| 1 | // The relay's write side (relay spec §5): the writer lease that makes one |
| 2 | // console the only process writing relay state, the operations that turn it |
| 3 | // on and off for an app, the etag-first publish of the authority record every |
| 4 | // relay box reads, and the five-minute reconcile that converges both derived |
| 5 | // artefacts from the records whatever happened in between. |
| 6 | // |
| 7 | // THREE OBJECTS, ONE SOURCE OF TRUTH. `apps/<acct>/<app>/relay.json` (the |
| 8 | // relay.Record) is the truth for an app. `control/relay.json` (the |
| 9 | // relay.Authority) is DERIVED from every record and never patched: |
| 10 | // publishing it is read-the-etag-first, list, build, CAS, and start over on |
| 11 | // conflict, so a disable that lands during the listing makes the write fail |
| 12 | // rather than quietly restore the app. `control/relay-issuer.json` is the |
| 13 | // unsealed pointer saying WHICH staged master a region issues from; adding a |
| 14 | // master to the console's secret issues nothing until that pointer moves. |
| 15 | // |
| 16 | // ONE WRITER, ENFORCED, AND THE WALK IS RESUMABLE. Every operation here |
| 17 | // writes its intent to the record first (`op` bumped, `phase` set), then |
| 18 | // writes config, then publishes, advancing `phase` after each step. That is |
| 19 | // what makes an interrupted operation finishable: the reconcile does not look |
| 20 | // for unfinished records, it re-derives what every record implies and fixes |
| 21 | // whatever disagrees. The lease is the fence around all of it — a process |
| 22 | // that loses a renewal stops before its next bucket write and abandons the |
| 23 | // operation, and the next holder's reconcile finishes it. |
| 24 | // |
| 25 | // THE CONSOLE HOLDS MASTERS BUT OPENS NOTHING. relayMasters reads the |
| 26 | // region's masters out of this console's OWN environment |
| 27 | // (CARLOS_RELAY_MASTERS_<REGION>, delivered like every other console |
| 28 | // variable) and derives an app key from them; the value it delivers is |
| 29 | // SEALED with CARLOS_SECRETS_PUBKEY, which this console cannot open — the |
| 30 | // same one-way street apiconfig.go's package comment describes, and the |
| 31 | // reason a console with no sealing key refuses to enable rather than mint a |
| 32 | // key it could never deliver. |
| 33 | // |
| 34 | // Spec: docs/superpowers/specs/2026-09-07-relay-stun-turn-design.md §4, §5. |
| 35 | package console |
| 36 | |
| 37 | import ( |
| 38 | "context" |
| 39 | "crypto/rand" |
| 40 | "crypto/sha256" |
| 41 | "encoding/base64" |
| 42 | "encoding/hex" |
| 43 | "encoding/json" |
| 44 | "errors" |
| 45 | "fmt" |
| 46 | "log" |
| 47 | "maps" |
| 48 | "os" |
| 49 | "slices" |
| 50 | "sort" |
| 51 | "strings" |
| 52 | "sync" |
| 53 | "time" |
| 54 | |
| 55 | "github.com/carlosframework/platform/internal/appconfig" |
| 56 | "github.com/carlosframework/platform/internal/provision" |
| 57 | "github.com/carlosframework/platform/internal/registry" |
| 58 | "github.com/carlosframework/platform/internal/relay" |
| 59 | "github.com/carlosframework/platform/internal/release" |
| 60 | ) |
| 61 | |
| 62 | const ( |
| 63 | // relayWriterKey is the writer lease: {holder, expires_at}, CAS-written. |
| 64 | relayWriterKey = "control/relay-writer.json" |
| 65 | // relayIssuerKey is the issuer pointer: {"<region>": "<kid>"}. Unsealed |
| 66 | // on purpose — it names a key, it is not one. |
| 67 | relayIssuerKey = "control/relay-issuer.json" |
| 68 | // relayAckPrefix is where each armed relay writes its heartbeat. |
| 69 | relayAckPrefix = "control/relay-ack/" |
| 70 | // configAckPrefix is where each box's config sync writes a per-route |
| 71 | // config-adoption acknowledgement (spec §5, internal/appconfig/ |
| 72 | // configack.go): configAckPrefix + "<box>/<route host>.json". |
| 73 | configAckPrefix = "control/config-ack/" |
| 74 | |
| 75 | // The lease's two numbers, from spec §5: sixty seconds of validity, |
| 76 | // renewed every twenty. Three renewals inside one TTL means a single |
| 77 | // slow bucket round trip costs nothing, while a console that dies is |
| 78 | // replaced within a minute. |
| 79 | relayLeaseTTL = 60 * time.Second |
| 80 | relayLeaseRenew = 20 * time.Second |
| 81 | // relayLeaseWait bounds how long StartRelay waits for the lease before |
| 82 | // its first reconcile. A console that does not get it simply reconciles |
| 83 | // nothing — the holder is doing it. |
| 84 | relayLeaseWait = 5 * time.Second |
| 85 | |
| 86 | // relayReconcileInterval is the convergence floor from spec §5: a stale |
| 87 | // config write is undone, an interrupted enable delivered, and a |
| 88 | // half-done deletion repaired, within this. |
| 89 | relayReconcileInterval = 5 * time.Minute |
| 90 | |
| 91 | // relayHeartbeatFresh is how recent a relay's last SUCCESSFUL authority |
| 92 | // read must be for its heartbeat to count as a live answer — the same |
| 93 | // five minutes `carlos relay status` calls "stale" past. |
| 94 | relayHeartbeatFresh = 5 * time.Minute |
| 95 | // relayRetireAfter is the earliest a rotated-away master may be retired |
| 96 | // (spec §5, step 4): 28 hours, comfortably past the four-hour ceiling on |
| 97 | // any credential minted just before the pointer moved. |
| 98 | relayRetireAfter = 28 * time.Hour |
| 99 | |
| 100 | // relayRouteMax caps the instance records one delivery check reads. An |
| 101 | // app with more routes than this has bigger questions than the relay. |
| 102 | relayRouteMax = 500 |
| 103 | |
| 104 | // relayVar is the one variable this whole file exists to deliver. |
| 105 | relayVar = "CARLOS_RELAY" |
| 106 | ) |
| 107 | |
| 108 | // The three phases of an operation's walk (spec §5). `done` is the only one |
| 109 | // that means "nothing is owed". |
| 110 | const ( |
| 111 | relayPhaseConfig = "config" |
| 112 | relayPhaseAggregate = "aggregate" |
| 113 | relayPhaseDone = "done" |
| 114 | ) |
| 115 | |
| 116 | // The refusals callers (and, in Task 11, doors) match on. Each is a |
| 117 | // configuration fact rather than a failure of the request, which is why they |
| 118 | // are sentinels rather than sentences built at the call site: the door turns |
| 119 | // them into a 503 naming the variable, and the CLI prints them verbatim. |
| 120 | var ( |
| 121 | errRelayNoBucket = errors.New("no deployment bucket on this host — this console cannot serve the relay") |
| 122 | errRelayNoRegions = errors.New( |
| 123 | "this deployment runs no relay (CARLOS_RELAY_REGIONS is empty) — ask the operator to arm a region before enabling it for an app") |
| 124 | errRelayNoSealKey = errors.New( |
| 125 | "this console has no secrets key (CARLOS_SECRETS_PUBKEY) — it cannot deliver a relay key, so it will not issue one") |
| 126 | // The commonest cause is not another console at all: it is THIS console |
| 127 | // having just restarted, leaving its own previous process's lease to |
| 128 | // expire. That takes up to relayLeaseTTL, so the message says to retry |
| 129 | // and names the wait rather than leaving an operator to guess whether |
| 130 | // the deployment is wedged (discussion #120). |
| 131 | errRelayNotHeld = errors.New( |
| 132 | "this console does not hold the relay writer lease. another console is driving the relay, " + |
| 133 | "or this one restarted and its previous lease has not expired yet. the operation was not started, " + |
| 134 | "so retry in up to a minute") |
| 135 | errRelayNotEnabled = errors.New("the relay is not enabled for this app") |
| 136 | errRelayUndelivered = errors.New("a route has no config delivery") |
| 137 | // errRelayUnknownHost is `carlos relay adopted --host` naming something |
| 138 | // that is not a route of this app — the door's 404. |
| 139 | errRelayUnknownHost = errors.New("no such route on this app") |
| 140 | // errRelayNotMemberRoute is `carlos relay adopted --host` naming a route |
| 141 | // the config sync restarts itself. Spec §5, step 5 gives a route exactly |
| 142 | // one way to establish its cutover, decided by shape: a unit or |
| 143 | // socket-exec route's cutover is the ack the config sync writes after it |
| 144 | // restarted the process, and a member's word would override an observed |
| 145 | // fact with a claim. |
| 146 | errRelayNotMemberRoute = errors.New("this route's cutover is observed, not declared") |
| 147 | ) |
| 148 | |
| 149 | // relayLease is control/relay-writer.json. |
| 150 | type relayLease struct { |
| 151 | Holder string `json:"holder"` |
| 152 | ExpiresAt time.Time `json:"expires_at"` |
| 153 | } |
| 154 | |
| 155 | // relayHolderID draws this process's lease identity: 16 random hex |
| 156 | // characters, like an incarnation and for the same reason — two consoles |
| 157 | // must never collide, and a counter or a hostname can. |
| 158 | func relayHolderID() string { |
| 159 | var b [8]byte |
| 160 | if _, err := rand.Read(b[:]); err != nil { |
| 161 | panic("console: crypto/rand: " + err.Error()) |
| 162 | } |
| 163 | return hex.EncodeToString(b[:]) |
| 164 | } |
| 165 | |
| 166 | // relayTime is every relay path's clock, seamed for the lease tests. |
| 167 | func (s *Server) relayTime() time.Time { |
| 168 | if s.relayNow != nil { |
| 169 | return s.relayNow() |
| 170 | } |
| 171 | return time.Now() |
| 172 | } |
