| 1 | // The CLI's object-storage API (issue #55): `carlos store create` declares |
| 2 | // that an app wants a bucket, `carlos store status` answers what became of |
| 3 | // that request, and grant/rotate/finish drive the walks that make it real. |
| 4 | // Every door here sits behind the same apiAccountApp gate as |
| 5 | // ship/promote/config — an unknown account, an unknown app and someone else's |
| 6 | // account are one byte-identical 404. |
| 7 | // |
| 8 | // THREE AUTHORS, TWO AUTHORITIES. A store record has three blocks for three |
| 9 | // actors (spec §4): the member DECLARES, a deployment operator GRANTS, and |
| 10 | // the console's own walk records what it PROVISIONED. Declaring and rotating |
| 11 | // are member verbs. Provisioning also admits members once a deployment |
| 12 | // operator has approved their account (apistore_provision.go). |
| 13 | // Per-app granting needs an address on CARLOS_STORE_OPERATORS as well |
| 14 | // (Config.StoreOperators). That check runs AFTER the member gate on purpose: |
| 15 | // a 403 handed to a stranger would confirm which accounts and apps exist. |
| 16 | // This is the deployment's first explicit operator-identity gate (spec §3). |
| 17 | // |
| 18 | // NOTHING SECRET CROSSES THESE DOORS, in either direction. The record holds |
| 19 | // bucket names, an IAM user, and access key IDs; the secret half of a key |
| 20 | // exists only sealed inside the app's config bundle and is never written to |
| 21 | // the record at all (objstore.Record's doc comment). Key IDs are echoed |
| 22 | // deliberately — `status` prints them so a member can tell which credential |
| 23 | // their app is running on, and match it against a rotation. |
| 24 | // |
| 25 | // THE SERIAL BUMPS EXACTLY WHERE CONFIG MOVED. Declaring changes nothing a |
| 26 | // box reads, and neither does finishing a rotation (it deletes a key in the |
| 27 | // backend and forgets it in the record); both are silent. Grant and rotate |
| 28 | // write the app's config bundle, so both announce, and the fleet picks the new |
| 29 | // credentials up in seconds instead of at the hourly floor. |
| 30 | // |
| 31 | // Spec: docs/superpowers/specs/2026-08-10-object-storage-primitive-design.md |
| 32 | // §3, §4, §5, §7, §8. |
| 33 | package console |
| 34 | |
| 35 | import ( |
| 36 | "context" |
| 37 | "crypto/ecdh" |
| 38 | "encoding/json" |
| 39 | "errors" |
| 40 | "fmt" |
| 41 | "log" |
| 42 | "net/http" |
| 43 | "os" |
| 44 | "strings" |
| 45 | "time" |
| 46 | |
| 47 | "github.com/carlosframework/platform/internal/appconfig" |
| 48 | "github.com/carlosframework/platform/internal/objstore" |
| 49 | "github.com/carlosframework/platform/internal/release" |
| 50 | "github.com/carlosframework/platform/internal/usage" |
| 51 | ) |
| 52 | |
| 53 | // storeTimeout bounds every store round trip these handlers make — the same |
| 54 | // 15s its siblings in apicli.go and apiconfig.go use, for the same reason: a |
| 55 | // wedged bucket must cost one request 15 seconds, not a console goroutine |
| 56 | // forever. Each handler derives its context from the request's, so a client |
| 57 | // that hangs up still cancels immediately; this only adds a ceiling. |
| 58 | const storeTimeout = 15 * time.Second |
| 59 | |
| 60 | // storeNoBucket is the 503 both doors give when this console has no |
| 61 | // deployment bucket. The record IS a bucket object: there is nowhere to write |
| 62 | // one and nothing to read one from, and answering "no store is declared" |
| 63 | // instead would be a lie a member could act on (they'd declare again, forever). |
| 64 | func storeNoBucket(w http.ResponseWriter) { |
| 65 | jsonError(w, http.StatusServiceUnavailable, |
| 66 | "no deployment bucket on this host — this console cannot serve object storage") |
| 67 | } |
| 68 | |
| 69 | // storeEffectivePrefix is the prefix a record's variables are delivered |
| 70 | // under. A record with no declaration at all is not a shape this file can |
| 71 | // write — the create door always writes `declared` — but the operator walks |
| 72 | // and hand-repaired records both exist, so the fallback is the same default a |
| 73 | // fresh declaration would take rather than an empty prefix that would render |
| 74 | // variables named "_BUCKET". |
| 75 | // storeEnvParam reads the environment a store door is aimed at, from |
| 76 | // ?environment=. Absent is "" — the app's own store, and every request made |
| 77 | // before environment stores existed. |
| 78 | // |
| 79 | // One helper for all six doors on purpose: the create door is the only one |
| 80 | // with a body, and rotate and finish document that they read none at all |
| 81 | // ("a rotation has no arguments"), so a body field would have made the |
| 82 | // spelling differ per door for no reason a member could predict. It is the |
| 83 | // same query parameter, with the same name, that the config plane already |
| 84 | // uses (apiconfig.go's ?environment=) — a store delivers INTO a bundle, so |
| 85 | // the two planes naming it differently would be a seam with no meaning. |
| 86 | // |
| 87 | // Validated here rather than in the walk: a refusal costs nothing, and an |
| 88 | // invalid name reaching objstore would be refused there anyway, one round |
| 89 | // trip later and with a less useful sentence. |
| 90 | func (s *Server) storeEnvParam(w http.ResponseWriter, r *http.Request) (string, bool) { |
| 91 | env := strings.TrimSpace(r.URL.Query().Get("environment")) |
| 92 | if env == "" { |
| 93 | return "", true |
| 94 | } |
| 95 | if !appconfig.ValidEnvironmentName(env) { |
| 96 | jsonError(w, http.StatusBadRequest, |
| 97 | fmt.Sprintf("invalid environment %q — want ^[a-z0-9-]{1,32}$, and not \"default\"", env)) |
| 98 | return "", false |
| 99 | } |
| 100 | return env, true |
| 101 | } |
| 102 | |
| 103 | func storeEffectivePrefix(rec *objstore.Record) string { |
| 104 | if rec != nil && rec.Declared != nil && rec.Declared.EnvPrefix != "" { |
| 105 | return rec.Declared.EnvPrefix |
| 106 | } |
| 107 | return objstore.DefaultEnvPrefix |
| 108 | } |
| 109 | |
| 110 | // storeReply is the ONE answer shape both doors give, and the one the CLI |
| 111 | // (`carlos store status`) and the operator doors parse. Kept in a single |
| 112 | // function so create and status can never drift into describing the same |
| 113 | // record differently — a create that reported a different `pending_finish` |
| 114 | // from the status call right after it would be worse than no field at all. |
| 115 | // |
| 116 | // The record is echoed whole. Its three blocks are null-or-present, which is |
| 117 | // exactly the tri-state a client needs to say "declared, waiting for an |
| 118 | // operator" apart from "provisioned": an empty object for an absent block |
