err, errNoTimeForVerdict) { |
| 349 | t.Fatalf("wake = %v, want the gate's refusal (errPlanStale), never a lack of time", err) |
| 350 | } |
| 351 | if got := classifyStep(nil, err); got != stepStale { |
| 352 | t.Fatalf("classifyStep = %v, want stepStale: abandonment", got) |
| 353 | } |
| 354 | if gens := b.started(); len(gens) != 0 { |
| 355 | t.Fatalf("spawned %+v: the good build booted untried after B's gate refused", gens) |
| 356 | } |
| 357 | time.Sleep(50 * time.Millisecond) |
| 358 | if k := evs.kinds(); len(k) != 0 { |
| 359 | t.Fatalf("events %v: an abandoned short-window wake reported something (an untried B reads as GenHealthy)", k) |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | // The exec gate rides on the context through SplitBackend to its exec half, |
| 364 | // which asks it at its launch (genbackend.go's contract): a refusal there |
| 365 | // reaches the caller unchanged. |
| 366 | func TestSplitBackendCarriesTheExecGate(t *testing.T) { |
| 367 | exec := &recordingGenExec{fakeBackend: &fakeBackend{}} |
| 368 | s := &SplitBackend{Exec: exec, Unit: &UnitBackend{}} |
| 369 | refused := errors.New("refused at the launch") |
| 370 | ctx := withExecGate(context.Background(), func() error { return refused }) |
| 371 | _, p, err := s.StartStaged(ctx, registry.Route{Host: "e.example.com"}, Gen{N: 7, StagedPath: "/opt/carlos/versions/v2/app"}) |
| 372 | if !errors.Is(err, refused) || p != nil || exec.started != 0 { |
| 373 | t.Fatalf("StartStaged = %v, %v (started %d): want the gate's refusal and nothing started", p, err, exec.started) |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | // gatedCgroups gives b a cgroup manager on a temp-dir subtree (the |
| 378 | // fixture delegatedBackend uses) whose log calls flip once a limit is |
| 379 | // written: a log line naming a memory.max is the manager's preparation |
| 380 | // (the parent's tier, the child's limits) under way. |
| 381 | func gatedCgroups(t *testing.T, b *ExecBackend, host string, flip func()) string { |
| 382 | t.Helper() |
| 383 | root := t.TempDir() |
| 384 | for _, f := range []string{"cgroup.procs", "cgroup.subtree_control"} { |
| 385 | if err := os.WriteFile(filepath.Join(root, f), nil, 0o644); err != nil { |
| 386 | t.Fatal(err) |
| 387 | } |
| 388 | } |
| 389 | var once sync.Once |
| 390 | m, err := cgroup.SetupAt(root, func(format string, args ...any) { |
| 391 | t.Logf(format, args...) |
| 392 | if flip != nil && strings.Contains(format, "memory.max=") { |
| 393 | once.Do(flip) |
| 394 | } |
| 395 | }) |
| 396 | if err != nil { |
| 397 | t.Fatal(err) |
| 398 | } |
| 399 | parent := filepath.Join(root, cgroup.ExecName(host)) |
| 400 | if err := os.MkdirAll(parent, 0o755); err != nil { |
| 401 | t.Fatal(err) |
| 402 | } |
| 403 | if err := os.WriteFile(filepath.Join(parent, "cgroup.controllers"), []byte("cpu io memory pids\n"), 0o644); err != nil { |
| 404 | t.Fatal(err) |
| 405 | } |
| 406 | b.Cgroups = m |
| 407 | return parent |
| 408 | } |
| 409 | |
| 410 | // workerGateService is a Service over a REAL ExecBackend worker spawn |
| 411 | // (realWorkerBackend) whose cgroup manager is gatedCgroups' temp-dir |
| 412 | // subtree: a worker start runs the production preparation (the build's |
| 413 | // publication, the cgroup and its limits) and its start attempts. It |
| 414 | // returns the worker's row, the file its build writes its $0 to the moment |
| 415 | // it runs, and the host's cgroup directory. |
| 416 | func workerGateService(t *testing.T, flip func()) (*Service, *ExecBackend, registry.Route, string, string) { |
| 417 | t.Helper() |
| 418 | eb := testExecBackend(t) |
| 419 | host := "bot.example.com" |
| 420 | out := filepath.Join(t.TempDir(), "argv0") |
| 421 | stageArgv0Worker(t, eb, host, "v1", out) |
| 422 | dir := gatedCgroups(t, eb, host, flip) |
| 423 | reg, err := registry.Open(filepath.Join(t.TempDir(), "registry.db")) |
| 424 | if err != nil { |
| 425 | t.Fatal(err) |
| 426 | } |
| 427 | a := New(Config{Bucket: "test-bucket", Idle: time.Hour}, &realWorkerBackend{fakeBackend: &fakeBackend{}, exec: eb}, reg) |
| 428 | a.GenerationsFile = release.GenerationsPath(t.TempDir()) |
| 429 | t.Cleanup(func() { a.Shutdown(context.Background()) }) |
| 430 | return a, eb, workerRoute(host), out, dir |
| 431 | } |
| 432 | |
| 433 | // Astra R3 §5 item 4 (ruling 206, "park worker launch hook"): a worker |
| 434 | // start's ownership gate is the cgroup manager's before-start hook, as an |