| 173 | |
| 174 | // relayHold reports whether this process currently holds the writer lease. |
| 175 | func (s *Server) relayHold() bool { return s.relayHeld.Load() } |
| 176 | |
| 177 | // requireRelayLease is the check every relay bucket write makes FIRST. A |
| 178 | // process that lost its renewal stops here rather than completing a walk |
| 179 | // whose next step another console has already started — spec §5's "stops |
| 180 | // every relay write in flight before its next bucket write". |
| 181 | func (s *Server) requireRelayLease() error { |
| 182 | if !s.relayHold() { |
| 183 | return errRelayNotHeld |
| 184 | } |
| 185 | return nil |
| 186 | } |
| 187 | |
| 188 | // relayAcquire is one acquire-or-renew attempt, by CAS on the lease object. |
| 189 | // Absent, expired, or already ours means we take it; anyone else's live lease |
| 190 | // means we do not, and — this is the load-bearing half — we drop `held` so |
| 191 | // any operation in flight in this process fails its next check. |
| 192 | func (s *Server) relayAcquire(ctx context.Context) error { |
| 193 | if s.cfg.Releases == nil { |
| 194 | return errRelayNoBucket |
| 195 | } |
| 196 | now := s.relayTime() |
| 197 | body, etag, err := s.cfg.Releases.Get(ctx, relayWriterKey) |
| 198 | switch { |
| 199 | case errors.Is(err, os.ErrNotExist): |
| 200 | body, etag = nil, "" |
| 201 | case err != nil: |
| 202 | s.relayHeld.Store(false) |
| 203 | return err |
| 204 | } |
| 205 | var cur relayLease |
| 206 | if len(body) > 0 { |
| 207 | if err := json.Unmarshal(body, &cur); err != nil { |
| 208 | // An unparseable lease is not a live one: it names nobody this |
| 209 | // process can defer to. Taking it over with the etag we read is |
| 210 | // still a CAS, so a concurrent repair wins the race, not us. |
| 211 | log.Printf("console: relay writer lease is unreadable, taking it over: %v", err) |
| 212 | cur = relayLease{} |
| 213 | } |
| 214 | } |
| 215 | if cur.Holder != "" && cur.Holder != s.relayHolder && now.Before(cur.ExpiresAt) { |
| 216 | s.relayHeld.Store(false) |
| 217 | return fmt.Errorf("relay writer lease is held by %s until %s: %w", |
| 218 | cur.Holder, cur.ExpiresAt.UTC().Format(time.RFC3339), errRelayNotHeld) |
| 219 | } |
| 220 | next, err := json.Marshal(relayLease{Holder: s.relayHolder, ExpiresAt: now.Add(relayLeaseTTL)}) |
| 221 | if err != nil { |
| 222 | s.relayHeld.Store(false) |
| 223 | return err |
| 224 | } |
| 225 | if _, err := s.cfg.Releases.Put(ctx, relayWriterKey, next, etag); err != nil { |
| 226 | // A lost CAS is another console taking the same lapsed lease. Held |
| 227 | // goes false either way: a write we did not land is a lease we do |
| 228 | // not hold. |
| 229 | s.relayHeld.Store(false) |
| 230 | return err |
| 231 | } |
| 232 | s.relayHeld.Store(true) |
| 233 | return nil |
| 234 | } |
| 235 | |
| 236 | // relayWriterLoop acquires the lease now and renews it every twenty seconds |
| 237 | // until ctx ends. It is the ONLY writer of s.relayHeld outside a test. |
| 238 | // |
| 239 | // A failed attempt is logged at most as a line, never fatal: the ordinary |
| 240 | // steady state of a two-console deployment is one process failing this call |
| 241 | // forever while the other holds the lease, and a deployment overlap must not |
| 242 | // fill a log with alarm. |
| 243 | func (s *Server) relayWriterLoop(ctx context.Context) { |
| 244 | t := time.NewTicker(relayLeaseRenew) |
| 245 | defer t.Stop() |
| 246 | held := false |
| 247 | for { |
| 248 | err := s.relayAcquire(ctx) |
| 249 | switch { |
| 250 | case err == nil && !held: |
| 251 | log.Printf("console: relay writer lease acquired (holder %s)", s.relayHolder) |
| 252 | held = true |
| 253 | case err != nil && held: |
| 254 | log.Printf("console: relay writer lease lost: %v", err) |
| 255 | held = false |
| 256 | case err != nil && !errors.Is(err, errRelayNotHeld) && !errors.Is(err, release.ErrCASConflict): |
| 257 | log.Printf("console: relay writer lease: %v", err) |
| 258 | } |
| 259 | select { |
| 260 | case <-ctx.Done(): |
| 261 | return |
| 262 | case <-t.C: |
| 263 | } |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | // StartRelay runs the relay's two background loops: the writer lease, and the |
| 268 | // converging reconcile every five minutes and once at start. Started from |
| 269 | // cmd/console beside the other passes — never from NewServer, which stays |
| 270 | // goroutine-free so constructing a Server in a test starts nothing. |
| 271 | // |
| 272 | // Gated on a configured region list rather than on holding the lease: a |
| 273 | // deployment with no relay has nothing to converge and should not be reading |
| 274 | // a lease object every twenty seconds forever. |
| 275 | func (s *Server) StartRelay(ctx context.Context) { |
| 276 | if s.cfg.Releases == nil || len(s.relayRegions()) == 0 { |
| 277 | return |
| 278 | } |
| 279 | go s.relayWriterLoop(ctx) |
| 280 | // The first reconcile is the one that finishes whatever the previous |
| 281 | // holder abandoned, so it is worth a short wait for the lease rather |
| 282 | // than five minutes of drift. |
| 283 | s.relayAwaitLease(ctx, relayLeaseWait) |
| 284 | t := time.NewTicker(relayReconcileInterval) |
| 285 | defer t.Stop() |
| 286 | for { |
| 287 | if s.relayHold() { |
| 288 | if err := s.relayReconcile(ctx); err != nil { |
| 289 | log.Printf("console: relay reconcile: %v", err) |
| 290 | } |
| 291 | } |
| 292 | select { |
| 293 | case <-ctx.Done(): |
| 294 | return |
| 295 | case <-t.C: |
| 296 | } |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | // relayAwaitLease waits up to d for the writer loop's first acquire. |
| 301 | func (s *Server) relayAwaitLease(ctx context.Context, d time.Duration) { |
| 302 | deadline := time.Now().Add(d) |
| 303 | for !s.relayHold() && time.Now().Before(deadline) { |
| 304 | select { |
| 305 | case <-ctx.Done(): |
| 306 | return |
| 307 | case <-time.After(25 * time.Millisecond): |
| 308 | } |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | // ---- masters and the issuer pointer ----------------------------------------- |
| 313 | |
| 314 | // relayMastersEnv is the console variable holding one region's masters: |
| 315 | // CARLOS_RELAY_MASTERS_EU_WEST_1 for eu-west-1 (uppercased, hyphens to |
| 316 | // underscores, since variable names refuse hyphens — spec §4). |
| 317 | func relayMastersEnv(region string) string { |
| 318 | return "CARLOS_RELAY_MASTERS_" + strings.ToUpper(strings.ReplaceAll(region, "-", "_")) |
| 319 | } |
| 320 | |
| 321 | // relayStagedMaster is one `kid:key` pair out of that variable, in the order |
| 322 | // the operator wrote them. |
| 323 | type relayStagedMaster struct { |
| 324 | Kid string |
| 325 | Key []byte |
| 326 | } |
| 327 | |
| 328 | // relayStagedMasters parses CARLOS_RELAY_MASTERS_<REGION> — read from the |
| 329 | // process environment AT USE TIME rather than cached at boot, so an operator |
| 330 | // who stages a master and restarts nothing still has it available to the next |
| 331 | // rotate. |
| 332 | // |
| 333 | // Every defect is a refusal naming the variable: this is key material, and a |
| 334 | // half-parsed list would silently issue from the wrong master. |
| 335 | func relayStagedMasters(region string) ([]relayStagedMaster, error) { |
| 336 | name := relayMastersEnv(region) |
| 337 | raw := strings.TrimSpace(os.Getenv(name)) |
| 338 | if raw == "" { |
| 339 | return nil, fmt.Errorf("this console holds no master for region %s (%s is unset) — ask the operator to stage one", region, name) |
| 340 | } |
| 341 | var out []relayStagedMaster |
| 342 | for _, pair := range strings.Split(raw, ",") { |
| 343 | pair = strings.TrimSpace(pair) |
| 344 | if pair == "" { |
| 345 | continue |
| 346 | } |
| 347 | kid, keyB64, ok := strings.Cut(pair, ":") |
| 348 | if !ok { |
| 349 | return nil, fmt.Errorf("%s holds an entry that is not <kid>:<base64 key>", name) |
| 350 | } |
| 351 | kid = strings.TrimSpace(kid) |
| 352 | if !relay.ValidKid(kid) { |
| 353 | return nil, fmt.Errorf("%s holds an invalid kid %q (1-8 lowercase letters or digits)", name, kid) |
| 354 | } |
| 355 | key, err := base64.StdEncoding.DecodeString(strings.TrimSpace(keyB64)) |
| 356 | if err != nil || len(key) != relay.MasterKeyLen { |
| 357 | return nil, fmt.Errorf("%s: master %s must be base64 of %d bytes", name, kid, relay.MasterKeyLen) |
| 358 | } |
| 359 | if slices.ContainsFunc(out, func(m relayStagedMaster) bool { return m.Kid == kid }) { |
| 360 | return nil, fmt.Errorf("%s names kid %s twice", name, kid) |
| 361 | } |
| 362 | out = append(out, relayStagedMaster{Kid: kid, Key: key}) |
| 363 | } |
| 364 | if len(out) == 0 { |
| 365 | return nil, fmt.Errorf("%s is set but holds no <kid>:<base64 key> entry", name) |
| 366 | } |
| 367 | return out, nil |
| 368 | } |
| 369 | |
| 370 | // relayIssuers reads the issuer pointer with its etag. Absent is an empty |
| 371 | // map, not an error: a deployment that has never rotated has never written |
| 372 | // one, and the fallback below (the first staged master) is what it means. |
| 373 | func (s *Server) relayIssuers(ctx context.Context) (map[string]string, string, error) { |
| 374 | body, etag, err := s.cfg.Releases.Get(ctx, relayIssuerKey) |
| 375 | if errors.Is(err, os.ErrNotExist) { |
| 376 | return map[string]string{}, "", nil |
| 377 | } |
