| 1 | package instancectl |
| 2 | |
| 3 | import ( |
| 4 | "bufio" |
| 5 | "context" |
| 6 | "encoding/json" |
| 7 | "errors" |
| 8 | "flag" |
| 9 | "fmt" |
| 10 | "log" |
| 11 | "net" |
| 12 | "net/http" |
| 13 | "net/netip" |
| 14 | "os" |
| 15 | "os/exec" |
| 16 | "path/filepath" |
| 17 | "strconv" |
| 18 | "strings" |
| 19 | "sync" |
| 20 | "sync/atomic" |
| 21 | "time" |
| 22 | |
| 23 | "github.com/carlosframework/platform/internal/cgroup" |
| 24 | "github.com/carlosframework/platform/internal/cliflag" |
| 25 | "github.com/carlosframework/platform/internal/controlsock" |
| 26 | "github.com/carlosframework/platform/internal/provision" |
| 27 | "github.com/carlosframework/platform/internal/version" |
| 28 | ) |
| 29 | |
| 30 | // Server is the root instance-control service (design §5). It does four |
| 31 | // things and nothing else: starts, stops and sweeps the four templates' |
| 32 | // instances, sets a host slice's memory tier before a start, runs the |
| 33 | // readiness probe, and answers /ready. Every systemctl argument it passes |
| 34 | // is built from a validated host by UnitName/SliceName; nothing from a |
| 35 | // request reaches systemctl as a string. |
| 36 | type Server struct { |
| 37 | ClaimStop http.Handler |
| 38 | // ClaimHeldRelease is the moves held-release route's handler (moves |
| 39 | // Task 7b), registered only when set: the host sets it with both |
| 40 | // CARLOS_HOST_CLAIM and CARLOS_MOVES_PHASE1 on, so with moves off the |
| 41 | // route does not exist and main's 404 stands (whole-branch fixes, |
| 42 | // item 6). |
| 43 | ClaimHeldRelease http.Handler |
| 44 | // Systemctl runs one systemctl invocation and returns its combined |
| 45 | // output; production is exec.Command, tests record. |
| 46 | Systemctl func(ctx context.Context, args ...string) (string, error) |
| 47 | // PeerOK is asked once per accepted connection, before a byte is read. |
| 48 | // Production requires SO_PEERCRED uid carlos and pid == MainPID of |
| 49 | // carlos-edge.service (design §5.4). A refused connection is closed |
| 50 | // with no response. |
| 51 | PeerOK func(conn net.Conn) error |
| 52 | // CgroupRoot is /sys/fs/cgroup; tests point it at a fixture tree. |
| 53 | CgroupRoot string |
| 54 | // ProbeOut is where the probe unit's StandardOutput=truncate: lands. |
| 55 | ProbeOut string |
| 56 | // FixtureOK verifies /etc/carlos/probe-secret is root:carlos 0640 |
| 57 | // and /dev/shm/carlos-probe is a root-owned file before a probe |
| 58 | // verdict is accepted (design §8.5, codex R11-1, #126 R15). |
| 59 | FixtureOK func() error |
| 60 | // ProbeEvery is the re-run cadence (five minutes in production). |
| 61 | ProbeEvery time.Duration |
| 62 | Logf func(string, ...any) |
| 63 | |
| 64 | // ExeRoot is where ensureExe lays out each instance's view of its own |
| 65 | // build (#214 item 2); empty skips it, for tests of other verbs. |
| 66 | ExeRoot string |
| 67 | // OptDir is the version store's root (/opt/carlos), and HostApp reads a |
| 68 | // host's account and app from the registry. Together they are what a |
| 69 | // nested generation's staged executable is validated against |
| 70 | // (validateStagedExe). Either one unset refuses every staged start. |
| 71 | OptDir string |
| 72 | HostApp func(host string) (acct, app string, err error) |
| 73 | // DropInDir is where generation units' runtime drop-ins go |
| 74 | // (/run/systemd/system). Empty refuses every generation start: the |
| 75 | // template alone runs nothing (zero-downtime plan 2 D1). |
| 76 | DropInDir string |
| 77 | // ResolvConf is the box's resolv.conf and ResolvedConf resolved's |
| 78 | // upstream list; whichever an instance sees (resolverFile) has its |
| 79 | // nameservers re-opened through the egress fence before every probe |
| 80 | // (#214 item 3). |
| 81 | ResolvConf string |
| 82 | ResolvedConf string |
| 83 | // LocalAddrs lists what routes to this box itself; a nameserver |
| 84 | // inside it is never opened (#126). Production reads procfs |
| 85 | // (localPrefixes); nil closes every resolver. |
| 86 | LocalAddrs func() ([]netip.Prefix, error) |
| 87 | // DataRoot is the instance data tree, and Statfs reports its free and |
| 88 | // total bytes; a nil Statfs skips the disk floors (#214 item 1). |
| 89 | DataRoot string |
| 90 | Statfs func(path string) (avail, total uint64, err error) |
