| 1 | // The `carlos agent` verb — the single-box role: edge (TLS + host routing) + |
| 2 | // hibernation activator + hourly release adoption (plus a 2s fast-converge |
| 3 | // loop, fastConvergeLoop: serial poll + canary/unstaged sweep) + the fleet status |
| 4 | // page's 5-minute self-report composed in one process. The composition is |
| 5 | // new here, but each part is a port: the edge from Eleven's router.go, the |
| 6 | // activator from lchat activator.go, the adopt loop is release.Pass (or, |
| 7 | // for canaries, release.PassFilter) on the cadence of Eleven's updater |
| 8 | // timer, and the status loop is internal/status's |
| 9 | // CollectBox/WriteBox/BuildFleet/PublishFleet run on a timer with one |
| 10 | // publisher elected via internal/lease. |
| 11 | // |
| 12 | // Fleet status page config surface (every env var any part of this file |
| 13 | // reads for it, gathered in one place since no single function owns the |
| 14 | // whole contract): |
| 15 | // |
| 16 | // - CARLOS_STATUS_BUCKET / CARLOS_STATUS_PREFIX / CARLOS_STATUS_REGION — |
| 17 | // read by statusLoop. CARLOS_STATUS_BUCKET names the S3 bucket each |
| 18 | // box writes its own status/boxes/<label>.json snapshot to and (when |
| 19 | // it wins the lease) publishes status/fleet.json + the viewer assets |
| 20 | // to. When CARLOS_STATUS_BUCKET is unset, it DEFAULTS to whatever |
| 21 | // CARLOS_DEPLOYMENT_BUCKET is set to (resolveStatusBucket) — the edge |
| 22 | // and console both serve the fleet status page out of the same |
| 23 | // deployment bucket adoptLoop already publishes releases to, so the |
| 24 | // common single-bucket deployment needs to set only one env var; an |
| 25 | // explicit CARLOS_STATUS_BUCKET still always wins, for the deployments |
| 26 | // that deliberately want a second, independent bucket. Both unset -> |
| 27 | // the loop never starts, logging one line and returning — the same |
| 28 | // "inert until configured" convention adoptLoop uses for |
| 29 | // CARLOS_DEPLOYMENT_BUCKET/_DIR, and litestream's own env-gated |
| 30 | // replication before that. CARLOS_STATUS_REGION feeds lease.Acquire |
| 31 | // (the single-writer election for the publish half of each tick); it |
| 32 | // may be empty (falls back to the AWS default chain). CARLOS_STATUS_PREFIX |
| 33 | // does double duty: it namespaces that same lease key (path.Join(prefix, |
| 34 | // "status-lease")) AND, since Task 5 of the customer-fleets plan, |
| 35 | // rebases every key statusLoop's AND metricsLoop's store touches |
| 36 | // (release.NewPrefixStore) — a fleet box set to |
| 37 | // CARLOS_STATUS_PREFIX=fleets/<name>/ writes its own status snapshot at |
| 38 | // fleets/<name>/status/boxes/<label>.json and its metrics/alerts under |
| 39 | // fleets/<name>/metrics/boxes/... and fleets/<name>/alerts/boxes/... |
| 40 | // instead of the shared paths every platform box also writes — so |
| 41 | // neither its status heartbeat nor its metrics/alerts ever collide with |
| 42 | // (or leak into) the platform's own fleet.json, metrics view, or alert |
| 43 | // webhook. Empty (every platform box today) is a pure passthrough — |
| 44 | // EXCEPT on a door-backed fleet box, where empty defaults to |
| 45 | // "fleets/<fleet>/" because the door admits no other write scope |
| 46 | // (resolveStatusPrefix). Log shipping is deliberately NOT rebased: see |
| 47 | // logship.go. A |
| 48 | // PUBLISHING box with a prefix set would relocate status/fleet.json |
| 49 | // itself under that prefix too — moot today since nothing sets a prefix |
| 50 | // on a publishing (non-fleet) box, and a fleet box never publishes at |
| 51 | // all (CARLOS_FLEET's guard, below). |
| 52 | // - CARLOS_FLEET — read by statusLoop (Task 5, customer-fleets plan). Set |
| 53 | // (to the fleet's name — its value is otherwise unused here) marks this |
| 54 | // agent as a fleet box: shouldSkipPublisherElection short-circuits |
| 55 | // publish to false BEFORE lease.Acquire is ever called, so a fleet box |
| 56 | // never contends for the cross-box publish lease and never runs |
| 57 | // BuildFleet/PublishFleet — it still self-reports via statusTick's |
| 58 | // unconditional WriteBox, just under its own CARLOS_STATUS_PREFIX. |
| 59 | // HONEST LIMIT: the CARLOS_DEPLOYMENT_BUCKET fallback only ever |
| 60 | // resolves to an S3 bucket. A deployment adopting from |
| 61 | // CARLOS_DEPLOYMENT_DIR (a plain directory — tests, or a single-box |
| 62 | // no-AWS deployment) has nothing for the fallback to resolve to, so |
| 63 | // DIR-based deployments do not publish fleet status yet; give them |
| 64 | // their own CARLOS_STATUS_BUCKET explicitly if they want the page. |
| 65 | // - CARLOS_FLEET_BOX — this box's operator-chosen fleet_boxes.label (the |
| 66 | // channel identity the edge dials, "<fleet>/<label>", NEVER the public |
| 67 | // status page's hashed pseudonym). Optional: defaults to |
| 68 | // CARLOS_TUNNEL_NAME with the "<fleet>/" prefix stripped, since a fleet |
| 69 | // box's tunnel name already IS "<fleet>/<label>" (see |
| 70 | // internal/host/tunnel_client.go). resolveFleetIdentity is the one |
| 71 | // place both env vars are read and validated (registry.ValidTunnelName) |
| 72 | // — an invalid result refuses to arm the wake-time/startup binding |
| 73 | // publish (fleetbind.go) loudly, and degrades the heartbeat's label back |
| 74 | // to the hashed pseudonym rather than disabling it outright. |
| 75 | // - CARLOS_TUNNEL_URL / CARLOS_TUNNEL_TOKEN — the tunnel env |
| 76 | // tunnel_client.go already dials with, read a SECOND time here (spec |
| 77 | // 2026-08-11-fleet-api-surface, F5) as the fleet store door's address |
| 78 | // and credential: on a fleet box with NO bucket env set, openDeployStore |
| 79 | // and the status/metrics store both become internal/fleetapi clients |
| 80 | // speaking release.Store over the edge's /store/ routes as |
| 81 | // "<fleet>/<label>", so the box needs no AWS credentials on the |
| 82 | // platform's deployment bucket at all. A bucket still set WINS |
| 83 | // everywhere (the migration fallback): every platform box, and any |
| 84 | // grandfathered fleet box that still carries bucket credentials, keeps |
| 85 | // direct S3 byte for byte. Missing tunnel env on a fleet box with no |
| 86 | // bucket is a loud error, not a silent degrade — that box has no way to |
| 87 | // reach a store at all. |
| 88 | // - STATUS_BOX_REGION — an operator-set, coarse region word (e.g. |
| 89 | // "EU · Ireland") carried into status.Probes.Region and published |
| 90 | // verbatim on this box's snapshot; "" omits the field. Deliberately |
| 91 | // free text chosen by the operator, never derived from cloud metadata — |
| 92 | // the public commitment is a human-readable region, not a specific |
| 93 | // datacenter. |
| 94 | // - CARLOS_SYSTEM_CHANNEL — platform channel this box follows for |
| 95 | // self-update, canary|stable (default stable). |
| 96 | // - CARLOS_LAYOUT_DEFAULT — the instance layout this box converges to |
| 97 | // when no layout intent is recorded, nested|flat (default nested; |
| 98 | // isolated-exec design §8.6, "Nested by default"). Nested applies only |
| 99 | // on a box whose isolation is ready; a recorded intent always wins. |
| 100 | // - STATUS_HOST / STATUS_PUBLIC — NOT read anywhere in this file. They |
| 101 | // gate whether ops/tofu fronts the bucket's published status/ prefix |
| 102 | // with a public hostname; recorded here only so the fleet status |
| 103 | // page's entire config surface is documented in one place rather than |
| 104 | // split across this file and the tofu module. |
| 105 | // |
| 106 | // The status-lease above is one of TWO elections this process runs, and the |
| 107 | // only per-tick one. The other is the HOUSEKEEPER lease (housekeeper.go): |
| 108 | // the same internal/lease mechanism, held long-term, at the fixed |
| 109 | // deployment-bucket key control/housekeeper-lease/lock.json, electing the one |
| 110 | // box that performs this deployment's global duties — today the releases |
| 111 | // retention prune (retentionLoop), from Part 2 of the multi-region spec |
| 112 | // certificate issuance. It reads CARLOS_DEPLOYMENT_BUCKET (or |
| 113 | // CARLOS_DEPLOYMENT_DIR, which makes the box its own housekeeper) and the |
| 114 | // same CARLOS_FLEET guard, and adds no env var of its own. |
| 115 | // |
| 116 | // Domains config surface (read by domainsLoop / domainsTickConfig): |
| 117 | // |
| 118 | // - PLATFORM_APPS_DOMAIN — the deployment's platform apps domain |
| 119 | // (flagship: oncarlos.com). UNSET = the domains feature is inert: no |
| 120 | // platform domains minted, no bucket manifests read, no health stamps — |
| 121 | // zero behavior change, same litestream-convention gate as the loops |
| 122 | // above. |
| 123 | // - CARLOS_WILDCARD_DOMAINS — comma list of static wildcard-cert parent |
| 124 | // domains for the edge (tito: go.tito.io); per-account parents under |
| 125 | // PLATFORM_APPS_DOMAIN are derived automatically (edge wildcard.go) |
| 126 | // - CARLOS_DNS_PROVIDER / CARLOS_DNSIMPLE_TOKEN — DNS-01 backend for |
| 127 | // wildcard certs: unset/dnsimple needs the token (SSM → unit env), |
| 128 | // route53 rides the box IAM role; nothing set = wildcards off, |
| 129 | // per-host HTTP-01 only |
| 130 | // - CARLOS_ACME_DELEGATION_ZONE — our own zone that delegated wildcard |
| 131 | // parents' challenge TXTs publish into (spec §4); unset or an invalid |
| 132 | // shape = delegation off, per-host and direct DNS-01 continue |
| 133 | // - PLATFORM_EDGE_IPS — comma-separated edge IPs custom-domain DNS checks |
| 134 | // compare against. Default: this box's public IPv4 from IMDS |
| 135 | // (best-effort). It is no longer the WHOLE comparison set: each pass |
| 136 | // unions it with the deployment's LIVE edges — every box's published |
| 137 | // control/edges/<box>.json whose status snapshot is still fresh |
| 138 | // (edgesfile.EdgeIPs) — because under latency steering a correctly- |
| 139 | // pointed domain resolves to whichever region's edge is nearest, and |
| 140 | // one box's boot-frozen list reads that as "elsewhere". The union only |
| 141 | // ever WIDENS, so an unreadable bucket degrades to exactly this |
| 142 | // variable. A pass that ends up knowing no addresses at all stands |
| 143 | // down rather than guessing "elsewhere" about every host. |
| 144 | package host |
| 145 | |
| 146 | import ( |
| 147 | "github.com/carlosframework/platform/internal/cliflag" |
| 148 | |
| 149 | "context" |
| 150 | "crypto/ecdh" |
| 151 | "crypto/ed25519" |
| 152 | "errors" |
| 153 | "flag" |
| 154 | "fmt" |
| 155 | "io" |
| 156 | "log" |
| 157 | "math/rand/v2" |
| 158 | "net" |
| 159 | "net/http" |
| 160 | "os" |
| 161 | "os/exec" |
| 162 | "os/signal" |
| 163 | "path" |
| 164 | "path/filepath" |
| 165 | "runtime" |
| 166 | "slices" |
| 167 | "sort" |
| 168 | "strings" |
| 169 | "sync" |
| 170 | "sync/atomic" |
| 171 | "syscall" |
| 172 | "time" |
| 173 | |
| 174 | "github.com/carlosframework/platform/internal/activator" |
| 175 | "github.com/carlosframework/platform/internal/agentsock" |
| 176 | "github.com/carlosframework/platform/internal/analytics" |
| 177 | "github.com/carlosframework/platform/internal/appconfig" |
| 178 | "github.com/carlosframework/platform/internal/apperr" |
| 179 | "github.com/carlosframework/platform/internal/applog" |
| 180 | "github.com/carlosframework/platform/internal/caps" |
| 181 | "github.com/carlosframework/platform/internal/cgroup" |
| 182 | "github.com/carlosframework/platform/internal/controlsock" |
| 183 | "github.com/carlosframework/platform/internal/domains" |
| 184 | "github.com/carlosframework/platform/internal/edge" |
| 185 | "github.com/carlosframework/platform/internal/edgesfile" |
| 186 | "github.com/carlosframework/platform/internal/fleetapi" |
| 187 | "github.com/carlosframework/platform/internal/hostbox" |
| 188 | "github.com/carlosframework/platform/internal/hostwrite" |
| 189 | "github.com/carlosframework/platform/internal/instancectl" |
| 190 | "github.com/carlosframework/platform/internal/instancefs" |
| 191 | "github.com/carlosframework/platform/internal/layout" |
| 192 | "github.com/carlosframework/platform/internal/lease" |
| 193 | "github.com/carlosframework/platform/internal/moves" |
| 194 | "github.com/carlosframework/platform/internal/provision" |
| 195 | "github.com/carlosframework/platform/internal/registry" |
| 196 | "github.com/carlosframework/platform/internal/release" |
| 197 | "github.com/carlosframework/platform/internal/rootparticipation" |
| 198 | "github.com/carlosframework/platform/internal/rootstate" |
| 199 | "github.com/carlosframework/platform/internal/schedule" |
| 200 | "github.com/carlosframework/platform/internal/status" |
| 201 | "github.com/carlosframework/platform/internal/version" |
| 202 | ) |
| 203 | |
| 204 | // dataDir resolves the fleet data directory (CARLOS_DATA_DIR, default /data). |
| 205 | func dataDir() string { |
| 206 | if d := os.Getenv("CARLOS_DATA_DIR"); d != "" { |
| 207 | return d |
| 208 | } |
| 209 | return "/data" |
| 210 | } |
| 211 | |
| 212 | // RunAgent runs edge + activator + adoption in one process. Flags are exactly |
| 213 | // edge.Run's (-registry, -dev, -http, -https, ...): the args pass through to |
| 214 | // edge.RunWithActivator, which owns the serving loop and blocks until |
| 215 | // SIGTERM/SIGINT. |
| 216 | // |
| 217 | // With CARLOS_HOST_CLAIM and CARLOS_REGION_HA on, a box that is not yet |
| 218 | // enrolled with its guard armed runs region HA's joining mode instead, and |
| 219 | // RunAgent returns once it can host, so systemd starts it again as the |
| 220 | // claim agent (hostingGate, joining.go). |
| 221 | // |
| 222 | // Before either, a box its region's group launched resolves its group |
| 223 | // reference, whatever the switches say: the loops runAgent starts read |
| 224 | // the box's label once and keep it (agentEntry, joining.go). |
| 225 | func RunAgent(args []string) { |
| 226 | agentEntry(os.Getenv, resolveGroupBoxReference, groupReferenceBackoff, log.Printf, |
| 227 | func() bool { return hostingGate(args) }, |
| 228 | func() { runAgent(args, productionClaimFactory) }) |
| 229 | } |
| 230 | |
| 231 | func runAgent(args []string, buildClaims ClaimFactory) { |
| 232 | agentCtx, cancelAgent := context.WithCancel(context.Background()) |
| 233 | defer cancelAgent() |
| 234 | // Startup can stop before children drain; rollout keeps agentCtx until Close. |
| 235 | startupCtx, cancelStartup := context.WithCancel(agentCtx) |
| 236 | defer cancelStartup() |
| 237 | claimCtx, err := claimEntryContext(agentCtx, os.Getenv, openDeployStore) |
| 238 | if err != nil { |
| 239 | log.Fatalf("agent: host claims: %v", err) |
| 240 | } |
| 241 | dd := dataDir() |
| 242 | regPath := peekRegistryPath(args, dd) |
| 243 | // A second handle on the same registry file the edge opens — the registry |
| 244 | // is designed for concurrent handles (fresh-connection reads), and this |
| 245 | // one feeds the activator's due-time index and the adopt loop. |
| 246 | reg, err := registry.Open(regPath) |
| 247 | if err != nil { |
| 248 | if hint := registryOpenHint(regPath, err); hint != "" { |
| 249 | log.Fatalf("agent: open registry %s: %v\nagent: %s", regPath, err, hint) |
| 250 | } |
| 251 | log.Fatalf("agent: open registry %s: %v", regPath, err) |
| 252 | } |
| 253 | |
| 254 | // Constructed once here so the activator's ExecBackend (purges) and the |
| 255 | // edge (reads + serves cached hits) share the very same store — see the |
| 256 | // package doc comment above about caching applying before the |
| 257 | // hibernation branch. |
| 258 | cache := edge.NewCache(edge.DefaultCacheMaxEntries, edge.DefaultCacheMaxBodyBytes) |
| 259 | |
| 260 | // Exec-child accounting: claim the cgroup subtree the unit file |
| 261 | // delegates to us (Delegate= in carlos-edge.service). Nil when there is |
| 262 | // no delegation — an old unit file, a dev run outside systemd, darwin — |
| 263 | // in which case children spawn exactly as before and simply have no |
| 264 | // per-instance readings. |
| 265 | var cgm *cgroup.Manager |
| 266 | if hostwrite.Enabled(claimEntryEnv(claimCtx)) { |
| 267 | cgm = cgroup.SetupRetaining(log.Printf) |
| 268 | } else { |
| 269 | cgm = cgroup.Setup(log.Printf) |
| 270 | } |
| 271 | |
| 272 | // Alert sinks and the operator paging hook (moves spec §13), built from |
| 273 | // the sink env alone and armed here, before armMovesPass and before any |
| 274 | // loop that runs a moves pass starts, so a stuck mark never finds paging |
| 275 | // unarmed. They do not wait on the metrics store: the metrics loop is |
| 276 | // handed this same instance, and a box whose store fails still pages. |
| 277 | alerts := armAgentAlerts(agentCtx, os.Getenv, &processAlertCounters, newSNSClient) |
| 278 | |
| 279 | // Log shipping (`carlos logs`): one sink for every producer in this |
| 280 | // process, armed only when there is somewhere to ship to — a nil sink |
| 281 | // leaves every producer exactly as before (no tees, no pipes, no |
| 282 | // events). The box identity is the one metrics publishes under. |
| 283 | // |
| 284 | // The gate is openStatusStore, not a bucket test: a cut-over fleet box has |
| 285 | // no bucket at all and ships through the store door, and gating on the |
| 286 | // bucket alone would leave `carlos logs` silently empty for exactly the |
| 287 | // boxes an operator can't SSH into. Constructing the store here to throw |
| 288 | // it away is deliberate and cheap (neither branch makes a call), and it |
| 289 | // asks precisely the question logshipLoop will ask again for itself. |
| 290 | var logSink *applog.Sink |
| 291 | var errSink *apperr.Sink |
| 292 | if _, _, err := openStatusStore(context.Background(), os.Getenv); err == nil { |
| 293 | logSink = applog.NewSink(status.Label(readMachineID()), time.Now) |
| 294 | errSink = apperr.NewSink(status.Label(readMachineID()), time.Now) |
| 295 | } |
| 296 | |
| 297 | cfg, err := activator.ConfigFromEnv() |
| 298 | if err != nil { |
| 299 | log.Fatalf("agent: %v", err) |
| 300 | } |
| 301 | claimCtx = withClaimRuntimeResources(claimCtx, ClaimRuntimeResources{Mode: "agent", Registry: reg, Manager: cgm, DataDir: dd, OptDir: release.OptDir(), UnitsDir: "/etc/systemd/system"}) |
| 302 | // The moves park providers (Plan 5 Task 9a) register before the claim |
| 303 | // assembly reads them, then the moves wake source right after them |
| 304 | // (Plan 5 Task 9b): only with both switches on. The parks refuse by |
| 305 | // name until the wake source registers, which opens the one gate |
| 306 | // (movesWakeSourceRegistered), so the parks arm in the same start as |
| 307 | // the wake owners. Task 5's expected refusals are quiet in the regional |
| 308 | // worker's log and its anomalies page once (movesrefusals.go). |
| 309 | movesRefusalLog := newMovesRefusals(metricsBoxLabel(os.Getenv, readMachineID()), PageOperator, log.Printf, time.Now) |
| 310 | if parks, err := registerProductionMovesParks(claimCtx, claimEntryEnv(claimCtx), claimStore(claimCtx), *cfg, productionRosterRegion(claimStore(claimCtx), os.Getenv), movesRefusalLog); err != nil { |
| 311 | log.Printf("agent: moves: the park providers are not registered: %v", err) |
| 312 | } else if err := registerProductionMovesWakeSource(claimEntryEnv(claimCtx), parks, dd); err != nil { |
| 313 | log.Printf("agent: moves: the wake source is not registered: %v", err) |
| 314 | } |
| 315 | // The C8 files manifest provider (moves Task 6) registers under the |
| 316 | // claim switch alone: every claim-on serving wake establishes its files |
| 317 | // manifest through it (claimfiles.go). |
| 318 | if err := registerProductionClaimFiles(claimEntryEnv(claimCtx), dd, reg); err != nil { |
| 319 | log.Printf("agent: host claims: the files manifest provider is not registered: %v", err) |
| 320 | } |
| 321 | if hostwrite.Enabled(claimEntryEnv(claimCtx)) { |
| 322 | // Region HA (stage B Task 10): with CARLOS_REGION_HA on, the |
| 323 | // relocation owner source is registered before the claim's owners |
| 324 | // open. In claim mode the assembly is opened here, as the claim |
| 325 | // factory would open it, so the relocation pass, or with region HA |
| 326 | // off the completion pass (ruling 82), binds to the same one. |
| 327 | if regionHAEnabled(claimEntryEnv(claimCtx)) { |
| 328 | if err := registerRelocationSource(claimCtx, dd); err != nil { |
| 329 | log.Fatalf("agent: region HA: %v", err) |
| 330 | } |
| 331 | } |
| 332 | if claimCtx, err = ensureProductionClaimAssembly(claimCtx); err != nil { |
| 333 | log.Fatalf("agent: host claims: %v", err) |
| 334 | } |
| 335 | } |
| 336 | claims, err := newClaimWiring(claimCtx, claimEntryEnv(claimCtx), buildClaims) |
| 337 | if err != nil { |
| 338 | log.Fatalf("agent: host claims: %v", err) |
| 339 | } |
| 340 | if err := configureWakeClaims(claims, cfg); err != nil { |
| 341 | log.Fatalf("agent: host claims: %v", err) |
| 342 | } |
| 343 | // The build-state store (settings spec §4.2), built once after the claim |
| 344 | // wiring so it locates claimed hosts through cfg.Claims. nil, nil on a box |
| 345 | // with no replica bucket. The activator's BeforeExec hook comes from bs |
| 346 | // (buildStateHook, below): a failure here is one operator line, the |
| 347 | // agent keeps running, and every spawn that records its build is refused |
| 348 | // until it is fixed: every exec-backed wake, worker, sidecar, swap and |
| 349 | // claimed start (Tasks 5a-5d). A unit-backed row is started by systemd |
| 350 | // on its live link and records nothing (activator identityFor is nil |
| 351 | // for it), so it still starts. |
| 352 | bs, bsErr := newBuildStates(cfg, reg, release.OptDir()) |
| 353 | // The status report says this box enforces update settings only while |
| 354 | // the build-state store is up (settings spec §4.2; Astra I3 #1). |
| 355 | noteBuildStatesInit(bsErr) |
| 356 | if bsErr != nil { |
| 357 | log.Printf("agent: build state store unavailable: %v. every exec-backed instance, worker and sidecar spawn is refused until it is fixed and the agent restarted (unit-backed rows still start)", bsErr) |
| 358 | } |
| 359 | |
| 360 | // Fleet box identity (customer-fleets plan, Task 6 fix round — controller |
| 361 | // ruling): resolved ONCE, here, before the activator is even constructed. |
| 362 | // identErr is errNotFleetBox for the ordinary platform-box case |
| 363 | // (CARLOS_FLEET unset — no log, nothing to arm) or names the specific |
| 364 | // validation failure for a fleet box whose identity doesn't resolve to a |
| 365 | // usable channel label, in which case the fleet machinery — OnWake and |
| 366 | // the startup pass, both — stays disarmed (loud log) rather than run |
| 367 | // with a label the edge could never dial. fleetStore is opened here too |
| 368 | // (once) so it can be threaded into svc.OnWake BEFORE svc.Start() below — |
| 369 | // assigning it after Start() would race the sweeper goroutine's own use |
| 370 | // of the field. |
| 371 | var fleetStore release.Store |
| 372 | identity, identErr := resolveFleetIdentity(os.Getenv) |
| 373 | switch { |
| 374 | case errors.Is(identErr, errNotFleetBox): |
| 375 | // Ordinary platform box; nothing to arm. |
| 376 | case identErr != nil: |
| 377 | log.Printf("agent: %v — refusing to arm fleet machinery", identErr) |
| 378 | default: |
| 379 | fs, err := openDeployStore(context.Background()) |
| 380 | if err != nil { |
| 381 | log.Printf("agent: CARLOS_FLEET=%s set but deploy store unavailable: %v — fleet bindings disabled", identity.Fleet, err) |
| 382 | } else { |
| 383 | fleetStore = fs |
| 384 | } |
| 385 | } |
| 386 | |
| 387 | // The config guard (amadan #213), armed before the activator and every |
| 388 | // reconcile loop that asks it: a box takes an instance's lease, or mints |
| 389 | // its record, only if it can provide that instance's config. Armed |
| 390 | // whether or not the activator is, because the mint gate is the reconcile |
| 391 | // pass's. A box with no deployment store reads it as "nothing sealed |
| 392 | // anywhere" and admits as before. |
| 393 | guard := newConfigGuard(os.Getenv, openDeployStore) |
| 394 | armConfigGuard(guard) |
| 395 | guard.Announce() |
| 396 | |
| 397 | var act edge.Activator |
| 398 | var svc *activator.Service |
| 399 | started := make(chan struct{}) |
| 400 | var startupDone <-chan struct{} |
| 401 | if cfg != nil { |
| 402 | // Exec children and carlos-app@ unit tenants share one state machine; |
| 403 | // the split backend dispatches per route on Backing. Note the arming |
| 404 | // gate is still the hibernate bucket env even though unit tenants |
| 405 | // never touch the bucket — one opt-in switch for the whole activator. |
| 406 | eb := &activator.ExecBackend{ |
| 407 | Cfg: *cfg, |
| 408 | OptDir: release.OptDir(), |
| 409 | DataDir: dd, |
| 410 | // The box's placement code, into every exec child's env |
| 411 | // (issue #329) — same env the reconcile pass scopes by. |
| 412 | Region: strings.TrimSpace(os.Getenv("CARLOS_REGION")), |
| 413 | Cache: cache, |
| 414 | Materialize: materializeHook(reg, dd), |
| 415 | Cgroups: cgm, |
| 416 | // The memory tier for each spawned child's cgroup, read off the |
| 417 | // staged manifest of the route's live version (memory-pressure |
| 418 | // spec §2.3): 512M, or 1G for a family that declares children. |
| 419 | Tier: memTier(stagedManifestOfLiveVersion(release.OptDir())), |
| 420 | // The same read for one named version, for a generation that runs |
| 421 | // a build other than the live link's (zero-downtime spec §7.3). |
| 422 | TierOf: memTierOf(func(rt registry.Route, v string) (*release.Manifest, error) { |
| 423 | return release.ReadStagedManifest(release.OptDir(), release.RouteAccount(rt), v, rt.App) |
| 424 | }), |
| 425 | // The scheduled-work verbs on every spawned instance's control |
| 426 | // socket write to the SAME registry this agent's converger and |
| 427 | // runner read (schedule spec §7.2): an app that asks for a |
| 428 | // once-timer from inside itself is asking this box's timer table, |
| 429 | // so the row is due-indexed by the next tick with nothing in |
| 430 | // between. reg satisfies schedule.TimerStore directly. |
| 431 | Timers: reg, |
| 432 | // The control socket's read-only GET /updates answers from the |
| 433 | // same registry copy (instance update settings spec §5.1). |
| 434 | UpdatesReader: updatesReader(reg, release.OptDir()), |
| 435 | // Asked inside Acquire, before the lease (amadan #213). |
| 436 | ConfigGate: configGate(), |
| 437 | } |
| 438 | if claims.Writes.Enabled { |
| 439 | eb.Timers = schedule.ClaimTimers{TimerStore: reg, Writes: claims.Writes, Context: agentCtx} |
| 440 | } |
| 441 | if logSink != nil { // a typed-nil in the interface would NOT be nil |
| 442 | eb.Logs = logSink |
| 443 | } |
| 444 | // The verified home keys of this box's replica rows (L21 rulings 26 |
| 445 | // and 30), read again at every replica launch until O8 reads them |
| 446 | // from each home's published env item, and the bucket's |
| 447 | // configuration, read fresh before a mint (ruling 5) and at each |
| 448 | // replica launch whose key is configured (ruling 27). The claim |
| 449 | // path's backend gets the same pair (claimExecBackend). |
| 450 | productionInstanceKeyReaders().wire(eb) |
| 451 | // The root instance-control channel (design §5.3) and the box facts |
| 452 | // a nested wake needs beyond the row (design §8.1, Task 6): the |
| 453 | // journal check and the peer list both read this SAME registry |
| 454 | // handle, so a nested row's classification here always agrees with |
| 455 | // what the rest of this process already knows about the box. |
| 456 | eb.Control = instancectl.NewClient(controlsock.Path) |
| 457 | eb.LayoutFacts = activator.LayoutInputs{ |
| 458 | DataDir: dd, |
| 459 | // A journal that cannot be read is not proof there isn't one: an |
| 460 | // EACCES or EIO on /data/layout must fence the row rather than |
| 461 | // wake it as if nothing were in flight, so only a definite |
| 462 | // ErrNotExist counts as "no journal". It is the pass's own |
| 463 | // predicate, an Lstat, so a symlink planted at the journal path |
| 464 | // fences the host rather than being followed. |
| 465 | Journal: func(host string) bool { return layout.JournalPresent(dd, host) }, |
| 466 | Peers: reg.List, |
| 467 | } |
| 468 | // The layout pass is handed this same gate below, so it never makes |
| 469 | // a nested tenant this backend would refuse to start. The gate's |
| 470 | // first question is about this process itself: the probe of design |
| 471 | // §8.5 runs under the instance templates' unit text and cannot see |
| 472 | // the edge's, and RestrictSUIDSGID= on carlos-edge.service once made |
| 473 | // every openat2 the agent issues answer ENOSYS (#126). A box in |
| 474 | // that state wakes nothing nested and migrates nothing, and says |
| 475 | // why, instead of failing on the first host's first open. |
| 476 | selfCheck := func() error { return instancefs.SelfCheck(dd) } |
| 477 | if err := selfCheck(); err != nil { |
| 478 | log.Printf("agent: %v (nested wakes and the layout pass refuse until the unit is fixed)", err) |
| 479 | } |
| 480 | eb.NestedOK = nestedGate(selfCheck, instancefs.MountNoSymfollow, eb.Control.Ready) |
| 481 | // CARLOS_SWAP=overlap turns on generations (plan 1), and with them |
| 482 | // plan 3's Swap, before the Service exists, which reads the switch |
| 483 | // once in New. Plan 2's per-generation containment goes on with |
| 484 | // them: an overlap needs each generation in its own child. A box |
| 485 | // that cannot contain generations refuses the flag here and stays |
| 486 | // on today's paths (swapArming), so all three stay off. |
| 487 | swapOn := swapArming(os.Getenv, eb) |
| 488 | activator.SetGenerations(swapOn) |
| 489 | eb.GenContainment = swapOn |
| 490 | rolloutMaterialize = eb.Materialize |
| 491 | svc = activator.New(*cfg, &activator.SplitBackend{ |
| 492 | Exec: eb, |
| 493 | Unit: &activator.UnitBackend{OptDir: release.OptDir()}, |
| 494 | }, reg) // the registry satisfies activator.DueStore directly |
| 495 | // doWake's own re-read of the committed row before it spawns |
| 496 | // (design §6.8). The registry satisfies RouteLookup with its own |
| 497 | // Get, exactly as DueStore above. |
| 498 | svc.Routes = reg |
| 499 | // Plans 2: an account's plan can shorten its entries' idle window. With |
| 500 | // plans off (no marker, CARLOS_PLAN_LIMITS=off, or no record) the hook |
| 501 | // answers 0 and every window is the box's own. |
| 502 | svc.SetIdleFor(planIdleFor) |
| 503 | // Settings spec §4.2: before any process runs a build against an |
| 504 | // instance's database, the build state names it, under the lease. A |
| 505 | // store that failed to start refuses every spawn (never "run |
| 506 | // unrecorded"). Set before Start, like Routes. |
| 507 | svc.BeforeExec = buildStateHook(bs, bsErr) |
| 508 | // The held wake check (settings spec §4.3 rule 3): a held instance |
| 509 | // never wakes on an installed build other than its target. The copy |
| 510 | // is refreshed before every wake that could be held (wakeTarget), and |
| 511 | // a target not installed is asked of root through the stage relay, |
| 512 | // one throttled requester for every wake. As every stage relay, not |
| 513 | // in host-writes mode (ruling R-AGENT-RELAY; release's OnStageBlocked |
| 514 | // callers): there a held wake waits for root's own pass. A box with |
| 515 | // no build-state store keeps today's wake. |
| 516 | if bs != nil { |
| 517 | svc.HeldTarget = func(h string) activator.HeldWake { return bs.wakeTarget(context.Background(), h) } |
| 518 | if !cfg.HostWrites.Enabled { |
| 519 | askStage := stageRequester(dataDir()) |
| 520 | // A wake waiting for root's decision on new settings names |
| 521 | // no build (ruling R-HELD-PURE): one marker for a root pass, |
| 522 | // nothing the relay would keep asking for. |
| 523 | askPass := stageMarkerRequester(dataDir()) |
| 524 | svc.RequestStage = func(host, build string) { |
| 525 | if build == "" { |
| 526 | askPass(host, build, nil) |
| 527 | return |
| 528 | } |
| 529 | askStage(host, build, nil) |
| 530 | } |
| 531 | } |
| 532 | } |
| 533 | // Every worker's captured serving build, published for root's pin |
| 534 | // set (zero-downtime deploys spec §4.2, ruling R17), whatever |
| 535 | // CARLOS_SWAP says. dd is dataDir(), the same data dir root's |
| 536 | // RunAdopt and this agent's passes hand release.BakeDir, which is |
| 537 | // where the pass reads release.GenerationsPath from. Set before |
| 538 | // Start and before any wake. |
| 539 | svc.GenerationsFile = release.GenerationsPath(dd) |
| 540 | // A retiring generation's schedule writes are refused once a swap |
| 541 | // has switched away from it, through each host's write fence, which |
| 542 | // the switch takes (zero-downtime spec §7.10). Wired before Start, |
| 543 | // like Routes. With swaps off nothing is ever retired or serialized, |
| 544 | // so every write applies as before. |
| 545 | eb.ScheduleFence = svc.ScheduleWrite |
| 546 | if fleetStore != nil { |
| 547 | // MUST be assigned before Start() below spawns the sweeper |
| 548 | // goroutine, which reads svc.OnWake on every successful wake from |
| 549 | // then on — setting it any later is a data race. |
| 550 | svc.OnWake = fleetBindingHook(fleetStore, identity.Fleet, identity.Label) |
| 551 | } |
| 552 | if logSink != nil { |
| 553 | svc.Logs = logSink |
| 554 | } |
| 555 | |
| 556 | // The startup sweep (design §6.7): a leftover instance unit from a |
| 557 | // previous agent process, one this box's registry no longer holds |
| 558 | // a live wake for, is stopped before this process starts spawning |
| 559 | // its own. Best-effort, like every other startup duty here: a |
| 560 | // channel that never answers logs and moves on, and the box is no |
| 561 | // worse off than before this existed (a leftover unit that keeps |
| 562 | // running is exactly what an agent restart left before this pass). |
| 563 | // |
| 564 | // Then, on a legacy box, layout recovery before Start can wake |
| 565 | // anything, and the instance-key backfill in the background after |
| 566 | // it, which this does not wait for (the edge below must listen |
| 567 | // within the update gate, whatever the box's size); on a claim |
| 568 | // box, Start alone. The order and the branch are activatorStart's. |
| 569 | svc.Startup = started |
| 570 | startupDone = activatorStart{ |
| 571 | Configure: func(_ context.Context) { |
| 572 | armAlways(svc, reg, dd) // flag on or off |
| 573 | var rolloutSink activator.EventSink |
| 574 | if logSink != nil { // a typed nil in the interface would not be nil |
| 575 | rolloutSink = logSink |
| 576 | } |
| 577 | armRolloutAndRelay(agentCtx, os.Getenv, svc, reg, dd, rolloutSink, openDeployStore) |
| 578 | // The residence-move machinery (issue #331), armed with the |
| 579 | // activator: the staging-fallback restore needs a same-region |
| 580 | // mirror client, and the move worker needs the service's quiesce |
| 581 | // answer. Both are best-effort — a mirror client that cannot |
| 582 | // build logs and leaves the fallback dark, exactly like every |
| 583 | // other optional duty here. |
| 584 | if m, err := lease.NewS3Mirror(agentCtx, cfg.Region, cfg.Region); err != nil { |
| 585 | log.Printf("agent: staging-fallback mirror client: %v (restore fallback disarmed)", err) |
| 586 | } else { |
| 587 | eb.Mirror = m |
| 588 | } |
| 589 | // Seeded instances (seeded-instances spec §3.4): the seed reader, |
| 590 | // the placed recorder, the provision pass's hooks and the claim |
| 591 | // flag the seed fallback refuses under. In Configure, like the |
| 592 | // mirror, as it was before the startup moved to the background |
| 593 | // (#281): before the sweep, before Start, and so before Ready |
| 594 | // closes the startup latch. Every wake waits on that latch |
| 595 | // (awaitStartup) and the reconcile loops start after it, so no |
| 596 | // seeded row is woken, and no seeded record acted on, before |
| 597 | // this has run. The import swap and rollback (spec §5.4, §5.6) |
| 598 | // fence through this activator, delete the replica under |
| 599 | // cfg.Prefix, and trust an unproven stop only after the startup |
| 600 | // sweep below (seedSweptAtStart, set in Recover, which also |
| 601 | // comes before the latch). |
| 602 | wireSeedPlacement(eb, claims.Writes.Enabled, seedDataOpener(cfg.Bucket, cfg.Region), |
| 603 | func() (release.Store, error) { return openDeployStore(context.Background()) }, |
| 604 | seedBox{Fence: svc, Prefix: cfg.Prefix, Swept: seedSweptAtStart.Load, NextFenceGen: fenceGenerations(reg)}) |