| 119 | // would make those indistinguishable without inspecting fields. |
| 120 | // |
| 121 | // env_names is the SET OF NAMES this store delivers under its effective |
| 122 | // prefix, in delivery order, and it includes the endpoint name only when the |
| 123 | // store is provisioned WITH one. A deployment that provisions no endpoint |
| 124 | // never writes that variable, so listing it would send someone hunting for a |
| 125 | // variable nothing will ever set. |
| 126 | // |
| 127 | // missing_vars is the same list filtered to what has NOT arrived in the app's |
| 128 | // config bundle, and it is computed ONLY for a provisioned store: before |
| 129 | // provisioning there is no bucket, no key and nothing owed, so reporting four |
| 130 | // missing variables would describe a healthy waiting-for-an-operator store as |
| 131 | // a broken one. Both lists are always present and never null — a client reads |
| 132 | // them rather than testing for absence. |
| 133 | // |
| 134 | // backend is the DEPLOYMENT's backend name, "" when none is configured, and |
| 135 | // it is here because without it `status` cannot answer the question a waiting |
| 136 | // member actually has: is my declaration queued behind an operator, or does |
| 137 | // this deployment have no object storage at all and never will? The |
| 138 | // no-infra-access rule (AGENTS.md) says the command alone must complete the |
| 139 | // job, and "declared, nothing happening" with no way to tell those apart |
| 140 | // fails it. Disclosure-wise it is the same class as secrets_pubkey and the |
| 141 | // apps domain: a deployment-wide fact about the platform, carrying nothing |
| 142 | // about any tenant, and every member of it can already see the consequences. |
| 143 | // storeEntitlementWire is the TENANT-FACING projection of an Entitlement, and |
| 144 | // exists because the record is operator commentary as much as it is tenant |
| 145 | // fact. Decision.By is the address of the person who decided; Decision.Reason |
| 146 | // is what they said about this account. §4 of the self-serve spec keeps both |
| 147 | // out of the fleet read ACL by putting the record under store/ — this keeps |
| 148 | // them out of the console's own answer, which is the other road a tenant could |
| 149 | // read them by, and the one an Admin service credential travels. |
| 150 | // |
| 151 | // What travels, and to whom: |
| 152 | // |
| 153 | // Decision.By never. The operator stays answerable to the deployment, |
| 154 | // not to the account they just declined. |
| 155 | // Decision.Reason to a PERSON only. Requiring a reason on decline is |
| 156 | // pointless if the account never learns it — that is the |
| 157 | // four-day silence of §1 in a new hat — but it is written |
| 158 | // for someone to read and act on, not for an app to branch |
| 159 | // on. |
| 160 | // Applied.* the applicant's own text about themselves, so it goes |
| 161 | // back to a person untouched. A machine gets the At alone: |
| 162 | // it needs the STATE, and free text a colleague typed is |
| 163 | // not state. |
| 164 | // |
| 165 | // The shape mirrors Entitlement's tags exactly, so the CLI decodes this into |
| 166 | // an objstore.Entitlement and the four-state printers keep working — a |
| 167 | // withheld field simply arrives as its zero value. |
| 168 | type storeEntitlementWire struct { |
| 169 | Applied *storeAppliedWire `json:"applied,omitempty"` |
| 170 | Decision *storeDecisionWire `json:"decision,omitempty"` |
| 171 | } |
| 172 | |
| 173 | type storeAppliedWire struct { |
| 174 | By string `json:"by,omitempty"` |
| 175 | At time.Time `json:"at"` |
| 176 | Note string `json:"note,omitempty"` |
| 177 | } |
| 178 | |
| 179 | type storeDecisionWire struct { |
| 180 | Approved bool `json:"approved"` |
| 181 | At time.Time `json:"at"` |
| 182 | Reason string `json:"reason,omitempty"` |
| 183 | } |
| 184 | |
| 185 | // The caller is read from ctx rather than passed: every door that renders a |
| 186 | // store reply builds its ctx from the request, so the principal is already |
| 187 | // there, and a parameter threaded through five call sites is five chances to |
| 188 | // pass the wrong one. A ctx with NO principal is treated as a machine — the |
| 189 | // fail-closed direction, and the operator walk doors are the only callers in |
| 190 | // that position, where the withheld reason is one they wrote themselves. |
| 191 | func tenantEntitlement(e *objstore.Entitlement, ctx context.Context) *storeEntitlementWire { |
| 192 | if e == nil { |
| 193 | return nil |
| 194 | } |
| 195 | p, ok := principalFrom(ctx) |
| 196 | person := ok && p.Kind == PrincipalUser |
| 197 | out := &storeEntitlementWire{} |
| 198 | if e.Applied != nil { |
| 199 | out.Applied = &storeAppliedWire{At: e.Applied.At} |
| 200 | if person { |
| 201 | out.Applied.By, out.Applied.Note = e.Applied.By, e.Applied.Note |
| 202 | } |
| 203 | } |
| 204 | if e.Decision != nil { |
| 205 | out.Decision = &storeDecisionWire{Approved: e.Decision.Approved, At: e.Decision.At} |
| 206 | if person { |
| 207 | out.Decision.Reason = e.Decision.Reason |
| 208 | } |
| 209 | } |
| 210 | return out |
| 211 | } |
| 212 | |
| 213 | func (s *Server) storeReply(ctx context.Context, acct *Account, app, env string, rec *objstore.Record) (map[string]any, error) { |
| 214 | entitlement, _, err := objstore.GetEntitlement(ctx, s.cfg.Releases, acct.Sqid) |
| 215 | if err != nil { |
| 216 | return nil, err |
| 217 | } |
| 218 | prefix := storeEffectivePrefix(rec) |
| 219 | wantEndpoint := rec.Provisioned != nil && rec.Provisioned.Endpoint != "" |
| 220 | bucket, region, endpoint, accessKey, secretKey := objstore.EnvNames(prefix) |
| 221 | names := []string{bucket, region} |
| 222 | if wantEndpoint { |
| 223 | names = append(names, endpoint) |
| 224 | } |
| 225 | names = append(names, accessKey, secretKey) |
| 226 | missing := []string{} |
| 227 | if rec.Provisioned != nil { |
| 228 | // The bundle this store DELIVERS into, which for an environment store |
| 229 | // is not the app's default one. Asking the default bundle would |
| 230 | // report every delivered variable missing, on every status call, for |
| 231 | // every environment store. |
| 232 | got, err := objstore.EnvDeliveredIn(ctx, s.cfg.Releases, acct.Sqid, app, env, prefix, wantEndpoint) |
| 233 | if err != nil { |
| 234 | return nil, err |
| 235 | } |
| 236 | missing = append(missing, got...) |
| 237 | } |
| 238 | return map[string]any{ |
| 239 | "app": app, |
| 240 | // "" for the app's own store. Omitting it when empty would make an |
| 241 | // old CLI and a new one disagree about which store a reply describes, |
| 242 | // so it is always present and always says which. |
| 243 | "environment": env, |