| 435 | // instance's is, so it is asked AFTER the cgroup preparation (the host's |
| 436 | // cgroup and its limits) and before the start. Ownership lost during that |
| 437 | // preparation starts nothing: no process, no $0 written, nothing placed in |
| 438 | // the cgroup, the gate's own error back, and the build's pending |
| 439 | // publication dropped. Before the fix the gate rode the before-exec hook, |
| 440 | // asked once before the cgroup preparation, and the worker started. |
| 441 | func TestAWorkerStartIsGatedAfterItsCgroupPreparation(t *testing.T) { |
| 442 | t.Parallel() |
| 443 | var refuse atomic.Bool |
| 444 | a, _, rt, out, dir := workerGateService(t, func() { refuse.Store(true) }) |
| 445 | lost := errors.New("lease renewal failed") |
| 446 | var asked atomic.Int32 |
| 447 | gate := func() error { |
| 448 | asked.Add(1) |
| 449 | if refuse.Load() { |
| 450 | return lost |
| 451 | } |
| 452 | return nil |
| 453 | } |
| 454 | p, _, _, err := a.startWorker(context.Background(), rt, nil, nil, gate) |
| 455 | if p != nil { |
| 456 | _ = p.Stop(context.Background()) |
| 457 | } |
| 458 | if !errors.Is(err, lost) || p != nil { |
| 459 | t.Fatalf("startWorker = %v, %v; want no process and the gate's refusal", p, err) |
| 460 | } |
| 461 | if !refuse.Load() || asked.Load() == 0 { |
| 462 | t.Fatalf("the cgroup preparation ran: %v; the gate was asked %d times", refuse.Load(), asked.Load()) |
| 463 | } |
| 464 | time.Sleep(200 * time.Millisecond) |
| 465 | if _, err := os.Stat(out); err == nil { |
| 466 | t.Fatal("the worker ran after its ownership was lost during the cgroup preparation") |
| 467 | } |
| 468 | if b, err := os.ReadFile(filepath.Join(dir, "cgroup.procs")); err == nil && len(b) != 0 { |
| 469 | t.Fatalf("the worker's cgroup was given pid %s after a refusal at its start", b) |
| 470 | } |
| 471 | if got := readGenerations(t, a.GenerationsFile); len(got[rt.Host]) != 0 { |
| 472 | t.Fatalf("generations file %v names a build that never ran", got) |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | // The same gate is the manager's before hook at the PID-move fallback a |
| 477 | // flat start takes after a failed placement (cgroup.StartGated's |
| 478 | // contract), so ownership lost by then starts nothing. The temp-dir subtree |
| 479 | // is no cgroupfs: the shim's checks refuse the placement before any fork, |
| 480 | // so the gate (asked after a placement) is first asked at the fallback, and |
| 481 | // refuses there. That a placement which got past the gate and then failed |
| 482 | // asks it again at the fallback is internal/cgroup's |
| 483 | // TestFlatPlacementFailureFallsBackToPIDMove: the hook passed through is |
| 484 | // this same gate. Before the clone3 start left (amadan #272) this asked |
| 485 | // twice, at the clone3 start and its retry. |
| 486 | func TestAWorkerStartFallbackAsksTheGate(t *testing.T) { |
| 487 | t.Parallel() |
| 488 | a, eb, rt, out, dir := workerGateService(t, nil) |
| 489 | var asked atomic.Int32 |
| 490 | eb.launchHook = func(string) { asked.Add(1) } |
| 491 | superseded := errors.New("the attempt was superseded") |
| 492 | gate := func() error { |
| 493 | if asked.Load() >= 1 { |
| 494 | return superseded |
| 495 | } |
| 496 | return nil |
| 497 | } |
| 498 | p, _, _, err := a.startWorker(context.Background(), rt, nil, nil, gate) |
| 499 | if p != nil { |
| 500 | _ = p.Stop(context.Background()) |
| 501 | } |
| 502 | if n := asked.Load(); n != 1 { |
| 503 | t.Fatalf("the gate was asked %d times at the starts, want once: the PID-move fallback's", n) |
| 504 | } |
| 505 | if b, err := os.ReadFile(filepath.Join(dir, "cgroup.procs")); err == nil && len(b) != 0 { |
| 506 | t.Fatalf("the worker's cgroup was given pid %s after a refusal at its start", b) |
| 507 | } |
| 508 | if !errors.Is(err, superseded) || p != nil { |
| 509 | t.Fatalf("startWorker = %v, %v; want no process and the gate's refusal", p, err) |
| 510 | } |
| 511 | time.Sleep(200 * time.Millisecond) |
| 512 | if _, err := os.Stat(out); err == nil { |
| 513 | t.Fatal("the retry started the worker after its gate refused") |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | // Final review wave 2 (F1 completion, Astra round 2 finding 1): a flat |
| 518 | // staged spawn's exec gate is the cgroup manager's before-start hook, so |
| 519 | // it is asked AFTER the cgroup preparation (the parent's tier, the child |
| 520 | // cgroup and its limits, their log lines) and the placement, before EACH |
