| 1 | package host |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "log" |
| 9 | "os" |
| 10 | "path" |
| 11 | "path/filepath" |
| 12 | "sort" |
| 13 | "strings" |
| 14 | "sync" |
| 15 | "sync/atomic" |
| 16 | "syscall" |
| 17 | "time" |
| 18 | |
| 19 | "github.com/carlosframework/platform/internal/activator" |
| 20 | "github.com/carlosframework/platform/internal/hostwrite" |
| 21 | "github.com/carlosframework/platform/internal/instancectl" |
| 22 | "github.com/carlosframework/platform/internal/layout" |
| 23 | "github.com/carlosframework/platform/internal/lease" |
| 24 | "github.com/carlosframework/platform/internal/moves" |
| 25 | "github.com/carlosframework/platform/internal/park" |
| 26 | "github.com/carlosframework/platform/internal/provision" |
| 27 | "github.com/carlosframework/platform/internal/registry" |
| 28 | "github.com/carlosframework/platform/internal/release" |
| 29 | ) |
| 30 | |
| 31 | // The offline operator move on a box (docs/superpowers/plans/2026-10-20- |
| 32 | // offline-operator-move.md, Task 3): the hold reconcile asks, the keep view, |
| 33 | // the capability, the moves pass's dispatch, and the source boxes' steps. |
| 34 | // The sealer's step is movesoffline_seal.go (Task 4) and the target's is |
| 35 | // movesoffline_land.go (Task 5). |
| 36 | // |
| 37 | // The stop has three layers. This file holds the first two on every |
| 38 | // source-region box: reconcile holds the row from the record alone |
| 39 | // (offlineHold), and each pass fences the activator and writes the box's |
| 40 | // source report. The third, the tombstone lease, is the sealer's. |
| 41 | // |
| 42 | // Dormant, with no offline record anywhere, nothing here acts: the hold |
| 43 | // answers false for every record, keep answers false once the startup |
| 44 | // listing is read, the dispatch matches no record, and the edges file |
| 45 | // carries one more capability string. |
| 46 | |
| 47 | // capOfflineMove1 is the edges-file capability a box publishes when it can |
| 48 | // take part in an offline move (offlineCapable). The console's start door |
| 49 | // refuses a region holding a box that does not publish it. |
| 50 | const capOfflineMove1 = "offline-move-1" |
| 51 | |
| 52 | // offlineCapable is the box's own witness that it can take part in an |
| 53 | // offline move: hibernation is configured and neither CARLOS_HOST_CLAIM nor |
| 54 | // CARLOS_MOVES_PHASE1 is set to anything at all. The guard mode is not read |
| 55 | // (platform-1e's ruling of 2026-10-20, the plan's Global Constraints): the |
| 56 | // target's build stages through ordinary adoption, which root's own floor |
| 57 | // check already covers on a guarded box. A box that is not capable publishes |
| 58 | // no capability and runs no offline step. |
| 59 | func offlineCapable(getenv func(string) string, hibernation bool) bool { |
| 60 | if !hibernation || hostwrite.Enabled(getenv) { |
| 61 | return false |
| 62 | } |
| 63 | return strings.TrimSpace(getenv(moves.SwitchEnv)) == "" && !moves.Enabled(getenv) |
| 64 | } |
| 65 | |
| 66 | // hibernationConfigured is the edges file's HibernationConfigured, read the |
| 67 | // same way, so the capability and the field beside it never disagree. |
| 68 | func hibernationConfigured(getenv func(string) string) bool { |
| 69 | return strings.TrimSpace(getenv("CARLOS_HIBERNATE_BUCKET")) != "" |
| 70 | } |
| 71 | |
| 72 | // offlineFenceOwner is the activator fence owner of one offline move. |
| 73 | func offlineFenceOwner(id string) string { return "offline:" + id } |
| 74 | |
| 75 | // offlineHold is reconcile's Hold for offline moves: a pure function of the |
| 76 | // record it is handed and this box's placement, never of a snapshot (Fable's |
| 77 | // B1). A target box that restarts after the flip reconciles before its first |
| 78 | // moves pass, and must mint the landing row held even then. |
| 79 | // |
| 80 | // - A source-region box (placement equals move.from) holds the host's row |
| 81 | // "moving" under the move id from offline-stopping through |
| 82 | // offline-sealed. After the flip the record names the target, the row |
| 83 | // is out of this box's scope and offlineView.keep keeps it, held as it |
| 84 | // is, until offline-finishing. |
| 85 | // - A target-region box (placement equals move.to) holds it "landing" |
| 86 | // while the record names the target and the move is offline-landing or |
| 87 | // offline-landed. offline-open is the target's to release (Task 5). |
| 88 | // |
| 89 | // Placements are compared as movesNamesBox compares them, by |
| 90 | // park.PlacementMatches under the verified home binding, so the home's "" |
| 91 | // and its roster name are one. |
| 92 | func offlineHold(box moveCfg) func(rec provision.Record) (status, op string, ok bool) { |
| 93 | home, here := box.ParkHomeRegion, box.parkRegion() |
| 94 | return func(rec provision.Record) (string, string, bool) { |
| 95 | m := rec.Move |
| 96 | if !m.Offline() || m.ID == "" || here == "" { |
| 97 | return "", "", false |
| 98 | } |
| 99 | switch m.State { |
| 100 | case provision.OfflineStopping, provision.OfflineSealing, provision.OfflineSealed: |
| 101 | if park.PlacementMatches(home, m.From, here) { |
| 102 | return "moving", m.ID, true |
| 103 | } |
| 104 | case provision.OfflineLanding, provision.OfflineLanded: |
| 105 | if rec.Region == m.To && park.PlacementMatches(home, m.To, here) { |
| 106 | return "landing", m.ID, true |
| 107 | } |
| 108 | } |
| 109 | return "", "", false |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | // offlineView is what reconcile's KeepParticipant asks about offline moves: |
| 114 | // the offline moves in flight, as the last listing read them. Only keep |
| 115 | // reads it; the hold never does (offlineHold). |
| 116 | // |
| 117 | // A source box's view has seen the move before the record flips: the seal |
| 118 | // door needs this box's report for the move, and the pass that writes it |
| 119 | // refreshes the view first; an agent restart reads the listing again at |
| 120 | // start. From then until finishing every listing carries the move. |
| 121 | // |
| 122 | // It is refreshed from every moves pass's record list and once at agent |
| 123 | // start, before the reconcile loops (refreshOfflineView). keep fails closed, |
| 124 | // keeping every row, only while no listing has been read (reconcileSerialized- |
| 125 | // With reads one first when the view is unbuilt or its last listing failed, |
| 126 | // so a failed startup listing heals on the next reconcile), and for a host |