| 378 | if err != nil { |
| 379 | return nil, "", err |
| 380 | } |
| 381 | out := map[string]string{} |
| 382 | if err := json.Unmarshal(body, &out); err != nil { |
| 383 | return nil, "", fmt.Errorf("console: %s: %w", relayIssuerKey, err) |
| 384 | } |
| 385 | return out, etag, nil |
| 386 | } |
| 387 | |
| 388 | // relayMasters answers which master a region ISSUES from right now: the kid |
| 389 | // the issuer pointer names, or — when the pointer says nothing about this |
| 390 | // region — the FIRST entry in the console's variable, which is what a |
| 391 | // deployment that has never rotated has. |
| 392 | // |
| 393 | // A pointer naming a kid this console does not hold is an error rather than a |
| 394 | // fallback. It means the operator moved the pointer without staging the key |
| 395 | // here (or removed the key too early), and issuing from a different master |
| 396 | // than the pointer names would mint credentials every relay refuses. |
| 397 | func (s *Server) relayMasters(ctx context.Context, region string) (string, []byte, error) { |
| 398 | staged, err := relayStagedMasters(region) |
| 399 | if err != nil { |
| 400 | return "", nil, err |
| 401 | } |
| 402 | issuers, _, err := s.relayIssuers(ctx) |
| 403 | if err != nil { |
| 404 | return "", nil, err |
| 405 | } |
| 406 | kid, ok := issuers[region] |
| 407 | if !ok || kid == "" { |
| 408 | return staged[0].Kid, staged[0].Key, nil |
| 409 | } |
| 410 | for _, m := range staged { |
| 411 | if m.Kid == kid { |
| 412 | return m.Kid, m.Key, nil |
| 413 | } |
| 414 | } |
| 415 | return "", nil, fmt.Errorf("%s names kid %s as region %s's issuer, but %s does not hold it", |
| 416 | relayIssuerKey, kid, region, relayMastersEnv(region)) |
| 417 | } |
| 418 | |
| 419 | // relayRegionMasters gathers the issuing master for each of an app's regions |
| 420 | // — everything relay.BuildConfig needs, and nothing it does not. |
| 421 | func (s *Server) relayRegionMasters(ctx context.Context, regions []string) ([]relay.RegionMaster, error) { |
| 422 | out := make([]relay.RegionMaster, 0, len(regions)) |
| 423 | for _, r := range regions { |
| 424 | kid, master, err := s.relayMasters(ctx, r) |
| 425 | if err != nil { |
| 426 | return nil, err |
| 427 | } |
| 428 | out = append(out, relay.RegionMaster{Region: r, Kid: kid, Master: master}) |
| 429 | } |
| 430 | return out, nil |
| 431 | } |
| 432 | |
| 433 | // ---- regions, hostnames, and the config value ------------------------------- |
| 434 | |
| 435 | // relayRegions is the deployment's region list, normalized and deduped, in |
| 436 | // the operator's order. |
| 437 | func (s *Server) relayRegions() []string { |
| 438 | var out []string |
| 439 | for _, r := range s.cfg.RelayRegions { |
| 440 | r = strings.ToLower(strings.TrimSpace(r)) |
| 441 | if r != "" && !slices.Contains(out, r) { |
| 442 | out = append(out, r) |
| 443 | } |
| 444 | } |
| 445 | return out |
| 446 | } |
| 447 | |
| 448 | // relayDomain is the domain relay hostnames hang off. CARLOS_PLATFORM_DOMAIN |
| 449 | // when the operator set it; otherwise the apps domain with a leading "apps." |
| 450 | // label stripped, which is the one shape this console can derive safely |
| 451 | // (apps.oncarlos.com -> oncarlos.com; oncarlos.com -> oncarlos.com). |
| 452 | func (s *Server) relayDomain() string { |
| 453 | if d := strings.ToLower(strings.TrimSpace(s.cfg.PlatformDomain)); d != "" { |
| 454 | return d |
| 455 | } |
| 456 | d := strings.ToLower(strings.TrimSpace(s.cfg.AppsDomain)) |
| 457 | if label, rest, ok := strings.Cut(d, "."); ok && label == "apps" { |
| 458 | return rest |
| 459 | } |
| 460 | return d |
| 461 | } |
| 462 | |
| 463 | // relayURLs is the ICE server list for one region (spec §2): the STUN |
| 464 | // listener and its two TURN transports, all on one port. |
| 465 | func (s *Server) relayURLs(region string) []string { |
| 466 | host := "relay." + region + "." + s.relayDomain() + ":3478" |
| 467 | return []string{ |
| 468 | "stun:" + host, |
| 469 | "turn:" + host + "?transport=udp", |
| 470 | "turn:" + host + "?transport=tcp", |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | // relayReady is the three configuration facts every operation needs before it |
| 475 | // touches anything: a bucket to write, a region to issue for, and a key to |
| 476 | // seal with. Checked together, and BEFORE the first write, for storeSealKey's |
| 477 | // reason: a key minted without delivery is worse than no key at all. |
| 478 | func (s *Server) relayReady() error { |
| 479 | if s.cfg.Releases == nil { |
| 480 | return errRelayNoBucket |
| 481 | } |
| 482 | if len(s.relayRegions()) == 0 { |
| 483 | return errRelayNoRegions |
| 484 | } |
| 485 | if s.cfg.SecretsPub == "" { |
| 486 | return errRelayNoSealKey |
| 487 | } |
| 488 | return nil |
| 489 | } |
| 490 | |
| 491 | // relayResolveRegions turns an app's request into the region list its record |
| 492 | // carries: every armed region in the CONSOLE's order when the caller named |
| 493 | // none, or the named subset — still in the console's order, since that order |
| 494 | // is the one every browser tries them in. |
| 495 | func (s *Server) relayResolveRegions(want []string) ([]string, error) { |
| 496 | armed := s.relayRegions() |
| 497 | if len(armed) == 0 { |
| 498 | return nil, errRelayNoRegions |
| 499 | } |
| 500 | var asked []string |
| 501 | for _, r := range want { |
| 502 | if r = strings.ToLower(strings.TrimSpace(r)); r != "" { |
| 503 | asked = append(asked, r) |
| 504 | } |
| 505 | } |
| 506 | if len(asked) == 0 { |
| 507 | return armed, nil |
| 508 | } |
| 509 | for _, r := range asked { |
| 510 | if !slices.Contains(armed, r) { |
| 511 | return nil, fmt.Errorf("this deployment runs no relay in %s — armed regions are %s", |
| 512 | r, strings.Join(armed, ", ")) |
| 513 | } |
| 514 | } |
| 515 | out := make([]string, 0, len(asked)) |
| 516 | for _, r := range armed { |
| 517 | if slices.Contains(asked, r) { |
| 518 | out = append(out, r) |
| 519 | } |
| 520 | } |
| 521 | return out, nil |
| 522 | } |
| 523 | |
| 524 | // relayConfigValue builds the CARLOS_RELAY value for a record and the |
| 525 | // username suffixes it issues, from the masters each region issues from now. |
| 526 | func (s *Server) relayConfigValue(ctx context.Context, acct, app string, rec *relay.Record) (string, map[string]string, error) { |
| 527 | masters, err := s.relayRegionMasters(ctx, rec.Regions) |
| 528 | if err != nil { |
| 529 | return "", nil, err |
| 530 | } |
| 531 | return relay.BuildConfig(acct, app, rec, masters, s.relayURLs) |
| 532 | } |
| 533 | |
| 534 | // relayExpectedSuffixes is what an enabled record SHOULD have issued, given |
| 535 | // the masters in force right now — the reconcile's comparison value. |
| 536 | // |
| 537 | // Suffixes, not the sealed value: the console cannot open an envelope it |
| 538 | // wrote, and two seals of identical plaintext differ anyway (each carries a |
| 539 | // fresh ephemeral key). The suffix carries account, app, incarnation and kid |
| 540 | // — every identity the value derives from — so a suffix set that matches is |
| 541 | // a value that matches, and a rotation, a re-enable or a region change all |
| 542 | // move it. |
| 543 | func (s *Server) relayExpectedSuffixes(ctx context.Context, acct, app string, rec *relay.Record) (map[string]string, error) { |
| 544 | masters, err := s.relayRegionMasters(ctx, rec.Regions) |
| 545 | if err != nil { |
| 546 | return nil, err |
| 547 | } |
| 548 | out := make(map[string]string, len(masters)) |
| 549 | for _, m := range masters { |
| 550 | out[m.Region] = relay.Suffix(acct, app, rec.Incarnation, m.Kid) |
| 551 | } |
| 552 | return out, nil |
| 553 | } |
| 554 | |
| 555 | // relayEnvelopeDigest identifies a sealed envelope without opening it: |
| 556 | // hex(sha256) of the envelope string as it sits in the bundle. It is what |
| 557 | // makes "the bundle still holds exactly what we wrote" checkable by a console |
| 558 | // that holds no decryption key — see relay.Record.IssuedDigest. |
| 559 | func relayEnvelopeDigest(envelope string) string { |
| 560 | sum := sha256.Sum256([]byte(envelope)) |
| 561 | return hex.EncodeToString(sum[:]) |
| 562 | } |
| 563 | |
| 564 | // relayConfigVar reads the app's delivered CARLOS_RELAY entry, if any. |
| 565 | func (s *Server) relayConfigVar(ctx context.Context, acct, app string) (appconfig.Var, bool, error) { |
| 566 | m, err := appconfig.Get(ctx, s.cfg.Releases, acct, app) |
| 567 | if err != nil { |
| 568 | return appconfig.Var{}, false, err |
| 569 | } |
| 570 | v, ok := m.Vars[relayVar] |
| 571 | return v, ok, nil |
| 572 | } |
| 573 | |
| 574 | // relayApplyConfig derives the app's key, seals it and writes it into the |
| 575 | // DEFAULT config bundle, recording on the record which identities it just |
| 576 | // issued. It mutates rec.IssuedSuffixes; the caller writes the record. |
| 577 | func (s *Server) relayApplyConfig(ctx context.Context, acct, app string, rec *relay.Record) error { |
| 578 | if err := s.requireRelayLease(); err != nil { |
| 579 | return err |
| 580 | } |
| 581 | if s.cfg.SecretsPub == "" { |
| 582 | return errRelayNoSealKey |
| 583 | } |
| 584 | value, suffixes, err := s.relayConfigValue(ctx, acct, app, rec) |
| 585 | if err != nil { |
| 586 | return err |
| 587 | } |
