| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "os/exec" |
| 8 | |
| 9 | "github.com/carlosframework/platform/internal/cgroup" |
| 10 | "github.com/carlosframework/platform/internal/instancectl" |
| 11 | "github.com/carlosframework/platform/internal/registry" |
| 12 | ) |
| 13 | |
| 14 | var ( |
| 15 | errNoGenContainment = errors.New("activator: generation containment is off") |
| 16 | errNoGenControl = errors.New("activator: the instance control channel cannot run generation units") |
| 17 | // errGenUnproven is a generation start that failed and whose kill did |
| 18 | // not prove the generation gone. It may still be running: its listeners |
| 19 | // and launch file are kept, and a caller must keep whatever it owns for |
| 20 | // the host (a later plan's swap keeps the instance's ownership). |
| 21 | errGenUnproven = errors.New("activator: generation start failed and its termination is unproven") |
| 22 | ) |
| 23 | |
| 24 | // startUncertain marks a start's error when the start was handed to |
| 25 | // something that may have run it before failing (final review wave 2, N4): |
| 26 | // root's instance-control start call answered an error other than |
| 27 | // instancectl.ErrStagedExe, or no invocation. Root may have started the |
| 28 | // unit, and the build may have run, whatever became of it after. The |
| 29 | // wrapper changes neither the text nor errors.Is, so every existing |
| 30 | // classification and member text is the same; only startMayHaveRun reads |
| 31 | // it. |
| 32 | type startUncertain struct{ err error } |
| 33 | |
| 34 | func (e startUncertain) Error() string { return e.err.Error() } |
| 35 | func (e startUncertain) Unwrap() error { return e.err } |
| 36 | |
| 37 | // markStartUncertain wraps err as startUncertain. |
| 38 | func markStartUncertain(err error) error { return startUncertain{err: err} } |
| 39 | |
| 40 | // legacyStartUncertain marks a failed start of a nested host's original |
| 41 | // run unit (generation 0, carlos-instance@) that root was asked for and may |
| 42 | // still carry out (#292): cancelling the client's /start does not cancel a |
| 43 | // start job systemd has queued, and that job can launch a writer later. |
| 44 | // The failed wake's cleanup must stop the unit and prove it gone |
| 45 | // (legacyRunProver) before the lease goes; unproven, the lease is kept. |
| 46 | // Like startUncertain it changes neither the text nor errors.Is, so the |
| 47 | // member's wording is unchanged. |
| 48 | type legacyStartUncertain struct{ err error } |
| 49 | |
| 50 | func (e legacyStartUncertain) Error() string { return e.err.Error() } |
| 51 | func (e legacyStartUncertain) Unwrap() error { return e.err } |
| 52 | |
| 53 | // markLegacyStartUncertain wraps err as legacyStartUncertain. |
| 54 | func markLegacyStartUncertain(err error) error { return legacyStartUncertain{err: err} } |
| 55 | |
| 56 | // launchRefused marks a start that its pre-exec gate refused |
| 57 | // (runExecGate): the caller's exec gate, the wake's pending record or the |
| 58 | // recheck after it answered no, immediately before the exec or before root |
| 59 | // was asked. Nothing of that launch ran, so a failed wake whose start comes |
| 60 | // back marked is known not started, not uncertain (seeded pre-land review, |
| 61 | // P2). Like startUncertain it changes neither the text nor errors.Is. |
| 62 | type launchRefused struct{ err error } |
