| 1 | // SyncExec is the exec-backend half of config delivery: for every |
| 2 | // activator-spawned route (socket-upstream, non-unit), materialize the |
| 3 | // route's merged manifest into <dataDir>/<host>.config.env — the file |
| 4 | // ExecBackend.Start layers UNDER the instance's own <host>.env at spawn. |
| 5 | // Sync owns this namespace outright (nothing else writes *.config.env), |
| 6 | // which is why an absent manifest REMOVES the file here, where the |
| 7 | // unit-path Sync leaves /etc/carlos/apps files alone: there is no |
| 8 | // hand-staged-file migration hazard by construction. See spec |
| 9 | // 2026-08-02-exec-config-delivery-design.md. |
| 10 | // |
| 11 | // A third eligibility case (issue #18): a TCP-addr exec route (Backing == |
| 12 | // "", Socket == "", Addr or Addrs set) whose BOUND BUNDLE has opted in via |
| 13 | // additive "exec_delivery": "file" manifest field (Manifest.ExecDelivery) |
| 14 | // gets the identical file, materialized with the same layering guarantees, |
| 15 | // environments-awareness, reap discipline, and mass-removal/path-traversal |
| 16 | // guards as a socket-exec route — but is NEVER bounced (see syncExecHost |
| 17 | // and execAddrCandidate): carlos routes to these processes without |
| 18 | // spawning or owning them, so the restart half belongs to the app's own |
| 19 | // supervisor. An un-opted TCP-addr route is skipped exactly as before this |
| 20 | // case existed — see execAddrCandidate and SyncExec's checkOptedIn closure. |
| 21 | // |
| 22 | // "Bound bundle" is issue #195: the flag is read from the route's named |
| 23 | // environment when that bundle STATES a value, and from the app's default |
| 24 | // bundle otherwise. Before that it was app-level only, so opting one |
| 25 | // relay route in wrote a file for every other TCP-addr route of the same |
| 26 | // app — a fleet-wide switch thrown to deliver one service's credential. |
| 27 | // |
| 28 | // Resolution mirrors sync.go's exactly — the group is (Account, App, |
| 29 | // Environment) and a route's bound environment is an overlay on the app's |
| 30 | // default bundle (environments.go, pipelines design §4a). Only the file |
| 31 | // semantics differ, and only where sync's ownership of the namespace |
| 32 | // makes them differ: absence removes rather than skips, and there is no |
| 33 | // key-loss guard (see syncExecHost). |
| 34 | package appconfig |
| 35 | |
| 36 | import ( |
| 37 | "bytes" |
| 38 | "context" |
| 39 | "crypto/ecdh" |
| 40 | "crypto/sha256" |
| 41 | "encoding/hex" |
| 42 | "errors" |
| 43 | "fmt" |
| 44 | "os" |
| 45 | "path/filepath" |
| 46 | "sort" |
| 47 | "strings" |
| 48 | "time" |
| 49 | |
| 50 | "github.com/carlosframework/platform/internal/hostwrite" |
| 51 | "github.com/carlosframework/platform/internal/instancefs" |
| 52 | "github.com/carlosframework/platform/internal/provision" |
| 53 | "github.com/carlosframework/platform/internal/registry" |
| 54 | "github.com/carlosframework/platform/internal/release" |
| 55 | ) |
| 56 | |
| 57 | // configEnvSuffix names sync's file namespace under dataDir. Everything |
| 58 | // with this suffix belongs to sync; nothing else in dataDir does. |
| 59 | const configEnvSuffix = ".config.env" |
| 60 | |
| 61 | // nestedRoute reports whether rt's DB and socket are exactly the nested |
| 62 | // layout's defaults (isolated-exec design §8.1). It is the same two-column |
| 63 | // test activator.Classify makes, spelled out here rather than imported so |
| 64 | // this package need not take a dependency on internal/activator for one |
| 65 | // comparison. |
| 66 | func nestedRoute(dataDir string, rt registry.Route) bool { |
| 67 | return rt.DBPath == provision.NestedDBPath(dataDir, rt.Host) && rt.Socket == provision.NestedSocket(rt.Host) |
| 68 | } |
| 69 | |
| 70 | // ConfigEnvPath is where a route's rendered bundle lands: inside the |
| 71 | // instance directory for a nested row (isolated-exec design §7), flat for |
| 72 | // everything else, including opted-in addr routes whose supervisor reads |
| 73 | // the flat path. |
| 74 | func ConfigEnvPath(dataDir string, rt registry.Route) string { |
| 75 | if nestedRoute(dataDir, rt) { |
| 76 | return filepath.Join(provision.InstanceDir(dataDir, rt.Host), rt.Host+configEnvSuffix) |
| 77 | } |
| 78 | return filepath.Join(dataDir, rt.Host+configEnvSuffix) |
| 79 | } |
| 80 | |
| 81 | // SyncExec resolves config once per distinct (Account, App, Environment) |
| 82 | // among the box's exec-syncable routes and converges every such route's |
| 83 | // <dataDir>/<host>.config.env, calling bounce(host) whenever a file's |
| 84 | // content changed (including removal) so a running instance respawns |
| 85 | // with fresh env on its next wake. Die-free per group, like Sync. |
| 86 | // |
| 87 | // The group carries Environment for the same reason Sync's does |
| 88 | // (pipelines design §4a): two routes of ONE app bound to different |
| 89 | // environments resolve to different manifests, and keying on (acct, app) |
| 90 | // alone would render whichever route was seen first into both files. |
| 91 | // |
| 92 | // ack, if non-nil, is called once per route this pass actually WRITES a |
| 93 | // changed file for, with that route's config-adoption acknowledgement |
| 94 | // (spec §5, configack.go) — nil means no acks, same contract as Sync's. A |
| 95 | // route whose file is REMOVED (manifest absent) gets no ack: there is |
| 96 | // nothing for a route to have adopted once its config is gone, and a |
| 97 | // removal is not one of spec §5's three delivery shapes. |
| 98 | // |
| 99 | // opts adjust one pass; see SkipFenced. |
| 100 | func SyncExec(ctx context.Context, reg *registry.Registry, s release.Store, priv *ecdh.PrivateKey, dataDir string, bounce func(host string) error, appsDomain string, ack func(ConfigAck), opts ...SyncOption) error { |
| 101 | return SyncExecWithCallbacks(ctx, reg, s, priv, dataDir, appsDomain, Callbacks{Restart: bounce, Ack: ack}, opts...) |
| 102 | } |
| 103 | |
| 104 | func SyncExecWithCallbacks(ctx context.Context, reg *registry.Registry, s release.Store, priv *ecdh.PrivateKey, dataDir, appsDomain string, callbacks Callbacks, opts ...SyncOption) error { |
| 105 | var o syncOptions |
| 106 | for _, opt := range opts { |
| 107 | opt(&o) |
| 108 | } |
| 109 | if o.hostWriteCount > 1 { |
| 110 | return errHostWriteOptions |
| 111 | } |
| 112 | hostWrites := o.hostWrites |
| 113 | bounce, ack := callbacks.Restart, callbacks.Ack |
| 114 | if hostWrites.Enabled && bounce == nil && callbacks.SelectedRestart != nil { |
| 115 | bounce = func(string) error { return hostwrite.ErrUnavailable } |
| 116 | } |
| 117 | routes, err := reg.List() |
| 118 | if err != nil { |
| 119 | return err |
| 120 | } |
| 121 | callbacks.enabled = hostWrites.Enabled |
| 122 | selected := map[string]hostwrite.Work{} |
| 123 | for _, rt := range routes { |
| 124 | if !execSyncable(rt) && !execSidecarCandidate(rt) && !execAddrCandidate(rt) { |
| 125 | continue |
| 126 | } |
| 127 | if !o.kept(rt) { |
| 128 | continue |
| 129 | } |
| 130 | if o.fenced != nil && o.fenced(rt.Host) { |
| 131 | continue |
| 132 | } |
| 133 | work, selectErr := hostWrites.Select(ctx, rt.Host, "config") |
| 134 | if selectErr != nil { |
| 135 | return fmt.Errorf("%s: select config work: %w", rt.Host, selectErr) |
| 136 | } |
| 137 | selected[rt.Host] = work |
| 138 | } |
| 139 | // hostSet names every route on this box, in or out of scope, so the |
| 140 | // namespace-collision guard below sees a route x.config even though |
| 141 | // x.config itself may never reach syncExecHost. Built once: the guard |
| 142 | // runs once per in-scope route in the materialization loop, and every |
| 143 | // route's Host is fixed for the pass. |
| 144 | hostSet := make(map[string]bool, len(routes)) |
| 145 | for _, rt := range routes { |
| 146 | hostSet[rt.Host] = true |
| 147 | } |
| 148 | type groupKey struct{ acct, app, env string } |
| 149 | // The same three states Sync resolves to, with this path's own file |
| 150 | // semantics for the middle one: a manifest to converge to; "never |
| 151 | // configured at all" (m == nil, err == nil), which REMOVES the file |
| 152 | // here because sync owns the *.config.env namespace outright (nothing |
| 153 | // hand-staged can live in it); and "resolution failed" (err != nil), |
| 154 | // which leaves the file exactly as it is and says so, loudly, per |
| 155 | // route. An error is never absence — folding the two together would |
| 156 | // delete a live route's config on a bucket blip or a deleted bundle. |
| 157 | type resolution struct { |
| 158 | m *Manifest |
| 159 | err error |
| 160 | } |
| 161 | resolved := map[groupKey]resolution{} |
| 162 | // The default bundle is read once per (acct, app) and reused across |
| 163 | // that app's environments — same cache, same reason, as Sync's. |
| 164 | type appKey struct{ acct, app string } |
| 165 | type defaultRes struct { |
| 166 | m *Manifest |
| 167 | found bool |
| 168 | err error |
| 169 | } |
| 170 | defaults := map[appKey]defaultRes{} |
| 171 | // getIfExists, not Get: an app with no manifest object in the bucket |
| 172 | // at all must be distinguishable from one whose every var was unset. |
| 173 | // Get folds those together into an empty manifest; here the first |
| 174 | // removes the route's config file and the second converges it to an |
| 175 | // empty one. |
| 176 | loadDefault := func(acct, app string) defaultRes { |
| 177 | ak := appKey{acct, app} |
| 178 | if d, done := defaults[ak]; done { |
| 179 | return d |
| 180 | } |
| 181 | m, found, err := getIfExists(ctx, s, acct, app) |
| 182 | if err != nil { |
| 183 | err = fmt.Errorf("%s/%s: config fetch failed: %w", acct, app, err) |
| 184 | } |
| 185 | d := defaultRes{m: m, found: found, err: err} |
| 186 | defaults[ak] = d |
| 187 | return d |
| 188 | } |
| 189 | // optedInRes is checkOptedIn's cached verdict. known is false ONLY for |
| 190 | // a fetch error: v is a placeholder in that case, never a |
| 191 | // determination — see checkOptedIn's def.err branch for why that |
| 192 | // distinction matters to the stale-file notice below. |
| 193 | type optedInRes struct { |
| 194 | v bool |
| 195 | known bool |
| 196 | } |
| 197 | // optKey keys the opt-in decision on the BUNDLE a route's config |
| 198 | // actually resolves through, not on the app alone (issue #195). Two |
| 199 | // routes of one app bound to different environments can now differ, so |
| 200 | // the cache — and the once-per-pass logs below — must too. |
| 201 | type optKey struct{ acct, app, env string } |
| 202 | // optedIn caches the exec_delivery decision so a box with many TCP-addr |
| 203 | // routes on one bundle doesn't cost more than one manifest read for it |
| 204 | // (loadDefault/loadEnvBundle already dedupe the fetches; this dedupes |
| 205 | // the value-interpretation, including the logs below, which must fire |
| 206 | // at most once per (acct, app, env) per pass, not once per route). |
| 207 | optedIn := map[optKey]optedInRes{} |
| 208 | // envBundles caches the named-bundle read the same way loadDefault |
| 209 | // caches the default's, and for the same reason. |
| 210 | type envRes struct { |
| 211 | m *Manifest |
| 212 | found bool |
| 213 | err error |
| 214 | } |
| 215 | envBundles := map[optKey]envRes{} |
| 216 | loadEnvBundle := func(acct, app, env string) envRes { |
| 217 | k := optKey{acct, app, env} |
| 218 | if e, done := envBundles[k]; done { |