| 588 | pub, err := appconfig.ParsePublicKey(s.cfg.SecretsPub, "CARLOS_SECRETS_PUBKEY") |
| 589 | if err != nil { |
| 590 | return err |
| 591 | } |
| 592 | env, err := appconfig.Seal(pub, appconfig.Binding{Account: acct, App: app}, []byte(value)) |
| 593 | if err != nil { |
| 594 | return err |
| 595 | } |
| 596 | if err := s.requireRelayLease(); err != nil { |
| 597 | return err |
| 598 | } |
| 599 | if _, err := appconfig.Apply(ctx, s.cfg.Releases, acct, app, |
| 600 | map[string]appconfig.Var{relayVar: {Kind: "secret", Envelope: env}}, nil); err != nil { |
| 601 | return err |
| 602 | } |
| 603 | rec.IssuedSuffixes = suffixes |
| 604 | rec.IssuedDigest = relayEnvelopeDigest(env) |
| 605 | return nil |
| 606 | } |
| 607 | |
| 608 | // relayUnsetConfig removes CARLOS_RELAY from the app's default bundle, and |
| 609 | // reports whether there was anything to remove. It READS first because an |
| 610 | // unset-only Apply against an app with no bundle at all is an error |
| 611 | // (appconfig.applyAt), and "this app has no config" is not a failed disable. |
| 612 | func (s *Server) relayUnsetConfig(ctx context.Context, acct, app string) (bool, error) { |
| 613 | if err := s.requireRelayLease(); err != nil { |
| 614 | return false, err |
| 615 | } |
| 616 | _, present, err := s.relayConfigVar(ctx, acct, app) |
| 617 | if err != nil || !present { |
| 618 | return false, err |
| 619 | } |
| 620 | if err := s.requireRelayLease(); err != nil { |
| 621 | return false, err |
| 622 | } |
| 623 | n, err := appconfig.Apply(ctx, s.cfg.Releases, acct, app, nil, []string{relayVar}) |
| 624 | return n > 0, err |
| 625 | } |
| 626 | |
| 627 | // ---- the record walk -------------------------------------------------------- |
| 628 | |
| 629 | // relayAppLock is the per-app mutex every operation and the reconcile's |
| 630 | // per-record work takes. The lease keeps two CONSOLES apart; this keeps two |
| 631 | // handlers inside one console apart, which the lease says nothing about. |
| 632 | func (s *Server) relayAppLock(acct, app string) *sync.Mutex { |
| 633 | v, _ := s.relayAppLocks.LoadOrStore(acct+"/"+app, &sync.Mutex{}) |
| 634 | return v.(*sync.Mutex) |
| 635 | } |
| 636 | |
| 637 | // relayPutRecord is relay.PutRecord behind the lease check. EVERY record |
| 638 | // write in this file goes through it rather than calling PutRecord directly: |
| 639 | // a walk is several writes long, the lease can lapse between any two of them, |
| 640 | // and spec §5's rule is a check before each bucket write, not one per |
| 641 | // operation. |
| 642 | func (s *Server) relayPutRecord(ctx context.Context, acct, app string, rec *relay.Record, etag string) (string, error) { |
| 643 | if err := s.requireRelayLease(); err != nil { |
| 644 | return "", err |
| 645 | } |
| 646 | return relay.PutRecord(ctx, s.cfg.Releases, acct, app, rec, etag) |
| 647 | } |
| 648 | |
| 649 | // relayAppRecord is one app's record with the coordinates that name it and |
| 650 | // the etag it was read at — the unit both the publish listing and the |
| 651 | // reconcile walk over. |
| 652 | type relayAppRecord struct { |
| 653 | Account string |
| 654 | App string |
| 655 | Record *relay.Record |
| 656 | Etag string |
| 657 | } |
| 658 | |
| 659 | // relayListRecords walks every app in the bucket and returns the ones that |
| 660 | // have a relay record. Two delimiter-aware listings (accounts, then their |
| 661 | // apps) rather than one flat List of the whole `apps/` prefix: a deployment's |
| 662 | // object count is dominated by releases, and this must not read them. |
| 663 | // |
| 664 | // A record that will not parse is logged and skipped, not fatal. The |
| 665 | // alternative is one corrupt object stopping the aggregate from being |
| 666 | // published for every other app in the deployment. |
| 667 | func (s *Server) relayListRecords(ctx context.Context) ([]relayAppRecord, error) { |
| 668 | accts, err := s.cfg.Releases.ListDirs(ctx, "apps/") |
| 669 | if err != nil { |
| 670 | return nil, err |
| 671 | } |
| 672 | var out []relayAppRecord |
| 673 | for _, acctPrefix := range accts { |
| 674 | acct := strings.Trim(strings.TrimPrefix(acctPrefix, "apps/"), "/") |
| 675 | if !release.ValidAccount(acct) { |
| 676 | continue |
| 677 | } |
| 678 | apps, err := s.cfg.Releases.ListDirs(ctx, acctPrefix) |
| 679 | if err != nil { |
| 680 | return nil, err |
| 681 | } |
| 682 | for _, appPrefix := range apps { |
| 683 | app := strings.Trim(strings.TrimPrefix(appPrefix, acctPrefix), "/") |
| 684 | if !release.ValidApp(app) { |
| 685 | continue |
| 686 | } |
| 687 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, acct, app) |
| 688 | if err != nil { |
| 689 | log.Printf("console: relay record %s/%s: %v", acct, app, err) |
| 690 | continue |
| 691 | } |
| 692 | if rec == nil { |
| 693 | continue |
| 694 | } |
| 695 | out = append(out, relayAppRecord{Account: acct, App: app, Record: rec, Etag: etag}) |
| 696 | } |
| 697 | } |
| 698 | sort.Slice(out, func(i, j int) bool { |
| 699 | if out[i].Account != out[j].Account { |
| 700 | return out[i].Account < out[j].Account |
| 701 | } |
| 702 | return out[i].App < out[j].App |
| 703 | }) |
| 704 | return out, nil |
| 705 | } |
| 706 | |
| 707 | // relayPublishOnce is ONE attempt at spec §5's publish sequence, from an etag |
| 708 | // the CALLER read before it did anything else: list every record, build the |
| 709 | // authority from the enabled ones whatever their phase, and write with that |
| 710 | // etag. A conflict comes straight back — this is the function the retry loop |
| 711 | // wraps, and the function a test can hand a deliberately stale etag. |
| 712 | // |
| 713 | // The etag being an argument is the whole design. Read after the listing, it |
| 714 | // would let a disable that landed mid-listing be overwritten by an aggregate |
| 715 | // built before it; read before, that same disable makes this write fail. |
| 716 | func (s *Server) relayPublishOnce(ctx context.Context, etag string) error { |
| 717 | if err := s.requireRelayLease(); err != nil { |
| 718 | return err |
| 719 | } |
| 720 | recs, err := s.relayListRecords(ctx) |
| 721 | if err != nil { |
| 722 | return err |
| 723 | } |
| 724 | a := relay.Authority{WrittenAt: s.relayTime().UTC()} |
| 725 | for _, r := range recs { |
| 726 | if !r.Record.Enabled { |
| 727 | continue |
| 728 | } |
| 729 | a.Apps = append(a.Apps, relay.AuthorityApp{ |
| 730 | Account: r.Account, |
| 731 | App: r.App, |
| 732 | Incarnation: r.Record.Incarnation, |
| 733 | Regions: r.Record.Regions, |
| 734 | }) |
| 735 | } |
| 736 | body, err := json.MarshalIndent(a, "", " ") |
| 737 | if err != nil { |
| 738 | return err |
| 739 | } |
| 740 | if s.relayGated() { |
| 741 | return s.relayPublishGated(ctx, etag, append(body, '\n')) |
| 742 | } |
| 743 | if err := s.requireRelayLease(); err != nil { |
| 744 | return err |
| 745 | } |
| 746 | _, err = s.cfg.Releases.Put(ctx, relay.AuthorityKey, append(body, '\n'), etag) |
| 747 | return err |
| 748 | } |
| 749 | |
| 750 | // relayPublishAggregate publishes control/relay.json: read the etag FIRST, |
| 751 | // then the whole sequence, restarting it on conflict, four attempts. Then a |
| 752 | // serial bump so relays notice in seconds. |
| 753 | // |
| 754 | // The bump is a LATENCY HINT, not the lease (spec §5): a relay that never |
| 755 | // sees it still re-reads on its own timer, so a failed bump is a log line and |
| 756 | // a successful publish, never a failed operation. |
| 757 | func (s *Server) relayPublishAggregate(ctx context.Context) error { |
| 758 | if err := s.requireRelayLease(); err != nil { |
| 759 | return err |
| 760 | } |
| 761 | var lastErr error |
| 762 | for attempt := 0; attempt < 4; attempt++ { |
| 763 | _, etag, err := s.cfg.Releases.Get(ctx, relay.AuthorityKey) |
| 764 | switch { |
| 765 | case errors.Is(err, os.ErrNotExist): |
| 766 | etag = "" |
| 767 | case err != nil: |
| 768 | return err |
| 769 | } |
| 770 | err = s.relayPublishOnce(ctx, etag) |
| 771 | if err == nil { |
| 772 | if err := release.BumpSerial(ctx, s.cfg.Releases, s.relayTime()); err != nil { |
| 773 | log.Printf("console: relay serial bump (relays converge on their own timer anyway): %v", err) |
| 774 | } |
| 775 | return nil |
| 776 | } |
| 777 | if !errors.Is(err, release.ErrCASConflict) { |
| 778 | return err |
| 779 | } |
| 780 | lastErr = err |
| 781 | } |
| 782 | return fmt.Errorf("publishing %s: %w", relay.AuthorityKey, lastErr) |
| 783 | } |
| 784 | |
| 785 | // ---- route delivery --------------------------------------------------------- |
| 786 | |
| 787 | // relayRoute is one of an app's routes as the delivery check sees it: enough |
| 788 | // to say which of spec §5's three shapes it is, and nothing else. |
| 789 | type relayRoute struct { |
| 790 | Host string |
| 791 | Kind string // instance | service | static ("" reads as instance) |
| 792 | Addr string // a TCP upstream this platform does not own |
| 793 | Socket string |
| 794 | Unit string |
| 795 | Environment string |
| 796 | } |
| 797 | |
| 798 | // relayAppRoutes enumerates an app's routes from both places the console can |
| 799 | // see them: the instance records it writes itself, and — for routes an |
| 800 | // operator added directly on a box — the platform registry. |
| 801 | // |
| 802 | // Records first, registry second, deduped on host: a route that exists in |