| 521 | // exec, a flat start's PID-move fallback included. Ownership lost during |
| 522 | // that preparation, or between a failed placement and its fallback, |
| 523 | // starts nothing: no process, no marker, the control socket unbound, the |
| 524 | // generation's child cgroup removed, abandonment with no execution |
| 525 | // evidence. Before the fix the gate ran once, before the cgroup |
| 526 | // preparation, and each of these launched. |
| 527 | func TestTheExecGateIsAskedAtEachCgroupStart(t *testing.T) { |
| 528 | t.Run("overlap successor, superseded during the cgroup preparation", func(t *testing.T) { |
| 529 | f := newSwapFixture(t) |
| 530 | b := newLaunchBackend(t, "flat", f.b) |
| 531 | b.eb.GenContainment = true |
| 532 | parent := gatedCgroups(t, b.eb, b.rt.Host, func() { f.rev.Add(1) }) |
| 533 | cgroup.FakePlacementForTest(t) // a temp dir is no cgroupfs, and a generation is never moved by PID |
| 534 | f.a.backend = b |
| 535 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 536 | id := markerScript(t, t.TempDir(), "v2", mark) |
| 537 | at := f.attempt(ModeOverlap) |
| 538 | at.StagedPath, at.Digest = id.StagedPath, id.Digest |
| 539 | res := f.a.Swap(context.Background(), at, f.newest) |
| 540 | if res.Outcome != SwapAbandoned || !strings.Contains(res.Reason, errSwapSuperseded.Error()) || res.Executed { |
| 541 | t.Fatalf("Swap = %+v, want abandoned (superseded), nothing executed", res) |
| 542 | } |
| 543 | gens := b.gens() |
| 544 | if len(gens) != 1 { |
| 545 | t.Fatalf("the backend was entered for generations %v, want the successor's alone", gens) |
| 546 | } |
| 547 | time.Sleep(100 * time.Millisecond) |
| 548 | if b.launched(t, id, mark) { |
| 549 | t.Fatal("the successor was launched after its attempt was superseded during the cgroup preparation") |
| 550 | } |
| 551 | b.prepGone(t, gens[0]) |
| 552 | // The preparation did run (the flip is its log line), and nothing was |
| 553 | // placed in the child it made. (On cgroupfs the empty child is |
| 554 | // removed; in this fixture its limit files are ordinary files, so the |
| 555 | // rmdir is refused, as CleanupHost's comment explains.) |
| 556 | if b, err := os.ReadFile(filepath.Join(parent, cgroup.GenChild(gens[0]), "cgroup.procs")); err == nil && len(b) != 0 { |
| 557 | t.Fatalf("generation %d's child cgroup was given pid %s after a refusal at its start", gens[0], b) |
| 558 | } |
| 559 | }) |
| 560 | t.Run("hold's pinned build, lease lost during the cgroup preparation", func(t *testing.T) { |
| 561 | f := newSwapFixture(t) |
| 562 | b := newLaunchBackend(t, "flat", f.b) |
| 563 | gatedCgroups(t, b.eb, b.rt.Host, func() { f.b.lastRep().setErr(errors.New("lease renewal failed")) }) |
| 564 | f.a.backend = b |
| 565 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 566 | id := markerScript(t, t.TempDir(), "v2", mark) |
| 567 | at := f.attempt(ModeHold) |
| 568 | at.StagedPath, at.Digest = id.StagedPath, id.Digest |
| 569 | res := f.a.Swap(context.Background(), at, f.newest) |
| 570 | if res.Outcome != SwapAbandoned || !strings.Contains(res.Reason, errSwapLeaseLost.Error()) || res.Executed { |
| 571 | t.Fatalf("Swap = %+v, want abandoned (the lost lease), nothing executed", res) |
| 572 | } |
| 573 | time.Sleep(100 * time.Millisecond) |
| 574 | if b.launched(t, id, mark) { |
| 575 | t.Fatal("the hold's build was launched after the lease was lost during the cgroup preparation") |
| 576 | } |
| 577 | b.prepGone(t, b.gens()[0]) |
| 578 | }) |
| 579 | t.Run("overlap successor, a failed placement is not retried by PID move", func(t *testing.T) { |
| 580 | f := newSwapFixture(t) |
| 581 | b := newLaunchBackend(t, "flat", f.b) |
| 582 | b.eb.GenContainment = true |
| 583 | // The temp-dir subtree is no cgroupfs: the shim's checks refuse the |
| 584 | // placement before any fork, so the gate, asked only after a |
| 585 | // placement (amadan #272, Ruling 13), is never asked. A generation is |
| 586 | // never started and then moved by PID (readiness round 2 Task 7), so |
| 587 | // there is no retry and no gate at all: the start is refused as the |
| 588 | // box's condition. |
| 589 | parent := gatedCgroups(t, b.eb, b.rt.Host, nil) |
| 590 | var asked atomic.Int32 |
| 591 | b.eb.launchHook = func(string) { asked.Add(1) } |
| 592 | f.a.backend = b |
| 593 | mark := filepath.Join(t.TempDir(), "v2.ran") |
| 594 | id := markerScript(t, t.TempDir(), "v2", mark) |
| 595 | at := f.attempt(ModeOverlap) |
| 596 | at.StagedPath, at.Digest = id.StagedPath, id.Digest |
