| 1 | package host |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "log" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "slices" |
| 10 | "strings" |
| 11 | "sync" |
| 12 | "time" |
| 13 | |
| 14 | "github.com/carlosframework/platform/internal/hostactors" |
| 15 | "github.com/carlosframework/platform/internal/hostbox" |
| 16 | "github.com/carlosframework/platform/internal/hostclaim" |
| 17 | execlease "github.com/carlosframework/platform/internal/hostexec" |
| 18 | "github.com/carlosframework/platform/internal/hostguard" |
| 19 | "github.com/carlosframework/platform/internal/hostwrite" |
| 20 | "github.com/carlosframework/platform/internal/release" |
| 21 | ) |
| 22 | |
| 23 | type claimParticipation struct { |
| 24 | Binding execlease.Binding |
| 25 | Controller claimExecutionController |
| 26 | Ready chan struct{} |
| 27 | Err error |
| 28 | Config execlease.Config |
| 29 | Captured hostclaim.Observed |
| 30 | CleanupSucceeded bool |
| 31 | // AckController is the controller ID this slot's controller recorded in |
| 32 | // its first execution acknowledgement. Only that acknowledgement is ever |
| 33 | // refreshed: a restarted agent opens a new slot and leaves an older |
| 34 | // controller's acknowledgement alone. Guarded by productionClaimLocal.mu. |
| 35 | AckController string |
| 36 | } |
| 37 | type claimExecutionController interface { |
| 38 | Ready(context.Context) (execlease.State, error) |
| 39 | Close(context.Context) error |
| 40 | } |
| 41 | |
| 42 | type productionClaimLocal struct { |
| 43 | executionCall func(context.Context, execlease.ExecutionRequest) (execlease.ExecutionResponse, error) |
| 44 | serving map[string]execlease.ExecutionRequest |
| 45 | legacyParticipation bool // Only legacy-controller fixtures select this path. |
| 46 | |
| 47 | executionRead func(context.Context, *hostclaim.Claim) (execlease.ExecutionRecord, error) |
| 48 | Resources ClaimRuntimeResources |
| 49 | Identity hostbox.IdentityProvider |
| 50 | Directory hostactors.Directory |
| 51 | Store release.Store |
| 52 | mu sync.Mutex |
| 53 | slots map[string]*claimParticipation |
| 54 | assembly *ClaimAssembly |
| 55 | lifetime context.Context |
| 56 | closed bool |
| 57 | loopCancel context.CancelFunc |
| 58 | loopDone chan struct{} |
| 59 | stopIdentity *claimStopLocal |
| 60 | // lockRoot is a test seam (region HA N12). Set, the participation's |
| 61 | // hostguard and hostexec locks live under it and its controllers skip |
| 62 | // the update-floor check (their UpdateGuardDir is empty; main's floor |
| 63 | // tests cover it). Empty is production: /run/carlos's locks and the |
| 64 | // floor under OptDir. |
| 65 | lockRoot string |
| 66 | // now is a test seam for the acknowledgement refresh's wall clock; nil |
| 67 | // is time.Now. |
| 68 | now func() time.Time |
| 69 | } |
| 70 | |
| 71 | // lockDir is the named lock directory under lockRoot, or "" (the |
| 72 | // package's default) when lockRoot is unset. |
| 73 | func (p *productionClaimLocal) lockDir(name string) string { |
| 74 | if p.lockRoot == "" { |
| 75 | return "" |
| 76 | } |
| 77 | return filepath.Join(p.lockRoot, name) |
| 78 | } |
| 79 | |
| 80 | // updateFloor is the controller's update-floor guard directory and |
| 81 | // journal: production's under OptDir, none under the N12 test seam. No |
| 82 | // production constructor sets lockRoot (OpenClaimOwners builds the owner |
| 83 | // without it; only tests assign it), so the floor is never bypassed in |
| 84 | // production. |
| 85 | func (p *productionClaimLocal) updateFloor() (string, string) { |