| 219 | return e |
| 220 | } |
| 221 | m, found, err := getIfExistsAt(ctx, s, EnvironmentKey(acct, app, env)) |
| 222 | if err != nil { |
| 223 | err = fmt.Errorf("%s/%s (environment %q): config fetch failed: %w", acct, app, env, err) |
| 224 | } |
| 225 | e := envRes{m: m, found: found, err: err} |
| 226 | envBundles[k] = e |
| 227 | return e |
| 228 | } |
| 229 | checkOptedIn := func(acct, app, env string) optedInRes { |
| 230 | k := optKey{acct, app, env} |
| 231 | if c, done := optedIn[k]; done { |
| 232 | return c |
| 233 | } |
| 234 | // The route's OWN bundle first, and only when it STATES something: |
| 235 | // a named bundle silent on the field inherits the app default, |
| 236 | // exactly as a key it does not set inherits the default bundle's |
| 237 | // value. That PRECEDENCE now lives in resolveExecDelivery, which |
| 238 | // AwaitingRestart resolves through too, so the notice an operator |
| 239 | // reads and the file this sweep writes can never disagree about |
| 240 | // who is opted in. What stays here is what THIS caller owes the |
| 241 | // two non-verdict outcomes: a loud line in the sweep's log, once |
| 242 | // per bundle per pass, and every file left exactly as it was. |
| 243 | var eb envRes |
| 244 | if env != "" { |
| 245 | eb = loadEnvBundle(acct, app, env) |
| 246 | if eb.err != nil { |
| 247 | // NOT a determination — the same contract the default |
| 248 | // bundle's fetch-failure branch below gives: say so loudly, |
| 249 | // leave every file alone, retry next pass. |
| 250 | fmt.Printf(" %s/%s: exec_delivery unresolved for environment %q (config fetch failed), leaving this bundle's TCP-addr route file(s) as-is: %v\n", acct, app, env, eb.err) |
| 251 | c := optedInRes{v: false, known: false} |
| 252 | optedIn[k] = c |
| 253 | return c |
| 254 | } |
| 255 | // found-but-silent, and absent, both fall through to the app |
| 256 | // default inside resolveExecDelivery. A bundle bound but |
| 257 | // MISSING is a defect the route's own resolution reports |
| 258 | // (LoadMerged refuses it) and its file is left as-is there — |
| 259 | // the verdict reached here never gets to a write in that case. |
| 260 | } |
| 261 | v, err := resolveExecDelivery(eb.m, eb.found, func() (*Manifest, bool, error) { |
| 262 | d := loadDefault(acct, app) |
| 263 | return d.m, d.found, d.err |
| 264 | }) |
| 265 | if err != nil { |
| 266 | // NOT a determination: this app may well be opted in and this |
| 267 | // box just failed to see it this pass — indistinguishable here |
| 268 | // from any other route's ordinary fetch-failure blip. Silence |
| 269 | // would mean an opted-in app's delivery can stall with zero |
| 270 | // trace in the log, on the exact path this feature exists to |
| 271 | // light up, so say so loudly — the same "unresolved, left |
| 272 | // as-is" contract the resolution loop below already gives a |
| 273 | // bound-group fetch failure. |
| 274 | // |
| 275 | // It is safe to leave any existing file exactly as it is |
| 276 | // regardless of how long this persists: NOT because a fetch |
| 277 | // failure and "never opted in" are indistinguishable (they |
| 278 | // aren't, that's the whole point of this branch existing |
| 279 | // separately from the ones below) but because reapConfigEnv |
| 280 | // keys on ROUTE EXISTENCE, never on scope (its live map is |
| 281 | // built from every route's Host, full stop — see its own doc |
| 282 | // comment). A route that drops out of scope for one pass, for |
| 283 | // ANY reason, is simply not visited by the resolution loop, |
| 284 | // the materialization loop, or reap; its file is untouched by |
| 285 | // construction, not because this function decided to spare |
| 286 | // it. Next pass retries from scratch. |
| 287 | fmt.Printf(" %s/%s: exec_delivery unresolved (config fetch failed), leaving this app's TCP-addr route file(s) as-is: %v\n", acct, app, err) |
| 288 | c := optedInRes{v: false, known: false} |
| 289 | optedIn[k] = c |
| 290 | return c |
| 291 | } |
| 292 | if v.Unknown != "" { |
| 293 | // A manifest a NEWER carlos wrote with a value THIS binary |
| 294 | // doesn't understand must not brick this box's sync (additive |
| 295 | // discipline, AGENTS.md) — ignore it, loudly, once, and treat |
| 296 | // the bundle as not opted in rather than refusing the whole |
| 297 | // pass. Which bundle said it is named, because "upgrade |
| 298 | // carlos" is useless advice without knowing where to look. |
| 299 | where := "the default bundle" |
| 300 | if v.FromEnvironment { |
| 301 | where = fmt.Sprintf("environment %q", env) |
| 302 | } |
| 303 | fmt.Printf(" %s/%s: config: exec_delivery %q in %s is not a value this binary understands — ignoring it (not opted in); upgrade carlos if this is expected\n", acct, app, v.Unknown, where) |
| 304 | } |
| 305 | c := optedInRes{v: v.On, known: true} |
| 306 | optedIn[k] = c |
| 307 | return c |
| 308 | } |
| 309 | // inScope is the single door both loops below use to decide whether a |
| 310 | // route is in scope this pass, and — for the addr and sidecar shapes — |
| 311 | // that it must never be bounced. One function so the two loops |
| 312 | // structurally cannot diverge on either question (see the "if they |
| 313 | // ever diverge" comment on the second loop's map lookup, which depends |
| 314 | // on the same key having been computed the same way here). |
| 315 | // |
| 316 | // sidecarShape never bounces for a DIFFERENT reason than addrShape |
| 317 | // does: carlos owns a sidecar's process (unlike the addr shape, whose |
| 318 | // app-supervised process carlos never touches), but Bounce alone has |
| 319 | // no way back for a worker — nothing dials one the way a request |
| 320 | // dials an instance, so an automatic bounce here would either strand |
| 321 | // an always-on sidecar until the agent restarts or leave a |
| 322 | // hibernating one silent until its schedule next comes due, on every |
| 323 | // ordinary config-sync pass. Restarting a sidecar on purpose is |
| 324 | // `carlos restart -app X` (host/restart.go), which pairs the stop |
| 325 | // with an explicit wake; this pass only ever writes the file. |
| 326 | inScope := func(rt registry.Route) (ok, addrShape, sidecarShape bool) { |
| 327 | // A route WithRouteFilter leaves out is out of scope before |
| 328 | // anything is read for it: no bundle, no opt-in, no overlay. |
| 329 | if !o.kept(rt) { |
| 330 | return false, false, false |
| 331 | } |
| 332 | if execSyncable(rt) { |
| 333 | return true, false, false |
| 334 | } |
| 335 | if execSidecarCandidate(rt) { |
| 336 | return true, false, true |
| 337 | } |
| 338 | if execAddrCandidate(rt) && checkOptedIn(release.RouteAccount(rt), rt.App, rt.Environment).v { |
| 339 | return true, true, false |
| 340 | } |
| 341 | return false, false, false |
| 342 | } |
| 343 | for _, rt := range routes { |
| 344 | if in, _, _ := inScope(rt); !in { |
| 345 | continue |
| 346 | } |
| 347 | k := groupKey{release.RouteAccount(rt), rt.App, rt.Environment} |
| 348 | if _, done := resolved[k]; done { |
| 349 | continue |
| 350 | } |
| 351 | def := loadDefault(k.acct, k.app) |
| 352 | if def.err != nil { |
| 353 | // A fetch failure is NOT "absent": leave this app's files |
| 354 | // alone this pass rather than removing config on a blip. |
| 355 | resolved[k] = resolution{err: def.err} |
| 356 | continue |
| 357 | } |
| 358 | if k.env == "" { |
| 359 | if !def.found { |
| 360 | resolved[k] = resolution{} // absent -> remove |
| 361 | continue |
| 362 | } |
| 363 | resolved[k] = resolution{m: def.m} |
| 364 | continue |
| 365 | } |
| 366 | // A bound environment asserts the bundle exists, so LoadMerged |
| 367 | // treats its absence as an error rather than an empty overlay — |
| 368 | // and that error stays contained to the routes bound to it. There |
| 369 | // is no never-configured case here: an app can lack a default |
| 370 | // bundle and still resolve (environment-only), and a missing |
| 371 | // environment is the error above, not silence. Critically for |
| 372 | // THIS path, that error must not read as absence: removing the |
| 373 | // file would strip a live instance's config because someone |
| 374 | // deleted a bundle. |
| 375 | m, err := LoadMerged(ctx, s, k.acct, k.app, k.env) |
| 376 | if err != nil || m == nil { |
| 377 | if err == nil { |
| 378 | err = fmt.Errorf("config: resolved a nil manifest for environment %q", k.env) |
| 379 | } |
| 380 | resolved[k] = resolution{err: err} |
| 381 | continue |
| 382 | } |
| 383 | resolved[k] = resolution{m: m} |
| 384 | } |
| 385 | |
| 386 | // Mass-removal misconfiguration guard. A registry rebuild that mangles |
| 387 | // the account column makes release.RouteAccount fall back to "ops" for |
| 388 | // every app at once, so every group's bucket key 404s and every group |
| 389 | // in this pass looks "absent" — indistinguishable, from here, from |
| 390 | // every manifest object having genuinely been deleted. An unguarded |
| 391 | // pass would then remove every config file on the box in one sweep. |
| 392 | // Require MORE THAN ONE group before refusing: a single-app box (a |
| 393 | // self-hoster who really did delete their one manifest) must still |
| 394 | // converge normally, and a lone group can't produce a false "mass" |
| 395 | // signal — only a registry-wide fault spanning multiple apps can look |
| 396 | // like this. |
| 397 | // |
| 398 | // Groups in the ERROR state are counted on neither side: they are |
| 399 | // neither absent nor present, they never remove anything anyway, and |
| 400 | // letting one sit in the denominator would silently disarm the guard |
| 401 | // on a box where one app's bundle happens to be broken. That choice |
| 402 | // has a direction: a PARTIALLY errored mangled box can still remove |
| 403 | // the remaining absent groups' files, which counting errors as absent |
| 404 | // would have prevented — but counting them as absent would also freeze |
| 405 | // legitimate removals for as long as any one app's bundle stays |
| 406 | // broken, indefinitely and silently. Bounded, recoverable over-removal |
| 407 | // beats an unbounded silent freeze; routes bound to an environment are |
| 408 | // protected by the error path regardless, since an unresolved group |
| 409 | // never reaches a removal at all. |
| 410 | // |
| 411 | // The instance layer does not enter this count, and does not need to. |
| 412 | // The fault this guard exists for is registry-wide — a mangled account |
| 413 | // column makes EVERY key of every app 404 at once, the overlay listing |
| 414 | // included, since it hangs off the same mangled prefix — so an overlay |
| 415 | // could not survive the fault it is counting. A route that does have an |
| 416 | // overlay never reaches a removal branch at all (see below), so it is |
| 417 | // neither protected nor endangered by this number. |
| 418 | eligible, absent := 0, 0 |
| 419 | for _, r := range resolved { |
| 420 | if r.err != nil { |
| 421 | continue |
| 422 | } |
| 423 | eligible++ |
| 424 | if r.m == nil { |
| 425 | absent++ |
| 426 | } |
| 427 | } |
| 428 | skipRemovals := eligible > 1 && absent == eligible |
| 429 | if skipRemovals { |
| 430 | // Name what was actually checked — "every app" would overclaim, |
| 431 | // since errored groups were never in this count. |
| 432 | fmt.Printf(" config sync: all %d resolvable (non-errored) config groups on this box resolved absent — refusing to remove any config files this pass; check the registry's account column and the deployment bucket\n", eligible) |
