| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "log" |
| 7 | "sync" |
| 8 | "time" |
| 9 | |
| 10 | "github.com/carlosframework/platform/internal/instancectl" |
| 11 | "github.com/carlosframework/platform/internal/instancekey" |
| 12 | "github.com/carlosframework/platform/internal/registry" |
| 13 | ) |
| 14 | |
| 15 | // Why an instance stop interrupted a run (run spec §5.3). The text is what |
| 16 | // the member reads after "the run was interrupted:" (RUN-C1 to RUN-C8); |
| 17 | // beginStopLocked records one on every stop. |
| 18 | const ( |
| 19 | StopIdle = "the instance hibernated" |
| 20 | StopRestart = "the instance was restarted" |
| 21 | StopRelease = "the instance moved to a new release" |
| 22 | StopExited = "the instance's process exited" |
| 23 | StopLeaseLost = "the instance lost its database lease" |
| 24 | StopFenced = "the instance was stopped for a move or a layout change" |
| 25 | StopRetired = "the instance was taken off this box" |
| 26 | StopShutdown = "the box's agent stopped" |
| 27 | ) |
| 28 | |
| 29 | // ErrCmdClaim: claim control is armed on this box. Runs need the isolated |
| 30 | // root protocol's reservations (#250), so they are refused (run spec §3.1). |
| 31 | var ErrCmdClaim = errors.New("activator: claim control is armed; runs wait for the reservation protocol (#250)") |
| 32 | |
| 33 | // ErrCmdBusy: a run is already attached to the instance, or an earlier |
| 34 | // run's unit was never proven settled. |
| 35 | var ErrCmdBusy = errors.New("activator: a run is already attached to this instance") |
| 36 | |
| 37 | // ErrCmdNotInstance: runs are for instances; the route is a worker or |
| 38 | // another kind. |
| 39 | var ErrCmdNotInstance = errors.New("activator: runs are only for instances, and this route is not one") |
| 40 | |
| 41 | // CmdWakeError: the wake failed or did not finish within cmdWakeBound. |
| 42 | // Reason is member-safe (memberSafeWakeReason). |
| 43 | type CmdWakeError struct { |
| 44 | Reason string |
| 45 | Err error |
| 46 | } |
| 47 | |
| 48 | func (e *CmdWakeError) Error() string { return e.Reason } |
| 49 | func (e *CmdWakeError) Unwrap() error { return e.Err } |
| 50 | |
| 51 | // CmdStoppedError: the instance began stopping before the unit started. The |
| 52 | // run is interrupted with Reason, never queued for later (run spec §5.3). |
| 53 | type CmdStoppedError struct{ Reason string } |
| 54 | |
| 55 | func (e *CmdStoppedError) Error() string { |
| 56 | return "activator: the instance stopped before the run started: " + e.Reason |
| 57 | } |
| 58 | |
| 59 | // cmdWakeBound is how long RunCmd waits for a wake (run spec §5.3). |
| 60 | var cmdWakeBound = 60 * time.Second |
| 61 | |
| 62 | // cmdStopBound bounds finishStop's wait for a run: the start in flight |
| 63 | // (cmdStartBound), the stop (DeadlineControl), the cgroup (actStopTimeout) |
| 64 | // and both streams (cmdAcceptGrace + cmdDrainBound each), with room. |
| 65 | // Shutdown runs inside systemd's 90 s TimeoutStopSec for the agent: a |
| 66 | // settle (about 62 s) fits; a start in flight on top may not, and an agent |
| 67 | // killed mid-stop leaves the lease to its TTL. |
| 68 | const cmdStopBound = 2 * time.Minute |
| 69 | |
| 70 | // cmdStartBound bounds RunCmd's StartCmd, well below cmdStopBound, so a |
| 71 | // stop waiting on the start still has the settle's budget after it. |
| 72 | const cmdStartBound = 45 * time.Second |
| 73 | |
| 74 | // CmdHandle is a started run, as the box's run pass sees it. |
| 75 | type CmdHandle interface { |
| 76 | Invocation() string |
| 77 | Poll(ctx context.Context) (instancectl.CmdStatus, error) |
| 78 | // Interrupted is the instance stop's reason when one stopped the unit, |
| 79 | // "" otherwise, and when that stop was asked for: a command whose done |