| 86 | if p.lockRoot != "" { |
| 87 | return "", "" |
| 88 | } |
| 89 | return filepath.Join(p.Resources.OptDir, "system", "guard"), filepath.Join(p.Resources.OptDir, "system", "router", "journal.json") |
| 90 | } |
| 91 | |
| 92 | func (p *productionClaimLocal) bind(ctx context.Context, a *ClaimAssembly) error { |
| 93 | local, err := a.localStopIdentity(ctx) |
| 94 | if err != nil { |
| 95 | return err |
| 96 | } |
| 97 | p.mu.Lock() |
| 98 | defer p.mu.Unlock() |
| 99 | p.assembly = a |
| 100 | p.lifetime = ctx |
| 101 | p.stopIdentity = &local |
| 102 | return nil |
| 103 | } |
| 104 | |
| 105 | // Prepare performs lease I/O before WorkSource takes hostguard. Serving never |
| 106 | // opens a lease. A failed controller is retained as denial until cleanup and |
| 107 | // CC9 positively discharge it, never replaced by age. |
| 108 | func (p *productionClaimLocal) Prepare(ctx context.Context, o hostclaim.Observed) error { |
| 109 | if o.Claim == nil || (!o.Claim.Free() && o.Claim.Holder == nil) { |
| 110 | return hostbox.ErrChanged |
| 111 | } |
| 112 | actor, err := p.Identity.Verify(ctx) |
| 113 | if err != nil { |
| 114 | return err |
| 115 | } |
| 116 | return p.prepareWith(ctx, actor, o) |
| 117 | } |
| 118 | func (p *productionClaimLocal) prepareLegacyWith(ctx context.Context, actor hostbox.Actor, o hostclaim.Observed) error { |
| 119 | var err error |
| 120 | c := o.Claim |
| 121 | p.mu.Lock() |
| 122 | old := p.slots[o.Host] |
| 123 | a := p.assembly |
| 124 | lifetime := p.lifetime |
| 125 | closed := p.closed |
| 126 | p.mu.Unlock() |
| 127 | if closed || a == nil || lifetime == nil { |
| 128 | return hostwrite.ErrUnavailable |
| 129 | } |
| 130 | if p.Resources.Mode != "agent" { |
| 131 | return p.checkRootParticipation(ctx, actor, c) |
| 132 | } |
| 133 | if c.Op.Kind == hostclaim.KindServing || c.Free() || !slices.Contains(c.Op.Participants, actor.Box) { |
| 134 | |
| 135 | if old == nil { |
| 136 | state, err := execlease.InspectLocal(p.Resources.DataDir, o.Host) |
| 137 | if errors.Is(err, os.ErrNotExist) { |
| 138 | return nil |
| 139 | } |
| 140 | if err != nil { |
| 141 | return err |
| 142 | } |
| 143 | return p.clearUntracked(ctx, actor, o, state, execlease.Config{Binding: state.Binding, Guard: hostguard.Guard{DataDir: p.Resources.DataDir, LockDir: p.lockDir("hostguard"), Claims: hostguard.StoreReader{Store: p.Store}}, LockDir: p.lockDir("hostexec")}) |
| 144 | } |
| 145 | select { |
| 146 | case <-old.Ready: |
| 147 | case <-ctx.Done(): |
| 148 | return ctx.Err() |
| 149 | } |
| 150 | if old.Controller != nil { |
| 151 | // Close joins before it reports cleanup or conditional-release |
| 152 | // errors. A timed-out join retains the slot and lifetime lock. |
| 153 | closeErr := old.Controller.Close(ctx) |
| 154 | if ctx.Err() != nil { |
| 155 | return errors.Join(closeErr, ctx.Err()) |
| 156 | } |
| 157 | } |
| 158 | // A previous failed cleanup remains denial until this fresh positive |
| 159 | // proof succeeds. Its historical error must not veto CC9 forever. |
| 160 | if !old.CleanupSucceeded { |
| 161 | if err = p.cleanup(ctx, old.Captured, old.Config.Binding); err != nil { |
| 162 | return err |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | if err = execlease.ClearStale(ctx, old.Config); err != nil { |
| 167 | return err |
| 168 | } |
| 169 | p.mu.Lock() |
| 170 | if p.slots[o.Host] == old { |
| 171 | delete(p.slots, o.Host) |
| 172 | } |
| 173 | p.mu.Unlock() |
| 174 | return nil |
| 175 | } |
| 176 | if old != nil { |
| 177 | select { |
| 178 | case <-old.Ready: |
| 179 | case <-ctx.Done(): |
| 180 | return ctx.Err() |