| 605 | armMovesPass(*cfg, dd, reg, svc, eb, claims.Writes, claims.Providers, rolloutSink, moveCarry(bs, bsErr)) |
| 606 | // carlos run's worker, armed with the moves pass as it was |
| 607 | // before the startup moved to the background (#281). Its |
| 608 | // passes ride the converge and instances loops, which the |
| 609 | // duties start only once Ready has closed the startup |
| 610 | // latch: no run is picked up before the startup sweep has |
| 611 | // stopped a previous agent's leftover units, the cmd units |
| 612 | // among them. |
| 613 | armRunsPass(svc, reg, dd, rolloutSink) |
| 614 | // The db-reads loop (db region forward spec §7.1): armed |
| 615 | // only on a box that can serve; it takes nothing until a |
| 616 | // console asks. |
| 617 | armDBReads(*cfg) |
| 618 | }, |
| 619 | Ready: func() { |
| 620 | close(started) |
| 621 | log.Printf("agent: hibernation activator armed (bucket %s, idle %s)", cfg.Bucket, cfg.Idle) |
| 622 | }, |
| 623 | Claims: claims.Writes.Enabled, |
| 624 | Sweep: func(ctx context.Context) error { |
| 625 | sw, err := eb.Control.Sweep(ctx) |
| 626 | if err != nil { |
| 627 | log.Printf("agent: instance sweep: %v (nested wakes refuse until the channel answers)", err) |
| 628 | } else if len(sw.Units) > 0 { |
| 629 | log.Printf("agent: swept %d instance unit(s) left from a previous agent: %v", len(sw.Units), sw.Units) |
| 630 | } |
| 631 | return err |
| 632 | }, |
| 633 | Recover: func(ctx context.Context, swept bool) *layout.Pass { |
| 634 | seedSweptAtStart.Store(swept) |
| 635 | pass := newLayoutPass(reg, svc, dd, eb.NestedOK, swept) |
| 636 | intent := startupRecoveryIntent(ctx, dd, openDeployStore, metricsBoxLabel(os.Getenv, readMachineID())) |
| 637 | recoverLayoutAtStart(ctx, pass, intent) |
| 638 | return pass |
| 639 | }, |
| 640 | Routes: reg.List, |
| 641 | Route: reg.Get, |
| 642 | Keys: eb, |
| 643 | Slot: svc.PrepareScheduledToken, |
| 644 | Publish: armedLayoutPass.Store, |
| 645 | Start: svc.Start, |
| 646 | }.Run(startupCtx) |
| 647 | act = svc |
| 648 | |
| 649 | // The agent socket: the one channel a sidecar dials back over |
| 650 | // (internal/agentsock, Task 7). Constructed only in this branch — |
| 651 | // the same condition that constructs the activator — because with |
| 652 | // hibernation off there are no workers on the box at all, and |
| 653 | // nothing would ever call in. cfg.AgentSocket is never empty here: |
| 654 | // activator.ConfigFromEnv defaults it to DefaultAgentSocket. |
| 655 | // |
| 656 | // The cancel is deferred right here, not called explicitly at the |
| 657 | // shutdown site below: RunAgent's defers run LIFO at return, so this |
| 658 | // one (registered first) fires AFTER the shutdown block's own |
| 659 | // `defer cancel()` for svc.Shutdown's context — i.e. once the |
| 660 | // children are already down — while staying the one place this |
| 661 | // context's lifetime is decided, which is what satisfies go vet's |
| 662 | // lostcancel check. |
| 663 | agentSockCtx, agentSockCancel := context.WithCancel(context.Background()) |
| 664 | defer agentSockCancel() |
| 665 | agentSrv := agentsock.New(svc) |
| 666 | go serveAgentSocket(agentSockCtx, cfg.AgentSocket, agentSrv.Serve) |
| 667 | log.Printf("agent: agent socket armed at %s", cfg.AgentSocket) |
| 668 | |
| 669 | // The boot pass (Task 8): a sidecar that does not hibernate must be |
| 670 | // running whenever the agent is — an agent restart otherwise leaves |
| 671 | // it down until a request or its due time happens to arrive, and an |
| 672 | // always-on worker has no request path at all to wake it. Run off |
| 673 | // the startup path (its own goroutine, not blocking RunAgent) since |
| 674 | // WakeWorker can block for the settle window per host and this must |
| 675 | // never delay the edge coming up; runWorkerBootPass itself wakes |
| 676 | // its candidates concurrently, bounded, rather than serially. |
| 677 | go func() { |
| 678 | select { |
| 679 | case <-started: |
| 680 | case <-startupCtx.Done(): |
| 681 | return |
| 682 | } |
| 683 | routes, err := reg.List() |
| 684 | if err != nil { |
| 685 | log.Printf("agent: worker boot pass: list routes: %v", err) |
| 686 | return |
| 687 | } |
| 688 | runWorkerBootPass(routes, svc.WakeWorker) |
| 689 | }() |
| 690 | } |
| 691 | |
| 692 | if svc == nil { |
| 693 | close(started) |
| 694 | } |
| 695 | |
| 696 | seedEnabled := claims.Writes.Enabled && os.Getenv("CARLOS_HOST_CLAIM_ORIGINS") == "1" && svc != nil |
| 697 | safeClaimClose := true |
| 698 | var shutdownCtx context.Context |
| 699 | var shutdownCancel context.CancelFunc |
| 700 | defer func() { |
| 701 | if shutdownCancel != nil { |
| 702 | defer shutdownCancel() |
| 703 | } |
| 704 | cancelAgent() |
| 705 | if !safeClaimClose { |
| 706 | return |
| 707 | } |
| 708 | base := shutdownCtx |
| 709 | if !seedEnabled || base == nil { |
| 710 | base = context.Background() |
| 711 | } |
| 712 | ctx, cancel := context.WithTimeout(base, 5*time.Second) |
| 713 | defer cancel() |
| 714 | if err := claims.close(ctx); err != nil { |
| 715 | log.Printf("agent: regional claims shutdown: %v", err) |
| 716 | } |
| 717 | }() |
| 718 | var stopHousekeeper func() |
| 719 | if !claims.Writes.Enabled { |
| 720 | // Certificate issuance must not wait for instance recovery. |
| 721 | stopHousekeeper = startHousekeeper() |
| 722 | } |
| 723 | dutiesDone := make(chan struct{}) |
| 724 | runDuties := func() { |
| 725 | defer close(dutiesDone) |
| 726 | select { |
| 727 | case <-started: |
| 728 | case <-startupCtx.Done(): |
| 729 | return |
| 730 | } |
| 731 | if startupCtx.Err() != nil { |
| 732 | return |
| 733 | } |
| 734 | |
| 735 | // Fleet box binding publish, startup half (customer-fleets plan, Task 6): |
| 736 | // a one-shot pass over every route already resident in this box's local |
| 737 | // registry, covering an agent restart (activator.Service's in-memory |
| 738 | // entries are gone, but an already-running child isn't) and routes that |
| 739 | // never go through activator.wake at all. Non-authoritative |
| 740 | // (publishResidentFleetBindings calls WriteFleetBinding with |
| 741 | // takeover=false) — see fleetbind.go's doc comments for why a bare |
| 742 | // "this box has this route resident" pass must never steal a binding a |
| 743 | // wake-success write (lease-backed, authoritative) already holds for a |
| 744 | // DIFFERENT box. |
| 745 | if fleetStore != nil { |
| 746 | if svc == nil { |
| 747 | log.Printf("agent: CARLOS_FLEET=%s set but the hibernation activator is not armed — wake-time fleet bindings will never publish", identity.Fleet) |
| 748 | } |
| 749 | go publishResidentFleetBindings(fleetStore, identity.Fleet, identity.Label, reg) |
| 750 | } |
| 751 | |
| 752 | // Scheduled work (schedule spec §6.2/§6.3), built ONCE here and shared by |
| 753 | // all three legs below: the seed and the fast leg in fastConvergeLoop, the |
| 754 | // hourly leg in adoptLoop, and the runner started here. One converger is a |
| 755 | // correctness requirement, not tidiness — its etag memory is what makes a |
| 756 | // declaration change cheap and a run signal exactly-once, and a second |
| 757 | // instance elsewhere would replay or swallow one. The three legs run in |
| 758 | // two different goroutines and do overlap (an hourly pass mid-flight when |
| 759 | // a console bump fires the fast one); Converger.Pass serializes itself on |
| 760 | // its own mutex, so no call site here needs one. |
| 761 | // |
| 762 | // The two halves arm on different conditions — armSchedule |
| 763 | // (schedule_wiring.go) is that switch, and carries the why. In short: |
| 764 | // declared schedules live in the bucket, so the converger needs the |
| 765 | // deploy store; the RUNNER needs only the registry, because once-timers |
| 766 | // are registered by instances on their control socket (mounted above, |
| 767 | // Timers: reg, whenever the activator is armed) straight into this box's |
| 768 | // rows, and a box that mounts the verb must also fire what it accepts. |
| 769 | // |
| 770 | // Guarded assignment, not a direct pass: logSink is a *applog.Sink, and |
| 771 | // a nil one handed to an interface parameter is NOT a nil interface — |
| 772 | // the same trap the activator's Logs assignment above avoids. |
| 773 | var sink activator.EventSink |
| 774 | if logSink != nil { |
| 775 | sink = logSink |
| 776 | } |
| 777 | schedStore, schedErr := openDeployStore(agentCtx) |
| 778 | sd := armSchedule(reg, schedStore, schedErr, svc, sink, dd) |
| 779 | // Move holds (moves spec §9.4): one gate for the runner, the converger |
| 780 | // and a cut's drain, and the converger's freeze. Nothing with the |
| 781 | // Phase 1 switch off. |
| 782 | armScheduleHolds(sd, armedHoldView()) |
| 783 | if err := configureScheduleClaims(claims, sd); err != nil { |
| 784 | log.Fatalf("agent: host claims: %v", err) |
| 785 | } |
| 786 | if sd != nil { |
| 787 | sd.runner.Start(context.Background()) // runs until the process ends, like every loop below |
| 788 | } |
| 789 | serialClaims := make(chan struct{}, 1) |
| 790 | regionalReport := func(err error) { log.Printf("agent: regional claims: %v", err) } |
| 791 | if hostwrite.Enabled(claimEntryEnv(claimCtx)) && moves.Enabled(os.Getenv) { |
| 792 | regionalReport = movesRefusalLog.report |
| 793 | } |
| 794 | if err := claims.startRegional(agentCtx, serialClaims, regionalReport); err != nil { |
| 795 | log.Fatalf("agent: regional claims: %v", err) |
| 796 | } |
| 797 | |
| 798 | // The relocation pass (stage B Task 10): it says itself what it started. |
| 799 | startRelocationPass(agentCtx, claimCtx, claims, reg, svc) |
| 800 | // The change index (S3 change-index design §4): armed before the |
| 801 | // loops that read through it, with its backstop running from here. |
| 802 | // Unset CARLOS_CHANGE_INDEX, or a fleet box: nothing. |
| 803 | if v := armChangeIndex(agentCtx, os.Getenv, openDeployStore); v != nil { |
| 804 | go recordViewBackstopLoop(agentCtx, v) |
| 805 | } |
| 806 | go adoptLoop(reg, svc, sd, claims.Writes) |
| 807 | go fastConvergeLoop(reg, svc, sd, claims.Writes, serialClaims, bs, claims.Providers.Provision) |
| 808 | if logSink != nil { |
| 809 | go logshipLoop(logSink, status.Label(readMachineID())) |
| 810 | } |
| 811 | if errSink != nil { |
| 812 | go errshipLoop(errSink, status.Label(readMachineID())) |
| 813 | } |
| 814 | // The build-state report (settings spec rev 11 §4.4) names the rows |
| 815 | // keepsBuildState selects, with host claims as the build-state store |
| 816 | // sees them (cfg.HostWrites.Enabled, set by configureWakeClaims). |
| 817 | buildStateClaims := cfg != nil && cfg.HostWrites.Enabled |
| 818 | collectBuildStates := func() (map[string]status.BuildStatePins, error) { return buildStatePins(reg, buildStateClaims) } |
| 819 | go statusLoop(reg, svc, collectBuildStates) |
| 820 | // The memory early-warning (memory-pressure spec §2.4): per box, gated |
| 821 | // by CARLOS_HOST_METRICS — one log line and nothing else where unset. |
| 822 | go memwatchLoop() |
| 823 | go domainsLoop(reg, peekAcmeCache(args), claims.Writes.Enabled) |
| 824 | go instancesLoop(reg, claims.Writes, bs, claims.Providers.Provision) |
| 825 | // The deployment-global election, armed BEFORE the loops it gates so a |
| 826 | // single-box deployment is holding well before their first tick (a |
| 827 | // minute away) — see housekeeper.go. Not a `go` of its own: it publishes |
| 828 | // the process's housekeeper and spawns its loop itself. Its stand-down |
| 829 | // runs in the post-drain block at the end of this function. |
| 830 | if claims.Writes.Enabled { |
| 831 | stopHousekeeper = startHousekeeper() |
| 832 | } |
| 833 | go retentionLoop(reg, collectBuildStates) |
| 834 | go registrySnapshotLoop(reg) |
| 835 | go edgesFileLoop(reg) |
| 836 | // Region HA enrollment evidence (spec 1 §5.1): once, at start, and only |
| 837 | // with CARLOS_REGION_HA on. See boxrecord.go. |
| 838 | go boxRecordLoop() |
| 839 | // The mesh key's own loop (multi-region backhaul spec, Part 3 Task 2) — |
| 840 | // deliberately NOT a duty riding retentionLoop's tick: that loop returns |
| 841 | // early, never arming at all, on a box missing CARLOS_RELEASE_PUBKEY, |
| 842 | // and such a box still WINS the housekeeper election. A mint duty living |
| 843 | // only inside retentionLoop would then silently never run fleet-wide. |
| 844 | // See mintMeshKeyLoop's own comment (housekeeper.go). |
| 845 | go mintMeshKeyLoop() |
| 846 | // The DNS steering converger (multi-region ingress spec, Part 4) — its |
| 847 | // own loop for mintMeshKeyLoop's exact reason, and dormant on every box |
| 848 | // today: CARLOS_STEERING_ZONE_ID unset means one log line at arm and |
| 849 | // nothing else, ever. See steering_loop.go. |
| 850 | go steeringLoop() |
| 851 | // The sweep reads the registry only through this closure — the stamped |
| 852 | // bespoke units to sample beside the carlos- ones, re-read each tick. |
| 853 | go metricsLoop(cgm.Root(), func() []string { return stampedUnitNames(reg) }, alerts) |
| 854 | go configSyncLoop(reg, svc, dd, claims.Writes) |
| 855 | go configGenLoop(reg, svc, dd, claims.Writes) |
| 856 | go socketAdvisoryLoop(reg, svc) |
| 857 | |
| 858 | } |
| 859 | if claims.Writes.Enabled { |
| 860 | runDuties() |
| 861 | } else { |
| 862 | go runDuties() |
| 863 | } |
| 864 | |
| 865 | // Seed handlers borrow this agent's resources and finish before they close. |
| 866 | var seedCtx context.Context |
| 867 | var seedCancel context.CancelFunc |
| 868 | var seedServer *filesSeedServer |
| 869 | if seedEnabled { |
| 870 | seedCtx, seedCancel = signal.NotifyContext(claimCtx, syscall.SIGTERM, os.Interrupt) |
| 871 | defer seedCancel() |
| 872 | var seedFinish func() |
| 873 | seedCtx, seedFinish = activator.WithCleanupBudget(seedCtx, 30*time.Second) |
| 874 | defer seedFinish() |
| 875 | stopSeed := context.AfterFunc(startupCtx, seedCancel) |
| 876 | defer stopSeed() |
| 877 | seedServer, err = startAgentFilesSeed(seedCtx, svc.Startup, svc) |
| 878 | if err != nil { |
| 879 | log.Printf("Files seed failed: %s.", err) |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | // The same serving loop as `carlos edge`, with the activator injected. |
| 884 | // Caching applies here too: cached hits still avoid the activator, which |
| 885 | // is the entire point of checking the cache before the hibernation branch. |
| 886 | // The ready hook fires synchronously before listening, so |
| 887 | // tunnelLoop is handed off to its own goroutine rather than called here — |
| 888 | // calling it directly would block the front door forever. |
| 889 | edgeOpts := append([]edge.RunOption{ |
| 890 | // Certificate issuance is a deployment-global duty, so it runs where |
| 891 | // the housekeeper runs (housekeeper.go). The gate goes DOWN into the |
| 892 | // edge as a func — this package imports internal/edge and never the |
| 893 | // reverse — and it is the GATE, not the bare flag: an ACME order runs |
| 894 | // for seconds to minutes, and the context ends the instant this box |
| 895 | // stops being the housekeeper. Armed above, so this closure is live |
| 896 | // and correct by the time the edge reads it. |
| 897 | edge.WithIssuerGate(housekeeperGate), |
| 898 | // Whether this box contends in that election AT ALL — the cert |
| 899 | // backfill's eligibility (#318). Mode is decided once from the |
| 900 | // environment, so the closure is stable; see housekeeperContends |
| 901 | // for the fleet-box exclusion it exists to state. |
| 902 | edge.WithIssuerContender(housekeeperContends), |
| 903 | // This box's own label, for the home map's own-box exclusion |
| 904 | // (internal/edge/homemap.go). metricsBoxLabel is the SAME call |
| 905 | // edgesFileLoop publishes under, deliberately: the map excludes our |
| 906 | // hosts by matching the box field of the file we ourselves wrote, so |
| 907 | // the two derivations must be one derivation. Without it the edge |
| 908 | // would offer nothing (fail-closed), never dial itself. |
| 909 | edge.WithBoxLabel(metricsBoxLabel(os.Getenv, readMachineID())), |
| 910 | // The change index (#336): the cert mirror and the home map read |
| 911 | // through the process's one watch. A loader, not the watch: on a box |
| 912 | // without claim writes the index is armed on runDuties' goroutine, |
| 913 | // which may not have run yet. Disarmed, it returns nil for good and |
| 914 | // both keep today's full reads. |
| 915 | edge.WithChangeWatch(armedChangeWatch.Load), |
| 916 | }, |
| 917 | // The edge-local door (#175 P7), only when CARLOS_EDGE_DOOR names a |
| 918 | // socket. It presents this box's public address, the one the edges |
| 919 | // file publishes; with none, the edge keeps the door shut and says so. |
| 920 | edgeDoorOptions(os.Getenv, readPublicIP)...) |
| 921 | // A tunnel client takes the hub's GeoIP answer and mirrors nothing (geoip |
| 922 | // spec §5.1); the same rule tunnelLoop arms on. |
| 923 | if tunnelEnvComplete(os.Getenv) { |
| 924 | edgeOpts = append(edgeOpts, edge.WithGeoTunnelClient()) |
| 925 | } |
| 926 | edge.RunWithActivatorReady(args, act, cache, func(e *edge.Edge) { |
| 927 | go tunnelLoop(e) |
| 928 | // The edge owns the health probers, so its state table is only |
| 929 | // reachable once it exists — statusLoop (already running above) reads |
| 930 | // it through this published source on its next tick. Method value, so |
| 931 | // the loop always sees the CURRENT probe set: Reload replaces probers |
| 932 | // behind this handle without the heartbeat needing to be re-wired. |
| 933 | setUpstreamSource(e.UpstreamStates) |
| 934 | // The GeoIP block for the heartbeat (geoip spec §5.5), the same way. |
| 935 | geoIPSource.Store(e.GeoIPState) |
| 936 | geoIPAlertSource.Store(e.GeoIPAlertCode) |
| 937 | // The layout pass reloads routing through this as soon as a move |
| 938 | // commits, instead of waiting for the 2s registry poll. |
| 939 | reload := func() { _ = e.Reload() } |
| 940 | layoutReload.Store(&reload) |
| 941 | m := edge.NewUsageMeter() |
| 942 | e.EnableUsage(m) |
| 943 | // Same arming point as the meter: the edge observes an upstream 5xx |
| 944 | // on the serving path and files it here (spec 2026-09-07 §3.2), and |
| 945 | // errshipLoop above ships what this fills. Guarded, because a typed |
| 946 | // nil inside the interface would NOT be nil to the edge. |
| 947 | if errSink != nil { |
| 948 | e.EnableErrors(errSink) |
| 949 | } |
| 950 | // The analytics meter (spec 2026-09-07 §3.3): the edge counts |
| 951 | // pageviews per (account, app) while telemetryPass (fullPass and |
| 952 | // instancesLoop) keeps its switch map current, and usageLoop |
| 953 | // drains it into day objects on the same 5-minute tick as usage. |
| 954 | am := analytics.NewMeter(dataDir(), status.Label(readMachineID()), time.Now) |
| 955 | e.EnableAnalytics(am) |
| 956 | setAnalyticsSink(e.SetAnalyticsOn) |
| 957 | go usageLoop(m.Drain, am) |
| 958 | }, edgeOpts...) |
| 959 | |
| 960 | var seedErr error |
| 961 | shutdownCtx, shutdownCancel, seedErr = joinAgentShutdown(seedCtx, seedCancel, seedServer, cancelStartup, startupDone, dutiesDone) |
| 962 | if seedErr != nil { |
| 963 | safeClaimClose = false |
| 964 | log.Printf("Files seed failed: %s.", seedErr) |
| 965 | os.Exit(1) |
| 966 | } |
| 967 | |
| 968 | // The edge has drained; take the children down too (they're this process's |
| 969 | // children — orphaning them would strand their leases until TTL expiry). |
| 970 | if svc != nil { |
| 971 | ctx := shutdownCtx |
| 972 | svc.Shutdown(ctx) |
| 973 | // Then the scheduler, under the same budget (spec §8.6, Task 11): an |
| 974 | // attempt whose swap Shutdown just cancelled finishes its done(), its |
| 975 | // outcome and the end of its intent on disk, and so does a wake event |
| 976 | // under way, before this process exits. A planned restart mid-swap so |
| 977 | // leaves nothing for the next agent to reconcile. |
| 978 | if m := armedRollout(); m != nil { |
| 979 | m.Close(ctx) |
| 980 | } |
| 981 | // The agent socket's own defer (above, at its construction) stops it |
| 982 | // once this function returns — after Shutdown, per the LIFO ordering |
| 983 | // explained there. That gap is real, not zero-width: svc.Shutdown |
| 984 | // returning here and agentSockCancel actually firing are two |
| 985 | // different points in the stack unwind, so the socket is technically |
| 986 | // still accepting connections for the width of that unwind. It is |
| 987 | // harmless: Shutdown has already stopped every entry and set |
| 988 | // shuttingDown, so a call landing in the gap finds no live entry — |
| 989 | // WorkerToken returns "" (authorize's unmatchable() then refuses it), |
| 990 | // and NoteWorkerContact/NoteWorkerReady/SleepWorker are all no-ops on |
| 991 | // a nil entry lookup. Nothing in the gap can reactivate a stopped |
| 992 | // worker or authorize against a dead token. |
| 993 | } |
| 994 | // Same principle one level up: this process may hold the DEPLOYMENT's |
| 995 | // housekeeper lease, and a restart that walked away from it would leave |
| 996 | // control/housekeeper-lease/lock.json standing for up to its 15-minute |
| 997 | // TTL — no pruner, and from Part 2 no cert issuer, across an ordinary |
| 998 | // roll of a single-box deployment. Stopping cancels every duty in flight |
| 999 | // and releases the lock, so the next box (or this one, a second later) |
| 1000 | // takes over immediately. |
| 1001 | if stopHousekeeper != nil { |
| 1002 | stopHousekeeper() |
| 1003 | } |
| 1004 | } |
| 1005 | |
| 1006 | // statusSocketProbe is the kernel-table probe the status loop hands |
| 1007 | // wrapListeningProbe: listeningProbe itself with generations off (or no |
| 1008 | // activator), and pointed at each managed route's current generation with |
| 1009 | // them on (currentSocketProbe). |
| 1010 | func statusSocketProbe(routes []registry.Route, svc *activator.Service) func(string) *bool { |
| 1011 | if svc == nil || !svc.GenerationsOn() { |
| 1012 | return listeningProbe |
| 1013 | } |
| 1014 | return currentSocketProbe(routes, svc.AppSocket, listeningProbe) |
| 1015 | } |
| 1016 | |
| 1017 | func configureWakeClaims(w *ClaimWiring, cfg *activator.Config) error { |
| 1018 | if w == nil { |
| 1019 | return hostwrite.ErrUnavailable |
| 1020 | } |
| 1021 | if !w.Writes.Enabled { |
| 1022 | return nil |
| 1023 | } |
| 1024 | if cfg == nil || w.Providers.Wake == nil || w.Writes.Selector == nil { |
| 1025 | return hostwrite.ErrUnavailable |
| 1026 | } |
| 1027 | cfg.HostWrites, cfg.Claims = w.Writes, w.Providers.Wake |
| 1028 | return nil |
| 1029 | } |
| 1030 | |
| 1031 | func configureScheduleClaims(w *ClaimWiring, sd *scheduleDeps) error { |
| 1032 | if w == nil { |
| 1033 | return hostwrite.ErrUnavailable |
| 1034 | } |
| 1035 | if !w.Writes.Enabled { |
| 1036 | return nil |
| 1037 | } |
| 1038 | if sd == nil || sd.runner == nil || sd.conv == nil || w.Writes.Selector == nil { |
| 1039 | return hostwrite.ErrUnavailable |
| 1040 | } |
| 1041 | sd.runner.HostWrites, sd.conv.HostWrites = w.Writes, w.Writes |
| 1042 | sd.runner.StartDelivery = w.Providers.Delivery |
| 1043 | return nil |
| 1044 | } |
| 1045 | |
| 1046 | // --- Task 8: the agent socket's retry/backoff, and the worker boot pass ---- |
| 1047 | |
| 1048 | // workersToStart is the boot pass's selection: a sidecar that does not |
| 1049 | // hibernate must be running whenever the agent is. A hibernating sidecar is |
| 1050 | // deliberately NOT started here — it waits for its due time, an operator |
| 1051 | // command, or (later) a delivery signal, which is the whole point of the |
| 1052 | // class. Status is checked the same way the due-time sweep checks it |
| 1053 | // (activator.go's sweepOnce): only "ready" (or the legacy "" some pre-status |
| 1054 | // rows carry) is eligible — a row still provisioning, or draining, must not |
| 1055 | // be spawned. |
| 1056 | func workersToStart(routes []registry.Route) []registry.Route { |
| 1057 | var out []registry.Route |
| 1058 | for _, rt := range routes { |
| 1059 | if rt.IsSidecar() && !rt.Hibernate && (rt.Status == "" || rt.Status == "ready") { |
| 1060 | out = append(out, rt) |
| 1061 | } |
| 1062 | } |
| 1063 | return out |
| 1064 | } |
| 1065 | |
| 1066 | // workerBootConcurrency bounds how many sidecars the boot pass wakes at |
| 1067 | // once. WakeWorker blocks for the settle window (activator's sidecarSettle, |
| 1068 | // 10s in production) whenever a child never calls /v1/ready — a slow or |
| 1069 | // silent sidecar, or simply one that only survives the window rather than |
| 1070 | // reporting early — so an unbounded fan-out could still spawn every child on |
| 1071 | // the box in the same instant (a thundering herd against S3 leases and |
| 1072 | // restore reads), while a serial pass would cost up to 10 seconds PER |
| 1073 | // sidecar of agent startup. This bound trades both off: bounded fan-out, not |
| 1074 | // one-at-a-time. |
| 1075 | const workerBootConcurrency = 8 |
| 1076 | |
| 1077 | // runWorkerBootPass wakes every route workersToStart selects out of routes, |
| 1078 | // concurrently (bounded by workerBootConcurrency) rather than one at a time |
| 1079 | // — see workerBootConcurrency's doc comment for why serial is too slow and |
| 1080 | // unbounded is too bursty. wake is svc.WakeWorker in production; a seam here |
| 1081 | // so the concurrency shape is testable without a real activator.Service. A |
| 1082 | // wake that errors is logged and does not stop the rest — one sidecar that |
| 1083 | // can't come up must not strand every other one on the box at boot. |
| 1084 | func runWorkerBootPass(routes []registry.Route, wake func(registry.Route) error) { |
| 1085 | toStart := workersToStart(routes) |
| 1086 | if len(toStart) == 0 { |
| 1087 | return |
| 1088 | } |
| 1089 | sem := make(chan struct{}, workerBootConcurrency) |
| 1090 | var wg sync.WaitGroup |
| 1091 | for _, rt := range toStart { |
| 1092 | rt := rt |
| 1093 | wg.Add(1) |
| 1094 | sem <- struct{}{} |
| 1095 | go func() { |
| 1096 | defer wg.Done() |
| 1097 | defer func() { <-sem }() |
| 1098 | if err := wake(rt); err != nil { |
| 1099 | log.Printf("agent: worker boot pass: wake %s: %v", rt.Host, err) |
| 1100 | } |
| 1101 | }() |
| 1102 | } |
| 1103 | wg.Wait() |
| 1104 | } |
| 1105 | |
| 1106 | // agentSocketRetryBase/Cap bound the agent socket's rebind backoff — vars, |
| 1107 | // not consts, so a test can shrink them (sidecarSettle's precedent in |
| 1108 | // internal/activator). |
| 1109 | var ( |
| 1110 | agentSocketRetryBase = time.Second |
| 1111 | agentSocketRetryCap = 30 * time.Second |
| 1112 | ) |
| 1113 | |
| 1114 | // serveAgentSocket runs serve (agentsock.Server.Serve in production) in a |
| 1115 | // retry loop with doubling backoff, forever, until ctx is cancelled. serve |
| 1116 | // is a seam so the retry/backoff shape is testable without a real unix |
| 1117 | // socket. |
| 1118 | // |
| 1119 | // A failed bind must NEVER be fatal to the agent: agentsock.Serve |
| 1120 | // deliberately REFUSES a socket another listener already owns (rather than |
| 1121 | // unlinking it) to prevent two agents both believing they own the box's |
| 1122 | // workers, and on an agent restart the old process's listener can take a |
| 1123 | // few seconds to release — that refusal is expected and transient. An agent |
| 1124 | // that exited because a socket was briefly busy would take down every route |
| 1125 | // on the box, so this loop only ever logs and retries; it has no path that |
| 1126 | // calls log.Fatal or returns an error to a caller who might. |
| 1127 | // |
| 1128 | // A clean return from serve (nil, once ctx is cancelled — agentsock.Serve's |
| 1129 | // own contract) ends the loop immediately with no retry: that is a normal |
| 1130 | // shutdown, not a failure. |
| 1131 | func serveAgentSocket(ctx context.Context, path string, serve func(context.Context, string) error) { |
| 1132 | backoff := agentSocketRetryBase |
| 1133 | for { |
| 1134 | err := serve(ctx, path) |
| 1135 | if err == nil || ctx.Err() != nil { |
| 1136 | return |
| 1137 | } |
| 1138 | log.Printf("agent: agent socket %s: %v — retrying in %s", path, err, backoff) |
| 1139 | timer := time.NewTimer(backoff) |
| 1140 | select { |
| 1141 | case <-timer.C: |
| 1142 | case <-ctx.Done(): |
| 1143 | timer.Stop() |
| 1144 | return |
| 1145 | } |
| 1146 | backoff *= 2 |
| 1147 | if backoff > agentSocketRetryCap { |
| 1148 | backoff = agentSocketRetryCap |
| 1149 | } |
| 1150 | } |
| 1151 | } |
| 1152 | |
| 1153 | // nestedGate is what an isolated wake requires of the box (design §8.5, |
| 1154 | // §8.6): this process able to make the openat2 call instancefs is built on |
| 1155 | // (self; the edge unit's own hardening is what can take it away, #126), |
| 1156 | // the three nosymfollow mounts a nested instance's tree depends on, and a |
| 1157 | // control service that reports ready. The answer is recomputed on |
| 1158 | // every wake (the mounts are cheap to read; /ready is one short request), |
| 1159 | // so a box that becomes ready needs no agent restart and a box that stops |
| 1160 | // being ready refuses the very next wake rather than the stale verdict a |
| 1161 | // once-at-startup check would give. |
| 1162 | func nestedGate(self func() error, mounts func(string) (bool, error), ready func(context.Context) error) func() error { |
| 1163 | return func() error { |
| 1164 | if err := self(); err != nil { |
| 1165 | return err |
| 1166 | } |
| 1167 | for _, m := range []string{"/data/instances", "/run/carlos/instances", "/run/carlos/logs"} { |
| 1168 | ok, err := mounts(m) |
| 1169 | if err != nil { |
| 1170 | return fmt.Errorf("%s: %v", m, err) |
| 1171 | } |
| 1172 | if !ok { |
| 1173 | return fmt.Errorf("%s is not a nosymfollow mount", m) |
| 1174 | } |
| 1175 | } |
| 1176 | ctx, cancel := context.WithTimeout(context.Background(), instancectl.DeadlineReady) |
| 1177 | defer cancel() |
| 1178 | return ready(ctx) |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | // peekRegistryPath pre-parses the agent args for -registry (default |
| 1183 | // <dataDir>/registry.db) without disturbing them — edge.RunWithActivator does |
| 1184 | // the authoritative parse. The set mirrors edge.Run's flags so a full command |
| 1185 | // line parses cleanly; an unknown flag just ends the peek early (the edge's |
| 1186 | // own ExitOnError parse then reports it properly). |
| 1187 | func peekRegistryPath(args []string, dd string) string { |
| 1188 | fs := flag.NewFlagSet("agent", flag.ContinueOnError) |
| 1189 | fs.SetOutput(io.Discard) |
| 1190 | regPath := fs.String("registry", filepath.Join(dd, "registry.db"), "") |
| 1191 | fs.String("acme-cache", "", "") |
| 1192 | fs.String("https", "", "") |
| 1193 | fs.String("http", "", "") |
| 1194 | fs.String("fallback", "", "") |
| 1195 | fs.Bool("dev", false, "") |
| 1196 | fs.Usage = cliflag.Usage(fs, "") |
| 1197 | _ = fs.Parse(args) |
| 1198 | return *regPath |
| 1199 | } |
| 1200 | |
| 1201 | // materializeHook builds the ExecBackend's cold-wake staging hook, or nil |
| 1202 | // when this box has no release store/pubkey — nil keeps today's behavior |
| 1203 | // ("no live binary for <host>" wake failure), the safe degrade for a box |
| 1204 | // whose routes are versioned by hand or not at all. |
| 1205 | // |
| 1206 | // The hook is the adopt pass scoped to one host (release.MaterializeHost), |
| 1207 | // so a cold wake with nothing staged pays one slow request instead of 503ing |
| 1208 | // until the hourly pass comes round (issue #77). Everything the pass refuses |
| 1209 | // it still refuses here: a bake-held head is NOT materialized, and the wake |
| 1210 | // fails with that reason rather than the head being installed early. |
| 1211 | // |
| 1212 | // Store and key are resolved ONCE, at agent start, exactly like adoptLoop's: |
| 1213 | // a wake is on a request's critical path, and re-reading env plus re-opening |
| 1214 | // an S3 client per cold wake would put that work there for nothing. |
| 1215 | // |
| 1216 | // OnStageBlocked is wired so a cold wake on a hardened box (this sandboxed |
| 1217 | // process cannot write /opt/carlos) relays the stage it cannot perform to |
| 1218 | // the root pass (carlos-stage.path) instead of just failing the wake. NO |
| 1219 | // RestartHibernating here, deliberately: the host is cold, restart stays |
| 1220 | // nil — the activator exec'ing the freshly-pointed binary IS the restart. |
| 1221 | // The requester is built once with the hook, so its once-a-minute budget is |
| 1222 | // the hook's own, same shape as the loops'. |
| 1223 | func materializeHook(reg *registry.Registry, dd string) func(context.Context, string) error { |
| 1224 | store, err := openDeployStore(context.Background()) |
| 1225 | if err != nil { |
| 1226 | log.Printf("agent: cold-wake materialize disabled: %v", err) |
| 1227 | return nil |
| 1228 | } |
| 1229 | pub, err := release.ParsePublicKey(os.Getenv("CARLOS_RELEASE_PUBKEY"), "CARLOS_RELEASE_PUBKEY") |
| 1230 | if err != nil { |
| 1231 | log.Printf("agent: cold-wake materialize disabled: %v", err) |
| 1232 | return nil |
| 1233 | } |
| 1234 | optDir := release.OptDir() |
| 1235 | stageReq := stageRequester(dd) |
| 1236 | return func(ctx context.Context, host string) error { |
| 1237 | // The process's one proof gate (proofask.go), so a cold wake cannot |
| 1238 | // adopt an unproven build of a listed app. |
| 1239 | return release.MaterializeHost(ctx, reg, store, pub, optDir, dd, host, |
| 1240 | release.OnStageBlocked(stageReq), release.WithProofGate(agentProofGate()), |
| 1241 | release.WithDigestRatchet(pointerRatchet(false, optDir, dd, os.Getenv)), |
| 1242 | updatesPassOption(dd, false)) |
| 1243 | } |
| 1244 | } |
| 1245 | |
| 1246 | // adoptLoop runs a release-adoption pass roughly hourly (plus jitter, so a |
| 1247 | // fleet never hits the store in the same second), forever. Missing deployment |
| 1248 | // config disables the loop with one log line — an agent with no release store |
| 1249 | // is a valid box (routes managed by hand, adopted by the timer unit, or not |
| 1250 | // versioned at all). |
| 1251 | // The activator is taken so the pass can cycle exec-backed instances the way |
| 1252 | // only this process can (see makeRestartRoute); nil is fine — a box with no |
| 1253 | // hibernation armed simply has no children to bounce. |
| 1254 | // |
| 1255 | // sd carries the scheduled-work converger; this loop is its HOURLY floor (the |
| 1256 | // fast one rides the serial in fastConvergeLoop). sd.converge is a no-op when |
| 1257 | // the box has no deploy store — the agent logged that once at arming. |
| 1258 | func adoptLoop(reg *registry.Registry, svc *activator.Service, sd *scheduleDeps, writes hostwrite.Config) { |
| 1259 | ctx := context.Background() |
| 1260 | store, err := openDeployStore(ctx) |
| 1261 | if err != nil { |
| 1262 | log.Printf("agent: adopt loop disabled: %v", err) |
| 1263 | return |
| 1264 | } |
| 1265 | pub, err := release.ParsePublicKey(os.Getenv("CARLOS_RELEASE_PUBKEY"), "CARLOS_RELEASE_PUBKEY") |
| 1266 | if err != nil { |
| 1267 | log.Printf("agent: adopt loop disabled: %v", err) |
| 1268 | return |
| 1269 | } |
| 1270 | optDir := release.OptDir() |
| 1271 | restart := makeRestartRoute(reg, systemctlTryRestart, agentCheckedUnitRestart(systemdUnits{}, optDir, os.Getenv, nil), adoptBounceFn(svc)) |
| 1272 | log.Printf("agent: adopt loop armed (hourly + jitter)") |
| 1273 | // First pass shortly after boot (Eleven's OnBootSec=5min, compressed — |
| 1274 | // the edge is already up by the time this fires), then hourly + jitter. |
| 1275 | delay := time.Minute |
| 1276 | for { |
| 1277 | time.Sleep(delay) |
| 1278 | delay = time.Hour + rand.N(10*time.Minute) |
| 1279 | // DO NOT add release.NagExpiredCanaries(...) to this call. This loop |
| 1280 | // runs hourly, but it is NOT "the hourly root adopt pass" the option's |
| 1281 | // own doc comment names — that's RunAdopt (carlos-adopt.service) |
| 1282 | // below, root-privileged and the one place config sync also runs. Two |
| 1283 | // hourly callers both nagging would double the log line every hour; |
| 1284 | // worse, if this loop's cadence ever changes, a nag wired here changes |
| 1285 | // frequency with it silently. One caller, one nag. |
| 1286 | // |
| 1287 | // DO NOT add release.WithCapabilities(...) to this call either, and |
| 1288 | // for a harder reason than the nag's: that option WRITES |
| 1289 | // /etc/systemd/system, and this loop runs inside carlos-edge.service, |
| 1290 | // which is ProtectSystem=strict with ReadWritePaths=/data /run/carlos. |
| 1291 | // It could not converge a drop-in if it tried; all it would produce is |
| 1292 | // a permission-denied line every hour. The capability converge belongs |
| 1293 | // to RunAdopt (carlos-adopt.service, root) alone — see below. |