| 597 | res := f.a.Swap(context.Background(), at, f.newest) |
| 598 | if n := asked.Load(); n != 0 { |
| 599 | t.Fatalf("the gate was asked %d times, want never: no placement, and no retry", n) |
| 600 | } |
| 601 | if res.Outcome != SwapAbandoned || !strings.Contains(res.Reason, cgroup.ErrLimitNotApplied.Error()) || res.Executed { |
| 602 | t.Fatalf("Swap = %+v, want abandoned (limits not applied), nothing executed", res) |
| 603 | } |
| 604 | time.Sleep(100 * time.Millisecond) |
| 605 | if b.launched(t, id, mark) { |
| 606 | t.Fatal("the successor was launched after its placement failed") |
| 607 | } |
| 608 | gens := b.gens() |
| 609 | b.prepGone(t, gens[0]) |
| 610 | // No PID move. In this fixture the child's limit files are ordinary |
| 611 | // files, so the refusal's rmdir leaves the child in place: it must |
| 612 | // still be there with its memory.max written (else the check below |
| 613 | // proves nothing), and with no cgroup.procs at all, which is the |
| 614 | // file a start-and-move writes the PID into. |
| 615 | child := filepath.Join(parent, cgroup.GenChild(gens[0])) |
| 616 | if _, err := os.Stat(filepath.Join(child, "memory.max")); err != nil { |
| 617 | t.Fatalf("generation %d's child cgroup has no limit written (%v): the start never reached its placement", gens[0], err) |
| 618 | } |
| 619 | if got, err := os.ReadFile(filepath.Join(child, "cgroup.procs")); !os.IsNotExist(err) { |
| 620 | t.Fatalf("generation %d's child cgroup has a cgroup.procs (%q, %v): a PID move wrote it", gens[0], got, err) |
| 621 | } |
| 622 | }) |
| 623 | } |
| 624 | |
| 625 | // liveGate is a LiveExecGate for tests: it records every build it is asked |
| 626 | // about and refuses the builds in refuse, as a declined box's failure holds |
| 627 | // would (internal/host). |
| 628 | type liveGate struct { |
| 629 | mu sync.Mutex |
| 630 | asked []string |
| 631 | refuse map[string]bool |
| 632 | } |
| 633 | |
| 634 | var errLiveHeld = errors.New("a failure record holds this build") |
| 635 | |
| 636 | func (g *liveGate) gate(_ registry.Route, build string) error { |
| 637 | g.mu.Lock() |
| 638 | defer g.mu.Unlock() |
| 639 | g.asked = append(g.asked, build) |
| 640 | if g.refuse[build] { |
| 641 | return fmt.Errorf("%w: %s", errLiveHeld, build) |
| 642 | } |
| 643 | return nil |
| 644 | } |
| 645 | |
| 646 | func (g *liveGate) builds() []string { |
| 647 | g.mu.Lock() |
| 648 | defer g.mu.Unlock() |
| 649 | return slices.Clone(g.asked) |
| 650 | } |
| 651 | |
| 652 | // repointLive points b's live/<host> at bin, as root's adopt pass does. |
| 653 | func repointLive(t *testing.T, b *ExecBackend, host, bin string) { |
| 654 | t.Helper() |
| 655 | tmp := b.liveLink(host) + ".tmp" |
| 656 | os.Remove(tmp) |
| 657 | if err := os.Symlink(bin, tmp); err != nil { |
| 658 | t.Error(err) |
| 659 | return |
| 660 | } |
| 661 | if err := os.Rename(tmp, b.liveLink(host)); err != nil { |
| 662 | t.Error(err) |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | // flatLiveFixture is a flat row on a real ExecBackend whose live link names |
| 667 | // v1, with v2 staged beside it. Each build writes its own $0 to out the |
| 668 | // moment it runs, so a test can tell which bytes were exec'd and by which |
| 669 | // path. |
| 670 | func flatLiveFixture(t *testing.T) (b *ExecBackend, rt registry.Route, v1, v2, out string) { |
| 671 | t.Helper() |
| 672 | b = testExecBackend(t) |
| 673 | b.RunDirOverride = t.TempDir() |
| 674 | host := "live.example.com" |
| 675 | out = filepath.Join(t.TempDir(), "argv0") |
| 676 | v2 = stageArgv0Worker(t, b, host, "v2", out) |
| 677 | v1 = stageArgv0Worker(t, b, host, "v1", out) |
| 678 | rt = registry.Route{Host: host, Kind: "instance", Hibernate: true, |
| 679 | DBPath: filepath.Join(b.DataDir, host+".db"), Socket: filepath.Join(b.RunDirOverride, host+".sock")} |
| 680 | return b, rt, v1, v2, out |
| 681 | } |
| 682 | |
| 683 | // Task 9 (Astra R1 #6, the R2/R3 declined residuals; plan review #6): on a |
| 684 | // box that declined CARLOS_SWAP=overlap and honours failure records, the |
| 685 | // flat live-link spawn resolves live/<host> ONCE, asks LiveExecGate about |
| 686 | // that build at the launch (after the cgroup's preparation), and execs the |
| 687 | // staged path it resolved, so a repoint landing between the gate and the |
| 688 | // exec cannot start a build the gate never saw. A refusal starts nothing. |
| 689 | // With the gate nil the spawn execs the live link exactly as before. |