| 80 | // status the caller had before then finished on its own. |
| 81 | Interrupted() (reason string, at time.Time) |
| 82 | // Stop stops the unit and waits for it to settle and its streams to |
| 83 | // end, leaving the instance running: a cancel, a timeout, or the end |
| 84 | // of a command that exited. Safe to call more than once. |
| 85 | Stop(ctx context.Context) error |
| 86 | // Release clears the run from the entry once its result is written. |
| 87 | Release() |
| 88 | } |
| 89 | |
| 90 | // cmdHold is the entry's cmd field. |
| 91 | type cmdHold struct { |
| 92 | runID string |
| 93 | // starting is set under a.mu by RunCmd's last awake check: from then |
| 94 | // on a stop waits for the start. Before it, a stop does not wait: the |
| 95 | // check will see the entry stopping and abandon the run. |
| 96 | starting bool |
| 97 | // cancelPrepare cancels spec.Prepare's context; beginStopLocked calls |
| 98 | // it, under a.mu, so a stop interrupts a download. |
| 99 | cancelPrepare context.CancelFunc |
| 100 | started chan struct{} // closed once StartCmd returned, either way |
| 101 | unit CmdUnit // written before started is closed; nil when no unit can be running |
| 102 | now func() time.Time |
| 103 | mu sync.Mutex |
| 104 | asked bool // a stop has been asked for; under mu |
| 105 | reason string |
| 106 | askedAt time.Time // when the first stop was asked for; under mu |
| 107 | // settling is the settle attempt in flight, nil when none is. A stop |
| 108 | // joins it rather than starting a second one. Under mu. |
| 109 | settling *cmdSettle |
| 110 | // proven: a settle attempt succeeded, so nothing of the unit runs. It |
| 111 | // is never unset. A failed attempt leaves it false, and the next stop |
| 112 | // tries again (merge ruling M1): a failed settle is never cached as |
| 113 | // the run's last word while the instance's ownership is given up. |
| 114 | // Under mu. |
| 115 | proven bool |
| 116 | // lastTry is when the last settle attempt ended, by h.now; the sweep's |
| 117 | // retry keeps actStopTimeout after it (retrySettle). Under mu. |
| 118 | lastTry time.Time |
| 119 | // gen is the generation the run was admitted against (merge ruling |
| 120 | // M2): its number and its exe pin. RunCmd sets it under a.mu with |
| 121 | // starting, and it is never written again. The command execs gen's |
| 122 | // build; a swap or a scheduled restart that retires gen interrupts |
| 123 | // the run (interruptCmdsLocked), and the generations file names gen's |
| 124 | // build until the run is proven settled (entryGenerationsLocked). |
| 125 | gen Gen |
| 126 | // onProven, when set, runs once a settle has proven the unit gone, |
| 127 | // with no lock held: RunCmd's republication of the generations file, |
| 128 | // whose pin for gen may have been this run's alone. Set before the |
| 129 | // unit starts and never written again. |
| 130 | onProven func() |
| 131 | } |
| 132 | |
| 133 | // cmdSettle is one settle attempt: done closes when it has ended, and err |
| 134 | // is its result, written before done closes. |
| 135 | type cmdSettle struct { |
| 136 | done chan struct{} |
| 137 | err error |
| 138 | } |
| 139 | |
| 140 | func newCmdHold(runID string, cancelPrepare context.CancelFunc, now func() time.Time) *cmdHold { |
| 141 | return &cmdHold{runID: runID, cancelPrepare: cancelPrepare, now: now, started: make(chan struct{})} |
| 142 | } |
| 143 | |
| 144 | // stop waits for the start to return, records reason and the time if it is |
| 145 | // the first stop asked for (the run's own Stop passes "", so an instance |
| 146 | // stop after it is not what ended the run), and stops and settles the unit. |
| 147 | // A unit already proven settled returns nil at once. Otherwise stop joins |