| 244 | "entitlement": tenantEntitlement(entitlement, ctx), |
| 245 | "backend": s.cfg.StoreDeployment.Backend, |
| 246 | "declared": rec.Declared, |
| 247 | "grant": rec.Grant, |
| 248 | "provisioned": rec.Provisioned, |
| 249 | "env_names": names, |
| 250 | // A rotation that minted a new key but has not yet deleted the old |
| 251 | // one. The member-visible half of §6's two-step: until it is finished, |
| 252 | // two credentials open this bucket. |
| 253 | "pending_finish": rec.Provisioned != nil && rec.Provisioned.PrevKeyID != "", |
| 254 | "missing_vars": missing, |
| 255 | // Additive, and additive is the ONLY safe move here: release.PurgeApp |
| 256 | // parses this same object with its own struct and freezes the `grant` |
| 257 | // and `provisioned` names. A new key it does not know about is |
| 258 | // ignored; a renamed one it does know about is a purge that stops |
| 259 | // working. Once scanning exists, a status that said nothing about it |
| 260 | // would be a half-truth — but it is nil on every store that never |
| 261 | // opted in, which is all of them by default. |
| 262 | "scanning": rec.Scanning, |
| 263 | // The bucket's size from the daily storage sweep, absent until a |
| 264 | // sweep has measured this store. See storeSize. |
| 265 | "size": s.storeSize(ctx, acct, app, rec), |
| 266 | }, nil |
| 267 | } |
| 268 | |
| 269 | // storeSizeLookback is how many UTC days storeSize walks back looking for a |
| 270 | // storage sweep. The sweep runs on the console's first hourly economics wake |
| 271 | // of each day, so today or yesterday normally answers; a week covers a |
| 272 | // console that was down for a while, and anything older is too stale to |
| 273 | // print as the bucket's size. |
| 274 | const storeSizeLookback = 7 |
| 275 | |
| 276 | // storeSizeReply is the "size" block of `carlos store status`. It is copied |
| 277 | // from the storage sweep (usage/storage/<date>.json) and is NOT measured on |
| 278 | // this request: sizing a bucket means listing every object in it, which is |
| 279 | // what the daily sweep already does once. Unswept means the sweep ran but |
| 280 | // could not size this store (bad credential, endpoint down), and then |
| 281 | // StoreBytes is meaningless. BucketBytes is the app's own files in the |
| 282 | // deployment bucket (releases, config), a separate figure from its store. |
| 283 | type storeSizeReply struct { |
| 284 | AsOf string `json:"as_of"` |
| 285 | StoreBytes int64 `json:"store_bytes"` |
| 286 | Unswept bool `json:"unswept,omitempty"` |
| 287 | BucketBytes int64 `json:"bucket_bytes"` |
| 288 | } |
| 289 | |
| 290 | // storeSize finds the newest storage sweep from the last storeSizeLookback |
| 291 | // days that has an answer for this app's store, or nil. A sweep that did not |
| 292 | // see the store provisioned (it ran before the grant, or predates the |
| 293 | // store_provisioned field) has no answer: its StoreBytes of 0 would read as |
| 294 | // an empty bucket. Read failures are logged and answer nil; the size is an |
| 295 | // annotation on status, never a reason for status to fail. |
| 296 | func (s *Server) storeSize(ctx context.Context, acct *Account, app string, rec *objstore.Record) *storeSizeReply { |
| 297 | if rec.Provisioned == nil { |
| 298 | return nil |
| 299 | } |
| 300 | key := acct.Sqid + "/" + app |
| 301 | now := time.Now() |
| 302 | for i := 0; i < storeSizeLookback; i++ { |
| 303 | date := usage.DayOf(now.AddDate(0, 0, -i)) |
| 304 | raw, _, err := s.cfg.Releases.Get(ctx, usage.StorageDayKey(date)) |
| 305 | if errors.Is(err, os.ErrNotExist) { |
| 306 | continue |
| 307 | } |
| 308 | if err != nil { |
| 309 | log.Printf("console: api store size %s: %v", usage.StorageDayKey(date), err) |
| 310 | return nil |
| 311 | } |
| 312 | var day usage.StorageDay |
| 313 | if err := json.Unmarshal(raw, &day); err != nil { |
| 314 | log.Printf("console: api store size: parse %s: %v", usage.StorageDayKey(date), err) |
| 315 | return nil |
| 316 | } |
| 317 | st, ok := day.Apps[key] |
| 318 | if !ok || (!st.StoreProvisioned && !st.StoreUnswept) { |
| 319 | // The newest sweep did not see this store; an older one cannot |
| 320 | // have either. |
| 321 | return nil |
| 322 | } |
| 323 | return &storeSizeReply{AsOf: date, StoreBytes: st.StoreBytes, |
| 324 | Unswept: st.StoreUnswept, BucketBytes: st.BucketBytes} |
| 325 | } |
| 326 | return nil |
| 327 | } |
| 328 | |
| 329 | // handleAPIStoreCreate answers `carlos store create`: record that this app |
| 330 | // wants object storage, under a chosen (or default) env prefix. It is a |
| 331 | // REQUEST, not a provisioning — nothing exists in the backend when this |
| 332 | // returns 200, and on a deployment with no backend configured at all the |
| 333 | // declaration simply waits for an operator who can grant it. That is |
| 334 | // deliberate: a member cannot see whether their deployment has a backend, and |
| 335 | // making them find out by being refused would teach them nothing they could |
| 336 | // act on, while the declaration itself is what an operator needs to act. |
| 337 | // |
| 338 | // The doors, in order, and why this order — handleAPIRestart's exactly: |
| 339 | // |
| 340 | // token — cliUser's one 401, before any name is resolved. |
| 341 | // account+app 404 — apiAccountApp: an unknown account, someone else's |
| 342 | // account and an app that isn't there are one answer, so |
| 343 | // probing this door teaches a stranger nothing. It also |
| 344 | // means a TRASHED app cannot declare a store: the live-apps |
| 345 | // lookup simply doesn't answer for it. |
| 346 | // rate limit — the shared `write` budget, keyed on the ADDRESS like |
| 347 | // promote's: a budget belongs to the person, not the |
| 348 | // terminal. |
| 349 | // Releases nil — 503. The declaration IS a bucket object; a console with |
| 350 | // no bucket has nowhere to put it and must not answer as |
| 351 | // if it had. |
| 352 | // prefix — refused before the first store call, so a corrected |
| 353 | // retry starts from exactly the state the caller left. |
| 354 | // |
| 355 | // A JSON BODY IS REQUIRED, even to take the default: readJSON's |
| 356 | // content-type-and-decode is shared by every CLI door, so `{}` is the empty |
| 357 | // request here — not a zero-length body, which decodes as malformed like |
| 358 | // anywhere else. |
| 359 | // |
| 360 | // Re-declaring the same prefix preserves the author, timestamp and etag. |
| 361 | // A no-op must not make an in-flight provision lose its record write. Members |
| 362 | // re-run create after a failed walk, from second terminals and deploy scripts. |
| 363 | func (s *Server) handleAPIStoreCreate(w http.ResponseWriter, r *http.Request) { |