| 181 | } |
| 182 | if old.Err != nil { |
| 183 | return old.Err |
| 184 | } |
| 185 | if old.Binding.Op != c.Op.ID || old.Binding.Incarnation != c.Holder.Incarnation { |
| 186 | return execlease.ErrLost |
| 187 | } |
| 188 | _, err = old.Controller.Ready(ctx) |
| 189 | return err |
| 190 | } |
| 191 | boot, err := os.ReadFile("/proc/sys/kernel/random/boot_id") |
| 192 | if err != nil { |
| 193 | return err |
| 194 | } |
| 195 | slot := &claimParticipation{Ready: make(chan struct{}), Captured: o} |
| 196 | p.mu.Lock() |
| 197 | if p.closed || p.slots[o.Host] != nil { |
| 198 | p.mu.Unlock() |
| 199 | return execlease.ErrControllerBusy |
| 200 | } |
| 201 | p.slots[o.Host] = slot |
| 202 | p.mu.Unlock() |
| 203 | binding := execlease.Binding{Box: actor.Box, Enrollment: actor.Enrollment, Host: o.Host, Incarnation: c.Holder.Incarnation, Op: c.Op.ID, BootID: strings.TrimSpace(string(boot))} |
| 204 | floor, journal := p.updateFloor() |
| 205 | slot.Config = execlease.Config{Binding: binding, Store: p.Store, Guard: hostguard.Guard{DataDir: p.Resources.DataDir, LockDir: p.lockDir("hostguard"), Claims: hostguard.StoreReader{Store: p.Store}}, Directory: p.Directory, Clock: execlease.SystemClock{}, UpdateGuardDir: floor, UpdateJournal: journal, LockDir: p.lockDir("hostexec"), |
| 206 | CloseAdmission: func() { p.cancelWork(o.Host) }, CancelWork: func() { p.cancelWork(o.Host) }, Cleanup: func(ctx context.Context) error { |
| 207 | err := p.cleanup(ctx, o, binding) |
| 208 | slot.CleanupSucceeded = err == nil |
| 209 | return err |
| 210 | }} |
| 211 | slot.Controller, slot.Err = execlease.Open(lifetime, slot.Config) |
| 212 | if slot.Err == nil { |
| 213 | var state execlease.State |
| 214 | state, slot.Err = slot.Controller.Ready(ctx) |
| 215 | slot.Binding = state.Binding |
| 216 | } |
| 217 | close(slot.Ready) |
| 218 | return slot.Err |
| 219 | } |
| 220 | func (p *productionClaimLocal) cancelWork(host string) { |
| 221 | p.mu.Lock() |
| 222 | a := p.assembly |
| 223 | p.mu.Unlock() |
| 224 | if a != nil && a.Wake != nil { |
| 225 | if pending := a.Wake.get(host); pending != nil { |
| 226 | pending.cancel() |
| 227 | } |
| 228 | } |
| 229 | } |
| 230 | func (p *productionClaimLocal) cleanup(ctx context.Context, o hostclaim.Observed, binding execlease.Binding) error { |
| 231 | p.mu.Lock() |
| 232 | a := p.assembly |
| 233 | local := p.stopIdentity |
| 234 | p.mu.Unlock() |
| 235 | if a != nil && local != nil { |
| 236 | copy := *a |
| 237 | saved := *local |
| 238 | copy.stopLocal = func(context.Context) (claimStopLocal, error) { return saved, nil } |
| 239 | a = © |
| 240 | } |
| 241 | if a == nil || a.Wake == nil || a.Launches == nil || a.Reconcile == nil || a.Reconcile.Census == nil { |
| 242 | return hostwrite.ErrUnavailable |
| 243 | } |
| 244 | p.cancelWork(o.Host) |
| 245 | if !p.legacyParticipation { |
| 246 | if done, err := a.completedRetirementDischarge(ctx, o, binding); done || err != nil { |
| 247 | return err |
| 248 | } |
| 249 | } |
| 250 | // Claim ask C10 (region HA rulings 56, 66): a relocation that completed |
| 251 | // on this box is discharged by the persisted participation binding and |
| 252 | // its exact completed probe evidence, before any ordinary cleanup, so |
| 253 | // no serving-shaped stop ever replaces the probe's fence. The captured |
| 254 | // observation is not consulted: after a restart it is the current one. |
| 255 | if done, err := a.completedRelocationDischarge(ctx, binding); done || err != nil { |
| 256 | return err |
| 257 | } |
| 258 | // Cleanup uses the journal and producer set directly. It must never call |