| 690 | func TestALiveExecGateGatesTheFlatLiveLinkSpawn(t *testing.T) { |
| 691 | t.Parallel() |
| 692 | t.Run("gated build is the exec'd build", func(t *testing.T) { |
| 693 | b, rt, v1, v2, out := flatLiveFixture(t) |
| 694 | g := &liveGate{refuse: map[string]bool{"v2": true}} |
| 695 | b.LiveExecGate = g.gate |
| 696 | // Root repoints at held v2 after the gate has been asked about v1 and |
| 697 | // before the exec. |
| 698 | b.launchHook = onCall(1, func() { repointLive(t, b, rt.Host, v2) }) |
| 699 | p, ver, err := b.spawnChildFrom(context.Background(), rt, rt.Socket, 0, "", cgroup.Limit{}) |
| 700 | if err != nil { |
| 701 | t.Fatalf("spawn = %v", err) |
| 702 | } |
| 703 | defer p.Stop(context.Background()) |
| 704 | if got := strings.TrimSpace(waitForFile(t, out, 5*time.Second)); got != v1 { |
| 705 | t.Fatalf("$0 = %q, want the staged v1 the gate was asked about (%s)", got, v1) |
| 706 | } |
| 707 | if ver != "v1" || !slices.Equal(g.builds(), []string{"v1"}) { |
| 708 | t.Fatalf("spawn version %q, gate asked %v; want v1 and [v1]", ver, g.builds()) |
| 709 | } |
| 710 | }) |
| 711 | t.Run("a held build starts nothing", func(t *testing.T) { |
| 712 | b, rt, _, v2, out := flatLiveFixture(t) |
| 713 | repointLive(t, b, rt.Host, v2) |
| 714 | g := &liveGate{refuse: map[string]bool{"v2": true}} |
| 715 | b.LiveExecGate = g.gate |
| 716 | p, _, err := b.spawnChildFrom(context.Background(), rt, rt.Socket, 0, "", cgroup.Limit{}) |
| 717 | if p != nil { |
| 718 | _ = p.Stop(context.Background()) |
| 719 | } |
| 720 | if !errors.Is(err, errLiveHeld) || p != nil { |
| 721 | t.Fatalf("spawn = %v, %v; want no process and the gate's refusal", p, err) |
| 722 | } |
| 723 | time.Sleep(200 * time.Millisecond) |
| 724 | if _, err := os.Stat(out); err == nil { |
| 725 | t.Fatal("held v2 ran") |
| 726 | } |
| 727 | if _, err := os.Stat(rt.Socket + ".control"); !os.IsNotExist(err) { |
| 728 | t.Fatalf("control socket after a refusal: %v, want it unbound", err) |
| 729 | } |
| 730 | if !slices.Equal(g.builds(), []string{"v2"}) { |
| 731 | t.Fatalf("gate asked %v, want [v2]", g.builds()) |
| 732 | } |
| 733 | }) |
| 734 | t.Run("no gate is today's spawn", func(t *testing.T) { |
| 735 | b, rt, _, _, out := flatLiveFixture(t) |
| 736 | p, ver, err := b.spawnChildFrom(context.Background(), rt, rt.Socket, 0, "", cgroup.Limit{}) |
| 737 | if err != nil { |
| 738 | t.Fatalf("spawn = %v", err) |
| 739 | } |
| 740 | defer p.Stop(context.Background()) |
| 741 | if got := strings.TrimSpace(waitForFile(t, out, 5*time.Second)); got != b.liveLink(rt.Host) || ver != "v1" { |
| 742 | t.Fatalf("$0 = %q, version %q; want today's exec of %s, v1", got, ver, b.liveLink(rt.Host)) |
| 743 | } |
| 744 | }) |
| 745 | } |
| 746 | |
| 747 | // Task 9: a nested live-link start asks LiveExecGate about the build the |
| 748 | // live link names, right before root is asked to start the unit. A refusal |
| 749 | // asks root for nothing. |
| 750 | func TestALiveExecGateGatesTheNestedLiveLinkStart(t *testing.T) { |
| 751 | ctl := &fakeControl{inv: "inv1"} |
| 752 | b, rt := nestedBackend(t, ctl, &fakeUnit{}) |
| 753 | g := &liveGate{refuse: map[string]bool{"v1": true}} |
| 754 | b.LiveExecGate = g.gate |
| 755 | _, _, err := b.Start(context.Background(), rt) |
| 756 | if !errors.Is(err, errLiveHeld) { |
| 757 | t.Fatalf("Start = %v, want the gate's refusal", err) |
| 758 | } |
| 759 | if strings.Contains(ctl.joined(), "start ") { |
| 760 | t.Fatalf("root was asked to start the unit after the gate refused: %q", ctl.joined()) |
| 761 | } |
| 762 | if !slices.Equal(g.builds(), []string{"v1"}) { |
| 763 | t.Fatalf("gate asked %v, want [v1]", g.builds()) |
| 764 | } |
| 765 | if _, err := os.Stat(rt.Socket + ".control"); !os.IsNotExist(err) { |
| 766 | t.Fatalf("control socket after a refusal: %v, want it unbound", err) |
| 767 | } |
| 768 | } |
| 769 | |
| 770 | // Task 9: a worker spawned from the live link (onto nil) asks LiveExecGate |
| 771 | // about the build it resolved, before its publication and its exec; a |
| 772 | // refusal publishes and runs nothing. A recycle onto the captured build |
| 773 | // (onto set) is not a move onto the live link, and is not asked. |
| 774 | func TestALiveExecGateGatesAWorkerSpawnFromTheLiveLink(t *testing.T) { |
| 775 | t.Parallel() |
| 776 | b := testExecBackend(t) |
| 777 | host := "bot.example.com" |
| 778 | out := filepath.Join(t.TempDir(), "argv0") |
| 779 | v1 := stageArgv0Worker(t, b, host, "v1", out) |
| 780 | stageArgv0Worker(t, b, host, "v2", out) |