| 364 | addr, acct, app, ok := s.apiAccountApp(w, r) |
| 365 | if !ok { |
| 366 | return |
| 367 | } |
| 368 | if s.refuseTemporary(w, acct, app, "give it object storage") { |
| 369 | return |
| 370 | } |
| 371 | p, _ := principalFrom(r.Context()) |
| 372 | if !s.allowWrite(p, time.Now()) { |
| 373 | jsonError(w, http.StatusTooManyRequests, "too many requests — wait a minute and try again") |
| 374 | return |
| 375 | } |
| 376 | if s.cfg.Releases == nil { |
| 377 | storeNoBucket(w) |
| 378 | return |
| 379 | } |
| 380 | var req struct { |
| 381 | EnvPrefix string `json:"env_prefix"` |
| 382 | } |
| 383 | if !readJSON(w, r, &req) { |
| 384 | return |
| 385 | } |
| 386 | prefix := strings.TrimSpace(req.EnvPrefix) |
| 387 | if prefix == "" { |
| 388 | prefix = objstore.DefaultEnvPrefix |
| 389 | } |
| 390 | if !objstore.ValidEnvPrefix(prefix) { |
| 391 | // The charset is named rather than described: the caller has to be |
| 392 | // able to fix their own prefix from this sentence alone, and the |
| 393 | // regexp is shorter than the English for it. |
| 394 | jsonError(w, http.StatusBadRequest, |
| 395 | fmt.Sprintf("invalid env prefix %q — want ^[A-Z][A-Z0-9_]{0,31}$ (e.g. %s)", req.EnvPrefix, objstore.DefaultEnvPrefix)) |
| 396 | return |
| 397 | } |
| 398 | env, ok := s.storeEnvParam(w, r) |
| 399 | if !ok { |
| 400 | return |
| 401 | } |
| 402 | ctx, cancel := context.WithTimeout(r.Context(), storeTimeout) |
| 403 | defer cancel() |
| 404 | // The environment-store ceiling (issue #412 §8) is enforced HERE as well |
| 405 | // as on the automatic path. A cap that only the create-an-instance door |
| 406 | // respected would be a cap anyone could step over by hand, and the |
| 407 | // footprint it bounds — a bucket and an IAM user per store, against |
| 408 | // IAM's per-account limits — does not care which door minted it. |
| 409 | // Re-declaring a store that already exists is never capped; storeCeiling |
| 410 | // carries that carve-out. |
| 411 | if ref := s.storeCeiling(ctx, acct, app, env); ref != nil { |
| 412 | jsonError(w, ref.Status, ref.Msg) |
| 413 | return |
| 414 | } |
| 415 | rec, etag, err := objstore.GetIn(ctx, s.cfg.Releases, acct.Sqid, app, env) |
| 416 | if err != nil { |
| 417 | // Logged, not returned: a store error carries bucket names, endpoints |
| 418 | // and request ids. The caller gets the fixed sentence, and a retry is |
| 419 | // safe — nothing was written. |
| 420 | log.Printf("console: api store read %s/%s: %v", acct.Sqid, app, err) |
| 421 | jsonError(w, http.StatusBadGateway, "could not read the store record — try again") |
| 422 | return |
| 423 | } |
| 424 | if rec == nil { |
| 425 | rec = &objstore.Record{} |
| 426 | } |
| 427 | // THE PREFIX IS FIXED ONCE AN OPERATOR HAS GRANTED. The delivered names |
| 428 | // are what a RUNNING app reads; renaming them would leave it looking for |
| 429 | // variables nothing will write again, while the old ones linger in its |
| 430 | // config pointing at a live credential. The refusal names the only honest |
| 431 | // way out rather than pretending this is a retry-able race. |
| 432 | // |
| 433 | // The guard keys on the GRANT, not on `provisioned`, because the grant |
| 434 | // walk is a sequence and its middle is exactly where a rename does the |
| 435 | // damage: the grant block is written first (§5 step 1) and the delivered |
| 436 | // variables land several steps before `provisioned` does. A re-declare |
| 437 | // squeezed into that window would be accepted under the old rule, and the |
| 438 | // vars already written under the OLD prefix would then be orphaned — |
| 439 | // nothing rewrites them, nothing reaps them, and they name a key the next |
| 440 | // pass revokes. Same widening Tasks 5 and 6 made to their own guards, for |
| 441 | // the same window. |
| 442 | if (rec.Grant != nil || rec.Provisioned != nil) && storeEffectivePrefix(rec) != prefix { |
| 443 | // Two sentences for two true states: an operator has GRANTED but the |
| 444 | // walk has not finished, versus a store that is fully provisioned. The |
| 445 | // way out is the same either way, and it is the half of the sentence |
| 446 | // that matters — but telling someone their store is "already |
| 447 | // provisioned" when no bucket exists yet is exactly the kind of |
| 448 | // confident lie an operator then has to talk them out of. |
| 449 | if rec.Provisioned == nil { |
| 450 | jsonError(w, http.StatusConflict, |
| 451 | "a store grant is already under way — the env prefix is fixed; destroy and re-create to change it") |
| 452 | return |
| 453 | } |
| 454 | jsonError(w, http.StatusConflict, |
| 455 | "store already provisioned — the env prefix is fixed; destroy and re-create to change it") |
| 456 | return |
| 457 | } |
| 458 | if rec.Declared == nil || storeEffectivePrefix(rec) != prefix { |
| 459 | rec.Declared = &objstore.Declared{EnvPrefix: prefix, At: time.Now().UTC(), By: addr} |
| 460 | if _, err := claimAppCall(s, ctx, acct, app, "store-declare", func(ctx context.Context) (string, error) { |
| 461 | return objstore.PutIn(ctx, s.cfg.Releases, acct.Sqid, app, env, rec, etag) |
| 462 | }); err != nil { |
| 463 | if errors.Is(err, release.ErrCASConflict) { |
| 464 | // Someone else wrote this record between our read and our write — |
| 465 | // plausibly the operator's grant landing. Re-read and re-ask; this |
| 466 | // door deliberately does not retry, because the thing it would be |
| 467 | // overwriting is the half it doesn't own. |
| 468 | jsonError(w, http.StatusConflict, "the store record changed underneath you. try again") |
| 469 | return |
| 470 | } |
| 471 | log.Printf("console: api store create %s/%s: %v", acct.Sqid, app, err) |
| 472 | jsonError(w, http.StatusBadGateway, "could not record the store. nothing was declared; try again") |
| 473 | return |
| 474 | } |
| 475 | // Audited, unlike a read: this is the request an operator will act on by |
| 476 | // spending money in the deployment's own cloud account, so who asked and |
| 477 | // under what prefix has to be visible to anyone reading rows. |
| 478 | s.audit(acct, app, addr, "store-create", "declared a store with env prefix "+prefix+storeEnvSuffix(env)+" (cli)") |
| 479 | } |
| 480 | // No bumpSerial: nothing a box reads has changed. See the file comment. |
| 481 | out, err := s.storeReply(ctx, acct, app, env, rec) |
| 482 | if err != nil { |
| 483 | log.Printf("console: api store reply %s/%s: %v", acct.Sqid, app, err) |
| 484 | jsonError(w, http.StatusBadGateway, "the store was declared, but its status could not be read — try `carlos store status`") |
| 485 | return |
| 486 | } |
| 487 | writeJSON(w, http.StatusOK, out) |
| 488 | } |
| 489 | |
| 490 | // handleAPIStoreGet answers `carlos store status`: the whole record plus the |