| 148 | // the attempt in flight, or starts one, and returns that attempt's result: |
| 149 | // an attempt that failed is tried again by the next stop. |
| 150 | func (h *cmdHold) stop(ctx context.Context, reason string) error { |
| 151 | select { |
| 152 | case <-h.started: |
| 153 | case <-ctx.Done(): |
| 154 | return ctx.Err() |
| 155 | } |
| 156 | h.mu.Lock() |
| 157 | if !h.asked { |
| 158 | h.asked = true |
| 159 | h.reason, h.askedAt = reason, h.now() |
| 160 | } |
| 161 | if h.unit == nil || h.proven { |
| 162 | h.mu.Unlock() |
| 163 | return nil |
| 164 | } |
| 165 | at := h.settling |
| 166 | if at == nil { |
| 167 | at = &cmdSettle{done: make(chan struct{})} |
| 168 | h.settling = at |
| 169 | go h.settle(at) |
| 170 | } |
| 171 | h.mu.Unlock() |
| 172 | select { |
| 173 | case <-at.done: |
| 174 | return at.err |
| 175 | case <-ctx.Done(): |
| 176 | return ctx.Err() |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | // settle is one attempt: the unit's StopAndSettle, unbounded by any |
| 181 | // caller's context (StopAndSettle bounds itself), so a caller that gives |
| 182 | // up waiting never cuts a settle short. |
| 183 | func (h *cmdHold) settle(at *cmdSettle) { |
| 184 | err := h.unit.StopAndSettle(context.Background()) |
| 185 | h.mu.Lock() |
| 186 | at.err = err |
| 187 | if err == nil { |
| 188 | h.proven = true |
| 189 | } |
| 190 | h.settling, h.lastTry = nil, h.now() |
| 191 | h.mu.Unlock() |
| 192 | close(at.done) |
| 193 | if err == nil && h.onProven != nil { |
| 194 | h.onProven() |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | // retrySettle starts one settle attempt in the background, for the sweep |
| 199 | // (re-review 6), when the unit has started and is not proven settled, no |
| 200 | // attempt is in flight, and the last one ended at least every ago. A hold |
| 201 | // still attached to its run is retried only once a stop has been asked |
| 202 | // for it (an interruption, or the run's own Stop that failed): a running |
| 203 | // command nobody asked to stop is left alone. A released hold is retried |
| 204 | // whatever: its run is over, and only its proof is missing. The reason a |
| 205 | // stop recorded is kept. It never blocks. |
| 206 | func (h *cmdHold) retrySettle(released bool, every time.Duration) { |
| 207 | select { |
| 208 | case <-h.started: |
| 209 | default: |
| 210 | return |
| 211 | } |
| 212 | h.mu.Lock() |
| 213 | defer h.mu.Unlock() |
| 214 | if h.unit == nil || h.proven || h.settling != nil || (!released && !h.asked) || h.now().Sub(h.lastTry) < every { |
| 215 | return |
| 216 | } |
| 217 | if !h.asked { |
| 218 | h.asked, h.askedAt = true, h.now() |
| 219 | } |
| 220 | at := &cmdSettle{done: make(chan struct{})} |
| 221 | h.settling = at |
| 222 | go h.settle(at) |
| 223 | } |
| 224 | |
| 225 | // provenSettled reports whether the hold leaves nothing running: the start |
| 226 | // returned no unit, or a settle of the unit succeeded. A failed start that |
| 227 | // reached the control service returns a unit, since the unit may be |
| 228 | // running. |
| 229 | func (h *cmdHold) provenSettled() bool { |
| 230 | select { |
| 231 | case <-h.started: |
| 232 | default: |
| 233 | return false |
| 234 | } |
| 235 | if h.unit == nil { |
| 236 | return true |
| 237 | } |
| 238 | h.mu.Lock() |
| 239 | defer h.mu.Unlock() |
| 240 | return h.proven |
| 241 | } |
| 242 | |
| 243 | type cmdRun struct { |
| 244 | a *Service |
| 245 | e *actEntry |
| 246 | h *cmdHold |
| 247 | once sync.Once |
| 248 | } |
| 249 | |
| 250 | func (r *cmdRun) Invocation() string { return r.h.unit.Invocation() } |
| 251 | func (r *cmdRun) Poll(ctx context.Context) (instancectl.CmdStatus, error) { |
| 252 | return r.h.unit.Poll(ctx) |
| 253 | } |
| 254 | func (r *cmdRun) Interrupted() (string, time.Time) { |
| 255 | r.h.mu.Lock() |