| 433 | } |
| 434 | |
| 435 | // The instance layer (senders design §6.2), resolved per route after the |
| 436 | // group base. Same loader, same cost, same failure contract as the unit |
| 437 | // pass's: one listing per app, a Get only for a host it named, and an |
| 438 | // unresolved overlay leaves the file exactly as it is. |
| 439 | overlayFor := instanceOverlayLoader(ctx, s) |
| 440 | // One line per app when the LISTING fails, not one per route: every |
| 441 | // route of that app fails identically and for the same reason, since |
| 442 | // instanceOverlayLoader caches that failure per (account, app) — |
| 443 | // independent of Environment, which is why this is keyed on appKey and |
| 444 | // not groupKey. A per-host Get or parse failure is a DIFFERENT claim — |
| 445 | // two routes can have two separately corrupt overlay objects, each |
| 446 | // wrong for its own reason — so that case is never folded in here; see |
| 447 | // the errors.As branch below. |
| 448 | said := map[appKey]bool{} |
| 449 | |
| 450 | // done is WithHostOutcome's answer per in-scope host: false until the |
| 451 | // host's file is as this pass resolved it. |
| 452 | done := map[string]bool{} |
| 453 | for _, rt := range routes { |
| 454 | in, addrShape, sidecarShape := inScope(rt) |
| 455 | if !in { |
| 456 | continue |
| 457 | } |
| 458 | done[rt.Host] = false |
| 459 | // A host the layout pass holds has its files between two |
| 460 | // directories, and either path this pass could pick would be the |
| 461 | // wrong one by the time the pass finishes. Nothing is written or |
| 462 | // removed for it; the next pass after the move renders it where the |
| 463 | // row then says. |
| 464 | if _, ok := selected[rt.Host]; !ok { |
| 465 | fmt.Printf(" %s: layout migration in flight; bundle not written this pass\n", rt.Host) |
| 466 | continue |
| 467 | } |
| 468 | k := groupKey{release.RouteAccount(rt), rt.App, rt.Environment} |
| 469 | // Insist the group was really resolved rather than taking the map's |
| 470 | // zero value: a zero resolution reads as "absent", which on this |
| 471 | // path REMOVES the file. The two loops filter through the same |
| 472 | // inScope today, so this cannot fire — but if they ever diverge, |
| 473 | // this turns a silent config deletion into a skip. |
| 474 | r, ok := resolved[k] |
| 475 | if !ok { |
| 476 | continue |
| 477 | } |
| 478 | if r.err != nil { |
| 479 | // Loud, per route, every pass: this route is running on a |
| 480 | // stale config file until an operator fixes it. Its file is |
| 481 | // left exactly as it is — nothing below this line runs for it. |
| 482 | env := k.env |
| 483 | if env == "" { |
| 484 | env = "default" |
| 485 | } |
| 486 | fmt.Printf(" %s: config unresolved (environment %q), config file left as-is: %v\n", rt.Host, env, r.err) |
| 487 | continue |
| 488 | } |
| 489 | // Fetched here, used at the bottom, because a failure must be able to |
| 490 | // stop a REMOVAL too: deciding a file is orphaned while unable to see |
| 491 | // whether this instance has an overlay would delete a live tenant's |
| 492 | // delivered secrets on a listing blip. |
| 493 | inst, instETag, instErr := overlayFor(k.acct, k.app, rt.Host) |
| 494 | if instErr != nil { |
| 495 | var listErr *instanceOverlayListErr |
| 496 | if errors.As(instErr, &listErr) { |
| 497 | // Said ONCE per (account, app), not once per route: the |
| 498 | // listing this failed on is cached per app inside the |
| 499 | // loader, so every route of the app, across every |
| 500 | // environment it is bound to, fails identically and for |
| 501 | // the same reason; a line each would bury the cause under |
| 502 | // its own repetitions on a box with many routes. |
| 503 | ak := appKey{k.acct, k.app} |
| 504 | if !said[ak] { |
| 505 | said[ak] = true |
| 506 | // rt.Host anchors the line to a real route (the first |
| 507 | // one this pass happened to reach), matching sync.go's |
| 508 | // own wording, so an operator scanning the log has a |
| 509 | // hostname to grep for — every other route of this app |
| 510 | // is silently covered by the same line. |
| 511 | fmt.Printf(" %s/%s: instance config unresolved (seen at %s), config files left as-is: %v\n", k.acct, k.app, rt.Host, instErr) |
| 512 | } |
| 513 | } else { |
| 514 | // A per-host Get or parse failure, unlike the listing |
| 515 | // above, is specific to THIS route: the listing already |
| 516 | // succeeded (it named this host), so another route of |
| 517 | // the same app can have a perfectly good overlay, or a |
| 518 | // failure of its own for a different reason. Folding this |
| 519 | // into the app-level line would leave that route's stale |
| 520 | // config file invisible, so it is reported every time. |
| 521 | fmt.Printf(" %s: instance config unresolved, config file left as-is: %v\n", rt.Host, instErr) |
| 522 | } |
| 523 | continue |
| 524 | } |
| 525 | // An overlay is configuration, so a route that has one is never |
| 526 | // "absent" for the MASS-REMOVAL guard below, however its group base |
| 527 | // resolved: writing a file a bundle asked for is not a removal. |
| 528 | // |
| 529 | // This narrowing (the `inst == nil` clause, added by this branch) |
| 530 | // makes the guard itself strictly narrower than it was before: a |
| 531 | // route with an overlay now sails through where it used to be held |
| 532 | // with everything else. That is backstopped, everywhere this |
| 533 | // narrowing actually matters (r.m == nil, i.e. a mangled account |
| 534 | // column made the group base read absent), by the #344 guard just |
| 535 | // below — EXCEPT for a custom-domain route. release.HostAccountLabel |
| 536 | // only derives a <sqid> to check from a PLATFORM-domain hostname; a |
| 537 | // custom-domain host's label is "", the #344 check's `sq != ""` test |
| 538 | // is false, and the guard simply does not fire. So a custom-domain |
| 539 | // route with an overlay, on a box whose account column is mangled in |
| 540 | // the exact way #344 exists to catch, has no backstop here at all — |
| 541 | // it sails through this guard AND past #344, and the wholesale |
| 542 | // overlay-alone replacement both guards were meant to prevent for it |
| 543 | // can still happen. Known and accepted (final review, plan 4): the |
| 544 | // fix would need a host-independent way to validate the account |
| 545 | // column, which #344 does not have either. |
| 546 | if r.m == nil && inst == nil && skipRemovals { |
| 547 | continue |
| 548 | } |
| 549 | // The #344 guard: an absent manifest removes the config file ONLY |
| 550 | // when the account it was looked up under agrees with the account |
| 551 | // the host itself names. A platform host's <sqid> label is not a |
| 552 | // moving part; the row's account column is (the domains reconcile |
| 553 | // stamps it from bucket listings, and a mis-stamp deleted a live |
| 554 | // tenant's delivered secrets on 2026-08-28). Held, loudly, like an |
| 555 | // unresolved group — the fix is the row's account, not this file. |
| 556 | // |
| 557 | // UNLIKE the mass-removal guard above, this hold is NOT narrowed by |
| 558 | // an overlay (review finding, plan 4 task 7 fix round 1: the brief's |
| 559 | // own reasoning — "a route with an overlay never reaches a removal |
| 560 | // branch anyway" — does not transfer here, because this is not a |
| 561 | // removal branch. It is "this account column cannot be trusted, |
| 562 | // touch nothing." A mis-stamp that makes the group's default-bundle |
| 563 | // READ land under the wrong prefix makes the instance layer's own |
| 564 | // WRITE (ApplyInstance) land under that same wrong prefix too, so an |
| 565 | // overlay resolving here under the mis-stamped account is exactly |
| 566 | // what a mis-stamped fleet produces — not evidence the account |
| 567 | // column is safe to act on for this route. Wholesale-replacing the |
| 568 | // file with that overlay's contents would be the 2026-08-28 failure |
| 569 | // again, just reached through the overlay instead of the base. |
| 570 | if r.m == nil { |
| 571 | if sq := release.HostAccountLabel(rt.Host, appsDomain); sq != "" && sq != k.acct { |
| 572 | fmt.Printf(" %s: no config manifest under account %q, but the host names account %q — config file left as-is; fix the route row's account column (#344)\n", |
| 573 | rt.Host, k.acct, sq) |
| 574 | continue |
| 575 | } |
| 576 | } |
| 577 | // Namespace-collision guard, checked before EITHER shape can reach |
| 578 | // syncExecHost below (the single call site both shapes share): if a |
| 579 | // route "<rt.Host>.config" also lives on this box, its hand-staged |
| 580 | // instance env (<host>.env for THAT route) is named |
| 581 | // "<rt.Host>.config.env" — byte-identical to what converging rt |
| 582 | // would write here. Refuse loudly rather than clobber another |
| 583 | // owner's file; see reapConfigEnv's twin guard for the |
| 584 | // deletion-side half of the same hazard. |
| 585 | if hostSet[rt.Host+".config"] { |
| 586 | fmt.Printf(" %s: not converging — route %q exists on this box and %s%s is its instance-env file (<host>.env); refusing to write into another owner's namespace\n", rt.Host, rt.Host+".config", rt.Host, configEnvSuffix) |
| 587 | continue |
| 588 | } |
| 589 | // The bucket key whose absence would drive a removal, named in the |
| 590 | // removal log so cause and effect land on one line. Always the |
| 591 | // DEFAULT bundle's key: removal is reachable only from the absent |
| 592 | // state, and only an unbound group can resolve absent (a bound one |
| 593 | // resolves through LoadMerged, which errors on a missing bundle |
| 594 | // rather than returning nothing). A bound group's key would be |
| 595 | // dead in every branch that reads this. |
| 596 | // |
| 597 | // addrShape routes get nil in bounce's place, not the real |
| 598 | // closure with a runtime check skipping the call: carlos does not |
| 599 | // own an app-supervised process, so there is structurally nothing |
| 600 | // to hand syncExecHost that COULD invoke it for this route (issue |
| 601 | // #18 design comment) — bounceLogged's existing nil-is-a-no-op |
| 602 | // contract does the rest. Same shape for the header: addrShape |
| 603 | // routes get execAddrConfigHeader, socket routes get "" (no |
| 604 | // header at all) — see execAddrConfigHeader's doc comment for why |
| 605 | // this must never apply to the socket shape. |
| 606 | // |
| 607 | // sidecarShape ALSO gets nil in bounce's place, but for the |
| 608 | // opposite reason addrShape does: carlos owns this process, but |
| 609 | // Bounce cannot safely stand in for a wake here — see inScope's |
| 610 | // doc comment. No header either; a sidecar's file is read by |
| 611 | // ExecBackend.instanceEnv, the same reader an instance's is. |
| 612 | hostBounce := bounce |
| 613 | header := "" |
| 614 | switch { |
| 615 | case addrShape: |
| 616 | hostBounce = nil |
| 617 | header = execAddrConfigHeader |
| 618 | case sidecarShape: |
| 619 | hostBounce = nil |
| 620 | } |
| 621 | // The group's merge, overlaid by this instance's own layer. A nil |
| 622 | // base with an overlay is the overlay ALONE, which is a real |
| 623 | // configuration and not the absence that drives a removal; nil and |
| 624 | // nil together stay nil, which still removes the file. |
| 625 | m := r.m |