| 1294 | // |
| 1295 | // OnStageBlocked, by contrast, is exactly right here: it fires only |
| 1296 | // when this sandboxed process cannot write the version store, and all |
| 1297 | // it does is ask the root pass to do the part this one cannot. |
| 1298 | // |
| 1299 | // RestartHibernating, by contrast again, belongs HERE and not in |
| 1300 | // RunAdopt: this process owns the exec children, so it is the only one |
| 1301 | // whose restart callback can reach a warm hibernator at all. Without |
| 1302 | // it the pass would split on the Hibernate FLAG and hand every exec |
| 1303 | // instance the pure repoint — the flag says "may sleep", and a busy |
| 1304 | // instance would go on serving the binary its symlink no longer names |
| 1305 | // until it happened to go idle. Safe because makeRestartRoute cannot |
| 1306 | // start a process on either of its paths; see release.RestartHibernating. |
| 1307 | // |
| 1308 | // WithStartUnit rides wherever the named-unit restart arm can fire: |
| 1309 | // a failed named-unit restart that this pass rolled back must not |
| 1310 | // leave the service down (release.WithStartUnit). It asks for no |
| 1311 | // authority the restart above did not already need — both are one |
| 1312 | // systemctl verb on the same unit — and it cannot wake anything: it |
| 1313 | // fires only after a try-restart RETURNED AN ERROR, which a stopped |
| 1314 | // unit's exit-0 no-op never does. |
| 1315 | // |
| 1316 | // DO NOT add the platform self-update here either: the system leg |
| 1317 | // lives in the fast loop and re-observes hourly on its own, so an |
| 1318 | // hourly call here would only double the pointer read. |
| 1319 | // |
| 1320 | // agentUnitsOption hands named units to root when this process is |
| 1321 | // not root (issue #315), which leaves WithStartUnit and |
| 1322 | // WithRollbackRestart to an agent run as root. |
| 1323 | opts := append([]release.PassOption{ |
| 1324 | release.BakeDir(dataDir()), release.OnStageBlocked(stageRequester(dataDir())), |
| 1325 | release.RestartHibernating(restart), |
| 1326 | release.WithStartUnit(makeStartRoute(reg, systemctlStartUnit)), |
| 1327 | release.WithRollbackRestart(agentRollbackUnitRestart(reg, systemdUnits{}, optDir, os.Getenv)), |
| 1328 | release.WithProofGate(agentProofGate()), release.WithHostWrites(writes), |
| 1329 | release.WithDigestRatchet(pointerRatchet(false, optDir, dataDir(), os.Getenv)), |
| 1330 | updatesPassOption(dataDir(), false)}, agentUnitsOption()...) |
| 1331 | if err := release.Pass(ctx, reg, store, pub, optDir, restart, opts...); err != nil { |
| 1332 | log.Printf("agent: adopt pass: %v", err) |
| 1333 | } |
| 1334 | // The schedule converge's hourly floor (schedule spec §6.2), AFTER the |
| 1335 | // release pass: an app whose schedules.json arrived with a new version |
| 1336 | // gets its rows against the version that is actually staged. Die-free |
| 1337 | // per app inside Pass — the joined error is one log line, never fatal. |
| 1338 | if err := sd.converge(ctx); err != nil { |
| 1339 | log.Printf("agent: schedule converge (hourly): %v", err) |
| 1340 | } |
| 1341 | } |
| 1342 | } |
| 1343 | |
| 1344 | // configSyncLoop delivers app config to exec-backed routes: hourly (plus |
| 1345 | // a first pass ~1min after boot, like adoptLoop), materialize every |
| 1346 | // exec-syncable route's <dataDir>/<host>.config.env from the bucket and |
| 1347 | // bounce running instances whose config changed. Runs IN-PROCESS on |
| 1348 | // purpose — /data is inside this unit's sandbox and the exec children |
| 1349 | // are this process's own, so it can both write the files and restart |
| 1350 | // the readers; the root adopt path keeps the unit-backed half |
| 1351 | // (/etc/carlos/apps needs root). Inert until CARLOS_SECRETS_KEY is set, |
| 1352 | // the same convention as every other optional loop. |
| 1353 | func configSyncLoop(reg *registry.Registry, svc *activator.Service, dd string, writes hostwrite.Config) { |
| 1354 | raw := os.Getenv("CARLOS_SECRETS_KEY") |
| 1355 | if raw == "" { |
| 1356 | log.Printf("agent: config sync loop disabled: CARLOS_SECRETS_KEY not set") |
| 1357 | return |
| 1358 | } |
| 1359 | priv, err := appconfig.ParsePrivateKey(raw, "CARLOS_SECRETS_KEY") |
| 1360 | if err != nil { |
| 1361 | log.Printf("agent: config sync loop disabled: %v", err) |
| 1362 | return |
| 1363 | } |
| 1364 | ctx := context.Background() |
| 1365 | store, err := openDeployStore(ctx) |
| 1366 | if err != nil { |
| 1367 | log.Printf("agent: config sync loop disabled: %v", err) |
| 1368 | return |
| 1369 | } |
| 1370 | log.Printf("agent: config sync loop armed (hourly + jitter)") |
| 1371 | delay := time.Minute |
| 1372 | for { |
| 1373 | time.Sleep(delay) |
| 1374 | delay = time.Hour + rand.N(10*time.Minute) |
| 1375 | if err := syncExecSerialized(ctx, reg, store, priv, dd, svc, writes); err != nil { |
| 1376 | log.Printf("agent: config sync: %v", err) |
| 1377 | } |
| 1378 | } |
| 1379 | } |
| 1380 | |
| 1381 | // configSyncMu serializes the two config loops' SyncExec passes against each |
| 1382 | // other. They can genuinely overlap — hourly pass mid-flight when the 3s |
| 1383 | // loop notices a fresh generation — and syncExecHost writes every route's |
| 1384 | // file through a FIXED temp path (<host>.config.env.tmp, syncexec.go), so |
| 1385 | // two interleaved passes could cross-rename each other's content: the loser |
| 1386 | // lands STALE bytes after the winner's fresh ones, with the generation cache |
| 1387 | // already advanced past the change, leaving nothing to re-fire until the |
| 1388 | // next hourly pass. Serializing in-process closes that outright; the root |
| 1389 | // adopt path is a different process writing a different namespace |
| 1390 | // (/etc/carlos/apps) and is not this mutex's problem. |
| 1391 | var configSyncMu sync.Mutex |
| 1392 | |
| 1393 | // syncExecSerialized is the one way either config loop runs SyncExec — the |
| 1394 | // mutex lives in a shared wrapper, not at the call sites, so a future third |
| 1395 | // caller cannot forget it. |
| 1396 | func syncExecSerialized(ctx context.Context, reg *registry.Registry, store release.Store, priv *ecdh.PrivateKey, dd string, svc *activator.Service, writes ...hostwrite.Config) error { |
| 1397 | return syncExecSerializedFiltered(ctx, nil, reg, store, priv, dd, svc, writes...) |
| 1398 | } |
| 1399 | |
| 1400 | // syncExecSerializedFiltered is syncExecSerialized over the routes keep |
| 1401 | // accepts (nil: every route): the scoped pass's config leg (S3 change-index |
| 1402 | // design §4.5). The filter is applied where SyncExec chooses its routes |
| 1403 | // (appconfig.WithRouteFilter), so a route it leaves out costs no read. |
| 1404 | func syncExecSerializedFiltered(ctx context.Context, keep func(registry.Route) bool, reg *registry.Registry, store release.Store, priv *ecdh.PrivateKey, dd string, svc *activator.Service, writes ...hostwrite.Config) error { |
| 1405 | configSyncMu.Lock() |
| 1406 | defer configSyncMu.Unlock() |
| 1407 | pending := map[string]string{} |
| 1408 | outcomes := map[string]bool{} |
| 1409 | opts := []appconfig.SyncOption{ |
| 1410 | appconfig.SkipFenced(func(host string) bool { return layout.JournalPresent(dd, host) }), |
| 1411 | appconfig.WhenInstanceDirAbsent(func(host, dir string) { pending[host] = dir }), |
| 1412 | appconfig.WithHostOutcome(func(host string, done bool) { outcomes[host] = done }), |
| 1413 | } |
| 1414 | if keep != nil { |
| 1415 | opts = append(opts, appconfig.WithRouteFilter(keep)) |
| 1416 | } |
| 1417 | for _, cfg := range writes { |
| 1418 | opts = append(opts, appconfig.WithHostWrite(cfg)) |
| 1419 | } |
| 1420 | // CARLOS_SWAP=overlap: the scheduler restarts, paced, and acknowledges a |
| 1421 | // revision once it is active; refusing: no restart. configSyncMu above |
| 1422 | // is what makes ConfigDelivered serial per host. |
| 1423 | var err error |
| 1424 | if cb, ok := rolloutConfigCallbacks(ctx, store, configLegs{awake: managedAwake(svc, reg), restart: bounce(svc)}); ok { |
| 1425 | err = appconfig.SyncExecWithCallbacks(ctx, reg, store, priv, dd, os.Getenv("PLATFORM_APPS_DOMAIN"), cb, opts...) |
| 1426 | } else { |
| 1427 | err = appconfig.SyncExec(ctx, reg, store, priv, dd, bounce(svc), os.Getenv("PLATFORM_APPS_DOMAIN"), configAckFunc(ctx, store, "agent"), opts...) |
| 1428 | } |
| 1429 | // A filtered pass names only the routes it kept, so it is never |
| 1430 | // complete: a host it left out keeps its wait (pendingDirs.record). |
| 1431 | configDirsPending.record(pending, outcomes, err == nil && keep == nil) |
| 1432 | return err |
| 1433 | } |
| 1434 | |
| 1435 | // configDirsPending is what the last config pass could not write because a |
| 1436 | // nested row's instance directory did not exist yet (harness Run 9, finding |
| 1437 | // 6): host -> the directory. The activator makes that directory at the |
| 1438 | // first wake, and before this the only pass that came back for the file was |
| 1439 | // configSyncLoop's hourly one, so a nested instance created with settings |
| 1440 | // woke and ran without them, its store binding included, for up to an hour. |
| 1441 | // configGenLoop now looks at these directories on every tick and runs the |
| 1442 | // pass again as soon as one is there. |
| 1443 | var configDirsPending pendingDirs |
| 1444 | |
| 1445 | type pendingDirs struct { |
| 1446 | mu sync.Mutex |
| 1447 | dirs map[string]*pendingDir |
| 1448 | } |
| 1449 | |
| 1450 | // pendingDir is one host the config pass waits for: its instance directory, |
| 1451 | // and, once the directory is there, how many passes since have left the host |
| 1452 | // undone and how many ticks to let go by before the next retry. |
| 1453 | type pendingDir struct { |
| 1454 | dir string |
| 1455 | failures int |
| 1456 | waitTicks int |
| 1457 | } |
| 1458 | |
| 1459 | // pendingRetryCap bounds the retry backoff: at most 16 generation-loop ticks |
| 1460 | // (about a minute) between passes for a host whose directory is there but |
| 1461 | // whose config a pass keeps leaving undone (a bundle that will not read). |
| 1462 | const pendingRetryCap = 16 |
| 1463 | |
| 1464 | // record keeps what one pass found: absent, the hosts whose instance |
| 1465 | // directory was not there; outcomes, every in-scope host the pass named, |
| 1466 | // done or not (appconfig.WithHostOutcome). A host stays waited for until a |
| 1467 | // pass reports it done (fix round 1, Astra's finding 3: a pass that left it |
| 1468 | // undone on a read failure still returned nil), or until a complete pass no |
| 1469 | // longer has it in scope (its row is gone or no longer takes a bundle). A |
| 1470 | // pass that failed part way names fewer hosts than it might have, so it only |
| 1471 | // adds. A host a pass leaves undone with its directory there backs off, |
| 1472 | // doubling from the next tick up to pendingRetryCap ticks. |
| 1473 | func (p *pendingDirs) record(absent map[string]string, outcomes map[string]bool, complete bool) { |
| 1474 | p.mu.Lock() |
| 1475 | defer p.mu.Unlock() |
| 1476 | if p.dirs == nil { |
| 1477 | p.dirs = map[string]*pendingDir{} |
| 1478 | } |
| 1479 | for host, e := range p.dirs { |
| 1480 | if _, again := absent[host]; again { |
| 1481 | continue |
| 1482 | } |
| 1483 | done, named := outcomes[host] |
| 1484 | switch { |
| 1485 | case named && done, !named && complete: |
| 1486 | delete(p.dirs, host) |
| 1487 | case named: |
| 1488 | // The count saturates once the delay reaches the cap, so the |
| 1489 | // shift below never overflows however long a host keeps failing. |
| 1490 | if 1<<e.failures <= pendingRetryCap { |
| 1491 | e.failures++ |
| 1492 | } |
| 1493 | e.waitTicks = min(1<<(e.failures-1), pendingRetryCap) - 1 |
| 1494 | } |
| 1495 | } |
| 1496 | for host, dir := range absent { |
| 1497 | if e, ok := p.dirs[host]; ok { |
| 1498 | e.dir = dir |
| 1499 | continue |
| 1500 | } |
| 1501 | p.dirs[host] = &pendingDir{dir: dir} |
| 1502 | } |
| 1503 | } |
| 1504 | |
| 1505 | // appeared names a host whose instance directory a pass waited for, which |
| 1506 | // exists now, and whose retry is due; ok is false when there is none. Each |
| 1507 | // call is one tick of the backoff. Lstat only: the directory is the |
| 1508 | // activator's, and nothing here follows or opens it. |
| 1509 | func (p *pendingDirs) appeared() (host string, ok bool) { |
| 1510 | p.mu.Lock() |
| 1511 | defer p.mu.Unlock() |
| 1512 | for h, e := range p.dirs { |
| 1513 | if _, err := os.Lstat(e.dir); err != nil { |
| 1514 | continue |
| 1515 | } |
| 1516 | if e.waitTicks > 0 { |
| 1517 | e.waitTicks-- |
| 1518 | continue |
| 1519 | } |
| 1520 | if !ok { |
| 1521 | host, ok = h, true |
| 1522 | } |
| 1523 | } |
| 1524 | return host, ok |
| 1525 | } |
| 1526 | |
| 1527 | // configAckFunc builds the callback config sync's two passes (the root |
| 1528 | // pass's appconfig.Sync and this agent's own appconfig.SyncExec, called |
| 1529 | // through syncExecSerialized above) invoke for a route they just converged |
| 1530 | // (relay spec §5, internal/appconfig/configack.go): it stamps this box's |
| 1531 | // label onto the ack — BoxLabel (metrics_loop.go) is the same identity |
| 1532 | // every other per-box artifact this platform writes uses, so a config-ack |
| 1533 | // and this box's heartbeat/metrics rows agree about which box they belong |
| 1534 | // to — and writes it. A write failure is logged, never returned: an ack is |
| 1535 | // best-effort telemetry for `carlos relay status`, not something config |
| 1536 | // delivery itself depends on, and failing the whole sync pass over it would |
| 1537 | // make a bucket blip on the ack object cost every OTHER route its config. |
| 1538 | func configAckFunc(ctx context.Context, store release.Store, logPrefix string) func(appconfig.ConfigAck) { |
| 1539 | box := BoxLabel() |
| 1540 | return func(a appconfig.ConfigAck) { |
| 1541 | a.Box = box |
| 1542 | if err := appconfig.WriteConfigAck(ctx, store, a); err != nil { |
| 1543 | log.Printf("%s: config ack %s: %v", logPrefix, a.Host, err) |
| 1544 | } |
| 1545 | } |
| 1546 | } |
| 1547 | |
| 1548 | // configAckWriter is configAckFunc for the armed scheduler's ack writer |
| 1549 | // (rollout.Manager.DeferAck): the same stamp and write, under the context |
| 1550 | // the writer hands each write, which bounds it (Task 11, Astra R1 Flag |
| 1551 | // triage T8). configAckFunc itself stays as it is, unbounded, for flag-off |
| 1552 | // sync and the root pass (plan review #11). |
| 1553 | func configAckWriter(store release.Store, logPrefix string) func(context.Context, appconfig.ConfigAck) { |
| 1554 | box := BoxLabel() |
| 1555 | return func(ctx context.Context, a appconfig.ConfigAck) { |
| 1556 | a.Box = box |
| 1557 | if err := appconfig.WriteConfigAck(ctx, store, a); err != nil { |
| 1558 | log.Printf("%s: config ack %s: %v", logPrefix, a.Host, err) |
| 1559 | } |
| 1560 | } |
| 1561 | } |
| 1562 | |
| 1563 | // reconcileMu serializes the two provisioning loops' Reconcile passes |
| 1564 | // against each other, the same shape as configSyncMu above. They can |
| 1565 | // genuinely overlap — instancesLoop's 5-minute pass mid-flight when a |
| 1566 | // serial bump fires fastConvergeLoop's full pass set — and Reconcile's |
| 1567 | // removal phase deletes any Provisioned, in-scope route that is absent from |
| 1568 | // ITS OWN record snapshot, which it lists at the top of the pass |
| 1569 | // (provision/reconcile.go): a route the newer pass minted between the older |
| 1570 | // pass's apps/ listing and its registry listing looks exactly like a |
| 1571 | // retired record to the older pass, which would delete it. The re-mint |
| 1572 | // minutes later comes back with its capability grants revoked |
| 1573 | // (registry.Delete purges grants — that is Delete's contract, not a bug). |
| 1574 | // Serializing in-process closes the race outright; the root adopt path runs |
| 1575 | // no Reconcile and is not this mutex's problem. |
| 1576 | var reconcileMu sync.Mutex |
| 1577 | |
| 1578 | // reconcileSerialized is the one way any agent loop runs the record-family |
| 1579 | // reconcile passes — the mutex lives in a shared wrapper, not at the call |
| 1580 | // sites, so a future third caller cannot forget it. Pinned by |
| 1581 | // TestReconcileSerializedHoldsTheMutex. |
| 1582 | // |
| 1583 | // BOTH families run here, under ONE hold of the mutex: the instance pass and |
| 1584 | // the sidecar pass. Each pass's removal phase is scoped to its own route Kind |
| 1585 | // (provision.ReconcileSidecars' doc comment), so they cannot delete each |
| 1586 | // other's rows — but they read and write the same registry, and the |
| 1587 | // overlapping-pass hazard the mutex exists for (a route minted by a newer |
| 1588 | // pass looking exactly like a retired record to an older one) applies within |
| 1589 | // each family just as much. Running them as one serialized unit also means a |
| 1590 | // caller cannot converge instances while leaving sidecars a tick behind. |
| 1591 | // |
| 1592 | // Errors are JOINED rather than short-circuited: a sidecar-side failure must |
| 1593 | // not hide an instance-side one, and neither must stop the other running — |
| 1594 | // each pass is independently safe to run and independently valuable. |
| 1595 | func reconcileSerialized(ctx context.Context, s release.Store, reg *registry.Registry, cfg provision.ReconcileConfig) error { |
| 1596 | return reconcileSerializedWith(ctx, s, reg, cfg, armedMovesWorker.Load()) |
| 1597 | } |
| 1598 | |
| 1599 | // reconcileSerializedWith is reconcileSerialized with the moves worker whose |
| 1600 | // offline hooks the pass takes (movesoffline.go): the armed worker in the |
| 1601 | // agent, a box's own worker in a test. A nil worker, or one on a box that |
| 1602 | // cannot take part in an offline move, adds nothing. |
| 1603 | func reconcileSerializedWith(ctx context.Context, s release.Store, reg *registry.Registry, cfg provision.ReconcileConfig, w *movesWorker) error { |
| 1604 | reconcileMu.Lock() |
| 1605 | defer reconcileMu.Unlock() |
| 1606 | // With the change index armed, the record walks below read the record |
| 1607 | // view's snapshot, refreshed here under reconcileMu so a pass's snapshot |
| 1608 | // is never older than the previous pass's (S3 change-index design §4.3). |
| 1609 | // Writes still go to s, and a key written through the snapshot store is |
| 1610 | // read from s afterwards. Disarmed, or when the view cannot vouch, rs is |
| 1611 | // s and the pass walks the bucket as before. |
| 1612 | rs, _, _ := viewStore(ctx, s) |
| 1613 | // A fresh row is minted at the layout this box's intent names |
| 1614 | // (ReconcileConfig.Layout): nested on a nested box, so a new tenant |
| 1615 | // never runs flat there. Set here, the one place both callers pass |
| 1616 | // through, because the intent is read from the store this pass has in |
| 1617 | // hand and scoped to the label this box publishes under. |
| 1618 | if cfg.Layout == nil { |
| 1619 | cfg.Layout = func(ctx context.Context, s release.Store) (string, error) { |
| 1620 | return mintLayoutIntentWith(ctx, s, reg.List) |
| 1621 | } |
| 1622 | } |
| 1623 | // Seeded records (seeded-instances spec §3.2): acted on only once the |
| 1624 | // activator has armed the data bucket (wireSeedPlacement). |
| 1625 | if cfg.Seeds == nil { |
| 1626 | cfg.Seeds = armedSeedHooks.Load() |
| 1627 | } |
| 1628 | // Move holds (moves spec §9.6): participant rows minted held and kept. |
| 1629 | // Unchanged with the Phase 1 switch off. |
| 1630 | cfg = withMovesHolds(cfg) |
| 1631 | // Offline move holds: the hold from the record alone, the keep from the |
| 1632 | // view, which is read here first if no listing has been read yet or the |
| 1633 | // last one failed, so keep never fails closed on a box whose listing |
| 1634 | // reads. That listing reads the bucket, never the record view: what a |
| 1635 | // move keeps is decided on live data (S3 change-index final review R1, |
| 1636 | // finding 1). Nothing on a box that cannot take part, and nothing where |
| 1637 | // Phase 1's hooks are set. |
| 1638 | if w != nil && w.cfg.OfflineCapable && w.cfg.Offline != nil { |
| 1639 | if !w.cfg.Offline.built() || w.cfg.Offline.failed() { |
| 1640 | refreshOfflineView(ctx, s, w.cfg.Offline) |
| 1641 | } |
| 1642 | cfg = withOfflineHolds(cfg, offlineHold(w.cfg), w.cfg.Offline) |
| 1643 | } |
| 1644 | // Runs ride this pass's record reads (run spec §5.2): the worker notes |
| 1645 | // every in-scope record with a run, and acts after the pass. |
| 1646 | cfg = withRunsHook(cfg) |
| 1647 | return errors.Join( |
| 1648 | provision.Reconcile(ctx, rs, reg, cfg), |
| 1649 | provision.ReconcileSidecars(ctx, rs, reg, cfg), |
| 1650 | ) |
| 1651 | } |
| 1652 | |
| 1653 | // bounce builds the restart callback SyncExec calls for a route whose |
| 1654 | // rendered config actually changed. Shared by BOTH config loops (hourly |
| 1655 | // configSyncLoop and the fast configGenLoop) rather than duplicated per |
| 1656 | // loop, unlike the adopt pair: those two differ from each other (filter, |
| 1657 | // nag) where these two must not — the delivery must be identical whichever |
| 1658 | // loop happened to notice a change first, and a divergence would mean an |
| 1659 | // app converges one way at 3s and another way on the hour. A nil svc (no |
| 1660 | // activator armed) still returns a working no-op: the files converge for |
| 1661 | // future spawns, there is simply nothing running to restart. |
| 1662 | func bounce(svc *activator.Service) func(string) error { |
| 1663 | return func(host string) error { |
| 1664 | if svc == nil { |
| 1665 | return nil |
| 1666 | } |
| 1667 | if svc.Bounce(host) { |
| 1668 | log.Printf("agent: config changed, bounced %s", host) |
| 1669 | } |
| 1670 | return nil |
| 1671 | } |
| 1672 | } |
| 1673 | |
| 1674 | // genAbsent marks an app whose generation object was READ and found |
| 1675 | // missing, distinguishing it from an app never seen at all — the two are |
| 1676 | // different in the only way this loop cares about: a never-seen app must |
| 1677 | // record silently (first sight), while a known-absent app that suddenly |
| 1678 | // has a generation object is an app whose config was just written for the |
| 1679 | // first time, which must sync. A real etag can never collide with it: no |
| 1680 | // store produces one containing a NUL byte (S3 etags are hex/quoted, |
| 1681 | // DirStore's is hex). |
| 1682 | const genAbsent = "\x00absent" |
| 1683 | |
| 1684 | // genCache is everything configGenLoop remembers between ticks: one |
| 1685 | // generation etag per config-carrying app, and the previous tick's |
| 1686 | // route-shape digests. One object rather than two parameters because the |
| 1687 | // two are remembered for exactly one reason — deciding what "changed" |
| 1688 | // means — and a caller holding one without the other would silently lose |
| 1689 | // half the loop's sensitivity. |
| 1690 | type genCache struct { |
| 1691 | etags map[[2]string]string |
| 1692 | // exec/unit are appconfig.ConfigRouteFingerprints' two digests as of |
| 1693 | // the last pass that computed them; shapeSeen is false until the first |
| 1694 | // such pass. The bool is not redundant with an empty string: a box with |
| 1695 | // no config-carrying routes at all has a real, stable digest of the |
| 1696 | // empty set, and treating that as "never seen" would make its first |
| 1697 | // route look like a change on the tick after boot rather than on the |
| 1698 | // tick it appeared. |
| 1699 | exec, unit string |
| 1700 | shapeSeen bool |
| 1701 | |
| 1702 | // gate skips the read of an app whose scope has not moved since its last |
| 1703 | // successful read, younger than the backstop (S3 change-index design §5). |
| 1704 | // nil, or a gate over a nil or unusable watch, reads every app every |
| 1705 | // tick, as before the index. now is the gate's clock (nil: time.Now). |
| 1706 | gate *release.ReadGate |
| 1707 | now func() time.Time |
| 1708 | } |
| 1709 | |
| 1710 | func newGenCache() *genCache { return &genCache{etags: map[[2]string]string{}} } |
| 1711 | |
| 1712 | // genPollPass is one tick of configGenLoop, split out for tests: read the |
| 1713 | // generation etag of every app the box can converge — exec-config apps and |
| 1714 | // unit-backed apps alike — plus the registry's own route shape, and report |
| 1715 | // which HALVES moved since the cache was built. |
| 1716 | // |
| 1717 | // Semantics, each of which is a deliberate choice: |
| 1718 | // |
| 1719 | // - First sight of an app (no cache entry) records silently and does NOT |
| 1720 | // sync. Boot must not re-bounce the world — every app would look |
| 1721 | // "changed" on the first tick after a restart. configSyncLoop's ~1min |
| 1722 | // first pass owns boot convergence. |
| 1723 | // - An unchanged etag does nothing. |
| 1724 | // - Absent that stays absent does nothing, forever: an app last written |
| 1725 | // by a pre-generation CLI has no generation object and never will |
| 1726 | // until someone writes its config again. That is legal, not a fault. |
| 1727 | // - Any movement syncs: the etag changed, or absent -> present. |
| 1728 | // present -> absent shouldn't happen (generation objects are only ever |
| 1729 | // created and overwritten, never deleted), but it's treated as a change |
| 1730 | // too — one extra SyncExec pass is cheap, and pretending we didn't see |
| 1731 | // it would be a lie about the only thing this function reports. |
| 1732 | // - ONE bool PER HALF for the whole pass, however many apps moved: the |
| 1733 | // caller runs one SyncExec (which resolves every route itself) and asks |
| 1734 | // for at most one root config pass (which sweeps every unit route |
| 1735 | // itself). Per-app syncs would multiply bucket reads to deliver exactly |
| 1736 | // the same files, and per-app root requests would ask the same pass to |
| 1737 | // run several times over. |
| 1738 | // - The halves are independent: an app whose routes are all exec-backed |
| 1739 | // sets execChanged alone, an all-unit app sets unitChanged alone, and an |
| 1740 | // app holding both shapes at once sets both — off the SAME single read, |
| 1741 | // because its whole config lives in one generation object however its |
| 1742 | // routes are backed. |
| 1743 | // - Apps that vanish from the route set lose their cache entries, so a |
| 1744 | // re-added app is a first sight again rather than a comparison against |
| 1745 | // an etag remembered from before it left (a route being re-added gets |
| 1746 | // fresh config from the spawn path anyway). |
| 1747 | // - The ROUTE SHAPE is watched too, not just the bucket. A route's |
| 1748 | // Environment binding lives in the box's own registry and moves no etag |
| 1749 | // anywhere, so without this the loop cannot see a rebinding at all and |
| 1750 | // delivery waits out configSyncLoop's hourly floor: issue #196, where a |
| 1751 | // live route served the wrong merge for ten minutes and the only symptom |
| 1752 | // was the app reporting a credential missing — which reads as a BAD |
| 1753 | // credential rather than an undelivered one. First sight of the shape is |
| 1754 | // silent for the same reason first sight of an app is. This covers BOTH |
| 1755 | // doors onto the column: `carlos route -environment` on the box, and the |
| 1756 | // provisioning reconciler converging Environment from an instance record |
| 1757 | // — which is the sanctioned door for a Provisioned row, and which the |
| 1758 | // issue's own suggested fix (bump the generation from `carlos route`) |
| 1759 | // would have missed entirely. It costs no bucket read: see |
| 1760 | // appconfig.ConfigRouteFingerprints. |
| 1761 | // - A read error that isn't "not found" leaves that app's cache entry |
| 1762 | // exactly as it was and keeps evaluating the rest: the errors come back |
| 1763 | // for the caller to aggregate into ONE log line per pass, never one per |
| 1764 | // app per tick. Clearing the entry instead would turn a recovered blip |
| 1765 | // into a bogus first sight or a bogus change. |
| 1766 | // - With the change index armed (cache.gate), an app already in the cache |
| 1767 | // is read only when its scope moved, its last read failed, or that read |
| 1768 | // is the backstop's age (S3 change-index design §5). A skipped app keeps |
| 1769 | // its entry and reports no movement. |
| 1770 | // |
| 1771 | // The cache is mutated in place — it lives across ticks in the loop, which |
| 1772 | // is the whole point (there is nowhere else to remember what "changed" |
| 1773 | // means). |
| 1774 | func genPollPass(ctx context.Context, routes []registry.Route, store release.Store, cache *genCache) (execChanged, unitChanged bool, readErrs []error) { |
| 1775 | // The registry-side half (issue #196). Free — no bucket read at all — |
| 1776 | // and evaluated first so a rebinding converges on the same tick a var |
| 1777 | // change would, rather than one behind it. |
| 1778 | execFP, unitFP := appconfig.ConfigRouteFingerprints(routes) |
| 1779 | if cache.shapeSeen { |
| 1780 | execChanged = execChanged || execFP != cache.exec |
| 1781 | unitChanged = unitChanged || unitFP != cache.unit |
| 1782 | } |
| 1783 | cache.exec, cache.unit, cache.shapeSeen = execFP, unitFP, true |
| 1784 | |
| 1785 | execApps := appconfig.ExecConfigApps(routes) |
| 1786 | unitApps := appconfig.UnitConfigApps(routes) |
| 1787 | // The poll set is the UNION, read once per app: the two sets overlap |
| 1788 | // whenever an app has both route shapes, and its config is one object |
| 1789 | // either way. inExec/inUnit then say which half a movement belongs to. |
| 1790 | inExec := make(map[[2]string]bool, len(execApps)) |
| 1791 | for _, k := range execApps { |
| 1792 | inExec[k] = true |
| 1793 | } |
| 1794 | inUnit := make(map[[2]string]bool, len(unitApps)) |
| 1795 | for _, k := range unitApps { |
| 1796 | inUnit[k] = true |
| 1797 | } |
| 1798 | apps := make([][2]string, 0, len(execApps)+len(unitApps)) |
| 1799 | apps = append(apps, execApps...) |
| 1800 | for _, k := range unitApps { |
| 1801 | if !inExec[k] { |
| 1802 | apps = append(apps, k) |
| 1803 | } |
| 1804 | } |
| 1805 | now := time.Now |
| 1806 | if cache.now != nil { |
| 1807 | now = cache.now |
| 1808 | } |
| 1809 | at := now() |
| 1810 | live := make(map[[2]string]bool, len(apps)) |
| 1811 | for _, k := range apps { |
| 1812 | live[k] = true |
| 1813 | key := appconfig.GenKey(k[0], k[1]) |
| 1814 | // The gate (spec §5): an app already in the cache is read again only |
| 1815 | // when its scope moved, its last read failed, or that read is the |
| 1816 | // backstop's age. An app not in the cache is always read: its first |
| 1817 | // sight is what a later change is compared against. The generation |
| 1818 | // is taken before the read, so a write landing during it stays due. |
| 1819 | scope, gated := release.ChangeScope(key) |
| 1820 | gen, due := cache.gate.Due(scope, at) |
| 1821 | if _, seen := cache.etags[k]; seen && gated && !due { |
| 1822 | continue |
| 1823 | } |
| 1824 | _, etag, err := store.Get(ctx, key) |
| 1825 | if gated { |
| 1826 | cache.gate.Mark(scope, gen, at, err == nil || errors.Is(err, os.ErrNotExist)) |
| 1827 | } |
| 1828 | switch { |
| 1829 | case errors.Is(err, os.ErrNotExist): |
| 1830 | etag = genAbsent |
| 1831 | case err != nil: |
| 1832 | readErrs = append(readErrs, fmt.Errorf("%s/%s: %w", k[0], k[1], err)) |
| 1833 | continue // cache entry untouched: this app is unknown this tick, not changed |
| 1834 | } |
| 1835 | prev, seen := cache.etags[k] |
| 1836 | cache.etags[k] = etag |
| 1837 | if seen && prev != etag { |
| 1838 | execChanged = execChanged || inExec[k] |
| 1839 | unitChanged = unitChanged || inUnit[k] |
| 1840 | } |
| 1841 | } |
| 1842 | for k := range cache.etags { |
| 1843 | if !live[k] { |
| 1844 | delete(cache.etags, k) |
| 1845 | } |
| 1846 | } |
| 1847 | return execChanged, unitChanged, readErrs |
| 1848 | } |
| 1849 | |
| 1850 | // configSyncMarker is the file configGenLoop touches to ask a root config |
| 1851 | // pass to run now: carlos-adopt.path watches it (PathChanged) and starts |
| 1852 | // carlos-config-sync.service (`carlos adopt -config-only`), which owns the |
| 1853 | // unit half of config delivery (/etc/carlos/apps + try-restart) that this |
| 1854 | // unprivileged process cannot — and only that half. |
| 1855 | // It lives in the data dir, deliberately outside the *.config.env namespace |
| 1856 | // reapConfigEnv sweeps — the reaper matches that suffix alone and must go |
| 1857 | // on ignoring this. |
| 1858 | const configSyncMarker = "config-sync-requested" |
| 1859 | |
| 1860 | // requestRootSync touches the marker. The content is a timestamp only so an |
| 1861 | // operator reading /data can see when the last request went out; systemd |
| 1862 | // watches the path, not the bytes. Errors are the caller's to log and |
| 1863 | // swallow: a request that cannot be written costs latency, not correctness, |
| 1864 | // because carlos-adopt.timer's hourly pass is the floor underneath it. |
| 1865 | func requestRootSync(dd string) error { |
| 1866 | return os.WriteFile(filepath.Join(dd, configSyncMarker), |
| 1867 | []byte(time.Now().UTC().Format(time.RFC3339)+"\n"), 0o644) |
| 1868 | } |
| 1869 | |
| 1870 | // stageRequestMarker is the release half of configSyncMarker. The agent's |
| 1871 | // adopt loops run inside carlos-edge.service (ProtectSystem=strict, no /opt |
| 1872 | // write), so they can SEE a promotion within seconds and stage nothing; the |
| 1873 | // only process that can write the version store is the root pass on |
| 1874 | // carlos-adopt.timer, an hour away. Touching this file relays "there is a |
| 1875 | // version here I cannot stage" to carlos-stage.path, exactly as the config |
| 1876 | // half relays to carlos-adopt.path. Lives in the data dir beside its config |
| 1877 | // twin, outside the *.config.env namespace reapConfigEnv sweeps. |
| 1878 | const stageRequestMarker = "stage-requested" |
| 1879 | |
| 1880 | // stageRequestEvery throttles those requests. Unlike the config marker — |
| 1881 | // touched only when a generation etag MOVES, which is naturally rare — a |
| 1882 | // blocked stage repeats on every tick of a loop that runs every 2-3 seconds, |
| 1883 | // so an untouched throttle would ask systemd to start a root pass ~30 times a |
| 1884 | // minute and each pass would re-block behind the one before it. |
| 1885 | // |
| 1886 | // Fifteen seconds, and a request inside the window is DEFERRED to its end, |
| 1887 | // never dropped (stageThrottle). It was a minute and dropped, and a second |
| 1888 | // release inside the minute then waited for whatever next moved the serial: |
| 1889 | // 98 extra seconds on the flagship on 2026-10-06 (#231). A root pass takes |
| 1890 | // about eight seconds, and the window has to outlast it: carlos-stage.path |
| 1891 | // watches with PathChanged, and systemd does not act on a change made while |
| 1892 | // carlos-stage.service is still running. |
| 1893 | const stageRequestEvery = 15 * time.Second |
| 1894 | |
| 1895 | // stageRequester returns a release.OnStageBlocked callback that touches the |
| 1896 | // marker at most once per stageRequestEvery. The closure holds the clock, so |
| 1897 | // each loop gets its own budget: the fast loop asking does not silence the |
| 1898 | // hourly one, and neither can starve the other. |
| 1899 | // |
| 1900 | // Failures are logged and swallowed, like requestRootSync's: a request that |
| 1901 | // cannot be written costs latency, never correctness, because |
| 1902 | // carlos-adopt.timer remains the floor underneath it. |
| 1903 | func stageRequester(dd string) func(host, ver string, err error) { |
| 1904 | ask := stageMarkerRequester(dd) |
| 1905 | return func(host, ver string, err error) { |
| 1906 | ask(host, ver, err) |
| 1907 | // The stage relay (stagerelay.go) learns the head here and not in |
| 1908 | // markerRequester, which mintRequester shares: a proof-mint request |
| 1909 | // is not a head that failed to stage. |
| 1910 | if r := armedStageRelay.Load(); r != nil { |
| 1911 | r.note(host, ver) |
| 1912 | } |
| 1913 | } |
| 1914 | } |
| 1915 | |
| 1916 | // stageMarkerRequester is stageRequester without the relay: the same marker, |
| 1917 | // throttle and log lines, and nothing remembered. The moves worker uses it |
| 1918 | // (movespass.go): its version is a move's held version, not a channel head, |
| 1919 | // and a held stage neither moves the installed Version nor leaves a refusal, |
| 1920 | // so the relay would have no way to know the request was answered. |
| 1921 | func stageMarkerRequester(dd string) func(host, ver string, err error) { |