| 256 | defer r.h.mu.Unlock() |
| 257 | return r.h.reason, r.h.askedAt |
| 258 | } |
| 259 | func (r *cmdRun) Stop(ctx context.Context) error { return r.h.stop(ctx, "") } |
| 260 | |
| 261 | // Release clears the hold and stamps activity, so the instance idles a |
| 262 | // full window after a run rather than hibernating at the next sweep. A |
| 263 | // unit not proven settled (its Stop failed or never came) stays the |
| 264 | // instance's: the next stop settles it before replication closes, and |
| 265 | // keeps replication and the lease until it has (run spec §5.3, codex 1; |
| 266 | // merge ruling M1). |
| 267 | func (r *cmdRun) Release() { |
| 268 | r.once.Do(func() { |
| 269 | r.a.mu.Lock() |
| 270 | if r.e.cmd == r.h { |
| 271 | r.e.cmd = nil |
| 272 | r.e.lastActive = r.a.now() |
| 273 | if !r.h.provenSettled() { |
| 274 | r.e.cmdLeft = append(r.e.cmdLeft, r.h) |
| 275 | } |
| 276 | } |
| 277 | r.a.mu.Unlock() |
| 278 | }) |
| 279 | } |
| 280 | |
| 281 | // cmdBusyLocked reports whether e cannot take a run: one is attached, or |
| 282 | // an earlier one's unit was never proven settled. It drops the released |
| 283 | // holds that have since settled. The caller holds a.mu. |
| 284 | func cmdBusyLocked(e *actEntry) bool { |
| 285 | left := e.cmdLeft[:0] |
| 286 | for _, h := range e.cmdLeft { |
| 287 | if !h.provenSettled() { |
| 288 | left = append(left, h) |
| 289 | } |
| 290 | } |
| 291 | clear(e.cmdLeft[len(left):]) |
| 292 | e.cmdLeft = left |
| 293 | return e.cmd != nil || len(left) > 0 |
| 294 | } |
| 295 | |
| 296 | // cmdHoldsLocked is every run of e whose unit may still be running: the |
| 297 | // released runs not proven settled, and the attached run once it is |
| 298 | // starting. An attached run not yet starting is not one: RunCmd's last |
| 299 | // awake check sees a stop and abandons it before any unit starts. The |
| 300 | // caller holds a.mu. |
| 301 | func cmdHoldsLocked(e *actEntry) []*cmdHold { |
| 302 | var holds []*cmdHold |
| 303 | for _, h := range e.cmdLeft { |
| 304 | if !h.provenSettled() { |
| 305 | holds = append(holds, h) |
| 306 | } |
| 307 | } |
| 308 | if e.cmd != nil && e.cmd.starting && !e.cmd.provenSettled() { |
| 309 | holds = append(holds, e.cmd) |
| 310 | } |
| 311 | return holds |
| 312 | } |
| 313 | |
| 314 | // cmdStartInFlightLocked reports whether e's run has passed RunCmd's last |
| 315 | // check and its StartCmd has not returned: a unit may be starting against |
| 316 | // e's current generation right now. An overlap swap is not admitted, and |
| 317 | // does not switch, while one is (merge ruling M2). The caller holds a.mu. |
| 318 | func cmdStartInFlightLocked(e *actEntry) bool { |
| 319 | if e.cmd == nil || !e.cmd.starting { |
| 320 | return false |
| 321 | } |
| 322 | select { |
| 323 | case <-e.cmd.started: |
| 324 | return false |
| 325 | default: |
| 326 | return true |
| 327 | } |
| 328 | } |
| 329 | |
| 330 | // switchPendingLocked is a live overlap successor of e: a member spawning |
| 331 | // or health-checking whose owner has not given up on it (an unproven |
| 332 | // member's attempt is over). While there is one, e's current generation |
| 333 | // may retire at any moment, so RunCmd starts no unit against it, and waits |
| 334 | // for the switch or the successor's end (merge ruling M2). nil when there |
| 335 | // is none. The caller holds a.mu. |
| 336 | func switchPendingLocked(e *actEntry) *genMember { |
| 337 | for _, m := range e.ops.members { |
| 338 | if (m.role == roleSpawning || m.role == roleSuccessor) && !m.unproven { |
| 339 | return m |
| 340 | } |
| 341 | } |
| 342 | return nil |
| 343 | } |
| 344 | |
| 345 | // cmdOwnershipLocked is why e must not start a run now, or "": the entry |
| 346 | // is the host's and awake with its lease (ownedLocked), and its current |