| 626 | if m != nil || inst != nil { |
| 627 | m = applyInstanceOverlay(k.app, m, inst) |
| 628 | } |
| 629 | hostCallbacks := callbacks |
| 630 | // With Managed set (the release scheduler is armed or refusing, |
| 631 | // final review F7), the scheduler's own routes and the rest part |
| 632 | // here. A socket route it does not manage is restarted as it is |
| 633 | // with the flag off, by Unmanaged, and its delivery is none of the |
| 634 | // scheduler's business. A native sidecar it manages keeps Delivered |
| 635 | // below though config sync has no restart for it: the scheduler's |
| 636 | // config attempt is how its new file reaches a running worker. |
| 637 | managed := callbacks.Managed != nil && callbacks.Managed(rt) |
| 638 | if callbacks.Managed != nil && hostBounce != nil && !managed { |
| 639 | hostBounce = callbacks.Unmanaged |
| 640 | hostCallbacks.Restart = callbacks.Unmanaged |
| 641 | hostCallbacks.Delivered = nil |
| 642 | } |
| 643 | if hostBounce == nil { |
| 644 | hostCallbacks.Restart = nil |
| 645 | hostCallbacks.SelectedRestart = nil |
| 646 | if !sidecarShape || !managed { |
| 647 | hostCallbacks.Delivered = nil |
| 648 | } |
| 649 | } |
| 650 | err := syncExecHostClaimed(ctx, selected[rt.Host], rt, m, Key(k.acct, k.app), priv, dataDir, hostBounce, header, instETag, ack, hostCallbacks) |
| 651 | // A nested row's bundle lives in its instance directory, which the |
| 652 | // activator makes at the first wake. Before then the write cannot |
| 653 | // land, so the row is named and the pass goes on to the other |
| 654 | // routes, claims on or off; the agent runs the pass again once the |
| 655 | // directory is there (WhenInstanceDirAbsent). A render that failed |
| 656 | // fails as before, directory or not. |
| 657 | if errors.Is(err, errInstanceDirAbsent) { |
| 658 | fmt.Printf(" %s: instance directory not made yet; config is written once it is\n", rt.Host) |
| 659 | if o.dirAbsent != nil { |
| 660 | o.dirAbsent(rt.Host, filepath.Dir(ConfigEnvPath(dataDir, rt))) |
| 661 | } |
| 662 | continue |
| 663 | } |
| 664 | if err != nil && hostWrites.Enabled { |
| 665 | return err |
| 666 | } |
| 667 | if err == nil { |
| 668 | done[rt.Host] = true |
| 669 | } |
| 670 | } |
| 671 | if o.outcome != nil { |
| 672 | for host, ok := range done { |
| 673 | o.outcome(host, ok) |
| 674 | } |
| 675 | } |
| 676 | // Stale-secrets notice (issue #18 design ruling, ratified 2026-08-03 |
| 677 | // review): an addr-shaped route whose app is DEFINITIVELY not opted in |
| 678 | // this pass (off, never configured, or an unrecognized exec_delivery |
| 679 | // value — checkOptedIn's known == true, as opposed to a transient |
| 680 | // fetch error) but still has a route row keeps any earlier |
| 681 | // <host>.config.env EXACTLY as it is: it was never added to |
| 682 | // inScope/resolved above, so neither the resolution loop nor the |
| 683 | // materialization loop touched it, and reap (below) never removes a |
| 684 | // file whose route still has a row. |
| 685 | // |
| 686 | // RULING: do not auto-delete it. Removing the file would change what |
| 687 | // the app's OWN supervisor layers in on its next respawn — a live |
| 688 | // environment change caused by a flag flip, taking effect at a moment |
| 689 | // carlos does not control and cannot announce, with no bounce (there |
| 690 | // never is one for this shape) to make it observable. That is worse |
| 691 | // than a stale file. It is also consistent with every other |
| 692 | // ineligibility transition in this file: a route losing |
| 693 | // Status: "ready" or gaining Platform: true doesn't get its file |
| 694 | // deleted either, only stops being written to. |
| 695 | // |
| 696 | // BUT staying silent about it is wrong, more than for those other |
| 697 | // transitions, because this file holds DECRYPTED secrets, and once an |
| 698 | // app opts out nothing will EVER update it again: rotate a secret in |
| 699 | // the bucket afterward and this file keeps serving yesterday's |
| 700 | // plaintext, indefinitely, with carlos reporting nothing. "Off" reads |
| 701 | // to an operator as "stop delivering config"; what it actually does |
| 702 | // is "freeze whatever was last delivered, forever". Name the file and |
| 703 | // the manual remedy, once per pass, so an operator doesn't mistake |
| 704 | // the one for the other — the remedy has to be manual (delete it once |
| 705 | // the app's own supervisor has stopped reading it) because carlos has |
| 706 | // no way to know when that is. |
| 707 | for _, rt := range routes { |
| 708 | if !execAddrCandidate(rt) || !o.kept(rt) { |
| 709 | continue |
| 710 | } |
| 711 | c := checkOptedIn(release.RouteAccount(rt), rt.App, rt.Environment) |
| 712 | if c.v || !c.known { |
| 713 | continue |
| 714 | } |
| 715 | path := ConfigEnvPath(dataDir, rt) |
| 716 | if _, err := os.Stat(path); err != nil { |
| 717 | continue // nothing was ever staged for this route: nothing to warn about |
| 718 | } |
| 719 | fmt.Printf(" %s: exec_delivery is off (or was never set) but %s still exists from an earlier opt-in — it holds DECRYPTED secrets and carlos will never update or remove it while this app stays un-opted-in; delete it by hand once the app's own supervisor has stopped reading it\n", rt.Host, path) |
| 720 | } |
| 721 | // An empty route list is never authority to delete anything — |
| 722 | // registry.Open silently creates an empty database on a wrong path or |
| 723 | // an unmounted /data (same guard, same reason, as the unit-path Sync). |
| 724 | if len(routes) == 0 { |
| 725 | return nil |
| 726 | } |
| 727 | if !hostWrites.Enabled { |
| 728 | reapConfigEnv(routes, dataDir) |
| 729 | } else { |
| 730 | reapSelectedConfigTemps(ctx, dataDir, selected) |
| 731 | } |
| 732 | return nil |
| 733 | } |
| 734 | |
| 735 | func reapSelectedConfigTemps(ctx context.Context, dataDir string, selected map[string]hostwrite.Work) { |
| 736 | ents, err := os.ReadDir(dataDir) |
| 737 | if err != nil { |
| 738 | return |
| 739 | } |
| 740 | for _, entry := range ents { |
| 741 | if entry.IsDir() || !strings.HasSuffix(entry.Name(), configEnvSuffix+".tmp") { |
| 742 | continue |
| 743 | } |
| 744 | host := strings.TrimSuffix(entry.Name(), configEnvSuffix+".tmp") |
| 745 | work := selected[host] |
| 746 | if work == nil { |
| 747 | continue |
| 748 | } |
| 749 | if err := guarded(ctx, work, renderedIdentity(nil, false, false, false), func() error { |
| 750 | return os.Remove(filepath.Join(dataDir, entry.Name())) |
| 751 | }); err != nil { |
| 752 | fmt.Printf(" %s: remove stale config temp file: %v\n", entry.Name(), err) |
| 753 | } |
| 754 | } |
| 755 | } |
| 756 | |
| 757 | // SyncOption adjusts one SyncExec pass. The options are variadic so the |
| 758 | // ordinary call keeps its shape. |
| 759 | type SyncOption func(*syncOptions) |
| 760 | |
| 761 | type syncOptions struct { |
| 762 | fenced func(host string) bool |
| 763 | keep func(rt registry.Route) bool |
| 764 | dirAbsent func(host, dir string) |
| 765 | outcome func(host string, done bool) |
| 766 | hostWrites hostwrite.Config |
| 767 | hostWriteCount int |
| 768 | } |
| 769 | |
| 770 | // WithRouteFilter narrows one SyncExec pass to the routes keep accepts. A |
| 771 | // route it leaves out is out of scope for the pass: nothing is read for it, |
| 772 | // and its file is neither written nor removed (reap keys on route existence, |
| 773 | // not scope). The agent's scoped pass uses it to converge only the apps the |
| 774 | // change index says moved (S3 change-index design §4.5). Nil keeps every |
| 775 | // route. |
| 776 | func WithRouteFilter(keep func(registry.Route) bool) SyncOption { |
| 777 | return func(o *syncOptions) { o.keep = keep } |
| 778 | } |
| 779 | |
| 780 | func (o *syncOptions) kept(rt registry.Route) bool { return o.keep == nil || o.keep(rt) } |
| 781 | |
| 782 | // WhenInstanceDirAbsent hands SyncExec a callback for every nested row whose |
| 783 | // bundle it could not write because the row's instance directory does not |
| 784 | // exist yet: the activator makes that directory at the instance's first |
| 785 | // wake (prepare), so a nested row created with settings has nowhere to put |
| 786 | // its <host>.config.env until then. The pass skips such a row cleanly, says |
| 787 | // so, and calls dirAbsent with the host and the directory it waits for, so |
| 788 | // the agent can run the pass again as soon as the directory appears rather |
| 789 | // than leaving the instance without its settings until the hourly floor |
| 790 | // (harness Run 9, finding 6). |
| 791 | func WhenInstanceDirAbsent(dirAbsent func(host, dir string)) SyncOption { |
| 792 | return func(o *syncOptions) { o.dirAbsent = dirAbsent } |
| 793 | } |
| 794 | |
| 795 | func WithHostWrite(cfg hostwrite.Config) SyncOption { |
| 796 | return func(o *syncOptions) { |
| 797 | o.hostWrites = cfg |
| 798 | o.hostWriteCount++ |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | // SkipFenced makes SyncExec leave alone every in-scope route for which |
| 803 | // fenced reports true: no bundle is written, removed or bounced for it, |
| 804 | // and the pass says so. The agent passes the layout pass's journal |
| 805 | // predicate. The pass also holds the agent's config-sync lock while it |
| 806 | // moves a host, so this is the second net: it covers a journal that |
| 807 | // outlives one pass, such as a completed revert held until the intent can |
| 808 | // be read. |
| 809 | func SkipFenced(fenced func(host string) bool) SyncOption { |
| 810 | return func(o *syncOptions) { o.fenced = fenced } |
| 811 | } |
| 812 | |
| 813 | // WithHostOutcome hands SyncExec a callback it calls, once a pass has gone |
| 814 | // through every route without stopping, for each in-scope route: done is |
| 815 | // true when the route's file is as the pass resolved it (written, already up |
| 816 | // to date, removed, or rightly left absent), and false when the pass left it |
| 817 | // as it was for any reason (an unresolved bundle or overlay, a guard, a |
| 818 | // layout migration in flight, a write that failed, an instance directory not |
| 819 | // made yet). A route out of scope is not named. The agent keeps a host whose |
| 820 | // instance directory it waits for until the host is done (fix round 1, |
| 821 | // Astra's finding 3: a pass that skipped a host on a transient read failure |
| 822 | // returned nil, and the agent took that as delivered). |
| 823 | func WithHostOutcome(outcome func(host string, done bool)) SyncOption { |
| 824 | return func(o *syncOptions) { o.outcome = outcome } |
| 825 | } |
| 826 | |
| 827 | // execRouteBase is the lifecycle-managed-instance shape every exec route |
| 828 | // kind shares: the unit path's syncable() exclusions, narrowed to |
| 829 | // non-unit backing (unit routes belong to the root adopt path). Neither |
| 830 | // SyncExec eligibility case on its own — see execSyncable, execAddrCandidate |
| 831 | // and execSidecarCandidate, which each narrow it by upstream shape. |
| 832 | // |
| 833 | // Kind "sidecar" is included alongside "instance" (Task 11, design §9.1): |
| 834 | // this is the direct replacement for the hand-staged |
| 835 | // <host>-sidecar.env — `carlos env set` must reach a sidecar, secrets are |
| 836 | // sealed and delivered the same way an instance's are, and the file this |
| 837 | // predicate gates is exactly what ExecBackend.instanceEnv reads at |