| 63 | |
| 64 | func (e launchRefused) Error() string { return e.err.Error() } |
| 65 | func (e launchRefused) Unwrap() error { return e.err } |
| 66 | |
| 67 | // launchNotStarted reports whether err is a start its pre-exec gate |
| 68 | // refused (launchRefused), and nothing marks it as one that may have run. |
| 69 | func launchNotStarted(err error) bool { |
| 70 | var r launchRefused |
| 71 | return errors.As(err, &r) && !legacyStartMayRun(err) && !startMayHaveRun(nil, err) |
| 72 | } |
| 73 | |
| 74 | // legacyStartMayRun reports whether err is a legacy run unit's start whose |
| 75 | // queued job may still run (legacyStartUncertain). |
| 76 | func legacyStartMayRun(err error) bool { |
| 77 | var l legacyStartUncertain |
| 78 | return errors.As(err, &l) |
| 79 | } |
| 80 | |
| 81 | // legacyRunProver is the backend half of that proof: stop the host's |
| 82 | // original run unit through instance-control and prove it gone (stop |
| 83 | // answered, unit inactive, its cgroup empty). nil is the proof. |
| 84 | // ExecBackend has it; SplitBackend passes it to its exec side. |
| 85 | type legacyRunProver interface { |
| 86 | LegacyRunGone(ctx context.Context, rt registry.Route) error |
| 87 | } |
| 88 | |
| 89 | // errNoLegacyProof: the backend cannot stop a legacy run unit, so nothing |
| 90 | // proves an aborted start gone and its lease is kept. |
| 91 | var errNoLegacyProof = errors.New("activator: this backend cannot prove a run unit stopped") |
| 92 | |
| 93 | // legacyStopBudget bounds one legacy proof: the stop call and the wait for |
| 94 | // the unit to read stopped, on a context of its own (a cancelled wake's |
| 95 | // context is often why the start was aborted). |
| 96 | const legacyStopBudget = instancectl.DeadlineControl |
| 97 | |
| 98 | // legacyRunGone runs the backend's proof for rt under legacyStopBudget and |
| 99 | // never past Shutdown's own budget (Astra, rounds 2 and 3): a proof started |
| 100 | // once Shutdown has begun runs under its context, and every proof is |
| 101 | // registered so that one already running when Shutdown began is cancelled |
| 102 | // the moment that budget ends (cancelProofs). A proof cut short is |
| 103 | // unproven, so the lease is kept. |
| 104 | func (a *Service) legacyRunGone(rt registry.Route) error { |
| 105 | p, ok := a.backend.(legacyRunProver) |
| 106 | if !ok { |
| 107 | return errNoLegacyProof |
| 108 | } |
| 109 | parent := context.Background() |
| 110 | token := new(byte) |
| 111 | a.mu.Lock() |
| 112 | if a.shutdownCtx != nil { |
| 113 | parent = a.shutdownCtx |
| 114 | } |
| 115 | ctx, cancel := context.WithTimeout(parent, legacyStopBudget) |
| 116 | if a.proofCancels == nil { |
| 117 | a.proofCancels = make(map[*byte]context.CancelFunc) |
| 118 | } |
| 119 | a.proofCancels[token] = cancel |
| 120 | a.mu.Unlock() |
| 121 | defer func() { |
| 122 | a.mu.Lock() |
| 123 | delete(a.proofCancels, token) |
| 124 | a.mu.Unlock() |
| 125 | cancel() |
| 126 | }() |
| 127 | return p.LegacyRunGone(ctx, rt) |
| 128 | } |
| 129 | |
| 130 | // cancelProofs cancels every legacy proof still running: Shutdown arms it |
| 131 | // on its own context, so none outlives the shutdown budget. |
| 132 | func (a *Service) cancelProofs() { |
| 133 | a.mu.Lock() |
| 134 | defer a.mu.Unlock() |
| 135 | for _, cancel := range a.proofCancels { |
| 136 | cancel() |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | // errStartedThenExited is the one uncertain start whose outcome is known to |