| 347 | // generation's process is running. Each refusal is the stop reason it |
| 348 | // will become (re-review 5). The caller holds a.mu. |
| 349 | func cmdOwnershipLocked(a *Service, e *actEntry) string { |
| 350 | switch err := a.ownedLocked(e, false); { |
| 351 | case errors.Is(err, errSwapLeaseLost): |
| 352 | return StopLeaseLost |
| 353 | case errors.Is(err, errSwapFenced): |
| 354 | return StopFenced |
| 355 | case errors.Is(err, errSwapShutdown): |
| 356 | return StopShutdown |
| 357 | case err != nil: |
| 358 | return StopIdle |
| 359 | } |
| 360 | if e.proc == nil || e.proc.Exited() { |
| 361 | return StopExited |
| 362 | } |
| 363 | return "" |
| 364 | } |
| 365 | |
| 366 | // cmdSwitchPoll is how often RunCmd re-checks a pending switch it waits on, |
| 367 | // besides the successor's own done. |
| 368 | var cmdSwitchPoll = 100 * time.Millisecond |
| 369 | |
| 370 | // interruptCmdsLocked interrupts every run of e admitted against |
| 371 | // generation n, which is retiring (merge ruling M2): each unit is stopped |
| 372 | // and settled off a.mu, and the run learns why, which the member reads as |
| 373 | // interrupted (exit 125, RUN-C75). A settle that fails leaves the run |
| 374 | // unproven: its own Stop or the instance's next stop tries again, and the |
| 375 | // generations file keeps naming n's build until one succeeds. The caller |
| 376 | // holds a.mu. |
| 377 | func (a *Service) interruptCmdsLocked(e *actEntry, n uint64, why string) { |
| 378 | for _, h := range cmdHoldsLocked(e) { |
| 379 | if h.gen.N != n { |
| 380 | continue |
| 381 | } |
| 382 | go func(h *cmdHold) { |
| 383 | ctx, cancel := context.WithTimeout(context.Background(), cmdStopBound) |
| 384 | defer cancel() |
| 385 | if err := h.stop(ctx, why); err != nil { |
| 386 | log.Printf("activator: %s: run %s on retired generation %d is not proven stopped: %v; the sweep tries again", e.host, h.runID, n, err) |
| 387 | } |
| 388 | }(h) |
| 389 | } |
| 390 | } |
| 391 | |
| 392 | // retryCmdSettlesLocked is the sweep's retry of every run of e whose stop |
| 393 | // was asked for and is not proven (re-review 6): a run interrupted when a |
| 394 | // swap retired its generation, a run whose own Stop failed, and a released |
| 395 | // run left unsettled. It runs whatever state e is in, pending stop or |
| 396 | // awake, so an instance that stays awake still gets its runs proven, their |
| 397 | // pins dropped and later runs admitted. Each attempt runs off a.mu, one at |
| 398 | // a time per run, actStopTimeout apart. The caller holds a.mu. |
| 399 | func (a *Service) retryCmdSettlesLocked(e *actEntry) { |
| 400 | for _, h := range e.cmdLeft { |
| 401 | h.retrySettle(true, actStopTimeout) |
| 402 | } |
| 403 | if e.cmd != nil && e.cmd.starting { |
| 404 | e.cmd.retrySettle(false, actStopTimeout) |
| 405 | } |
| 406 | } |
| 407 | |
| 408 | // settleCmds is a run's part of an instance's termination proof (run spec |
| 409 | // §5.3; merge ruling M1): every run of e whose unit may still be running is |
| 410 | // stopped and settled, bounded by cmdStopBound, and it reports whether none |
| 411 | // may be running now. Proven means the unit stopped and its cgroup is empty |
| 412 | // (cmdUnit.StopAndSettle). A settle that fails is not given up on: the stop |
| 413 | // stays pending with replication and the lease held, and the sweep's next |
| 414 | // pass calls this again (finishPendingStops). Never call it with a.mu held. |
| 415 | func (a *Service) settleCmds(e *actEntry) bool { |
| 416 | a.mu.Lock() |
| 417 | holds, why := cmdHoldsLocked(e), e.stopWhy |
| 418 | a.mu.Unlock() |
| 419 | if len(holds) == 0 { |
| 420 | return true |
| 421 | } |
| 422 | ctx, cancel := context.WithTimeout(context.Background(), cmdStopBound) |