| 781 | g := &liveGate{refuse: map[string]bool{"v2": true}} |
| 782 | b.LiveExecGate = g.gate |
| 783 | published := 0 |
| 784 | ctx := withBeforeExec(context.Background(), func(release.StagedIdentity) error { published++; return nil }) |
| 785 | p, _, err := b.StartWorkerStaged(ctx, workerRoute(host), nil, nil) |
| 786 | if p != nil { |
| 787 | _ = p.Stop(context.Background()) |
| 788 | } |
| 789 | if !errors.Is(err, errLiveHeld) || p != nil || published != 0 { |
| 790 | t.Fatalf("StartWorkerStaged = %v, %v (published %d); want the gate's refusal, nothing published or run", p, err, published) |
| 791 | } |
| 792 | time.Sleep(200 * time.Millisecond) |
| 793 | if _, err := os.Stat(out); err == nil { |
| 794 | t.Fatal("held v2 ran as the worker") |
| 795 | } |
| 796 | sum, err := release.FileSHA256(v1) |
| 797 | if err != nil { |
| 798 | t.Fatal(err) |
| 799 | } |
| 800 | onto := release.StagedIdentity{Version: "v1", StagedPath: v1, Digest: sum} |
| 801 | p, id, err := b.StartWorkerStaged(ctx, workerRoute(host), nil, &onto) |
| 802 | if err != nil { |
| 803 | t.Fatalf("recycle onto the captured v1 = %v", err) |
| 804 | } |
| 805 | defer p.Stop(context.Background()) |
| 806 | if id != onto || strings.TrimSpace(waitForFile(t, out, 5*time.Second)) != v1 { |
| 807 | t.Fatalf("recycled %+v, want the captured v1", id) |
| 808 | } |
| 809 | if !slices.Equal(g.builds(), []string{"v2"}) { |
| 810 | t.Fatalf("gate asked %v, want only the live-link spawn's [v2]", g.builds()) |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | // Task 9 (brief (b)): a declined worker whose live link moves to a held |
| 815 | // build is not stopped by the sweep's converge. The activator asks the gate |
| 816 | // before it retires the old child, so a refusal leaves the worker running |
| 817 | // its current build instead of stopping it into a refused spawn. |
| 818 | func TestADeclinedWorkerIsNotConvergedOntoAHeldBuild(t *testing.T) { |
| 819 | t.Parallel() |
| 820 | a, b, clk, _, ver := stagedActivator(t) |
| 821 | g := &liveGate{refuse: map[string]bool{"v2": true}} |
| 822 | a.SetLiveExecGate(g.gate) |
| 823 | host := "bot.x.internal" |
| 824 | if err := a.WakeWorker(alwaysOnWorkerRoute(host)); err != nil { |
| 825 | t.Fatalf("WakeWorker: %v", err) |
| 826 | } |
| 827 | ver.Store("v2") |
| 828 | a.sweepOnce() |
| 829 | clk.advance(ConvergeGrace + time.Minute) |
| 830 | a.sweepOnce() |
| 831 | a.sweepOnce() |
| 832 | time.Sleep(100 * time.Millisecond) |
| 833 | if calls := b.calls(); len(calls) != 1 || a.entryRespawning(host) { |
| 834 | t.Fatalf("spawns %v (respawning %v): the worker was converged onto held v2", calls, a.entryRespawning(host)) |
| 835 | } |
| 836 | if b.lastWorker().Exited() { |
| 837 | t.Fatal("the converge stopped the worker serving v1") |
| 838 | } |
| 839 | if got := a.entrySpawnVersion(host); got != "v1" { |
| 840 | t.Fatalf("spawn version %q, want v1", got) |
| 841 | } |
| 842 | if !slices.Contains(g.builds(), "v2") { |
| 843 | t.Fatalf("gate asked %v, want v2 asked", g.builds()) |
| 844 | } |
| 845 | // Once the link names a build the gate allows, the converge runs as |
| 846 | // before. |
| 847 | ver.Store("v3") |
| 848 | a.sweepOnce() |
| 849 | clk.advance(ConvergeGrace + time.Minute) |
| 850 | a.sweepOnce() |
| 851 | waitFor(t, "the converge onto v3", func() bool { return a.entrySpawnVersion(host) == "v3" && !a.entryRespawning(host) }) |
| 852 | } |
| 853 | |
| 854 | // Task 9 (brief (c)): a declined worker that crashes while its live link |
| 855 | // names a held build respawns its captured build, however long the drift |
| 856 | // lasts: the recycle's converge is refused before anything is stopped. |
| 857 | func TestADeclinedWorkerCrashRecyclesItsCapturedBuildWhenTheLiveLinkIsHeld(t *testing.T) { |
| 858 | t.Parallel() |
| 859 | a, b, clk, _, ver := stagedActivator(t) |
| 860 | a.SetLiveExecGate((&liveGate{refuse: map[string]bool{"v2": true}}).gate) |
| 861 | host := "bot.x.internal" |
| 862 | if err := a.WakeWorker(alwaysOnWorkerRoute(host)); err != nil { |
| 863 | t.Fatalf("WakeWorker: %v", err) |
| 864 | } |
| 865 | ver.Store("v2") |
| 866 | for i := 1; i <= 5; i++ { |
| 867 | b.lastWorker().kill() |
| 868 | clk.advance(sidecarRestartMax + time.Second) |
| 869 | a.sweepOnce() |
| 870 | waitForRestart(t, a, b.fakeBackend, host, i+1) |
| 871 | } |
| 872 | for i, c := range b.calls()[1:] { |
| 873 | if c == nil || *c != fakeIdentity("v1") { |
| 874 | t.Fatalf("respawn %d asked for %v; want the captured v1 every time", i+1, c) |