| 1922 | return markerRequester(dd, func(host, ver string, err error) { |
| 1923 | if err != nil { |
| 1924 | log.Printf("agent: cannot request root stage for %s %s: %v", host, ver, err) |
| 1925 | return |
| 1926 | } |
| 1927 | // One line for every cause the marker relays (Task 12a): a build |
| 1928 | // this process cannot stage, an install request only root |
| 1929 | // resolves (ErrInstallRequestNeedsRoot), a named unit only root |
| 1930 | // restarts (ErrUnitRestartNeedsRoot), and a held wake's target only |
| 1931 | // root installs. The throttle does not carry the cause. |
| 1932 | log.Printf("agent: %s needs a root pass for %s (a build only root can stage or install, a unit only root can restart, or an install request only root resolves); asked for one", host, ver) |
| 1933 | }) |
| 1934 | } |
| 1935 | |
| 1936 | // mintRequester is the mint asker's (proofask.go) request: the same marker |
| 1937 | // and the same throttle shape as stageRequester's, with its own budget and |
| 1938 | // its own words. The asker logs why it asks; this logs only a failed write. |
| 1939 | func mintRequester(dd string) func(host, ver string, err error) { |
| 1940 | return markerRequester(dd, func(host, ver string, err error) { |
| 1941 | if err != nil { |
| 1942 | log.Printf("agent: cannot request a root pass to mint a proof for %s %s: %v", host, ver, err) |
| 1943 | } |
| 1944 | }) |
| 1945 | } |
| 1946 | |
| 1947 | // markerRequester touches the stage-request marker through a fresh |
| 1948 | // stageThrottle and hands said each write's outcome. |
| 1949 | func markerRequester(dd string, said func(host, ver string, err error)) func(host, ver string, err error) { |
| 1950 | t := &stageThrottle{every: stageRequestEvery, now: time.Now, after: func(d time.Duration, f func()) { time.AfterFunc(d, f) }, |
| 1951 | write: func(host, ver string) { |
| 1952 | said(host, ver, os.WriteFile(filepath.Join(dd, stageRequestMarker), |
| 1953 | []byte(time.Now().UTC().Format(time.RFC3339)+"\n"), 0o644)) |
| 1954 | }} |
| 1955 | return func(host, ver string, _ error) { t.request(host, ver) } |
| 1956 | } |
| 1957 | |
| 1958 | // stageThrottle writes at most once per every, measured from the end of one |
| 1959 | // write to the start of the next. While a write is scheduled (busy), every |
| 1960 | // request rides it: a deferred write owns the window however late its timer |
| 1961 | // fires. A request inside the window schedules one write for the window's |
| 1962 | // end, so none waits longer than every after the write before it. |
| 1963 | type stageThrottle struct { |
| 1964 | every time.Duration |
| 1965 | now func() time.Time |
| 1966 | after func(time.Duration, func()) |
| 1967 | write func(host, ver string) |
| 1968 | mu sync.Mutex |
| 1969 | last time.Time // when the last write finished |
| 1970 | busy bool |
| 1971 | } |
| 1972 | |
| 1973 | func (t *stageThrottle) request(host, ver string) { |
| 1974 | t.mu.Lock() |
| 1975 | if t.busy { |
| 1976 | t.mu.Unlock() |
| 1977 | return |
| 1978 | } |
| 1979 | t.busy = true |
| 1980 | // The write runs under the lock: a request made once the marker has |
| 1981 | // changed may concern a pointer root has already read, so it waits for |
| 1982 | // the write to finish and schedules the next window, rather than |
| 1983 | // riding a write that cannot answer it. |
| 1984 | do := func() { |
| 1985 | t.mu.Lock() |
| 1986 | t.write(host, ver) |
| 1987 | t.busy, t.last = false, t.now() |
| 1988 | t.mu.Unlock() |
| 1989 | } |
| 1990 | if since := t.now().Sub(t.last); !t.last.IsZero() && since < t.every { |
| 1991 | t.mu.Unlock() |
| 1992 | t.after(t.every-since, do) |
| 1993 | return |
| 1994 | } |
| 1995 | t.mu.Unlock() |
| 1996 | do() |
| 1997 | } |
| 1998 | |
| 1999 | // configGenLoop is the fast-converge shape applied to issue #44: a config |
| 2000 | // set should converge like a canary adopts — seconds, not the next hour. |
| 2001 | // Every ~3s it reads one small object per config-carrying app (the |
| 2002 | // generation object every config write touches) and, on movement, delivers |
| 2003 | // the halves it can: ONE SyncExec pass for the whole tick for exec-backed |
| 2004 | // instances, and one touch of the root-sync marker for unit-backed ones. |
| 2005 | // Spurious movement costs a pass that finds nothing to write: SyncExec (and |
| 2006 | // the root pass's appconfig.Sync) bounces only routes whose rendered file |
| 2007 | // actually changed, so a redundant touch can never cost a restart. |
| 2008 | // |
| 2009 | // Gated on CARLOS_SECRETS_KEY exactly as configSyncLoop is, for a reason |
| 2010 | // beyond convention: without the key SyncExec cannot render an app's |
| 2011 | // environment at all, so this loop would poll the bucket every 3 seconds |
| 2012 | // forever to signal no one. openDeployStore failing is the same story with |
| 2013 | // nothing to read. The gate covers the unit half too, which is an accepted |
| 2014 | // asymmetry rather than a considered one: a box whose agent lacks the key |
| 2015 | // gets no fast path for its unit apps either, even though the root pass |
| 2016 | // that would run holds its own key. In practice both units read the same |
| 2017 | // /etc/carlos/host.env, so the case is theoretical — and the hourly timer |
| 2018 | // floor stands regardless. |
| 2019 | // |
| 2020 | // Cardinality: unlike fastConvergeLoop's sweep, which is bounded to the |
| 2021 | // deliberately rare canary/* routes plus still-unstaged ones, this watches |
| 2022 | // EVERY config-carrying app on the box. That is accepted knowingly — one GET |
| 2023 | // of the smallest object in the bucket |
| 2024 | // per app per 3s, against the store fastConvergeLoop already hits every 2s. What |
| 2025 | // keeps it bounded is that the pass reads generation objects only: no |
| 2026 | // manifests, no pointers, no per-environment resolution. All of that |
| 2027 | // happens inside the SyncExec a real change triggers. |
| 2028 | // |
| 2029 | // Deliberately NOT done here: |
| 2030 | // |
| 2031 | // - No fast first pass. The loop sleeps BEFORE its first tick and the |
| 2032 | // first sight of every app is silent (see genPollPass) — boot |
| 2033 | // convergence belongs to configSyncLoop's ~1min first pass, and a box |
| 2034 | // coming up must not bounce every instance it has because it has never |
| 2035 | // seen their etags before. |
| 2036 | // - No unit-backed rendering or restarting. This process still writes |
| 2037 | // nothing under /etc and restarts no unit (see RunAdopt): a unit app's |
| 2038 | // movement only touches <dataDir>/config-sync-requested, which |
| 2039 | // carlos-adopt.path relays into a root carlos-config-sync.service run |
| 2040 | // (`carlos adopt -config-only`) that does both. The relay reaches ONLY |
| 2041 | // the config half — never release adoption, never the canary-TTL nag, |
| 2042 | // which would otherwise fire off an unrelated app's config write |
| 2043 | // instead of hourly. The relay can drop an event that lands while that |
| 2044 | // service is mid-run — accepted: the next config write re-touches the |
| 2045 | // marker, and the timer floor delivers within the hour regardless. |
| 2046 | // - No hourly floor replaced. configSyncLoop stays exactly as it was: it |
| 2047 | // catches apps written by older CLIs (no generation object at all), |
| 2048 | // ticks this loop missed, and the boot-retry gap of issue #27. |
| 2049 | // |
| 2050 | // Thin and untested by convention, like adoptLoop/statusLoop — the tick's |
| 2051 | // whole decision lives in genPollPass, which is tested. |
| 2052 | func configGenLoop(reg *registry.Registry, svc *activator.Service, dd string, writes hostwrite.Config) { |
| 2053 | raw := os.Getenv("CARLOS_SECRETS_KEY") |
| 2054 | if raw == "" { |
| 2055 | log.Printf("agent: config generation loop disabled: CARLOS_SECRETS_KEY not set") |
| 2056 | return |
| 2057 | } |
| 2058 | priv, err := appconfig.ParsePrivateKey(raw, "CARLOS_SECRETS_KEY") |
| 2059 | if err != nil { |
| 2060 | log.Printf("agent: config generation loop disabled: %v", err) |
| 2061 | return |
| 2062 | } |
| 2063 | ctx := context.Background() |
| 2064 | store, err := openDeployStore(ctx) |
| 2065 | if err != nil { |
| 2066 | log.Printf("agent: config generation loop disabled: %v", err) |
| 2067 | return |
| 2068 | } |
| 2069 | // One cache for the life of the process: what "changed" means is |
| 2070 | // exactly "differs from the last tick that could read it". Its gate is |
| 2071 | // built here, after RunAgent armed the change index. |
| 2072 | cache := newGenCache() |
| 2073 | cache.gate = newChangeGate() |
| 2074 | log.Printf("agent: config generation loop armed (3s + jitter, exec-config and unit-backed apps)") |
| 2075 | for { |
| 2076 | time.Sleep(3*time.Second + rand.N(time.Second)) |
| 2077 | routes, err := reg.List() |
| 2078 | if err != nil { |
| 2079 | // The registry is a local file this process already holds open; |
| 2080 | // a failing List is a broken box, not a passing blip, so it is |
| 2081 | // worth a line even at this cadence (the same call is logged per |
| 2082 | // tick by statusLoop). |
| 2083 | log.Printf("agent: config generation poll: list routes: %v", err) |
| 2084 | continue |
| 2085 | } |
| 2086 | configGenTick(ctx, routes, reg, store, priv, dd, svc, writes, cache) |
| 2087 | } |
| 2088 | } |
| 2089 | |
| 2090 | // configGenTick is one tick of configGenLoop over the routes it listed, split |
| 2091 | // out so a test can run the loop's own decision. |
| 2092 | func configGenTick(ctx context.Context, routes []registry.Route, reg *registry.Registry, store release.Store, priv *ecdh.PrivateKey, dd string, svc *activator.Service, writes hostwrite.Config, cache *genCache) { |
| 2093 | execChanged, unitChanged, readErrs := genPollPass(ctx, routes, store, cache) |
| 2094 | if len(readErrs) > 0 { |
| 2095 | // ONE line per pass, never one per app: a bucket having a bad |
| 2096 | // minute would otherwise flood the journal at 20 lines a minute |
| 2097 | // per app. The count carries the scale, the first error carries |
| 2098 | // the cause; they are the same cause in practice. |
| 2099 | log.Printf("agent: config generation poll: %d apps unreadable: %v", len(readErrs), readErrs[0]) |
| 2100 | } |
| 2101 | if unitChanged { |
| 2102 | // Requested first and never fatal: the write is instant while |
| 2103 | // syncExecSerialized can block on a serialized pass, so the root |
| 2104 | // run gets to start alongside it rather than behind it. A failed |
| 2105 | // marker write costs those apps the wait until |
| 2106 | // carlos-adopt.timer's next hourly pass — never correctness. |
| 2107 | if err := requestRootSync(dd); err != nil { |
| 2108 | log.Printf("agent: request root config sync: %v", err) |
| 2109 | } else { |
| 2110 | log.Printf("agent: config changed for unit-backed app(s), requested root config sync pass") |
| 2111 | } |
| 2112 | } |
| 2113 | // A nested row an earlier pass could not write for because its instance |
| 2114 | // directory was not there yet: the first wake has made it, so its |
| 2115 | // bundle goes in now (finding 6). The pass bounces the instance onto |
| 2116 | // its settings once the wake it is in has finished (Service.Bounce). |
| 2117 | if host, ok := configDirsPending.appeared(); ok && !execChanged { |
| 2118 | log.Printf("agent: %s: its instance directory is there now, writing its config", host) |
| 2119 | execChanged = true |
| 2120 | } |
| 2121 | if !execChanged { |
| 2122 | return |
| 2123 | } |
| 2124 | if err := syncExecSerialized(ctx, reg, store, priv, dd, svc, writes); err != nil { |
| 2125 | log.Printf("agent: config generation sync: %v", err) |
| 2126 | } |
| 2127 | } |
| 2128 | |
| 2129 | // Fast-converge cadence and clamps (carlos-deploy spec §3-§4). The tick is |
| 2130 | // the serial poll's latency promise (~2s from console bump to pass); the |
| 2131 | // clamp bounds what a burst of console writes can cost (one full pass set |
| 2132 | // per 5s, with the last bump deferred, never dropped); the backoff bounds |
| 2133 | // what a permanently-unstageable route can cost (one look per 5min at the |
| 2134 | // cap, instead of one per tick forever). |
| 2135 | const ( |
| 2136 | fastTickEvery = 2 * time.Second |
| 2137 | fastFullClamp = 5 * time.Second |
| 2138 | fastBackoffBase = 2 * time.Second |
| 2139 | fastBackoffCap = 5 * time.Minute |
| 2140 | |
| 2141 | // The restart-failure hold (noteRestartFailed) starts higher and caps |
| 2142 | // higher than the staging backoff above, because what it bounds is not |
| 2143 | // a retry's COST to this box but a service CYCLE on a tenant's process: |
| 2144 | // re-attempting a version whose restart just failed repoints, performs a |
| 2145 | // REAL restart of the running old service, fails, rolls back and starts |
| 2146 | // it again — retrying that 2 seconds later is near-certain to do it all |
| 2147 | // again. 30s → 15m doubling means a persistently broken head costs a |
| 2148 | // handful of honest cycles in its first half hour and then four an hour |
| 2149 | // from this loop, with the hourly floors' one-each on top — instead of |
| 2150 | // one every other second for as long as the broken head stands. |
| 2151 | fastRestartBackoffBase = 30 * time.Second |
| 2152 | fastRestartBackoffCap = 15 * time.Minute |
| 2153 | ) |
| 2154 | |
| 2155 | // fastDeps is one fastConvergeLoop's state and seams — a struct so the tick |
| 2156 | // is testable without goroutines, env vars, or sleeps. |
| 2157 | type fastDeps struct { |
| 2158 | reg *registry.Registry |
| 2159 | store release.Store |
| 2160 | pub ed25519.PublicKey |
| 2161 | optDir string |
| 2162 | restart func(string) error |
| 2163 | passOpts []release.PassOption // fastPassOpts — every sweep pass carries all of them |
| 2164 | fullPass func(ctx context.Context) // reconcile → adopt → config sync, or a test hook |
| 2165 | restartLeg func(ctx context.Context) // the `carlos restart` check (issue #113), or a test hook |
| 2166 | wakeLeg func(ctx context.Context) // the sidecar wake check (design: sidecar members), or a test hook |
| 2167 | scheduleLeg func(ctx context.Context) // the schedule converge (spec §6.2), or a test hook; nil = not armed |
| 2168 | layoutLeg func(ctx context.Context) // the layout pass (isolated-exec design §8.2), or a test hook; nil = not armed |
| 2169 | systemLeg *systemLeg // the platform self-update leg (spec §4); nil = not armed |
| 2170 | proof *release.ProofGate // D3's gate (agentProofGate); nil = no proof waits to keep |
| 2171 | mintLeg func(ctx context.Context) // D3's mint asker (proofask.go); nil = not armed |
| 2172 | lastSerial int64 |
| 2173 | seeded bool |
| 2174 | lastFull time.Time |
| 2175 | notBefore map[string]time.Time |
| 2176 | step map[string]time.Duration |
| 2177 | restartHold map[string]*restartFailure // hosts held out of the sweep after a failed restart; see noteRestartFailed |
| 2178 | serialBroken bool // last ReadSerial failed and was logged; reset on success |
| 2179 | serialNotify chan<- struct{} |
| 2180 | |
| 2181 | // The change index (CARLOS_CHANGE_INDEX=read, S3 change-index design |
| 2182 | // §4.5); watch nil leaves the tick exactly as before. The watch is shared |
| 2183 | // with the record view, so Poll's moved list is advisory: the tick diffs |
| 2184 | // Observed against indexActed, the generations it last acted on, and |
| 2185 | // moves that map on only once a pass has covered them. |
| 2186 | watch *release.ChangeWatch |
| 2187 | scopedPass func(ctx context.Context, moved []string) // passLegs.scoped, or a test hook |
| 2188 | fullDue *atomic.Bool // the backstop's request for a full pass; nil = none |
| 2189 | indexActed map[string]uint64 // nil until the first good observation seeds it |
| 2190 | indexBroken bool // last Poll failed and was logged; reset on success |
| 2191 | |
| 2192 | // canaryGate gates the sweep's canary pointer reads (spec §5): a canary |
| 2193 | // route stays a member of every tick, but its group's pointer is read |
| 2194 | // only when the app's scope moved, the last read failed or is the |
| 2195 | // backstop's age, or a route of the group is not live on the head that |
| 2196 | // read returned (canaryHeads). nil: every pointer is read, as before. |
| 2197 | canaryGate *release.ReadGate |
| 2198 | canaryHeads map[[3]string]canaryRead // (account, app, channel) -> its group's last good read |
| 2199 | } |
| 2200 | |
| 2201 | // canaryRead is one canary group's last good pointer read: the head it |
| 2202 | // returned, the app scope's generation captured before it, and when. The |
| 2203 | // gate's mark is per app scope, so another channel of the same app can mark |
| 2204 | // it; a group is settled only on its own read (final review R1, finding 6). |
| 2205 | type canaryRead struct { |
| 2206 | head string |
| 2207 | gen uint64 |
| 2208 | at time.Time |
| 2209 | } |
| 2210 | |
| 2211 | // restartFailure is one host's cross-pass memo of a version its restart |
| 2212 | // refused: which version failed, the current hold step, and when the sweep |
| 2213 | // may try that host again. Kept on the loop (fastDeps), never in |
| 2214 | // release.Pass — the pass stays stateless; this loop is the caller whose |
| 2215 | // cadence needs the bound. |
| 2216 | type restartFailure struct { |
| 2217 | ver string |
| 2218 | step time.Duration |
| 2219 | notBefore time.Time |
| 2220 | } |
| 2221 | |
| 2222 | // fastPassOpts is the option set EVERY fast-converge pass carries — the |
| 2223 | // serial-triggered full release.Pass and the sweep's release.PassFilter |
| 2224 | // both, built in one place so neither call site can drift from the other. |
| 2225 | // |
| 2226 | // DO NOT add release.NagExpiredCanaries(...) to this set, ever. This |
| 2227 | // is THE loop release.NagExpiredCanaries' doc comment warns about by |
| 2228 | // name: every 2-3 seconds, canary/* always in the set. An expired canary sitting |
| 2229 | // still would nag on every single tick — a journal-flooding poll storm |
| 2230 | // out of what is supposed to be an hourly, easy-to-skim log line. The |
| 2231 | // nag belongs to RunAdopt (carlos-adopt.service) alone. |
| 2232 | // |
| 2233 | // DO NOT add release.WithCapabilities(...) to this set, ever. Same |
| 2234 | // sandbox as adoptLoop above (ProtectSystem=strict, |
| 2235 | // ReadWritePaths=/data /run/carlos), and this one ticks every 2-3 |
| 2236 | // SECONDS: arming the converge here would write nothing and log a |
| 2237 | // permission denial forever. The accepted consequence, ruled on in |
| 2238 | // the 2026-08-03 review of the capability design: a freshly promoted |
| 2239 | // canary runs its new binary against the unit shape the previous |
| 2240 | // version left until the hourly root pass converges it — up to an |
| 2241 | // hour. release.adoptHost prints one line per such repoint so the |
| 2242 | // window is visible in the journal rather than silent. |
| 2243 | // |
| 2244 | // OnStageBlocked IS wired here, and this is the loop it matters most |
| 2245 | // for: a canary is promoted to be picked up in seconds, and staging is |
| 2246 | // the one step of that pickup this process cannot perform. Throttled |
| 2247 | // inside stageRequester — this tick is 2-3 seconds, the request is at |
| 2248 | // most one per stageRequestEvery (the requester is built ONCE per loop, so the budget |
| 2249 | // is the loop's, not one tick's). |
| 2250 | // |
| 2251 | // RestartHibernating IS wired here too, and for the same reason as on |
| 2252 | // adoptLoop: this process owns the exec children. It matters more on |
| 2253 | // this loop than on any other — a promotion is bumped or fast-swept |
| 2254 | // precisely to be picked up in seconds, and the routes this loop reaches |
| 2255 | // are the warm ones somebody is watching. Pinned by |
| 2256 | // TestFastSweepBouncesAWarmHibernatingCanary: dropping it would restore |
| 2257 | // a dead-branch production bug without failing any other test. |
| 2258 | // |
| 2259 | // WithStartUnit IS wired here too (start — makeStartRoute in production, |
| 2260 | // nil in tests that never reach a rollback): any pass whose restart |
| 2261 | // callback carries the named-unit arm must also carry the start that |
| 2262 | // undoes that arm's failed restart, or its rollback leaves the service |
| 2263 | // down and the next pass stamps a never-run build |
| 2264 | // (release.WithStartUnit). It asks for no authority the try-restart arm |
| 2265 | // did not already need, and it cannot fire for a sleeper: only an |
| 2266 | // ERRORED restart reaches it, and every arm that can meet a sleeping |
| 2267 | // route swallows its failure. |
| 2268 | // |
| 2269 | // unitsOption hands named units to root when this process is not root |
| 2270 | // (issue #315); see agentUnitsOption. |
| 2271 | func fastPassOpts(dd string, stageReq func(host, ver string, err error), restart, start func(string) error) []release.PassOption { |
| 2272 | return append([]release.PassOption{ |
| 2273 | release.BakeDir(dd), |
| 2274 | release.OnStageBlocked(stageReq), |
| 2275 | release.RestartHibernating(restart), |
| 2276 | release.WithStartUnit(start), |
| 2277 | release.WithDigestRatchet(pointerRatchet(false, release.OptDir(), dd, os.Getenv)), |
| 2278 | // Settings spec §4: the agent applies root's decisions, never |
| 2279 | // resolves. The book is opened per pass, not here: these options are |
| 2280 | // built once per agent run. |
| 2281 | updatesPassOption(dd, false), |
| 2282 | }, agentUnitsOption()...) |
| 2283 | } |
| 2284 | |
| 2285 | // agentRunsAsRoot is the agent's euid test, a seam for tests. |
| 2286 | var agentRunsAsRoot = func() bool { return os.Geteuid() == 0 } |
| 2287 | |
| 2288 | // agentUnitsOption is release.RootRestartsUnits for an agent that is not |
| 2289 | // root (issue #315). carlos-edge.service runs as User=carlos, and systemd |
| 2290 | // refuses that user a restart of any unit a route names, so a named-unit |
| 2291 | // route goes to the root pass through the stage relay before anything moves. |
| 2292 | // An agent run as root restarts the unit itself, as before. |
| 2293 | func agentUnitsOption() []release.PassOption { |
| 2294 | if agentRunsAsRoot() { |
| 2295 | return nil |
| 2296 | } |
| 2297 | return []release.PassOption{release.RootRestartsUnits()} |
| 2298 | } |
| 2299 | |
| 2300 | // fastMembers computes the fast set for one tick — the routes with a reason |
| 2301 | // to be looked at FASTER than the hourly floor: |
| 2302 | // |
| 2303 | // - canary/* routes, always: a direct-bucket operator promote |
| 2304 | // (scripts/promote-app.sh, or `carlos promote` against the bucket) |
| 2305 | // never bumps the serial, so the sweep is the only thing keeping the |
| 2306 | // "canary adopts in seconds" promise (spec §4 correction). |
| 2307 | // - unstaged routes (no live symlink), once their backoff window has |
| 2308 | // passed: a route was just minted, or its stage failed; the sweep |
| 2309 | // picks it up as soon as there is something to stage. |
| 2310 | // - routes waiting for a dev proof (D3), each time the proof gate's |
| 2311 | // cached miss runs out: they have a live version, but their head is |
| 2312 | // held until its proof arrives. |
| 2313 | // |
| 2314 | // Computed BEFORE any store work so an empty set costs zero store reads — |
| 2315 | // the idle box pays one serial GET per tick and nothing else. A registry |
| 2316 | // list failure returns the empty set (a broken box, logged per tick by |
| 2317 | // statusLoop's identical call — not worth a second line every 2s here). |
| 2318 | func (d *fastDeps) fastMembers(now time.Time) []string { |
| 2319 | routes, err := d.reg.List() |
| 2320 | if err != nil { |
| 2321 | return nil |
| 2322 | } |
| 2323 | var members []string |
| 2324 | seen := map[string]bool{} |
| 2325 | for _, rt := range routes { |
| 2326 | // release.Adoptable, not a local copy: the sweep must count as |
| 2327 | // outstanding work exactly the routes a pass would consider — |
| 2328 | // anything else (platform aliases, non-ready routes, routes with no |
| 2329 | // app) would be swept forever, since the pass refuses it on shape. |
| 2330 | if !release.Adoptable(rt) || seen[rt.Host] { |
| 2331 | continue |
| 2332 | } |
| 2333 | // A host inside its restart-failure hold is outside the sweep |
| 2334 | // entirely — canary or unstaged. Sweeping it again at this cadence |
| 2335 | // would repoint, cycle the RUNNING old service through a restart |
| 2336 | // that is going to fail, roll back and start it again, every 2-3 |
| 2337 | // seconds for as long as the broken head stands (before the |
| 2338 | // rollback-start fix that flap self-terminated after two ticks — by |
| 2339 | // stamping the lie instead). The hourly floors still retry held |
| 2340 | // hosts; a console promote's serial bump still converges them |
| 2341 | // immediately (fullPass is not membership-gated); and a restart that |
| 2342 | // SUCCEEDS clears the hold at once (observeRestart). |
| 2343 | if d.restartHeld(rt.Host, now) { |
| 2344 | continue |
| 2345 | } |
| 2346 | if isCanaryRoute(rt) { |
| 2347 | seen[rt.Host] = true |
| 2348 | members = append(members, rt.Host) |
| 2349 | continue |
| 2350 | } |
| 2351 | // D3 (spec §4.5, codex 14): a listed route waiting for a dev proof |
| 2352 | // stays in the sweep although it has a live version, whenever its |
| 2353 | // 30 s miss has run out, so it adopts soon after the proof arrives |
| 2354 | // rather than at the next serial bump or the hourly floor. A wait the |
| 2355 | // last sweep left standing (fastTick) is inside a backoff window, and |
| 2356 | // sits it out like an unstaged route. |
| 2357 | if d.proof.WaitingDue(rt.Host) { |
| 2358 | if nb, ok := d.notBefore[rt.Host]; ok && now.Before(nb) { |
| 2359 | continue |
| 2360 | } |
| 2361 | seen[rt.Host] = true |
| 2362 | members = append(members, rt.Host) |
| 2363 | continue |
| 2364 | } |
| 2365 | // A LiveVersion error reads as unstaged: the pass is the thing that |
| 2366 | // can repair a broken link, and the backoff bounds the retry cost. |
| 2367 | if v, err := release.LiveVersion(d.optDir, rt.Host); err == nil && v != "" { |
| 2368 | continue // staged: nothing outstanding at this cadence |
| 2369 | } |
| 2370 | if nb, ok := d.notBefore[rt.Host]; ok && now.Before(nb) { |
| 2371 | continue // still inside this host's backoff window |
| 2372 | } |
| 2373 | seen[rt.Host] = true |
| 2374 | members = append(members, rt.Host) |
| 2375 | } |
| 2376 | return members |
| 2377 | } |
| 2378 | |
| 2379 | // noteUnstaged records that host came through a sweep still unstaged: |
| 2380 | // double its backoff (base 2s, capped at 5min) and push its next look out. |
| 2381 | // Held routes cost almost nothing this way: the pass's own bake logic |
| 2382 | // refuses them and the backoff keeps re-checks rare; the serial-triggered |
| 2383 | // and hourly passes re-notice them when the window expires. |
| 2384 | func (d *fastDeps) noteUnstaged(host string, now time.Time) { |
| 2385 | if d.step == nil { |
| 2386 | d.step = map[string]time.Duration{} |
| 2387 | d.notBefore = map[string]time.Time{} |
| 2388 | } |
| 2389 | s := d.step[host] |
| 2390 | if s == 0 { |
| 2391 | s = fastBackoffBase |
| 2392 | } else if s *= 2; s > fastBackoffCap { |
| 2393 | s = fastBackoffCap |
| 2394 | } |
| 2395 | d.step[host] = s |
| 2396 | d.notBefore[host] = now.Add(s) |
| 2397 | } |
| 2398 | |
| 2399 | // proofWaitRenewed reports whether host is on the proof gate's waiting list |
| 2400 | // with a cached miss still running: the sweep just asked the gate about it |
| 2401 | // and the gate held it again. |
| 2402 | func (d *fastDeps) proofWaitRenewed(host string) bool { |
| 2403 | if d.proof.WaitingDue(host) { |
| 2404 | return false |
| 2405 | } |
| 2406 | for _, w := range d.proof.Waits() { |
| 2407 | if w.Host == host { |
| 2408 | return true |
| 2409 | } |
| 2410 | } |
| 2411 | return false |
| 2412 | } |
| 2413 | |
| 2414 | // backoffFor reports host's current backoff step (zero when none recorded). |
| 2415 | func (d *fastDeps) backoffFor(host string) time.Duration { return d.step[host] } |
| 2416 | |
| 2417 | // noteRestartFailed records that host's restart onto ver just failed, and |
| 2418 | // holds the host out of the sweep for a growing window: same version failing |
| 2419 | // again doubles the step (capped), a DIFFERENT version starts the hold fresh |
| 2420 | // at the base — a new promotion is new work, not a continuation of the old |
| 2421 | // failure. Keyed per host with the version recorded, rather than per |
| 2422 | // (host, version) pairs accumulating: one live symlink per host means one |
| 2423 | // version under attempt at a time, and the map stays bounded by the fleet. |
| 2424 | func (d *fastDeps) noteRestartFailed(host, ver string, now time.Time) { |
| 2425 | if d.restartHold == nil { |
| 2426 | d.restartHold = map[string]*restartFailure{} |
| 2427 | } |
| 2428 | m := d.restartHold[host] |
| 2429 | if m == nil || m.ver != ver { |
| 2430 | m = &restartFailure{ver: ver, step: fastRestartBackoffBase} |
| 2431 | } else if m.step *= 2; m.step > fastRestartBackoffCap { |
| 2432 | m.step = fastRestartBackoffCap |
| 2433 | } |
| 2434 | m.notBefore = now.Add(m.step) |
| 2435 | d.restartHold[host] = m |
| 2436 | } |
| 2437 | |
| 2438 | // restartHeld reports whether host is inside its restart-failure hold. |
| 2439 | func (d *fastDeps) restartHeld(host string, now time.Time) bool { |
| 2440 | m := d.restartHold[host] |
| 2441 | return m != nil && now.Before(m.notBefore) |
| 2442 | } |
| 2443 | |
| 2444 | // observeRestart wraps the restart dispatch so this loop SEES its outcomes — |
| 2445 | // the seam the restart-failure hold feeds on. On an error the version under |
| 2446 | // attempt is read from the live symlink (adoptHost repoints BEFORE it |
| 2447 | // consults the callback, so at callback time the symlink names the version |
| 2448 | // being tried; the read failing degrades the memo to host-keyed, which only |
| 2449 | // makes the hold slightly coarser). On success the hold is cleared: the host |
| 2450 | // is converged, or a recovery promote just landed, and either way there is |
| 2451 | // nothing left to hold it out of the sweep for. |
| 2452 | // |
| 2453 | // The wrapper never SUPPRESSES a restart — suppression inside the callback |
| 2454 | // would return nil to adoptHost, which would stamp SetVersions over a |
| 2455 | // process nothing cycled: the exact lie this whole mechanism exists to |
| 2456 | // prevent. The hold gates sweep MEMBERSHIP (fastMembers) instead, so a held |
| 2457 | // host is simply not repointed at this cadence at all. |
| 2458 | // |
| 2459 | // Single-goroutine by construction, like every other fastDeps map: the |
| 2460 | // callback runs inside PassFilter/Pass inside fastTick, on the loop's own |
| 2461 | // goroutine. |
| 2462 | func (d *fastDeps) observeRestart(restart func(string) error) func(string) error { |
| 2463 | return func(host string) error { |
| 2464 | err := restart(host) |
| 2465 | if err != nil { |
| 2466 | ver, _ := release.LiveVersion(d.optDir, host) |
| 2467 | d.noteRestartFailed(host, ver, time.Now()) |
| 2468 | } else { |
| 2469 | delete(d.restartHold, host) |
| 2470 | } |
| 2471 | return err |
| 2472 | } |
| 2473 | } |
| 2474 | |
| 2475 | // convergeAdopted finishes the adoptions this process delegated or never saw. |
| 2476 | // |
| 2477 | // The agent is sandboxed (carlos-edge.service: ProtectSystem=strict, no /opt |
| 2478 | // write), so on a real box a new version is staged, pointed live and stamped |
| 2479 | // by a ROOT pass — carlos-stage.service on this loop's own request, or the |
| 2480 | // hourly carlos-adopt.service. That pass cannot cycle an exec child |
| 2481 | // (rootRestartRoute: the children are this process's), and once it has run, |
| 2482 | // this process's own pass finds the host up to date and never reaches the |
| 2483 | // restart callback RestartHibernating exists to hand it. The warm child was |
| 2484 | // left on the old build until the activator sweep's ConvergeGrace ran out: |
| 2485 | // 110-200s per deploy on the flagship, past `carlos deploy`'s own watch |
| 2486 | // (discussion #231). |
| 2487 | // |
| 2488 | // So, for each warm instance the activator reports drifted, cycle it once the |
| 2489 | // route is STAMPED with the version its symlink now names. The stamp is the |
| 2490 | // adopting pass saying it is finished with the host (adoptHost writes it last, |
| 2491 | // after its restart and any rollback); a drift the stamp does not yet agree |
| 2492 | // with is an adoption still in flight, and this leg waits a tick for it rather |
| 2493 | // than race it. execBounceable keeps the usual line between the two halves: a |
| 2494 | // route whose process the root pass's systemctl arms can reach is never this |
| 2495 | // leg's to cycle. |
| 2496 | // |
| 2497 | // At most limit instances are cycled per call, in host order. One pass can |
| 2498 | // adopt every warm instance of an app at once — a root pass repoints them all |
| 2499 | // inside a few seconds — and stopping them together would drain, checkpoint |
| 2500 | // and later restore them together too: every host of the app blips at the |
| 2501 | // same moment. The activator sweep spends the same care (one convergence |
| 2502 | // cycle per sweep); this leg keeps it at its own cadence, so N instances |
| 2503 | // converge over N ticks rather than in one. Only a cycle that actually |
| 2504 | // started counts against the limit — a drifted host this leg must wait on |
| 2505 | // costs the others nothing. |
| 2506 | // |
| 2507 | // Never wakes anything — ConvergeAdopted is a stop, and a sleeper execs the |
| 2508 | // new build at its own next wake. drifted returning nothing is the ordinary |
| 2509 | // tick and costs no registry read. Returns how many it cycled. |
| 2510 | func convergeAdopted(drifted func() map[string]string, get func(host string) (*registry.Route, error), converge func(host, ver string) bool, limit int) int { |
| 2511 | stale := drifted() |
| 2512 | if len(stale) == 0 { |
| 2513 | return 0 |
| 2514 | } |
| 2515 | hosts := make([]string, 0, len(stale)) |
| 2516 | for host := range stale { |
| 2517 | hosts = append(hosts, host) |
| 2518 | } |
| 2519 | sort.Strings(hosts) |
| 2520 | n := 0 |
| 2521 | for _, host := range hosts { |
| 2522 | if n >= limit { |
| 2523 | break |
| 2524 | } |
| 2525 | live := stale[host] |
| 2526 | rt, err := get(host) |
| 2527 | if err != nil || rt == nil || rt.Version != live || !execBounceable(*rt) { |
| 2528 | continue |
| 2529 | } |
| 2530 | if converge(host, live) { |
| 2531 | n++ |
| 2532 | log.Printf("agent: %s was adopted onto %s by another pass — cycled its warm instance onto it", host, live) |
| 2533 | } |
| 2534 | } |
| 2535 | return n |
| 2536 | } |
| 2537 | |
| 2538 | // adoptedCyclesPerTick is convergeAdopted's limit per run of its loop: one |
| 2539 | // warm instance every adoptedConvergeEvery. A single-instance app — the common |
| 2540 | // case — converges on the first run after its adoption is stamped; an app with |
| 2541 | // a dozen warm instances takes under half a minute, one blip at a time. |
| 2542 | const adoptedCyclesPerTick = 1 |
| 2543 | |
| 2544 | // adoptedConvergeEvery is how often the converge leg looks for a finished |
| 2545 | // adoption to act on. |
| 2546 | const adoptedConvergeEvery = 2 * time.Second |
| 2547 | |
| 2548 | // convergeLegFor builds the converge leg over the box's activator. nil for a |
| 2549 | // box with none armed: there are no exec children to cycle. On a box that |
| 2550 | // declined CARLOS_SWAP=overlap but honours failure records (heldHolds), a |
| 2551 | // host whose installed build a record holds for it is passed by: cycling it |
| 2552 | // would stop it into a wake its Planner refuses (final review F6). |
| 2553 | func convergeLegFor(svc *activator.Service, reg *registry.Registry) func() { |
| 2554 | if svc == nil { |
| 2555 | return nil |
| 2556 | } |
| 2557 | get := reg.Get |
| 2558 | if h := heldHolds(); h != nil { |
| 2559 | get = h.skipGet(reg.Get) |
| 2560 | } |
| 2561 | return func() { |
| 2562 | convergeAdopted(svc.Drifted, get, svc.ConvergeAdopted, adoptedCyclesPerTick) |
| 2563 | } |
| 2564 | } |
| 2565 | |
| 2566 | // runConvergeLoop runs the converge leg every `every` until stop closes (nil: |
| 2567 | // for the life of the process). |
| 2568 | // |
| 2569 | // Its own loop, not a leg of fastTick, and the reason is measured: on the |
| 2570 | // flagship's Ireland box a fast tick ran 25 to 50 seconds whenever the serial |
| 2571 | // moved (a full reconcile and adopt pass over ~180 routes), and console |
| 2572 | // activity across every tenant moves it about once a minute. Riding the tick, |
| 2573 | // a stamped adoption waited ~48s for its cycle (discussion #231, 2026-10-05). |