| 491 | // two things a member with no box access cannot see any other way — which |
| 492 | // delivered variables have not arrived in the app's config, and whether a |
| 493 | // rotation is still half-done. |
| 494 | // |
| 495 | // THE 404 HERE SPEAKS, and that is not a weakening of the gate above it. It |
| 496 | // fires only AFTER apiAccountApp has proved the caller is a member of the |
| 497 | // account and that the app exists, so the sentence tells them nothing they |
| 498 | // were not already entitled to. apiAccountApp's OWN refusals stay |
| 499 | // byte-identical with http.NotFound's, which is what keeps a stranger from |
| 500 | // learning that an account, an app or a store exists at all. |
| 501 | // |
| 502 | // Not rate limited, matching every other read door: an exhausted write budget |
| 503 | // must not blind a member to the state of what they already asked for. |
| 504 | func (s *Server) handleAPIStoreGet(w http.ResponseWriter, r *http.Request) { |
| 505 | _, acct, app, ok := s.apiAccountApp(w, r) |
| 506 | if !ok { |
| 507 | return |
| 508 | } |
| 509 | if s.cfg.Releases == nil { |
| 510 | storeNoBucket(w) |
| 511 | return |
| 512 | } |
| 513 | env, ok := s.storeEnvParam(w, r) |
| 514 | if !ok { |
| 515 | return |
| 516 | } |
| 517 | ctx, cancel := context.WithTimeout(r.Context(), storeTimeout) |
| 518 | defer cancel() |
| 519 | rec, _, err := objstore.GetIn(ctx, s.cfg.Releases, acct.Sqid, app, env) |
| 520 | if err != nil { |
| 521 | log.Printf("console: api store status %s/%s: %v", acct.Sqid, app, err) |
| 522 | jsonError(w, http.StatusBadGateway, "could not read the store record — try again") |
| 523 | return |
| 524 | } |
| 525 | if rec == nil { |
| 526 | // A state, not a failure — and a 404 rather than an empty record, |
| 527 | // because "no store" and "a store with nothing filled in yet" are |
| 528 | // genuinely different things for the CLI to print. |
| 529 | if env != "" { |
| 530 | jsonError(w, http.StatusNotFound, |
| 531 | "no store is declared for this app's "+env+" environment — run `carlos store create --environment "+env+"`") |
| 532 | return |
| 533 | } |
| 534 | // The app has no store of its OWN — but it may have environment |
| 535 | // stores, and this is the shape where answering "no store" is most |
| 536 | // wrong. An app whose storage is entirely per-environment (a managed |
| 537 | // product's, where every tenant is a bundle) would have every one of |
| 538 | // its buckets hidden behind a 404 that says it has none. That is the |
| 539 | // exact opposite of what this listing exists for: a store nobody |
| 540 | // reads still bills, and teardown is explicit. |
| 541 | envs, lerr := objstore.StoreEnvironments(ctx, s.cfg.Releases, acct.Sqid, app) |
| 542 | if lerr != nil { |
| 543 | log.Printf("console: api store environments %s/%s: %v", acct.Sqid, app, lerr) |
| 544 | jsonError(w, http.StatusBadGateway, "could not read the store record — try again") |
| 545 | return |
| 546 | } |
| 547 | if len(envs) == 0 { |
| 548 | jsonError(w, http.StatusNotFound, "no store is declared for this app — run `carlos store create`") |
| 549 | return |
| 550 | } |
| 551 | // A zero record, not a nil one: every block reads as absent, which is |
| 552 | // true, and the environments are named beside it. |
| 553 | out, rerr := s.storeReply(ctx, acct, app, "", &objstore.Record{}) |
| 554 | if rerr != nil { |
| 555 | log.Printf("console: api store delivery %s/%s: %v", acct.Sqid, app, rerr) |
| 556 | jsonError(w, http.StatusBadGateway, "could not read the store's status. try again") |
| 557 | return |
| 558 | } |
| 559 | out["environments"] = envs |
| 560 | writeJSON(w, http.StatusOK, out) |
| 561 | return |
| 562 | } |
| 563 | out, err := s.storeReply(ctx, acct, app, env, rec) |
| 564 | if err != nil { |
| 565 | log.Printf("console: api store delivery %s/%s: %v", acct.Sqid, app, err) |
| 566 | jsonError(w, http.StatusBadGateway, "could not read the store's status. try again") |
| 567 | return |
| 568 | } |
| 569 | // Asked about the app's own store, a member is told what ELSE this app |
| 570 | // has: an environment store nobody remembers is a bucket nobody |
| 571 | // remembers, and this is the one screen where it can be seen (spec §8). |
| 572 | // Only on the app-store call — a request aimed at one environment |
| 573 | // answers about that environment. |
| 574 | if env == "" { |
| 575 | envs, err := objstore.StoreEnvironments(ctx, s.cfg.Releases, acct.Sqid, app) |
| 576 | if err != nil { |
| 577 | log.Printf("console: api store environments %s/%s: %v", acct.Sqid, app, err) |
| 578 | jsonError(w, http.StatusBadGateway, "could not read the store's status. try again") |
| 579 | return |
| 580 | } |
| 581 | out["environments"] = envs |
| 582 | } |
| 583 | writeJSON(w, http.StatusOK, out) |
| 584 | } |
| 585 | |
| 586 | // storeEnvSuffix names the environment in an audit row, or says nothing at |
| 587 | // all for the app's own store — so rows written before environment stores |
| 588 | // existed and rows written after read identically for the same action. |
| 589 | func storeEnvSuffix(env string) string { |
| 590 | if env == "" { |
| 591 | return "" |
| 592 | } |
| 593 | return " in environment " + env |
| 594 | } |
| 595 | |
| 596 | // ---- the walk doors: grant, rotate, finish ----------------------------------- |
| 597 | |
| 598 | // storeNameTakenMsg answers objstore.ErrStoreNameTaken (#218). Bucket and IAM |
| 599 | // user names join account, app and environment with "-", so app "shop" with |
| 600 | // environment "eu" and app "shop-eu" would get one estate; the walk refuses |
| 601 | // the second before creating anything, and the sentence says why. |
| 602 | const storeNameTakenMsg = "another store already uses this store's bucket or IAM user name (app and environment names are joined with \"-\"). nothing was created" |
| 603 | |
| 604 | // storeGrantTimeout bounds the grant walk. Longer than storeTimeout because a |
| 605 | // grant is not one round trip: it creates a bucket, an IAM principal with an |
| 606 | // inline policy, and a key, then writes config and the record. Bucket creation |
| 607 | // alone can take seconds on a cold region, and a walk cut off mid-sequence |
| 608 | // costs the operator a re-run they shouldn't need. |
| 609 | const storeGrantTimeout = 60 * time.Second |
| 610 | |
| 611 | // storeRotateTimeout bounds rotate and finish. Both act on an estate that |
| 612 | // already exists — delete, reap, mint, deliver — so they are shorter than a |
| 613 | // grant but still well past the 15s a single object write gets. |
| 614 | const storeRotateTimeout = 30 * time.Second |
| 615 | |
| 616 | // storeTeardownTimeout bounds a Teardown walk, wherever it is driven from (the |