| 875 | } |
| 876 | } |
| 877 | if got := a.entrySpawnVersion(host); got != "v1" { |
| 878 | t.Fatalf("spawn version %q, want v1", got) |
| 879 | } |
| 880 | } |
| 881 | |
| 882 | // Task 9 (brief (c)): a worker with no captured build (a backend without |
| 883 | // the staged capability) whose live link names a held build is not |
| 884 | // respawned from it: a crashed one stays down until a wake starts it, and |
| 885 | // one whose max_run falls due keeps running. |
| 886 | func TestADeclinedUncapturedWorkerIsNotRespawnedOntoAHeldBuild(t *testing.T) { |
| 887 | t.Parallel() |
| 888 | t.Run("crashed", func(t *testing.T) { |
| 889 | a, b, clk, _ := testActivator(t, time.Hour) |
| 890 | a.liveVer = func(string) (string, error) { return "v2", nil } |
| 891 | a.SetLiveExecGate((&liveGate{refuse: map[string]bool{"v2": true}}).gate) |
| 892 | host := "bot.x.internal" |
| 893 | if err := a.WakeWorker(alwaysOnWorkerRoute(host)); err != nil { |
| 894 | t.Fatalf("WakeWorker: %v", err) |
| 895 | } |
| 896 | b.lastWorker().kill() |
| 897 | for range 3 { |
| 898 | a.sweepOnce() |
| 899 | clk.advance(sidecarRestartMax + time.Second) |
| 900 | } |
| 901 | time.Sleep(100 * time.Millisecond) |
| 902 | if n := b.workerStarts(); n != 1 || a.entryRespawning(host) { |
| 903 | t.Fatalf("workerStarts = %d (respawning %v): respawned onto held v2", n, a.entryRespawning(host)) |
| 904 | } |
| 905 | if !a.Owns(host) { |
| 906 | t.Fatal("the entry went: it stays, down, for a wake to start") |
| 907 | } |
| 908 | }) |
| 909 | t.Run("max_run", func(t *testing.T) { |
| 910 | a, b, clk, _ := testActivator(t, time.Hour) |
| 911 | a.liveVer = func(string) (string, error) { return "v2", nil } |
| 912 | a.SetLiveExecGate((&liveGate{refuse: map[string]bool{"v2": true}}).gate) |
| 913 | host := "bot.x.internal" |
| 914 | rt := alwaysOnWorkerRoute(host) |
| 915 | rt.MaxRunSeconds = 30 |
| 916 | if err := a.WakeWorker(rt); err != nil { |
| 917 | t.Fatalf("WakeWorker: %v", err) |
| 918 | } |
| 919 | for range 3 { |
| 920 | clk.advance(31 * time.Second) |
| 921 | a.sweepOnce() |
| 922 | } |
| 923 | time.Sleep(100 * time.Millisecond) |
| 924 | if n := b.workerStarts(); n != 1 || a.entryRespawning(host) { |
| 925 | t.Fatalf("workerStarts = %d (respawning %v): recycled onto held v2", n, a.entryRespawning(host)) |
| 926 | } |
| 927 | if b.lastWorker().Exited() { |
| 928 | t.Fatal("the running child was stopped by a recycle the gate refused") |
| 929 | } |
| 930 | }) |
| 931 | } |
| 932 | |
| 933 | // Task 9 fix round 1 (ruling 1): the legacy co-spawned sidecar execs the |
| 934 | // live link too, and its wake tail, late attachment and the sweep's |
| 935 | // supervision respawn it while the instance keeps running. On a declined box |
| 936 | // with the gate set, root's repoint of the link at a held build must not |
| 937 | // start that build as the sidecar: the start resolves the link (no hash), |
| 938 | // asks the gate about that build, and execs the path it resolved. A refusal |
| 939 | // starts nothing (the supervisor's backoff keeps it down). With the gate nil |
| 940 | // the sidecar execs the link, as before. |
| 941 | func TestALiveExecGateGatesTheLegacySidecarStart(t *testing.T) { |
| 942 | t.Parallel() |
| 943 | sidecarEnv := func(t *testing.T, b *ExecBackend, rt registry.Route) { |
| 944 | t.Helper() |
| 945 | if err := os.WriteFile(b.pathsFor(rt).sidecarEnv, nil, 0o600); err != nil { |
| 946 | t.Fatal(err) |
| 947 | } |
| 948 | } |
| 949 | t.Run("flat, gated build is the exec'd build", func(t *testing.T) { |
| 950 | b, rt, v1, _, out := flatLiveFixture(t) |
| 951 | sidecarEnv(t, b, rt) |
| 952 | g := &liveGate{refuse: map[string]bool{"v2": true}} |
| 953 | b.LiveExecGate = g.gate |
| 954 | p, err := b.StartSidecar(context.Background(), rt) |
| 955 | if err != nil { |
| 956 | t.Fatalf("StartSidecar = %v", err) |
| 957 | } |
| 958 | defer p.Stop(context.Background()) |
| 959 | if got := strings.TrimSpace(waitForFile(t, out, 5*time.Second)); got != v1 { |
| 960 | t.Fatalf("sidecar $0 = %q, want the staged v1 the gate was asked about (%s)", got, v1) |
| 961 | } |
| 962 | if !slices.Equal(g.builds(), []string{"v1"}) { |
| 963 | t.Fatalf("gate asked %v, want [v1]", g.builds()) |
| 964 | } |
| 965 | }) |
| 966 | t.Run("flat, a held build starts nothing", func(t *testing.T) { |
| 967 | b, rt, _, v2, out := flatLiveFixture(t) |
| 968 | sidecarEnv(t, b, rt) |
| 969 | repointLive(t, b, rt.Host, v2) |
| 970 | b.LiveExecGate = (&liveGate{refuse: map[string]bool{"v2": true}}).gate |