| 803 | // both is one route, and the record is the authority on its shape. |
| 804 | func (s *Server) relayAppRoutes(ctx context.Context, acct, app string) ([]relayRoute, error) { |
| 805 | recs, truncated, err := provision.ListRecords(ctx, s.cfg.Releases, acct, app, relayRouteMax) |
| 806 | if err != nil { |
| 807 | return nil, err |
| 808 | } |
| 809 | // A truncated listing is a REFUSAL, not a partial answer. The one caller |
| 810 | // that matters is the delivery check, and a check that only saw the first |
| 811 | // 500 routes would enable the relay over an undeliverable route it never |
| 812 | // read — the exact state spec §5 refuses because it blocks retirement |
| 813 | // forever, arrived at silently. |
| 814 | if truncated { |
| 815 | return nil, fmt.Errorf("%s/%s has more than %d instance records — the relay cannot check them all, so it will not act on a partial list", |
| 816 | acct, app, relayRouteMax) |
| 817 | } |
| 818 | var out []relayRoute |
| 819 | seen := map[string]bool{} |
| 820 | for _, rec := range recs { |
| 821 | seen[rec.Host] = true |
| 822 | // Fold Addrs the way the registry branch does (routeAddr, below). A |
| 823 | // record picks ONE upstream shape — the legacy Addr or the Addrs set, |
| 824 | // never both (provision.ValidateShape) — so a multi-upstream external |
| 825 | // route carries Addr == "", and "no address" is exactly what the |
| 826 | // delivery check reads as deliverable. Left unfolded, such a route |
| 827 | // slips past the refusal enable exists to make, shows as shape |
| 828 | // "unknown" in status, and can be acknowledged with `carlos relay |
| 829 | // adopted`. What every caller here asks is "is there a process at the |
| 830 | // far end this platform never spawned", and the first upstream |
| 831 | // answers that as well as five. |
| 832 | addr := rec.Addr |
| 833 | if all := rec.AddrsOrDefault(); len(all) > 0 { |
| 834 | addr = all[0] |
| 835 | } |
| 836 | out = append(out, relayRoute{Host: rec.Host, Kind: rec.Kind, Addr: addr, |
| 837 | Unit: rec.Unit, Environment: rec.Environment}) |
| 838 | } |
| 839 | if s.cfg.Registry == "" { |
| 840 | return out, nil |
| 841 | } |
| 842 | reg, err := openRegistry(s.cfg.Registry) |
| 843 | if err != nil { |
| 844 | // A registry this console cannot open is not a reason to refuse an |
| 845 | // enable: the records above are the platform's own writing, and the |
| 846 | // registry copy is the belt to their braces. |
| 847 | log.Printf("console: relay routes %s/%s: registry: %v", acct, app, err) |
| 848 | return out, nil |
| 849 | } |
| 850 | routes, err := reg.List() |
| 851 | if err != nil { |
| 852 | log.Printf("console: relay routes %s/%s: registry list: %v", acct, app, err) |
| 853 | return out, nil |
| 854 | } |
| 855 | for _, rt := range routes { |
| 856 | if rt.App != app || release.RouteAccount(rt) != acct || seen[rt.Host] { |
| 857 | continue |
| 858 | } |
| 859 | seen[rt.Host] = true |
| 860 | out = append(out, relayRoute{Host: rt.Host, Kind: rt.Kind, Addr: routeAddr(rt), |
| 861 | Socket: rt.Socket, Unit: rt.Unit, Environment: rt.Environment}) |
| 862 | } |
| 863 | sort.Slice(out, func(i, j int) bool { return out[i].Host < out[j].Host }) |
| 864 | return out, nil |
| 865 | } |
| 866 | |
| 867 | // routeAddr folds a route's two upstream shapes into "does it have one". |
| 868 | func routeAddr(rt registry.Route) string { |
| 869 | if rt.Addr != "" { |
| 870 | return rt.Addr |
| 871 | } |
| 872 | return rt.Addrs |
| 873 | } |
| 874 | |
| 875 | // relayExecDelivery answers whether a route's BOUND bundle opts into file |
| 876 | // delivery, applying the same precedence appconfig's own resolver does |
| 877 | // (#195): the route's named bundle when it states something, the app default |
| 878 | // otherwise. Absence is inheritance, never "off". |
| 879 | func (s *Server) relayExecDelivery(ctx context.Context, acct, app, envName string) (bool, error) { |
| 880 | if envName != "" && appconfig.ValidEnvironmentName(envName) { |
| 881 | m, err := appconfig.LoadEnvironment(ctx, s.cfg.Releases, acct, app, envName) |
| 882 | if err != nil { |
| 883 | return false, err |
| 884 | } |
| 885 | if m.ExecDelivery != "" { |
| 886 | return m.ExecDelivery == appconfig.ExecDeliveryFile, nil |
| 887 | } |
| 888 | } |
| 889 | m, err := appconfig.Get(ctx, s.cfg.Releases, acct, app) |
| 890 | if err != nil { |
| 891 | return false, err |
| 892 | } |
| 893 | return m.ExecDelivery == appconfig.ExecDeliveryFile, nil |
| 894 | } |
| 895 | |
| 896 | // relayNoDelivery answers the ONE question three callers need about a route: |
| 897 | // can the config sync put a relay key in front of its process at all? |
| 898 | // |
| 899 | // The three shapes: an `instance` route is the platform's own — it has a |
| 900 | // socket or a unit and the config sync restarts it, so it is deliverable. A |
| 901 | // `static` route runs no process at all and is ignored. Everything else is |
| 902 | // deliverable only if it has a socket or a unit to bounce, or its bundle opts |
| 903 | // into file delivery — a bare TCP address whose process this platform never |
| 904 | // spawned is exactly the route the spec refuses. |
| 905 | // |
| 906 | // It lives in one function because the three callers must agree: enable |
| 907 | // REFUSES such a route (below), `carlos relay status` NAMES it (relayReply, |
| 908 | // apirelay.go), and `carlos relay adopted` refuses to record a cutover for |
| 909 | // one (relayAdopted). Three copies of this switch would eventually let a |
| 910 | // member acknowledge a key that had never been written for them. |
| 911 | func (s *Server) relayNoDelivery(ctx context.Context, acct, app string, rt relayRoute) (bool, error) { |
| 912 | switch { |
| 913 | case rt.Kind == "static", |
| 914 | rt.Kind == "" || rt.Kind == "instance", |
| 915 | rt.Socket != "" || rt.Unit != "", |
| 916 | rt.Addr == "": |
| 917 | return false, nil |
| 918 | } |
| 919 | on, err := s.relayExecDelivery(ctx, acct, app, rt.Environment) |
| 920 | return !on, err |
| 921 | } |
| 922 | |
| 923 | // relayDeliverable is enable's second refusal (spec §5, step 2): a route that |
| 924 | // could never receive a rotated key must not be enabled, because it would |
| 925 | // block every future retirement forever and nobody would find out until the |
| 926 | // rotation stalled. |
| 927 | func (s *Server) relayDeliverable(ctx context.Context, acct, app string) error { |
| 928 | routes, err := s.relayAppRoutes(ctx, acct, app) |
| 929 | if err != nil { |
| 930 | return err |
| 931 | } |
| 932 | for _, rt := range routes { |
| 933 | none, err := s.relayNoDelivery(ctx, acct, app, rt) |
| 934 | if err != nil { |
| 935 | return err |
| 936 | } |
| 937 | if none { |
| 938 | return fmt.Errorf("route %s is a TCP address this platform does not own and its config bundle does not opt into file delivery, "+ |
| 939 | "so it could never pick up a rotated relay key — set it with `carlos env exec-delivery file` first: %w", rt.Host, errRelayUndelivered) |
| 940 | } |
| 941 | } |
| 942 | return nil |
| 943 | } |
| 944 | |
| 945 | // ---- the operations --------------------------------------------------------- |
| 946 | |
| 947 | // relayEnable turns the relay on for one app and delivers its key (spec §5, |
| 948 | // steps 3-7). The transition table decides what happens to the record; the |
| 949 | // walk after it is always the same three steps, each recorded in `phase` so |
| 950 | // the reconcile can finish it. |
| 951 | // |
| 952 | // A repeat enable of a FINISHED, identically-configured record is the one |
| 953 | // path that writes nothing at all: no op bump, no config write, no new |
| 954 | // envelope. Anything else — an unfinished operation, a region change, a |
| 955 | // re-enable after a disable — walks. |
| 956 | func (s *Server) relayEnable(ctx context.Context, acct *Account, app string, regions []string, by string) (*relay.Record, error) { |
| 957 | if err := s.relayReady(); err != nil { |
| 958 | return nil, err |
| 959 | } |
| 960 | want, err := s.relayResolveRegions(regions) |
| 961 | if err != nil { |
| 962 | return nil, err |
| 963 | } |
| 964 | mu := s.relayAppLock(acct.Sqid, app) |
| 965 | mu.Lock() |
| 966 | defer mu.Unlock() |
| 967 | var out *relay.Record |
| 968 | err = s.relayAppMutation(ctx, acct.Sqid, app, "relay-enable", func(ctx context.Context) error { |
| 969 | var err error |
| 970 | out, err = s.relayEnableWalk(ctx, acct, app, want, by) |
| 971 | return err |
| 972 | }) |
| 973 | if err != nil { |
| 974 | return nil, err |
| 975 | } |
| 976 | return out, nil |
| 977 | } |
| 978 | |
| 979 | // relayEnableWalk is relayEnable's body under the app's lock and permit. |
| 980 | func (s *Server) relayEnableWalk(ctx context.Context, acct *Account, app string, want []string, by string) (*relay.Record, error) { |
| 981 | if err := s.requireRelayLease(); err != nil { |
| 982 | return nil, err |
| 983 | } |
| 984 | if err := s.relayDeliverable(ctx, acct.Sqid, app); err != nil { |
| 985 | return nil, err |
| 986 | } |
| 987 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, acct.Sqid, app) |
| 988 | if err != nil { |
| 989 | return nil, err |
| 990 | } |
| 991 | now := s.relayTime().UTC() |
| 992 | switch { |
| 993 | case rec != nil && rec.Enabled && rec.Phase == relayPhaseDone && slices.Equal(rec.Regions, want): |