| 423 | defer cancel() |
| 424 | proven := true |
| 425 | for _, h := range holds { |
| 426 | if err := h.stop(ctx, why); err != nil { |
| 427 | log.Printf("activator: %s: run %s is not proven stopped: %v", e.host, h.runID, err) |
| 428 | proven = false |
| 429 | } |
| 430 | } |
| 431 | return proven |
| 432 | } |
| 433 | |
| 434 | // RunCmd runs one command against rt's instance (run spec §5.2 and §5.3): |
| 435 | // refused with claim control armed, for anything but an instance, or on a |
| 436 | // layout other than nested; the instance woken through the ordinary wake |
| 437 | // path, waiting up to 60 s; the run attached to the awake entry; |
| 438 | // spec.Prepare run under the hold, its context cancelled by a stop; and |
| 439 | // the unit started only while the entry is still awake. A stop that begins |
| 440 | // before the last awake check does not wait for the run, which then ends |
| 441 | // here (CmdStoppedError, no unit). One that begins after it waits in |
| 442 | // finishStop for the start to return and then stops the unit before |
| 443 | // replication closes. |
| 444 | // |
| 445 | // ctx is the run's own: its deadline and cancellation span the whole |
| 446 | // admission, the wake, Prepare, any wait for a swap's switch and the last |
| 447 | // checks (re-review 5). A run whose ctx ends before the start returns |
| 448 | // ctx's error and starts nothing. |
| 449 | func (a *Service) RunCmd(ctx context.Context, rt registry.Route, spec CmdSpec) (CmdHandle, error) { |
| 450 | if a.cfg.HostWrites.Enabled { |
| 451 | return nil, ErrCmdClaim |
| 452 | } |
| 453 | if rt.Kind != "instance" { |
| 454 | return nil, ErrCmdNotInstance |
| 455 | } |
| 456 | cb, ok := a.backend.(CmdBackend) |
| 457 | if !ok { |
| 458 | return nil, ErrCmdNotNested |
| 459 | } |
| 460 | if c, ok := a.backend.(Classifier); !ok { |
| 461 | return nil, ErrCmdNotNested |
| 462 | } else if l, _ := c.Layout(rt); l != LayoutNested { |
| 463 | return nil, ErrCmdNotNested |
| 464 | } |
| 465 | wctx, cancel := context.WithTimeout(ctx, cmdWakeBound) |
| 466 | defer cancel() |
| 467 | var release func() |
| 468 | for { |
| 469 | r, err := a.begin(wctx, rt, true) |
| 470 | if err == nil { |
| 471 | release = r |
| 472 | break |
| 473 | } |
| 474 | if errors.Is(err, errWakeTimeout) && wctx.Err() == nil { |
| 475 | continue // begin's own bound is 30 s; a run waits 60 |
| 476 | } |
| 477 | if wctx.Err() != nil && ctx.Err() == nil { |
| 478 | err = errWakeTimeout |
| 479 | } |
| 480 | return nil, &CmdWakeError{Reason: memberSafeWakeReason(err), Err: err} |
| 481 | } |
| 482 | pctx, pcancel := context.WithCancel(ctx) |
| 483 | defer pcancel() |
| 484 | h := newCmdHold(spec.RunID, pcancel, a.now) |
| 485 | a.mu.Lock() |
| 486 | e := a.entries[rt.Host] |
| 487 | switch { |
| 488 | case e == nil || e.state != actAwake: |
| 489 | why := StopIdle |
| 490 | if e != nil && e.stopWhy != "" { |
| 491 | why = e.stopWhy |
| 492 | } |
| 493 | a.mu.Unlock() |
| 494 | release() |
| 495 | return nil, &CmdStoppedError{Reason: why} |
| 496 | case e.worker: |
| 497 | a.mu.Unlock() |
| 498 | release() |
| 499 | return nil, ErrCmdNotInstance |
| 500 | case cmdBusyLocked(e): |
| 501 | a.mu.Unlock() |
| 502 | release() |
| 503 | return nil, ErrCmdBusy |
| 504 | } |
| 505 | e.cmd = h |
| 506 | a.mu.Unlock() |
| 507 | release() // the hold, not the in-flight count, keeps the idle timer off now |
| 508 | run := &cmdRun{a: a, e: e, h: h} |
| 509 | abandon := func(err error) (CmdHandle, error) { |
| 510 | close(h.started) |
| 511 | run.Release() |
| 512 | return nil, err |
| 513 | } |
| 514 | var prepErr error |
| 515 | if spec.Prepare != nil { |
| 516 | prepErr = spec.Prepare(pctx) |
| 517 | } |
| 518 | a.mu.Lock() |
| 519 | // waited: the run has waited for a switch since launch eligibility was |