| 127 | // whose record does not decode. Once a listing is read, a host with no |
| 128 | // offline move is not kept, so removals happen on the first pass as they |
| 129 | // always have. |
| 130 | type offlineView struct { |
| 131 | home, region string |
| 132 | // written is reconcile's HoldWritten for the offline hooks: the |
| 133 | // stage-requested marker after a hold write (registry.SetHold's |
| 134 | // contract, G10). Nil writes nothing. |
| 135 | written func(host string) |
| 136 | snap atomic.Pointer[offlineSnapshot] |
| 137 | logged sync.Map // host -> time.Time: the last unreadable-record line |
| 138 | } |
| 139 | |
| 140 | type offlineSnapshot struct { |
| 141 | moves map[string]provision.MoveState // host -> its offline move |
| 142 | owners map[string][2]string // host -> the account and app its record lives under |
| 143 | failed bool // the listing failed: keep everything |
| 144 | // unreadable are the hosts whose record read but did not decode, or |
| 145 | // disagreed with its path (provision.ListAllRecordsStrict): their move |
| 146 | // cannot be known, so their rows are kept. |
| 147 | unreadable map[string]bool |
| 148 | } |
| 149 | |
| 150 | // newOfflineView builds the view for the box cfg describes. Unbuilt until |
| 151 | // its first refresh. |
| 152 | func newOfflineView(cfg moveCfg, written func(host string)) *offlineView { |
| 153 | return &offlineView{home: cfg.ParkHomeRegion, region: cfg.parkRegion(), written: written} |
| 154 | } |
| 155 | |
| 156 | // refresh rebuilds the view from a record list that was read, with the |
| 157 | // hosts provision.ListAllRecordsStrict could not decode: those are kept |
| 158 | // (fail closed for keep only; the hold never reads the view) and logged at |
| 159 | // most once a minute each. |
| 160 | func (v *offlineView) refresh(recs []provision.Record, unreadable map[string]bool) { |
| 161 | next := &offlineSnapshot{moves: map[string]provision.MoveState{}, owners: map[string][2]string{}, unreadable: map[string]bool{}} |
| 162 | for _, rec := range recs { |
| 163 | if rec.Move.Offline() { |
| 164 | next.moves[rec.Host] = *rec.Move |
| 165 | next.owners[rec.Host] = [2]string{rec.Account, rec.App} |
| 166 | } |
| 167 | } |
| 168 | now := time.Now() |
| 169 | for host := range unreadable { |
| 170 | next.unreadable[host] = true |
| 171 | if last, ok := v.logged.Load(host); !ok || now.Sub(last.(time.Time)) >= time.Minute { |
| 172 | v.logged.Store(host, now) |
| 173 | log.Printf("moves: offline view: %s: the record does not decode; its row is kept until it does", host) |
| 174 | } |
| 175 | } |
| 176 | v.snap.Store(next) |
| 177 | } |
| 178 | |
| 179 | // fail records a listing that could not be read, before any was: keep |
| 180 | // answers true for every host until a listing is read. |
| 181 | func (v *offlineView) fail() { v.snap.Store(&offlineSnapshot{failed: true}) } |
| 182 | |
| 183 | // built reports whether any listing, read or failed, has been recorded. |
| 184 | func (v *offlineView) built() bool { return v.snap.Load() != nil } |
| 185 | |
| 186 | // failed reports whether the last listing recorded failed: keep answers true |
| 187 | // for every host until one reads. |
| 188 | func (v *offlineView) failed() bool { |
| 189 | snap := v.snap.Load() |
| 190 | return snap != nil && snap.failed |
| 191 | } |
| 192 | |
| 193 | // keep is reconcile's KeepParticipant: the row of a host whose offline move |
| 194 | // names this box's region stays, whatever the record's region says. The |
| 195 | // source keeps it until offline-finishing (when it is removed, as the plan's |
| 196 | // finishing step relies on); the target keeps it until the move is cleared |
| 197 | // (the target's revert deletes it itself). An unknown state keeps. |
| 198 | func (v *offlineView) keep(host string) bool { |
| 199 | snap := v.snap.Load() |
| 200 | if snap == nil || snap.failed || snap.unreadable[host] { |
| 201 | return true |
| 202 | } |
| 203 | m, ok := snap.moves[host] |
| 204 | if !ok || v.region == "" { |
| 205 | return false |
| 206 | } |
| 207 | if park.PlacementMatches(v.home, m.To, v.region) { |
| 208 | return true |
| 209 | } |
| 210 | if park.PlacementMatches(v.home, m.From, v.region) { |
| 211 | return m.State != provision.OfflineFinishing |
| 212 | } |
| 213 | return false |
| 214 | } |
| 215 | |
| 216 | // elsewhere reports whether host's offline move has its record name a region |
| 217 | // other than this box's: the target's from landing to finishing, the source's |
| 218 | // while reverting. Such a host is not this box's to touch whatever its row |
| 219 | // says (scheduleSkip). A view with no listing yet, and a host with no |
| 220 | // offline move, answer false: the row's own hold answers for those. |
| 221 | func (v *offlineView) elsewhere(host string) bool { |
| 222 | snap := v.snap.Load() |
| 223 | if snap == nil || snap.failed || v.region == "" { |
| 224 | return false |
| 225 | } |
| 226 | m, ok := snap.moves[host] |
| 227 | if !ok { |
| 228 | return false |
| 229 | } |
| 230 | region := m.From |
| 231 | switch m.State { |
| 232 | case provision.OfflineLanding, provision.OfflineLanded, provision.OfflineOpen, provision.OfflineFinishing: |
| 233 | region = m.To |
| 234 | } |
| 235 | return !park.PlacementMatches(v.home, region, v.region) |
| 236 | } |
| 237 | |
| 238 | // refreshOfflineView reads the record list once, outside a moves pass: at |
| 239 | // agent start, before the reconcile loops, next to refreshMovesHolds. A |
| 240 | // failed listing leaves keep failing closed until a listing is read. |
| 241 | func refreshOfflineView(ctx context.Context, s release.Store, v *offlineView) { |
| 242 | if v == nil || s == nil { |
| 243 | return |
| 244 | } |
| 245 | recs, unreadable, err := provision.ListAllRecordsStrict(ctx, s) |
| 246 | if err != nil { |
| 247 | log.Printf("moves: offline view: list records: %v; every row is kept until a listing reads", err) |
| 248 | v.fail() |
| 249 | return |
| 250 | } |
| 251 | v.refresh(recs, unreadable) |
| 252 | } |
| 253 | |
| 254 | // armedOfflineView is the armed worker's view, nil when no worker is armed |
| 255 | // or the box cannot take part in an offline move. |
| 256 | func armedOfflineView() *offlineView { |
| 257 | w := armedMovesWorker.Load() |
| 258 | if w == nil || !w.cfg.OfflineCapable { |
| 259 | return nil |
| 260 | } |
| 261 | return w.cfg.Offline |
| 262 | } |
| 263 | |
| 264 | // armedOfflineCapable is whether this box publishes offline-move-1: the |
| 265 | // armed worker is capable and has its keep view, exactly the condition |