| 259 | // WorkSource.Select: Open is still waiting for this cleanup to finish. |
| 260 | has, err := a.Launches.HasHost(ctx, o.Host) |
| 261 | if err != nil { |
| 262 | return err |
| 263 | } |
| 264 | if has { |
| 265 | if o.Claim == nil || o.Claim.Holder == nil || o.Claim.Holder.Session == 0 { |
| 266 | return ErrClaimStopInventory |
| 267 | } |
| 268 | return a.stopCapturedSession(ctx, o) |
| 269 | } |
| 270 | if o.Claim == nil || o.Claim.Holder == nil { |
| 271 | return ErrClaimStopInventory |
| 272 | } |
| 273 | if !p.legacyParticipation && a.Launches.Remote() && a.Reconcile.Inventory.Remote() { |
| 274 | return ctx.Err() |
| 275 | } |
| 276 | return a.Reconcile.Census.CheckLocal(ctx, o, hostbox.RecordSelection{Host: o.Host, Incarnation: o.Claim.Holder.Incarnation}) |
| 277 | } |
| 278 | func (p *productionClaimLocal) checkLegacy(ctx context.Context, a hostbox.Actor, c *hostclaim.Claim, s hostbox.Selection) error { |
| 279 | if ctx.Err() != nil { |
| 280 | return ctx.Err() |
| 281 | } |
| 282 | if c == nil || c.Holder == nil || a.Box != s.Box || a.Region != s.Region || c.Holder.Incarnation != s.Captured.Incarnation || slices.Contains(c.Lost, a.Box) { |
| 283 | return hostbox.ErrChanged |
| 284 | } |
| 285 | p.mu.Lock() |
| 286 | slot := p.slots[c.Host] |
| 287 | closed := p.closed |
| 288 | p.mu.Unlock() |
| 289 | if closed { |
| 290 | return execlease.ErrLost |
| 291 | } |
| 292 | if p.Resources.Mode != "agent" { |
| 293 | return p.checkRootParticipation(ctx, a, c) |
| 294 | } |
| 295 | if c.Op.Kind == hostclaim.KindServing { |
| 296 | if slot != nil { |
| 297 | return execlease.ErrLost |
| 298 | } |
| 299 | _, err := execlease.InspectLocal(p.Resources.DataDir, c.Host) |
| 300 | if !errors.Is(err, os.ErrNotExist) { |
| 301 | return errors.Join(execlease.ErrLost, err) |
| 302 | } |
| 303 | return nil |
| 304 | } |
| 305 | if slot == nil { |
| 306 | state, err := execlease.InspectLocal(p.Resources.DataDir, c.Host) |
| 307 | if err != nil { |
| 308 | return err |
| 309 | } |
| 310 | if state.Op != c.Op.ID || state.Incarnation != c.Holder.Incarnation || state.Box != a.Box || state.Enrollment != a.Enrollment { |
| 311 | return execlease.ErrState |
| 312 | } |
| 313 | return execlease.CheckLocal(p.Resources.DataDir, state.Binding, execlease.SystemClock{}, "") |
| 314 | } |
| 315 | select { |
| 316 | case <-slot.Ready: |
| 317 | default: |
| 318 | return execlease.ErrInactive |
| 319 | } |
| 320 | if slot.Err != nil { |
| 321 | return slot.Err |
| 322 | } |
| 323 | if slot.Binding.Op != c.Op.ID || slot.Binding.Incarnation != c.Holder.Incarnation || slot.Binding.Box != a.Box || slot.Binding.Enrollment != a.Enrollment { |
| 324 | return execlease.ErrLost |
| 325 | } |
| 326 | return execlease.CheckLocal(p.Resources.DataDir, slot.Binding, execlease.SystemClock{}, "") |
| 327 | } |
| 328 | func (p *productionClaimLocal) Close(ctx context.Context) error { |
| 329 | p.mu.Lock() |
| 330 | p.closed = true |
| 331 | cancel, done := p.loopCancel, p.loopDone |
| 332 | slots := make([]*claimParticipation, 0, len(p.slots)) |
| 333 | for _, s := range p.slots { |
| 334 | slots = append(slots, s) |
| 335 | } |
| 336 | p.mu.Unlock() |
| 337 | if cancel != nil { |
| 338 | cancel() |
| 339 | select { |
| 340 | case <-done: |
| 341 | case <-ctx.Done(): |
| 342 | return ctx.Err() |
| 343 | } |
| 344 | } |
| 345 | var errs []error |
| 346 | for _, slot := range slots { |
| 347 | select { |
| 348 | case <-slot.Ready: |
| 349 | case <-ctx.Done(): |
| 350 | return errors.Join(errors.Join(errs...), ctx.Err()) |
| 351 | } |
| 352 | if slot.Controller != nil { |
