| 1 | package engines |
| 2 | |
| 3 | import ( |
| 4 | "fmt" |
| 5 | "sort" |
| 6 | ) |
| 7 | |
| 8 | // The launch gate (spec §5.1 "The invariant", "Setting the mark", "Starts |
| 9 | // systemd makes on its own"). |
| 10 | // |
| 11 | // I1: a process of a build whose manifest declares doc_engine starts only |
| 12 | // after the edge has fenced its group as internal. Every activator launch |
| 13 | // path ends in the gate; systemd's own starts (Restart=always, socket |
| 14 | // activation) execute the live link and pass no gate, so they are safe only |
| 15 | // because adoption gates before it repoints and nothing else moves the link. |
| 16 | |
| 17 | // fenceWaitTicks is the gate's bound: 30 s of the edge's reloads, one per |
| 18 | // tick. |
| 19 | const fenceWaitTicks = 3 |
| 20 | |
| 21 | // gateRule names the rule that gates a launch path; systemd and socket |
| 22 | // starts are not activator launches and have none. |
| 23 | var gateRule = map[string]string{ |
| 24 | "adopt": "gate-adopt", |
| 25 | "swap": "gate-swap", |
| 26 | "fallback": "gate-fallback", |
| 27 | "recovery": "gate-recovery", |
| 28 | "seed": "gate-seed", |
| 29 | "cmd": "gate-cmd", |
| 30 | "systemd": "", |
| 31 | "socket": "", |
| 32 | } |
| 33 | |
| 34 | // gateSeesDeclaring is what the gate reads from the build's manifest: its |
| 35 | // declaration, or, when the manifest cannot be read, declaring. |
| 36 | func (w *world) gateSeesDeclaring(ver int) bool { |
| 37 | bd := w.builds[ver] |
| 38 | if !bd.ManifestOK { |
| 39 | return on("unreadable-declares") |
| 40 | } |
| 41 | return bd.Declares |
| 42 | } |
| 43 | |
| 44 | // gate is the launch gate for a build about to start in group g: a build it |
| 45 | // sees as declaring writes 1 if 1 is not already written, then always waits |
| 46 | // for the edge's fence to cover the group's mark, within the bound. False is |
| 47 | // a refused launch (the wake fails). |
| 48 | func (w *world) gate(b *box, g string, ver int, path string) bool { |
| 49 | if !w.gateSeesDeclaring(ver) { |
| 50 | return true |
| 51 | } |
| 52 | if !w.dedicatedLaunchCheck(b, g, ver, path) { |
| 53 | return false |
| 54 | } |
| 55 | wrote := false |
| 56 | if w.groupMark(b, g) != markInternal { |
| 57 | w.writeMark(b, g, markInternal) |
| 58 | wrote = true |
| 59 | } |
| 60 | if !wrote && !on("gate-waits-when-written") { |
| 61 | w.logf("box%d: gate(%s) %s v%d: 1 already written; not waiting", b.ID, path, g, ver) |
| 62 | return true |
| 63 | } |
| 64 | for i := 0; i < fenceWaitTicks && !w.fenced(b, g); i++ { |
| 65 | w.reload(b) |
| 66 | } |
| 67 | if !w.fenced(b, g) { |
| 68 | w.logf("box%d: gate(%s) %s v%d: no fence within the bound; refused", b.ID, path, g, ver) |
| 69 | return false |
| 70 | } |
| 71 | return true |
| 72 | } |
| 73 | |
| 74 | // launch starts build ver in group g by path, through the gate where the |
| 75 | // path has one. It returns false when the gate refuses. |
| 76 | func (w *world) launch(b *box, g string, ver int, path string) bool { |
| 77 | rule, known := gateRule[path] |
| 78 | if !known { |
| 79 | panic(fmt.Sprintf("launch: unknown path %q", path)) |
| 80 | } |
| 81 | // the gate holds the group's lock from its check through process |
| 82 | // creation; --public holds the same lock for its whole run, so a launch |
| 83 | // that finds it held waits (here: is refused this pass) |
| 84 | if rule != "" && b.Locked[g] && on("clear-launch-lock") { |
| 85 | w.logf("box%d: %s %s v%d: --public holds the group's lock; not this pass", b.ID, path, g, ver) |
| 86 | return false |
| 87 | } |
| 88 | if rule != "" && on(rule) && !w.gate(b, g, ver, path) { |