| 266 | // under which reconcileSerializedWith installs the offline hooks. |
| 267 | func armedOfflineCapable() bool { return armedOfflineView() != nil } |
| 268 | |
| 269 | // withOfflineHolds gives a reconcile pass the offline hooks, only when the |
| 270 | // caller and Phase 1's view set none (withMovesHolds runs first): the two |
| 271 | // are never armed together, since a box with Phase 1 on is not capable. |
| 272 | func withOfflineHolds(cfg provision.ReconcileConfig, hold func(provision.Record) (string, string, bool), v *offlineView) provision.ReconcileConfig { |
| 273 | if hold == nil || v == nil || cfg.Hold != nil || cfg.KeepParticipant != nil { |
| 274 | return cfg |
| 275 | } |
| 276 | cfg.Hold, cfg.KeepParticipant = hold, v.keep |
| 277 | if cfg.HoldWritten == nil && v.written != nil { |
| 278 | cfg.HoldWritten = v.written |
| 279 | } |
| 280 | return cfg |
| 281 | } |
| 282 | |
| 283 | // newMovesWorker is the worker for cfg: armMovesPass's, and each test box's. |
| 284 | func newMovesWorker(cfg moveCfg) *movesWorker { |
| 285 | return &movesWorker{cfg: cfg, mirrors: map[string]lease.MirrorStore{}} |
| 286 | } |
| 287 | |
| 288 | // fenceAsWith is moveCfg.FenceAs over an activator: FenceAndStopAs under a |
| 289 | // short deadline, so one slow stop does not hold up the pass. |
| 290 | // ErrStopUnproven still means the fence is installed, which is all the step |
| 291 | // asks of it; Quiesced answers the rest. |
| 292 | func fenceAsWith(svc *activator.Service) func(ctx context.Context, host string, generation uint64, owner string) error { |
| 293 | return func(ctx context.Context, host string, generation uint64, owner string) error { |
| 294 | fctx, cancel := context.WithTimeout(ctx, 2*time.Second) |
| 295 | defer cancel() |
| 296 | if _, err := svc.FenceAndStopAs(fctx, host, generation, owner); err != nil && !errors.Is(err, activator.ErrStopUnproven) { |
| 297 | return err |
| 298 | } |
| 299 | return nil |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | // hostSliceCgroup is the absolute cgroup directory of a nested host's |
| 304 | // systemd slice under cgroupRoot (the cgroup v2 mount): every role and |
| 305 | // generation unit of the host runs in it. |
| 306 | func hostSliceCgroup(cgroupRoot, host string) string { |
| 307 | return filepath.Join(cgroupRoot, filepath.FromSlash(path.Dir(instancectl.ClaimUnitControlGroup(host)))) |
| 308 | } |
| 309 | |
| 310 | // offlineRound is the moves pass's round for a record with an offline move, |
| 311 | // reached before every other branch of the pass (the Phase 1 skip and the |
| 312 | // lost-box gate included). A box that is not capable logs one notice per |
| 313 | // host and does nothing. Otherwise it runs the source step when its |
| 314 | // placement is move.from, the sealer's when it is move.sealer, and the |
| 315 | // target's when its placement is move.to, in that order, so the sealer |
| 316 | // reads the report this box has just written. |
| 317 | func (w *movesWorker) offlineRound(ctx context.Context, s release.Store, rec provision.Record) error { |
| 318 | cfg := w.cfg |
| 319 | m := rec.Move |
| 320 | if !m.Offline() { |
| 321 | return nil |
| 322 | } |
| 323 | here := cfg.parkRegion() |
| 324 | source := here != "" && park.PlacementMatches(cfg.ParkHomeRegion, m.From, here) |
| 325 | target := here != "" && park.PlacementMatches(cfg.ParkHomeRegion, m.To, here) |
| 326 | sealer := cfg.Box != "" && cfg.Box == m.Sealer |
| 327 | if !cfg.OfflineCapable { |
| 328 | if source || target || sealer { |
| 329 | w.notice(rec.Host+"#offline", "moves: %s: offline move %s waits: this box cannot take part in an offline move (it needs hibernation configured, and %s and %s unset)", rec.Host, m.ID, "CARLOS_HOST_CLAIM", moves.SwitchEnv) |
| 330 | } |
| 331 | return nil |
| 332 | } |
| 333 | if m.State == provision.OfflineStaging { |
| 334 | // The pre-copy (movesoffline_stage.go): nothing holds or fences |
| 335 | // the instance, and only the stager acts. |
| 336 | if source { |
| 337 | return w.offlineStageStep(ctx, s, rec) |
| 338 | } |
| 339 | return nil |
| 340 | } |
| 341 | var errs []error |
| 342 | if source { |
| 343 | errs = append(errs, w.offlineSourceStep(ctx, s, rec)) |
| 344 | } |
| 345 | if sealer { |
| 346 | errs = append(errs, w.offlineSealerStep(ctx, s, rec)) |
| 347 | } |
| 348 | if target { |
| 349 | errs = append(errs, w.offlineTargetStep(ctx, s, rec)) |
| 350 | } |
| 351 | return errors.Join(errs...) |
| 352 | } |
| 353 | |
| 354 | // offlineReassertEvery is how often a box that takes part in an offline move |
| 355 | // rewrites the reports a door judges by age (harness Run 9, finding 7). The |
| 356 | // doors believe a report for two minutes (offlineReportFresh, internal/ |
| 357 | // console), and the moves pass runs on a serial move or every five to six |
| 358 | // minutes. So a report written on the pass one door's bump started was stale |
| 359 | // by the time an operator who waited over two minutes ran the next door, and |
| 360 | // flip and open refused until the timed pass: most of Run 9's 366 s offline. |
| 361 | // At this cadence a participant's report is never much more than 20 s old. |
| 362 | const offlineReassertEvery = 15 * time.Second |
| 363 | |
| 364 | // offlineReassertRoles says which of this box's steps keep its report current |
| 365 | // while a move is in m's state: a source box's stop from stopping to landed, |
| 366 | // and the sealer's keep from sealed to landed. Those are the reports seal, |
| 367 | // flip and open judge by age. The seal itself (sealing) and the target's |
| 368 | // landing are not repeated here: they are long, and they run on the pass a |
| 369 | // door's serial bump starts. From open on no door reads a report by age. |
| 370 | // The sealer only re-reads and reports (offlineSealerRefresh): its ordinary |
| 371 | // pass re-asserts the tombstone, and the refresh never does. |
| 372 | func offlineReassertRoles(m *provision.MoveState, source, sealer bool) (asSource, asSealer bool) { |
| 373 | switch m.State { |
| 374 | case provision.OfflineStopping, provision.OfflineSealing: |
| 375 | return source, false |
| 376 | case provision.OfflineSealed, provision.OfflineLanding, provision.OfflineLanded: |
| 377 | return source, sealer |
| 378 | } |
| 379 | return false, false |
| 380 | } |
| 381 | |
| 382 | // offlineReassert is one tick of offlineReassertLoop: for each offline move |
| 383 | // the view holds in which this box's report is judged by age, it reads the |