| 617 | // browser destroy door and the purge door, appsettings.go and ownership.go). |
| 618 | // A grant's ceiling for a walk with a grant's shape and one step a grant does |
| 619 | // not have: EmptyBucket is PAGINATED precisely because a real bucket does not |
| 620 | // empty in one call, so a customer's store can be many round trips before the |
| 621 | // bucket, the principal and the config write even begin. Under storeTimeout's |
| 622 | // 15s — a ceiling documented for a SINGLE round trip — the deadline would land |
| 623 | // mid-walk and be reported as a backend that could not be reached, which is |
| 624 | // the wrong thing to tell someone whose store is simply large. |
| 625 | const storeTeardownTimeout = 60 * time.Second |
| 626 | |
| 627 | // storeOperator reports whether this address may drive the grant door. |
| 628 | // |
| 629 | // EXACT MATCH ON THE CANONICAL FORM, never a fold. Every address entry point |
| 630 | // in this codebase ASCII-lowercases on the way in (sign-in, team-add, the |
| 631 | // materialize pass, `console member`), so the address on a CLI token is |
| 632 | // already canonical; the list entry is lowercased and trimmed here so an |
| 633 | // operator who typed Alice@Keymail.dev into a unit file still matches. |
| 634 | // |
| 635 | // strings.EqualFold would be the obvious way to write that and is WRONG for |
| 636 | // the highest-authority gate in the API: it applies Unicode simple folding, in |
| 637 | // which U+017F (ſ, long s) folds to "s" — and ſ is already lowercase, so it |
| 638 | // survives every ToLower on the way in. ValidAddress permits it (anything |
| 639 | // without whitespace and with an @). With a fold, owner@example.teſt would |
| 640 | // match an allowlist entry of owner@example.test and be handed the authority |
| 641 | // to spend the deployment's money. |
| 642 | // |
| 643 | // Empty list is fail-closed — the loop simply never matches — which is the |
| 644 | // right direction for a gate on spending. |
| 645 | func (s *Server) storeOperator(addr string) bool { |
| 646 | for _, op := range s.cfg.StoreOperators { |
| 647 | if strings.ToLower(strings.TrimSpace(op)) == addr { |
| 648 | return true |
| 649 | } |
| 650 | } |
| 651 | return false |
| 652 | } |
| 653 | |
| 654 | // storeWalkGates are the gates every store WALK door shares, after the caller |
| 655 | // has resolved the member (and, for grant, checked the operator allowlist): |
| 656 | // the write budget, a bucket to record into, and a backend to act on. |
| 657 | // |
| 658 | // The provisioner check is here rather than inside the walks because a |
| 659 | // deployment with no object storage is a configuration fact, not a failure of |
| 660 | // this request: the member's declaration still stands and the sentence names |
| 661 | // the variable that would wire it up, which is the only thing the person |
| 662 | // reading it can do about it. |
| 663 | func (s *Server) storeWalkGates(w http.ResponseWriter, p Principal) bool { |
| 664 | if !s.allowWrite(p, time.Now()) { |
| 665 | jsonError(w, http.StatusTooManyRequests, "too many requests — wait a minute and try again") |
| 666 | return false |
| 667 | } |
| 668 | if s.cfg.Releases == nil { |
| 669 | storeNoBucket(w) |
| 670 | return false |
| 671 | } |
| 672 | if s.cfg.StoreProvisioner == nil { |
| 673 | jsonError(w, http.StatusServiceUnavailable, |
| 674 | "object storage is not configured for this deployment (CARLOS_STORE_BACKEND)") |
| 675 | return false |
| 676 | } |
| 677 | return true |
| 678 | } |
| 679 | |
| 680 | // storeSealKey is the gate that says: no sealing key, no minting. |
| 681 | // |
| 682 | // A credential this console cannot seal is one it cannot DELIVER, and a key |
| 683 | // minted without delivery is the worst state in the system — it exists in the |
| 684 | // backend, it is billable, its secret died in this process, and only a reap |
| 685 | // finds it again. So grant and rotate refuse before the first backend call |
| 686 | // rather than after. Finish does not call this at all: it seals nothing. |
| 687 | func (s *Server) storeSealKey(w http.ResponseWriter) (*ecdh.PublicKey, bool) { |
| 688 | if s.cfg.SecretsPub == "" { |
| 689 | jsonError(w, http.StatusServiceUnavailable, |
| 690 | "this console has no secrets key (CARLOS_SECRETS_PUBKEY) — it cannot deliver a store credential, so it will not mint one") |
| 691 | return nil, false |
| 692 | } |
| 693 | pub, err := appconfig.ParsePublicKey(s.cfg.SecretsPub, "CARLOS_SECRETS_PUBKEY") |
| 694 | if err != nil { |
| 695 | // Logged with the parse detail, answered without it: the caller cannot |
| 696 | // fix this console's configuration and the sentence they get is the one |
| 697 | // that tells the operator where to look. |
| 698 | log.Printf("console: api store seal key: %v", err) |
| 699 | jsonError(w, http.StatusServiceUnavailable, |
| 700 | "this console's secrets key (CARLOS_SECRETS_PUBKEY) cannot be read — it cannot deliver a store credential, so it will not mint one") |
| 701 | return nil, false |
| 702 | } |
| 703 | return pub, true |
| 704 | } |
| 705 | |
| 706 | // storeBucketName is the bucket a record names, for an audit line: what was |
| 707 | // provisioned if the walk got that far, otherwise what the grant chose. |
| 708 | func storeBucketName(rec *objstore.Record) string { |
| 709 | switch { |
| 710 | case rec == nil: |
| 711 | return "" |
| 712 | case rec.Provisioned != nil: |
| 713 | return rec.Provisioned.Bucket |
| 714 | case rec.Grant != nil: |
| 715 | return rec.Grant.Bucket |
| 716 | } |
| 717 | return "" |
| 718 | } |
| 719 | |
| 720 | // storeWalkReply is the shared tail of all three walk doors: render the same |
| 721 | // answer shape the member doors give, from the record the walk just wrote. |
| 722 | // A walk that succeeded and a reply that could not be rendered are different |
| 723 | // events, and the sentence says so — the thing was done, go and look at it — |
| 724 | // because telling someone their grant failed when it didn't is how a second |
| 725 | // grant gets run over a live estate. |
| 726 | func (s *Server) storeWalkReply(ctx context.Context, w http.ResponseWriter, acct *Account, app, env string, rec *objstore.Record) { |
| 727 | out, err := s.storeReply(ctx, acct, app, env, rec) |
| 728 | if err != nil { |
| 729 | log.Printf("console: api store reply %s/%s: %v", acct.Sqid, app, err) |
| 730 | jsonError(w, http.StatusBadGateway, |
| 731 | "the store changed, but its status could not be read — run `carlos store status`") |
| 732 | return |
| 733 | } |
| 734 | writeJSON(w, http.StatusOK, out) |
| 735 | } |
| 736 | |
| 737 | // handleAPIStoreGrant answers `carlos store grant`: provision the declared |
| 738 | // store and deliver its credentials. THIS IS THE AUTHORITY DOOR — the one |