| 971 | p, err := b.StartSidecar(context.Background(), rt) |
| 972 | if p != nil { |
| 973 | _ = p.Stop(context.Background()) |
| 974 | } |
| 975 | if !errors.Is(err, errLiveHeld) || p != nil { |
| 976 | t.Fatalf("StartSidecar = %v, %v; want no process and the gate's refusal", p, err) |
| 977 | } |
| 978 | time.Sleep(200 * time.Millisecond) |
| 979 | if _, err := os.Stat(out); err == nil { |
| 980 | t.Fatal("held v2 ran as the sidecar") |
| 981 | } |
| 982 | }) |
| 983 | t.Run("flat, no gate is today's", func(t *testing.T) { |
| 984 | b, rt, _, _, out := flatLiveFixture(t) |
| 985 | sidecarEnv(t, b, rt) |
| 986 | p, err := b.StartSidecar(context.Background(), rt) |
| 987 | if err != nil { |
| 988 | t.Fatalf("StartSidecar = %v", err) |
| 989 | } |
| 990 | defer p.Stop(context.Background()) |
| 991 | if got := strings.TrimSpace(waitForFile(t, out, 5*time.Second)); got != b.liveLink(rt.Host) { |
| 992 | t.Fatalf("sidecar $0 = %q, want today's exec of %s", got, b.liveLink(rt.Host)) |
| 993 | } |
| 994 | }) |
| 995 | t.Run("nested, a held build asks root for nothing", func(t *testing.T) { |
| 996 | ctl := &fakeControl{inv: "inv1"} |
| 997 | b, rt := nestedBackend(t, ctl, &fakeUnit{}) |
| 998 | g := &liveGate{refuse: map[string]bool{"v1": true}} |
| 999 | b.LiveExecGate = g.gate |
| 1000 | p, err := b.StartSidecar(context.Background(), rt) |
| 1001 | if !errors.Is(err, errLiveHeld) || p != nil { |
| 1002 | t.Fatalf("StartSidecar = %v, %v; want the gate's refusal", p, err) |
| 1003 | } |
| 1004 | if strings.Contains(ctl.joined(), "start ") { |
| 1005 | t.Fatalf("root was asked to start the sidecar after the gate refused: %q", ctl.joined()) |
| 1006 | } |
| 1007 | if !slices.Equal(g.builds(), []string{"v1"}) { |
| 1008 | t.Fatalf("gate asked %v, want [v1]", g.builds()) |
| 1009 | } |
| 1010 | }) |
| 1011 | } |
| 1012 | |
| 1013 | // Task 9 fix round 1 (ruling 2): the sweep's drift cycle of an awake |
| 1014 | // instance, or a hibernating worker, stops it so its next wake runs the new |
| 1015 | // build. On a declined box a held build's next wake is refused, so the gate |
| 1016 | // is asked first: a refusal keeps it running and spends no budget. A build |
| 1017 | // the gate allows cycles as before. |
| 1018 | func TestADeclinedSweepDoesNotCycleOntoAHeldBuild(t *testing.T) { |
| 1019 | t.Parallel() |
| 1020 | t.Run("instance", func(t *testing.T) { |
| 1021 | a, b, clk, _ := testActivator(t, time.Hour) |
| 1022 | var ver atomic.Value |
| 1023 | ver.Store("v1") |
| 1024 | a.liveVer = func(string) (string, error) { return ver.Load().(string), nil } |
| 1025 | b.execVerFn = func() string { return ver.Load().(string) } |
| 1026 | a.SetLiveExecGate((&liveGate{refuse: map[string]bool{"v2": true}}).gate) |
| 1027 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 1028 | if err != nil { |
| 1029 | t.Fatalf("beginRequest: %v", err) |
| 1030 | } |
| 1031 | release() |
| 1032 | ver.Store("v2") |
| 1033 | a.sweepOnce() |
| 1034 | clk.advance(ConvergeGrace + time.Second) |
| 1035 | a.sweepOnce() |
| 1036 | a.sweepOnce() |
| 1037 | time.Sleep(50 * time.Millisecond) |
| 1038 | if a.entryCount() != 1 || b.lastProc().isStopped() { |
| 1039 | t.Fatal("the sweep stopped an instance serving v1 into a wake onto held v2") |
| 1040 | } |
| 1041 | ver.Store("v3") |
| 1042 | a.sweepOnce() |
| 1043 | clk.advance(ConvergeGrace + time.Second) |
| 1044 | a.sweepOnce() |
| 1045 | waitFor(t, "the cycle onto v3", func() bool { return a.entryCount() == 0 }) |
| 1046 | }) |
| 1047 | t.Run("hibernating worker", func(t *testing.T) { |
| 1048 | a, b, clk, _ := testActivator(t, time.Hour) |
| 1049 | var ver atomic.Value |
| 1050 | ver.Store("v1") |
| 1051 | a.liveVer = func(string) (string, error) { return ver.Load().(string), nil } |
| 1052 | b.execVerFn = func() string { return ver.Load().(string) } |
| 1053 | a.SetLiveExecGate((&liveGate{refuse: map[string]bool{"v2": true}}).gate) |
| 1054 | host := "bot.x.internal" |
| 1055 | if err := a.WakeWorker(testWorkerRoute(host)); err != nil { |
| 1056 | t.Fatalf("WakeWorker: %v", err) |
| 1057 | } |
| 1058 | ver.Store("v2") |
| 1059 | a.sweepOnce() |
| 1060 | clk.advance(ConvergeGrace + time.Second) |
| 1061 | a.sweepOnce() |
| 1062 | a.sweepOnce() |
| 1063 | time.Sleep(50 * time.Millisecond) |
| 1064 | if !a.Owns(host) || b.lastWorker().Exited() { |
| 1065 | t.Fatal("the sweep stopped a worker serving v1 into a wake onto held v2") |
| 1066 | } |
| 1067 | }) |
| 1068 | } |
| 1069 | |
amadan