| 520 | // last checked, by Prepare or by spec.Recheck. A wait can be long, so |
| 521 | // eligibility is checked again before anything starts (re-review 5). |
| 522 | waited := false |
| 523 | for { |
| 524 | if e.state != actAwake { |
| 525 | why := e.stopWhy |
| 526 | a.mu.Unlock() |
| 527 | return abandon(&CmdStoppedError{Reason: why}) |
| 528 | } |
| 529 | if prepErr != nil { |
| 530 | a.mu.Unlock() |
| 531 | return abandon(prepErr) |
| 532 | } |
| 533 | // ctx carries the run's deadline and cancellation (the box's run |
| 534 | // pass derives it from the record): a run cancelled or expired at |
| 535 | // any point of its admission never starts (re-review 5). |
| 536 | if err := ctx.Err(); err != nil { |
| 537 | a.mu.Unlock() |
| 538 | return abandon(err) |
| 539 | } |
| 540 | // Ownership, and a generation that is serving: the lease is held, |
| 541 | // no fence or shutdown holds the host, and the current |
| 542 | // generation's process is running. |
| 543 | if why := cmdOwnershipLocked(a, e); why != "" { |
| 544 | a.mu.Unlock() |
| 545 | return abandon(&CmdStoppedError{Reason: why}) |
| 546 | } |
| 547 | // An overlap successor is on its way: the current generation may |
| 548 | // retire before the unit is even up, so nothing starts against it. |
| 549 | // The run waits for the switch (and then runs the successor's |
| 550 | // build) or for the successor's end (merge ruling M2). A stop |
| 551 | // cancels pctx, as do the run's own cancel and deadline, and the |
| 552 | // checks above then abandon the run. |
| 553 | if m := switchPendingLocked(e); m != nil { |
| 554 | a.mu.Unlock() |
| 555 | select { |
| 556 | case <-m.done: |
| 557 | case <-time.After(cmdSwitchPoll): |
| 558 | case <-pctx.Done(): |
| 559 | } |
| 560 | waited = true |
| 561 | a.mu.Lock() |
| 562 | continue |
| 563 | } |
| 564 | if !waited || spec.Recheck == nil { |
| 565 | break |
| 566 | } |
| 567 | // The run's own eligibility, which Prepare checked before the |
| 568 | // wait, read again off a.mu (the record), then every check above |
| 569 | // once more, since the entry may have moved meanwhile. |
| 570 | a.mu.Unlock() |
| 571 | prepErr = spec.Recheck(pctx) |
| 572 | waited = false |
| 573 | a.mu.Lock() |
| 574 | } |
| 575 | // Admitted against the current generation, under the lock a swap's |
| 576 | // admission and switch take (merge ruling M2): no successor is live, |
| 577 | // and a swap admitted from here on is refused until StartCmd returns |
| 578 | // (cmdStartInFlightLocked). |
| 579 | h.starting = true |
| 580 | h.gen = e.gen |
| 581 | h.onProven = func() { |
| 582 | // The pin for h.gen is this run's alone once that generation is |
| 583 | // no longer the entry's current one. |
| 584 | a.mu.Lock() |
| 585 | stale := a.entries[e.host] != e || e.gen.N != h.gen.N |
| 586 | a.mu.Unlock() |
| 587 | if stale { |
| 588 | _ = a.persistGenerations() |
| 589 | } |
| 590 | } |
| 591 | // What the command runs is what the instance serves now: the current |
| 592 | // generation's recorded file, by path, never the live link, which a |
| 593 | // repoint, a fallback or a held update can move off it (zero-downtime |
| 594 | // spec §7.5). That holds whatever CARLOS_SWAP says, as it does for the |
| 595 | // legacy sidecar (settings ruling R-EXEC-SIDECAR): since settings Task |
| 596 | // 5c a flag-off wake runs its recorded file by path too, so a flag-off |
| 597 | // command on the live link could run a different build than its |
| 598 | // instance. Only a generation that names no file (a backend that cannot |
| 599 | // pin one) leaves the run the live link. |
| 600 | spec.serving = nil |
| 601 | if h.gen.StagedPath != "" { |
| 602 | spec.serving = &instancectl.StagedExe{Path: h.gen.StagedPath, Version: h.gen.Build, Digest: h.gen.Digest} |