| 384 | // record again and runs this box's source or sealer step for it, in |
| 385 | // offlineRound's order, under the pass lock. Each step is an ensure that ends |
| 386 | // by rewriting the box's report (offlineSourceStop, and offlineSealerRefresh, |
| 387 | // which checks the tombstone but never re-asserts it). With no |
| 388 | // such move in flight it reads nothing at all. It returns how many records it |
| 389 | // acted on. |
| 390 | func (w *movesWorker) offlineReassert(ctx context.Context, s release.Store) int { |
| 391 | cfg := w.cfg |
| 392 | v := cfg.Offline |
| 393 | if !cfg.OfflineCapable || v == nil || s == nil || cfg.Box == "" { |
| 394 | return 0 |
| 395 | } |
| 396 | snap := v.snap.Load() |
| 397 | if snap == nil || snap.failed || len(snap.moves) == 0 { |
| 398 | return 0 |
| 399 | } |
| 400 | here := cfg.parkRegion() |
| 401 | roles := func(m *provision.MoveState) (bool, bool) { |
| 402 | source := here != "" && park.PlacementMatches(cfg.ParkHomeRegion, m.From, here) |
| 403 | return offlineReassertRoles(m, source, m.Sealer == cfg.Box) |
| 404 | } |
| 405 | var hosts []string |
| 406 | for host, m := range snap.moves { |
| 407 | if asSource, asSealer := roles(&m); asSource || asSealer { |
| 408 | hosts = append(hosts, host) |
| 409 | } |
| 410 | } |
| 411 | if len(hosts) == 0 { |
| 412 | return 0 |
| 413 | } |
| 414 | sort.Strings(hosts) |
| 415 | w.pass.Lock() |
| 416 | defer w.pass.Unlock() |
| 417 | acted := 0 |
| 418 | for _, host := range hosts { |
| 419 | owner := snap.owners[host] |
| 420 | rec, err := provision.GetRecord(ctx, s, owner[0], owner[1], host) |
| 421 | if err != nil { |
| 422 | w.logTrouble(host, fmt.Errorf("offline move: read the record to report again: %w", err)) |
| 423 | continue |
| 424 | } |
| 425 | if !rec.Move.Offline() { |
| 426 | continue |
| 427 | } |
| 428 | asSource, asSealer := roles(rec.Move) |
| 429 | if !asSource && !asSealer { |
| 430 | continue |
| 431 | } |
| 432 | acted++ |
| 433 | var errs []error |
| 434 | if asSource { |
| 435 | errs = append(errs, w.offlineSourceStep(ctx, s, *rec)) |
| 436 | } |
| 437 | if asSealer { |
| 438 | errs = append(errs, w.offlineSealerRefresh(ctx, s, *rec)) |
| 439 | } |
| 440 | if err := errors.Join(errs...); err != nil { |
| 441 | w.logTrouble(host, err) |
| 442 | } |
| 443 | } |
| 444 | return acted |
| 445 | } |
| 446 | |
| 447 | // now is cfg.Now, or the wall clock when a test left it unset. |
| 448 | func (c moveCfg) now() time.Time { |
| 449 | if c.Now != nil { |
| 450 | return c.Now() |
| 451 | } |
| 452 | return time.Now() |
| 453 | } |
| 454 | |
| 455 | // writeOfflineReport writes r at its key (moves.OfflineReportKey), stamped |
| 456 | // with the report version, after the shape check every reader makes. The |
| 457 | // source, sealer and target steps all write through it. |
| 458 | func writeOfflineReport(ctx context.Context, s release.Store, r moves.OfflineReport) error { |
| 459 | r.V = moves.OfflineReportVersion |
| 460 | key := moves.OfflineReportKey(r.Host, r.Op, r.Role, r.Box) |
| 461 | if key == "" { |
| 462 | return fmt.Errorf("offline report: host %q, op %q, role %q or box %q is not usable", r.Host, r.Op, r.Role, r.Box) |
| 463 | } |
| 464 | if err := moves.ValidateOfflineReport(r); err != nil { |
| 465 | return err |
| 466 | } |
| 467 | body, err := json.Marshal(r) |
| 468 | if err != nil { |
| 469 | return err |
| 470 | } |
| 471 | // A report is rewritten every pass: read its etag and replace it |
| 472 | // (release.Store.Put with no etag is create-only). |
| 473 | return moves.PutReplace(ctx, s, key, body) |
| 474 | } |
| 475 | |
| 476 | // putOfflineCreateOnly writes body at key only if nothing is there yet: |
| 477 | // the immutable seal record (control/offline-moves/<host>/<id>/seal.json, |
| 478 | // moves.OfflineSealKey, written once by the sealer and never rewritten). |
| 479 | // It is an ensure: when the key already exists it writes nothing and |
| 480 | // returns what is stored, with created false, so a sealer re-running after |
| 481 | // a crash compares rather than overwrites, and a reader (flip, the landing) |
| 482 | // always sees the first write. key must be non-empty. |
| 483 | func putOfflineCreateOnly(ctx context.Context, s release.Store, key string, body []byte) (stored []byte, created bool, err error) { |
| 484 | if key == "" { |
| 485 | return nil, false, errors.New("offline move: no key to write") |
| 486 | } |
| 487 | _, err = s.Put(ctx, key, body, "") |
| 488 | if err == nil { |
| 489 | return body, true, nil |
| 490 | } |
| 491 | if !errors.Is(err, release.ErrCASConflict) { |
| 492 | return nil, false, fmt.Errorf("write %s: %w", key, err) |
| 493 | } |
| 494 | stored, _, err = s.Get(ctx, key) |
| 495 | if err != nil { |
| 496 | return nil, false, fmt.Errorf("read %s after a create-only write found it: %w", key, err) |
| 497 | } |
| 498 | return stored, false, nil |
| 499 | } |
| 500 | |
| 501 | // readOfflineReports reads every report of one role for one move, keyed by |
| 502 | // box. A report that does not validate is an error, never skipped. |
| 503 | func readOfflineReports(ctx context.Context, s release.Store, host, op, role string) (map[string]moves.OfflineReport, error) { |
| 504 | prefix := moves.OfflineReportPrefix(host, op) |
| 505 | if prefix == "" { |
| 506 | return nil, fmt.Errorf("offline report: host %q or op %q is not usable", host, op) |
| 507 | } |
| 508 | prefix += role + "/" |
| 509 | keys, err := s.List(ctx, prefix) |
| 510 | if err != nil { |
| 511 | return nil, err |
| 512 | } |
| 513 | out := map[string]moves.OfflineReport{} |
| 514 | for _, key := range keys { |
| 515 | if !strings.HasSuffix(key, ".json") { |
| 516 | continue |
| 517 | } |
| 518 | body, _, err := s.Get(ctx, key) |
| 519 | if err != nil { |
| 520 | return nil, err |
| 521 | } |
| 522 | var r moves.OfflineReport |
| 523 | if err := json.Unmarshal(body, &r); err != nil { |
| 524 | return nil, fmt.Errorf("%s: %w", key, err) |
| 525 | } |
| 526 | if err := moves.ValidateOfflineReport(r); err != nil { |
| 527 | return nil, fmt.Errorf("%s: %w", key, err) |
| 528 | } |
| 529 | if moves.OfflineReportKey(r.Host, r.Op, r.Role, r.Box) != key { |
| 530 | return nil, fmt.Errorf("%s: the report names %s/%s/%s/%s", key, r.Host, r.Op, r.Role, r.Box) |
| 531 | } |
| 532 | out[r.Box] = r |
| 533 | } |
| 534 | return out, nil |
| 535 | } |
| 536 | |
| 537 | // offlineSourceStep is a source-region box's part of an offline move. Every |
| 538 | // state's step is an ensure: it reads what is on disk, in the registry and |