| 838 | // StartWorker's spawn. Every other clause is untouched: a sidecar row is |
| 839 | // still excluded by Platform, a non-ready Status, or Backing == |
| 840 | // BackingUnit, same as an instance would be. |
| 841 | func execRouteBase(rt registry.Route) bool { |
| 842 | return (rt.Kind == "instance" || rt.IsSidecar()) && rt.App != "" && !rt.Platform && |
| 843 | (rt.Status == "" || rt.Status == "ready") && |
| 844 | rt.Backing != registry.BackingUnit |
| 845 | } |
| 846 | |
| 847 | // execSyncable: lifecycle-managed instances the ACTIVATOR spawns — and |
| 848 | // therefore whose environment this sweep can actually deliver and whose |
| 849 | // process it can bounce. Narrows execRouteBase to a socket upstream; |
| 850 | // unconditionally in scope, opt-in or not (this is the original, PR #16 |
| 851 | // shape). |
| 852 | func execSyncable(rt registry.Route) bool { |
| 853 | return execRouteBase(rt) && rt.Socket != "" |
| 854 | } |
| 855 | |
| 856 | // execAddrCandidate: the TCP-addr exec shape (issue #18) — an |
| 857 | // app-supervised instance family carlos routes to but neither spawns nor |
| 858 | // owns the process of (titogo's router fleet is the motivating case: |
| 859 | // titogo-router + controld own the instance processes, carlos just |
| 860 | // proxies to 127.0.0.1:<port>). Narrows execRouteBase to an addr upstream |
| 861 | // with no socket. Both registry spellings are accepted: Addr is the legacy |
| 862 | // single-upstream field, while Addrs is the HA field used even when a route |
| 863 | // currently has only one upstream. Treating only Addr as eligible silently |
| 864 | // skipped config delivery when an existing route migrated to Addrs. |
| 865 | // |
| 866 | // This predicate is the ROUTE-SHAPE half only — it says nothing about |
| 867 | // whether the route is actually in scope this pass. Unlike execSyncable, |
| 868 | // that also requires the app's default bundle to carry "exec_delivery": |
| 869 | // "file" (Manifest.ExecDelivery): the reader of the file this shape |
| 870 | // produces is the app's OWN supervisor, not carlos, so writing it for an |
| 871 | // app that never reads it would be dead weight at best and, if the app's |
| 872 | // supervisor treats the file's mere presence or mtime as a restart |
| 873 | // signal, an unwanted bounce at worst. See SyncExec's checkOptedIn/ |
| 874 | // inScope closures, which gate this predicate on that opt-in. |
| 875 | func execAddrCandidate(rt registry.Route) bool { |
| 876 | return execRouteBase(rt) && rt.Socket == "" && (rt.Addr != "" || rt.Addrs != "") |
| 877 | } |
| 878 | |
| 879 | // execSidecarCandidate: the standalone-worker exec shape (Task 11, design |
| 880 | // §9.1) — a sidecar row carries neither Socket nor Addr by construction |
| 881 | // (sidecarRouteFor's contract: a sidecar binds nothing and nothing dials |
| 882 | // it), so it can never satisfy execSyncable or execAddrCandidate; this is |
| 883 | // its own narrowing of execRouteBase, keyed on Kind alone. |
| 884 | // |
| 885 | // Unconditionally in scope, like execSyncable and UNLIKE execAddrCandidate: |
| 886 | // a sidecar's process is spawned and owned by THIS box's activator |
| 887 | // (ExecBackend.StartWorker), the same as a socket-exec instance, so there |
| 888 | // is no "app-supervised, may not read the file" case here to opt into — |
| 889 | // carlos always owns delivery for a route it owns the process of. |
| 890 | func execSidecarCandidate(rt registry.Route) bool { |
| 891 | return execRouteBase(rt) && rt.IsSidecar() |
| 892 | } |
| 893 | |
| 894 | // ExecConfigApps is the distinct (account, app) set the agent's fast |
| 895 | // generation loop (host package) polls: every app with at least one route |
| 896 | // SyncExec could deliver config to. Accounts are resolved through |
| 897 | // release.RouteAccount, the same way every other read of a route's account |
| 898 | // resolves, so the key the loop watches is the key a write actually touches. |
| 899 | // |
| 900 | // The SHAPE half only — the TCP-addr case's opt-in (Manifest.ExecDelivery) |
| 901 | // is NOT consulted, because consulting it costs a bucket read per app, |
| 902 | // which is exactly what the fast loop exists to avoid. The cost of that |
| 903 | // choice is bounded: polling an un-opted app's generation is one wasted |
| 904 | // small read per tick, and the SyncExec pass a change triggers applies the |
| 905 | // opt-in properly before writing anything. |
| 906 | // |
| 907 | // execSidecarCandidate is included alongside execSyncable and |
| 908 | // execAddrCandidate (Task 11 fix round 1): left out, a sidecar-only app's |
| 909 | // config — delivered by a direct-bucket write that never bumps |
| 910 | // control/serial, exactly what the pinfra scripts do — would fall back to |
| 911 | // the hourly floor while an instance next to it converges on this loop's |
| 912 | // fast cadence. Not an error, just work happening late; the same |
| 913 | // half-migrated-consumer shape this branch has repeatedly found. Costs |
| 914 | // nothing extra to include unconditionally: a sidecar's process is spawned |
| 915 | // and owned by carlos exactly like a socket-exec instance's (see |
| 916 | // execSidecarCandidate's own doc comment), so there is no opt-in question |
| 917 | // to defer here the way there is for the addr shape. |
| 918 | // |
| 919 | // Exported from THIS file, beside the predicates it composes, so the loop's |
| 920 | // notion of eligibility can never drift from SyncExec's own — an app the |
| 921 | // loop watched but SyncExec skipped would burn reads forever, and an app it |
| 922 | // didn't watch would silently fall back to the hourly pass. |
| 923 | func ExecConfigApps(routes []registry.Route) [][2]string { |
| 924 | seen := map[[2]string]bool{} |
| 925 | var out [][2]string |
| 926 | for _, rt := range routes { |
| 927 | if !execSyncable(rt) && !execAddrCandidate(rt) && !execSidecarCandidate(rt) { |
| 928 | continue |
| 929 | } |
| 930 | k := [2]string{release.RouteAccount(rt), rt.App} |
| 931 | if seen[k] { |
| 932 | continue |
| 933 | } |
| 934 | seen[k] = true |
| 935 | out = append(out, k) |
| 936 | } |
| 937 | // Sorted so the loop's per-tick read order — and any log line built |
| 938 | // from it — is stable across passes, where registry order is not. |
| 939 | sort.Slice(out, func(i, j int) bool { |
| 940 | if out[i][0] != out[j][0] { |
| 941 | return out[i][0] < out[j][0] |
| 942 | } |
| 943 | return out[i][1] < out[j][1] |
| 944 | }) |
| 945 | return out |
| 946 | } |
| 947 | |
| 948 | // execDeliveryVerdict is the DECIDED half of exec_delivery resolution |
| 949 | // (issue #195): given a route's bound bundle and its app's default one, |
| 950 | // what does the flag say? Pure and log-free, so that the two callers who |
| 951 | // need the answer — SyncExec's checkOptedIn, which writes the files, and |
| 952 | // AwaitingRestart, which tells an operator a file was written that nothing |
| 953 | // has read — reach it through one function rather than through two |
| 954 | // readings of one rule. A notice that disagreed with the sweep about who |
| 955 | // gets a file would be worse than no notice: it would name the wrong |
| 956 | // route, in the exact place an operator went looking for certainty. |
| 957 | // |
| 958 | // Fetch FAILURES are deliberately not modelled here. They are not |
| 959 | // determinations (see checkOptedIn's def.err branch), and the two callers |
| 960 | // owe them different things — a loud log line and a left-alone file on one |
| 961 | // side, a named "could not tell" host on the other. A caller that could |
| 962 | // not read a bundle must not call this. |
| 963 | type execDeliveryVerdict struct { |
| 964 | // On is the only question SyncExec asks: does this bundle get files? |
| 965 | On bool |
| 966 | // Unknown is non-empty when the DECIDING bundle stated a value this |
| 967 | // binary does not understand — decided "off", but for a reason worth |
| 968 | // saying out loud (additive discipline, AGENTS.md: a manifest a newer |
| 969 | // carlos wrote must not brick an older box, and must not pass silently |
| 970 | // either). Carried out rather than logged in place, because the two |
| 971 | // callers word it differently and print it to different places. |
| 972 | Unknown string |
| 973 | // FromEnvironment records WHICH bundle decided, so a caller's message |
| 974 | // can name it. False means the app default decided — including the |
| 975 | // common case where a named bundle was silent and inherited. |
| 976 | FromEnvironment bool |
| 977 | } |
| 978 | |
| 979 | // resolveExecDelivery applies the #195 precedence: the route's OWN bundle |
| 980 | // first, and only when it STATES something; the app default otherwise. |
| 981 | // Absence is inheritance, never "off" — an operator who has never |
| 982 | // mentioned exec_delivery in a bundle has not asked for delivery to stop. |
| 983 | // |
| 984 | // loadDefault is a thunk rather than a value because the default bundle |
| 985 | // must not be FETCHED when the named one already decided: eager fetching |
| 986 | // would let an unrelated blip on the default's read turn a decided verdict |
| 987 | // into an unresolved one, and would cost a bucket read per call on the |
| 988 | // path where the answer was already in hand. |
| 989 | func resolveExecDelivery(envM *Manifest, envFound bool, loadDefault func() (*Manifest, bool, error)) (execDeliveryVerdict, error) { |
| 990 | if envFound && envM.ExecDelivery == ExecDeliveryFile { |
| 991 | return execDeliveryVerdict{On: true, FromEnvironment: true}, nil |
| 992 | } |
| 993 | if envFound && envM.ExecDelivery != "" { |
| 994 | return execDeliveryVerdict{Unknown: envM.ExecDelivery, FromEnvironment: true}, nil |
| 995 | } |
| 996 | defM, defFound, err := loadDefault() |
| 997 | if err != nil { |
| 998 | return execDeliveryVerdict{}, err |
| 999 | } |
| 1000 | switch { |
| 1001 | case !defFound: |
| 1002 | return execDeliveryVerdict{}, nil |
| 1003 | case defM.ExecDelivery == ExecDeliveryFile: |
| 1004 | return execDeliveryVerdict{On: true}, nil |
| 1005 | case defM.ExecDelivery == "": |
| 1006 | return execDeliveryVerdict{}, nil |
| 1007 | default: |
| 1008 | return execDeliveryVerdict{Unknown: defM.ExecDelivery}, nil |
| 1009 | } |
| 1010 | } |
| 1011 | |
| 1012 | // RestartHost names one route a config write delivered a file to and that |
| 1013 | // nothing bounced. Unit is carried because it decides the REMEDY, and the |
| 1014 | // two remedies are not interchangeable: an addr route naming its own unit |
| 1015 | // (issue #110's consoles) is the one addr shape `carlos restart` can |
| 1016 | // actually cycle, where a plain one reaches nothing at all — the agent's |
| 1017 | // restart leg skips it (internal/host/restart.go, the rt.Unit != "" test) |
| 1018 | // and the activator it would otherwise ask has no child for a process |
| 1019 | // carlos never spawned. Telling an operator to run a command that is a |
| 1020 | // documented no-op for their route would be a second silent failure |
| 1021 | // answering the first. |
| 1022 | type RestartHost struct { |
| 1023 | Host string |
| 1024 | Unit string |
| 1025 | } |
| 1026 | |
| 1027 | // RestartPending is what a config write left behind on ONE box: files that |
| 1028 | // landed for processes carlos does not own, and therefore does not restart. |
| 1029 | // |
| 1030 | // Unresolved is the honest half and the reason this is not a plain |
| 1031 | // []string. A route whose exec_delivery could not be resolved this call — |