| 89 | return false |
| 90 | } |
| 91 | // a group keeps at most three processes: the oldest has exited |
| 92 | n := 0 |
| 93 | for _, p := range b.Procs { |
| 94 | if p.Group == g { |
| 95 | n++ |
| 96 | } |
| 97 | } |
| 98 | if n >= 3 { |
| 99 | for i, p := range b.Procs { |
| 100 | if p.Group == g { |
| 101 | b.Procs = append(b.Procs[:i], b.Procs[i+1:]...) |
| 102 | w.logf("box%d: process %d of %s exited", b.ID, p.ID, g) |
| 103 | break |
| 104 | } |
| 105 | } |
| 106 | } |
| 107 | w.stage(b, g, ver) // a build runs from /opt/carlos/versions: it is staged |
| 108 | w.nextID++ |
| 109 | p := &proc{ID: w.nextID, Group: g, Ver: ver, Slice: "inst:" + g} |
| 110 | b.Procs = append(b.Procs, p) |
| 111 | w.logf("box%d: %s started %s v%d as process %d", b.ID, path, g, ver, p.ID) |
| 112 | w.applyTamper(b, p) |
| 113 | w.checkStart(b, p) |
| 114 | return true |
| 115 | } |
| 116 | |
| 117 | // stopGroup stops every process of group g. |
| 118 | func (w *world) stopGroup(b *box, g string) { |
| 119 | kept := b.Procs[:0] |
| 120 | for _, p := range b.Procs { |
| 121 | if p.Group != g { |
| 122 | kept = append(kept, p) |
| 123 | } |
| 124 | } |
| 125 | b.Procs = kept |
| 126 | } |
| 127 | |
| 128 | func (w *world) running(b *box, g string) bool { |
| 129 | for _, p := range b.Procs { |
| 130 | if p.Group == g { |
| 131 | return true |
| 132 | } |
| 133 | } |
| 134 | return false |
| 135 | } |
| 136 | |
| 137 | // adopt makes ver the live release of group g: set-before (the gate, reached |
| 138 | // before the repoint; on a timeout the link does not move and the mark stays |
| 139 | // 1), then the repoint, then the restart, itself a gated launch. between runs |
| 140 | // after the repoint and before the restart: whatever systemd does there. |
| 141 | func (w *world) adopt(b *box, g string, ver int, between func()) bool { |
| 142 | w.logf("box%d: adopt %s v%d (live v%d)", b.ID, g, ver, b.Live[g]) |
| 143 | if !w.repoint(b, g, ver) { |
| 144 | return false |
| 145 | } |
| 146 | if between != nil { |
| 147 | between() |
| 148 | } |
| 149 | w.stopGroup(b, g) |
| 150 | return w.launch(b, g, ver, "adopt") |
| 151 | } |
| 152 | |
| 153 | // repointAwake is adoption that repoints the live link and does not restart: |
| 154 | // an awake hibernating route (release/adopt.go:1174), or a generation |
| 155 | // rollout that only schedules the swap. What ran keeps running. |
| 156 | func (w *world) repointAwake(b *box, g string, ver int) bool { |
| 157 | w.logf("box%d: adopt %s v%d without a restart (live v%d)", b.ID, g, ver, b.Live[g]) |
| 158 | if !w.repoint(b, g, ver) { |
| 159 | return false |
| 160 | } |
| 161 | w.logf("box%d: %s keeps its running processes until their next start", b.ID, g) |
| 162 | return true |
| 163 | } |
| 164 | |
| 165 | // repoint is adoption up to the repoint: the group's lock (public-holds-lock), |
| 166 | // set-before (the gate, reached before the repoint; on a timeout the link |
| 167 | // does not move and the mark stays 1), then the repoint. |
| 168 | func (w *world) repoint(b *box, g string, ver int) bool { |
| 169 | if b.Locked[g] && on("public-holds-lock") { |
| 170 | w.logf("box%d: adopt %s v%d: --public holds the group's lock; not this pass", b.ID, g, ver) |
| 171 | return false |
| 172 | } |
| 173 | if on("set-before-repoint") && !w.gate(b, g, ver, "adopt") { |
| 174 | w.logf("box%d: adopt %s v%d: not repointed this pass", b.ID, g, ver) |
| 175 | return false |
| 176 | } |
| 177 | b.Live[g] = ver |
| 178 | w.stage(b, g, ver) |
| 179 | if b.Good[g] == 0 { |
| 180 | b.Good[g] = ver |
| 181 | } |
| 182 | w.logf("box%d: live link of %s -> v%d", b.ID, g, ver) |