| 353 | errs = append(errs, slot.Controller.Close(ctx)) |
| 354 | } |
| 355 | errs = append(errs, slot.Err) |
| 356 | } |
| 357 | return errors.Join(errs...) |
| 358 | } |
| 359 | |
| 360 | // start observes participants even when they have no pending local writes. |
| 361 | // Root and adopt only inspect the agent's controller and never own a lease. |
| 362 | func (p *productionClaimLocal) start(ctx context.Context) { |
| 363 | if p.Resources.Mode != "agent" { |
| 364 | return |
| 365 | } |
| 366 | worker, cancel := context.WithCancel(ctx) |
| 367 | p.mu.Lock() |
| 368 | if p.closed || p.loopCancel != nil { |
| 369 | p.mu.Unlock() |
| 370 | cancel() |
| 371 | return |
| 372 | } |
| 373 | p.loopCancel, p.loopDone = cancel, make(chan struct{}) |
| 374 | done := p.loopDone |
| 375 | p.mu.Unlock() |
| 376 | go func() { |
| 377 | defer close(done) |
| 378 | ticker := time.NewTicker(30 * time.Second) |
| 379 | defer ticker.Stop() |
| 380 | for { |
| 381 | err := p.participate(worker) |
| 382 | if err != nil && worker.Err() == nil { |
| 383 | log.Printf("host claim participation: %v", err) |
| 384 | } |
| 385 | select { |
| 386 | case <-worker.Done(): |
| 387 | return |
| 388 | case <-ticker.C: |
| 389 | } |
| 390 | } |
| 391 | }() |
| 392 | } |
| 393 | func (p *productionClaimLocal) participate(ctx context.Context) error { |
| 394 | identity, finishIdentity := context.WithTimeout(ctx, 30*time.Second) |
| 395 | actor, err := p.Identity.Verify(identity) |
| 396 | finishIdentity() |
| 397 | if err != nil { |
| 398 | return err |
| 399 | } |
| 400 | discovery, finishDiscovery := context.WithTimeout(ctx, 30*time.Second) |
| 401 | keys, err := p.Store.List(discovery, "control/hosts/") |
| 402 | finishDiscovery() |
| 403 | if err != nil { |
| 404 | return err |
| 405 | } |
| 406 | var errs []error |
| 407 | for _, key := range keys { |
| 408 | host := strings.TrimSuffix(strings.TrimPrefix(key, "control/hosts/"), ".json") |
| 409 | canonical, err := hostclaim.Key(host) |
| 410 | if err != nil || canonical != key { |
| 411 | continue |
| 412 | } |
| 413 | // Each host gets a fresh budget from the worker, including its claim read. |
| 414 | // One slow earlier host cannot consume every later host's next pass. |
| 415 | bounded, cancel := context.WithTimeout(ctx, hostbox.RequestLimit) |
| 416 | err = p.participateHost(bounded, actor, host) |
| 417 | cancel() |
| 418 | if err != nil { |
| 419 | errs = append(errs, err) |
| 420 | } |
| 421 | if ctx.Err() != nil { |
| 422 | return errors.Join(errors.Join(errs...), ctx.Err()) |
| 423 | } |
| 424 | } |
| 425 | return errors.Join(errs...) |
| 426 | } |
| 427 | func (p *productionClaimLocal) participateHost(ctx context.Context, actor hostbox.Actor, host string) error { |
| 428 | o, err := hostclaim.Read(ctx, p.Store, host) |
| 429 | if err != nil { |
| 430 | return err |
| 431 | } |
| 432 | if o.Absent || o.Claim == nil { |
| 433 | return nil |
| 434 | } |
| 435 | if p.legacyParticipation { |
| 436 | p.mu.Lock() |
| 437 | tracked := p.slots[host] != nil |
| 438 | p.mu.Unlock() |
| 439 | if !tracked && (o.Claim.Op.Kind == hostclaim.KindServing || !slices.Contains(o.Claim.Op.Participants, actor.Box)) { |
| 440 | if _, err := execlease.InspectLocal(p.Resources.DataDir, host); errors.Is(err, os.ErrNotExist) { |
| 441 | return nil |
| 442 | } |
| 443 | } |
| 444 | } |
| 445 | if err := p.prepareWith(ctx, actor, o); err != nil { |
| 446 | return err |
| 447 | } |
| 448 | if p.legacyParticipation { |
| 449 | return nil |
| 450 | } |
| 451 | return p.acknowledgeOperation(ctx, actor, o) |
| 452 | } |
| 453 | |
amadan