| 1032 | // its bundle unreadable, or the bundle it is bound to missing outright — |
| 1033 | // is a route nothing can say "delivered" or "not delivered" about. Dropping |
| 1034 | // it silently would commit the very conflation this notice exists to end: |
| 1035 | // "nothing needs a restart" and "I could not tell" rendered identically is |
| 1036 | // how discussion #16 got filed in the first place. |
| 1037 | type RestartPending struct { |
| 1038 | Hosts []RestartHost |
| 1039 | Unresolved []string |
| 1040 | } |
| 1041 | |
| 1042 | // Empty reports whether there is nothing to say — no host delivered to and |
| 1043 | // none unresolved. The renderers' one test, so "say nothing" is one |
| 1044 | // judgement rather than four. |
| 1045 | func (r RestartPending) Empty() bool { return len(r.Hosts) == 0 && len(r.Unresolved) == 0 } |
| 1046 | |
| 1047 | // AwaitingRestart answers, for a config write to app's envName bundle |
| 1048 | // ("" = the default), which of this box's routes were handed a file that |
| 1049 | // nothing will make their process read. It is the reporting half of the |
| 1050 | // ownership line SyncExec draws: carlos writes these files and |
| 1051 | // deliberately never bounces the processes behind them, so the only thing |
| 1052 | // that can close the loop is telling the person who just ran the command. |
| 1053 | // See discussion #16 — `secrets set` reported success, `secrets list` |
| 1054 | // confirmed the key, and the value reached nothing for want of one line. |
| 1055 | // |
| 1056 | // keys are the var names the write touched (set and unset together), used |
| 1057 | // ONLY to skip routes whose merged config the write cannot have changed — |
| 1058 | // see the shadowing test below. An empty slice disables that narrowing, |
| 1059 | // which is what a write with no vars in it (turning exec_delivery on) |
| 1060 | // needs: there is nothing to shadow, and the file appears regardless. |
| 1061 | // |
| 1062 | // Eligibility is execAddrCandidate plus the opt-in, which is exactly |
| 1063 | // SyncExec's inScope for the addr shape — reached through the same |
| 1064 | // predicate and the same resolveExecDelivery, from the same file, for the |
| 1065 | // same reason ExecConfigApps is exported from here. |
| 1066 | // |
| 1067 | // SCOPE, and it is a real limit: routes is whatever the caller could see, |
| 1068 | // which is one box's registry (the #145 blind spot internal/console/ |
| 1069 | // apiroutes.go documents). This function cannot know what it was not |
| 1070 | // shown, so its callers say what they looked at rather than letting a |
| 1071 | // short list read as a complete one. |
| 1072 | func AwaitingRestart(ctx context.Context, s release.Store, routes []registry.Route, acct, app, envName string, keys []string) RestartPending { |
| 1073 | var out RestartPending |
| 1074 | // The instance layer, through the same loader the sweeps use, so this |
| 1075 | // notice and the pass that writes the files agree about what sits above |
| 1076 | // a bundle. One listing for the app, then a Get per route that has an |
| 1077 | // overlay; both live for one call, like the caches below. |
| 1078 | overlayFor := instanceOverlayLoader(ctx, s) |
| 1079 | // Both caches exist for the same reason SyncExec's do: a box with many |
| 1080 | // TCP-addr routes on one bundle must not cost more than one read for |
| 1081 | // it. Unlike SyncExec's, they live for one call. |
| 1082 | var ( |
| 1083 | defM *Manifest |
| 1084 | defFound bool |
| 1085 | defErr error |
| 1086 | defDone bool |
| 1087 | ) |
| 1088 | loadDefault := func() (*Manifest, bool, error) { |
| 1089 | if !defDone { |
| 1090 | defM, defFound, defErr = getIfExists(ctx, s, acct, app) |
| 1091 | defDone = true |
| 1092 | } |
| 1093 | return defM, defFound, defErr |
| 1094 | } |
| 1095 | type envRes struct { |
| 1096 | m *Manifest |
| 1097 | found bool |
| 1098 | err error |
| 1099 | } |
| 1100 | envBundles := map[string]envRes{} |
| 1101 | loadEnv := func(env string) envRes { |
| 1102 | if e, done := envBundles[env]; done { |
| 1103 | return e |
| 1104 | } |
| 1105 | m, found, err := getIfExistsAt(ctx, s, EnvironmentKey(acct, app, env)) |
| 1106 | e := envRes{m: m, found: found, err: err} |
| 1107 | envBundles[env] = e |
| 1108 | return e |
| 1109 | } |
| 1110 | for _, rt := range routes { |
| 1111 | if !execAddrCandidate(rt) || rt.App != app || release.RouteAccount(rt) != acct { |
| 1112 | continue |
| 1113 | } |
| 1114 | // Which routes a write REACHES. A named bundle is read by the |
| 1115 | // routes bound to it and by nothing else. The default bundle is |
| 1116 | // read by every route of the app, bound or not, because a named |
| 1117 | // bundle is an overlay ON it (environments.go, LoadMerged) rather |
| 1118 | // than a replacement for it. |
| 1119 | if envName != "" && rt.Environment != envName { |
| 1120 | continue |
| 1121 | } |
| 1122 | eb := envRes{} |
| 1123 | if rt.Environment != "" { |
| 1124 | eb = loadEnv(rt.Environment) |
| 1125 | if eb.err != nil { |
| 1126 | out.Unresolved = append(out.Unresolved, rt.Host) |
| 1127 | continue |
| 1128 | } |
| 1129 | // A route bound to a bundle that does not exist resolves to an |
| 1130 | // ERROR, not to the app default: LoadMerged refuses it, and |
| 1131 | // SyncExec leaves the route's file exactly as it was. Nothing |
| 1132 | // was delivered, so "delivered" would be a lie — but neither |
| 1133 | // is this route fine, so it is not a skip either. |
| 1134 | if !eb.found { |
| 1135 | out.Unresolved = append(out.Unresolved, rt.Host) |
| 1136 | continue |
| 1137 | } |
| 1138 | } |
| 1139 | v, err := resolveExecDelivery(eb.m, eb.found, loadDefault) |
| 1140 | if err != nil { |
| 1141 | out.Unresolved = append(out.Unresolved, rt.Host) |
| 1142 | continue |
| 1143 | } |
| 1144 | // Not opted in — including the Unknown case, which decides "off" |
| 1145 | // — means no file was ever written for this route, by this write |
| 1146 | // or any other. Nothing to say. |
| 1147 | if !v.On { |
| 1148 | continue |
| 1149 | } |
| 1150 | // The shadowing test, and the reason keys is a parameter: where every |
| 1151 | // touched key is shadowed, the merged manifest is byte-identical, |
| 1152 | // syncExecHost's bytes.Equal writes nothing, and naming the route |
| 1153 | // would send someone to restart a process for a value it was never |
| 1154 | // going to see. An empty keys disables this whole test (nothing to |
| 1155 | // shadow — see keys' own doc comment), so the instance layer is |
| 1156 | // consulted only when it could actually change the verdict: a route |
| 1157 | // whose overlay is unreadable must not be downgraded to Unresolved |
| 1158 | // over a call that was never going to ask it anything. |
| 1159 | if len(keys) > 0 { |
| 1160 | // Which layers sit ABOVE the bundle that was written, for this |
| 1161 | // route. A write cannot change what a route resolves to for a |
| 1162 | // key some higher layer states, and the third layer sits above |
| 1163 | // both of the other two. |
| 1164 | inst, _, instErr := overlayFor(acct, app, rt.Host) |
| 1165 | if instErr != nil { |
| 1166 | out.Unresolved = append(out.Unresolved, rt.Host) |
| 1167 | continue |
| 1168 | } |
| 1169 | var above []*Manifest |
| 1170 | if envName == "" && rt.Environment != "" { |
| 1171 | above = append(above, eb.m) |
| 1172 | } |
| 1173 | above = append(above, inst) |
| 1174 | if allShadowed(keys, above...) { |
| 1175 | continue |
| 1176 | } |
| 1177 | } |
| 1178 | out.Hosts = append(out.Hosts, RestartHost{Host: rt.Host, Unit: rt.Unit}) |
| 1179 | } |
| 1180 | // Sorted so a caller's output — and any test pinning it — is stable |
| 1181 | // across passes, where registry order is not. Same discipline as |
| 1182 | // ExecConfigApps'. |
| 1183 | sort.Slice(out.Hosts, func(i, j int) bool { return out.Hosts[i].Host < out.Hosts[j].Host }) |
| 1184 | sort.Strings(out.Unresolved) |
| 1185 | return out |
| 1186 | } |
| 1187 | |
| 1188 | // allShadowed reports whether every one of keys is STATED by at least one |
| 1189 | // of the layers above the bundle a write touched — the test that decides |
| 1190 | // whether that write can have changed what a route actually resolves to. An |
| 1191 | // unset key counts the same as a set one: removing a key a higher layer |
| 1192 | // shadows changes the merge by nothing. |
| 1193 | // |
| 1194 | // above is variadic because how many layers sit over a write depends on |
| 1195 | // which bundle was written: a default-bundle write to a bound route is |
| 1196 | // under both the named bundle and the instance overlay, a named-bundle |
| 1197 | // write is under the overlay alone, and a write to the overlay itself has |
| 1198 | // nothing over it. A nil layer (no bundle of that kind for this route) is |
| 1199 | // skipped rather than treated as stating nothing, so an empty list answers |
| 1200 | // false and the route is named. |
| 1201 | func allShadowed(keys []string, above ...*Manifest) bool { |
| 1202 | for _, k := range keys { |
| 1203 | stated := false |
| 1204 | for _, m := range above { |
| 1205 | if m == nil { |
| 1206 | continue |
| 1207 | } |
| 1208 | if _, ok := m.Vars[k]; ok { |
| 1209 | stated = true |
| 1210 | break |
| 1211 | } |
| 1212 | } |
| 1213 | if !stated { |
| 1214 | return false |
| 1215 | } |
| 1216 | } |
| 1217 | return true |
| 1218 | } |
| 1219 | |
| 1220 | // execAddrConfigHeader is prepended to a TCP-addr (issue #18) route's |
| 1221 | // <host>.config.env — and ONLY that shape. RenderEnv itself stays |
| 1222 | // header-free, deliberately: it is shared by sync.go's unit path AND this |
| 1223 | // file's socket-exec path, and either one growing a header would change |
| 1224 | // the byte content of every EXISTING file of that shape on upgrade — |
| 1225 | // sync.go's own comment on lostKeys already weighs a "# managed by |
| 1226 | // carlos" marker for the unit path and rejects it for exactly this |
| 1227 | // churn. The socket-exec shape has the same installed-base problem: every |
| 1228 | // currently-running instance's file would look "changed" on the pass |
| 1229 | // after this ships, and — unlike the addr shape — a socket route DOES |
| 1230 | // bounce on a content change, so that churn would restart every |
| 1231 | // socket-exec instance on the fleet once, for a comment. The addr shape |
| 1232 | // has no installed base at all (this is its first pass ever) and never |
| 1233 | // bounces, so the header costs it nothing. |
| 1234 | // |
| 1235 | // Format: "#"-prefixed and blank lines are the universal EnvironmentFile |
| 1236 | // comment convention — systemd's own EnvironmentFile= parser skips them, |
| 1237 | // and this repo's own guard for the unit path, lostKeys (sync.go), |
| 1238 | // already relies on exactly that tolerance when parsing a hand-staged |
| 1239 | // file. The issue #18 design comment specifies titogo's controld layers |
| 1240 | // this file the same way systemd's EnvironmentFile= does (default bundle |
| 1241 | // under the app-supervisor's own env), so a plain KEY=value-plus-#-comments |
| 1242 | // reader is the assumed parse contract on that side too; this repo has no |
| 1243 | // titogo checkout to verify it against directly, so treat this as an |
| 1244 | // assumption to confirm against titogo's actual parser before or during |
| 1245 | // the app-half PR, not a verified fact. |