| 183 | return true |
| 184 | } |
| 185 | |
| 186 | // stage records that build ver of group g is staged on the box. A version is |
| 187 | // staged once per app (/opt/carlos/versions holds the app's versions): if |
| 188 | // another instance of the app already staged it, it stays recorded there |
| 189 | // only, and this group runs it from that copy. |
| 190 | func (w *world) stage(b *box, g string, ver int) { |
| 191 | for _, v := range w.stagedForApp(b, g) { |
| 192 | if v == ver { |
| 193 | return |
| 194 | } |
| 195 | } |
| 196 | b.Staged[g] = append(b.Staged[g], ver) |
| 197 | } |
| 198 | |
| 199 | // appGroups is group g and every other instance group on the box whose own |
| 200 | // row has the same non-empty account/app, sorted. |
| 201 | func (w *world) appGroups(b *box, g string) []string { |
| 202 | own := b.Rows[g] |
| 203 | if own == nil || own.App == "" { |
| 204 | return []string{g} |
| 205 | } |
| 206 | out := []string{} |
| 207 | seen := false |
| 208 | for _, h := range w.groups(b) { |
| 209 | if r := b.Rows[h]; r != nil && r.App == own.App { |
| 210 | out = append(out, h) |
| 211 | seen = seen || h == g |
| 212 | } |
| 213 | } |
| 214 | if !seen { |
| 215 | out = append(out, g) |
| 216 | sort.Strings(out) |
| 217 | } |
| 218 | return out |
| 219 | } |
| 220 | |
| 221 | // stagedForApp is every version of g's app staged on the box, whichever |
| 222 | // instance of the app staged it. |
| 223 | func (w *world) stagedForApp(b *box, g string) []int { |
| 224 | var out []int |
| 225 | for _, h := range w.appGroups(b, g) { |
| 226 | out = append(out, b.Staged[h]...) |
| 227 | } |
| 228 | return out |
| 229 | } |
| 230 | |
| 231 | // wake is the activator waking group g by its wake plan: "recovery" starts |
| 232 | // the recovery build, "fallback" first fails to start the live build and |
| 233 | // falls back; both choose GoodVersion. The triggering request, if any, is |
| 234 | // delivered afterwards (or finds the fence cancelled it). It reports whether |
| 235 | // there was a build to launch. |
| 236 | func (w *world) wake(b *box, g string, mode string, trigger *req) bool { |
| 237 | ver := b.Good[g] |
| 238 | if mode == "fallback" { |
| 239 | w.logf("box%d: wake %s: live v%d fails to start; falling back to good v%d", b.ID, g, b.Live[g], ver) |
| 240 | } else { |
| 241 | w.logf("box%d: wake %s: recovery build good v%d", b.ID, g, ver) |
| 242 | } |
| 243 | if ver != 0 { |
| 244 | w.launch(b, g, ver, mode) |
| 245 | } |
| 246 | if trigger != nil { |
| 247 | w.deliver(trigger) |
| 248 | } |
| 249 | return ver != 0 |
| 250 | } |
| 251 | |
| 252 | // moveLive is anything other than adoption moving group g's live link (a |
| 253 | // rollout tool, a hand repair). Only adoption, through the gate, and |
| 254 | // --public may move a live link. |
| 255 | func (w *world) moveLive(b *box, g string, ver int, who string) bool { |
| 256 | if on("live-link-only-adoption") { |
| 257 | w.logf("box%d: %s may not move the live link of %s; refused", b.ID, who, g) |
| 258 | return false |
| 259 | } |
| 260 | b.Live[g] = ver |
| 261 | w.logf("box%d: %s moved the live link of %s -> v%d", b.ID, who, g, ver) |
| 262 | return true |
| 263 | } |
| 264 | |
| 265 | // systemdRestart is Restart=always: the unit's main process exits and |
| 266 | // systemd starts /opt/carlos/live/<host> again, with no activator. False |
| 267 | // when the route has no live release. |
| 268 | func (w *world) systemdRestart(b *box, g string) bool { |
| 269 | if b.Live[g] == 0 { |
| 270 | return false |
| 271 | } |
| 272 | w.stopGroup(b, g) |
| 273 | w.launch(b, g, b.Live[g], "systemd") |