| 91 | DiskEvery time.Duration |
| 92 | |
| 93 | // cmdPollEvery, cmdWaitBound and cmdStartBound pace the cmd routes: |
| 94 | // how often GET /cmd/{host} re-reads the unit (250 ms), how long it |
| 95 | // holds a request (30 s, run spec §5.4), and how long POST /cmd waits |
| 96 | // for the queued start to show a new invocation (10 s). Zero is the |
| 97 | // default; tests shorten them. |
| 98 | cmdPollEvery, cmdWaitBound, cmdStartBound time.Duration |
| 99 | // cmdLocks holds one *sync.Mutex per host, taken by handleCmd from its |
| 100 | // ActiveState read to its answer, so two starts for one host cannot both |
| 101 | // pass the check. |
| 102 | cmdLocks sync.Map |
| 103 | |
| 104 | // Run executes one command as root and returns its combined output; |
| 105 | // production is exec.CommandContext, tests record. Only |
| 106 | // inspectMountpoints uses it, and every argument it passes is a |
| 107 | // constant: nothing from a request reaches it. |
| 108 | Run func(ctx context.Context, args ...string) (string, error) |
| 109 | |
| 110 | // Build is the build this process runs (version.Build), answered on |
| 111 | // every /ready as the BuildHeader so the agent can see it talks to |
| 112 | // its own release (#256). Empty sends no header, which a client of |
| 113 | // this release reads as an older service. |
| 114 | Build string |
| 115 | // ExeCurrent reports whether this process still runs the installed |
| 116 | // platform binary; nil skips the check. Production compares |
| 117 | // /proc/self/exe with /opt/carlos/carlos by identity |
| 118 | // (installedBinaryCurrent). Asked on every /ready, not once at probe |
| 119 | // time, because the binary can change at any moment after the probe |
| 120 | // ran: that is exactly the stale-service case of #256. |
| 121 | ExeCurrent func() error |
| 122 | |
| 123 | // cgroupWaitBudget bounds a wait for an empty cgroup; tests shorten it. |
| 124 | cgroupWaitBudget time.Duration |
| 125 | |
| 126 | // genMu guards genStarting, the generation units a start is preparing |
| 127 | // now. The sweep never removes the files of one of them. |
| 128 | genMu sync.Mutex |
| 129 | genStarting map[string]bool |
| 130 | |
| 131 | // Legacy run start attempts per host (legacystart.go, #292): the |
| 132 | // starts in flight, the stops in progress, and each host's barrier, |
| 133 | // the accept order of the latest stop's request. connSeq numbers |
| 134 | // connections in accept order. |
| 135 | legacyMu sync.Mutex |
| 136 | legacyStarts map[string]map[*legacyAttempt]struct{} |
| 137 | legacyStopping map[string]int |
| 138 | legacyBarrier map[string]uint64 |
| 139 | connSeq atomic.Uint64 |
| 140 | // legacyJoin overrides legacyJoinBudget; tests shorten it. |
| 141 | legacyJoin time.Duration |
| 142 | |
| 143 | mu sync.Mutex |
| 144 | report ProbeReport |
| 145 | probed bool |
| 146 | // startLines is inspectMountpoints' verdict, taken once at start and |
| 147 | // folded into every report RunProbe stores. A failing line keeps the |
| 148 | // box not ready whatever the probe says, until an operator clears the |
| 149 | // files and restarts the service. |
| 150 | startLines []VerdictLine |
| 151 | } |
| 152 | |
| 153 | const ( |
| 154 | probeUnit = "carlos-instance-probe.service" |
| 155 | edgeUnit = "carlos-edge.service" |
| 156 | parentSlices = "carlos.slice/carlos-instances.slice" |
| 157 | ) |
| 158 | |
| 159 | // Run is `carlos instance-control`. |
| 160 | func Run(args []string) { runWithClaim(args, nil, nil, nil) } |
| 161 | |
| 162 | func RunWithRegistry(args []string, hostApp func(string) (string, string, error)) { |
| 163 | runWithClaim(args, nil, nil, hostApp) |
| 164 | } |
| 165 | |
| 166 | func runWithClaim(args []string, claim, heldRelease http.Handler, hostApp func(string) (string, string, error)) { |
| 167 | fs := flag.NewFlagSet("instance-control", flag.ExitOnError) |
| 168 | fs.Usage = cliflag.Usage(fs, "usage: carlos instance-control (socket-activated by carlos-instance-control.socket)") |
| 169 | _ = fs.Parse(args) |
| 170 | if os.Geteuid() != 0 { |
| 171 | log.Fatal("instance-control: this is the root service behind carlos-instance-control.socket; it is not for operators") |
| 172 | } |
| 173 | s := &Server{ |
| 174 | ClaimStop: claim, |
| 175 | ClaimHeldRelease: heldRelease, |