| 539 | // in the activator, does only what is missing, and reports what it found. |
| 540 | func (w *movesWorker) offlineSourceStep(ctx context.Context, s release.Store, rec provision.Record) error { |
| 541 | if w.cfg.Box == "" { |
| 542 | return errors.New("offline move: this box has no label to report under") |
| 543 | } |
| 544 | switch rec.Move.State { |
| 545 | case provision.OfflineStopping, provision.OfflineSealing, provision.OfflineSealed, provision.OfflineLanding, provision.OfflineLanded: |
| 546 | // The pinned stager's stop acknowledgement first: the seal door |
| 547 | // needs it beside this box's stop (movesoffline_stage.go). |
| 548 | return errors.Join(w.offlineStagerStopped(ctx, s, rec), w.offlineSourceStop(ctx, s, rec)) |
| 549 | case provision.OfflineOpen: |
| 550 | return w.offlineSourcePark(ctx, s, rec) |
| 551 | case provision.OfflineFinishing: |
| 552 | return w.offlineSourceGone(ctx, s, rec) |
| 553 | case provision.OfflineReverting: |
| 554 | return w.offlineSourceRevert(ctx, s, rec) |
| 555 | } |
| 556 | return nil |
| 557 | } |
| 558 | |
| 559 | // offlineFence installs, or meets again, this move's activator fence for |
| 560 | // host. Fences live in the activator's memory, so every pass installs it |
| 561 | // again and an agent restart regains it on its next pass. True when the |
| 562 | // fence is this move's now. |
| 563 | func (w *movesWorker) offlineFence(ctx context.Context, host string, m *provision.MoveState) bool { |
| 564 | if w.cfg.FenceAs == nil { |
| 565 | return false |
| 566 | } |
| 567 | if err := w.cfg.FenceAs(ctx, host, moveFenceGeneration(m), offlineFenceOwner(m.ID)); err != nil { |
| 568 | w.notice(host+"#offline-fence", "moves: %s: offline move %s: the wake fence is not installed: %v", host, m.ID, err) |
| 569 | return false |
| 570 | } |
| 571 | return true |
| 572 | } |
| 573 | |
| 574 | // offlineLiftFence lifts this move's fence, and only it. |
| 575 | func (w *movesWorker) offlineLiftFence(host string, m *provision.MoveState) { |
| 576 | if w.cfg.LiftFenceAs != nil && w.cfg.LiftFenceAs(host, moveFenceGeneration(m), offlineFenceOwner(m.ID)) { |
| 577 | log.Printf("moves: %s: offline move %s: wake fence lifted", host, m.ID) |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | // offlineRow reads the host's row on this box; nil when it has none. |
| 582 | func (w *movesWorker) offlineRow(host string) (*registry.Route, error) { |
| 583 | if w.cfg.Route == nil { |
| 584 | return nil, errors.New("offline move: no registry on this box") |
| 585 | } |
| 586 | return w.cfg.Route(host) |
| 587 | } |
| 588 | |
| 589 | // heldBy reports whether rt is held by op. An absent row counts: nothing |
| 590 | // can serve or wake a host this box has no row for. |
| 591 | func heldBy(rt *registry.Route, op string) bool { |
| 592 | return rt == nil || (rt.HoldOp == op && registry.IsHold(rt.Status)) |
| 593 | } |
| 594 | |
| 595 | // offlineLocal is what a box keeps on disk for a host, judged by lstat and |
| 596 | // directory listings only, never by following a link. A row's own layout is |
| 597 | // read at the row's paths; a box with no row, or a leftover of the other |
| 598 | // layout, is read at both. |
| 599 | type offlineLocal struct { |
| 600 | LocalDB, Files, Env, Sidecar bool |
| 601 | LocalTXID uint64 |
| 602 | } |
| 603 | |
| 604 | // flatDBPath is where a flat row's database is: the row's own path when it |
| 605 | // has one and is flat, else the default. |
| 606 | func (c moveCfg) flatDBPath(host string, rt *registry.Route) string { |
| 607 | if rt != nil && rt.DBPath != "" && !provision.NestedPaths(c.DataDir, *rt) { |
| 608 | return rt.DBPath |
| 609 | } |
| 610 | return provision.DefaultDBPath(c.DataDir, host) |
| 611 | } |
| 612 | |
| 613 | // activatorMetaDir is the activator's litestream meta directory beside a |
| 614 | // database (localLTXPosition reads it). |
| 615 | func activatorMetaDir(dbPath string) string { |
| 616 | return filepath.Join(filepath.Dir(dbPath), ".activator-"+filepath.Base(dbPath)+"-litestream") |
| 617 | } |
| 618 | |
| 619 | func lexists(p string) bool { |
| 620 | _, err := os.Lstat(p) |
| 621 | return err == nil |
| 622 | } |
| 623 | |
| 624 | func isDir(p string) bool { |
| 625 | st, err := os.Lstat(p) |
| 626 | return err == nil && st.IsDir() |
| 627 | } |
| 628 | |
| 629 | // offlinePlatformName reports a nested instance directory's platform entry |
| 630 | // (layout.IsNestedPlatformName: the database and its files, the env files, |
| 631 | // the generation, seed and files markers, the adoption record, and the |
| 632 | // launch descriptors, launch.gen.<n>.json included): everything else in it |
| 633 | // is the app's files. |
| 634 | func offlinePlatformName(host, name string) bool { return layout.IsNestedPlatformName(host, name) } |
| 635 | |
| 636 | // localState reads what this box keeps for host: the DB file or either |
| 637 | // litestream meta dir (S8: a meta dir alone is local state, since |
| 638 | // localLTXPosition reads it), the files directory, the env file and a |
| 639 | // sidecar env file, under the row's layout and any leftover of the other. |
| 640 | func (w *movesWorker) localState(host string, rt *registry.Route) (offlineLocal, error) { |
| 641 | cfg := w.cfg |
| 642 | var out offlineLocal |
| 643 | nestedRow := rt != nil && provision.NestedPaths(cfg.DataDir, *rt) |
| 644 | // Flat: a flat row's own paths, or a nested row's flat leftovers. |
| 645 | { |
| 646 | db := cfg.flatDBPath(host, rt) |
| 647 | out.LocalDB = lexists(db) || lexists(db+"-litestream") || lexists(activatorMetaDir(db)) |
| 648 | out.Files = isDir(filepath.Join(cfg.DataDir, host)) |
| 649 | out.Env = lexists(filepath.Join(cfg.DataDir, host+".env")) |
| 650 | out.Sidecar = lexists(filepath.Join(cfg.DataDir, host+"-sidecar.env")) |
| 651 | pos, err := localLTXPosition(db) |
| 652 | if err != nil { |
| 653 | return out, err |
| 654 | } |
| 655 | out.LocalTXID = pos |
| 656 | } |
| 657 | // Nested, read through the raw directory (layout.RawInstanceDir). |
| 658 | raw := layout.RawInstanceDir(cfg.DataDir, host) |
| 659 | if nestedRow || isDir(raw) { |
| 660 | db := filepath.Join(raw, host+".db") |
| 661 | out.LocalDB = out.LocalDB || lexists(db) || lexists(db+"-litestream") || lexists(activatorMetaDir(db)) |
| 662 | out.Env = out.Env || lexists(filepath.Join(raw, host+".env")) |
| 663 | out.Sidecar = out.Sidecar || lexists(filepath.Join(raw, host+"-sidecar.env")) |
| 664 | ents, err := os.ReadDir(raw) |
| 665 | if err != nil && !errors.Is(err, os.ErrNotExist) { |
| 666 | return out, err |