| 274 | return true |
| 275 | } |
| 276 | |
| 277 | // socketActivate is carlos-app@.socket starting a stopped unit from the live |
| 278 | // link, with no activator. False when there is nothing to start. |
| 279 | func (w *world) socketActivate(b *box, g string) bool { |
| 280 | if b.Live[g] == 0 || w.running(b, g) { |
| 281 | return false |
| 282 | } |
| 283 | w.launch(b, g, b.Live[g], "socket") |
| 284 | return true |
| 285 | } |
| 286 | |
| 287 | // ---------- the random walk ---------- |
| 288 | |
| 289 | func (w *world) pickBox() *box { return w.boxes[w.rng.Intn(len(w.boxes))] } |
| 290 | |
| 291 | // groups is every instance group of b's registry, sorted: the groups of the |
| 292 | // instance rows that are not platform-minted (adoption skips minted rows, |
| 293 | // release/adopt.go:1210; their processes are their source's). |
| 294 | func (w *world) groups(b *box) []string { |
| 295 | seen := map[string]bool{} |
| 296 | var out []string |
| 297 | for h, r := range b.Rows { |
| 298 | if r.Kind != "instance" || r.Platform { |
| 299 | continue |
| 300 | } |
| 301 | if g := w.group(b, h); !seen[g] { |
| 302 | seen[g] = true |
| 303 | out = append(out, g) |
| 304 | } |
| 305 | } |
| 306 | sort.Strings(out) |
| 307 | return out |
| 308 | } |
| 309 | |
| 310 | func (w *world) pickGroup(b *box) string { |
| 311 | gs := w.groups(b) |
| 312 | if len(gs) == 0 { |
| 313 | return "" |
| 314 | } |
| 315 | return gs[w.rng.Intn(len(gs))] |
| 316 | } |
| 317 | |
| 318 | func (w *world) pickVer() int { return 1 + w.rng.Intn(nBuilds) } |
| 319 | |
| 320 | // randomPublic sends one public request to a host of b's edge map, by its |
| 321 | // own name, an alias or a path map. |
| 322 | func (w *world) randomPublic(b *box) *req { |
| 323 | hosts := make([]string, 0, len(b.EdgeRows)) |
| 324 | for h := range b.EdgeRows { |
| 325 | hosts = append(hosts, h) |
| 326 | } |
| 327 | if len(hosts) == 0 { |
| 328 | return nil |
| 329 | } |
| 330 | sort.Strings(hosts) |
| 331 | h := hosts[w.rng.Intn(len(hosts))] |
| 332 | paths := []string{"/"} |
| 333 | for p := range b.EdgeRows[h].PathTo { |
| 334 | paths = append(paths, p) |
| 335 | } |
| 336 | sort.Strings(paths[1:]) |
| 337 | return w.admitPublic(b, h, paths[w.rng.Intn(len(paths))]) |
| 338 | } |
| 339 | |
| 340 | // heldFor is the oldest held request on b whose proxy reaches group g. |
| 341 | func (w *world) heldFor(b *box, g string) *req { |
| 342 | for _, r := range b.Admitted { |
| 343 | if r.Held && w.backendGroup(b, r.Backend) == g { |
| 344 | return r |
| 345 | } |
| 346 | } |
| 347 | return nil |
| 348 | } |
| 349 | |
| 350 | func init() { |
| 351 | actions = append(actions, |
| 352 | action{name: "adopt", weight: 3, run: func(w *world) bool { |
| 353 | b := w.pickBox() |
| 354 | g := w.pickGroup(b) |
| 355 | if g == "" { |
| 356 | return false |
| 357 | } |
| 358 | ver := w.pickVer() |
| 359 | var between func() |
| 360 | switch w.rng.Intn(3) { |
| 361 | case 1: |
| 362 | between = func() { w.systemdRestart(b, g) } |
| 363 | case 2: |
| 364 | between = func() { |
| 365 | if r := w.randomPublic(b); r != nil { |
| 366 | w.socketActivate(b, w.backendGroup(b, r.Backend)) |
| 367 | w.deliver(r) |
| 368 | } |
| 369 | } |
| 370 | } |
| 371 | w.adopt(b, g, ver, between) |
| 372 | return true |
| 373 | }}, |
| 374 | action{name: "reload", weight: 4, run: func(w *world) bool { w.reload(w.pickBox()); return true }}, |
| 375 | action{name: "reload-fails", weight: 1, run: func(w *world) bool { |
| 376 | b := w.pickBox() |
| 377 | b.ReloadFails = !b.ReloadFails && !w.calm() |
| 378 | w.logf("box%d: reload fails = %v", b.ID, b.ReloadFails) |