| 739 | // request in the whole API that creates a bucket, mints an IAM key and starts |
| 740 | // a bill in the deployment's own cloud account. |
| 741 | // |
| 742 | // The doors, in order, and why this order: |
| 743 | // |
| 744 | // token — cliUser's one 401, before any name is resolved. |
| 745 | // account+app 404 — apiAccountApp, FIRST and before the allowlist. An |
| 746 | // operator is not automatically a member: eve holding the |
| 747 | // operator role must still get the same 404 for an account |
| 748 | // she isn't in that a stranger gets, or this door becomes |
| 749 | // the one place in the API that confirms an account or an |
| 750 | // app exists. Spec §3 wants both — the operator allowlist |
| 751 | // AND account membership. |
| 752 | // allowlist — 403, naming CARLOS_STORE_OPERATORS. Safe to speak here: |
| 753 | // the caller has already proved membership, so all they |
| 754 | // learn is what authority THEY lack. |
| 755 | // rate limit — the shared write budget, keyed on the ADDRESS. |
| 756 | // Releases nil — 503; the record and the delivered config are both bucket |
| 757 | // objects. |
| 758 | // backend nil — 503, naming CARLOS_STORE_BACKEND: this deployment has no |
| 759 | // object storage at all, which is the one thing the member |
| 760 | // who declared could not see for themselves. |
| 761 | // seal key — 503. Cannot seal → cannot deliver → MUST NOT MINT. |
| 762 | // bucket name — 400 before any backend call. An adopt name is |
| 763 | // interpolated into the IAM policy's resource ARN, so an |
| 764 | // unchecked one is a scope nobody meant to grant. The walk |
| 765 | // validates it too (objstore.ValidBucketName, the same |
| 766 | // function) — this copy exists so the refusal is a clean |
| 767 | // 400 with the name in it rather than a 502. |
| 768 | // |
| 769 | // IDEMPOTENT, because the walk is: re-running over a healthy store is a no-op |
| 770 | // that mints nothing. A re-run over a DRIFTED one — a delivered variable |
| 771 | // unset, a key gone from the backend — repairs it by reaping and re-minting, |
| 772 | // which means the app's key CHANGES. That is the repair working as designed |
| 773 | // (§5), and it is why the client sentence for a failed grant says "re-run to |
| 774 | // resume" rather than promising nothing will move. |
| 775 | func (s *Server) handleAPIStoreGrant(w http.ResponseWriter, r *http.Request) { |
| 776 | addr, acct, app, ok := s.apiAccountApp(w, r) |
| 777 | if !ok { |
| 778 | return |
| 779 | } |
| 780 | if s.refuseTemporary(w, acct, app, "give it object storage") { |
| 781 | return |
| 782 | } |
| 783 | if !s.storeOperator(addr) { |
| 784 | jsonError(w, http.StatusForbidden, |
| 785 | "store grants need a deployment operator (CARLOS_STORE_OPERATORS)") |
| 786 | return |
| 787 | } |
| 788 | p, _ := principalFrom(r.Context()) |
| 789 | if !s.storeWalkGates(w, p) { |
| 790 | return |
| 791 | } |
| 792 | pub, ok := s.storeSealKey(w) |
| 793 | if !ok { |
| 794 | return |
| 795 | } |
| 796 | // The body carries one optional field: a bucket to ADOPT instead of |
| 797 | // minting. `{}` is the ordinary request, as on the create door. |
| 798 | var req struct { |
| 799 | Bucket string `json:"bucket"` |
| 800 | } |
| 801 | if !readJSON(w, r, &req) { |
| 802 | return |
| 803 | } |
| 804 | bucket := strings.TrimSpace(req.Bucket) |
| 805 | if bucket != "" && !objstore.ValidBucketName(bucket) { |
| 806 | jsonError(w, http.StatusBadRequest, |
| 807 | fmt.Sprintf("invalid bucket name %q — want 3-63 characters of lowercase letters, digits, dots and hyphens, starting and ending alphanumeric", req.Bucket)) |
| 808 | return |
| 809 | } |
| 810 | env, ok := s.storeEnvParam(w, r) |
| 811 | if !ok { |
| 812 | return |
| 813 | } |
| 814 | ctx, cancel := context.WithTimeout(r.Context(), storeGrantTimeout) |
| 815 | defer cancel() |
| 816 | defer s.storeWalks.lock(acct.Sqid, app, env)() |
| 817 | rec, err := claimAppCall(s, ctx, acct, app, "store-grant", func(ctx context.Context) (*objstore.Record, error) { |
| 818 | return objstore.GrantIn(ctx, s.cfg.Releases, s.cfg.StoreProvisioner, s.cfg.StoreDeployment, |
| 819 | pub, acct.Sqid, app, env, addr, objstore.GrantSourceOperator, bucket, time.Now()) |
| 820 | }) |
| 821 | switch { |
| 822 | case errors.Is(err, objstore.ErrNotDeclared): |
| 823 | // The operator's verb never invents the member's half: the env prefix |
| 824 | // is the member's choice and there is nothing to deliver into. |
| 825 | jsonError(w, http.StatusConflict, |
| 826 | "no store is declared for this app. run `carlos store create` first") |
| 827 | return |
| 828 | case errors.Is(err, objstore.ErrStoreNameTaken): |
| 829 | // Refused before anything was created (#218): the bucket, or the IAM |
| 830 | // user this store would be given, is already another store's. |
| 831 | log.Printf("console: api store grant %s/%s (bucket %q): %v", acct.Sqid, app, bucket, err) |
| 832 | jsonError(w, http.StatusConflict, storeNameTakenMsg) |
| 833 | return |
| 834 | case err != nil: |
| 835 | // Logged, not returned: a backend error carries bucket names, ARNs, |
| 836 | // endpoints and request ids. The client sentence promises exactly what |
| 837 | // the walk's ordering guarantees — every resource this walk created is |
| 838 | // named by a record it wrote before creating it, so nothing is stranded |
| 839 | // where only a bill would find it — and the fix is to run the same verb |
| 840 | // again, which resumes from wherever it stopped. |
| 841 | log.Printf("console: api store grant %s/%s (bucket %q): %v", acct.Sqid, app, bucket, err) |
| 842 | jsonError(w, http.StatusBadGateway, |
| 843 | "could not provision the store — nothing unrecorded was left behind; re-run the grant to resume") |
| 844 | return |
| 845 | } |
| 846 | s.audit(acct, app, addr, "store-grant", "granted a store in bucket "+storeBucketName(rec)+storeEnvSuffix(env)+" (cli)") |
| 847 | // The app's config bundle changed: boxes read config only when the serial |
| 848 | // moves, so without this the credentials sit in the bucket until the hourly |
| 849 | // pass and the app keeps failing to reach a store it has been given. |
| 850 | s.bumpSerial(ctx) |
| 851 | s.storeWalkReply(ctx, w, acct, app, env, rec) |
| 852 | } |
| 853 | |
| 854 | // handleAPIStoreRotate answers `carlos store rotate`: mint a fresh key, |
| 855 | // deliver it, and leave the OLD key live until an explicit finish. |
| 856 | // |
| 857 | // A MEMBER VERB, not an operator one (spec §8, and the route table in §10). |
| 858 | // Rotation spends nothing new — it cycles a credential inside an estate an |
| 859 | // operator already granted — and the person who needs it most is a member who |
| 860 | // has just leaked a key at 2am. Making them find a deployment operator first |