| 176 | Systemctl: func(ctx context.Context, args ...string) (string, error) { |
| 177 | out, err := exec.CommandContext(ctx, "systemctl", args...).CombinedOutput() |
| 178 | return string(out), err |
| 179 | }, |
| 180 | Run: func(ctx context.Context, args ...string) (string, error) { |
| 181 | out, err := exec.CommandContext(ctx, args[0], args[1:]...).CombinedOutput() |
| 182 | return string(out), err |
| 183 | }, |
| 184 | PeerOK: productionPeerOK(), |
| 185 | CgroupRoot: "/sys/fs/cgroup", |
| 186 | ProbeOut: filepath.Join(controlsock.Dir, "probe.out"), |
| 187 | FixtureOK: probeFixturesOK, |
| 188 | ProbeEvery: 5 * time.Minute, |
| 189 | Logf: log.Printf, |
| 190 | ExeRoot: filepath.Join(controlsock.Dir, "exe"), |
| 191 | OptDir: "/opt/carlos", |
| 192 | HostApp: hostApp, |
| 193 | DropInDir: "/run/systemd/system", |
| 194 | ResolvConf: "/etc/resolv.conf", |
| 195 | ResolvedConf: ResolvedConf, |
| 196 | LocalAddrs: func() ([]netip.Prefix, error) { return localPrefixes(fibTriePath, ifInet6Path) }, |
| 197 | DataRoot: "/data/instances", |
| 198 | Statfs: statfsBytes, |
| 199 | DiskEvery: 30 * time.Second, |
| 200 | Build: version.Build, |
| 201 | ExeCurrent: installedBinaryCurrent, |
| 202 | } |
| 203 | ln, activated, err := activatedListener() |
| 204 | if err != nil { |
| 205 | log.Fatalf("instance-control: %v", err) |
| 206 | } |
| 207 | if !activated { |
| 208 | // Not socket-activated (a hand run): bind the path ourselves with |
| 209 | // the same ownership the socket unit gives it. |
| 210 | ln, err = bindControlSocket(controlsock.Path) |
| 211 | if err != nil { |
| 212 | log.Fatalf("instance-control: %v", err) |
| 213 | } |
| 214 | } |
| 215 | ctx, cancel := context.WithCancel(context.Background()) |
| 216 | defer cancel() |
| 217 | s.InspectMountpoints(ctx) |
| 218 | go s.probeLoop(ctx) |
| 219 | go s.diskLoop(ctx) |
| 220 | if err := s.Serve(ctx, ln); err != nil { |
| 221 | log.Fatalf("instance-control: %v", err) |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | // Serve accepts connections, asks PeerOK about each, and hands the accepted |
| 226 | // ones to the HTTP server one at a time through a listener that only ever |
| 227 | // yields vetted connections. |
| 228 | func (s *Server) Serve(ctx context.Context, ln net.Listener) error { |
| 229 | vetted := &vettedListener{Listener: ln, ok: s.peerOK, logf: s.logf} |
| 230 | srv := &http.Server{Handler: s.Handler(), ReadHeaderTimeout: 5 * time.Second, ConnContext: s.connContext} |
| 231 | // One request per connection (#292): a connection's accept order is the |
| 232 | // request's (connContext), which the legacy start fence relies on, so a |
| 233 | // client may not reuse one. The agent's client never does. |
| 234 | srv.SetKeepAlivesEnabled(false) |
| 235 | go func() { |
| 236 | <-ctx.Done() |
| 237 | srv.Close() |
| 238 | }() |
| 239 | err := srv.Serve(vetted) |
| 240 | if errors.Is(err, http.ErrServerClosed) || ctx.Err() != nil { |
| 241 | return nil |
| 242 | } |
| 243 | return err |
| 244 | } |
| 245 | |
| 246 | // vettedListener runs the peer check inside Accept so a refused peer is |
| 247 | // closed before net/http ever reads from it. |
| 248 | type vettedListener struct { |
| 249 | net.Listener |
| 250 | ok func(net.Conn) error |
| 251 | logf func(string, ...any) |
| 252 | |
| 253 | // lastLog and suppressed rate-limit the refusal log to at most one |
| 254 | // line per second: a refusing peer that reconnects in a tight loop |
| 255 | // (the common shape while carlos-edge.service has no MainPID yet) |
| 256 | // would otherwise flood the log at the same rate. Accept is only ever |
| 257 | // called from the one goroutine net/http.Server.Serve runs its accept |
| 258 | // loop on, so these need no lock of their own. |
| 259 | lastLog time.Time |
| 260 | suppressed int |
| 261 | } |
| 262 | |
| 263 | func (v *vettedListener) Accept() (net.Conn, error) { |
| 264 | for { |
| 265 | c, err := v.Listener.Accept() |
| 266 | if err != nil { |
| 267 | return nil, err |
| 268 | } |
| 269 | if err := v.ok(c); err != nil { |
| 270 | v.logRefusal(err) |
| 271 | c.Close() |
| 272 | continue |
| 273 | } |
| 274 | return c, nil |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | func (v *vettedListener) logRefusal(err error) { |
| 279 | now := time.Now() |
| 280 | if !v.lastLog.IsZero() && now.Sub(v.lastLog) < time.Second { |
| 281 | v.suppressed++ |