| 603 | } |
| 604 | // And with the environment that generation launched with, from its |
| 605 | // handle: never the delivered files, which may hold a config revision |
| 606 | // it has not activated, a config swap queued or failed (merge ruling |
| 607 | // M3). A handle that kept none gets none, and StartCmd refuses. |
| 608 | spec.genEnv, spec.genKeyLayer = nil, instancekey.LayerNone |
| 609 | if lp, ok := e.proc.(launchEnvProc); ok { |
| 610 | spec.genEnv = lp.LaunchEnv() |
| 611 | } |
| 612 | if lp, ok := e.proc.(launchKeyLayerProc); ok { |
| 613 | spec.genKeyLayer = lp.LaunchKeyLayer() |
| 614 | } |
| 615 | a.mu.Unlock() |
| 616 | // One bound covers the build-state write and the start, so a stop |
| 617 | // waiting on the start still has the settle's budget after it |
| 618 | // (cmdStartBound). |
| 619 | sctx, scancel := context.WithTimeout(ctx, cmdStartBound) |
| 620 | defer scancel() |
| 621 | // The build about to touch the instance's database is recorded first |
| 622 | // (settings spec §4.2), as the sidecar's is (startSidecar): a command |
| 623 | // runs its serving generation's build against the data, and the build |
| 624 | // state must name that build before it does. A failed successor may |
| 625 | // have left the record naming its own build; without this write a |
| 626 | // command on the incumbent would run under a record that names |
| 627 | // another, and a held recovery could pick the wrong one. A write that |
| 628 | // fails refuses the launch (errBuildStateUnwritten): nothing starts. |
| 629 | // |
| 630 | // The starting hold is kept throughout. A stop that begins meanwhile |
| 631 | // waits for this start to return, and no overlap swap is admitted |
| 632 | // (cmdStartInFlightLocked), so the generation recorded is still the |
| 633 | // one serving when the unit starts. The write takes time, so |
| 634 | // ownership and the run's own cancel and deadline are checked again |
| 635 | // after it, before anything starts. |
| 636 | if err := a.recordBuild(sctx, rt.Host, h.gen.Build); err != nil { |
| 637 | return abandon(err) |
| 638 | } |
| 639 | a.mu.Lock() |
| 640 | var refused error |
| 641 | switch { |
| 642 | case e.state != actAwake: |
| 643 | refused = &CmdStoppedError{Reason: e.stopWhy} |
| 644 | case ctx.Err() != nil: |
| 645 | refused = ctx.Err() |
| 646 | case e.gen.N != h.gen.N: |
| 647 | // Not reached while the starting hold keeps swaps out; a |
| 648 | // generation that changed anyway is never started against. |
| 649 | refused = errSwapBusy |
| 650 | default: |
| 651 | if why := cmdOwnershipLocked(a, e); why != "" { |
| 652 | refused = &CmdStoppedError{Reason: why} |
| 653 | } |
| 654 | } |
| 655 | a.mu.Unlock() |
| 656 | if refused != nil { |
| 657 | return abandon(refused) |
| 658 | } |
| 659 | u, err := cb.StartCmd(sctx, rt, spec) |
| 660 | h.unit = u |
| 661 | close(h.started) |
| 662 | if err != nil { |
| 663 | if u != nil { |
| 664 | // The start reached the control service, so the unit may be |
| 665 | // running: settle it now. Release keeps it for the next stop |
| 666 | // if it does not settle. |
| 667 | sctx, scancel := context.WithTimeout(context.WithoutCancel(ctx), cmdStopBound) |
| 668 | _ = h.stop(sctx, "") |
| 669 | scancel() |
| 670 | } |
| 671 | run.Release() |
| 672 | return nil, err |
| 673 | } |
| 674 | a.mu.Lock() |
| 675 | stopping, why := e.state != actAwake, e.stopWhy |
| 676 | a.mu.Unlock() |
| 677 | if stopping { |
| 678 | // A stop began during the start. finishStop is waiting to stop the |
| 679 | // unit, unless its wait already timed out: stop it here too (once |
| 680 | // either way), so a late start never outlives the entry. |
| 681 | go h.stop(context.Background(), why) //nolint:errcheck |
| 682 | } |
| 683 | return run, nil |
| 684 | } |
| 685 | |
amadan