| 141 | // be a crash at the start: root started the unit and it had exited before |
| 142 | // root read its invocation (Astra R3 ruling 206, zdd-readiness Task 6). The |
| 143 | // build ran and died on its own, so a planned wake fails the build in |
| 144 | // overlap mode too, where any other uncertain start is only "may have run". |
| 145 | // It is wrapped by markStartUncertain, so startMayHaveRun still reads it. |
| 146 | var errStartedThenExited = errors.New("generation exited before it was observed") |
| 147 | |
| 148 | // exitedUnobserved is errStartedThenExited in the unit's own words. Its |
| 149 | // text is exactly the one spawnUnitFrom always wrote, because a flag-off |
| 150 | // nested wake fails with it and memberSafeWakeReason passes it to the |
| 151 | // member; only errors.Is learns the sentinel. |
| 152 | type exitedUnobserved struct{ unit string } |
| 153 | |
| 154 | func (e exitedUnobserved) Error() string { return e.unit + " exited before it was observed" } |
| 155 | func (e exitedUnobserved) Is(target error) bool { return target == errStartedThenExited } |
| 156 | |
| 157 | // startMayHaveRun is a spawn's execution evidence, tri-state collapsed to |
| 158 | // what an attempt may rely on (final review wave 2, N4): false only when it |
| 159 | // is PROVEN that nothing of the build ran (no process came back, and the |
| 160 | // error neither says the start's kill is unproven nor that the start was |
| 161 | // handed off with an uncertain outcome); true when it ran or may have. A |
| 162 | // flat start that failed is proven: the exec never happened. Every refusal |
| 163 | // before the start (a gate, a preparation step, root's ErrStagedExe) is |
| 164 | // proven too. |
| 165 | func startMayHaveRun(proc Proc, err error) bool { |
| 166 | if proc != nil || genStartUnproven(err) { |
| 167 | return true |
| 168 | } |
| 169 | var u startUncertain |
| 170 | return errors.As(err, &u) |
| 171 | } |
| 172 | |
| 173 | // spawnKind is which child shape a spawn is, for placeFor. |
| 174 | type spawnKind int |
| 175 | |
| 176 | const ( |
| 177 | spawnInstance spawnKind = iota |
| 178 | spawnSidecar // the legacy co-spawned sidecar (<host>-sidecar.env) |
| 179 | spawnSeed |
| 180 | spawnWorker // workers and native sidecar members (StartWorker) |
| 181 | ) |
| 182 | |
| 183 | // placeFor names the child cgroup a spawn lands in under the generation |
| 184 | // layout (zero-downtime spec §7.3, plan 2 D4): g<n> with the tier's full |
| 185 | // limits for an instance generation; role-sidecar and role-seed, with no |
| 186 | // limit of their own, for the legacy co-spawned sidecar and the seed, which |
| 187 | // share the instance's host cgroup and so cannot sit in a parent that holds |
| 188 | // no processes. Workers keep one cgroup per host because they never overlap |
| 189 | // (§7.9). "" (today's single-level cgroup) for every kind while |
| 190 | // GenContainment is off. |
| 191 | func (b *ExecBackend) placeFor(kind spawnKind, n uint64) (child string, ownTier bool, err error) { |
| 192 | if !b.GenContainment { |
| 193 | return "", false, nil |
| 194 | } |
| 195 | switch kind { |
| 196 | case spawnInstance: |
| 197 | if n == 0 { |
| 198 | return "", false, errors.New("activator: generation 0 is not a generation") |
| 199 | } |
| 200 | return cgroup.GenChild(n), true, nil |
| 201 | case spawnSidecar: |