| 667 | } |
| 668 | for _, e := range ents { |
| 669 | if !offlinePlatformName(host, e.Name()) { |
| 670 | out.Files = true |
| 671 | break |
| 672 | } |
| 673 | } |
| 674 | dbPath := db |
| 675 | if nestedRow && rt.DBPath != "" { |
| 676 | dbPath = rt.DBPath |
| 677 | } |
| 678 | pos, err := localLTXPosition(dbPath) |
| 679 | if err != nil { |
| 680 | return out, err |
| 681 | } |
| 682 | out.LocalTXID = max(out.LocalTXID, pos) |
| 683 | } |
| 684 | return out, nil |
| 685 | } |
| 686 | |
| 687 | // procs is the process count for the report: -1 when it cannot be read, |
| 688 | // which the flip door reads as not stopped. |
| 689 | func (w *movesWorker) procs(host string) int { |
| 690 | if w.cfg.Procs == nil { |
| 691 | return -1 |
| 692 | } |
| 693 | n, err := w.cfg.Procs(host) |
| 694 | if err != nil { |
| 695 | w.notice(host+"#offline-procs", "moves: %s: the host's processes cannot be counted on this box: %v", host, err) |
| 696 | return -1 |
| 697 | } |
| 698 | return n |
| 699 | } |
| 700 | |
| 701 | // sourceReport is the source report's evidence as it stands now. |
| 702 | func (w *movesWorker) sourceReport(host string, m *provision.MoveState, rt *registry.Route, fenced bool, phase string) (moves.OfflineReport, error) { |
| 703 | cfg := w.cfg |
| 704 | local, err := w.localState(host, rt) |
| 705 | if err != nil { |
| 706 | return moves.OfflineReport{}, err |
| 707 | } |
| 708 | quiesced := cfg.Quiesced != nil && cfg.Quiesced(host) |
| 709 | return moves.OfflineReport{ |
| 710 | Op: m.ID, Host: host, Box: cfg.Box, Role: moves.OfflineRoleSource, Phase: phase, At: cfg.now().UTC(), |
| 711 | Held: heldBy(rt, m.ID), Fenced: fenced, Quiesced: quiesced, RowPresent: rt != nil, |
| 712 | LocalDB: local.LocalDB, Files: local.Files, Env: local.Env, Sidecar: local.Sidecar, LocalTXID: local.LocalTXID, |
| 713 | Procs: w.procs(host), LayoutJournal: layout.JournalPresent(cfg.DataDir, host), |
| 714 | }, nil |
| 715 | } |
| 716 | |
| 717 | // offlineSourceStop is the source step from offline-stopping through |
| 718 | // offline-landed: install the fence, read the row and the local state, and |
| 719 | // rewrite this box's source report with At on every pass (the open door |
| 720 | // needs it current). The row's hold is reconcile's (offlineHold) and, |
| 721 | // after the flip, its keep (offlineView.keep). |
| 722 | func (w *movesWorker) offlineSourceStop(ctx context.Context, s release.Store, rec provision.Record) error { |
| 723 | m := rec.Move |
| 724 | fenced := w.offlineFence(ctx, rec.Host, m) |
| 725 | rt, err := w.offlineRow(rec.Host) |
| 726 | if err != nil { |
| 727 | return err |
| 728 | } |
| 729 | r, err := w.sourceReport(rec.Host, m, rt, fenced, moves.OfflinePhaseStopped) |
| 730 | if err != nil { |
| 731 | return err |
| 732 | } |
| 733 | return writeOfflineReport(ctx, s, r) |
| 734 | } |
| 735 | |
| 736 | // retirePath is one path of the manifest's source-retire row on this box: |
| 737 | // where it is, its name in the park directory, and, for a path the agent |
| 738 | // never renames, why. |
| 739 | type retirePath struct { |
| 740 | path, name, left string |
| 741 | } |
| 742 | |
| 743 | // Why a path is left where it is. |
| 744 | const ( |
| 745 | leftRootOwned = "root-owned: listed for the census, not renamed" |
| 746 | leftView = "the mount view of the instance directory: it stays, empty" |
| 747 | leftInstance = "the instance directory stays, empty" |
| 748 | ) |
| 749 | |
| 750 | // nestedParkPrefix is where a nested instance directory's entries go inside |
| 751 | // the park directory, apart from the flat names. |
| 752 | const nestedParkPrefix = "instance" |
| 753 | |
| 754 | // retirePaths is the manifest's source-retire row as paths on this box, in |
| 755 | // a fixed order: the row's layout in full, the other layout's paths only |
| 756 | // where they exist, the app's own entries of a nested instance directory as |
| 757 | // found there and in the park directory, then what is never renamed. |
| 758 | func (w *movesWorker) retirePaths(host, parkDir string, rt *registry.Route) ([]retirePath, error) { |
| 759 | cfg := w.cfg |
| 760 | nestedRow := rt != nil && provision.NestedPaths(cfg.DataDir, *rt) |
| 761 | raw := layout.RawInstanceDir(cfg.DataDir, host) |
| 762 | var out []retirePath |
| 763 | add := func(p, name string, always bool) { |
| 764 | if always || lexists(p) { |
| 765 | out = append(out, retirePath{path: p, name: name}) |
| 766 | } |
| 767 | } |
| 768 | flatDB := cfg.flatDBPath(host, rt) |
| 769 | flat := !nestedRow |
| 770 | add(flatDB, filepath.Base(flatDB), flat) |
| 771 | add(flatDB+"-wal", filepath.Base(flatDB)+"-wal", flat) |
| 772 | add(flatDB+"-shm", filepath.Base(flatDB)+"-shm", flat) |
| 773 | add(flatDB+"-litestream", filepath.Base(flatDB)+"-litestream", flat) |
| 774 | add(activatorMetaDir(flatDB), filepath.Base(activatorMetaDir(flatDB)), flat) |
| 775 | add(filepath.Join(cfg.DataDir, host), host, flat) |
| 776 | add(filepath.Join(cfg.DataDir, host+".env"), host+".env", flat) |
| 777 | add(filepath.Join(cfg.DataDir, host+".config.env"), host+".config.env", flat) |
| 778 | |
| 779 | nested := nestedRow || isDir(raw) |
| 780 | if nested { |
| 781 | named := map[string]bool{} |
| 782 | for _, n := range []string{host + ".db", host + ".db-wal", host + ".db-shm", host + ".db-litestream", ".activator-" + host + ".db-litestream", host + ".env", host + ".config.env"} { |
| 783 | named[n] = true |
| 784 | add(filepath.Join(raw, n), nestedParkPrefix+"/"+n, true) |
| 785 | } |
| 786 | // The app's files, and every other entry the directory holds: what |
| 787 | // is still there, and what an earlier pass already parked. |
| 788 | // A numbered copy in the park directory (parkName) belongs to the |
| 789 | // entry it is a copy of, not to the app. |
| 790 | others := map[string]bool{} |
| 791 | listed := map[string][]string{} |
| 792 | for _, dir := range []string{raw, filepath.Join(parkDir, nestedParkPrefix)} { |
| 793 | ents, err := os.ReadDir(dir) |
| 794 | if err != nil && !errors.Is(err, os.ErrNotExist) { |
| 795 | return nil, err |
| 796 | } |
| 797 | for _, e := range ents { |
| 798 | listed[dir] = append(listed[dir], e.Name()) |
| 799 | } |
| 800 | } |
| 801 | known := map[string]bool{} |
| 802 | for n := range named { |
| 803 | known[n] = true |
| 804 | } |
| 805 | for _, n := range listed[raw] { |
| 806 | known[n] = true |
| 807 | } |
| 808 | for _, n := range listed[filepath.Join(parkDir, nestedParkPrefix)] { |
| 809 | known[n] = true |
| 810 | } |
| 811 | for _, n := range listed[raw] { |