| 994 | return rec, nil |
| 995 | case rec != nil && rec.Enabled && slices.Equal(rec.Regions, want): |
| 996 | // An unfinished operation RESUMES rather than starting a new one: |
| 997 | // same incarnation, same op, and the walk below repeats the steps it |
| 998 | // may already have done, all of which are idempotent. |
| 999 | case rec != nil && rec.Enabled: |
| 1000 | // A region change keeps the incarnation — the app's identity did not |
| 1001 | // move, only where it may connect (spec §5's transition table). |
| 1002 | rec.Regions = want |
| 1003 | rec.Op++ |
| 1004 | rec.Phase = relayPhaseConfig |
| 1005 | default: |
| 1006 | op := int64(0) |
| 1007 | if rec != nil { |
| 1008 | op = rec.Op |
| 1009 | } |
| 1010 | // A NEW incarnation for a first enable and for every re-enable after |
| 1011 | // a disable, so a key from before the disable can never be mistaken |
| 1012 | // for a live one. Cutovers go with it: they acknowledged identities |
| 1013 | // that no longer exist. |
| 1014 | rec = &relay.Record{ |
| 1015 | Enabled: true, Incarnation: relay.NewIncarnation(), Regions: want, |
| 1016 | Op: op + 1, Phase: relayPhaseConfig, EnabledAt: now, |
| 1017 | } |
| 1018 | } |
| 1019 | if etag, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1020 | return nil, err |
| 1021 | } |
| 1022 | if err := s.relayApplyConfig(ctx, acct.Sqid, app, rec); err != nil { |
| 1023 | return nil, err |
| 1024 | } |
| 1025 | rec.Phase = relayPhaseAggregate |
| 1026 | if etag, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1027 | return nil, err |
| 1028 | } |
| 1029 | if err := s.relayPublishAggregate(ctx); err != nil { |
| 1030 | return nil, err |
| 1031 | } |
| 1032 | rec.Phase = relayPhaseDone |
| 1033 | if _, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1034 | return nil, err |
| 1035 | } |
| 1036 | log.Printf("console: relay enabled %s/%s in %s by %s (incarnation %s, op %d)", |
| 1037 | acct.Sqid, app, strings.Join(rec.Regions, ","), by, rec.Incarnation, rec.Op) |
| 1038 | return rec, nil |
| 1039 | } |
| 1040 | |
| 1041 | // relayDisable is the same walk in reverse (spec §5): record the intent, |
| 1042 | // remove the variable, publish an aggregate without the app, mark done. The |
| 1043 | // incarnation stays in the record — it is what `status` reports the pending |
| 1044 | // disable against — and a later enable draws a new one. |
| 1045 | func (s *Server) relayDisable(ctx context.Context, acct *Account, app string, by string) (*relay.Record, error) { |
| 1046 | if s.cfg.Releases == nil { |
| 1047 | return nil, errRelayNoBucket |
| 1048 | } |
| 1049 | mu := s.relayAppLock(acct.Sqid, app) |
| 1050 | mu.Lock() |
| 1051 | defer mu.Unlock() |
| 1052 | var out *relay.Record |
| 1053 | err := s.relayAppMutation(ctx, acct.Sqid, app, "relay-disable", func(ctx context.Context) error { |
| 1054 | var err error |
| 1055 | out, err = s.relayDisableWalk(ctx, acct, app, by) |
| 1056 | return err |
| 1057 | }) |
| 1058 | if err != nil { |
| 1059 | return nil, err |
| 1060 | } |
| 1061 | return out, nil |
| 1062 | } |
| 1063 | |
| 1064 | // relayDisableWalk is relayDisable's body under the app's lock and permit. |
| 1065 | func (s *Server) relayDisableWalk(ctx context.Context, acct *Account, app string, by string) (*relay.Record, error) { |
| 1066 | if err := s.requireRelayLease(); err != nil { |
| 1067 | return nil, err |
| 1068 | } |
| 1069 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, acct.Sqid, app) |
| 1070 | if err != nil { |
| 1071 | return nil, err |
| 1072 | } |
| 1073 | if rec == nil { |
| 1074 | return nil, errRelayNotEnabled |
| 1075 | } |
| 1076 | if !rec.Enabled && rec.Phase == relayPhaseDone { |
| 1077 | return rec, nil // already off and finished: nothing owed |
| 1078 | } |
| 1079 | if rec.Enabled { |
| 1080 | rec.Enabled = false |
| 1081 | rec.Op++ |
| 1082 | rec.Phase = relayPhaseConfig |
| 1083 | rec.DisabledAt = s.relayTime().UTC() |
| 1084 | // Nothing is issued any more. Cutovers stay: they are a historical |
| 1085 | // record of what a member acknowledged, and a re-enable clears them. |
| 1086 | rec.IssuedSuffixes = nil |
| 1087 | rec.IssuedDigest = "" |
| 1088 | if etag, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1089 | return nil, err |
| 1090 | } |
| 1091 | } |
| 1092 | if _, err := s.relayUnsetConfig(ctx, acct.Sqid, app); err != nil { |
| 1093 | return nil, err |
| 1094 | } |
| 1095 | rec.Phase = relayPhaseAggregate |
| 1096 | if etag, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1097 | return nil, err |
| 1098 | } |
| 1099 | if err := s.relayPublishAggregate(ctx); err != nil { |
| 1100 | return nil, err |
| 1101 | } |
| 1102 | rec.Phase = relayPhaseDone |
| 1103 | if _, err = s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1104 | return nil, err |
| 1105 | } |
| 1106 | log.Printf("console: relay disabled %s/%s by %s (op %d)", acct.Sqid, app, by, rec.Op) |
| 1107 | return rec, nil |
| 1108 | } |
| 1109 | |
| 1110 | // relayAdopted records a member's acknowledgement that a TCP-address route |
| 1111 | // they restarted by hand now runs the identities the console last issued |
| 1112 | // (spec §5, step 5). |
| 1113 | // |
| 1114 | // The identities are stamped from the record AT THIS MOMENT, not named by the |
| 1115 | // caller, and they are what makes the ack meaningful: an acknowledgement |
| 1116 | // given before a later rotation names the older suffixes and so does not |
| 1117 | // count for the newer ones. One cutover per host — a second ack for the same |
| 1118 | // route replaces the first. |
| 1119 | func (s *Server) relayAdopted(ctx context.Context, acct *Account, app, host, by string) (*relay.Record, error) { |
| 1120 | if s.cfg.Releases == nil { |
| 1121 | return nil, errRelayNoBucket |
| 1122 | } |
| 1123 | mu := s.relayAppLock(acct.Sqid, app) |
| 1124 | mu.Lock() |
| 1125 | defer mu.Unlock() |
| 1126 | var out *relay.Record |
| 1127 | err := s.relayAppMutation(ctx, acct.Sqid, app, "relay-adopted", func(ctx context.Context) error { |
| 1128 | var err error |
| 1129 | out, err = s.relayAdoptedWalk(ctx, acct, app, host, by) |
| 1130 | return err |
| 1131 | }) |
| 1132 | if err != nil { |
| 1133 | return nil, err |
| 1134 | } |
| 1135 | return out, nil |
| 1136 | } |
| 1137 | |
| 1138 | // relayAdoptedWalk is relayAdopted's body under the app's lock and permit. |
| 1139 | func (s *Server) relayAdoptedWalk(ctx context.Context, acct *Account, app, host, by string) (*relay.Record, error) { |
| 1140 | if err := s.requireRelayLease(); err != nil { |
| 1141 | return nil, err |
| 1142 | } |
| 1143 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, acct.Sqid, app) |
| 1144 | if err != nil { |
| 1145 | return nil, err |
| 1146 | } |
| 1147 | if rec == nil || !rec.Enabled { |
| 1148 | return nil, errRelayNotEnabled |
| 1149 | } |
| 1150 | // The host must be one of the app's own routes. Cheap, and it is what |
| 1151 | // keeps the record bounded: without it, an ack naming anything at all |
| 1152 | // appends a cutover forever, and `status` would report identities against |
| 1153 | // routes that do not exist. |
| 1154 | routes, err := s.relayAppRoutes(ctx, acct.Sqid, app) |
| 1155 | if err != nil { |
| 1156 | return nil, err |
| 1157 | } |
| 1158 | i := slices.IndexFunc(routes, func(rt relayRoute) bool { return rt.Host == host }) |
| 1159 | if i < 0 { |
| 1160 | return nil, fmt.Errorf("%s is not a route of this app: %w", host, errRelayUnknownHost) |
| 1161 | } |
| 1162 | // A route the config sync cannot reach has been sent NOTHING to adopt, so |
| 1163 | // an acknowledgement of it would be false — and worse than merely false: |
| 1164 | // a cutover is what the retirement of an old master is eventually checked |
| 1165 | // against, and one recorded here would say a process has a key it has |
| 1166 | // never seen. Such a route can only appear after an enable (enable itself |
| 1167 | // refuses it), which is exactly when a member might reach for this verb |
| 1168 | // to make the warning go away. |
| 1169 | if none, err := s.relayNoDelivery(ctx, acct.Sqid, app, routes[i]); err != nil { |
| 1170 | return nil, err |
| 1171 | } else if none { |
| 1172 | return nil, fmt.Errorf("%s has no config delivery, so no relay key has been written for it to adopt — "+ |
| 1173 | "opt it into file delivery with `carlos env exec-delivery file` or move it to a unit: %w", host, errRelayUndelivered) |
| 1174 | } |
| 1175 | // And the route must be the ONE shape whose cutover a member establishes. |
| 1176 | // Spec §5, step 5: a unit or socket-exec route is restarted by the config |
| 1177 | // sync, so its cutover is the ack that sync writes — an observed fact, |
| 1178 | // with the identities the host actually rendered. Accepting a member's |
| 1179 | // word for such a route would stamp the identities the console ISSUED |
| 1180 | // over that observation, and a route whose restart silently failed would |
| 1181 | // read as cut over. The member verb exists only for the process this |
| 1182 | // platform cannot restart. |
| 1183 | switch shape := relayRouteShape(routes[i]); shape { |
| 1184 | case relayShapeFile: // the one shape a member acknowledges |
| 1185 | case relayShapeUnit, relayShapeSocket: |
| 1186 | return nil, fmt.Errorf("%s is a %s route: the platform restarts this route itself; its adoption is observed, not declared — "+ |