| 202 | return cgroup.RoleChild("sidecar"), false, nil |
| 203 | case spawnSeed: |
| 204 | return cgroup.RoleChild("seed"), false, nil |
| 205 | } |
| 206 | return "", false, nil |
| 207 | } |
| 208 | |
| 209 | // spawnInCgroupAt is spawnInCgroupAtTier with the live tier: the spawn of a |
| 210 | // placed child whose artifact is the one the live link names. |
| 211 | func (b *ExecBackend) spawnInCgroupAt(rt registry.Route, child string, ownTier bool, build func() *exec.Cmd) (*exec.Cmd, func(), error) { |
| 212 | return b.spawnInCgroupAtTier(rt, child, ownTier, cgroup.Limit{}, build, nil) |
| 213 | } |
| 214 | |
| 215 | // spawnInCgroupAtTier is spawnInCgroup for a placed child, with the |
| 216 | // artifact's tier named: the generation child's own limit, and the parent's |
| 217 | // floor for StartChild (which never lowers a parent a swap has raised). Zero |
| 218 | // tier is b.tier(rt). With no child it IS spawnInCgroup at that tier, so a |
| 219 | // flag-off spawn with a zero tier is byte for byte today's, and a recorded |
| 220 | // build's (ruling R-TIER) caps the host cgroup at its own tier. |
| 221 | // |
| 222 | // A placed spawn is never strict: strict containment of a generation layout |
| 223 | // is the claim path's (PrepareStrictParent, plan 5). It still honours the |
| 224 | // host gate, though, so a placed spawn can never start beside a strict |
| 225 | // owner's group, and a host that is strict is refused rather than started |
| 226 | // outside its strict group. The returned settle is spawnInCgroup's: a no-op |
| 227 | // here. A spawn that fails removes the child it may have created. |
| 228 | // |
| 229 | // before, when set, is the cgroup manager's before-start hook (final review |
| 230 | // F1): asked immediately before each start attempt, after the parent and |
| 231 | // child are prepared, and on the PID-move retry too. A refusal starts |
| 232 | // nothing, is cleaned up as a failed start is, and comes back as before's |
| 233 | // own error. Nil is the spawn exactly as it was. |
| 234 | func (b *ExecBackend) spawnInCgroupAtTier(rt registry.Route, child string, ownTier bool, tier cgroup.Limit, build func() *exec.Cmd, before func() error) (*exec.Cmd, func(), error) { |
| 235 | if b.Cfg.HostWrites.Enabled { |
| 236 | return nil, nil, ErrClaimIsolationUnavailable |
| 237 | } |
| 238 | if child == "" { |
| 239 | return b.spawnInCgroupGated(rt, tier, build, before) |
| 240 | } |
| 241 | start := func() (*exec.Cmd, func(), error) { |
| 242 | t := tier |
| 243 | if t == (cgroup.Limit{}) { |
| 244 | t = b.tier(rt) |
| 245 | } |
| 246 | own := cgroup.Limit{} |
| 247 | if ownTier { |
| 248 | own = t |
| 249 | } |
| 250 | cmd, err := b.Cgroups.StartChildGated(rt.Host, child, t, own, build, before) |
| 251 | if err != nil { |
| 252 | b.Cgroups.RemoveChild(rt.Host, child) |
| 253 | return nil, nil, err |
| 254 | } |
| 255 | return cmd, func() {}, nil |
| 256 | } |
| 257 | if b.Strict == nil { |
| 258 | return start() |
| 259 | } |
| 260 | g := b.gate(rt.Host) |
| 261 | g.mu.Lock() |
| 262 | defer g.mu.Unlock() |
| 263 | if b.Strict(rt.Host) { |
| 264 | return nil, nil, fmt.Errorf("%s: %w: a generation layout is not started strict here", rt.Host, ErrStrictRequired) |
| 265 | } |
| 266 | if o := g.owner; o != nil { |
| 267 | cause := ErrStrictGroupInUse |
| 268 | if o.pending { |
| 269 | cause = ErrStrictCleanupPending |
| 270 | } |