| 812 | if !named[n] { |
| 813 | others[n] = true |
| 814 | } |
| 815 | } |
| 816 | for _, n := range listed[filepath.Join(parkDir, nestedParkPrefix)] { |
| 817 | if !named[n] && !isNumberedCopy(n, known) { |
| 818 | others[n] = true |
| 819 | } |
| 820 | } |
| 821 | names := make([]string, 0, len(others)) |
| 822 | for n := range others { |
| 823 | names = append(names, n) |
| 824 | } |
| 825 | sort.Strings(names) |
| 826 | for _, n := range names { |
| 827 | add(filepath.Join(raw, n), nestedParkPrefix+"/"+n, true) |
| 828 | } |
| 829 | } |
| 830 | add(filepath.Join(cfg.DataDir, "programs", host+".json"), host+".json", true) |
| 831 | if nested { |
| 832 | out = append(out, retirePath{path: provision.InstanceDir(cfg.DataDir, host), left: leftView}) |
| 833 | out = append(out, retirePath{path: raw, left: leftInstance}) |
| 834 | } |
| 835 | optDir := cfg.OptDir |
| 836 | if optDir == "" { |
| 837 | optDir = release.OptDir() |
| 838 | } |
| 839 | out = append(out, retirePath{path: filepath.Join(optDir, "live", host), left: leftRootOwned}) |
| 840 | return out, nil |
| 841 | } |
| 842 | |
| 843 | // parkCopySep separates a name from its copy number in the park directory |
| 844 | // (parkName). Chosen to be unlikely in an app's own file names. |
| 845 | const parkCopySep = ".parked-" |
| 846 | |
| 847 | // isNumberedCopy reports whether n is a known entry's numbered park copy |
| 848 | // (parkName's "<name>.parked-<n>"), which belongs to that entry. |
| 849 | func isNumberedCopy(n string, known map[string]bool) bool { |
| 850 | i := strings.LastIndex(n, parkCopySep) |
| 851 | if i <= 0 || !known[n[:i]] || i+len(parkCopySep) == len(n) { |
| 852 | return false |
| 853 | } |
| 854 | for _, c := range n[i+len(parkCopySep):] { |
| 855 | if c < '0' || c > '9' { |
| 856 | return false |
| 857 | } |
| 858 | } |
| 859 | return true |
| 860 | } |
| 861 | |
| 862 | // parkName is where name goes in dir: name itself, or name.parked-<n> for |
| 863 | // the first free n when something of that name was parked before (the |
| 864 | // source path came back after an earlier pass parked it). The park |
| 865 | // directory is the move's own, so the first park of every path lands at its |
| 866 | // plain name and a later pass reports it from there. |
| 867 | func parkName(dir, name string) string { |
| 868 | dst := filepath.Join(dir, name) |
| 869 | if !lexists(dst) { |
| 870 | return dst |
| 871 | } |
| 872 | for n := 1; ; n++ { |
| 873 | if c := dst + parkCopySep + fmt.Sprint(n); !lexists(c) { |
| 874 | return c |
| 875 | } |
| 876 | } |
| 877 | } |
| 878 | |
| 879 | // Why a rename the park attempted did not happen. A path whose rename fails |
| 880 | // with an error a later pass would meet again (a mount, EXDEV, EBUSY; a path |
| 881 | // the agent may not rename, EPERM or EACCES; a read-only filesystem) is |
| 882 | // left for good: reported once, with leftCannot and the error, and never |
| 883 | // tried again for this move, so a pass does not re-fence, fail and lift |
| 884 | // over it. Any other error is leftNotYet, and the next pass tries again. |
| 885 | const ( |
| 886 | leftCannot = "cannot be renamed: " |
| 887 | leftNotYet = "not renamed yet: " |
| 888 | ) |
| 889 | |
| 890 | // renameTerminal reports whether a rename error is one a later pass would |
| 891 | // meet again. |
| 892 | func renameTerminal(err error) bool { |
| 893 | return errors.Is(err, os.ErrPermission) || errors.Is(err, syscall.EXDEV) || |
| 894 | errors.Is(err, syscall.EBUSY) || errors.Is(err, syscall.EROFS) |
| 895 | } |
| 896 | |
| 897 | // retire reports one path's action from what is on disk, renaming it into |
| 898 | // the park directory first when park is true and it is there to rename. |
| 899 | // cannot is the reason an earlier pass recorded for a path it found it can |
| 900 | // never rename ("" when none): such a path is reported left again, as it |
| 901 | // is, and not tried. Nothing is ever deleted. |
| 902 | func retire(p retirePath, parkDir string, park bool, cannot string) moves.RetiredPath { |
| 903 | present := lexists(p.path) |
| 904 | switch { |
| 905 | case p.left != "" && present: |
| 906 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredLeft, Reason: p.left} |
| 907 | case p.left != "": |
| 908 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredAbsent} |
| 909 | case present && cannot != "": |
| 910 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredLeft, Reason: cannot} |
| 911 | case present && park: |
| 912 | dst := parkName(parkDir, p.name) |
| 913 | err := os.MkdirAll(filepath.Dir(dst), 0o700) |
| 914 | if err == nil { |
| 915 | err = os.Rename(p.path, dst) |
| 916 | } |
| 917 | switch { |
| 918 | case err == nil: |
| 919 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredParked} |
| 920 | case errors.Is(err, os.ErrNotExist) && !lexists(p.path): |
| 921 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredAbsent} |
| 922 | case renameTerminal(err): |
| 923 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredLeft, Reason: leftCannot + err.Error()} |
| 924 | } |
| 925 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredLeft, Reason: leftNotYet + err.Error()} |
| 926 | case present: |
| 927 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredLeft, Reason: leftNotYet + "the stop could not be confirmed this pass"} |
| 928 | case lexists(filepath.Join(parkDir, p.name)): |
| 929 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredParked} |
| 930 | } |
| 931 | return moves.RetiredPath{Path: p.path, Action: moves.RetiredAbsent} |
| 932 | } |
| 933 | |
| 934 | // cannotRename reads this box's own last report for the move and returns |
| 935 | // the paths it recorded as never renameable (leftCannot), with their |
| 936 | // reasons. Reading them from the report, not from memory, keeps the answer |
| 937 | // across an agent restart. |
| 938 | func (w *movesWorker) cannotRename(ctx context.Context, s release.Store, host, op string) (map[string]string, error) { |
| 939 | out := map[string]string{} |
| 940 | body, _, err := s.Get(ctx, moves.OfflineReportKey(host, op, moves.OfflineRoleSource, w.cfg.Box)) |
| 941 | if errors.Is(err, os.ErrNotExist) { |
| 942 | return out, nil |
| 943 | } |
| 944 | if err != nil { |
| 945 | return nil, err |
| 946 | } |
| 947 | var r moves.OfflineReport |
| 948 | if err := json.Unmarshal(body, &r); err != nil { |
| 949 | return nil, err |
| 950 | } |
| 951 | for _, rp := range r.Retired { |
| 952 | if rp.Action == moves.RetiredLeft && strings.HasPrefix(rp.Reason, leftCannot) { |