| 2574 | // The leg needs none of the tick's state: it reads the activator and the |
| 2575 | // registry, both safe to read beside the tick, and its one write is a stop |
| 2576 | // under the activator's own lock — where the tick's adopt bounce is |
| 2577 | // version-aware (BounceAdopted), so the two cannot stop one child twice for |
| 2578 | // one build. |
| 2579 | func runConvergeLoop(leg func(), every time.Duration, stop <-chan struct{}) { |
| 2580 | t := time.NewTicker(every) |
| 2581 | defer t.Stop() |
| 2582 | for { |
| 2583 | select { |
| 2584 | case <-stop: |
| 2585 | return |
| 2586 | case <-t.C: |
| 2587 | leg() |
| 2588 | } |
| 2589 | } |
| 2590 | } |
| 2591 | |
| 2592 | // slowLegAfter is when logSlow speaks: a fast-loop leg that takes longer than |
| 2593 | // this is worth a line, because the loop's promise is seconds. |
| 2594 | const slowLegAfter = 5 * time.Second |
| 2595 | |
| 2596 | // logSlow logs that what took longer than slowLegAfter, with detail. Meant |
| 2597 | // for a defer: the loop's latency was invisible in the journal until someone |
| 2598 | // reconstructed it from timestamps by hand (#231). |
| 2599 | func logSlow(what string, start time.Time, detail string) { |
| 2600 | if d := time.Since(start); d > slowLegAfter { |
| 2601 | if detail != "" { |
| 2602 | detail = " (" + detail + ")" |
| 2603 | } |
| 2604 | log.Printf("agent: fast converge: %s took %s%s", what, d.Round(100*time.Millisecond), detail) |
| 2605 | } |
| 2606 | } |
| 2607 | |
| 2608 | // fastTick is one fast-converge tick — two independent jobs (spec §3-§4): |
| 2609 | // |
| 2610 | // 1. Serial check: one ReadSerial. The FIRST successful read only seeds |
| 2611 | // lastSerial — a present object at its value (boot must not replay |
| 2612 | // history; adoptLoop's 1-minute first pass covers boot), an absent one |
| 2613 | // at zero (absence is not history, and the first-ever bump on a fresh |
| 2614 | // bucket must fire, not read as replay). Afterward any CHANGE |
| 2615 | // (not just increments — a reset must not wedge the box) runs the full |
| 2616 | // pass set, clamped to one per fastFullClamp: a bump inside the window |
| 2617 | // is deferred by NOT recording the acted serial, so the next tick past |
| 2618 | // the window re-fires. Read errors are tolerated silently (one log |
| 2619 | // line per state change, never per tick). |
| 2620 | // The restart leg (issue #113) rides that same move, AFTER the pass set: |
| 2621 | // a bump can carry a promote and a restart request at once, and cycling |
| 2622 | // processes before their new version and config land would cycle them onto |
| 2623 | // the old ones. It is a separate call rather than a fourth leg inside |
| 2624 | // fullPass because it is not release convergence — it restarts what is |
| 2625 | // already correct — and because it must still run on a box with no |
| 2626 | // CARLOS_SECRETS_KEY, which fullPass's config leg silently skips. The |
| 2627 | // sidecar wake leg (design: sidecar members) rides the same move |
| 2628 | // immediately after, for the identical reason: a bump can carry a |
| 2629 | // provisioning change and a wake request together, and it must run on |
| 2630 | // a box with no CARLOS_SECRETS_KEY too. |
| 2631 | // The schedule converge (schedule spec §6.2) is a further leg on that same |
| 2632 | // move, for both of the same reasons; see the call site. The layout |
| 2633 | // leg (isolated-exec design §8.2) follows it and only starts a pass in |
| 2634 | // the background; the system leg's observe comes last. |
| 2635 | // 2. Fast sweep: fastMembers, and only when non-empty one PassFilter over |
| 2636 | // exactly those hosts — an empty sweep costs zero store reads. After |
| 2637 | // the sweep, members still unstaged back off, and so do members whose |
| 2638 | // proof wait the sweep left standing; other staged members clear their |
| 2639 | // backoff state. |
| 2640 | func fastTick(ctx context.Context, d *fastDeps, now time.Time) { |
| 2641 | // The index is observed BEFORE any pass this tick runs, so a pass covers |
| 2642 | // at least what was observed and indexActed can take it afterwards. |
| 2643 | obs, indexOK := d.observeIndex(ctx) |
| 2644 | fullRan := false |
| 2645 | doc, _, err := release.ReadSerial(ctx, d.store) |
| 2646 | switch { |
| 2647 | case err != nil: |
| 2648 | if !d.serialBroken { |
| 2649 | log.Printf("agent: fast converge serial read: %v", err) |
| 2650 | d.serialBroken = true |
| 2651 | } |
| 2652 | case !d.seeded: |
| 2653 | // The first successful read seeds — the ABSENT object included, |
| 2654 | // which ReadSerial reports as the zero doc. Absence is not history: |
| 2655 | // there is nothing to replay, and every deployment's bucket starts |
| 2656 | // without the object, so seeding zero is what lets the first-ever |
| 2657 | // console bump (0 → 1) fire instead of being swallowed as boot |
| 2658 | // noise. A PRESENT object seeds at its current value, and THAT is |
| 2659 | // the no-replay rule: the bumps behind it are history a booting box |
| 2660 | // must not act on (adoptLoop's 1-minute first pass covers boot). |
| 2661 | d.lastSerial, d.seeded = doc.Serial, true |
| 2662 | d.serialBroken = false |
| 2663 | default: |
| 2664 | d.serialBroken = false |
| 2665 | if doc.Serial != d.lastSerial && now.Sub(d.lastFull) >= fastFullClamp { |
| 2666 | d.runFull(ctx) |
| 2667 | fullRan = true |
| 2668 | if d.restartLeg != nil { |
| 2669 | d.restartLeg(ctx) |
| 2670 | } |
| 2671 | if d.wakeLeg != nil { |
| 2672 | d.wakeLeg(ctx) |
| 2673 | } |
| 2674 | // The schedule converge rides the same move, after the pass set for |
| 2675 | // the same reason the restart leg does: a bump can carry a promote |
| 2676 | // and a `carlos schedule set` at once, and converging schedules |
| 2677 | // before the version they belong to has staged would write rows |
| 2678 | // against the outgoing build. Also a separate leg rather than a |
| 2679 | // member of fullPass — it is not release convergence, and it must |
| 2680 | // still run on a box with no CARLOS_SECRETS_KEY. |
| 2681 | if d.scheduleLeg != nil { |
| 2682 | d.scheduleLeg(ctx) |
| 2683 | } |
| 2684 | // The layout pass converges to the fleet's recorded layout on |
| 2685 | // every move, after the pass set has minted and staged whatever |
| 2686 | // the move carried, so a freshly minted row is seen. It only |
| 2687 | // starts the pass: the pass runs on its own goroutine, because |
| 2688 | // it can wait on stops, and nothing above or below may wait on |
| 2689 | // it. |
| 2690 | if d.layoutLeg != nil { |
| 2691 | d.layoutLeg(ctx) |
| 2692 | } |
| 2693 | if d.serialNotify != nil { |
| 2694 | select { |
| 2695 | case d.serialNotify <- struct{}{}: |
| 2696 | default: |
| 2697 | } |
| 2698 | } |
| 2699 | // The platform self-update leg rides the move too, last: a |
| 2700 | // `carlos system update` is a console action like any other, and |
| 2701 | // this is what makes a promote noticed in seconds instead of at |
| 2702 | // the leg's own hourly floor. It only READS the pointer and sets |
| 2703 | // a deadline — the fetch and the request happen in the tick |
| 2704 | // below, after the spread, so a serial move costs one Get. |
| 2705 | if d.systemLeg != nil { |
| 2706 | d.systemLeg.observe(ctx) |
| 2707 | } |
| 2708 | d.lastSerial = doc.Serial |
| 2709 | d.lastFull = now |
| 2710 | } |
| 2711 | } |
| 2712 | |
| 2713 | // The backstop's full pass (spec §4.4), on this goroutine: requested |
| 2714 | // after each rebuild, run once, under the same clamp. |
| 2715 | if !fullRan && d.fullDue != nil && d.fullDue.Load() && now.Sub(d.lastFull) >= fastFullClamp { |
| 2716 | d.runFull(ctx) |
| 2717 | fullRan = true |
| 2718 | d.lastFull = now |
| 2719 | } |
| 2720 | // An index-only change runs the scoped pass (spec §4.5). |
| 2721 | d.indexTick(ctx, now, obs, indexOK, fullRan) |
| 2722 | |
| 2723 | // The self-update tick, outside the serial switch and before the sweep's |
| 2724 | // early return: it is what runs a spread deadline down and re-touches an |
| 2725 | // unanswered request, and neither has anything to do with routes. It |
| 2726 | // reads no store unless a deadline has actually come due. |
| 2727 | if d.systemLeg != nil { |
| 2728 | d.systemLeg.tick(ctx, now) |
| 2729 | } |
| 2730 | |
| 2731 | // D3's mint asker: on a proving box, ask for a root pass every 2 minutes |
| 2732 | // while a listed build has baked with no proof. Elsewhere it reads one |
| 2733 | // absent file and returns. |
| 2734 | if d.mintLeg != nil { |
| 2735 | d.mintLeg(ctx) |
| 2736 | } |
| 2737 | |
| 2738 | members := d.fastMembers(now) |
| 2739 | if len(members) == 0 { |
| 2740 | return // zero store reads: no PassFilter, no bake-clock write |
| 2741 | } |
| 2742 | defer logSlow("fast sweep", time.Now(), fmt.Sprintf("%d routes", len(members))) |
| 2743 | member := map[string]bool{} |
| 2744 | for _, h := range members { |
| 2745 | member[h] = true |
| 2746 | } |
| 2747 | opts := d.passOpts |
| 2748 | // The canary gate only on a tick whose own index observation succeeded: |
| 2749 | // a tick that could not read the index reads every pointer, as before |
| 2750 | // the index. (A failed poll also leaves the watch unknown, so the gate |
| 2751 | // would find every group due anyway; this keeps it off whatever another |
| 2752 | // caller's poll did in between.) |
| 2753 | if d.canaryGate != nil && indexOK { |
| 2754 | opts = append(slices.Clip(opts), release.WithPointerGate(d.canaryPointerGate(now))) |
| 2755 | } |
| 2756 | if err := release.PassFilter(ctx, d.reg, d.store, d.pub, d.optDir, d.restart, |
| 2757 | func(rt registry.Route) bool { return member[rt.Host] }, opts...); err != nil { |
| 2758 | log.Printf("agent: fast converge sweep: %v", err) |
| 2759 | } |
| 2760 | for _, h := range members { |
| 2761 | switch v, err := release.LiveVersion(d.optDir, h); { |
| 2762 | case err != nil || v == "": |
| 2763 | d.noteUnstaged(h, now) |
| 2764 | // An unstaged route whose wait this sweep renewed is waiting |
| 2765 | // for its proof, not failing to stage: spec §4.5 has it rejoin |
| 2766 | // the sweep each time its 30 s miss runs out, so its backoff |
| 2767 | // stops doubling there. |
| 2768 | if d.proofWaitRenewed(h) && d.step[h] > release.ProofMissTTL { |
| 2769 | d.step[h] = release.ProofMissTTL |
| 2770 | d.notBefore[h] = now.Add(release.ProofMissTTL) |
| 2771 | } |
| 2772 | case d.proof.WaitingDue(h): |
| 2773 | // The sweep neither renewed this host's proof wait (which would |
| 2774 | // cache a fresh 30 s miss) nor cleared it: adoptHost returned |
| 2775 | // before the gate, say on a channel with no pointer. Nothing will |
| 2776 | // change that by the next tick, so it backs off like an unstaged |
| 2777 | // route instead of being swept every 2 s (plan 3 ruling 5). |
| 2778 | d.noteUnstaged(h, now) |
| 2779 | default: |
| 2780 | delete(d.notBefore, h) |
| 2781 | delete(d.step, h) |
| 2782 | } |
| 2783 | } |
| 2784 | } |
| 2785 | |
| 2786 | // canaryPointerGate is the sweep's release.PointerGate at now. Only canary |
| 2787 | // groups are gated; every other group is read as before. A canary group is |
| 2788 | // skipped when its app scope has not moved since the group's own last good |
| 2789 | // read (another channel's read of the same app does not count), that read is |
| 2790 | // younger than the backstop, and every route of the group is live on the |
| 2791 | // head it returned: then a read could only repeat what the box already |
| 2792 | // runs. Anything unsettled (an unstaged route, a failed restart rolled back, |
| 2793 | // a proof wait, a build the update settings chose over the head) keeps |
| 2794 | // today's read on every tick. |
| 2795 | func (d *fastDeps) canaryPointerGate(now time.Time) release.PointerGate { |
| 2796 | return func(acct, app, channel string, routes []registry.Route) (bool, func(*release.ChannelPointer, error)) { |
| 2797 | if !strings.HasPrefix(channel, "canary/") { |
| 2798 | return false, nil |
| 2799 | } |
| 2800 | scope, gated := release.ChangeScope(release.ChannelKey(acct, app, channel)) |
| 2801 | if !gated { |
| 2802 | return false, nil |
| 2803 | } |
| 2804 | k := [3]string{acct, app, channel} |
| 2805 | gen, due := d.canaryGate.Due(scope, now) |
| 2806 | if !due && d.canarySettled(k, routes, gen, now) { |
| 2807 | return true, nil |
| 2808 | } |
| 2809 | return false, func(ptr *release.ChannelPointer, err error) { |
| 2810 | if d.canaryHeads == nil { |
| 2811 | d.canaryHeads = map[[3]string]canaryRead{} |
| 2812 | } |
| 2813 | if err != nil || ptr == nil { |
| 2814 | delete(d.canaryHeads, k) |
| 2815 | d.canaryGate.Mark(scope, gen, now, false) |
| 2816 | return |
| 2817 | } |
| 2818 | d.canaryHeads[k] = canaryRead{head: ptr.Version, gen: gen, at: now} |
| 2819 | d.canaryGate.Mark(scope, gen, now, true) |
| 2820 | } |
| 2821 | } |
| 2822 | } |
| 2823 | |
| 2824 | // canarySettled reports whether group k's own last good read is at gen (the |
| 2825 | // generation Due returned) and younger than the gate's maximum age, and every |
| 2826 | // route of the group is live on the head that read returned. |
| 2827 | func (d *fastDeps) canarySettled(k [3]string, routes []registry.Route, gen uint64, now time.Time) bool { |
| 2828 | r, ok := d.canaryHeads[k] |
| 2829 | if !ok || r.head == "" || len(routes) == 0 || r.gen != gen || now.Sub(r.at) >= d.canaryGate.MaxAge() { |
| 2830 | return false |
| 2831 | } |
| 2832 | for _, rt := range routes { |
| 2833 | if v, err := release.LiveVersion(d.optDir, rt.Host); err != nil || v != r.head { |
| 2834 | return false |
| 2835 | } |
| 2836 | } |
| 2837 | return true |
| 2838 | } |
| 2839 | |
| 2840 | // runFull runs the full pass set's fullPass, taking any backstop request |
| 2841 | // first: a full pass that starts after the backstop's rebuild satisfies it. |
| 2842 | func (d *fastDeps) runFull(ctx context.Context) { |
| 2843 | if d.fullDue != nil { |
| 2844 | d.fullDue.Store(false) |
| 2845 | } |
| 2846 | d.fullPass(ctx) |
| 2847 | } |
| 2848 | |
| 2849 | // observeIndex polls the change index and returns its observed generations; |
| 2850 | // ok is false when no watch is armed or the index is unusable (then the |
| 2851 | // serial and the floors carry the tick, as before the index). A failure is |
| 2852 | // logged once per state change, never per tick. |
| 2853 | func (d *fastDeps) observeIndex(ctx context.Context) (map[string]uint64, bool) { |
| 2854 | if d.watch == nil { |
| 2855 | return nil, false |
| 2856 | } |
| 2857 | // Poll's moved list is advisory (shared watch): only Observed counts. |
| 2858 | // Bounded: the tick runs on a background context, and a hung index GET |
| 2859 | // would otherwise stall the tick and hold the watch's poll lock that the |
| 2860 | // edge's loops share. Running out is "unknown", like any other error. |
| 2861 | pctx, cancel := context.WithTimeout(ctx, release.ChangeIndexPollTimeout) |
| 2862 | _, err := d.watch.Poll(pctx) |
| 2863 | cancel() |
| 2864 | if err != nil { |
| 2865 | if !d.indexBroken { |
| 2866 | log.Printf("agent: fast converge change index: %v (scoped passes wait; the serial and the floors carry on)", err) |
| 2867 | d.indexBroken = true |
| 2868 | } |
| 2869 | return nil, false |
| 2870 | } |
| 2871 | obs, ok := d.watch.Observed() |
| 2872 | if ok && d.indexBroken { |
| 2873 | log.Printf("agent: fast converge change index readable again") |
| 2874 | d.indexBroken = false |
| 2875 | } |
| 2876 | return obs, ok |
| 2877 | } |
| 2878 | |
| 2879 | // indexTick acts on an index observation: the first one seeds indexActed |
| 2880 | // without a pass (boot must not replay history; the backstop's first full |
| 2881 | // pass and the loops' first passes cover boot), a full pass this tick |
| 2882 | // covers everything observed, and otherwise app and account scopes that |
| 2883 | // moved since indexActed run one scoped pass, clamped like the full pass: |
| 2884 | // a change inside the clamp leaves indexActed as it was, so the next tick |
| 2885 | // past the window acts on it. indexActed moves on only after a pass ran. |
| 2886 | func (d *fastDeps) indexTick(ctx context.Context, now time.Time, obs map[string]uint64, ok, fullRan bool) { |
| 2887 | if !ok { |
| 2888 | return |
| 2889 | } |
| 2890 | if d.indexActed == nil || fullRan { |
| 2891 | d.indexActed = obs |
| 2892 | return |
| 2893 | } |
| 2894 | moved := indexScopes(d.indexActed, obs) |
| 2895 | if len(moved) == 0 { |
| 2896 | d.indexActed = obs // control/ scopes only: not this loop's business |
| 2897 | return |
| 2898 | } |
| 2899 | if now.Sub(d.lastFull) < fastFullClamp || d.scopedPass == nil { |
| 2900 | return |
| 2901 | } |
| 2902 | d.scopedPass(ctx, moved) |
| 2903 | d.indexActed = obs |
| 2904 | d.lastFull = now |
| 2905 | } |
| 2906 | |
| 2907 | // fastConvergeLoop replaces the canary-only fast adopt loop with the |
| 2908 | // carlos-deploy fast set (spec §3-§4): a 2s serial poll that turns a console |
| 2909 | // bump (promote, instance create, config write) into the ordinary verified |
| 2910 | // pass set — reconcile → adopt → exec config sync — within seconds, plus a |
| 2911 | // fast sweep of the routes with genuinely outstanding work (canary/* always; |
| 2912 | // unstaged routes until they drain, with per-host backoff). Everything it |
| 2913 | // triggers is the same signature-verified, bake-gated work the hourly loops |
| 2914 | // run; the serial carries no authority, only earliness. |
| 2915 | // |
| 2916 | // Setup mirrors adoptLoop's: missing deployment config or pubkey disables |
| 2917 | // the loop silently (adoptLoop already logged the reason). The config-sync |
| 2918 | // leg degrades exactly as configSyncLoop does — no CARLOS_SECRETS_KEY, or a |
| 2919 | // broken one, and the full pass set simply omits that leg (configSyncLoop |
| 2920 | // already logged that too). Takes the activator for the same reason |
| 2921 | // adoptLoop does: a fresh version on a warm exec instance must bounce onto |
| 2922 | // its new binary, not wait for hibernation. Nil is fine. |
| 2923 | // |
| 2924 | // Thin and untested by convention, like adoptLoop/statusLoop — the tick's |
| 2925 | // whole decision lives in fastTick/fastMembers, which are tested. |
| 2926 | func fastConvergeLoop(reg *registry.Registry, svc *activator.Service, sd *scheduleDeps, writes hostwrite.Config, serialNotify chan<- struct{}, bs *buildStates, configure ...func(*provision.ReconcileConfig)) { |
| 2927 | ctx := context.Background() |
| 2928 | store, err := openDeployStore(ctx) |
| 2929 | if err != nil { |
| 2930 | declinedHoldsStay("no deploy store") |
| 2931 | return // adoptLoop already logged this |
| 2932 | } |
| 2933 | pub, err := release.ParsePublicKey(os.Getenv("CARLOS_RELEASE_PUBKEY"), "CARLOS_RELEASE_PUBKEY") |
| 2934 | if err != nil { |
| 2935 | declinedHoldsStay("no release key") |
| 2936 | return // ditto |
| 2937 | } |
| 2938 | var priv *ecdh.PrivateKey |
| 2939 | if raw := os.Getenv("CARLOS_SECRETS_KEY"); raw != "" { |
| 2940 | if p, err := appconfig.ParsePrivateKey(raw, "CARLOS_SECRETS_KEY"); err == nil { |
| 2941 | priv = p |
| 2942 | } |
| 2943 | } |
| 2944 | dd := dataDir() |
| 2945 | optDir := release.OptDir() |
| 2946 | // d is built empty first so the restart dispatch can be wrapped in its |
| 2947 | // observer (observeRestart): a failed restart holds its host out of the |
| 2948 | // sweep instead of the sweep re-cycling the service every tick. The |
| 2949 | // remaining fields are filled in right below. |
| 2950 | d := &fastDeps{reg: reg, store: store, pub: pub, optDir: optDir, serialNotify: serialNotify} |
| 2951 | restart := d.observeRestart(makeRestartRoute(reg, systemctlTryRestart, agentCheckedUnitRestart(systemdUnits{}, optDir, os.Getenv, nil), adoptBounceFn(svc))) |
| 2952 | // Built once, used by both pass shapes below — see fastPassOpts for the |
| 2953 | // DO-NOT list and why the stage requester is per-loop, not per-tick. |
| 2954 | opts := append(fastPassOpts(dd, stageRequester(dd), restart, makeStartRoute(reg, systemctlStartUnit)), release.WithHostWrites(writes)) |
| 2955 | // The rollback's restart (D1): checked like the forward one, and a replace |
| 2956 | // rather than a start, so a rejected build left up is cycled out. |
| 2957 | opts = append(opts, release.WithRollbackRestart(agentRollbackUnitRestart(reg, systemdUnits{}, optDir, os.Getenv))) |
| 2958 | // D3: the process's one proof gate rides every pass this loop runs, and |
| 2959 | // fastMembers reads its waiting list. The mint asker has its own |
| 2960 | // requester, so its 2-minute asks never spend the stage budget and its |
| 2961 | // journal line does not claim a stage. |
| 2962 | opts = append(opts, release.WithProofGate(agentProofGate())) |
| 2963 | d.proof = agentProofGate() |
| 2964 | d.mintLeg = (&mintAsker{reg: reg, store: store, sys: systemdUnits{}, optDir: optDir, |
| 2965 | policyPath: release.ProofPolicyPath, request: mintRequester(dd), now: time.Now}).tick |
| 2966 | // The restart check's memory (issue #113, narrowed requests added by |
| 2967 | // discussion #16), seeded ONCE here — two reads per app at arming (the |
| 2968 | // app-wide generation and the narrowed request log) and never again except |
| 2969 | // on a serial move, which is what keeps an idle box at exactly one serial |
| 2970 | // GET per tick. See restartSeed for why the baseline cannot wait for the |
| 2971 | // first request to establish itself. |
| 2972 | restartMem := newRestartMemory() |
| 2973 | // `carlos instances retry` (zero-downtime spec §8.4) reaches the rollout |
| 2974 | // manager through the host-retry log (a third read per app at arming, |
| 2975 | // and on a move, only where retryHost is set; hostRetryPass) or a legacy |
| 2976 | // entry in the narrowed-restart log, or, on a box that declined |
| 2977 | // the flag but honours its failure book, that book (Task 10); nil |
| 2978 | // otherwise. Each entry's grant is verified against the root key first |
| 2979 | // (spec §11). |
| 2980 | restartMem.pub = pub |
| 2981 | restartMem.retryHost = rolloutRetryHost() |
| 2982 | // The declined box's pointer pass (Task 10, ruling 194): `carlos promote |
| 2983 | // --retry` clears a record, a signed rollback a marker, read through |
| 2984 | // the same verified pointer read the armed scheduler uses, at most |
| 2985 | // every rollout.PointerReadEvery. |
| 2986 | if h := heldHolds(); h != nil { |
| 2987 | go h.pointerLoop(ctx, func(ctx context.Context, acct, app, channel string) (*release.ChannelPointer, error) { |
| 2988 | return release.Channel(ctx, store, pub, acct, app, channel) |
| 2989 | }) |
| 2990 | } |
| 2991 | restartSeed(ctx, reg, store, dd, restartMem) |
| 2992 | restartBounce := bounceFn(svc) |
| 2993 | restartWake := wakeFn(svc) |
| 2994 | // The sidecar wake leg's own memory, seeded the same way and for the |
| 2995 | // same reason: one silent read per served sidecar at arming so the |
| 2996 | // FIRST wake request after a boot is acted on rather than swallowed as |
| 2997 | // the baseline (sidecarWakeSeed's own doc comment). |
| 2998 | sidecarWakeMem := newSidecarWakeMemory(writes.Enabled) |
| 2999 | sidecarWakeSeed(ctx, reg, store, sidecarWakeMem) |
| 3000 | sidecarWake := wakeFn(svc) |
| 3001 | // The moves hold view, built before the converge below: that converge |
| 3002 | // runs before the first moves pass, and an unbuilt view freezes every |
| 3003 | // host (moves spec §9.4). Nothing with the Phase 1 switch off. |
| 3004 | refreshMovesHolds(ctx, store) |
| 3005 | // The offline keep view, likewise before any reconcile pass: once it is |
| 3006 | // read, keep answers false for every host with no offline move, so the |
| 3007 | // first pass removes what it always has (movesoffline.go). |
| 3008 | refreshOfflineView(ctx, store, armedOfflineView()) |
| 3009 | // carlos run's boot pass (run spec §5.2, index item 2): reconcile's |
| 3010 | // first pass is a minute away and the CLI expires a run at 60 s. |
| 3011 | runRunsBootPass(ctx) |
| 3012 | // The schedule converge's own seed, on restartSeed's precedent and for the |
| 3013 | // same shape of reason: the leg below runs only on a serial MOVE, so |
| 3014 | // without one pass here a box that booted after the bump would carry no |
| 3015 | // timer rows (and report no status) until the hourly floor in adoptLoop — |
| 3016 | // up to an hour of a schedule silently not running. It also seeds the |
| 3017 | // converger's etag memory, which is what makes a run signal older than this |
| 3018 | // process read as history rather than as a fresh request to run. |
| 3019 | if err := sd.converge(ctx); err != nil { |
| 3020 | log.Printf("agent: schedule converge (start): %v", err) |
| 3021 | } |
| 3022 | d.restart = restart |
| 3023 | d.passOpts = opts |
| 3024 | // The scheduler's pass with CARLOS_SWAP=overlap armed, nothing while |
| 3025 | // managed starts are refused, today's convergeAdopted otherwise. |
| 3026 | if leg := fastConvergeLeg(svc, reg); leg != nil { |
| 3027 | go runConvergeLoop(leg, adoptedConvergeEvery, nil) |
| 3028 | } |
| 3029 | // The serial-triggered full pass set, in spec order: mint routes |
| 3030 | // (reconcile), then reconcile aliases, then stage them (adopt), then deliver |
| 3031 | // their config. Each error is one log line, never fatal — the hourly floors |
| 3032 | // stand underneath every leg. Not membership-gated by the restart hold: |
| 3033 | // a serial bump is an explicit console action, bounded by whoever keeps |
| 3034 | // making it — and it is what lets a recovery promote through the console |
| 3035 | // converge immediately even mid-hold. |
| 3036 | legs := &passLegs{store: store, reg: reg, pub: pub, optDir: optDir, restart: restart, opts: opts, |
| 3037 | priv: priv, dd: dd, svc: svc, writes: writes, bs: bs, |
| 3038 | cfg: func() provision.ReconcileConfig { |
| 3039 | return configuredClaimProvision(provisionTickConfig(os.Getenv), configure) |
| 3040 | }} |
| 3041 | d.fullPass = legs.full |
| 3042 | // The change index (S3 change-index design §4.5): an index-only change |
| 3043 | // runs the scoped pass, and the backstop's request runs the full one, |
| 3044 | // both on this loop's goroutine. Nil watch: the tick is as before. |
| 3045 | if watch := armedChangeWatch.Load(); watch != nil { |
| 3046 | d.watch = watch |
| 3047 | d.scopedPass = legs.scoped |
| 3048 | d.fullDue = &backstopFullDue |
| 3049 | // The sweep's canary pointer reads are gated on the same watch |
| 3050 | // (spec §5). Built here, after RunAgent armed the index. |
| 3051 | d.canaryGate = newChangeGate() |
| 3052 | } |
| 3053 | // The restart leg: this process's own children bounced here, everything |
| 3054 | // unit-shaped relayed to a root pass. Deliberately NOT inside fullPass |
| 3055 | // above — see fastTick's comment for both reasons. |
| 3056 | d.restartLeg = func(ctx context.Context) { |
| 3057 | routes, err := reg.List() |
| 3058 | if err != nil { |
| 3059 | log.Printf("agent: restart check: list routes: %v", err) |
| 3060 | return |
| 3061 | } |
| 3062 | restartRequestLeg(ctx, routes, store, dd, restartBounce, restartWake, restartMem) |
| 3063 | } |
| 3064 | // The sidecar wake leg: apps/<acct>/<app>/sidecar-wake/<name>.json |
| 3065 | // (release.SidecarWakeKey) compared and dispatched here, riding the |
| 3066 | // same serial move as the restart leg immediately above it — see |
| 3067 | // fastTick's comment for why both must run after fullPass and on a |
| 3068 | // box with no CARLOS_SECRETS_KEY. |
| 3069 | d.wakeLeg = func(ctx context.Context) { |
| 3070 | routes, err := reg.List() |
| 3071 | if err != nil { |
| 3072 | log.Printf("agent: sidecar wake check: list routes: %v", err) |
| 3073 | return |
| 3074 | } |
| 3075 | sidecarWakeLeg(ctx, routes, store, sidecarWakeMem, sidecarWake) |
| 3076 | } |
| 3077 | // The schedule leg: converge this box's timer rows from the apps' |
| 3078 | // declarations. Nil when the box has no scheduled work armed, which |
| 3079 | // fastTick tolerates. |
| 3080 | d.scheduleLeg = scheduleLegFor(sd) |
| 3081 | // The layout leg: this box's label is the one its status publishes |
| 3082 | // under, which is the label an operator scopes a layout request to. |
| 3083 | layoutBox := metricsBoxLabel(os.Getenv, readMachineID()) |
| 3084 | layoutPasses := newLayoutLeg(ctx, func(ctx context.Context) bool { return runLayoutLeg(ctx, store, layoutBox) }) |
| 3085 | d.layoutLeg = layoutPasses.kick |
| 3086 | // The platform self-update leg (spec §4): the sandboxed half, which only |
| 3087 | // ever fetches into /data and writes a request. It lives here rather than |
| 3088 | // in a loop of its own because the spread deadline needs a seconds |
| 3089 | // cadence to run down on, and this is the loop that already has one. |
| 3090 | d.systemLeg = newSystemLeg(store, pub, os.Getenv, dd, optDir) |
| 3091 | // One observe at arming, on restartSeed's precedent: the leg below only |
| 3092 | // re-observes on a serial move or at its hourly floor, so a box that |
| 3093 | // (re)starts under an already-promoted pointer would otherwise sit on the |
| 3094 | // old build for up to an hour with nobody to bump the serial. observe |
| 3095 | // handles "no pointer" itself, which is every deployment before its first |
| 3096 | // `carlos system update`. |
| 3097 | d.systemLeg.observe(ctx) |
| 3098 | log.Printf("agent: fast converge loop armed (2s + jitter: serial poll + canary/* + unstaged routes + system pointer)") |
| 3099 | for { |
| 3100 | time.Sleep(fastTickEvery + rand.N(time.Second)) |
| 3101 | fastTick(ctx, d, time.Now()) |
| 3102 | } |
| 3103 | } |
| 3104 | |
| 3105 | // isCanaryRoute reports whether rt follows a canary/* channel — the same |
| 3106 | // two-segment convention bake.go's DefaultHolds wildcard matches (a bare |
| 3107 | // "canary" with no slash is NOT covered, same as bake.go: it falls through |
| 3108 | // to the unknown-channel default there, so it shouldn't fast-poll here |
| 3109 | // either). |
| 3110 | func isCanaryRoute(rt registry.Route) bool { |
| 3111 | return strings.HasPrefix(rt.Channel, "canary/") |
| 3112 | } |
| 3113 | |
| 3114 | // countExpiredCanaries counts routes (this box's own, platform aliases |
| 3115 | // excluded — same convention as CollectBox) on a canary/* channel whose |
| 3116 | // pointer has passed release.CanaryTTL, for the fleet status page's §5 |
| 3117 | // count (internal/status Probes.ExpiredCanaries). A canary route whose |
| 3118 | // pointer can't be resolved at all — store unavailable, pubkey unusable, or |
| 3119 | // the specific channel won't verify — nils the WHOLE count rather than |
| 3120 | // undercounting: "we couldn't check" and "there are none" are different |
| 3121 | // claims, and this function only ever makes the second one when it's sure. |
| 3122 | // No canary routes at all is a confident, known 0 — it needs no store to |
| 3123 | // answer "none" about routes that don't exist. |
| 3124 | // |
| 3125 | // Pointers are resolved at most once per distinct (account, app, channel) |
| 3126 | // among the routes given (the local `cache` map below), mirroring adopt.go's |
| 3127 | // Pass — this function is called once per 5-minute status tick |
| 3128 | // (statusLoop), so N routes on this box following the same canary channel |
| 3129 | // cost exactly ONE release.Channel() store read for that tick, not N |
| 3130 | // (2026-08-03 review: confirmed and pinned by |
| 3131 | // TestCountExpiredCanariesReadsEachChannelOnce). The cache is local to one |
| 3132 | // call/one tick — nothing persists across ticks, which is fine at a |
| 3133 | // 5-minute cadence. |
| 3134 | func countExpiredCanaries(ctx context.Context, routes []registry.Route, store release.Store, pub ed25519.PublicKey, now time.Time) *int { |
| 3135 | type key struct{ acct, app, channel string } |
| 3136 | cache := map[key]*release.ChannelPointer{} |
| 3137 | n := 0 |
| 3138 | for _, rt := range routes { |
| 3139 | if rt.Platform || !isCanaryRoute(rt) { |
| 3140 | continue |
| 3141 | } |
| 3142 | if store == nil || len(pub) != ed25519.PublicKeySize { |
| 3143 | return nil // a canary route exists but nothing here can resolve its pointer |
| 3144 | } |
| 3145 | k := key{release.RouteAccount(rt), rt.App, rt.Channel} |
| 3146 | ptr, cached := cache[k] |
| 3147 | if !cached { |
| 3148 | p, err := release.Channel(ctx, store, pub, k.acct, k.app, k.channel) |
| 3149 | if err != nil { |
| 3150 | return nil // this route's own pointer won't verify — unknown, not zero |
| 3151 | } |
| 3152 | ptr = p |
| 3153 | cache[k] = ptr |
| 3154 | } |
| 3155 | if release.CanaryExpired(ptr, now) { |
| 3156 | n++ |
| 3157 | } |
| 3158 | } |
| 3159 | return &n |
| 3160 | } |
| 3161 | |
| 3162 | // peekAcmeCache pre-parses the agent args for -acme-cache the same way |
| 3163 | // peekRegistryPath peeks -registry: the edge does the authoritative parse, |
| 3164 | // this just lets the domains loop stamp cert expiry from the SAME cache dir |
| 3165 | // the edge actually mints into when the operator moved it off the default. |
| 3166 | // "" = flag not given; domainsTickConfig then derives the edge's own default |
| 3167 | // ($CARLOS_DATA_DIR/acme). |
| 3168 | func peekAcmeCache(args []string) string { |
| 3169 | fs := flag.NewFlagSet("agent", flag.ContinueOnError) |
| 3170 | fs.SetOutput(io.Discard) |
| 3171 | fs.String("registry", "", "") |
| 3172 | acme := fs.String("acme-cache", "", "") |
| 3173 | fs.String("https", "", "") |
| 3174 | fs.String("http", "", "") |
| 3175 | fs.String("fallback", "", "") |
| 3176 | fs.Bool("dev", false, "") |
| 3177 | fs.Usage = cliflag.Usage(fs, "") |
| 3178 | _ = fs.Parse(args) |
| 3179 | return *acme |
| 3180 | } |
| 3181 | |
| 3182 | // domainsTickConfig builds one reconcile pass's Config from the environment |
| 3183 | // — pure (getenv + the already-read public IP injected) so the whole env |
| 3184 | // contract is testable without t.Setenv. PLATFORM_APPS_DOMAIN unset returns |
| 3185 | // the zero Config, on which Reconcile is inert by contract. |
| 3186 | // |
| 3187 | // The SECOND return is the parsed PLATFORM_EDGE_IPS list (with the public-IP |
| 3188 | // fallback applied). It is not a Config field: Config.EdgeIPs is now a func |
| 3189 | // answering the deployment's LIVE edge set, which needs a deployment store — |
| 3190 | // something this function must never touch if it is to stay pure and |
| 3191 | // getenv-injectable. domainsLoop folds the two together once its store is |
| 3192 | // open. A field only the agent's own wiring reads would be dead weight in |
| 3193 | // every other caller and every test fixture, hence a return value. |
| 3194 | // |
| 3195 | // Resolve is armed for every configured deployment now, no longer gated on |
| 3196 | // this box knowing an edge IP: a box with no PLATFORM_EDGE_IPS whose FOLD |
| 3197 | // supplies edges must still stamp. The stand-down when nothing is knowable |
| 3198 | // moved to where the set is actually known — domains.Reconcile's section 4, |
| 3199 | // per pass — because "no edge addresses" is now a per-pass fact, not a |
| 3200 | // boot-time one. |
| 3201 | // |
| 3202 | // The error return is ADVISORY, never fatal: it is non-nil only when |
| 3203 | // CARLOS_RELEASE_PUBKEY was SET but would not parse, and it exists so the |
| 3204 | // caller can log the real cause. release.ParsePublicKey says the same sentence |
| 3205 | // for an empty value as for a typo'd one, so an unset key is reported as nil |
| 3206 | // error and a nil ReleasePub — the two are different operator problems and the |
| 3207 | // one log line about them must not name the wrong one. The Config is fully |
| 3208 | // usable either way; only the auto-purge is off. |
| 3209 | func domainsTickConfig(getenv func(string) string, publicIP string) (domains.Config, []string, error) { |
| 3210 | domain := strings.TrimSpace(getenv("PLATFORM_APPS_DOMAIN")) |
| 3211 | if domain == "" { |
| 3212 | return domains.Config{}, nil, nil |
| 3213 | } |
| 3214 | var ips []string |
| 3215 | for _, ip := range strings.Split(getenv("PLATFORM_EDGE_IPS"), ",") { |
| 3216 | if ip = strings.TrimSpace(ip); ip != "" { |
| 3217 | ips = append(ips, ip) |
| 3218 | } |
| 3219 | } |
| 3220 | if len(ips) == 0 && publicIP != "" { |
| 3221 | ips = []string{publicIP} |
| 3222 | } |
| 3223 | dd := getenv("CARLOS_DATA_DIR") |
| 3224 | if dd == "" { |
| 3225 | dd = "/data" |
| 3226 | } |
| 3227 | cfg := domains.Config{ |
| 3228 | AppsDomain: domain, |
| 3229 | // The same default the edge's -acme-cache flag carries |
| 3230 | // (edge.RunWithActivator): $CARLOS_DATA_DIR/acme. domainsLoop |
| 3231 | // overrides this when the flag was given explicitly. |
| 3232 | AcmeDir: filepath.Join(dd, "acme"), |
| 3233 | Now: time.Now, |
| 3234 | Resolve: resolveHost, |
| 3235 | } |
| 3236 | // EdgeIPs is deliberately left nil here — a Config out of this function |
| 3237 | // alone stamps nothing. domainsLoop installs the fold once it has a |