| 282 | return |
| 283 | } |
| 284 | if v.suppressed > 0 { |
| 285 | v.logf("instance-control: refused connection: %v (%d more refusals suppressed)", err, v.suppressed) |
| 286 | } else { |
| 287 | v.logf("instance-control: refused connection: %v", err) |
| 288 | } |
| 289 | v.lastLog = now |
| 290 | v.suppressed = 0 |
| 291 | } |
| 292 | |
| 293 | func (s *Server) logf(format string, args ...any) { |
| 294 | if s.Logf != nil { |
| 295 | s.Logf(format, args...) |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | // peerOK calls through s.PeerOK per connection (never captured into a |
| 300 | // closure at Serve start) so a caller that sets PeerOK after Serve has |
| 301 | // begun — a test, or a production reconfiguration — takes effect on the |
| 302 | // very next Accept instead of being silently ignored (codex review, |
| 303 | // fix round 2 item 1). |
| 304 | func (s *Server) peerOK(c net.Conn) error { return s.PeerOK(c) } |
| 305 | |
| 306 | func (s *Server) Handler() http.Handler { |
| 307 | mux := http.NewServeMux() |
| 308 | if s.ClaimStop != nil { |
| 309 | mux.Handle("POST /claim/stop", s.ClaimStop) |
| 310 | mux.Handle("POST /claim/retire", s.ClaimStop) |
| 311 | mux.Handle("POST /claim/handover", s.ClaimStop) |
| 312 | mux.Handle("POST /claim/recovery-rebind", s.ClaimStop) |
| 313 | mux.Handle("POST /claim/journal", s.ClaimStop) |
| 314 | mux.Handle("POST /claim/execution", s.ClaimStop) |
| 315 | mux.Handle("POST /claim/writers-stopped", s.ClaimStop) |
| 316 | mux.Handle("POST /claim/launch", s.ClaimStop) |
| 317 | } |
| 318 | if s.ClaimHeldRelease != nil { |
| 319 | // moves Task 7b: the held release under the claim, only with both |
| 320 | // flags on (item 6); off, the route is absent as on main. |
| 321 | mux.Handle("POST /claim/held-release", s.ClaimHeldRelease) |
| 322 | } |
| 323 | mux.HandleFunc("POST /start", s.claimAwareLegacy(s.handleStart)) |
| 324 | mux.HandleFunc("POST /seed", s.claimAwareLegacy(s.handleSeed)) |
| 325 | mux.HandleFunc("POST /stop", s.claimAwareLegacy(s.handleStop)) |
| 326 | mux.HandleFunc("POST /reset", s.claimAwareLegacy(s.handleReset)) |
| 327 | mux.HandleFunc("POST /sweep", s.claimAwareLegacy(s.handleSweep)) |
| 328 | mux.HandleFunc("POST /parent", s.claimAwareLegacy(s.handleParent)) |
| 329 | mux.HandleFunc("POST /cmd", s.claimAwareLegacy(s.handleCmd)) |
| 330 | mux.HandleFunc("GET /cmd/{host}", s.claimAwareLegacy(s.handleCmdStatus)) |
| 331 | mux.HandleFunc("GET /ready", s.handleReady) |
| 332 | return mux |
| 333 | } |
| 334 | |
| 335 | // request is the validated form of a verb's query: nothing else reaches a |
| 336 | // unit name. |
| 337 | type request struct { |
| 338 | host string |
| 339 | role Role |
| 340 | tier cgroup.Limit |
| 341 | gen uint64 // a generation of the run role (zero-downtime spec §7.6); 0 is none |
| 342 | kill bool |
| 343 | exe StagedExe // a staged build to mount (zero-downtime plan 3); Path "" is none |
| 344 | } |
| 345 | |
| 346 | // verbShape is what a verb's query may carry. |
| 347 | type verbShape struct { |
| 348 | roles []Role |
| 349 | tier bool |
| 350 | gen bool |
| 351 | kill bool |
| 352 | exe bool // a start, seed or cmd verb: the staged build its unit binds |
| 353 | // seed is the /seed verb, which names no role: its unit is the seed |
| 354 | // unit, which may bind a staged build (settings spec §4.2). |
| 355 | seed bool |
| 356 | // cmd is /cmd, whose staged build is the serving generation's, as the |
| 357 | // legacy sidecar's is (run spec §5.4 under zero-downtime spec §7.5). |
| 358 | cmd bool |
| 359 | } |
| 360 | |
| 361 | func parseRequest(r *http.Request, v verbShape) (request, error) { |
| 362 | q := r.URL.Query() |
| 363 | host := q.Get("host") |
| 364 | if !provision.ValidHostShape(host) { |
| 365 | return request{}, fmt.Errorf("host %q is not a host", host) |
| 366 | } |
| 367 | req := request{host: host} |
| 368 | if len(v.roles) > 0 { |
| 369 | role, ok := ParseRole(q.Get("role")) |
| 370 | if !ok { |
| 371 | return request{}, fmt.Errorf("role %q is not run, seed, sidecar or cmd", q.Get("role")) |
| 372 | } |
| 373 | allowed := false |
| 374 | for _, r := range v.roles { |
| 375 | allowed = allowed || r == role |
| 376 | } |
| 377 | if !allowed { |
| 378 | return request{}, fmt.Errorf("role %s is not valid for this verb", role) |
| 379 | } |
| 380 | req.role = role |