| 379 | return true |
| 380 | }}, |
| 381 | // 9, not 6 (Task 6): fence-box-wide's sweep catch needs a request held |
| 382 | // through a service row that is deleted before a declaring adoption's |
| 383 | // fence; at 6 the 3000-seed sweep caught it 7 times, at 9 12 times |
| 384 | action{name: "public-request", weight: 9, run: func(w *world) bool { |
| 385 | b := w.pickBox() |
| 386 | r := w.randomPublic(b) |
| 387 | if r == nil { |
| 388 | return false |
| 389 | } |
| 390 | if len(b.Admitted) <= 8 && w.rng.Intn(2) == 0 { |
| 391 | r.Held = true |
| 392 | w.logf("box%d: request %d held in the activator's queue", b.ID, r.ID) |
| 393 | return true |
| 394 | } |
| 395 | w.deliver(r) |
| 396 | return true |
| 397 | }}, |
| 398 | action{name: "resume", weight: 3, run: func(w *world) bool { |
| 399 | var held []*req |
| 400 | for _, b := range w.boxes { |
| 401 | for _, r := range b.Admitted { |
| 402 | if r.Held { |
| 403 | held = append(held, r) |
| 404 | } |
| 405 | } |
| 406 | } |
| 407 | if len(held) == 0 { |
| 408 | return false |
| 409 | } |
| 410 | sort.Slice(held, func(i, j int) bool { return held[i].ID < held[j].ID }) |
| 411 | w.deliver(held[w.rng.Intn(len(held))]) |
| 412 | return true |
| 413 | }}, |
| 414 | action{name: "wake", weight: 3, run: func(w *world) bool { |
| 415 | b := w.pickBox() |
| 416 | g := w.pickGroup(b) |
| 417 | if g == "" { |
| 418 | return false |
| 419 | } |
| 420 | mode := "recovery" |
| 421 | if w.rng.Intn(2) == 0 { |
| 422 | mode = "fallback" |
| 423 | } |
| 424 | w.stopGroup(b, g) // the group had gone to sleep |
| 425 | return w.wake(b, g, mode, w.heldFor(b, g)) |
| 426 | }}, |
| 427 | action{name: "swap", weight: 2, run: func(w *world) bool { |
| 428 | b := w.pickBox() |
| 429 | g := w.pickGroup(b) |
| 430 | if g == "" { |
| 431 | return false |
| 432 | } |
| 433 | if w.launch(b, g, w.pickVer(), "swap") { |
| 434 | // the incoming generation serves; the outgoing ones stop |
| 435 | newest := b.Procs[len(b.Procs)-1] |
| 436 | kept := b.Procs[:0] |
| 437 | for _, p := range b.Procs { |
| 438 | if p.Group != g || p == newest { |
| 439 | kept = append(kept, p) |
| 440 | } |
| 441 | } |
| 442 | b.Procs = kept |
| 443 | } |
| 444 | return true |
| 445 | }}, |
| 446 | action{name: "seed", weight: 1, run: func(w *world) bool { |
| 447 | b := w.pickBox() |
| 448 | g := w.pickGroup(b) |
| 449 | if g == "" { |
| 450 | return false |
| 451 | } |
| 452 | w.stopGroup(b, g) |
| 453 | w.launch(b, g, w.pickVer(), "seed") |
| 454 | return true |
| 455 | }}, |
| 456 | action{name: "cmd", weight: 1, run: func(w *world) bool { |
| 457 | b := w.pickBox() |
| 458 | g := w.pickGroup(b) |
| 459 | if g == "" { |
| 460 | return false |
| 461 | } |
| 462 | w.launch(b, g, w.pickVer(), "cmd") |
| 463 | return true |
| 464 | }}, |
| 465 | action{name: "systemd-restart", weight: 2, run: func(w *world) bool { |
| 466 | b := w.pickBox() |
| 467 | return w.systemdRestart(b, w.pickGroup(b)) |
| 468 | }}, |
| 469 | action{name: "socket-activation", weight: 2, run: func(w *world) bool { |
| 470 | b := w.pickBox() |
| 471 | g := w.pickGroup(b) |
| 472 | w.stopGroup(b, g) // the unit exited idle; a connection arrives |
| 473 | ok := w.socketActivate(b, g) |
| 474 | if r := w.heldFor(b, g); r != nil { |
| 475 | w.deliver(r) |
| 476 | } |
| 477 | return ok |
| 478 | }}, |
| 479 | action{name: "repoint-live", weight: 1, run: func(w *world) bool { |
| 480 | b := w.pickBox() |
| 481 | g := w.pickGroup(b) |
| 482 | if g == "" { |
| 483 | return false |
| 484 | } |
| 485 | w.moveLive(b, g, w.pickVer(), "a rollout tool") |
| 486 | return true |
| 487 | }}, |
| 488 | ) |
| 489 | } |
| 490 | |
amadan