| 3238 | // store; see the doc comment. |
| 3239 | |
| 3240 | // The root release public key, same source and same parser as adoptLoop's. |
| 3241 | // Absence is TOLERATED and leaves ReleasePub nil, which switches the |
| 3242 | // auto-purge off: a box that cannot verify an app's production flag must |
| 3243 | // not destroy its objects. Everything else about the pass — including |
| 3244 | // hiding trashed apps, the reversible half — carries on regardless. A key |
| 3245 | // that is SET but unparseable is tolerated the same way, and RETURNED, so |
| 3246 | // the operator who typo'd it hears about the typo rather than being told |
| 3247 | // the variable is unset. |
| 3248 | if raw := strings.TrimSpace(getenv("CARLOS_RELEASE_PUBKEY")); raw != "" { |
| 3249 | pub, err := release.ParsePublicKey(raw, "CARLOS_RELEASE_PUBKEY") |
| 3250 | if err != nil { |
| 3251 | return cfg, ips, err |
| 3252 | } |
| 3253 | cfg.ReleasePub = pub |
| 3254 | } |
| 3255 | return cfg, ips, nil |
| 3256 | } |
| 3257 | |
| 3258 | // provisionTickConfig builds one instances reconcile pass's Config from the |
| 3259 | // environment — pure (getenv-injected), mirroring domainsTickConfig's shape. |
| 3260 | // Unlike domainsTickConfig, there is no inert zero-Config case: an unset |
| 3261 | // CARLOS_REGION/CARLOS_POOLS is a LEGAL state, not a "feature not |
| 3262 | // configured" one — see provision.Reconcile's doc comment — a box that |
| 3263 | // hasn't been told its region/pools still serves only ""-scoped records, |
| 3264 | // never someone else's residency promise. So this always returns a |
| 3265 | // populated Config, and CARLOS_DATA_DIR is honored (default "/data") |
| 3266 | // regardless of whether region/pools are set. CARLOS_INBOUND (design §8) is |
| 3267 | // the same shape again: unset/"any" is today's behaviour exactly, and it is |
| 3268 | // a declaration about what work THIS BOX will accept, independent of a |
| 3269 | // record's own "no-inbound" placement request — see ReconcileConfig.Inbound. |
| 3270 | func provisionTickConfig(getenv func(string) string) provision.ReconcileConfig { |
| 3271 | region := strings.TrimSpace(getenv("CARLOS_REGION")) |
| 3272 | var pools []string |
| 3273 | for _, p := range strings.Split(getenv("CARLOS_POOLS"), ",") { |
| 3274 | if p = strings.TrimSpace(p); p != "" { |
| 3275 | pools = append(pools, p) |
| 3276 | } |
| 3277 | } |
| 3278 | dd := getenv("CARLOS_DATA_DIR") |
| 3279 | if dd == "" { |
| 3280 | dd = "/data" |
| 3281 | } |
| 3282 | // CARLOS_INBOUND is the same apps-domain-style inert gate as |
| 3283 | // CARLOS_REGION/CARLOS_POOLS above: unset (or "any") is today's |
| 3284 | // behaviour exactly, a legal "no constraint" state, never a "feature not |
| 3285 | // configured" one. Unlike Region above it, this IS folded and validated |
| 3286 | // here (provision.NormalizeInbound) rather than left a pure passthrough: |
| 3287 | // review fix, Task 10 finding — an un-folded "None"/"NONE" or a typo'd |
| 3288 | // value compared exact-equal against InboundNone false, so a box meant |
| 3289 | // to be no-inbound would silently mint routed instances it cannot serve. |
| 3290 | // NormalizeInbound folds case/whitespace and maps any unrecognised value |
| 3291 | // to the restrictive reading (with its own log line), so |
| 3292 | // provision.Reconcile/ReconcileSidecars downstream see only the closed |
| 3293 | // vocabulary and their exact-equality comparisons stay correct. |
| 3294 | inbound := provision.NormalizeInbound(getenv("CARLOS_INBOUND")) |
| 3295 | cfg := provision.ReconcileConfig{ |
| 3296 | Region: region, |
| 3297 | Pools: pools, |
| 3298 | DataDir: dd, |
| 3299 | Now: time.Now, |
| 3300 | Inbound: inbound, |
| 3301 | // The armed config guard, asked before a mint (amadan #213). Nil |
| 3302 | // when none is armed (every test that does not arm one). |
| 3303 | ConfigGate: configGate(), |
| 3304 | } |
| 3305 | // The root release public key, same source and same parser as adoptLoop's |
| 3306 | // and domainsTickConfig's. It guards ONE decision in this pass: whether a |
| 3307 | // trash marker may retire a production-flagged app's routes (#340). |
| 3308 | // |
| 3309 | // Absence is tolerated and leaves ReleasePub nil, which retires trashed |
| 3310 | // routes UNGUARDED — deliberately the same end state domains.Config |
| 3311 | // documents and accepts for the same decision (§1c), because two passes |
| 3312 | // reaching opposite conclusions about one app is worse than either |
| 3313 | // answer. A key SET but unparseable lands in the same place, so it is |
| 3314 | // logged here rather than swallowed: it is a persistent misconfiguration, |
| 3315 | // and the operator who typo'd it must hear about the typo rather than |
| 3316 | // silently losing the guard. |
| 3317 | if raw := strings.TrimSpace(getenv("CARLOS_RELEASE_PUBKEY")); raw != "" { |
| 3318 | pub, err := release.ParsePublicKey(raw, "CARLOS_RELEASE_PUBKEY") |
| 3319 | if err != nil { |
| 3320 | log.Printf("agent: instances pass: %v — a trash marker will retire routes without checking the production flag", err) |
| 3321 | } else { |
| 3322 | cfg.ReleasePub = pub |
| 3323 | } |
| 3324 | } |
| 3325 | return cfg |
| 3326 | } |
| 3327 | |
| 3328 | // resolveHost is the production DNS seam: the default resolver with a hard |
| 3329 | // 3s timeout, so one dead nameserver can't stall a whole reconcile pass. |
| 3330 | func resolveHost(host string) ([]string, error) { |
| 3331 | ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second) |
| 3332 | defer cancel() |
| 3333 | return net.DefaultResolver.LookupHost(ctx, host) |
| 3334 | } |
| 3335 | |
| 3336 | // imdsToken fetches an EC2 IMDSv2 session token (60s TTL) — the shared |
| 3337 | // first step of every IMDS read in this file. Any failure (not on EC2, |
| 3338 | // IMDS disabled, a transient metadata-endpoint hiccup) returns "", and |
| 3339 | // every caller already treats a blank token as "no answer", never an |
| 3340 | // error. |
| 3341 | func imdsToken(client *http.Client) string { |
| 3342 | req, err := http.NewRequest(http.MethodPut, "http://169.254.169.254/latest/api/token", nil) |
| 3343 | if err != nil { |
| 3344 | return "" |
| 3345 | } |
| 3346 | req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", "60") |
| 3347 | resp, err := client.Do(req) |
| 3348 | if err != nil { |
| 3349 | return "" |
| 3350 | } |
| 3351 | defer resp.Body.Close() |
| 3352 | if resp.StatusCode != http.StatusOK { |
| 3353 | return "" |
| 3354 | } |
| 3355 | body, _ := io.ReadAll(resp.Body) |
| 3356 | return strings.TrimSpace(string(body)) |
| 3357 | } |
| 3358 | |
| 3359 | // imdsLeaf reads one IMDS leaf under the given token, trimmed. The status |
| 3360 | // code travels back with the body so a caller can tell "empty leaf" from |
| 3361 | // "leaf did not answer 200" without a second round trip. |
| 3362 | func imdsLeaf(client *http.Client, token, path string) (string, int) { |
| 3363 | req, err := http.NewRequest(http.MethodGet, "http://169.254.169.254"+path, nil) |
| 3364 | if err != nil { |
| 3365 | return "", 0 |
| 3366 | } |
| 3367 | req.Header.Set("X-aws-ec2-metadata-token", token) |
| 3368 | resp, err := client.Do(req) |
| 3369 | if err != nil { |
| 3370 | return "", 0 |
| 3371 | } |
| 3372 | defer resp.Body.Close() |
| 3373 | body, _ := io.ReadAll(resp.Body) |
| 3374 | return strings.TrimSpace(string(body)), resp.StatusCode |
| 3375 | } |
| 3376 | |
| 3377 | // readPublicIP best-effort reads this box's public IPv4 from EC2 IMDSv2 |
| 3378 | // (token PUT, then the public-ipv4 leaf). Any failure — not on EC2, IMDS |
| 3379 | // disabled, no public address — returns "", and the caller treats that as |
| 3380 | // "unknown", never as an error: a box without a discoverable public IP is a |
| 3381 | // valid deployment whose DNS checks simply stand down. |
| 3382 | func readPublicIP() string { |
| 3383 | client := &http.Client{Timeout: 2 * time.Second} |
| 3384 | token := imdsToken(client) |
| 3385 | if token == "" { |
| 3386 | return "" |
| 3387 | } |
| 3388 | ip, status := imdsLeaf(client, token, "/latest/meta-data/public-ipv4") |
| 3389 | if status != http.StatusOK { |
| 3390 | return "" |
| 3391 | } |
| 3392 | if net.ParseIP(ip) == nil { |
| 3393 | return "" |
| 3394 | } |
| 3395 | return ip |
| 3396 | } |
| 3397 | |
| 3398 | // edgeDoorOptions is the edge-local door's option (#175 P7): none unless |
| 3399 | // CARLOS_EDGE_DOOR names a socket path, and then edge.WithEdgeDoor with the |
| 3400 | // box's public address. publicAddr is a func so the address is read only for |
| 3401 | // a box that asked for the door: off EC2, readPublicIP can block for its |
| 3402 | // whole timeout. An empty address still passes through, and the edge logs |
| 3403 | // that the door is off for want of one. |
| 3404 | func edgeDoorOptions(getenv func(string) string, publicAddr func() string) []edge.RunOption { |
| 3405 | p := strings.TrimSpace(getenv("CARLOS_EDGE_DOOR")) |
| 3406 | if p == "" { |
| 3407 | return nil |
| 3408 | } |
| 3409 | return []edge.RunOption{edge.WithEdgeDoor(p, publicAddr())} |
| 3410 | } |
| 3411 | |
| 3412 | // ReadPublicIP is readPublicIP for callers outside this package — the relay |
| 3413 | // advertises the box's Elastic IP as its relay address and learns it exactly |
| 3414 | // as the edge does (relay spec §3, "Listeners"). Same contract: "" means |
| 3415 | // unknown, never an error. |
| 3416 | func ReadPublicIP() string { return readPublicIP() } |
| 3417 | |
| 3418 | // readPlacementRegion best-effort reads this box's AWS region from EC2 |
| 3419 | // IMDSv2 (token PUT, then the placement/region leaf) — the same idiom and |
| 3420 | // the same "" on any failure contract as readPublicIP: not on EC2, IMDS |
| 3421 | // disabled, or a metadata-endpoint hiccup all read as "unknown" here, and |
| 3422 | // the caller (edgesFileRegion) treats that identically to an unset |
| 3423 | // CARLOS_AWS_REGION override — an edge with neither is simply unarmed for |
| 3424 | // steering, not an error. |
| 3425 | func readPlacementRegion() string { |
| 3426 | client := &http.Client{Timeout: 2 * time.Second} |
| 3427 | token := imdsToken(client) |
| 3428 | if token == "" { |
| 3429 | return "" |
| 3430 | } |
| 3431 | region, status := imdsLeaf(client, token, "/latest/meta-data/placement/region") |
| 3432 | if status != http.StatusOK { |
| 3433 | return "" |
| 3434 | } |
| 3435 | return region |
| 3436 | } |
| 3437 | |
| 3438 | // ReadPlacementRegion is readPlacementRegion for callers outside this |
| 3439 | // package. The relay names its region in every heartbeat and checks it |
| 3440 | // against each app's region list, and it must be the SAME string the edges |
| 3441 | // file uses — hence the same probe, with the caller applying the same |
| 3442 | // CARLOS_AWS_REGION override edgesFileRegion applies. |
| 3443 | func ReadPlacementRegion() string { return readPlacementRegion() } |
| 3444 | |
| 3445 | // domainsLoop runs the domains reconcile pass forever on the STATUS tick's |
| 3446 | // cadence (5min + jitter), not the adopt loop's hour: the console's |
| 3447 | // custom-domain guidance promises "Checks run every few minutes", and a DNS |
| 3448 | // fix that takes up to an hour to turn green would read as broken. Thin and |
| 3449 | // untested by convention (like adoptLoop/statusLoop) — everything with |
| 3450 | // behavior lives in domainsTickConfig and domains.Reconcile. Missing config |
| 3451 | // disables the loop with one log line, the same inert-until-configured |
| 3452 | // shape as its siblings. |
| 3453 | func domainsLoop(reg *registry.Registry, acmeCacheFlag string, hostClaim bool) { |
| 3454 | ctx := context.Background() |
| 3455 | // The inert gate is checked BEFORE readPublicIP: an unconfigured |
| 3456 | // deployment must make no IMDS call at all (a 2s link-local probe on |
| 3457 | // every agent start, on boxes that aren't even on EC2, for a feature |
| 3458 | // that is switched off). Inert means inert — no network, not just no |
| 3459 | // writes. domainsTickConfig re-checks the same variable; that's the |
| 3460 | // authoritative gate, this is the "don't pay for it first" one. |
| 3461 | if strings.TrimSpace(os.Getenv("PLATFORM_APPS_DOMAIN")) == "" { |
| 3462 | log.Printf("agent: domains loop disabled: PLATFORM_APPS_DOMAIN is not set") |
| 3463 | return |
| 3464 | } |
| 3465 | cfg, envIPs, keyErr := domainsTickConfig(os.Getenv, readPublicIP()) |
| 3466 | cfg.HostClaim = hostClaim |
| 3467 | if cfg.AppsDomain == "" { |
| 3468 | log.Printf("agent: domains loop disabled: PLATFORM_APPS_DOMAIN is not set") |
| 3469 | return |
| 3470 | } |
| 3471 | if acmeCacheFlag != "" { |
| 3472 | cfg.AcmeDir = acmeCacheFlag |
| 3473 | } |
| 3474 | store, err := openDeployStore(ctx) |
| 3475 | if err != nil { |
| 3476 | log.Printf("agent: domains loop disabled: %v", err) |
| 3477 | return |
| 3478 | } |
| 3479 | // The edge set is folded PER PASS, here rather than in |
| 3480 | // domainsTickConfig, because it needs the store this line above just |
| 3481 | // opened — and because a set read once at arm is the bug this replaces: |
| 3482 | // under latency steering a correctly-pointed customer domain resolves to |
| 3483 | // whichever region's edge is nearest, which a single box's boot-frozen |
| 3484 | // PLATFORM_EDGE_IPS reads as "elsewhere". envIPs stays in the union, so |
| 3485 | // a deployment whose bucket is unreadable degrades to exactly the old |
| 3486 | // behaviour and never below it. The bound matters: a hung store must |
| 3487 | // cost this pass 30s, not the loop forever. |
| 3488 | cfg.EdgeIPs = func() []string { |
| 3489 | ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second) |
| 3490 | defer cancel() |
| 3491 | return edgesfile.EdgeIPs(ctx, store, envIPs, time.Now()) |
| 3492 | } |
| 3493 | log.Printf("agent: domains loop armed (5min + jitter), apps domain %s, configured edge IPs %v (the live set is folded from every box's edges file each pass) — health stamps every tick; claims reconcile only while this box holds the housekeeper lease", cfg.AppsDomain, envIPs) |
| 3494 | // One line, and it must name the REAL cause: a typo'd key and an absent |
| 3495 | // one are different problems with different fixes, and this is the only |
| 3496 | // place an operator learns that purging is not happening. |
| 3497 | if keyErr != nil { |
| 3498 | log.Printf("agent: domains loop auto-purge disabled: %v — trashed apps will be hidden but never destroyed", keyErr) |
| 3499 | } else if cfg.ReleasePub == nil { |
| 3500 | log.Printf("agent: domains loop auto-purge disabled: CARLOS_RELEASE_PUBKEY is not set — trashed apps will be hidden but never destroyed") |
| 3501 | } |
| 3502 | duty := domainsDuty(func(ctx context.Context) error { |
| 3503 | return domains.Reconcile(ctx, store, reg, cfg) |
| 3504 | }) |
| 3505 | fleetBox := shouldSkipPublisherElection(os.Getenv) |
| 3506 | delay := time.Minute |
| 3507 | for { |
| 3508 | time.Sleep(delay) |
| 3509 | delay = 5*time.Minute + rand.N(time.Minute) |
| 3510 | // The deployment-global half, gated; then the per-box half on EVERY |
| 3511 | // platform box (#303): DNS and TLS health stamp this box's own |
| 3512 | // registry from this box's own view, no lease involved, ONCE per |
| 3513 | // tick — after the duties so a host section 3 just minted is |
| 3514 | // stamped in the same tick it appeared. Never on a fleet box: it |
| 3515 | // never contends for the lease for the same reason (housekeeperNever), |
| 3516 | // and it publishes no edges file, so it has no standing to judge the |
| 3517 | // deployment's DNS. (The old reason — that its EdgeIPs fall back to |
| 3518 | // its own address — no longer holds now that the set is folded from |
| 3519 | // every box's file; the exclusion stands on the publication rule |
| 3520 | // instead, which is the durable one.) |
| 3521 | runHousekeeperDuties(housekeeperGate, duty) |
| 3522 | if !fleetBox { |
| 3523 | if err := domains.StampHealth(ctx, reg, cfg, nil); err != nil { |
| 3524 | log.Printf("agent: domains health: %v", err) |
| 3525 | } |
| 3526 | } |
| 3527 | } |
| 3528 | } |
| 3529 | |
| 3530 | // domainsDuty wraps one domains reconcile pass as a housekeeper duty. The |
| 3531 | // pass is deployment-wide state — materializing routes from claims, minting |
| 3532 | // platform-domain and canary hosts, purging trashed apps — with one correct |
| 3533 | // writer, and since multi-region (spec 2026-08-20) every platform box in a |
| 3534 | // deployment carries PLATFORM_APPS_DOMAIN (the per-account wildcards and the |
| 3535 | // steering probe need it), so an ungated loop would be N reconcilers racing |
| 3536 | // over the same records. Gating on the lease is what makes the env safe to |
| 3537 | // set fleet-wide. (The DNS/TLS health stamps are NOT in this pass: they are |
| 3538 | // per-box, domains.StampHealth, run by domainsLoop on every box — #303.) Single-box |
| 3539 | // deployments are unchanged: the one box always holds the lease. A box that |
| 3540 | // never contends (a fleet box) never runs the pass — which is right, the |
| 3541 | // pass is a platform-box job. |
| 3542 | func domainsDuty(pass func(context.Context) error) housekeeperDuty { |
| 3543 | return housekeeperDuty{"domains reconcile", func(ctx context.Context) error { |
| 3544 | if pass == nil { |
| 3545 | return nil |
| 3546 | } |
| 3547 | return pass(ctx) |
| 3548 | }} |
| 3549 | } |
| 3550 | |
| 3551 | // instancesLoop runs the instance-provisioning reconcile pass forever on the |
| 3552 | // same cadence as domainsLoop (5min + jitter, first tick after 1 minute) — |
| 3553 | // turning bucket instance records into registry routes for THIS agent's |
| 3554 | // region/pool scope (provision.Reconcile). Unlike domainsLoop there is NO |
| 3555 | // apps-domain-style inert gate: an unset CARLOS_REGION/CARLOS_POOLS is a |
| 3556 | // legal, still-armed state (see provisionTickConfig and |
| 3557 | // provision.Reconcile's doc comment — a box with no region deliberately |
| 3558 | // still serves ""-scoped records), so this loop arms whenever the deploy |
| 3559 | // store opens, full stop. Missing deployment config disables the loop with |
| 3560 | // one log line, the same "inert until configured" shape adoptLoop and |
| 3561 | // domainsLoop use for their own missing config. Per-pass errors are logged; |
| 3562 | // the loop lives on. |
| 3563 | // |
| 3564 | // Also runs ReconcileFleetBindings (customer-fleets plan, Task 6) on the same |
| 3565 | // cadence, platform boxes only: turns every fleet's published |
| 3566 | // fleets/*/bindings/*.json object into this platform box's own |
| 3567 | // registry.Route.FleetBox — see fleetbind.go's doc comment for the |
| 3568 | // fleet-placement/corrupt-object contract. Deliberately homed HERE rather |
| 3569 | // than domainsLoop (an earlier pass rode domainsLoop's ticker, since it |
| 3570 | // already had a store+reg in scope) — that coupled binding materialization |
| 3571 | // to PLATFORM_APPS_DOMAIN being set, which has nothing to do with fleets; |
| 3572 | // this loop has no such gate and shares the same store already opened above. |
| 3573 | func instancesLoop(reg *registry.Registry, writes hostwrite.Config, bs *buildStates, configure ...func(*provision.ReconcileConfig)) { |
| 3574 | ctx := context.Background() |
| 3575 | store, err := openDeployStore(ctx) |
| 3576 | if err != nil { |
| 3577 | log.Printf("agent: instances loop disabled: %v", err) |
| 3578 | return |
| 3579 | } |
| 3580 | cfg := configuredClaimProvision(provisionTickConfig(os.Getenv), configure) |
| 3581 | backfill := newClaimBackfill(writes, store, cfg.ReleasePub, os.Getenv("CARLOS_DEPLOYMENT_BUCKET")) |
| 3582 | log.Printf("agent: instances loop armed (5min + jitter), region %q pools %v", cfg.Region, cfg.Pools) |
| 3583 | // Platform-only: a fleet box (CARLOS_FLEET set) never reads other |
| 3584 | // fleets' bindings — it only ever WRITES its own (fleetbind.go's |
| 3585 | // fleetBindingHook/publishResidentFleetBindings, wired in RunAgent). |
| 3586 | // Reusing shouldSkipPublisherElection rather than re-testing the env var |
| 3587 | // here: both questions are exactly "is this box a fleet box", so one |
| 3588 | // predicate, already covered by TestShouldSkipPublisherElection, answers |
| 3589 | // both. |
| 3590 | isFleetBox := shouldSkipPublisherElection(os.Getenv) |
| 3591 | // An offline move's participants keep their reports current between the |
| 3592 | // passes below (finding 7); idle, it reads nothing. |
| 3593 | go offlineReassertLoop(ctx, store) |
| 3594 | delay := time.Minute |
| 3595 | for { |
| 3596 | time.Sleep(delay) |
| 3597 | delay = 5*time.Minute + rand.N(time.Minute) |
| 3598 | if err := backfill(ctx); err != nil { |
| 3599 | log.Printf("agent: claim backfill: %v", err) |
| 3600 | } |
| 3601 | if err := reconcileSerialized(ctx, store, reg, cfg); err != nil { |
| 3602 | log.Printf("agent: instances pass: %v", err) |
| 3603 | } |
| 3604 | // Every reconcile refreshes the build-state copies (settings spec |
| 3605 | // §4.2), this 5-minute one too: the fast leg above runs only on a |
| 3606 | // serial move. |
| 3607 | if bs != nil { |
| 3608 | bs.refreshAll(ctx) |
| 3609 | } |
| 3610 | telemetryPass(ctx, store, reg) |
| 3611 | plansPass(ctx, store, reg, os.Getenv) |
| 3612 | runMovesPass(ctx, store) |
| 3613 | runRunsPass(ctx) |
| 3614 | deploymentRoutes := edgesfile.DeploymentRoutes(ctx, store, time.Now()) |
| 3615 | if err := domains.ReconcileAliases(ctx, store, reg, time.Now, deploymentRoutes, &aliasClaimCache); err != nil { |
| 3616 | log.Printf("agent: instances alias pass: %v", err) |
| 3617 | } |
| 3618 | if !isFleetBox { |
| 3619 | if err := ReconcileFleetBindings(ctx, store, reg); err != nil { |
| 3620 | log.Printf("agent: fleet bindings reconcile: %v", err) |
| 3621 | } |
| 3622 | } |
| 3623 | } |
| 3624 | } |
| 3625 | |
| 3626 | // offlineReassertLoop runs the armed moves worker's offlineReassert every |
| 3627 | // offlineReassertEvery plus up to 5 s of jitter, for the life of the agent: |
| 3628 | // while this box takes part in an offline move, its source and sealer reports |
| 3629 | // stay current for the doors whatever the serial does. Thin and untested by |
| 3630 | // convention, like instancesLoop; the tick is offlineReassert. |
| 3631 | func offlineReassertLoop(ctx context.Context, s release.Store) { |
| 3632 | for { |
| 3633 | time.Sleep(offlineReassertEvery + rand.N(5*time.Second)) |
| 3634 | if w := armedMovesWorker.Load(); w != nil { |
| 3635 | w.offlineReassert(ctx, s) |
| 3636 | } |
| 3637 | } |
| 3638 | } |
| 3639 | |
| 3640 | // retentionLoop runs the release-retention pass forever on the slow |
| 3641 | // reconcile cadence (5min + jitter, first tick after 1 minute) — never the |
| 3642 | // 2s fast pass: pruning is not urgent, and a policy written now is enforced |
| 3643 | // within minutes, which is the promise the CLI makes ("boxes prune within |
| 3644 | // ~10 minutes"). Thin and untested by convention (like domainsLoop / |
| 3645 | // instancesLoop); everything with behavior lives in retentionPass. Missing |
| 3646 | // deployment config, a missing/unparseable release pubkey, or no box |
| 3647 | // reports to read (retentionReports: the status store, or on a sole |
| 3648 | // housekeeper with none this box's own build-state report) disables the |
| 3649 | // loop with one log line — fail closed, domainsLoop's auto-purge shape: a |
| 3650 | // box that cannot verify pointers, or cannot learn which builds the build |
| 3651 | // states name (settings spec rev 11 §4.4), must not delete anything. |
| 3652 | // |
| 3653 | // Since the multi-region writer map (spec |
| 3654 | // 2026-08-20-multi-region-ingress-design.md, Part 1) the pass itself is a |
| 3655 | // HOUSEKEEPER DUTY: it runs only on the one box holding the deployment's |
| 3656 | // housekeeper lease. releases/** is deployment-global state, and N boxes |
| 3657 | // pruning it concurrently is the bug class the spec refuses to have. Every |
| 3658 | // other box still arms this loop and still ticks — it just prunes nothing, |
| 3659 | // silently, until the day it wins the lease (housekeeper.go). On a |
| 3660 | // single-box deployment the holder is always this box, so the behaviour is |
| 3661 | // unchanged. |
| 3662 | func retentionLoop(reg *registry.Registry, buildStates func() (map[string]status.BuildStatePins, error)) { |
| 3663 | ctx := context.Background() |
| 3664 | // Platform boxes only (spec F6). Retention prunes the PLATFORM's release |
| 3665 | // objects — a customer's box scheduling deletes of them was already wrong |
| 3666 | // under direct S3, and through the store door it is refused outright (the |
| 3667 | // door's write/delete ACL scopes a fleet to fleets/<fleet>/**), so arming |
| 3668 | // here would buy nothing but an hourly loop of 404s. Reuses |
| 3669 | // shouldSkipPublisherElection because the question is exactly "is this a |
| 3670 | // fleet box", the same one instancesLoop asks — and it is deliberately the |
| 3671 | // bare CARLOS_FLEET test, not resolveFleetIdentity: a fleet box whose |
| 3672 | // LABEL is misconfigured is still a fleet box, and must still not prune. |
| 3673 | if shouldSkipPublisherElection(os.Getenv) { |
| 3674 | log.Printf("agent: retention loop is a platform-box job — skipped on a fleet box (spec F6)") |
| 3675 | return |
| 3676 | } |
| 3677 | store, err := openDeployStore(ctx) |
| 3678 | if err != nil { |
| 3679 | log.Printf("agent: retention loop disabled: %v", err) |
| 3680 | return |
| 3681 | } |
| 3682 | raw := strings.TrimSpace(os.Getenv("CARLOS_RELEASE_PUBKEY")) |
| 3683 | if raw == "" { |
| 3684 | log.Printf("agent: retention loop disabled: CARLOS_RELEASE_PUBKEY is not set — retention policies will not be enforced by this box") |
| 3685 | return |
| 3686 | } |
| 3687 | pub, err := release.ParsePublicKey(raw, "CARLOS_RELEASE_PUBKEY") |
| 3688 | if err != nil { |
| 3689 | log.Printf("agent: retention loop disabled: %v — retention policies will not be enforced by this box", err) |
| 3690 | return |
| 3691 | } |
| 3692 | // The box reports (settings spec rev 11 §4.4: the build-state builds |
| 3693 | // every box names) live in the status store, which is not always the |
| 3694 | // deployment store above (CARLOS_STATUS_BUCKET wins); a sole |
| 3695 | // housekeeper with none reads its own (retentionReports). Without them |
| 3696 | // no app can be pruned safely, so the loop does not arm. |
| 3697 | reports, err := retentionReports(ctx, os.Getenv, store, buildStates) |
| 3698 | if err != nil { |
| 3699 | log.Printf("agent: retention loop disabled: the box reports cannot be read (%v). retention policies will not be enforced by this box", err) |
| 3700 | return |
| 3701 | } |
| 3702 | if reports.local != nil { |
| 3703 | log.Printf("agent: retention: no status store; as the sole housekeeper this box reads its own build-state report") |
| 3704 | } |
| 3705 | log.Printf("agent: retention loop armed (5min + jitter) — prunes only while this box holds the housekeeper lease") |
| 3706 | // The pass runs under the HOLDING's context, not this loop's: a sweep |
| 3707 | // that is halfway through its deletes when renewal fails must stop, not |
| 3708 | // finish against a lock another box now owns (housekeeper.go). |
| 3709 | prune := housekeeperDuty{"retention pass", func(ctx context.Context) error { |
| 3710 | return retentionPass(ctx, store, reports, pub, reg) |
| 3711 | }} |
| 3712 | delay := time.Minute |
| 3713 | for { |
| 3714 | time.Sleep(delay) |
| 3715 | delay = 5*time.Minute + rand.N(time.Minute) |
| 3716 | runHousekeeperDuties(housekeeperGate, prune) |
| 3717 | } |
| 3718 | } |
| 3719 | |
| 3720 | // stampedUnitNames is the metrics sweep's one registry read: every distinct |
| 3721 | // bespoke unit stamped on a route (Route.Unit, the #110 column — an |
| 3722 | // operator's console.service, tito's titogo-canary@<tenant>.service), which |
| 3723 | // the box's cgroup walk cannot discover by name convention. |
| 3724 | // |
| 3725 | // Called once per 60s tick from metricsLoop's probe, so it is deliberately a |
| 3726 | // plain List + filter with no cache: the registry is a local SQLite file, one |
| 3727 | // read a minute is nothing next to the pass loops already reading it, and a |
| 3728 | // cache would be one more thing to invalidate when an operator stamps a unit. |
| 3729 | // |
| 3730 | // A read failure returns nil — this tick's stamped units are lost, the sweep |
| 3731 | // and the agent are not. The names are returned sorted so the slice handed |
| 3732 | // down each tick is stable (nothing below depends on the order; a stable one |
| 3733 | // just makes a log or a test diff mean something). |
| 3734 | func stampedUnitNames(reg *registry.Registry) []string { |
| 3735 | routes, err := reg.List() |
| 3736 | if err != nil { |
| 3737 | log.Printf("agent: metrics: read stamped units: %v", err) |
| 3738 | return nil |
| 3739 | } |
| 3740 | seen := make(map[string]bool, len(routes)) |
| 3741 | var out []string |
| 3742 | for _, rt := range routes { |
| 3743 | if rt.Unit == "" || seen[rt.Unit] { |
| 3744 | continue |
| 3745 | } |
| 3746 | seen[rt.Unit] = true |
| 3747 | out = append(out, rt.Unit) |
| 3748 | } |
| 3749 | sort.Strings(out) |
| 3750 | return out |
| 3751 | } |
| 3752 | |
| 3753 | // resolveStatusBucket applies the status/deployment bucket defaulting |
| 3754 | // documented in this file's package comment: an explicit CARLOS_STATUS_BUCKET |
| 3755 | // always wins, and when it's unset the status loop falls back to whatever |
| 3756 | // CARLOS_DEPLOYMENT_BUCKET is set to, so the common single-bucket deployment |
| 3757 | // is one env var instead of two that must be kept in sync by hand. Returns |
| 3758 | // "" when neither is set — the loop stays disabled, same as before this fix. |
| 3759 | // Deliberately does NOT fall back to CARLOS_DEPLOYMENT_DIR: the status page |
| 3760 | // has no directory-store serving path (the edge/console read status/ out of |
| 3761 | // an S3 bucket, not a local directory), so a DIR-based deployment simply |
| 3762 | // doesn't get this fallback — see the "HONEST LIMIT" note in the package |
| 3763 | // comment above. |
| 3764 | // |
| 3765 | // Takes a getenv func rather than reading os.Getenv itself so the defaulting |
| 3766 | // logic is testable without mutating process-wide environment variables — |
| 3767 | // t.Setenv serializes tests within a package and the host package already |
| 3768 | // has other env-reading tests; a plain map/func seam avoids that coupling |
| 3769 | // entirely. |
| 3770 | func resolveStatusBucket(getenv func(string) string) string { |
| 3771 | return ResolveStatusBucket(getenv) |
| 3772 | } |
| 3773 | |
| 3774 | // ResolveStatusBucket is resolveStatusBucket for callers outside this package |
| 3775 | // — `carlos status -live` reads the boxes an agent wrote, and it must land on |
| 3776 | // the SAME bucket the agent's status loop writes to, including the |
| 3777 | // CARLOS_DEPLOYMENT_BUCKET fallback. Exported as a second name over one |
| 3778 | // implementation rather than renamed at every internal call site, so the |
| 3779 | // defaulting rule documented above stays in exactly one function. |
| 3780 | func ResolveStatusBucket(getenv func(string) string) string { |
| 3781 | if b := getenv("CARLOS_STATUS_BUCKET"); b != "" { |
| 3782 | return b |
| 3783 | } |
| 3784 | return getenv("CARLOS_DEPLOYMENT_BUCKET") |
| 3785 | } |
| 3786 | |
| 3787 | // shouldSkipPublisherElection reports whether this tick must never contend |
| 3788 | // for the cross-box publish lease at all — true exactly when CARLOS_FLEET is |
| 3789 | // set (Task 5, customer-fleets plan): a fleet box self-reports its own Box |
| 3790 | // snapshot every tick (statusTick's WriteBox call is unconditional) but must |
| 3791 | // never win — or even attempt — the election that would have it merge and |
| 3792 | // publish status/fleet.json, which is the platform's own boxes' shared feed. |
| 3793 | // Checked BEFORE lease.Acquire is ever called, not after: a fleet box never |
| 3794 | // makes the network call, not merely one that's discarded on return. |
| 3795 | func shouldSkipPublisherElection(getenv func(string) string) bool { |
| 3796 | return getenv("CARLOS_FLEET") != "" |
| 3797 | } |
| 3798 | |
| 3799 | // upstreamSource holds the live edge's UpstreamStates method, published by |
| 3800 | // RunAgent's edge-ready callback and read by statusLoop's probe set — the same |
| 3801 | // shape tunnelStateVal uses for tunnel state (tunnel_client.go), and for the |
| 3802 | // same reason: the status loop starts BEFORE the edge exists (RunAgent spawns |
| 3803 | // the loops, then builds and serves the edge), so the two cannot be wired by |
| 3804 | // passing a handle down. |
| 3805 | // |
| 3806 | // A func rather than the *edge.Edge itself, so what this package's tests |
| 3807 | // inject is the same one value production sets — no fake Edge, no probe |
| 3808 | // goroutine a test would have no way to stop. |
| 3809 | var upstreamSource atomic.Value // func() map[string][]status.UpstreamState |
| 3810 | |
| 3811 | // geoIPSource is the edge's GeoIP state for the heartbeat, published from |
| 3812 | // the edge-ready callback like upstreamSource (geoip spec §5.5). |
| 3813 | var geoIPSource atomic.Value // func() edge.GeoIPState |
| 3814 | |
| 3815 | // geoIPAlertSource is the edge's geoip reason code for the box alert rule |
| 3816 | // (geoip spec §10.4), published from the same ready callback. |
| 3817 | var geoIPAlertSource atomic.Value // func(time.Time) (float64, bool) |
| 3818 | |
| 3819 | // geoIPAlertReading is the rule's reading, or nil before the edge is ready |
| 3820 | // (no data never fires and never clears). |
| 3821 | func geoIPAlertReading(now time.Time) *float64 { |
| 3822 | fn, _ := geoIPAlertSource.Load().(func(time.Time) (float64, bool)) |
| 3823 | if fn == nil { |
| 3824 | return nil |
| 3825 | } |
| 3826 | v, known := fn(now) |
| 3827 | if !known { |
| 3828 | return nil // no data never fires and never clears |
| 3829 | } |
| 3830 | return &v |
| 3831 | } |
| 3832 | |
| 3833 | // geoIPState reads it, or nil before the edge is ready. |
| 3834 | func geoIPState() *status.GeoIPState { |
| 3835 | fn, _ := geoIPSource.Load().(func() edge.GeoIPState) |
| 3836 | if fn == nil { |
| 3837 | return nil |
| 3838 | } |
| 3839 | s := fn() |
| 3840 | return &status.GeoIPState{Mode: s.Mode, SHA256: s.SHA256, Type: s.Type, Error: s.Error} |
| 3841 | } |
| 3842 | |
| 3843 | // tunnelEnvComplete is tunnelLoop's own arming rule: all three set. |
| 3844 | func tunnelEnvComplete(getenv func(string) string) bool { |
| 3845 | return getenv("CARLOS_TUNNEL_URL") != "" && getenv("CARLOS_TUNNEL_TOKEN") != "" && getenv("CARLOS_TUNNEL_NAME") != "" |
| 3846 | } |
| 3847 | |
| 3848 | // setUpstreamSource publishes (or, with nil, withdraws) the probe-state |
| 3849 | // source. Called once, from the edge-ready callback. |
| 3850 | func setUpstreamSource(fn func() map[string][]status.UpstreamState) { |
| 3851 | if fn == nil { |
| 3852 | upstreamSource.Store(func() map[string][]status.UpstreamState { return nil }) |
| 3853 | return |
| 3854 | } |
| 3855 | upstreamSource.Store(fn) |
| 3856 | } |
| 3857 | |
| 3858 | // upstreamStates reads the current probe table for the heartbeat, or nil when |
| 3859 | // no source has been published yet. |
| 3860 | // |
| 3861 | // nil is the honest answer for every box shape that has no edge probe table to |
| 3862 | // report: `carlos edge` run without the agent, an agent whose ready callback |
| 3863 | // hasn't fired yet, and the overwhelmingly common case of a box with no |
| 3864 | // multi-upstream route at all. All three publish no upstreams key rather than |
| 3865 | // an empty object (status.CollectBox's omitempty rule) — a box that isn't |
| 3866 | // probing and a box running code that predates this field stay |
| 3867 | // indistinguishable on the wire. |
| 3868 | func upstreamStates() map[string][]status.UpstreamState { |
| 3869 | fn, _ := upstreamSource.Load().(func() map[string][]status.UpstreamState) |
| 3870 | if fn == nil { |
| 3871 | return nil |
| 3872 | } |
| 3873 | return fn() |
| 3874 | } |
| 3875 | |
| 3876 | // statusLoop runs the fleet status page's 5-minute tick, forever, once armed |
| 3877 | // — see the package doc comment above for the full env contract. Missing |
| 3878 | // CARLOS_STATUS_BUCKET disables the loop with one log line, the same |
| 3879 | // "a box with no status store is still a valid deployment" shape adoptLoop |
| 3880 | // uses for its own missing config. |
| 3881 | // |
| 3882 | // Each tick: gather fresh probes (small OS-reading funcs in probes.go, none |
| 3883 | // of which this loop's own tests call — statusTick below takes a |
| 3884 | // status.Probes value, so tests inject fixtures instead of touching /proc or |