| 381 | } |
| 382 | if v.tier { |
| 383 | n, err := strconv.ParseInt(q.Get("memory_max"), 10, 64) |
| 384 | if err != nil { |
| 385 | return request{}, fmt.Errorf("memory_max %q is not a number", q.Get("memory_max")) |
| 386 | } |
| 387 | switch n { |
| 388 | case cgroup.TierDefault.MemoryMax, cgroup.TierChildren.MemoryMax: |
| 389 | req.tier = cgroup.Limit{MemoryMax: n} |
| 390 | default: |
| 391 | return request{}, fmt.Errorf("memory_max %d is not a platform tier", n) |
| 392 | } |
| 393 | } |
| 394 | if q.Has("gen") { |
| 395 | raw := q.Get("gen") |
| 396 | if !v.gen { |
| 397 | return request{}, errors.New("this verb takes no generation") |
| 398 | } |
| 399 | n, err := strconv.ParseUint(raw, 10, 64) |
| 400 | if err != nil || n == 0 || strconv.FormatUint(n, 10) != raw { |
| 401 | return request{}, fmt.Errorf("gen %q is not a generation", raw) |
| 402 | } |
| 403 | if req.role != RoleRun { |
| 404 | return request{}, fmt.Errorf("role %s has no generations", req.role) |
| 405 | } |
| 406 | req.gen = n |
| 407 | } |
| 408 | if q.Has("kill") { |
| 409 | if !v.kill || req.gen == 0 || q.Get("kill") != "1" { |
| 410 | return request{}, errors.New("kill=1 belongs to a generation stop only") |
| 411 | } |
| 412 | req.kill = true |
| 413 | } |
| 414 | if q.Has("exe") || q.Has("version") || q.Has("digest") { |
| 415 | if !v.exe { |
| 416 | return request{}, errors.New("this verb takes no staged executable") |
| 417 | } |
| 418 | if x := q.Get("exe"); x != "" { |
| 419 | // A generation start; the legacy sidecar (zero-downtime spec |
| 420 | // §7.5: it runs its instance's serving artifact); a run's |
| 421 | // command, which runs that same artifact; or the original |
| 422 | // template's run unit and the seed unit, which run the build a |
| 423 | // cold wake recorded (settings spec §4.2). Every one is |
| 424 | // validated before anything is laid out (refuseStagedExe). |
| 425 | if req.gen == 0 && !v.seed && !v.cmd && req.role != RoleRun && req.role != RoleSidecar { |
| 426 | return request{}, errors.New("exe is only for a run, seed or sidecar unit, a run's command, or a generation start") |
| 427 | } |
| 428 | req.exe = StagedExe{Path: x, Version: q.Get("version"), Digest: q.Get("digest")} |
| 429 | } else { |
| 430 | return request{}, errors.New("version and digest need an exe") |
| 431 | } |
| 432 | } |
| 433 | return req, nil |
| 434 | } |
| 435 | |
| 436 | // setTier writes the whole tier onto the host slice in one call: the |
| 437 | // memory caps, and the CPU and IO shares that follow from the same tier |
| 438 | // (#214 item 1). Seed, instance and sidecar share all of it. |
| 439 | func (s *Server) setTier(ctx context.Context, host string, tier cgroup.Limit) error { |
| 440 | return sliceTier(ctx, s.Systemctl, host, tier) |
| 441 | } |
| 442 | |
| 443 | // sliceTier is setTier's one systemctl call, shared with the claim launch. |
| 444 | func sliceTier(ctx context.Context, systemctl func(context.Context, ...string) (string, error), host string, tier cgroup.Limit) error { |
| 445 | max := strconv.FormatInt(tier.MemoryMax, 10) |
| 446 | high := strconv.FormatInt(cgroup.HighOf(tier.MemoryMax), 10) |
| 447 | sh := cgroup.SharesOf(tier) |
| 448 | _, err := systemctl(ctx, "set-property", "--runtime", SliceName(host), |
| 449 | "MemoryMax="+max, "MemoryHigh="+high, "MemorySwapMax="+max, |
| 450 | "CPUWeight="+strconv.Itoa(sh.CPUWeight), "CPUQuota="+strconv.Itoa(sh.CPUQuotaPercent)+"%", |
| 451 | "IOWeight="+strconv.Itoa(sh.IOWeight)) |
| 452 | return err |
| 453 | } |
| 454 | |
| 455 | func writeJSON(w http.ResponseWriter, status int, v any) { |
| 456 | w.Header().Set("Content-Type", "application/json") |
| 457 | w.WriteHeader(status) |
| 458 | _ = json.NewEncoder(w).Encode(v) |
| 459 | } |
| 460 | |
| 461 | func (s *Server) handleStart(w http.ResponseWriter, r *http.Request) { |
| 462 | req, err := parseRequest(r, verbShape{roles: []Role{RoleRun, RoleSidecar}, tier: true, gen: true, exe: true}) |
| 463 | if err != nil { |
| 464 | http.Error(w, err.Error(), http.StatusBadRequest) |
| 465 | return |
| 466 | } |
| 467 | if req.gen != 0 { |
| 468 | s.startGen(w, r, req) |
| 469 | return |
| 470 | } |
| 471 | ctx := r.Context() |
| 472 | if req.role == RoleRun { |
| 473 | // An attempt a /stop of this host can cancel and join, and that a |
| 474 | // stop which has already run refuses (legacystart.go, #292). |