| 1246 | const execAddrConfigHeader = `# Written by carlos (SyncExec, issue #18). Carlos may overwrite or remove |
| 1247 | # this file on any future sync pass, but NEVER restarts this process because |
| 1248 | # of it — carlos does not own it. This app's own supervisor owns restarts |
| 1249 | # and must notice and act on changes here on its own schedule. |
| 1250 | ` |
| 1251 | |
| 1252 | // syncExecHost converges one route's config file. m == nil means the |
| 1253 | // manifest object is absent from the bucket: remove the file (sync owns |
| 1254 | // the namespace; a file with no manifest is a leftover, and leaving it |
| 1255 | // would keep dead config flowing into spawns). key is the default |
| 1256 | // bundle's key, Key(acct, app) — the only key whose absence can reach |
| 1257 | // this branch — included in the removal log line so cause (which bucket |
| 1258 | // key came back absent) and effect (which file got removed) land |
| 1259 | // together on one line, instead of requiring a reader to cross-reference |
| 1260 | // the fetch-side log. |
| 1261 | // |
| 1262 | // bounce may be nil: bounceLogged treats that as "nothing to call", which |
| 1263 | // is how a TCP-addr (issue #18) route's file gets converged or removed |
| 1264 | // exactly like a socket-exec route's — same write/rename/remove path, |
| 1265 | // same logging — while never restarting anything, because carlos does |
| 1266 | // not own that process. The caller (SyncExec) passes nil for those routes |
| 1267 | // structurally, not a flag this function branches on. |
| 1268 | // |
| 1269 | // header is prepended to the rendered content when non-empty |
| 1270 | // (execAddrConfigHeader for the addr shape, "" for socket — see that |
| 1271 | // constant's doc comment for why the two must never be swapped or |
| 1272 | // unified) — folded into the SAME diff/write/rename path as the vars |
| 1273 | // themselves, so a header-only change (e.g. its text edited in a future |
| 1274 | // release) converges exactly like a var change would, once. |
| 1275 | // |
| 1276 | // instETag is the etag of the instance overlay this route's merge already |
| 1277 | // folded in (empty when the route has no overlay), carried through to the |
| 1278 | // ack unchanged — see ConfigAck.InstanceConfigETag. |
| 1279 | // |
| 1280 | // ack, if non-nil, is called once this route's file has actually changed |
| 1281 | // (written, never on removal — see SyncExec's own doc comment) with its |
| 1282 | // config-adoption acknowledgement (spec §5). Restarted/RestartError come |
| 1283 | // from bounceLogged's own outcome when bounce is non-nil (the socket-exec |
| 1284 | // shape: spec §5 step 5 counts that bounce as the cutover, "since the |
| 1285 | // platform restarted it") and are always false/empty when bounce is nil |
| 1286 | // (the tcp-exec-file shape: its process belongs to an external supervisor |
| 1287 | // this platform never restarts at all, so that shape's cutover is a |
| 1288 | // member acknowledgement, `carlos relay adopted`, instead). |
| 1289 | func syncExecHost(rt registry.Route, m *Manifest, key string, priv *ecdh.PrivateKey, dataDir string, bounce func(host string) error, header, instETag string, ack func(ConfigAck)) { |
| 1290 | _ = syncExecHostClaimed(context.Background(), legacyHostWork{}, rt, m, key, priv, dataDir, bounce, header, instETag, ack) |
| 1291 | } |
| 1292 | |
| 1293 | func syncExecHostClaimed(ctx context.Context, work hostwrite.Work, rt registry.Route, m *Manifest, key string, priv *ecdh.PrivateKey, dataDir string, bounce func(host string) error, header, instETag string, ack func(ConfigAck), options ...Callbacks) error { |
| 1294 | callbacks := configCallbacks(work, bounce, ack, options) |
| 1295 | if rt.Host == "" || rt.Host == "." || rt.Host == ".." || rt.Host != filepath.Base(rt.Host) { |
| 1296 | fmt.Printf(" %q: not a bare hostname, refusing to derive a config file path from it\n", rt.Host) |
| 1297 | return nil |
| 1298 | } |
| 1299 | path := ConfigEnvPath(dataDir, rt) |
| 1300 | if m == nil { |
| 1301 | id := renderedIdentity(nil, false, rt.Environment != "", instETag != "") |
| 1302 | if err := guarded(ctx, work, id, func() error { return os.Remove(path) }); err != nil { |
| 1303 | if !os.IsNotExist(err) { |
| 1304 | fmt.Printf(" %s: remove config for %s (manifest absent): %v\n", rt.Host, key, err) |
| 1305 | } |
| 1306 | if os.IsNotExist(err) { |
| 1307 | return nil |
| 1308 | } |
| 1309 | return err |
| 1310 | } |
| 1311 | fmt.Printf(" %s: manifest %s absent, removed its config file\n", rt.Host, key) |
| 1312 | if err := guarded(ctx, work, id, func() error { return nil }); err != nil { |
| 1313 | return err |
| 1314 | } |
| 1315 | if callbacks.Delivered != nil { |
| 1316 | callbacks.Delivered(rt.Host, id.Config) |
| 1317 | } |
| 1318 | if bounce != nil || callbacks.SelectedRestart != nil { |
| 1319 | if callbacks.enabled { |
| 1320 | if err := callbacks.restart(ctx, hostwrite.Selection{Host: rt.Host, Work: work, Identity: id}); !errors.Is(err, ErrDeliveredOnWake) { |
| 1321 | return err |
| 1322 | } |
| 1323 | return nil |
| 1324 | } |
| 1325 | bounceLogged(rt.Host, bounce) |
| 1326 | } |
| 1327 | return nil |
| 1328 | } |
| 1329 | // keepAbsent: no instance overlay. A bundle that renders nothing for a |
| 1330 | // route whose file is absent leaves it absent (renderExecConfig). |
| 1331 | r, err := renderExecConfig(ctx, work, rt, m, priv, dataDir, header, instETag, instETag == "") |
| 1332 | if err != nil { |
| 1333 | return err |
| 1334 | } |
| 1335 | if !r.changed { |
| 1336 | return nil // up to date — stay quiet on the periodic tick |
| 1337 | } |
| 1338 | id := r.id |
| 1339 | if err := guarded(ctx, work, id, func() error { return nil }); err != nil { |
| 1340 | return err |
| 1341 | } |
| 1342 | if callbacks.Delivered != nil { |
| 1343 | callbacks.Delivered(rt.Host, id.Config) |
| 1344 | } |
| 1345 | var bounceErr error |
| 1346 | if bounce != nil || callbacks.SelectedRestart != nil { |
| 1347 | if callbacks.enabled { |
| 1348 | bounceErr = callbacks.restart(ctx, hostwrite.Selection{Host: rt.Host, Work: work, Identity: id}) |
| 1349 | if bounceErr != nil && !errors.Is(bounceErr, ErrDeliveredOnWake) { |
| 1350 | return bounceErr |
| 1351 | } |
| 1352 | } else { |
| 1353 | bounceErr = bounceLogged(rt.Host, bounce) |
| 1354 | } |
| 1355 | } |
| 1356 | if ack != nil || callbacks.SelectedAck != nil { |
| 1357 | a := ConfigAck{Schema: ConfigAckSchema, Host: rt.Host, Shape: shapeOf(rt), At: time.Now(), InstanceConfigETag: instETag} |
| 1358 | if suffixes, err := relaySuffixesFromRelayVar(m, priv, routeBinding(rt)); err != nil { |
| 1359 | fmt.Printf(" %s: config updated, but CARLOS_RELAY could not be read for the config-ack: %v\n", rt.Host, err) |
| 1360 | } else { |
| 1361 | a.RelaySuffixes = suffixes |
| 1362 | } |
| 1363 | // bounce == nil (the tcp-exec-file/addr shape) leaves Restarted |
| 1364 | // false and RestartError empty — see this function's own doc |
| 1365 | // comment for why. bounce != nil is the socket-exec shape, and its |
| 1366 | // bounce IS this pass's restart (spec §5 step 5: a socket-exec |
| 1367 | // route's cutover is the ack "since the platform restarted it"), |
| 1368 | // so the ack has to carry what bounceLogged actually returned |
| 1369 | // rather than assume success the way the removal branch above may. |
| 1370 | // A sleeper's restart answers ErrDeliveredOnWake: nothing ran and |
| 1371 | // nothing failed, so the ack carries the unit pass's marker. |
| 1372 | if bounce != nil { |
| 1373 | switch { |
| 1374 | case errors.Is(bounceErr, ErrDeliveredOnWake): |
| 1375 | a.Hibernating = true |
| 1376 | case bounceErr != nil: |
| 1377 | a.RestartError = bounceErr.Error() |
| 1378 | default: |
| 1379 | a.Restarted = true |
| 1380 | } |
| 1381 | } |
| 1382 | if err := guarded(ctx, work, id, func() error { return nil }); err != nil { |
| 1383 | return err |
| 1384 | } |
| 1385 | return callbacks.acknowledge(ctx, hostwrite.Selection{Host: rt.Host, Work: work, Identity: id}, a) |
| 1386 | } |
| 1387 | return nil |
| 1388 | } |
| 1389 | |
| 1390 | // errInstanceDirAbsent is a nested row's write that found no instance |
| 1391 | // directory: the activator has not woken the instance here yet. |
| 1392 | var errInstanceDirAbsent = errors.New("the instance directory is not made yet") |
| 1393 | |
| 1394 | // execRender is what renderExecConfig did: the file's path, the bytes it |
| 1395 | // holds now, whether this call wrote them, and the identity the writes were |
| 1396 | // guarded under (zero when nothing was written). |
| 1397 | type execRender struct { |
| 1398 | path string |
| 1399 | rendered []byte |
| 1400 | changed bool |
| 1401 | id hostwrite.Identity |
| 1402 | } |
| 1403 | |
| 1404 | // renderExecConfig is the file-writing half of syncExecHostClaimed for a |
| 1405 | // present manifest: render m for rt, compare with what is on disk, and |
| 1406 | // write through work only when the bytes differ. Extracted so the held-row |
| 1407 | // render (RenderHeld, moves design §9.1) writes through the very same |
| 1408 | // reader, renderer and writer SyncExec does, rather than a second copy of |
| 1409 | // them. The bare-hostname check stays with each caller, before any path is |
| 1410 | // derived. |
| 1411 | // |
| 1412 | // keepAbsent (config sync's pass, for a route with no instance overlay): |
| 1413 | // when the file is absent and the render is empty, nothing is written and |
| 1414 | // the render counts as up to date. An absent file and an empty one give a |
| 1415 | // spawn the same environment (the activator reads a missing config env as |
| 1416 | // none), and an absent file is what every never-configured app's instances |
| 1417 | // already have. Writing the empty file would be a config change for every |
| 1418 | // running instance: a restart, a delivered revision and an ack, for a |
| 1419 | // bundle that only gained settings boxes read from the bucket, such as the |
| 1420 | // swap policy (Astra plan 6 R1 #1, plan Q3). The held render (RenderHeld) |
| 1421 | // passes false: a move's ack hashes the file it wrote. |
| 1422 | func renderExecConfig(ctx context.Context, work hostwrite.Work, rt registry.Route, m *Manifest, priv *ecdh.PrivateKey, dataDir, header, instETag string, keepAbsent bool) (execRender, error) { |
| 1423 | return renderExecConfigBefore(ctx, work, rt, m, priv, dataDir, header, instETag, nil, keepAbsent) |
| 1424 | } |
| 1425 | |
| 1426 | // renderExecConfigBefore is renderExecConfig with before, when not nil, |
| 1427 | // called with the hex sha256 of the rendered bytes once they are known and |
| 1428 | // before the file is written or found up to date (RenderHeldBefore). An |
| 1429 | // error from it ends the render with nothing written. |
| 1430 | // |
| 1431 | // before and keepAbsent are independent. before runs as soon as the bytes |
| 1432 | // are rendered, ahead of both the up-to-date check and the keepAbsent skip, |
| 1433 | // so it sees every render whatever becomes of the write: it is a witness of |
| 1434 | // what the render holds, not of what was written. Its only callers are held |
| 1435 | // renders (RenderHeldBefore), which pass keepAbsent=false and so never skip; |
| 1436 | // config sync passes before=nil and keepAbsent per its overlay. A caller |
| 1437 | // that someday passes both gets before called with the hash of an empty |
| 1438 | // render that was then deliberately not written. |