| 3885 | // a syscall), then contend for the cross-box status-lease. Acquire succeeding |
| 3886 | // means this box is the one publisher for this tick: statusTick both writes |
| 3887 | // this box's own snapshot AND merges+publishes the fleet-wide feed, and the |
| 3888 | // lease is released again right after (held only for the publish, not across |
| 3889 | // the 5-minute gap — the next tick, possibly on a different box, re-acquires |
| 3890 | // fresh). A *lease.HeldError means another box already won this tick's |
| 3891 | // election: completely normal contention, not a failure, so it's swallowed |
| 3892 | // without a log line — every box hits this on every tick it doesn't win, and |
| 3893 | // logging that every 5 minutes forever would bury the log in noise. Any |
| 3894 | // OTHER acquire error (a real S3/credential problem) is logged, same as any |
| 3895 | // other tick failure. |
| 3896 | func statusLoop(reg *registry.Registry, svc *activator.Service, buildStates func() (map[string]status.BuildStatePins, error)) { |
| 3897 | ctx := context.Background() |
| 3898 | // bucket is read for the lease key alone (lease.Acquire below), which a |
| 3899 | // fleet box never reaches: shouldSkipPublisherElection short-circuits the |
| 3900 | // election before it, and a fleet box is the only shape that gets a store |
| 3901 | // with no bucket behind it (openStatusStore). Every other box resolves a |
| 3902 | // real bucket here exactly as before. |
| 3903 | bucket := resolveStatusBucket(os.Getenv) |
| 3904 | prefix := resolveStatusPrefix(os.Getenv) |
| 3905 | region := os.Getenv("CARLOS_STATUS_REGION") |
| 3906 | boxRegion := os.Getenv("STATUS_BOX_REGION") |
| 3907 | skipElection := shouldSkipPublisherElection(os.Getenv) |
| 3908 | |
| 3909 | // Fleet boxes heartbeat under the operator-chosen channel label |
| 3910 | // (fleet_boxes.label — the identity the edge dials via |
| 3911 | // edge.fleetChannelFor), NEVER the public status page's hashed |
| 3912 | // pseudonym (status.Label(machineID)): that anonymization exists so a |
| 3913 | // customer's machine id is never recoverable from the PUBLIC |
| 3914 | // status/fleet.json feed, but a fleet box's own snapshot never reaches |
| 3915 | // that feed at all (skipElection, above) — it only ever lands at |
| 3916 | // fleets/<fleet>/status/boxes/<label>.json, an owner-facing path no |
| 3917 | // public page reads, so the anonymization rationale simply doesn't |
| 3918 | // apply here. An invalid identity degrades to the hashed label rather |
| 3919 | // than disabling the heartbeat outright (RunAgent already refuses, |
| 3920 | // loudly, to arm the binding-publish machinery that actually needs the |
| 3921 | // label to be dial-able). |
| 3922 | fleetLabel := "" |
| 3923 | if identity, err := resolveFleetIdentity(os.Getenv); err == nil { |
| 3924 | fleetLabel = identity.Label |
| 3925 | } else if !errors.Is(err, errNotFleetBox) { |
| 3926 | log.Printf("agent: %v — heartbeat will publish under the hashed pseudonym instead", err) |
| 3927 | } |
| 3928 | |
| 3929 | s, target, err := openStatusStore(ctx, os.Getenv) |
| 3930 | if err != nil { |
| 3931 | if errors.Is(err, errNoStatusStore) { |
| 3932 | log.Printf("agent: status loop disabled: neither CARLOS_STATUS_BUCKET nor CARLOS_DEPLOYMENT_BUCKET is set (CARLOS_DEPLOYMENT_DIR deployments don't publish status yet — set CARLOS_STATUS_BUCKET explicitly)") |
| 3933 | } else { |
| 3934 | log.Printf("agent: status loop disabled: %v", err) |
| 3935 | } |
| 3936 | return |
| 3937 | } |
| 3938 | // store is what statusTick actually reads/writes through: identical to s |
| 3939 | // (a pure passthrough) unless CARLOS_STATUS_PREFIX is set, in which case |
| 3940 | // every key it touches is rebased under that prefix — see |
| 3941 | // release.NewPrefixStore and the package doc comment's |
| 3942 | // CARLOS_STATUS_PREFIX entry. |
| 3943 | store := release.NewPrefixStore(s, prefix) |
| 3944 | // The effective destination is worth a log line on its own: it's easy to |
| 3945 | // set CARLOS_DEPLOYMENT_BUCKET and CARLOS_STATUS_BUCKET to different |
| 3946 | // values by accident (or on purpose) and end up with the agent writing |
| 3947 | // status to a bucket the edge/console never look at — this line is the |
| 3948 | // one place an operator can confirm which bucket (or, on a bucketless |
| 3949 | // fleet box, which door identity) actually won. |
| 3950 | log.Printf("agent: status loop armed (%s + jitter), publishing to %s", status.SnapshotPeriod, target) |
| 3951 | |
| 3952 | // canaryStore/canaryPub resolve the release deployment config |
| 3953 | // countExpiredCanaries needs to answer §5's "how many canaries on this |
| 3954 | // box have expired" — the SAME env pair adoptLoop reads |
| 3955 | // (CARLOS_DEPLOYMENT_BUCKET/_DIR + CARLOS_RELEASE_PUBKEY), not |
| 3956 | // necessarily the status bucket above, which may deliberately be a |
| 3957 | // separate one. Either can come back unusable: no deployment store |
| 3958 | // configured, or no pubkey set. That is a normal, valid box shape (the |
| 3959 | // same "routes managed by hand" case adoptLoop's own doc comment |
| 3960 | // names) — it just means countExpiredCanaries can't answer and reports |
| 3961 | // nil, same as every other unmeasured probe in this loop. Resolved once |
| 3962 | // here, not re-read every tick, mirroring adoptLoop's own one-time |
| 3963 | // resolution. |
| 3964 | // |
| 3965 | // The error is logged rather than swallowed: "no deployment store" is a |
| 3966 | // valid box shape, but since the store can now also fail to build for a |
| 3967 | // REAL misconfiguration (a fleet box with no bucket and no tunnel env — |
| 3968 | // openFleetDoorStore's loud error), silence here would turn that into an |
| 3969 | // unexplained permanently-nil ExpiredCanaries probe. |
| 3970 | canaryStore, err := openDeployStore(ctx) |
| 3971 | if err != nil { |
| 3972 | log.Printf("agent: status loop: expired-canary probe unavailable: %v", err) |
| 3973 | } |
| 3974 | canaryPub, _ := release.ParsePublicKey(os.Getenv("CARLOS_RELEASE_PUBKEY"), "CARLOS_RELEASE_PUBKEY") |
| 3975 | |
| 3976 | // status.SnapshotPeriod IS this cadence — the const is named for the tick |
| 3977 | // below, and every reader that judges a box stale (the published boxes |
| 3978 | // table, the console's private host detail, `carlos status -live`) measures |
| 3979 | // against it, so changing it here must move them together rather than |
| 3980 | // leaving them judging a heartbeat that arrives on a different beat. |
| 3981 | // Jittered on the same shape as adoptLoop's delay: every box on the same |
| 3982 | // cadence would all hit the lease and the bucket in the same second |
| 3983 | // otherwise. |
| 3984 | delay := time.Minute |
| 3985 | for { |
| 3986 | select { |
| 3987 | case <-time.After(delay): |
| 3988 | case <-statusNudge: |
| 3989 | // A rollout transition (flag on only): publish within seconds, |
| 3990 | // at most every two, so a busy rollout cannot hammer the bucket. |
| 3991 | time.Sleep(2 * time.Second) |
| 3992 | } |
| 3993 | delay = status.SnapshotPeriod + rand.N(time.Minute) |
| 3994 | |
| 3995 | // Routes are listed here (not left to statusTick's own reg.List()) |
| 3996 | // because the Listening probe needs to know which sockets are |
| 3997 | // systemd-socket-activated BEFORE it's built — see |
| 3998 | // wrapListeningProbe's doc comment. statusTick lists again for |
| 3999 | // itself since it stays store/lease-agnostic and self-contained; |
| 4000 | // one extra List() per 5-minute tick is not worth threading routes |
| 4001 | // through the pure function's signature for. |
| 4002 | routes, err := reg.List() |
| 4003 | if err != nil { |
| 4004 | log.Printf("agent: status tick: list routes: %v", err) |
| 4005 | continue |
| 4006 | } |
| 4007 | |
| 4008 | memTotal, memAvail := readMemBytes() |
| 4009 | p := status.Probes{ |
| 4010 | MachineID: readMachineID(), |
| 4011 | InstanceID: cachedInstanceID(), |
| 4012 | Region: boxRegion, |
| 4013 | MemFrac: readMemFrac(), |
| 4014 | DiskFrac: readDiskFrac(), |
| 4015 | Listening: wrapListeningProbe(unitBackedSockets(routes), statusSocketProbe(routes, svc)), |
| 4016 | // One systemctl call per 5-minute tick, asking the only question |
| 4017 | // that catches an enabled-but-never-armed adoption timer — the |
| 4018 | // flagship's silent no-adopt state, which nothing on any page |
| 4019 | // could see before this. |
| 4020 | AdoptRunning: adoptTimerActive(), |
| 4021 | // A second systemctl call per tick, asking the same question of |
| 4022 | // the config-sync trigger: an enabled-but-unstarted |
| 4023 | // carlos-adopt.path is a box whose unit-backed config silently |
| 4024 | // waits for the hourly timer (issue #44). |
| 4025 | ConfigPathRunning: configPathActive(), |
| 4026 | // A third, for the restart trigger: an enabled-but-unstarted |
| 4027 | // carlos-restart.path is a box where `carlos restart` answers |
| 4028 | // "requested" and nothing cycles until the hourly pass (issue #113). |
| 4029 | RestartPathRunning: restartPathActive(), |
| 4030 | // Whether this box's own alert rules have anywhere to deliver, |
| 4031 | // read from the env the metrics loop's sinks are built from |
| 4032 | // (discussion #192: an unset webhook was silent for weeks). |
| 4033 | Alerts: alertDeliveryProbe(os.Getenv, nil), |
| 4034 | // ExpiredCanaries answers spec §5 for this box's own routes; see |
| 4035 | // countExpiredCanaries and the canaryStore/canaryPub comment above. |
| 4036 | ExpiredCanaries: countExpiredCanaries(ctx, routes, canaryStore, canaryPub, time.Now()), |
| 4037 | // LagSecs stays nil: litestream's per-DB replication lag isn't |
| 4038 | // surfaced up to the agent process anywhere yet — the |
| 4039 | // replication machinery (internal/lease's Replicator) lives |
| 4040 | // inside the activator, one level down, and doesn't currently |
| 4041 | // export a "seconds since last successful sync" reading this |
| 4042 | // loop could poll. Publishing a guess (0? the box's own clock |
| 4043 | // skew?) would be worse than the honest null CollectBox already |
| 4044 | // knows how to carry — wiring this up is a seam for later, not |
| 4045 | // a gap to paper over now. |
| 4046 | MemTotalBytes: memTotal, |
| 4047 | MemAvailBytes: memAvail, |
| 4048 | Load1: readLoad1(), |
| 4049 | DiskFreeBytes: readDiskFreeBytes(), |
| 4050 | Release: version.Build, |
| 4051 | Platform: runtime.GOOS + "-" + runtime.GOARCH, |
| 4052 | // The invalid-value line is newSystemLeg's to say, once at |
| 4053 | // arming; this tick publishes the channel the box actually |
| 4054 | // follows, five-minutely, and has nowhere to say it again. |
| 4055 | SystemChannel: boxSystemChannel(os.Getenv), |
| 4056 | System: systemStateProbe(release.OptDir()), |
| 4057 | Tunnel: tunnelState(), |
| 4058 | GeoIP: geoIPState(), |
| 4059 | // The edge's probe table, read WHOLE every tick (never merged |
| 4060 | // with the last one), so a route this box has stopped probing |
| 4061 | // leaves no ghost behind in the next snapshot. nil until the |
| 4062 | // edge-ready callback publishes the source, and on any box that |
| 4063 | // probes nothing — both publish no key at all. |
| 4064 | Upstreams: upstreamStates(), |
| 4065 | // One stat (two with a WAL) per database route per tick, which |
| 4066 | // is what `carlos instances list` and `carlos routes` show as DB. |
| 4067 | DBBytes: dbFileBytes, |
| 4068 | // Layout is the armed pass's classification of every |
| 4069 | // instance-kind row, or, with no activator on this box, the |
| 4070 | // rows judged by path shape under Mode "unmanaged" — see |
| 4071 | // layoutStatusReport's own doc comment for why a box with no |
| 4072 | // pass still says something rather than nothing. |
| 4073 | Layout: layoutStatusReport(routes), |
| 4074 | } |
| 4075 | // The hosts this box is refusing because it cannot provide their |
| 4076 | // config (amadan #213): private Metrics only, nothing public. |
| 4077 | claimReport(&p) |
| 4078 | p.ConfigRefusals, p.ConfigRefusalsTotal = configRefusalProbe() |
| 4079 | p.ConfigGuard = configGuardProbe() |
| 4080 | // Builds the root pass rolled back (D1), as the metrics loop's |
| 4081 | // refusal watch last read them. Private Metrics only: hosts. |
| 4082 | p.ReleaseRefusals = releaseRefusalProbe() |
| 4083 | // The rollout scheduler's report (zero-downtime spec §8.4): nil |
| 4084 | // with CARLOS_SWAP off. Private Metrics only. |
| 4085 | p.Rollout = rolloutStatusProbe() |
| 4086 | // Hosts held for a dev proof (D3): private Metrics only. |
| 4087 | p.ProofWaits = proofWaitProbe() |
| 4088 | // The build-state builds of the instances this box hosts (settings |
| 4089 | // spec rev 11 §4.4), for the housekeeper's retention. A failed |
| 4090 | // collection is reported as such: the housekeeper prunes nothing |
| 4091 | // while it is fresh. Private Metrics only: hosts. |
| 4092 | if pins, err := buildStates(); err != nil { |
| 4093 | log.Printf("agent: status tick: build states: %v", err) |
| 4094 | p.BuildStatesErr = err.Error() |
| 4095 | } else { |
| 4096 | p.BuildStates = pins |
| 4097 | } |
| 4098 | // The update-settings protocol this box enforces, and root's last |
| 4099 | // decision per instance (settings spec §4.3). Private Metrics only. |
| 4100 | fillUpdateProbes(&p, reg, release.OptDir()) |
| 4101 | |
| 4102 | // A fleet box (CARLOS_FLEET set) never gets here as a contender: it |
| 4103 | // self-reports only, never attempting the election at all — no |
| 4104 | // lease.Acquire call, no BuildFleet, no PublishFleet. See |
| 4105 | // shouldSkipPublisherElection and the package doc comment's |
| 4106 | // CARLOS_FLEET entry. |
| 4107 | var held *lease.Held |
| 4108 | publish := false |
| 4109 | if !skipElection { |
| 4110 | // path.Join (an S3 key, not a filesystem path — matches |
| 4111 | // activator.go's Config.hibernatePrefix convention) rather than a |
| 4112 | // bare "+" concat: guards against a trailing-slash |
| 4113 | // CARLOS_STATUS_PREFIX producing a doubled "//status-lease" key. |
| 4114 | var acqErr error |
| 4115 | held, acqErr = lease.Acquire(ctx, bucket, path.Join(prefix, "status-lease"), region) |
| 4116 | publish = acqErr == nil |
| 4117 | if acqErr != nil { |
| 4118 | var heldErr *lease.HeldError |
| 4119 | if !errors.As(acqErr, &heldErr) { |
| 4120 | log.Printf("agent: status tick: lease acquire: %v", acqErr) |
| 4121 | } |
| 4122 | } |
| 4123 | } |
| 4124 | |
| 4125 | // groupStatusTick is statusTick for every box but one its region's |
| 4126 | // group launched, which publishes only under its group reference |
| 4127 | // (census ruling region-ha-2b, boxreference.go). |
| 4128 | if err := groupStatusTick(ctx, store, reg, p, publish, fleetLabel, os.Getenv); err != nil { |
| 4129 | log.Printf("agent: status tick: %v", err) |
| 4130 | } |
| 4131 | |
| 4132 | if publish { |
| 4133 | if err := held.Release(ctx); err != nil { |
| 4134 | log.Printf("agent: status tick: lease release: %v", err) |
| 4135 | } |
| 4136 | } |
| 4137 | } |
| 4138 | } |
| 4139 | |
| 4140 | // statusTick is the fleet status page's pure per-tick work, extracted out of |
| 4141 | // statusLoop so it needs no systemd, S3, or timers to test: given an |
| 4142 | // already-built Store (DirStore in tests, the real S3Store in production) |
| 4143 | // and an already-decided publish flag (statusLoop's lease result; a fixed |
| 4144 | // true/false in tests), it does exactly what one tick does and nothing else. |
| 4145 | // The box's own snapshot is written unconditionally — every box self-reports |
| 4146 | // every tick regardless of who's publishing — and BuildFleet+PublishFleet |
| 4147 | // only run when publish is true, which for a DirStore-backed dev/test caller |
| 4148 | // with no lease in play at all just means "always true, no election needed" |
| 4149 | // (a lone box is its own only publisher). |
| 4150 | // |
| 4151 | // labelOverride, when non-empty, replaces the box's hashed-pseudonym label |
| 4152 | // (status.CollectBox derives one from p.MachineID by default) with the given |
| 4153 | // value before WriteBox — statusLoop passes a fleet box's channel label |
| 4154 | // (resolveFleetIdentity) here so its heartbeat lands at |
| 4155 | // fleets/<fleet>/status/boxes/<label>.json under the SAME label the edge |
| 4156 | // dials, not an opaque pseudonym only the public status page ever wanted. |
| 4157 | // "" (every platform box) leaves CollectBox's own label untouched. |
| 4158 | func statusTick(ctx context.Context, s release.Store, reg *registry.Registry, p status.Probes, publish bool, labelOverride string) error { |
| 4159 | routes, err := reg.List() |
| 4160 | if err != nil { |
| 4161 | return err |
| 4162 | } |
| 4163 | now := time.Now() |
| 4164 | |
| 4165 | // Aliases carries route_aliases' DNS/cert stamps (Task 2/5) up to the box |
| 4166 | // snapshot (issue #151) — read fresh alongside routes above rather than |
| 4167 | // folded into the caller's Probes in statusLoop, since reg is already in |
| 4168 | // scope here and a failed read must not sink the whole snapshot: logged |
| 4169 | // and left nil (CollectBox already treats nil the same as "no alias rows |
| 4170 | // exist"), same best-effort shape as every other probe in this loop. |
| 4171 | if aliases, err := reg.ListAliases(); err != nil { |
| 4172 | log.Printf("agent: status tick: list aliases: %v", err) |
| 4173 | } else { |
| 4174 | p.Aliases = aliases |
| 4175 | } |
| 4176 | |
| 4177 | box := status.CollectBox(routes, p, now) |
| 4178 | if labelOverride != "" { |
| 4179 | box.Label = labelOverride |
| 4180 | } |
| 4181 | if err := status.WriteBox(ctx, s, box); err != nil { |
| 4182 | return err |
| 4183 | } |
| 4184 | if !publish { |
| 4185 | return nil |
| 4186 | } |
| 4187 | |
| 4188 | // status.SnapshotPeriod is this loop's own tick cadence, named once in the |
| 4189 | // status package so the console's private host table and `carlos status |
| 4190 | // -live` measure staleness against the same period this publish does. |
| 4191 | fleet, err := status.BuildFleet(ctx, s, routes, now, status.SnapshotPeriod) |
| 4192 | if err != nil { |
| 4193 | return err |
| 4194 | } |
| 4195 | return status.PublishFleet(ctx, s, fleet) |
| 4196 | } |
| 4197 | |
| 4198 | // makeRestartRoute builds the adopt pass's restart callback, dispatched on |
| 4199 | // what owns the route's process. THREE arms, in this order: |
| 4200 | // |
| 4201 | // 1. a route that NAMES its own unit (registry.Route.Unit — an operator-run |
| 4202 | // bespoke service, console.service) is try-restarted by that name, and |
| 4203 | // returns. This arm comes first because it is the only one that can be |
| 4204 | // right when it applies: such a route is addr-backed, so the activator |
| 4205 | // owns no child to Bounce and there is no carlos-app@<host> unit to |
| 4206 | // try-restart — both of the arms below would run and cycle nothing, |
| 4207 | // which is issue #110 (adoption stamped Version, and therefore |
| 4208 | // X-Carlos-Version, behind a console still serving the old build). |
| 4209 | // 2. a carlos-app@<host> unit tenant is try-restarted by host (a no-op when |
| 4210 | // the unit isn't running — hibernators are never woken by this). |
| 4211 | // 3. an exec child is Bounced through the activator, the same graceful |
| 4212 | // cycle config sync uses. systemctl cannot reach a child of THIS |
| 4213 | // process, which is why exec-backed warm instances used to keep serving |
| 4214 | // the old binary until they happened to hibernate: the symlink was |
| 4215 | // repointed under them and nothing cycled the process. |
| 4216 | // |
| 4217 | // Hosts the registry cannot answer for take arm 2: try-restart of a |
| 4218 | // nonexistent unit is a no-op, whereas a bounce of a process we might not own |
| 4219 | // is not. |
| 4220 | // |
| 4221 | // A bounce that cycled NOTHING falls through to try-restart rather than |
| 4222 | // returning: an exec-backed route with an ACTIVE carlos-app@<host> unit is a |
| 4223 | // real state — an unstamped unit tenant, the one adopt_advice.go warns about |
| 4224 | // (issue #14) — and try-restart is what has always cycled it. The same |
| 4225 | // fallthrough covers a box with no activator armed at all. When the route |
| 4226 | // really is a sleeping exec child the cost is one try-restart of a stopped or |
| 4227 | // absent unit, which starts nothing and at worst logs once: this callback |
| 4228 | // only fires when the version actually moves (adoptHost early-returns on |
| 4229 | // cur == ver), so it is one line per version change per host, never per tick. |
| 4230 | // |
| 4231 | // The named-unit arm takes NO part in that fallthrough, in either direction. |
| 4232 | // It does not fall through (a route that names its unit has said where its |
| 4233 | // process is; trying the other two afterwards could only cycle something |
| 4234 | // else), and nothing falls through INTO it (the other arms are reached only |
| 4235 | // when Unit is empty, i.e. when there is no name to try). |
| 4236 | // |
| 4237 | // No path can WAKE anything, which is what lets release.Pass hand this |
| 4238 | // callback hibernating routes (RestartHibernating): Bounce returns false |
| 4239 | // without starting anything when the host is not awake, and try-restart is |
| 4240 | // defined as a no-op on a unit that is not running — including the named one, |
| 4241 | // which is why it is try-restart and not restart (a console an operator has |
| 4242 | // deliberately stopped stays stopped). |
| 4243 | // |
| 4244 | // ERRORS split by arm, and the split is the point: |
| 4245 | // |
| 4246 | // Arms 2 and 3 log and return nil, unchanged: a failed try-restart of a |
| 4247 | // carlos-app@<host> unit must not poison the adopted version for the other |
| 4248 | // hosts in the pass (release/adopt.go's rule), and it cannot be told apart |
| 4249 | // from the ordinary case anyway — a sleeping exec child reaches arm 2 under |
| 4250 | // RestartHibernating precisely because there is nothing there to cycle. |
| 4251 | // Fire-and-forget is what makes that safe. |
| 4252 | // |
| 4253 | // Arm 1 RETURNS the failure, because for a named unit the ordinary case is |
| 4254 | // not indistinguishable: try-restart of a unit that is loaded and stopped |
| 4255 | // exits 0, so a non-zero exit means the name is wrong (exit 5, "not loaded") |
| 4256 | // or the service would not come back. Both are "the new build is NOT |
| 4257 | // running", and swallowing them would hand adoptHost a success it did not |
| 4258 | // have: it would stamp Version, the edge would serve the new |
| 4259 | // X-Carlos-Version, and `carlos deploy` would report live over the old |
| 4260 | // process — issue #110 again, now with the stamp actively lying rather than |
| 4261 | // merely unattended. Returned, adoptHost rolls the symlink back to the |
| 4262 | // version that IS running, skips SetVersions, prints the failure, and the |
| 4263 | // next pass retries; the header stays honest on the old version throughout. |
| 4264 | func makeRestartRoute(reg *registry.Registry, tryRestart func(host string), tryRestartUnit func(rt registry.Route) error, bounce func(host string) bool) func(host string) error { |
| 4265 | return func(host string) error { |
| 4266 | rt, err := reg.Get(host) |
| 4267 | if err != nil { |
| 4268 | log.Printf("adopt: registry lookup %s: %v (restarting as a unit)", host, err) |
| 4269 | } |
| 4270 | if err == nil && rt != nil && rt.Unit != "" { |
| 4271 | return tryRestartUnit(*rt) |
| 4272 | } |
| 4273 | if err == nil && rt != nil && rt.Backing != registry.BackingUnit { |
| 4274 | if bounce != nil && bounce(host) { |
| 4275 | log.Printf("adopt: bounced exec instance %s onto its new version", host) |
| 4276 | return nil |
| 4277 | } |
| 4278 | } |
| 4279 | tryRestart(host) |
| 4280 | return nil |
| 4281 | } |
| 4282 | } |
| 4283 | |
| 4284 | // makeStartRoute builds the adopt pass's rollback-start callback |
| 4285 | // (release.WithStartUnit): after a restart that FAILED has had its symlink |
| 4286 | // rolled back, start the unit the route NAMES on the restored version — |
| 4287 | // `systemctl start`, not try-restart, because the premise of that path is |
| 4288 | // that the unit is down (a try-restart there is a no-op and the box stays |
| 4289 | // serving nothing). |
| 4290 | // |
| 4291 | // ONE arm, deliberately, where makeRestartRoute has three: only a route that |
| 4292 | // names its own unit has anything for this callback to start. The other two |
| 4293 | // process owners never route here at all — their restart arms swallow their |
| 4294 | // failures and return nil, so the rollback path this callback belongs to is |
| 4295 | // unreachable for them — and a start of carlos-app@<host> or an activator |
| 4296 | // child from here could WAKE a sleeper, which no adopt path may do. A route |
| 4297 | // with an empty Unit, an unknown host, or a registry that will not answer |
| 4298 | // therefore starts NOTHING and returns nil: for those shapes "nothing was |
| 4299 | // stopped by the failed restart" is the only answer this callback can stand |
| 4300 | // over. (The converged-daemon rollback is caps.Restore's business, not |
| 4301 | // this callback's — see release.WithStartUnit.) |
| 4302 | // |
| 4303 | // This cannot wake anything, twice over. By ARGUMENT: it fires only after a |
| 4304 | // try-restart of the named unit returned non-zero, and try-restart exits 0 |
| 4305 | // on a loaded-but-stopped unit — so a deliberately stopped service never |
| 4306 | // produces the error that reaches this, and a unit whose name resolves to |
| 4307 | // nothing gives systemctl start nothing to bring up either. And by |
| 4308 | // STRUCTURE: a Hibernate-flagged route starts nothing here, full stop. That |
| 4309 | // guard costs nothing real — provision refuses the addr-with-hibernate |
| 4310 | // shape, so no legitimate named-unit route hibernates, and the only |
| 4311 | // hibernating+Unit row is the misconfiguration restart.go already documents |
| 4312 | // — and it turns "an adopt path cannot wake a sleeper" back into a property |
| 4313 | // of the code's shape rather than of an argument about systemd exit codes. |
| 4314 | func makeStartRoute(reg *registry.Registry, startUnit func(unit string) error) func(host string) error { |
| 4315 | return func(host string) error { |
| 4316 | rt, err := reg.Get(host) |
| 4317 | if err != nil || rt == nil || rt.Hibernate || rt.Unit == "" { |
| 4318 | return nil |
| 4319 | } |
| 4320 | return startUnit(rt.Unit) |
| 4321 | } |
| 4322 | } |
| 4323 | |
| 4324 | // systemctlTryRestart is makeRestartRoute's production unit path: cycle the |
| 4325 | // route's carlos-app@ unit iff it's running, best-effort. A missing or |
| 4326 | // stopped unit is not this pass's problem — try-restart is by definition a |
| 4327 | // no-op on one, which is what makes makeRestartRoute's fallthrough safe: |
| 4328 | // an exec route whose child is asleep returns false from bounce (nothing to |
| 4329 | // cycle) and lands here under RestartHibernating, where try-restart finds no |
| 4330 | // carlos-app@ unit for it and does nothing. Sleeping hosts DO reach this |
| 4331 | // line; they simply have nothing here to restart, and pick the new version |
| 4332 | // up on their next wake. |
| 4333 | func systemctlTryRestart(host string) { |
| 4334 | out, err := exec.Command("systemctl", "try-restart", activator.UnitName(host)).CombinedOutput() |
| 4335 | if err != nil { |
| 4336 | log.Printf("adopt: try-restart carlos-app@%s: %v: %s (continuing)", host, err, strings.TrimSpace(string(out))) |
| 4337 | } |
| 4338 | } |
| 4339 | |
| 4340 | // systemctlTryRestartUnit is makeRestartRoute's NAMED-unit path: cycle the |
| 4341 | // unit the route itself names, iff it's running. Same try-restart-not-restart |
| 4342 | // reasoning as systemctlTryRestart above — a service an operator has stopped |
| 4343 | // stays stopped, and try-restart on a loaded-but-stopped unit exits 0. |
| 4344 | // |
| 4345 | // It LOGS AND RETURNS, where its carlos-app@ sibling only logs, and the |
| 4346 | // asymmetry is deliberate: this name came from an operator rather than from |
| 4347 | // activator.UnitName, so a non-zero exit here means the stamp is wrong or the |
| 4348 | // service will not come back — never the routine "nothing was running" |
| 4349 | // systemctlTryRestart has to tolerate. The journal line is kept as well as |
| 4350 | // the error because the error's own destination is a pass that prints one |
| 4351 | // summary line per host; the exit status and systemctl's own words belong in |
| 4352 | // the journal where an operator greps for the unit name. |
| 4353 | // |
| 4354 | // The unit reaches exec.Command as ONE argv element, so there is no shell for |
| 4355 | // a name to escape into; the shape rail lives at the write (registry's |
| 4356 | // ValidUnitName), where a typo can still be reported to the person who typed |
| 4357 | // it. |
| 4358 | func systemctlTryRestartUnit(unit string) error { |
| 4359 | out, err := exec.Command("systemctl", "try-restart", unit).CombinedOutput() |
| 4360 | if err != nil { |
| 4361 | log.Printf("adopt: try-restart %s: %v: %s", unit, err, strings.TrimSpace(string(out))) |
| 4362 | return fmt.Errorf("try-restart %s: %w: %s", unit, err, strings.TrimSpace(string(out))) |
| 4363 | } |
| 4364 | return nil |
| 4365 | } |
| 4366 | |
| 4367 | // unitOnly adapts a unit-name restart to makeRestartRoute's route-aware arm. |
| 4368 | func unitOnly(f func(unit string) error) func(rt registry.Route) error { |
| 4369 | return func(rt registry.Route) error { return f(rt.Unit) } |
| 4370 | } |
| 4371 | |
| 4372 | // checkedUnitRestart is the named-unit arm with D1's settle check (spec |
| 4373 | // §2.2-§2.3). The binding is verified first: the check judges a restart only |
| 4374 | // when the unit's ExecStart runs THIS route's live link. Anything it cannot |
| 4375 | // establish (show denied, wrong binding, unit stopped) falls back to today's |
| 4376 | // plain try-restart and returns that command's own error. |
| 4377 | // |
| 4378 | // settled runs only when the settle check passed (never for a unit stopped |
| 4379 | // during it, a failed settle or an unchecked fallback). The agent loops pass |
| 4380 | // nil: the sandboxed agent cannot hash another unit's executable, and spec |
| 4381 | // §4.5 has root record the healthy history. |
| 4382 | func agentCheckedUnitRestart(sys unitSystem, optDir string, getenv func(string) string, settled func(rt registry.Route) error) func(rt registry.Route) error { |
| 4383 | return func(rt registry.Route) error { |
| 4384 | before, err := sys.Show(rt.Unit) |
| 4385 | if err != nil { |
| 4386 | log.Printf("adopt: %s: cannot read %s (%v); restart not checked", rt.Host, rt.Unit, err) |
| 4387 | return sys.TryRestart(rt.Unit) |
| 4388 | } |
| 4389 | link := filepath.Join(optDir, "live", rt.Host) |
| 4390 | if before.ExecPath != link { |
| 4391 | log.Printf("adopt: binding: %s runs %s, but the unit stamp is on %s; restart not checked", rt.Unit, before.ExecPath, rt.Host) |
| 4392 | return sys.TryRestart(rt.Unit) |
| 4393 | } |
| 4394 | if before.ActiveState != "active" { |
| 4395 | return sys.TryRestart(rt.Unit) // stopped by an operator: stays a no-op |
| 4396 | } |
| 4397 | want, err := filepath.EvalSymlinks(link) |
| 4398 | if err != nil { |
| 4399 | return sys.TryRestart(rt.Unit) |
| 4400 | } |
| 4401 | if err := sys.TryRestart(rt.Unit); err != nil { |
| 4402 | return err |
| 4403 | } |
| 4404 | timeout := unitSettleTimeout(getenv) |
| 4405 | err = settleUnit(sys, rt.Unit, before.InvocationID, settleOpts{ |
| 4406 | Timeout: timeout, Poll: 500 * time.Millisecond, Hold: 5 * time.Second, |
| 4407 | WantDir: filepath.Dir(want), Ready: routeReady(rt, 2*time.Second), |
| 4408 | Now: time.Now, Sleep: time.Sleep}) |
| 4409 | if errors.Is(err, errUnitStopped) { |
| 4410 | // An operator stopped it between the read above and now. Not a |
| 4411 | // verdict on the build, and a rollback restart would undo their |
| 4412 | // stop: same as a unit that was stopped before the restart. |
| 4413 | log.Printf("adopt: %s: %s was stopped during settle; not judged, not restarted", rt.Host, rt.Unit) |
| 4414 | return nil |
| 4415 | } |
| 4416 | if err == nil { |
| 4417 | log.Printf("adopt: %s: %s settled on %s", rt.Host, rt.Unit, filepath.Dir(want)) |
| 4418 | // D3 (proofobserve.go): a build that passed the settle check is a |
| 4419 | // build this box has run well. |
| 4420 | if settled != nil { |
| 4421 | err = settled(rt) |
| 4422 | } |
| 4423 | } |
| 4424 | return err |
| 4425 | } |
| 4426 | } |
| 4427 | |
| 4428 | // rollbackUnitRestart is D1's rollback restart (release.WithRollbackRestart): |
| 4429 | // the live link already names the restored build; replace whatever is |
| 4430 | // running (a deadline failure can leave the rejected build up) and check |
| 4431 | // the restored build comes up too. |
| 4432 | func agentRollbackUnitRestart(reg *registry.Registry, sys unitSystem, optDir string, getenv func(string) string) func(host string) error { |
| 4433 | return func(host string) error { |
| 4434 | rt, err := reg.Get(host) |
| 4435 | if err != nil || rt == nil || rt.Unit == "" { |
| 4436 | return nil |
| 4437 | } |
| 4438 | before, _ := sys.Show(rt.Unit) |
| 4439 | if err := sys.ResetFailed(rt.Unit); err != nil { |
| 4440 | log.Printf("adopt: %s: reset-failed %s: %v (continuing)", host, rt.Unit, err) |
| 4441 | } |
| 4442 | if err := sys.Restart(rt.Unit); err != nil { |
| 4443 | return err |
| 4444 | } |
| 4445 | want, err := filepath.EvalSymlinks(filepath.Join(optDir, "live", host)) |
| 4446 | if err != nil { |
| 4447 | return err |
| 4448 | } |
| 4449 | err = settleUnit(sys, rt.Unit, before.InvocationID, settleOpts{ |
| 4450 | Timeout: unitSettleTimeout(getenv), Poll: 500 * time.Millisecond, Hold: 5 * time.Second, |
| 4451 | WantDir: filepath.Dir(want), Ready: routeReady(*rt, 2*time.Second), |
| 4452 | Now: time.Now, Sleep: time.Sleep}) |
| 4453 | if errors.Is(err, errUnitStopped) { |
| 4454 | // The restored build is what the link names; a stop an operator |
| 4455 | // made while it settled is theirs to keep, not a failed recovery. |
| 4456 | log.Printf("adopt: %s: %s was stopped during the rollback's settle; not judged, not restarted", host, rt.Unit) |
| 4457 | return nil |
| 4458 | } |
| 4459 | if err != nil { |
| 4460 | return fmt.Errorf("rollback did not come up either: %w", err) |
| 4461 | } |
| 4462 | return nil |
| 4463 | } |
| 4464 | } |
| 4465 | |
| 4466 | // systemctlStartUnit is makeStartRoute's production path: bring the named |
| 4467 | // unit up after a rolled-back restart left it down. START, not try-restart — |
| 4468 | // try-restart is a no-op on a stopped unit, which on this path would leave |
| 4469 | // the box with a correct symlink and no process, and hand the NEXT pass's |
| 4470 | // try-restart an exit 0 to stamp a never-run build with. Idempotent where it |
| 4471 | // matters: `systemctl start` of an already-running unit exits 0 and does |
| 4472 | // nothing, so a service something else restored first is not cycled twice. |
| 4473 | // |
| 4474 | // LOGS AND RETURNS like systemctlTryRestartUnit, and for the same reason: |
| 4475 | // this name came from an operator, its failure means the box is serving |
| 4476 | // nothing, and the caller's one summary line per host must say so while the |
| 4477 | // exit status and systemd's own words stay in the journal under the unit |
| 4478 | // name an operator would grep for. Same no-shell shape too: the unit reaches |
| 4479 | // exec.Command as ONE argv element. |
| 4480 | func systemctlStartUnit(unit string) error { |
| 4481 | out, err := exec.Command("systemctl", "start", unit).CombinedOutput() |
| 4482 | if err != nil { |
| 4483 | log.Printf("adopt: start %s: %v: %s", unit, err, strings.TrimSpace(string(out))) |
| 4484 | return fmt.Errorf("start %s: %w: %s", unit, err, strings.TrimSpace(string(out))) |
| 4485 | } |
| 4486 | return nil |
| 4487 | } |
| 4488 | |
| 4489 | // rootRestartRoute is the root adopt pass's restart callback (RunAdopt, out of |
| 4490 | // carlos-adopt.service): no bounce. That process has no activator and no |
| 4491 | // children of its own — the exec instances belong to the agent, and another |
| 4492 | // process's children are not its to stop. With bounce nil the dispatch's arm |
| 4493 | // 3 is unreachable and every route it used to try-restart by host still is, |
| 4494 | // byte-identically; what it gains is arm 1. |
| 4495 | // |
| 4496 | // Arm 1 has to be HERE, not only on the agent's in-process loops, and the |
| 4497 | // reason is the sandbox. carlos-edge.service runs as User=carlos with |
| 4498 | // ProtectSystem=strict and ReadWritePaths=/data /run/carlos, so the agent |