| 475 | attempt, finish, err := s.beginLegacyStart(ctx, req.host, requestSeq(ctx)) |
| 476 | if err != nil { |
| 477 | http.Error(w, "start: "+err.Error(), http.StatusConflict) |
| 478 | return |
| 479 | } |
| 480 | defer finish() |
| 481 | ctx = attempt |
| 482 | } |
| 483 | unit := UnitName(req.role, req.host) |
| 484 | // A sidecar naming its generation's staged build, or a run unit naming |
| 485 | // the build its cold wake recorded, is validated before anything is |
| 486 | // laid out, under the generation rule (validateStagedExe). |
| 487 | if s.refuseStagedExe(w, req, unit) { |
| 488 | return |
| 489 | } |
| 490 | if req.role == RoleRun { |
| 491 | if err := s.clearLegacyClaimLaunch(ctx, req.host); err != nil { |
| 492 | http.Error(w, "start: "+err.Error(), http.StatusBadGateway) |
| 493 | return |
| 494 | } |
| 495 | } |
| 496 | if s.refuseStart(w, req.host, unit, func() error { return s.ensureHostExeAt(req.host, req.exe) }) { |
| 497 | return |
| 498 | } |
| 499 | // A sidecar start names its instance's staged build with generations |
| 500 | // on or off (settings ruling R-EXEC-SIDECAR). With them on, the host |
| 501 | // slice is the generation parent (spec §7.3): a swap may have raised it |
| 502 | // to two tiers, and lowering it here would put both generations under |
| 503 | // one tier's limit. Its restore is the swap owner's, so a populated |
| 504 | // slice is left alone, exactly as startGen does; a cold slice still gets |
| 505 | // the tier. With them off, a populated slice was tiered by the |
| 506 | // instance's own run start of that same recorded build, so leaving it |
| 507 | // alone keeps the right tier too. A start without exe (a live-link |
| 508 | // sidecar) sets the tier as before. The exemption is the sidecar's |
| 509 | // alone: a run unit naming its recorded build starts a cold instance, |
| 510 | // and sets its tier as a run start always has. |
| 511 | if req.role == RoleSidecar && req.exe.Path != "" && s.slicePopulated(req.host) { |
| 512 | s.logf("instance-control: %s: slice populated, leaving its limit to the generation owner", req.host) |
| 513 | } else if err := s.setTier(ctx, req.host, req.tier); err != nil { |
| 514 | http.Error(w, "set-property: "+err.Error(), http.StatusBadGateway) |
| 515 | return |
| 516 | } |
| 517 | // Never submitted once cancelled (#292): a /stop of this host cancels |
| 518 | // the attempt before it joins it, and the client may have gone. |
| 519 | if err := ctx.Err(); err != nil { |
| 520 | http.Error(w, "start: cancelled before it was submitted: "+err.Error(), http.StatusConflict) |
| 521 | return |
| 522 | } |
| 523 | if out, err := s.Systemctl(ctx, "start", unit); err != nil { |
| 524 | http.Error(w, "start: "+strings.TrimSpace(out+" "+err.Error()), http.StatusBadGateway) |
| 525 | return |
| 526 | } |
| 527 | inv, err := s.Systemctl(ctx, "show", "-p", "InvocationID", "--value", unit) |
| 528 | if err != nil { |
| 529 | http.Error(w, "show: "+err.Error(), http.StatusBadGateway) |
| 530 | return |
| 531 | } |
| 532 | s.logf("instance-control: %s: started %s memory.max=%d", req.host, unit, req.tier.MemoryMax) |
| 533 | writeJSON(w, http.StatusOK, StartResponse{InvocationID: strings.TrimSpace(inv)}) |
| 534 | } |
| 535 | |
| 536 | // parseShow reads systemctl show's KEY=value lines. |
| 537 | func parseShow(out string) map[string]string { |
| 538 | m := map[string]string{} |
| 539 | sc := bufio.NewScanner(strings.NewReader(out)) |
| 540 | for sc.Scan() { |
| 541 | if k, v, ok := strings.Cut(sc.Text(), "="); ok { |
| 542 | m[k] = strings.TrimSpace(v) |
| 543 | } |
| 544 | } |
| 545 | return m |
| 546 | } |
| 547 | |
| 548 | // handleSeed runs the seed oneshot once and returns its retained result. |
| 549 | // The unit's InvocationID is read before and after the first `start`: if |
| 550 | // start left the SAME invocation in place, an earlier request's stop |
| 551 | // never ran (its client disconnected before this fix existed, or the |
| 552 | // agent restarted mid-seed), so this handler stops the stale run and |
| 553 | // starts again rather than reading and returning someone else's result. |
| 554 | // |
| 555 | // The request context ctx is used only up to and including that first |
| 556 | // `start`. From the moment it returns, the unit is this server's |