| 1439 | func renderExecConfigBefore(ctx context.Context, work hostwrite.Work, rt registry.Route, m *Manifest, priv *ecdh.PrivateKey, dataDir, header, instETag string, before func(string) error, keepAbsent bool) (execRender, error) { |
| 1440 | path := ConfigEnvPath(dataDir, rt) |
| 1441 | // A nested row's current file lives on tenant-owned ground, so its read |
| 1442 | // goes through instancefs the same way its write below does: a plain |
| 1443 | // os.ReadFile would block forever on a FIFO a tenant planted at this |
| 1444 | // name (isolated-exec design §3.3). The flat directory is agent-owned |
| 1445 | // throughout and keeps the plain read it always used. |
| 1446 | var cur []byte |
| 1447 | var existed bool |
| 1448 | var err error |
| 1449 | if nestedRoute(dataDir, rt) { |
| 1450 | cur, err = instancefs.ReadBeneathLimit(filepath.Dir(path), filepath.Base(path), 1<<20) |
| 1451 | existed = err == nil |
| 1452 | if err != nil && !os.IsNotExist(err) { |
| 1453 | fmt.Printf(" %s: read %s: %v\n", rt.Host, path, err) |
| 1454 | return execRender{}, err |
| 1455 | } |
| 1456 | } else { |
| 1457 | cur, err = os.ReadFile(path) |
| 1458 | existed = err == nil |
| 1459 | if err != nil && !os.IsNotExist(err) { |
| 1460 | fmt.Printf(" %s: read %s: %v\n", rt.Host, path, err) |
| 1461 | return execRender{}, err |
| 1462 | } |
| 1463 | } |
| 1464 | // Where sync.go has its defaultMissing/lostKeys guard, this path has |
| 1465 | // nothing — deliberately. That guard exists because /etc/carlos/apps |
| 1466 | // files may be HAND-STAGED, so a bound-environment-only merged view |
| 1467 | // dropping keys could delete config no one put in the bucket. This |
| 1468 | // file is sync-owned by construction: every byte in it was rendered by |
| 1469 | // a previous pass of this function from a manifest. A merged view that |
| 1470 | // drops keys is therefore correct convergence, not truncation — and |
| 1471 | // the instance's own <host>.env, the one file here that IS |
| 1472 | // hand-staged, is a separate file layered on top at spawn and never |
| 1473 | // read or written by this sweep. |
| 1474 | // Deliberately NOT run through applyAgentVars (sync.go, spec §7.3): |
| 1475 | // this is <host>.config.env, one layer under the socket-exec |
| 1476 | // instance's own <host>.env, and the activator applies withVar/ |
| 1477 | // withoutVars to CARLOS_PURGE at spawn time (backend_exec.go's |
| 1478 | // Start) — cache-CONDITIONAL, set only when b.Cache != nil. Baking |
| 1479 | // CARLOS_PURGE=header into this file unconditionally would make the |
| 1480 | // var dishonest on disk on a cacheless box, the exact failure mode |
| 1481 | // agent-ownership exists to prevent. The asymmetry with sync.go's |
| 1482 | // unit path (which does apply it, unconditionally, because cache |
| 1483 | // construction is unconditional there) is correct; do not "fix" it. |
| 1484 | rendered, err := RenderEnv(m, priv, routeBinding(rt)) |
| 1485 | if err != nil { |
| 1486 | fmt.Printf(" %s: render config: %v\n", rt.Host, err) |
| 1487 | return execRender{}, err |
| 1488 | } |
| 1489 | if header != "" { |
| 1490 | rendered = append([]byte(header), rendered...) |
| 1491 | } |
| 1492 | out := execRender{path: path, rendered: rendered} |
| 1493 | if before != nil { |
| 1494 | sum := sha256.Sum256(rendered) |
| 1495 | if err := before(hex.EncodeToString(sum[:])); err != nil { |
| 1496 | return execRender{}, err |
| 1497 | } |
| 1498 | } |
| 1499 | if existed && bytes.Equal(cur, rendered) { |
| 1500 | return out, nil // up to date — stay quiet on the periodic tick |
| 1501 | } |
| 1502 | if keepAbsent && !existed && len(rendered) == 0 { |
| 1503 | return out, nil // absent, and nothing to put in it: as an unconfigured app's |
| 1504 | } |
| 1505 | id := renderedIdentity(rendered, true, rt.Environment != "", instETag != "") |
| 1506 | out.id = id |
| 1507 | // A nested row's config file lives inside its own instance directory, |
| 1508 | // tenant-owned ground (isolated-exec design §3.2). So it is written |
| 1509 | // through instancefs.WriteAtomic's O_EXCL temp name and openat2 |
| 1510 | // resolution rather than a predictable path.tmp: the same reasoning |
| 1511 | // backend_exec.go's own writes into that directory follow. The flat |
| 1512 | // directory is agent-owned throughout, so it keeps the plain |
| 1513 | // temp-then-rename it always used. |
| 1514 | if nestedRoute(dataDir, rt) { |
| 1515 | if err := guarded(ctx, work, id, func() error { return instancefs.WriteAtomic(filepath.Dir(path), filepath.Base(path), rendered, 0o600) }); err != nil { |
| 1516 | if _, lerr := os.Lstat(filepath.Dir(path)); errors.Is(lerr, os.ErrNotExist) { |
| 1517 | return execRender{}, fmt.Errorf("%w: %s", errInstanceDirAbsent, filepath.Dir(path)) |
| 1518 | } |
| 1519 | fmt.Printf(" %s: write %s: %v\n", rt.Host, path, err) |
| 1520 | return execRender{}, err |
| 1521 | } |
| 1522 | } else { |
| 1523 | tmp := path + ".tmp" |
| 1524 | if err := guarded(ctx, work, id, func() error { return os.WriteFile(tmp, rendered, 0o600) }); err != nil { |
| 1525 | fmt.Printf(" %s: write %s: %v\n", rt.Host, path, err) |
| 1526 | return execRender{}, err |
| 1527 | } |
| 1528 | if err := guarded(ctx, work, id, func() error { return os.Rename(tmp, path) }); err != nil { |
| 1529 | _ = guarded(ctx, work, id, func() error { return os.Remove(tmp) }) |
| 1530 | fmt.Printf(" %s: rename %s: %v\n", rt.Host, path, err) |
| 1531 | return execRender{}, err |
| 1532 | } |
| 1533 | } |
| 1534 | fmt.Printf(" %s: config updated\n", rt.Host) |
| 1535 | out.changed = true |
| 1536 | return out, nil |
| 1537 | } |
| 1538 | |
| 1539 | // bounceLogged calls bounce — best-effort in the sense that a failure is |
| 1540 | // logged rather than aborting the pass, the file on disk is already |
| 1541 | // correct and the next natural wake picks it up — and returns what bounce |
| 1542 | // itself returned, so a caller that also writes a config-ack (syncExecHost |
| 1543 | // above) can report whether THIS pass's restart actually succeeded rather |
| 1544 | // than silently discarding that fact. bounce == nil (the tcp-exec-file/ |
| 1545 | // addr shape, whose process this platform does not own) is a no-op |
| 1546 | // returning nil. |
| 1547 | func bounceLogged(host string, bounce func(host string) error) error { |
| 1548 | if bounce == nil { |
| 1549 | return nil |
| 1550 | } |
| 1551 | if err := bounce(host); err != nil { |
| 1552 | if errors.Is(err, ErrDeliveredOnWake) { |
| 1553 | fmt.Printf(" %s: config updated, asleep: delivered on wake\n", host) |
| 1554 | return err |
| 1555 | } |
| 1556 | fmt.Printf(" %s: bounce after config change: %v\n", host, err) |
| 1557 | return err |
| 1558 | } |
| 1559 | return nil |
| 1560 | } |
| 1561 | |
| 1562 | // reapConfigEnv deletes *.config.env (and stale *.config.env.tmp) for |
| 1563 | // hosts with no route row at all. It matches ONLY sync's own suffix: |
| 1564 | // <host>.env and <host>-sidecar.env are other owners' files and are |
| 1565 | // never candidates, whatever their host's route state. |
| 1566 | func reapConfigEnv(routes []registry.Route, dataDir string) { |
| 1567 | ents, err := os.ReadDir(dataDir) |
| 1568 | if err != nil { |
| 1569 | return |
| 1570 | } |
| 1571 | live := make(map[string]bool, len(routes)) |
| 1572 | for _, rt := range routes { |
| 1573 | live[rt.Host] = true |
| 1574 | } |
| 1575 | for _, e := range ents { |
| 1576 | if e.IsDir() { |
| 1577 | continue |
| 1578 | } |
| 1579 | if strings.HasSuffix(e.Name(), configEnvSuffix+".tmp") { |
| 1580 | // A crashed pass's leftover holds decrypted secrets and is |
| 1581 | // stale by construction once reap runs. |
| 1582 | if err := os.Remove(filepath.Join(dataDir, e.Name())); err != nil { |
| 1583 | fmt.Printf(" %s: remove stale config temp file: %v\n", e.Name(), err) |
| 1584 | } |
| 1585 | continue |
| 1586 | } |
| 1587 | if !strings.HasSuffix(e.Name(), configEnvSuffix) { |
| 1588 | continue |
| 1589 | } |
| 1590 | host := strings.TrimSuffix(e.Name(), configEnvSuffix) |
| 1591 | if live[host] { |
| 1592 | continue |
| 1593 | } |
| 1594 | if asInstanceEnvHost := strings.TrimSuffix(e.Name(), ".env"); live[asInstanceEnvHost] { |
| 1595 | // The name is ambiguous: <x>.config.env is host x's sync-owned |
| 1596 | // config file AND host x.config's hand-staged instance env |
| 1597 | // (<host>.env). With no route x but a live route x.config, the |
| 1598 | // bytes on disk may be the latter — another owner's file, never |
| 1599 | // deleted here. |
| 1600 | fmt.Printf(" %s: not reaped — live route %q claims this name as its instance env\n", e.Name(), asInstanceEnvHost) |
| 1601 | continue |
| 1602 | } |
| 1603 | if err := os.Remove(filepath.Join(dataDir, e.Name())); err != nil { |
| 1604 | fmt.Printf(" %s: remove orphaned config file: %v\n", host, err) |
| 1605 | continue |
| 1606 | } |
| 1607 | fmt.Printf(" %s: no route on this box, removed its config file\n", host) |
| 1608 | } |
| 1609 | reapNestedConfigEnv(live, dataDir) |
| 1610 | } |
| 1611 | |
| 1612 | // reapNestedConfigEnv is reapConfigEnv's nested half (isolated-exec design |
| 1613 | // §7): a nested row's rendered bundle lives inside its own instance |
| 1614 | // directory, one level under <dataDir>/instances, so a stale file there is |
| 1615 | // found by walking that tree rather than dataDir's own listing. The flat |
| 1616 | // directory's <host>.config / <host>.config.env name collision (reapConfigEnv's |
| 1617 | // own ambiguity guard, above) cannot happen inside an instance directory. |
| 1618 | // Nothing in there is named after some OTHER host, so no such guard is |
| 1619 | // needed here. |
| 1620 | func reapNestedConfigEnv(live map[string]bool, dataDir string) { |
| 1621 | root := filepath.Join(dataDir, "instances") |
| 1622 | dirs, err := os.ReadDir(root) |
| 1623 | if err != nil { |
| 1624 | return // no nested layout on this box (or nothing has migrated yet) |
| 1625 | } |
| 1626 | for _, d := range dirs { |
| 1627 | if !d.IsDir() { |
| 1628 | continue |
| 1629 | } |
| 1630 | instDir := filepath.Join(root, d.Name()) |
| 1631 | ents, err := os.ReadDir(instDir) |
| 1632 | if err != nil { |
| 1633 | continue |
| 1634 | } |
| 1635 | for _, e := range ents { |
| 1636 | if e.IsDir() { |
| 1637 | continue |
| 1638 | } |
| 1639 | // A nested config file is written through instancefs.WriteAtomic, |
| 1640 | // whose temp name is ".<name>.tmp-<hex>" beside the real name, not |
| 1641 | // "<name>.tmp" the way the flat directory's plain rename leaves |
| 1642 | // one: the leading dot and the "-<hex>" suffix are both load |
| 1643 | // bearing here, or a crashed pass's leftover (it holds decrypted |
| 1644 | // secrets) is never reaped. |
| 1645 | if strings.HasPrefix(e.Name(), ".") && strings.Contains(e.Name(), configEnvSuffix+".tmp-") { |
| 1646 | if err := os.Remove(filepath.Join(instDir, e.Name())); err != nil { |
| 1647 | fmt.Printf(" %s: remove stale config temp file: %v\n", e.Name(), err) |
| 1648 | } |
| 1649 | continue |
| 1650 | } |
| 1651 | if !strings.HasSuffix(e.Name(), configEnvSuffix) { |
| 1652 | continue |
| 1653 | } |
| 1654 | host := strings.TrimSuffix(e.Name(), configEnvSuffix) |
| 1655 | if live[host] { |
| 1656 | continue |
| 1657 | } |
| 1658 | if err := os.Remove(filepath.Join(instDir, e.Name())); err != nil { |
| 1659 | fmt.Printf(" %s: remove orphaned config file: %v\n", host, err) |
| 1660 | continue |
| 1661 | } |
| 1662 | fmt.Printf(" %s: no route on this box, removed its config file\n", host) |
| 1663 | } |
| 1664 | } |
| 1665 | } |
| 1666 | |
amadan