| 4499 | // cannot write /opt/carlos: on a real box its pass reports the stage as |
| 4500 | // blocked and relays to this verb (carlos-stage.path), which does the |
| 4501 | // staging, the repoint, the restart and the version stamp. A restart arm |
| 4502 | // wired only into the agent would therefore be a restart arm the consoles |
| 4503 | // never reach. |
| 4504 | // |
| 4505 | // settled is checkedUnitRestart's hook for a named unit whose settle check |
| 4506 | // passed (D3: recordSettled, which adds the build to the healthy history); |
| 4507 | // nil records nothing. |
| 4508 | // bounceFn adapts an activator to makeRestartRoute's bounce parameter, |
| 4509 | // nil-safe: a box with no activator armed gets a nil callback, which |
| 4510 | // makeRestartRoute reads as "nothing running to cycle" rather than falling |
| 4511 | // back to a systemctl path that could not work for an exec route anyway. |
| 4512 | func bounceFn(svc *activator.Service) func(host string) bool { |
| 4513 | if svc == nil { |
| 4514 | return nil |
| 4515 | } |
| 4516 | return svc.Bounce |
| 4517 | } |
| 4518 | |
| 4519 | // adoptBounceFn is bounceFn for an ADOPT pass's restart dispatch |
| 4520 | // (makeRestartRoute on adoptLoop and fastConvergeLoop): the same bounce, |
| 4521 | // except that a child already exec'd from the build live/<host> names is left |
| 4522 | // running — see activator.BounceAdopted for the interleaving that produces |
| 4523 | // one. `carlos restart` keeps bounceFn: a restart request cycles whatever is |
| 4524 | // running. |
| 4525 | func adoptBounceFn(svc *activator.Service) func(host string) bool { |
| 4526 | if svc == nil { |
| 4527 | return nil |
| 4528 | } |
| 4529 | if m := armedRollout(); m != nil { |
| 4530 | return m.Adopted |
| 4531 | } |
| 4532 | if rolloutUnavailable() != "" { |
| 4533 | return nil // no adopt bounce while managed starts are refused |
| 4534 | } |
| 4535 | if h := heldHolds(); h != nil { |
| 4536 | // Declined with failure records: a host a record holds keeps |
| 4537 | // running rather than stop into a refused wake (final review F6). |
| 4538 | return h.skipHost("adopt bounce", svc.BounceAdopted) |
| 4539 | } |
| 4540 | return svc.BounceAdopted |
| 4541 | } |
| 4542 | |
| 4543 | // wakeFn adapts an activator to cycleWorkers' wake parameter, nil-safe like |
| 4544 | // bounceFn: a box with no activator armed gets a nil callback, which |
| 4545 | // cycleWorkers reads as "nothing to wake" rather than panicking on a nil |
| 4546 | // activator. |
| 4547 | func wakeFn(svc *activator.Service) func(registry.Route) error { |
| 4548 | if svc == nil { |
| 4549 | return nil |
| 4550 | } |
| 4551 | return svc.WakeWorker |
| 4552 | } |
| 4553 | |
| 4554 | // openDeployStore opens the release store the box adopts from: |
| 4555 | // CARLOS_DEPLOYMENT_BUCKET (S3), CARLOS_DEPLOYMENT_DIR (a directory — tests |
| 4556 | // and single-box no-AWS deployments), or — on a fleet box with neither set — |
| 4557 | // the edge's store door over the tunnel env (spec F5, openFleetDoorStore). |
| 4558 | // Mirrors cli.go's openStore, but returns the error instead of dying: the |
| 4559 | // agent degrades, the CLI exits. |
| 4560 | func openDeployStore(ctx context.Context) (release.Store, error) { |
| 4561 | return openDeployStoreEnv(ctx, os.Getenv) |
| 4562 | } |
| 4563 | |
| 4564 | // OpenDeployStore is openDeployStore for callers outside this package. The |
| 4565 | // relay reads its authority record and writes its heartbeat and metering to |
| 4566 | // the same store the agent adopts from, chosen by the same rule, so a box |
| 4567 | // that has moved onto the fleet door does not leave the relay behind on a |
| 4568 | // bucket it can no longer reach. |
| 4569 | func OpenDeployStore(ctx context.Context) (release.Store, error) { return openDeployStore(ctx) } |
| 4570 | |
| 4571 | // openDeployStoreEnv is openDeployStore with the environment injected — the |
| 4572 | // same getenv seam resolveStatusBucket already uses in this file, so the |
| 4573 | // selection rule can be tested without mutating process-wide env. |
| 4574 | // |
| 4575 | // Order is the whole contract (spec F5): a bucket or dir still set WINS, |
| 4576 | // before fleet identity is even consulted, so every platform box and every |
| 4577 | // grandfathered fleet box that still has bucket credentials keeps exactly |
| 4578 | // today's store, byte for byte. Only a box that resolves a fleet identity |
| 4579 | // AND has no bucket/dir at all reaches the door — that is the cutover |
| 4580 | // switch, per box, no flag day: drop the bucket var and the box is on the |
| 4581 | // API; put it back and it is off. |
| 4582 | func openDeployStoreEnv(ctx context.Context, getenv func(string) string) (release.Store, error) { |
| 4583 | if b := getenv("CARLOS_DEPLOYMENT_BUCKET"); b != "" { |
| 4584 | return release.NewDeploymentS3Store(ctx, b) |
| 4585 | } |
| 4586 | if d := getenv("CARLOS_DEPLOYMENT_DIR"); d != "" { |
| 4587 | ds := release.NewDirStore(d) |
| 4588 | ds.EnableChangeIndex() |
| 4589 | return ds, nil |
| 4590 | } |
| 4591 | identity, err := resolveFleetIdentity(getenv) |
| 4592 | switch { |
| 4593 | case err == nil: |
| 4594 | return openFleetDoorStore(getenv, identity) |
| 4595 | case !errors.Is(err, errNotFleetBox): |
| 4596 | // CARLOS_FLEET is set but the identity is unusable, and there is no |
| 4597 | // bucket to fall back to: this box has no store at all. Return the |
| 4598 | // identity error itself rather than the generic "no bucket" line — |
| 4599 | // the actual fix is the label, and the generic message would send an |
| 4600 | // operator hunting for a bucket var that is deliberately absent. |
| 4601 | return nil, err |
| 4602 | } |
| 4603 | return nil, errNoDeployStore |
| 4604 | } |
| 4605 | |
| 4606 | // openFleetDoorStore builds the fleetapi client a bucketless fleet box adopts |
| 4607 | // through: the edge origin it already tunnels to (CARLOS_TUNNEL_URL), its |
| 4608 | // channel identity ("<fleet>/<label>" — the same string tunnel_client.go |
| 4609 | // registers with and edge.fleetChannelFor dials), and the tunnel bearer token. |
| 4610 | // |
| 4611 | // Missing tunnel env here is an ERROR, never a silent no-store: the caller |
| 4612 | // established that this box is a fleet box with no bucket, so there is no |
| 4613 | // other store it could possibly be meant to use, and a quiet nil would leave |
| 4614 | // every loop in this file reporting "not configured" about a box that is in |
| 4615 | // fact configured — just with one variable missing. |
| 4616 | func openFleetDoorStore(getenv func(string) string, identity FleetIdentity) (release.Store, error) { |
| 4617 | base := strings.TrimSpace(getenv("CARLOS_TUNNEL_URL")) |
| 4618 | token := strings.TrimSpace(getenv("CARLOS_TUNNEL_TOKEN")) |
| 4619 | name := identity.Fleet + "/" + identity.Label |
| 4620 | if base == "" || token == "" { |
| 4621 | return nil, fmt.Errorf("fleet box %s has no deployment bucket and cannot reach the store door: both CARLOS_TUNNEL_URL and CARLOS_TUNNEL_TOKEN must be set", name) |
| 4622 | } |
| 4623 | return fleetapi.New(base, name, token) |
| 4624 | } |
| 4625 | |
| 4626 | // resolveStatusPrefix returns the prefix every key statusLoop and metricsLoop |
| 4627 | // touch is rebased under (release.NewPrefixStore at both call sites). |
| 4628 | // |
| 4629 | // An explicit CARLOS_STATUS_PREFIX always wins, unchanged, on every box. What |
| 4630 | // this adds is the DOOR-branch default: a fleet box publishing through the |
| 4631 | // store door with an EMPTY prefix would otherwise write status/boxes/... and |
| 4632 | // metrics/boxes/... at the shared platform root, and the door's write ACL |
| 4633 | // admits only fleets/<own fleet>/** — every heartbeat and every flush would |
| 4634 | // 404, forever, on a box that looks correctly configured. So on that branch |
| 4635 | // alone the prefix defaults to "fleets/<fleet>/", which is exactly the layout |
| 4636 | // the console's fleet page already reads. |
| 4637 | // |
| 4638 | // S3-path semantics are deliberately untouched: an empty prefix on a |
| 4639 | // bucket-backed box is the platform box's shared root, which is correct |
| 4640 | // there, and a bucket-backed fleet box keeps whatever the operator set (the |
| 4641 | // customer-fleets rollout sets CARLOS_STATUS_PREFIX=fleets/<name>/ by hand). |
| 4642 | func resolveStatusPrefix(getenv func(string) string) string { |
| 4643 | if p := getenv("CARLOS_STATUS_PREFIX"); p != "" { |
| 4644 | return p |
| 4645 | } |
| 4646 | if resolveStatusBucket(getenv) != "" { |
| 4647 | return "" |
| 4648 | } |
| 4649 | identity, err := resolveFleetIdentity(getenv) |
| 4650 | if err != nil { |
| 4651 | return "" |
| 4652 | } |
| 4653 | return "fleets/" + identity.Fleet + "/" |
| 4654 | } |
| 4655 | |
| 4656 | // errNoStatusStore is openStatusStore's "this deployment simply hasn't |
| 4657 | // configured one" sentinel — distinct from a real construction failure, so |
| 4658 | // each loop can keep its own long-standing disabled-log wording for the |
| 4659 | // unconfigured case while sharing one selection rule. |
| 4660 | var errNoStatusStore = errors.New("no status store configured") |
| 4661 | |
| 4662 | // openStatusStore resolves the store statusLoop and metricsLoop publish |
| 4663 | // through, plus a human phrase naming where that is for their armed log |
| 4664 | // lines. Same F5 rule as openDeployStoreEnv, read through |
| 4665 | // resolveStatusBucket: an explicit CARLOS_STATUS_BUCKET (else the |
| 4666 | // CARLOS_DEPLOYMENT_BUCKET fallback) WINS and keeps direct S3; only a fleet |
| 4667 | // box with neither set speaks the door. |
| 4668 | // |
| 4669 | // LEDGER NOTE (spec F5): the door serves the DEPLOYMENT store. A fleet box |
| 4670 | // that sets CARLOS_STATUS_BUCKET to some OTHER bucket therefore keeps direct |
| 4671 | // S3 for status/metrics while its releases may come through the door — that |
| 4672 | // is the bucket-set-wins rule doing its job, not a bug: an operator who named |
| 4673 | // a bucket explicitly gets that bucket. |
| 4674 | // |
| 4675 | // The CARLOS_STATUS_PREFIX rebase both loops apply (release.NewPrefixStore) |
| 4676 | // is deliberately NOT applied here — it stays at each call site, unchanged, |
| 4677 | // and wraps whichever store this returns. |
| 4678 | func openStatusStore(ctx context.Context, getenv func(string) string) (release.Store, string, error) { |
| 4679 | if bucket := resolveStatusBucket(getenv); bucket != "" { |
| 4680 | // The fallback is the deployment bucket, so open it as one; status/ and |
| 4681 | // metrics/ are not change scopes and bump nothing. |
| 4682 | s, err := release.NewDeploymentS3Store(ctx, bucket) |
| 4683 | if err != nil { |
| 4684 | return nil, "", err |
| 4685 | } |
| 4686 | return s, "bucket " + bucket, nil |
| 4687 | } |
| 4688 | identity, err := resolveFleetIdentity(getenv) |
| 4689 | switch { |
| 4690 | case err == nil: |
| 4691 | s, derr := openFleetDoorStore(getenv, identity) |
| 4692 | if derr != nil { |
| 4693 | return nil, "", derr |
| 4694 | } |
| 4695 | return s, "the fleet store door as " + identity.Fleet + "/" + identity.Label, nil |
| 4696 | case !errors.Is(err, errNotFleetBox): |
| 4697 | return nil, "", err |
| 4698 | } |
| 4699 | return nil, "", errNoStatusStore |
| 4700 | } |
| 4701 | |
| 4702 | // RunAdopt is the `carlos adopt` verb: one immediate adoption pass — the |
| 4703 | // carlos-adopt.timer unit's ExecStart, and an operator's "converge now". |
| 4704 | // Runs as root (see carlos-adopt.service): this is the ONLY place the |
| 4705 | // UNIT-BACKED half of config sync (appconfig.Sync) may run, because it |
| 4706 | // writes /etc/carlos/apps and restarts units — the same filesystem-write |
| 4707 | // shape release staging needs root for, and exactly what the sandboxed |
| 4708 | // carlos-edge.service's in-process adopt loop (ProtectSystem=strict, |
| 4709 | // ReadWritePaths=/data /run/carlos only) cannot do. The EXEC-BACKED half |
| 4710 | // (appconfig.SyncExec) runs elsewhere, in the agent's in-process |
| 4711 | // configSyncLoop: it needs neither root nor /etc access, because it only |
| 4712 | // ever writes under /data and bounces its own children. |
| 4713 | // |
| 4714 | // -config-only runs the UNIT-BACKED CONFIG half ALONE: appconfig.Sync plus |
| 4715 | // its restarts, no release.Pass, no canary nag, no backing advice. That is |
| 4716 | // what carlos-config-sync.service execs, and it exists because the trigger |
| 4717 | // behind it is a per-config-write path unit, not the hourly timer. Letting |
| 4718 | // that trigger run a full adoption pass would mean one tenant's `carlos env |
| 4719 | // set` could stage and cut over a DIFFERENT app's pending release within |
| 4720 | // seconds of an unrelated write, and would fire the canary-TTL nag per |
| 4721 | // config write instead of hourly — the "hourly root pass alone" split |
| 4722 | // release/adopt.go and the nag comment below both depend on. |
| 4723 | // |
| 4724 | // -restart-only runs the RESTART-REQUEST half alone (issue #113): compare each |
| 4725 | // app's restart generation against this box's marks and cycle the ones that |
| 4726 | // moved, nothing else. That is what carlos-restart.service execs, and it is |
| 4727 | // separated from -config-only for the same reason -config-only is separated |
| 4728 | // from the full pass: the trigger behind it is a per-request path unit, and a |
| 4729 | // request to cycle one app's processes must not stage a release, run the |
| 4730 | // capability converge, or nag about canaries. It deliberately needs NO release |
| 4731 | // key and NO secrets key — restarting verifies no signature and renders no |
| 4732 | // config — so a box missing either still honours it. |
| 4733 | func RunAdopt(args []string) { runAdopt(args, productionClaimFactory) } |
| 4734 | |
| 4735 | func runAdopt(args []string, buildClaims ClaimFactory) { |
| 4736 | runAdoptWithEnsureSystemUnits(args, buildClaims, func() { |
| 4737 | ensureSystemUnits("/etc", filepath.Join(systemRootDir(release.OptDir()), systemLockFile), productionSysOps().systemctl, systemctlProbe, log.Printf) |
| 4738 | }) |
| 4739 | } |
| 4740 | |
| 4741 | // runAdoptWithEnsureSystemUnits keeps the root-only unit effect injectable; |
| 4742 | // the rest of adopt startup, including its fatal registry diagnostics, is real. |
| 4743 | func runAdoptWithEnsureSystemUnits(args []string, buildClaims ClaimFactory, ensureUnits func(), sources ...adoptRootSource) { |
| 4744 | open := func() (rootstate.Store, error) { return rootstate.Open(rootReleasePrivate, true, false) } |
| 4745 | probe := rootparticipation.OpenForUpdate |
| 4746 | if len(sources) == 1 { |
| 4747 | open = sources[0].OpenRelease |
| 4748 | probe = sources[0].Census |
| 4749 | } |
| 4750 | err := withClaimRootRelease(context.Background(), hostwrite.Enabled(os.Getenv), open, func() error { |
| 4751 | return runAdoptUnderRootRelease(args, buildClaims, ensureUnits, probe, productionRootUpdates(dataDir(), probe)) |
| 4752 | }, probe) |
| 4753 | if err != nil { |
| 4754 | log.Fatalf("adopt: %v", err) |
| 4755 | } |
| 4756 | } |
| 4757 | |
| 4758 | func runAdoptUnderRootRelease(args []string, buildClaims ClaimFactory, ensureUnits func(), census func() (rootparticipation.Ledger, bool, error), updateProviders ...release.RootUpdateOpener) error { |
| 4759 | reapStagingWrites(release.OptDir(), dataDir(), "/etc", log.Printf) |
| 4760 | if os.Geteuid() == 0 { |
| 4761 | if err := repairOffcloudCredentials("/etc/carlos/credentials"); err != nil { |
| 4762 | log.Printf("ALERT: adopt: offcloud credentials: %v", err) |
| 4763 | } |
| 4764 | } |
| 4765 | claimCtx, err := claimEntryContext(context.Background(), os.Getenv, openDeployStore) |
| 4766 | if err != nil { |
| 4767 | log.Fatalf("adopt: host claims: %v", err) |
| 4768 | } |
| 4769 | claimsEnabled := hostwrite.Enabled(claimEntryEnv(claimCtx)) |
| 4770 | var claims *ClaimWiring |
| 4771 | if !claimsEnabled { |
| 4772 | // Keep the flag-off startup effects in their old order: |
| 4773 | // disabled wiring, flags, pass lock, unit repair, registry open. |
| 4774 | claims, err = newClaimWiring(claimCtx, claimEntryEnv(claimCtx), buildClaims) |
| 4775 | if err != nil { |
| 4776 | log.Fatalf("adopt: host claims: %v", err) |
| 4777 | } |
| 4778 | } |
| 4779 | fs := flag.NewFlagSet("adopt", flag.ExitOnError) |
| 4780 | regPath := fs.String("registry", filepath.Join(dataDir(), "registry.db"), "registry database") |
| 4781 | configOnly := fs.Bool("config-only", false, "sync unit-backed config and restart changed routes only: no release adoption, no canary nag") |
| 4782 | stageOnly := fs.Bool("stage-only", false, "adopt releases without the canary nag: the agent-requested pass (carlos-stage.service)") |
| 4783 | restartOnly := fs.Bool("restart-only", false, "cycle the routes of apps whose restart generation moved, and nothing else: the agent-requested pass (carlos-restart.service)") |
| 4784 | fs.Usage = cliflag.Usage(fs, "") |
| 4785 | _ = fs.Parse(args) |
| 4786 | var reg *registry.Registry |
| 4787 | if claimsEnabled { |
| 4788 | reg, err = openServiceRegistry(*regPath) |
| 4789 | if err != nil { |
| 4790 | if hint := registryOpenHint(*regPath, err); hint != "" { |
| 4791 | log.Fatalf("adopt: open registry %s: %v\nadopt: %s", *regPath, err, hint) |
| 4792 | } |
| 4793 | log.Fatalf("adopt: open registry %s: %v", *regPath, err) |
| 4794 | } |
| 4795 | defer reg.Close() |
| 4796 | claimCtx = withClaimRuntimeResources(claimCtx, ClaimRuntimeResources{Mode: "adopt", Registry: reg, DataDir: dataDir(), OptDir: release.OptDir(), UnitsDir: "/etc/systemd/system"}) |
| 4797 | claims, err = newClaimWiring(claimCtx, claimEntryEnv(claimCtx), buildClaims) |
| 4798 | if err != nil { |
| 4799 | log.Fatalf("adopt: host claims: %v", err) |
| 4800 | } |
| 4801 | } |
| 4802 | defer func() { |
| 4803 | ctx, cancel := context.WithTimeout(context.Background(), hostbox.RequestLimit) |
| 4804 | defer cancel() |
| 4805 | if err := claims.close(ctx); err != nil { |
| 4806 | log.Printf("adopt: host claims shutdown: %v", err) |
| 4807 | } |
| 4808 | }() |
| 4809 | |
| 4810 | // One exclusive lock for the whole pass, in every mode, because the four |
| 4811 | // units that exec this verb can start in the same second and the sweep |
| 4812 | // below is the first thing a pass does that a concurrent pass cannot |
| 4813 | // safely repeat (lockAdoptPass). |
| 4814 | // |
| 4815 | // A lock that cannot be taken is NOT fatal: it drops the sweep and runs |
| 4816 | // the pass exactly as this verb ran before the sweep existed. Every other |
| 4817 | // half is idempotent against a concurrent pass — that is why they were |
| 4818 | // unserialized for a year — so the honest response to "I cannot serialize" |
| 4819 | // is to skip the one half that needs it, not to stop converging the box. |
| 4820 | unlockPass, passLockErr := lockAdoptPass(dataDir()) |
| 4821 | if passLockErr != nil { |
| 4822 | log.Printf("adopt: pass lock %s unavailable (%v) — continuing unserialized, staged-version sweep skipped", adoptPassLockPath(dataDir()), passLockErr) |
| 4823 | } else { |
| 4824 | defer unlockPass() |
| 4825 | } |
| 4826 | // Spec §7 belt 2: arm the self-update relay if the last hand roll did |
| 4827 | // not. Before the fatal gates below, and never fatal itself. |
| 4828 | if !*restartOnly { |
| 4829 | ensureUnits() |
| 4830 | } |
| 4831 | // Root keeps SQLite in the service-user child for every adoption mode, |
| 4832 | // including the early return for restart-only. |
| 4833 | if !claimsEnabled { |
| 4834 | reg, err = openServiceRegistry(*regPath) |
| 4835 | if err != nil { |
| 4836 | if hint := registryOpenHint(*regPath, err); hint != "" { |
| 4837 | log.Fatalf("adopt: open registry %s: %v\nadopt: %s", *regPath, err, hint) |
| 4838 | } |
| 4839 | log.Fatalf("adopt: open registry %s: %v", *regPath, err) |
| 4840 | } |
| 4841 | defer reg.Close() |
| 4842 | } |
| 4843 | // The moves protocol floor (moves spec §11): re-judge carlos.prev |
| 4844 | // against the armed floor and the held rows on every root pass, in every |
| 4845 | // mode, so a raised floor or a new hold reaches the units' choice within |
| 4846 | // the hour, or within seconds of a SetHold (its caller touches |
| 4847 | // stage-requested). Here, before the fatal gates below, so a pass that |
| 4848 | // dies later still judged. Never fatal; one stat on a box with no floor. |
| 4849 | restartChecks := productionRootRestartChecks(systemdUnits{}, release.OptDir()) |
| 4850 | restartChecks.legacy = func() (bool, error) { |
| 4851 | if census == nil { |
| 4852 | return false, os.ErrPermission |
| 4853 | } |
| 4854 | ledger, required, err := census() |
| 4855 | if ledger.State != nil { |
| 4856 | ledger.State.Close() |
| 4857 | } |
| 4858 | return !required && err == nil, err |
| 4859 | } |
| 4860 | if err := recoverRootRestarts(claimCtx, systemdUnits{}, release.OptDir(), os.Getenv, restartChecks, claims.Writes); err != nil { |
| 4861 | log.Printf("adopt: %v", err) |
| 4862 | } |
| 4863 | if err := rootJudgePrev(reg); err != nil { |
| 4864 | log.Printf("adopt: moves protocol floor: %v", err) |
| 4865 | } |
| 4866 | |
| 4867 | // Step 0 (discussion #111): evict the staged versions nothing on this box |
| 4868 | // still names, BEFORE anything that can fail because the disk is full. |
| 4869 | // |
| 4870 | // The placement is the whole point of the fix, not a detail. Below this |
| 4871 | // line are three fatal gates — openDeployStore, the release pubkey parse, |
| 4872 | // and release.Pass, whose bake clock is written with CreateTemp under |
| 4873 | // /data and which log.Fatalf's the pass on error. On the box this is for, |
| 4874 | // with 1.2 MB free, a sweep placed after any of them never runs, and the |
| 4875 | // box has no way to free itself. It needs only the registry and /opt, so |
| 4876 | // it goes as early as having a registry allows. restartPass sits below |
| 4877 | // release.Pass's fatality for exactly this reason and says so. |
| 4878 | // |
| 4879 | // Hourly only: -stage-only is triggered per relayed release, up to once a |
| 4880 | // minute per loop, and walking the version store that often to free a |
| 4881 | // day's worth of shipping is work for no gain. -config-only and |
| 4882 | // -restart-only are narrow triggers that must do their one job. |
| 4883 | generationPins := productionRootGenerationPins() |
| 4884 | if passLockErr == nil && !*stageOnly && !*configOnly && !*restartOnly { |
| 4885 | sweepStagedVersionsPass(reg, generationPins) |
| 4886 | } |
| 4887 | |
| 4888 | ctx := context.Background() |
| 4889 | store, err := openDeployStore(ctx) |
| 4890 | if err != nil { |
| 4891 | log.Fatalf("adopt: %v", err) |
| 4892 | } |
| 4893 | // One callback for both halves below, built once here now that it needs |
| 4894 | // the registry — it reads each route's `unit` column to decide what to |
| 4895 | // cycle. See rootRestartRoute for why the named-unit arm has to exist on |
| 4896 | // THIS pass. |
| 4897 | // |
| 4898 | // Only the RELEASE pass can actually reach that arm. appconfig.Sync skips |
| 4899 | // every route whose Backing is not BackingUnit before it restarts anything |
| 4900 | // (an exec-backed route's config is delivered by the agent's in-process |
| 4901 | // loop instead), and a unit-backed route is exactly what `carlos route |
| 4902 | // -unit` refuses to stamp — so config sync sees no unit-named route by |
| 4903 | // construction. It is handed the same callback anyway because there should |
| 4904 | // be ONE restart dispatch on this pass, not two that could drift. |
| 4905 | // |
| 4906 | // settled: D1's settle passing records the build in the healthy history |
| 4907 | // (D3, proofobserve.go). Root is the process that can hash the unit's |
| 4908 | // executable. |
| 4909 | observer := productionProofObserver(release.OptDir()) |
| 4910 | observer.resolve = restartChecks.resolve |
| 4911 | settled := observer.recordSettled(release.RootHealthyDir) |
| 4912 | restart := rootRestartRouteWith(reg, systemdUnits{}, release.OptDir(), os.Getenv, settled, restartChecks) |
| 4913 | // The restart-request half, and the ONE place it dispatches from — |
| 4914 | // including for exec-backed routes, whose children belong to the agent: |
| 4915 | // rootRestartRoute's try-restart arm reaches a carlos-app@ unit, and an |
| 4916 | // exec child that this pass cannot cycle is cycled by nothing else, which |
| 4917 | // is a real limit of the -restart-only path recorded in restartPass. |
| 4918 | // |
| 4919 | // -restart-only returns here: no config sync, no capability converge, no |
| 4920 | // release.Pass, no nag, and — above all — no release public key, which is |
| 4921 | // parsed below and is fatal when unusable. A restart request must not fail |
| 4922 | // on a key it never needs. |
| 4923 | if *restartOnly { |
| 4924 | logRestartRequestError(productionRestartRequestPass(ctx, store, claims.Writes, func() error { |
| 4925 | routes, err := reg.List() |
| 4926 | if err != nil { |
| 4927 | log.Fatalf("adopt: list routes: %v", err) |
| 4928 | } |
| 4929 | restartPass(ctx, routes, store, restartMarksPath(dataDir()), restart) |
| 4930 | return nil |
| 4931 | })) |
| 4932 | return nil |
| 4933 | } |
| 4934 | // Parsed here, before config sync, exactly as the full pass always has |
| 4935 | // (a box with an unusable release key must die before it converges |
| 4936 | // anything), but SKIPPED entirely for -config-only: that mode never |
| 4937 | // verifies a signature, and making a config-sync trigger fatal on a |
| 4938 | // release-key problem would fail every config write on a box whose |
| 4939 | // releases were already going nowhere. |
| 4940 | var pub ed25519.PublicKey |
| 4941 | if !*configOnly { |
| 4942 | if pub, err = release.ParsePublicKey(os.Getenv("CARLOS_RELEASE_PUBKEY"), "CARLOS_RELEASE_PUBKEY"); err != nil { |
| 4943 | log.Fatalf("adopt: %v", err) |
| 4944 | } |
| 4945 | } |
| 4946 | // Config sync is optional — inert until CARLOS_SECRETS_KEY is set, the |
| 4947 | // same convention every other optional loop in this package follows. It |
| 4948 | // runs before release adoption so a newly adopted route's very first |
| 4949 | // start already sees its env file, and a pass carrying both a release |
| 4950 | // and a config change never starts the new binary against stale config. |
| 4951 | if raw := os.Getenv("CARLOS_SECRETS_KEY"); raw != "" { |
| 4952 | secretsPriv, err := appconfig.ParsePrivateKey(raw, "CARLOS_SECRETS_KEY") |
| 4953 | if err != nil { |
| 4954 | log.Printf("adopt: config sync disabled: %v", err) |
| 4955 | } else if err := appconfig.Sync(ctx, reg, store, secretsPriv, appconfig.DefaultEnvDir, restart, configAckFunc(ctx, store, "adopt"), claims.Writes); err != nil { |
| 4956 | log.Printf("adopt: config sync: %v", err) |
| 4957 | } |
| 4958 | } |
| 4959 | if *configOnly { |
| 4960 | // Everything below is release adoption and its reporting — the half |
| 4961 | // carlos-config-sync.service must never reach, however often a |
| 4962 | // config write triggers it. The restart half is below for the same |
| 4963 | // reason: a config write is not a restart request, and running the |
| 4964 | // comparison here would only move it earlier by minutes at the cost of |
| 4965 | // firing on a trigger that has nothing to do with it. |
| 4966 | return nil |
| 4967 | } |
| 4968 | // The restart-request half again (issue #113), here as the FLOOR under |
| 4969 | // carlos-restart.path: that relay can drop an event that lands while the |
| 4970 | // service is mid-run, exactly as the config one can, and a dropped restart |
| 4971 | // with no floor would be a request that never happens at all. |
| 4972 | // |
| 4973 | // BEFORE release.Pass deliberately. That call is fatal on error, so a box |
| 4974 | // with an unstageable release would otherwise never reach this line — and |
| 4975 | // a wedged instance is most worth cycling precisely on a box where |
| 4976 | // something else is already wrong. |
| 4977 | logRestartRequestError(productionRestartRequestPass(ctx, store, claims.Writes, func() error { |
| 4978 | if routes, err := reg.List(); err != nil { |
| 4979 | log.Printf("adopt: list routes for the restart pass: %v", err) |
| 4980 | } else { |
| 4981 | restartPass(ctx, routes, store, restartMarksPath(dataDir()), restart) |
| 4982 | } |
| 4983 | return nil |
| 4984 | })) |
| 4985 | |
| 4986 | // Step 2 (spec §4d): the grant-drift capability converge, BETWEEN config |
| 4987 | // sync and release adoption. Before release.Pass because it repairs the |
| 4988 | // shape of what is ALREADY live — a revoked grant, a deleted drop-in, a |
| 4989 | // socket unit nobody enabled — and a route that release.Pass then moves |
| 4990 | // to a new version converges again inside adoptHost, which is a no-op |
| 4991 | // when nothing about the shape changed: same bytes, no write, no reload, |
| 4992 | // no restart. |
| 4993 | // |
| 4994 | // The zero caps.Ops is the production one (real /etc/systemd/system, real |
| 4995 | // systemctl, real useradd), for the same reason WithCapabilities below is |
| 4996 | // only ever set here: this verb is root out of carlos-adopt.service, and |
| 4997 | // the in-process loops in this file cannot write unit configuration at |
| 4998 | // all. |
| 4999 | // |
| 5000 | // Never fatal. A failure here is the registry not listing; adoption of |
| 5001 | // releases is a separate job and must still run. |
| 5002 | if err := capabilityPass(reg, release.OptDir(), productionClaimCapsOps(caps.Ops{}, restartChecks.allowLegacy)); err != nil { |
| 5003 | log.Printf("adopt: capability pass: %v", err) |
| 5004 | } |
| 5005 | |
| 5006 | // NagExpiredCanaries(os.Stdout) is set HERE ONLY (2026-08-03 review) — the |
| 5007 | // canary-TTL nag (spec §5) belongs to the hourly root pass |
| 5008 | // (carlos-adopt.service, this verb's ExecStart), the same split PR #9's |
| 5009 | // config sync above uses. adoptLoop and fastConvergeLoop in this same |
| 5010 | // package call release.Pass/PassFilter without the nag option, so an |
| 5011 | // expired canary nags once an hour from here, never |
| 5012 | // every 2-3 seconds from the in-process fast loop. |
| 5013 | // |
| 5014 | // WithCapabilities(caps.Ops{}) is likewise set HERE ONLY, and the zero Ops |
| 5015 | // is the production one (real /etc/systemd/system, real systemctl, real |
| 5016 | // useradd). This verb runs as root out of carlos-adopt.service; the |
| 5017 | // in-process loops in this same file run inside carlos-edge.service's |
| 5018 | // ProtectSystem=strict sandbox and cannot write unit configuration at all, |
| 5019 | // which is why adoptLoop and fastPassOpts each carry a comment saying |
| 5020 | // never to add this one. |
| 5021 | // |
| 5022 | // -stage-only drops the nag and keeps everything else. It is what |
| 5023 | // carlos-stage.service runs when the agent relays a version it could not |
| 5024 | // stage: that request arrives at most once per stageRequestEvery per loop, and the nag |
| 5025 | // is an HOURLY visibility feature — firing it per request would flood the |
| 5026 | // journal with the same expired-canary lines, which is precisely what the |
| 5027 | // option's own doc comment says never to do. The capability converge and |
| 5028 | // config sync stay: both are jobs only this root pass can do, and a |
| 5029 | // newly staged version wants its unit shape and env file converged with |
| 5030 | // it, not an hour later. |
| 5031 | // |
| 5032 | // WithStartUnit is wired on THIS pass above all: on a hardened box the |
| 5033 | // sandboxed agent's stage is blocked and relayed here, so this is the |
| 5034 | // pass that actually try-restarts a console's named unit — and therefore |
| 5035 | // the pass whose failed restart, rolled back, must start that unit again |
| 5036 | // on the restored version rather than leave a correct symlink over no |
| 5037 | // process (release.WithStartUnit; without it the NEXT pass's try-restart |
| 5038 | // no-ops on the stopped unit, exits 0, and stamps a build that has never |
| 5039 | // run). -stage-only keeps it for the same reason it keeps the converge: |
| 5040 | // a relayed stage is exactly the request that performs such a restart. |
| 5041 | pr := proofRoot{ |
| 5042 | obs: observer, reg: reg, store: store, bakeDir: dataDir(), |
| 5043 | healthyDir: release.RootHealthyDir, policyPath: release.ProofPolicyPath, |
| 5044 | readKey: readProvingKey, hourly: !*stageOnly, logf: log.Printf, |
| 5045 | } |
| 5046 | history, err := configureRootProof(&pr, restartChecks, func() (*release.RootHistoryPass, error) { |
| 5047 | return openRootHistory(store, release.OptDir()) |
| 5048 | }) |
| 5049 | if err != nil { |
| 5050 | log.Printf("adopt: %v", err) |
| 5051 | } else if history != nil { |
| 5052 | defer history.History.Store.Close() |
| 5053 | linkRootHistory(history, restartChecks, release.OptDir(), os.Getenv) |
| 5054 | } |
| 5055 | recovery := productionNamedUnitRecovery(store, pub, release.OptDir(), claims.Writes) |
| 5056 | passOpts := []release.PassOption{release.WithRootGenerationPins(generationPins), release.WithCapabilities(productionClaimCapsOps(caps.Ops{}, restartChecks.allowLegacy)), release.BakeDir(dataDir()), release.WithRestartAttemptSettled(restartChecks.finish), |
| 5057 | release.WithRunningVersion(restartChecks.runningVersion), |
| 5058 | release.WithSelectedRestarts( |
| 5059 | rootSelectedRestart(reg, systemdUnits{}, release.OptDir(), os.Getenv, settled, restartChecks, false), |
| 5060 | rootSelectedRestart(reg, systemdUnits{}, release.OptDir(), os.Getenv, nil, restartChecks, true)), |
| 5061 | release.WithStartUnit(rootStartRoute(reg, release.OptDir(), restartChecks)), release.WithHostWrites(claims.Writes), |
| 5062 | release.WithDigestRatchet(pointerRatchet(true, release.OptDir(), dataDir(), os.Getenv)), |
| 5063 | release.WithRollbackRestart(rollbackUnitRestart(reg, systemdUnits{}, release.OptDir(), os.Getenv, restartChecks)), |
| 5064 | // D3's gate: a fresh gate per root pass, so the policy file is read |
| 5065 | // at the start of this pass (BeginPass) like every other. Healthy |
| 5066 | // history is root-owned; diagnostic refusals still use BakeDir. |
| 5067 | release.WithProofGate(release.NewRootProofGate(release.ProofPolicyPath, dataDir())), |
| 5068 | // Settings spec §4.3: root alone resolves install requests and |
| 5069 | // writes /opt/carlos/consumed/<host>, before it stages. |
| 5070 | updatesPassOption(dataDir(), true, updateProviders...)} |
| 5071 | passOpts, err = rootAdoptionOptions(passOpts, restartChecks, history, func() release.PassOption { return recovery.passOption(reg, restartChecks, os.Getenv) }) |
| 5072 | if err != nil { |
| 5073 | return err |
| 5074 | } |
| 5075 | if !*stageOnly { |
| 5076 | passOpts = append(passOpts, release.NagExpiredCanaries(os.Stdout)) |
| 5077 | } |
| 5078 | // D3 (proofroot.go) wraps the pass. Before it, record every named-unit |
| 5079 | // route's live build that has run well as healthy, so the build live when |
| 5080 | // a requirement is switched on is in the history before this pass can |
| 5081 | // adopt a proven head over it (final review H4). After it, mint on a |
| 5082 | // proving box and, on the hourly pass, say when a requirement is off; |
| 5083 | // a route the pass changed is looked at again first. |
| 5084 | passErr := pr.around(ctx, func() error { |
| 5085 | return release.Pass(ctx, reg, store, pub, release.OptDir(), restart, passOpts...) |
| 5086 | }) |
| 5087 | if passErr != nil { |
| 5088 | log.Printf("adopt: %v", passErr) |
| 5089 | } |
| 5090 | |
| 5091 | // A move's held stage (moves design §7.3, T7): the sandboxed agent |
| 5092 | // cannot write /opt, so a move's target leaves a request under |
| 5093 | // /data/moves/held-stage and touches stage-requested. Served here, after |
| 5094 | // release adoption, on -stage-only and on the hourly pass (its floor). |
| 5095 | // Reads nothing and writes nothing unless a request exists, and none |
| 5096 | // exist while CARLOS_MOVES_PHASE1 is off. |
| 5097 | heldStagePass(ctx, dataDir(), release.OptDir(), pub, pointerRatchet(true, release.OptDir(), dataDir(), os.Getenv), reg, store, BoxLabel()) |
| 5098 | |
| 5099 | // Step 4: the advisories. The two capability advisories that fall out of |
| 5100 | // READING manifests (declared-but-ungranted, and capabilities on an |
| 5101 | // exec-backed route) are printed by capabilityPass above, where the |
| 5102 | // manifests are already in hand; this one needs the box's firewall and |
| 5103 | // its passwd database instead, so it lives here with the other |
| 5104 | // systemd-probing advisory. |
| 5105 | if routes, err := reg.List(); err == nil { |
| 5106 | adviseBackingMismatches(routes) |
| 5107 | adviseHostRoleGateMissing(routes, lookupUID, nftGateUIDs) |
| 5108 | } |
| 5109 | return passErr |
| 5110 | } |
| 5111 | |
amadan