| 557 | // responsibility to clean up, whether or not the client that asked for |
| 558 | // it is still connected: `start` is the blocking call, and production's |
| 559 | // Systemctl is exec.CommandContext, which kills the systemctl process |
| 560 | // outright when its context is cancelled. A client that hung up while |
| 561 | // `start` was running would otherwise cancel every systemctl call after |
| 562 | // it too: the very next `show` would fail, the handler would 502, and |
| 563 | // the cleanup `stop` after it would never run (fix round 1, item 4). So |
| 564 | // every call from here to the end of the handler runs on ONE context |
| 565 | // (dctx) detached from the request instead. The order is show (the |
| 566 | // fields RemainAfterExit retains), then stop, then wait for the cgroup |
| 567 | // to empty: a populated cgroup after the wait is still reported as a 502 |
| 568 | // naming the unit. |
| 569 | // |
| 570 | // A seed may name its executable (settings spec §4.2): the staged build |
| 571 | // the wake recorded. Root validates it as it does a generation's (403 on |
| 572 | // a refusal, before anything is laid out or started) and binds the host's |
| 573 | // app mount to it. Without one the mount is the live link, as before. |
| 574 | func (s *Server) handleSeed(w http.ResponseWriter, r *http.Request) { |
| 575 | req, err := parseRequest(r, verbShape{tier: true, exe: true, seed: true}) |
| 576 | if err != nil { |
| 577 | http.Error(w, err.Error(), http.StatusBadRequest) |
| 578 | return |
| 579 | } |
| 580 | ctx := r.Context() |
| 581 | unit := UnitName(RoleSeed, req.host) |
| 582 | if s.refuseStagedExe(w, req, unit) { |
| 583 | return |
| 584 | } |
| 585 | if s.refuseStart(w, req.host, unit, func() error { return s.ensureHostExeAt(req.host, req.exe) }) { |
| 586 | return |
| 587 | } |
| 588 | if err := s.setTier(ctx, req.host, req.tier); err != nil { |
| 589 | http.Error(w, "set-property: "+err.Error(), http.StatusBadGateway) |
| 590 | return |
| 591 | } |
| 592 | |
| 593 | // A oneshot's start job completes when ExecStart exits; RemainAfterExit |
| 594 | // keeps the unit loaded with its result until the stop below. |
| 595 | invBefore := s.showValue(ctx, "InvocationID", unit) |
| 596 | startOut, startErr := s.Systemctl(ctx, "start", unit) |
| 597 | |
| 598 | // See the handler doc comment: everything past this point belongs to |
| 599 | // the server, not the client, so it runs on a context of the server's |
| 600 | // own rather than one the client's disconnect can cancel. |
| 601 | dctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second) |
| 602 | defer cancel() |
| 603 | |
| 604 | invAfter := s.showValue(dctx, "InvocationID", unit) |
| 605 | if invBefore != "" && invBefore == invAfter { |
| 606 | s.stopUnit(dctx, unit) |
| 607 | startOut, startErr = s.Systemctl(dctx, "start", unit) |
| 608 | if retry := s.showValue(dctx, "InvocationID", unit); retry == "" || retry == invAfter { |
| 609 | // Still this server's unit to clean up even though the request |
| 610 | // is about to fail: leaving it running would strand the NEXT |
| 611 | // /seed on this same stale-invocation guard forever. |
| 612 | s.stopUnit(dctx, unit) |
| 613 | http.Error(w, "seed: "+unit+" would not leave its stale invocation "+invAfter, http.StatusBadGateway) |
| 614 | return |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | out, err := s.Systemctl(dctx, "show", "-p", "InvocationID,Result,ExecMainStatus,ExecMainCode", unit) |
| 619 | if err != nil { |
| 620 | http.Error(w, "show: "+err.Error(), http.StatusBadGateway) |
| 621 | return |
| 622 | } |
| 623 | m := parseShow(out) |
| 624 | s.stopUnit(dctx, unit) |
| 625 | wctx, waitCancel := context.WithTimeout(dctx, s.waitBudget()) |
| 626 | defer waitCancel() |
| 627 | if err := WaitCgroupEmpty(wctx, s.CgroupRoot, SliceName(req.host), unit, 50*time.Millisecond); err != nil { |
| 628 | http.Error(w, "seed: "+err.Error(), http.StatusBadGateway) |
| 629 | return |
| 630 | } |
| 631 | if startErr != nil && m["Result"] == "" { |
| 632 | http.Error(w, "start: "+strings.TrimSpace(startOut+" "+startErr.Error()), http.StatusBadGateway) |
| 633 | return |
| 634 | } |
| 635 | writeJSON(w, http.StatusOK, SeedResponse{InvocationID: m["InvocationID"], Result: m["Result"], Status: m["ExecMainStatus"], Code: m["ExecMainCode"]}) |
| 636 | } |
| 637 | |