| 271 | return nil, nil, fmt.Errorf("%s: %w", rt.Host, cause) |
| 272 | } |
| 273 | return start() |
| 274 | } |
| 275 | |
| 276 | // genControl is Control's generation half, or errNoGenControl. |
| 277 | func (b *ExecBackend) genControl() (GenControl, error) { |
| 278 | if b.Control == nil { |
| 279 | return nil, errNoGenControl |
| 280 | } |
| 281 | gc, ok := b.Control.(GenControl) |
| 282 | if !ok { |
| 283 | return nil, errNoGenControl |
| 284 | } |
| 285 | return gc, nil |
| 286 | } |
| 287 | |
| 288 | // errNoStagedExeControl refuses a start or seed of an original template |
| 289 | // that names a staged build when the control channel cannot name one. |
| 290 | // Binding the live link instead would run a build nobody recorded. |
| 291 | var errNoStagedExeControl = errors.New("activator: the instance control channel cannot name an executable for this unit") |
| 292 | |
| 293 | // stagedExeControl is Control's original-template executable half, or |
| 294 | // errNoStagedExeControl. |
| 295 | func (b *ExecBackend) stagedExeControl() (StagedExeControl, error) { |
| 296 | if b.Control == nil { |
| 297 | return nil, errNoStagedExeControl |
| 298 | } |
| 299 | xc, ok := b.Control.(StagedExeControl) |
| 300 | if !ok { |
| 301 | return nil, errNoStagedExeControl |
| 302 | } |
| 303 | return xc, nil |
| 304 | } |
| 305 | |
| 306 | // unitGen is the generation a nested run unit is started as: 0 (the |
| 307 | // original template) while GenContainment is off, n otherwise. n of 0 |
| 308 | // with containment on is refused. |
| 309 | func (b *ExecBackend) unitGen(n uint64) (uint64, error) { |
| 310 | if !b.GenContainment { |
| 311 | return 0, nil |
| 312 | } |
| 313 | if n == 0 { |
| 314 | return 0, errors.New("activator: generation 0 is not a generation") |
| 315 | } |
| 316 | return n, nil |
| 317 | } |
| 318 | |
| 319 | // startRole starts one role's unit and returns its invocation: the |
| 320 | // original template through Control, a generation through GenControl. exe |
| 321 | // names a staged build (nil: the live link, plan 2's StartGen), and tier |
| 322 | // is that build's own (zero: b.tier(rt)). |
| 323 | // |
| 324 | // At generation 0 a named exe is the build a cold wake recorded (settings |
| 325 | // spec §4.2), started through StagedExeControl. A channel without it |
| 326 | // refuses the start: dropping exe would bind the live link, a build |
| 327 | // nobody recorded. |
| 328 | func (b *ExecBackend) startRole(ctx context.Context, rt registry.Route, role instancectl.Role, gen uint64, exe *instancectl.StagedExe, tier cgroup.Limit) (string, error) { |
| 329 | if b.Cfg.HostWrites.Enabled { |
| 330 | return "", ErrClaimIsolationUnavailable |
| 331 | } |
| 332 | if gen == 0 { |
| 333 | if exe == nil { |
| 334 | return b.Control.Start(ctx, rt.Host, role, b.tier(rt).MemoryMax) |
| 335 | } |
| 336 | xc, err := b.stagedExeControl() |
| 337 | if err != nil { |
| 338 | return "", fmt.Errorf("%s: not started on %s: %w", rt.Host, exe.Version, err) |
| 339 | } |
| 340 | if tier == (cgroup.Limit{}) { |
| 341 | tier = b.tier(rt) |
| 342 | } |
| 343 | return xc.StartExe(ctx, rt.Host, role, tier.MemoryMax, *exe) |
| 344 | } |
| 345 | gc, err := b.genControl() |
| 346 | if err != nil { |
| 347 | return "", err |
| 348 | } |
| 349 | if tier == (cgroup.Limit{}) { |
| 350 | tier = b.tier(rt) |
| 351 | } |
| 352 | if exe != nil { |
| 353 | // The staged artifact's own tier, not the live link's. |