| 1187 | "`carlos relay status` reports it as adopted once the config sync acknowledges the write: %w", |
| 1188 | host, shape, errRelayNotMemberRoute) |
| 1189 | default: |
| 1190 | return nil, fmt.Errorf("%s has no unit, no socket and no address, so this console cannot say what would run a relay key, "+ |
| 1191 | "let alone that it has: %w", host, errRelayNotMemberRoute) |
| 1192 | } |
| 1193 | cut := relay.RouteCutover{ |
| 1194 | Host: host, |
| 1195 | Suffixes: maps.Clone(rec.IssuedSuffixes), |
| 1196 | At: s.relayTime().UTC(), |
| 1197 | By: by, |
| 1198 | } |
| 1199 | replaced := false |
| 1200 | for i := range rec.Cutovers { |
| 1201 | if rec.Cutovers[i].Host == host { |
| 1202 | rec.Cutovers[i] = cut |
| 1203 | replaced = true |
| 1204 | break |
| 1205 | } |
| 1206 | } |
| 1207 | if !replaced { |
| 1208 | rec.Cutovers = append(rec.Cutovers, cut) |
| 1209 | } |
| 1210 | if _, err := s.relayPutRecord(ctx, acct.Sqid, app, rec, etag); err != nil { |
| 1211 | return nil, err |
| 1212 | } |
| 1213 | return rec, nil |
| 1214 | } |
| 1215 | |
| 1216 | // ---- rotation --------------------------------------------------------------- |
| 1217 | |
| 1218 | // relayRotateResult is what `carlos relay rotate` prints: which kid the |
| 1219 | // region now issues from, how many apps were re-issued, and the earliest the |
| 1220 | // old master may be retired. |
| 1221 | type relayRotateResult struct { |
| 1222 | Region string |
| 1223 | Kid string |
| 1224 | Apps int |
| 1225 | RetireNoEarlierThan time.Time |
| 1226 | } |
| 1227 | |
| 1228 | // relayHeartbeats reads every armed relay's heartbeat. |
| 1229 | func (s *Server) relayHeartbeats(ctx context.Context) ([]relay.Heartbeat, error) { |
| 1230 | keys, err := s.cfg.Releases.List(ctx, relayAckPrefix) |
| 1231 | if err != nil { |
| 1232 | return nil, err |
| 1233 | } |
| 1234 | var out []relay.Heartbeat |
| 1235 | for _, key := range keys { |
| 1236 | if !strings.HasSuffix(key, ".json") { |
| 1237 | continue |
| 1238 | } |
| 1239 | body, _, err := s.cfg.Releases.Get(ctx, key) |
| 1240 | if err != nil { |
| 1241 | if errors.Is(err, os.ErrNotExist) { |
| 1242 | continue |
| 1243 | } |
| 1244 | return nil, err |
| 1245 | } |
| 1246 | var hb relay.Heartbeat |
| 1247 | if err := json.Unmarshal(body, &hb); err != nil { |
| 1248 | log.Printf("console: relay heartbeat %s: %v", key, err) |
| 1249 | continue |
| 1250 | } |
| 1251 | out = append(out, hb) |
| 1252 | } |
| 1253 | return out, nil |
| 1254 | } |
| 1255 | |
| 1256 | // relayConfigAcks reads every per-route config-adoption acknowledgement |
| 1257 | // under configAckPrefix whose HOST — a key's basename, ConfigAckKey's own |
| 1258 | // "<box>/<host>.json" shape — is in hosts. There is no per-host box list |
| 1259 | // to narrow the LISTING by, the same limit relayHeartbeats accepts for |
| 1260 | // control/relay-ack/, so this still lists every key under the prefix; but |
| 1261 | // hosts filters that listing down to this app's own routes before a |
| 1262 | // single Get is issued, so a status call costs one List plus a Get per |
| 1263 | // THIS APP's route, never a Get per ack in the whole fleet (review |
| 1264 | // finding 2026-09-07 — relayReply used to pass every ack for every app on |
| 1265 | // the deployment through this filter instead of narrowing the read |
| 1266 | // itself). |
| 1267 | func (s *Server) relayConfigAcks(ctx context.Context, hosts map[string]bool) ([]appconfig.ConfigAck, error) { |
| 1268 | keys, err := s.cfg.Releases.List(ctx, configAckPrefix) |
| 1269 | if err != nil { |
| 1270 | return nil, err |
| 1271 | } |
| 1272 | var out []appconfig.ConfigAck |
| 1273 | for _, key := range keys { |
| 1274 | if !strings.HasSuffix(key, ".json") { |
| 1275 | continue |
| 1276 | } |
| 1277 | host := strings.TrimSuffix(key, ".json") |
| 1278 | if idx := strings.LastIndexByte(host, '/'); idx >= 0 { |
| 1279 | host = host[idx+1:] |
| 1280 | } |
| 1281 | if !hosts[host] { |
| 1282 | continue |
| 1283 | } |
| 1284 | body, _, err := s.cfg.Releases.Get(ctx, key) |
| 1285 | if err != nil { |
| 1286 | if errors.Is(err, os.ErrNotExist) { |
| 1287 | continue |
| 1288 | } |
| 1289 | return nil, err |
| 1290 | } |
| 1291 | var a appconfig.ConfigAck |
| 1292 | if err := json.Unmarshal(body, &a); err != nil { |
| 1293 | log.Printf("console: relay config ack %s: %v", key, err) |
| 1294 | continue |
| 1295 | } |
| 1296 | out = append(out, a) |
| 1297 | } |
| 1298 | return out, nil |
| 1299 | } |
| 1300 | |
| 1301 | // relayRotate moves a region's issuer pointer to the master the operator |
| 1302 | // staged, then re-issues every app enabled in that region (spec §5, step 4). |
| 1303 | // |
| 1304 | // IT REFUSES UNTIL EVERY LIVE RELAY IN THE REGION HAS THE NEW KID. That gate |
| 1305 | // is the whole safety of a rotation: issuing from a master a box has not |
| 1306 | // loaded mints credentials that box refuses, and every call in the region |
| 1307 | // fails until it catches up. A heartbeat older than five minutes is not an |
| 1308 | // answer — a dead relay must not be able to consent by silence — so a region |
| 1309 | // with no fresh heartbeat at all refuses too. |
| 1310 | // |
| 1311 | // The new kid is the LAST staged entry that is not the current issuer, which |
| 1312 | // is where an operator following the runbook appends it. |
| 1313 | func (s *Server) relayRotate(ctx context.Context, region string) (relayRotateResult, error) { |
| 1314 | var res relayRotateResult |
| 1315 | if err := s.relayReady(); err != nil { |
| 1316 | return res, err |
| 1317 | } |
| 1318 | region = strings.ToLower(strings.TrimSpace(region)) |
| 1319 | if !slices.Contains(s.relayRegions(), region) { |
| 1320 | return res, fmt.Errorf("this deployment runs no relay in %s — armed regions are %s", |
| 1321 | region, strings.Join(s.relayRegions(), ", ")) |
| 1322 | } |
| 1323 | if err := s.requireRelayLease(); err != nil { |
| 1324 | return res, err |
| 1325 | } |
| 1326 | staged, err := relayStagedMasters(region) |
| 1327 | if err != nil { |
| 1328 | return res, err |
| 1329 | } |
| 1330 | current, issuerEtag, err := s.relayIssuers(ctx) |
| 1331 | if err != nil { |
| 1332 | return res, err |
| 1333 | } |
| 1334 | curKid := current[region] |
| 1335 | if curKid == "" { |
| 1336 | curKid = staged[0].Kid |
| 1337 | } |
| 1338 | // The new kid is the LAST staged entry, full stop — never "the last one |
| 1339 | // that is not the issuer". After a rotation the variable holds |
| 1340 | // "<old>,<new>" with the pointer on <new> for the 28 hours before the old |
| 1341 | // one may be retired, and a second rotate in that window must refuse, not |
| 1342 | // walk the pointer BACK onto the master about to be retired and re-issue |
| 1343 | // every app under it. |
| 1344 | next := staged[len(staged)-1].Kid |
| 1345 | if next == curKid { |
| 1346 | return res, fmt.Errorf("no new master is staged for %s: the last entry in %s is already the issuer", |
| 1347 | region, relayMastersEnv(region)) |
| 1348 | } |
| 1349 | hbs, err := s.relayHeartbeats(ctx) |
| 1350 | if err != nil { |
| 1351 | return res, err |
| 1352 | } |
| 1353 | now := s.relayTime() |
| 1354 | live := 0 |
| 1355 | for _, hb := range hbs { |
| 1356 | if hb.Region != region || now.Sub(hb.ReadAt) > relayHeartbeatFresh { |
| 1357 | continue |
| 1358 | } |
| 1359 | live++ |
| 1360 | if !slices.Contains(hb.Kids, next) { |
| 1361 | return res, fmt.Errorf("relay %s in %s has not loaded kid %s yet (it has %s) — add the line to /etc/carlos/relay-masters there and try again", |
| 1362 | hb.Box, region, next, strings.Join(hb.Kids, ",")) |
| 1363 | } |
| 1364 | } |
| 1365 | if live == 0 { |
| 1366 | return res, fmt.Errorf("no relay in %s has heartbeated in the last %s — rotating now would issue keys nothing can verify", |
| 1367 | region, relayHeartbeatFresh) |
| 1368 | } |
| 1369 | // The pointer moves FIRST, under CAS, and every key derived after this |
| 1370 | // line uses the new master. An interrupted rotation therefore leaves a |
| 1371 | // moved pointer and some apps still on the old kid, which is exactly what |
| 1372 | // the reconcile converges — and why the old kid stays accepted for 28 |
| 1373 | // hours. |
| 1374 | current[region] = next |
| 1375 | body, err := json.MarshalIndent(current, "", " ") |
| 1376 | if err != nil { |
| 1377 | return res, err |
| 1378 | } |
| 1379 | if err := s.requireRelayLease(); err != nil { |
| 1380 | return res, err |
| 1381 | } |
| 1382 | if _, err := s.cfg.Releases.Put(ctx, relayIssuerKey, append(body, '\n'), issuerEtag); err != nil { |
| 1383 | return res, err |
| 1384 | } |
| 1385 | recs, err := s.relayListRecords(ctx) |
| 1386 | if err != nil { |
| 1387 | return res, err |
| 1388 | } |
| 1389 | reissued := 0 |
| 1390 | for _, r := range recs { |
| 1391 | if !r.Record.Enabled || !slices.Contains(r.Record.Regions, region) { |
| 1392 | continue |
| 1393 | } |
| 1394 | if err := s.relayReissue(ctx, r); err != nil { |
| 1395 | // One app that cannot be re-issued must not stop the rest: the |
| 1396 | // pointer has already moved, so every remaining app is one the |
| 1397 | // reconcile would otherwise have to catch five minutes later. |
| 1398 | log.Printf("console: relay rotate %s: %s/%s: %v", region, r.Account, r.App, err) |
| 1399 | continue |
| 1400 | } |
| 1401 | reissued++ |