| 861 | // would make the safety verb the slow one. The gates are therefore create's, |
| 862 | // plus the two 503s that any minting door owes: a backend to mint in and a key |
| 863 | // to seal with. |
| 864 | // |
| 865 | // The overlap the walk leaves behind is the point: adoption swaps a symlink, |
| 866 | // it does not restart processes, so at the instant new credentials land there |
| 867 | // are still hosts and hibernating tenants holding the old pair. `pending_finish` |
| 868 | // in the reply is how a member sees that window is open. |
| 869 | func (s *Server) handleAPIStoreRotate(w http.ResponseWriter, r *http.Request) { |
| 870 | addr, acct, app, ok := s.apiAccountApp(w, r) |
| 871 | if !ok { |
| 872 | return |
| 873 | } |
| 874 | p, _ := principalFrom(r.Context()) |
| 875 | if !s.storeWalkGates(w, p) { |
| 876 | return |
| 877 | } |
| 878 | pub, ok := s.storeSealKey(w) |
| 879 | if !ok { |
| 880 | return |
| 881 | } |
| 882 | // No body: a rotation has no arguments, and handleAPIRestart — the other |
| 883 | // verb whose whole request is "do the thing to this app" — reads none |
| 884 | // either, so a CLI can POST with nothing at all. |
| 885 | env, ok := s.storeEnvParam(w, r) |
| 886 | if !ok { |
| 887 | return |
| 888 | } |
| 889 | ctx, cancel := context.WithTimeout(r.Context(), storeRotateTimeout) |
| 890 | defer cancel() |
| 891 | defer s.storeWalks.lock(acct.Sqid, app, env)() |
| 892 | rec, err := claimAppCall(s, ctx, acct, app, "store-rotate", func(ctx context.Context) (*objstore.Record, error) { |
| 893 | return objstore.RotateIn(ctx, s.cfg.Releases, s.cfg.StoreProvisioner, pub, acct.Sqid, app, env, time.Now()) |
| 894 | }) |
| 895 | switch { |
| 896 | case errors.Is(err, objstore.ErrNotProvisioned), errors.Is(err, objstore.ErrNotDeclared): |
| 897 | // Rotation edits an existing estate; it never creates one, because the |
| 898 | // coordinates it reseals come from the record and there is no |
| 899 | // defensible guess for them. The repair for this state is a grant. |
| 900 | jsonError(w, http.StatusConflict, |
| 901 | "no provisioned store to rotate — grant one first") |
| 902 | return |
| 903 | case errors.Is(err, release.ErrCASConflict): |
| 904 | // A concurrent write to the record. NOT a failed rotation: by the time |
| 905 | // the walk reaches its final Put it has already minted a key and |
| 906 | // delivered it into the app's config, so telling the caller nothing |
| 907 | // happened would send them looking for a key that is already live. The |
| 908 | // honest instruction is to run it again, which converges. |
| 909 | log.Printf("console: api store rotate %s/%s: %v", acct.Sqid, app, err) |
| 910 | jsonError(w, http.StatusConflict, |
| 911 | "another rotation is in progress — re-run to converge") |
| 912 | return |
| 913 | case err != nil: |
| 914 | log.Printf("console: api store rotate %s/%s: %v", acct.Sqid, app, err) |
| 915 | jsonError(w, http.StatusBadGateway, |
| 916 | "could not rotate the store credential — nothing unrecorded was left behind; re-run the rotation to resume") |
| 917 | return |
| 918 | } |
| 919 | key := "" |
| 920 | if rec.Provisioned != nil { |
| 921 | key = rec.Provisioned.KeyID |
| 922 | } |
| 923 | // Key ids are not secrets — `status` prints them — and the audit row is |
| 924 | // where a member later matches "which credential did we cut over to". |
| 925 | s.audit(acct, app, addr, "store-rotate", "rotated the store credential to key "+key+storeEnvSuffix(env)+" (cli)") |
| 926 | s.bumpSerial(ctx) |
| 927 | s.storeWalkReply(ctx, w, acct, app, env, rec) |
| 928 | } |
| 929 | |
| 930 | // handleAPIStoreRotateFinish answers `carlos store rotate -finish`: delete the |
| 931 | // previous key and stop naming it. The member's explicit "nothing is using it |
| 932 | // any more" — the one destructive step in the rotation state machine, which is |
| 933 | // why it is a verb someone asks for rather than a timer. |
| 934 | // |
| 935 | // TWO DIFFERENCES from its siblings, both deliberate: |
| 936 | // |
| 937 | // It needs NO SEALING KEY. Finishing delivers nothing — the app has been |
| 938 | // reading the current key since the rotation — so a console that has lost its |
| 939 | // CARLOS_SECRETS_PUBKEY can still close a two-key window. Requiring the key |
| 940 | // uniformly would be tidier and would strand exactly the deployment that most |
| 941 | // needs to retire a credential. |
| 942 | // |
| 943 | // It does NOT bump the serial. No object a box reads changed: the delivered |
| 944 | // config is untouched, and the record is console-side. Waking the fleet to |
| 945 | // tell it a key it never held is gone would be pure noise. |
| 946 | func (s *Server) handleAPIStoreRotateFinish(w http.ResponseWriter, r *http.Request) { |
| 947 | addr, acct, app, ok := s.apiAccountApp(w, r) |
| 948 | if !ok { |
| 949 | return |
| 950 | } |
| 951 | p, _ := principalFrom(r.Context()) |
| 952 | if !s.storeWalkGates(w, p) { |
| 953 | return |
| 954 | } |
| 955 | env, ok := s.storeEnvParam(w, r) |
| 956 | if !ok { |
| 957 | return |
| 958 | } |
| 959 | ctx, cancel := context.WithTimeout(r.Context(), storeRotateTimeout) |
| 960 | defer cancel() |
| 961 | defer s.storeWalks.lock(acct.Sqid, app, env)() |
| 962 | rec, err := claimAppCall(s, ctx, acct, app, "store-rotate-finish", func(ctx context.Context) (*objstore.Record, error) { |
| 963 | return objstore.RotateFinishIn(ctx, s.cfg.Releases, s.cfg.StoreProvisioner, acct.Sqid, app, env) |
| 964 | }) |
| 965 | switch { |
| 966 | case errors.Is(err, objstore.ErrNotProvisioned): |
| 967 | jsonError(w, http.StatusConflict, |
| 968 | "no provisioned store to rotate — grant one first") |
| 969 | return |
| 970 | case errors.Is(err, objstore.ErrNoPendingFinish): |
| 971 | // Refused rather than succeeding vacuously: a member who believes they |
| 972 | // just retired the old key, when an earlier finish already did, has a |
| 973 | // different picture of the estate than the estate does. |
| 974 | jsonError(w, http.StatusConflict, "no rotation is pending finish") |
| 975 | return |
| 976 | case err != nil: |
| 977 | log.Printf("console: api store rotate finish %s/%s: %v", acct.Sqid, app, err) |
| 978 | jsonError(w, http.StatusBadGateway, |
| 979 | "could not finish the rotation — the previous key may still be live; re-run to resume") |
| 980 | return |
| 981 | } |
| 982 | key := "" |
| 983 | if rec.Provisioned != nil { |
| 984 | key = rec.Provisioned.KeyID |
| 985 | } |
| 986 | s.audit(acct, app, addr, "store-rotate-finish", |
| 987 | "finished the rotation — the store now runs on key "+key+" alone"+storeEnvSuffix(env)+" (cli)") |
| 988 | // No bumpSerial: see the doc comment. |
| 989 | s.storeWalkReply(ctx, w, acct, app, env, rec) |
| 990 | } |
| 991 | |
amadan