| 354 | return gc.StartGenExe(ctx, rt.Host, gen, tier.MemoryMax, *exe) |
| 355 | } |
| 356 | return gc.StartGen(ctx, rt.Host, gen, tier.MemoryMax) |
| 357 | } |
| 358 | |
| 359 | // ParentLimit sets rt's generation parent to l: carlos-exec@<host> for a |
| 360 | // flat or foreign host, the host slice for a nested one (spec §7.3). The |
| 361 | // swap raises it to cgroup.OverlapLimit before the successor spawns and |
| 362 | // restores the tier once the retiring generation's child is empty. With |
| 363 | // GenContainment off there is no parent, and it refuses, so a caller can |
| 364 | // never believe it raised a limit that does not exist. |
| 365 | func (b *ExecBackend) ParentLimit(ctx context.Context, rt registry.Route, l cgroup.Limit) error { |
| 366 | if !b.GenContainment { |
| 367 | return errNoGenContainment |
| 368 | } |
| 369 | switch p := b.pathsFor(rt); p.layout { |
| 370 | case LayoutNested: |
| 371 | gc, err := b.genControl() |
| 372 | if err != nil { |
| 373 | return err |
| 374 | } |
| 375 | return gc.Parent(ctx, rt.Host, l.MemoryMax) |
| 376 | case LayoutFenced, LayoutConflict: |
| 377 | return fmt.Errorf("%s: refusing to set a parent limit: %s", rt.Host, p.reason) |
| 378 | } |
| 379 | // The manager's SetParentLimit answers nil with no manager; here that |
| 380 | // would admit an overlap with no parent raised and no generation child |
| 381 | // to kill, so it refuses. |
| 382 | if b.Cgroups == nil { |
| 383 | return cgroup.ErrNoDelegation |
| 384 | } |
| 385 | return b.Cgroups.SetParentLimit(rt.Host, l) |
| 386 | } |
| 387 | |
| 388 | // KillGen kills generation n outright and returns once its child is empty |
| 389 | // (spec §7.3, §8.2 step 1). It acts on that generation's child only. |
| 390 | func (b *ExecBackend) KillGen(ctx context.Context, rt registry.Route, n uint64) error { |
| 391 | if !b.GenContainment || n == 0 { |
| 392 | return errNoGenContainment |
| 393 | } |
| 394 | switch p := b.pathsFor(rt); p.layout { |
| 395 | case LayoutNested: |
| 396 | gc, err := b.genControl() |
| 397 | if err != nil { |
| 398 | return err |
| 399 | } |
| 400 | return gc.StopGen(ctx, rt.Host, n, true) |
| 401 | case LayoutFenced, LayoutConflict: |
| 402 | return fmt.Errorf("%s: refusing to kill generation %d: %s", rt.Host, n, p.reason) |
| 403 | } |
| 404 | if b.Cgroups == nil { |
| 405 | return cgroup.ErrNoDelegation |
| 406 | } |
| 407 | child := cgroup.GenChild(n) |
| 408 | if err := b.killChild(ctx, rt.Host, child); err != nil { |
| 409 | return err |
| 410 | } |
| 411 | b.Cgroups.RemoveChild(rt.Host, child) |
| 412 | return nil |
| 413 | } |
| 414 | |
| 415 | // killChild kills whatever is in one child of host's generation parent and |
| 416 | // returns once that child is empty. It touches no sibling. An absent child |
| 417 | // is empty; no delegated cgroup at all is cgroup.ErrNoDelegation. The child |
| 418 | // is kept: the caller removes it (RemoveChild) once it is done with it. |
| 419 | func (b *ExecBackend) killChild(ctx context.Context, host, child string) error { |
| 420 | c, err := b.Cgroups.OpenCleanupChild(host, child) |
| 421 | if errors.Is(err, cgroup.ErrNoCgroup) { |
| 422 | return nil // already removed, so already empty |
| 423 | } |
| 424 | if err != nil { |
| 425 | return err |
| 426 | } |
| 427 | if err := c.KillAndEmpty(ctx); err != nil { |
| 428 | _ = c.Close() |
| 429 | return err |
| 430 | } |
| 431 | return c.Close() |
| 432 | } |
| 433 | |
amadan