| 1402 | } |
| 1403 | log.Printf("console: relay rotated %s to kid %s, re-issued %d app(s)", region, next, reissued) |
| 1404 | return relayRotateResult{Region: region, Kid: next, Apps: reissued, |
| 1405 | RetireNoEarlierThan: now.Add(relayRetireAfter).UTC()}, nil |
| 1406 | } |
| 1407 | |
| 1408 | // relayReissue is a rotation's per-app operation: op bumped, config |
| 1409 | // re-applied from the master in force now, aggregate published, done. The |
| 1410 | // aggregate does not CHANGE on a rotation (the incarnation is untouched), but |
| 1411 | // it is published anyway because the phase walk is what makes the operation |
| 1412 | // resumable, and a publish is cheap next to a half-recorded rotation. |
| 1413 | func (s *Server) relayReissue(ctx context.Context, r relayAppRecord) error { |
| 1414 | mu := s.relayAppLock(r.Account, r.App) |
| 1415 | mu.Lock() |
| 1416 | defer mu.Unlock() |
| 1417 | return s.relayAppMutation(ctx, r.Account, r.App, "relay-reissue", func(ctx context.Context) error { |
| 1418 | return s.relayReissueWalk(ctx, r) |
| 1419 | }) |
| 1420 | } |
| 1421 | |
| 1422 | // relayReissueWalk is relayReissue's body under the app's lock and permit. |
| 1423 | func (s *Server) relayReissueWalk(ctx context.Context, r relayAppRecord) error { |
| 1424 | if err := s.requireRelayLease(); err != nil { |
| 1425 | return err |
| 1426 | } |
| 1427 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, r.Account, r.App) |
| 1428 | if err != nil { |
| 1429 | return err |
| 1430 | } |
| 1431 | if rec == nil || !rec.Enabled { |
| 1432 | return nil // disabled under us: the reconcile owns it now |
| 1433 | } |
| 1434 | rec.Op++ |
| 1435 | rec.Phase = relayPhaseConfig |
| 1436 | if etag, err = s.relayPutRecord(ctx, r.Account, r.App, rec, etag); err != nil { |
| 1437 | return err |
| 1438 | } |
| 1439 | if err := s.relayApplyConfig(ctx, r.Account, r.App, rec); err != nil { |
| 1440 | return err |
| 1441 | } |
| 1442 | rec.Phase = relayPhaseAggregate |
| 1443 | if etag, err = s.relayPutRecord(ctx, r.Account, r.App, rec, etag); err != nil { |
| 1444 | return err |
| 1445 | } |
| 1446 | if err := s.relayPublishAggregate(ctx); err != nil { |
| 1447 | return err |
| 1448 | } |
| 1449 | rec.Phase = relayPhaseDone |
| 1450 | _, err = s.relayPutRecord(ctx, r.Account, r.App, rec, etag) |
| 1451 | return err |
| 1452 | } |
| 1453 | |
| 1454 | // ---- the reconcile ---------------------------------------------------------- |
| 1455 | |
| 1456 | // relayReconcile converges BOTH derived artefacts from every record, every |
| 1457 | // five minutes and once at console start (spec §5). |
| 1458 | // |
| 1459 | // It deliberately does not look for unfinished records. For each enabled |
| 1460 | // record it derives what the config should hold from the record and the |
| 1461 | // masters in force, and re-applies when the config holds anything else; for |
| 1462 | // each disabled record it removes the variable if present; and it ALWAYS |
| 1463 | // rebuilds and publishes the aggregate, which is what repairs a deletion that |
| 1464 | // removed a record but died before publishing. |
| 1465 | // |
| 1466 | // One app's failure is logged and skipped, never fatal. The aggregate publish |
| 1467 | // is the last step and the only one whose failure is returned, because it is |
| 1468 | // the one that speaks for every app at once. |
| 1469 | func (s *Server) relayReconcile(ctx context.Context) error { |
| 1470 | if s.cfg.Releases == nil { |
| 1471 | return errRelayNoBucket |
| 1472 | } |
| 1473 | if err := s.requireRelayLease(); err != nil { |
| 1474 | return err |
| 1475 | } |
| 1476 | recs, err := s.relayListRecords(ctx) |
| 1477 | if err != nil { |
| 1478 | return err |
| 1479 | } |
| 1480 | // Records whose walk reached `aggregate` in this pass, with the etag to |
| 1481 | // mark them done once the publish below has landed. |
| 1482 | type pending struct { |
| 1483 | rec relayAppRecord |
| 1484 | etag string |
| 1485 | } |
| 1486 | var advanced []pending |
| 1487 | for _, r := range recs { |
| 1488 | etag, changed, err := s.relayConverge(ctx, r) |
| 1489 | if err != nil { |
| 1490 | log.Printf("console: relay reconcile %s/%s: %v", r.Account, r.App, err) |
| 1491 | continue |
| 1492 | } |
| 1493 | if changed { |
| 1494 | advanced = append(advanced, pending{rec: r, etag: etag}) |
| 1495 | } |
| 1496 | } |
| 1497 | if err := s.relayPublishAggregate(ctx); err != nil { |
| 1498 | return err |
| 1499 | } |
| 1500 | for _, p := range advanced { |
| 1501 | if err := s.relayMarkDone(ctx, p.rec, p.etag); err != nil { |
| 1502 | log.Printf("console: relay reconcile %s/%s: marking done: %v", p.rec.Account, p.rec.App, err) |
| 1503 | } |
| 1504 | } |
| 1505 | return nil |
| 1506 | } |
| 1507 | |
| 1508 | // relayConverge brings ONE record's config into agreement with it, returning |
| 1509 | // the etag it left the record at and whether the record still needs marking |
| 1510 | // done after the publish. |
| 1511 | func (s *Server) relayConverge(ctx context.Context, r relayAppRecord) (string, bool, error) { |
| 1512 | mu := s.relayAppLock(r.Account, r.App) |
| 1513 | mu.Lock() |
| 1514 | defer mu.Unlock() |
| 1515 | var ( |
| 1516 | etag string |
| 1517 | changed bool |
| 1518 | ) |
| 1519 | err := s.relayAppMutation(ctx, r.Account, r.App, "relay-converge", func(ctx context.Context) error { |
| 1520 | var err error |
| 1521 | etag, changed, err = s.relayConvergeWalk(ctx, r) |
| 1522 | return err |
| 1523 | }) |
| 1524 | if err != nil { |
| 1525 | return "", false, err |
| 1526 | } |
| 1527 | return etag, changed, nil |
| 1528 | } |
| 1529 | |
| 1530 | // relayConvergeWalk is relayConverge's body under the app's lock and permit. |
| 1531 | func (s *Server) relayConvergeWalk(ctx context.Context, r relayAppRecord) (string, bool, error) { |
| 1532 | if err := s.requireRelayLease(); err != nil { |
| 1533 | return "", false, err |
| 1534 | } |
| 1535 | // Re-read under the lock: the listing is a snapshot, and an operation may |
| 1536 | // have moved this record between it and here. |
| 1537 | rec, etag, err := relay.GetRecord(ctx, s.cfg.Releases, r.Account, r.App) |
| 1538 | if err != nil || rec == nil { |
| 1539 | return "", false, err |
| 1540 | } |
| 1541 | if rec.Enabled { |
| 1542 | want, err := s.relayExpectedSuffixes(ctx, r.Account, r.App, rec) |
| 1543 | if err != nil { |
| 1544 | return "", false, err |
| 1545 | } |
| 1546 | v, present, err := s.relayConfigVar(ctx, r.Account, r.App) |
| 1547 | if err != nil { |
| 1548 | return "", false, err |
| 1549 | } |
| 1550 | // Current means BOTH halves agree: the identities the record says it |
| 1551 | // issued are the ones the masters in force imply (a rotation, |
| 1552 | // re-enable or region change moves those), AND the bundle still holds |
| 1553 | // the exact envelope this console last wrote (a straggler write from |
| 1554 | // a superseded console, or a member setting CARLOS_RELAY themselves, |
| 1555 | // moves that). A record from before IssuedDigest existed has none, so |
| 1556 | // it re-applies once and then carries one. |
| 1557 | current := present && v.Kind == "secret" && v.Envelope != "" && |
| 1558 | maps.Equal(rec.IssuedSuffixes, want) && |
| 1559 | rec.IssuedDigest != "" && relayEnvelopeDigest(v.Envelope) == rec.IssuedDigest |
| 1560 | if current && rec.Phase == relayPhaseDone { |
| 1561 | return etag, false, nil |
| 1562 | } |
| 1563 | if !current { |
| 1564 | if err := s.relayApplyConfig(ctx, r.Account, r.App, rec); err != nil { |
| 1565 | return "", false, err |
| 1566 | } |
| 1567 | } |
| 1568 | } else { |
| 1569 | if _, err := s.relayUnsetConfig(ctx, r.Account, r.App); err != nil { |
| 1570 | return "", false, err |
| 1571 | } |
| 1572 | if rec.Phase == relayPhaseDone { |
| 1573 | return etag, false, nil |
| 1574 | } |
| 1575 | } |
| 1576 | rec.Phase = relayPhaseAggregate |
| 1577 | etag, err = s.relayPutRecord(ctx, r.Account, r.App, rec, etag) |
| 1578 | if err != nil { |
| 1579 | return "", false, err |
| 1580 | } |
| 1581 | return etag, true, nil |
| 1582 | } |
| 1583 | |
| 1584 | // relayMarkDone is the last step of a reconciled record's walk, after the |
| 1585 | // aggregate that carries it has been published. |
| 1586 | func (s *Server) relayMarkDone(ctx context.Context, r relayAppRecord, etag string) error { |
| 1587 | mu := s.relayAppLock(r.Account, r.App) |
| 1588 | mu.Lock() |
| 1589 | defer mu.Unlock() |
| 1590 | return s.relayAppMutation(ctx, r.Account, r.App, "relay-mark-done", func(ctx context.Context) error { |
| 1591 | return s.relayMarkDoneWalk(ctx, r, etag) |
| 1592 | }) |
| 1593 | } |
| 1594 | |
| 1595 | // relayMarkDoneWalk is relayMarkDone's body under the app's lock and permit. |
| 1596 | func (s *Server) relayMarkDoneWalk(ctx context.Context, r relayAppRecord, etag string) error { |
| 1597 | if err := s.requireRelayLease(); err != nil { |
| 1598 | return err |
| 1599 | } |
| 1600 | rec, cur, err := relay.GetRecord(ctx, s.cfg.Releases, r.Account, r.App) |
| 1601 | if err != nil || rec == nil { |
| 1602 | return err |
| 1603 | } |
| 1604 | if cur != etag { |
| 1605 | return nil // moved on under us; its own walk owns the phase now |
| 1606 | } |
| 1607 | rec.Phase = relayPhaseDone |
| 1608 | _, err = s.relayPutRecord(ctx, r.Account, r.App, rec, etag) |
| 1609 | return err |
| 1610 | } |
| 1611 | |
amadan