| 953 | out[rp.Path] = rp.Reason |
| 954 | } |
| 955 | } |
| 956 | return out, nil |
| 957 | } |
| 958 | |
| 959 | // offlineSourcePark is the source step at offline-open: with the fence held, |
| 960 | // the row held (or absent), the host quiesced, no process left and no layout |
| 961 | // journal, rename every path of the source-retire row that is still here |
| 962 | // into the park directory; then lift the fence and report parked with the |
| 963 | // inventory. A pass that finds nothing left to rename does not fence again: |
| 964 | // it reports from what is parked. A path that can never be renamed |
| 965 | // (leftCannot) is reported left with its reason and does not count as left |
| 966 | // to rename, so it neither blocks the parked report nor brings the fence |
| 967 | // back on every pass. A rename that may yet succeed (leftNotYet) reports |
| 968 | // failed, and the next pass tries again. Nothing is deleted. |
| 969 | func (w *movesWorker) offlineSourcePark(ctx context.Context, s release.Store, rec provision.Record) error { |
| 970 | cfg := w.cfg |
| 971 | m, host := rec.Move, rec.Host |
| 972 | parkDir := moves.OfflineParkDir(cfg.DataDir, host, m.ID, moves.OfflineRoleSource) |
| 973 | if parkDir == "" { |
| 974 | return fmt.Errorf("offline move %s: no park directory for %s under %q", m.ID, host, cfg.DataDir) |
| 975 | } |
| 976 | rt, err := w.offlineRow(host) |
| 977 | if err != nil { |
| 978 | return err |
| 979 | } |
| 980 | paths, err := w.retirePaths(host, parkDir, rt) |
| 981 | if err != nil { |
| 982 | return err |
| 983 | } |
| 984 | cannot, err := w.cannotRename(ctx, s, host, m.ID) |
| 985 | if err != nil { |
| 986 | return err |
| 987 | } |
| 988 | pending := false |
| 989 | for _, p := range paths { |
| 990 | if p.left == "" && cannot[p.path] == "" && lexists(p.path) { |
| 991 | pending = true |
| 992 | break |
| 993 | } |
| 994 | } |
| 995 | if pending { |
| 996 | fenced := w.offlineFence(ctx, host, m) |
| 997 | quiesced := cfg.Quiesced != nil && cfg.Quiesced(host) |
| 998 | procs := w.procs(host) |
| 999 | journal := layout.JournalPresent(cfg.DataDir, host) |
| 1000 | if !fenced || !heldBy(rt, m.ID) || !quiesced || procs != 0 || journal { |
| 1001 | return fmt.Errorf("offline move %s: not parking %s yet: fenced %v, held %v, quiesced %v, processes %d, layout journal %v", |
| 1002 | m.ID, host, fenced, heldBy(rt, m.ID), quiesced, procs, journal) |
| 1003 | } |
| 1004 | if err := os.MkdirAll(parkDir, 0o700); err != nil { |
| 1005 | return err |
| 1006 | } |
| 1007 | } |
| 1008 | retired := make([]moves.RetiredPath, 0, len(paths)) |
| 1009 | retry := false |
| 1010 | for _, p := range paths { |
| 1011 | rp := retire(p, parkDir, pending, cannot[p.path]) |
| 1012 | if rp.Action == moves.RetiredLeft && strings.HasPrefix(rp.Reason, leftNotYet) { |
| 1013 | retry = true |
| 1014 | } |
| 1015 | retired = append(retired, rp) |
| 1016 | } |
| 1017 | w.offlineLiftFence(host, m) |
| 1018 | phase := moves.OfflinePhaseParked |
| 1019 | if retry { |
| 1020 | phase = moves.OfflinePhaseFailed |
| 1021 | } |
| 1022 | r, err := w.sourceReport(host, m, rt, false, phase) |
| 1023 | if err != nil { |
| 1024 | return err |
| 1025 | } |
| 1026 | r.Retired = retired |
| 1027 | if err := writeOfflineReport(ctx, s, r); err != nil { |
| 1028 | return err |
| 1029 | } |
| 1030 | if retry { |
| 1031 | return fmt.Errorf("offline move %s: %s: some paths were not renamed this pass; the next pass tries again", m.ID, host) |
| 1032 | } |
| 1033 | return nil |
| 1034 | } |
| 1035 | |
| 1036 | // offlineSourceGone is the source step at offline-finishing: keep no longer |
| 1037 | // holds the row, so reconcile removes it; once it is gone, report gone. The |
| 1038 | // fence was lifted at open and is lifted again, should a restart have |
| 1039 | // re-fenced in between. |
| 1040 | func (w *movesWorker) offlineSourceGone(ctx context.Context, s release.Store, rec provision.Record) error { |
| 1041 | m, host := rec.Move, rec.Host |
| 1042 | rt, err := w.offlineRow(host) |
| 1043 | if err != nil { |
| 1044 | return err |
| 1045 | } |
| 1046 | w.offlineLiftFence(host, m) |
| 1047 | if rt != nil { |
| 1048 | w.notice(host+"#offline-gone", "moves: %s: offline move %s: waiting for reconcile to remove the row", host, m.ID) |
| 1049 | return nil |
| 1050 | } |
| 1051 | r, err := w.sourceReport(host, m, nil, false, moves.OfflinePhaseGone) |
| 1052 | if err != nil { |
| 1053 | return err |
| 1054 | } |
| 1055 | return writeOfflineReport(ctx, s, r) |
| 1056 | } |
| 1057 | |
| 1058 | // offlineSourceRevert is every source box's part of offline-reverting: the |
| 1059 | // row leaves this move's hold (ReleaseHold under the host's hold lock), the |
| 1060 | // fence is lifted, and the box reports reverted. The sealer's extra part is |
| 1061 | // its own step's (Task 4). |
| 1062 | // |
| 1063 | // The release also holds reconcileMu: a reconcile pass that listed the |
| 1064 | // records before the abort would otherwise mint a missing row held again |
| 1065 | // from its stale listing after the release, and a row held by a move that |
| 1066 | // has finished is never released. A pass that starts after the release |
| 1067 | // lists the record at offline-reverting, which holds nothing. |
| 1068 | func (w *movesWorker) offlineSourceRevert(ctx context.Context, s release.Store, rec provision.Record) error { |
| 1069 | cfg := w.cfg |
| 1070 | m, host := rec.Move, rec.Host |
| 1071 | if cfg.ReleaseHold == nil { |
| 1072 | return errors.New("offline move: no registry to release the hold through") |
| 1073 | } |
| 1074 | var relErr error |
| 1075 | reconcileMu.Lock() |
| 1076 | provision.WithHostHoldLock(host, func() { |
| 1077 | _, relErr = cfg.ReleaseHold(host, m.ID) |
| 1078 | }) |
| 1079 | reconcileMu.Unlock() |
| 1080 | if relErr != nil { |
| 1081 | return relErr |
| 1082 | } |
| 1083 | w.offlineLiftFence(host, m) |
| 1084 | rt, err := w.offlineRow(host) |
| 1085 | if err != nil { |
| 1086 | return err |
| 1087 | } |
| 1088 | if rt != nil && rt.HoldOp == m.ID { |
| 1089 | return fmt.Errorf("offline move %s: %s is still held by the move", m.ID, host) |
| 1090 | } |
| 1091 | // A staged move's pre-copy is removed before this box says reverted. |
| 1092 | if err := w.offlineUnstage(ctx, s, rec); err != nil { |
| 1093 | return err |
| 1094 | } |
| 1095 | r, err := w.sourceReport(host, m, rt, false, moves.OfflinePhaseReverted) |
| 1096 | if err != nil { |
| 1097 | return err |
| 1098 | } |
| 1099 | r.Held = false |
| 1100 | return writeOfflineReport(ctx, s, r) |
| 1101 | } |
| 1102 | |
amadan