| 638 | // showValue reads one systemd property's --value for unit, returning "" on |
| 639 | // any error. It is a probe, not a hard failure path: the caller's own |
| 640 | // later show call is what surfaces a real systemctl error to the client. |
| 641 | func (s *Server) showValue(ctx context.Context, prop, unit string) string { |
| 642 | out, err := s.Systemctl(ctx, "show", "-p", prop, "--value", unit) |
| 643 | if err != nil { |
| 644 | return "" |
| 645 | } |
| 646 | return strings.TrimSpace(out) |
| 647 | } |
| 648 | |
| 649 | // stopUnit issues `systemctl stop` for unit on ctx, logging rather than |
| 650 | // surfacing a failure to the caller: every call site in handleSeed is |
| 651 | // already past the point where the unit is this server's responsibility, |
| 652 | // not something a failed stop should turn into a different HTTP error for |
| 653 | // the client. |
| 654 | func (s *Server) stopUnit(ctx context.Context, unit string) { |
| 655 | if _, err := s.Systemctl(ctx, "stop", unit); err != nil { |
| 656 | s.logf("instance-control: stop %s: %v", unit, err) |
| 657 | } |
| 658 | } |
| 659 | |
| 660 | func (s *Server) handleStop(w http.ResponseWriter, r *http.Request) { |
| 661 | req, err := parseRequest(r, verbShape{roles: []Role{RoleRun, RoleSeed, RoleSidecar, RoleCmd}, gen: true, kill: true}) |
| 662 | if err != nil { |
| 663 | http.Error(w, err.Error(), http.StatusBadRequest) |
| 664 | return |
| 665 | } |
| 666 | if req.gen != 0 { |
| 667 | s.stopGen(w, r, req) |
| 668 | return |
| 669 | } |
| 670 | if req.role == RoleRun { |
| 671 | // Every start of this host's run unit in flight is cancelled and |
| 672 | // joined first, and earlier requests' starts are refused from here |
| 673 | // on (legacystart.go, #292): once this answers, no start from an |
| 674 | // earlier request can still be submitted, and the stop below |
| 675 | // replaces any job a joined start had submitted. |
| 676 | release, err := s.fenceLegacyStarts(r.Context(), req.host) |
| 677 | defer release() |
| 678 | if err != nil { |
| 679 | // Stopped anyway, but not proven final: a start may follow. |
| 680 | _, _ = s.Systemctl(r.Context(), "stop", UnitName(req.role, req.host)) |
| 681 | http.Error(w, "stop: "+err.Error(), http.StatusBadGateway) |
| 682 | return |
| 683 | } |
| 684 | } |
| 685 | if out, err := s.Systemctl(r.Context(), "stop", UnitName(req.role, req.host)); err != nil { |
| 686 | http.Error(w, "stop: "+strings.TrimSpace(out+" "+err.Error()), http.StatusBadGateway) |
| 687 | return |
| 688 | } |
| 689 | if req.role == RoleRun { |
| 690 | if err := s.clearLegacyClaimLaunch(r.Context(), req.host); err != nil { |
| 691 | http.Error(w, "stop: "+err.Error(), http.StatusBadGateway) |
| 692 | return |
| 693 | } |
| 694 | } |
| 695 | if req.role == RoleCmd { |
| 696 | s.cleanCmdExe(r.Context(), req.host) |
| 697 | } |
| 698 | w.WriteHeader(http.StatusNoContent) |
| 699 | } |
| 700 | |
| 701 | func (s *Server) handleReset(w http.ResponseWriter, r *http.Request) { |
| 702 | req, err := parseRequest(r, verbShape{roles: []Role{RoleRun, RoleSeed, RoleSidecar, RoleCmd}, gen: true}) |
| 703 | if err != nil { |
| 704 | http.Error(w, err.Error(), http.StatusBadRequest) |
| 705 | return |
| 706 | } |
| 707 | unit := UnitName(req.role, req.host) |
| 708 | if req.gen != 0 { |
| 709 | unit = GenUnitName(req.host, req.gen) |
| 710 | } |
| 711 | if out, err := s.Systemctl(r.Context(), "reset-failed", unit); err != nil && !strings.Contains(out+err.Error(), "not loaded") { |
| 712 | http.Error(w, "reset-failed: "+strings.TrimSpace(out+" "+err.Error()), http.StatusBadGateway) |
| 713 | return |
| 714 | } |
| 715 | w.WriteHeader(http.StatusNoContent) |
| 716 | } |
| 717 | |
| 718 | // handleSweep stops every instance of the four templates that is not |
| 719 | // inactive or has a job, keyed on the names systemd reports, so a unit |
| 720 | // whose identifier is hashed and whose row is gone is still stopped |
| 721 | // (design §6.7, codex R7-1). |
| 722 | func (s *Server) handleSweep(w http.ResponseWriter, r *http.Request) { |
| 723 | ctx := r.Context() |
| 724 | targets := map[string]bool{} |
| 725 | out, err := s.Systemctl(ctx, "list-units", "--all", "--plain", "--no-legend", "--no-pager", |
| 726 | UnitPrefixRun+"*.service", UnitPrefixSeed+"*.service", UnitPrefixSidecar+"*.service", UnitPrefixCmd+"*.service", UnitPrefixGen+"*.service") |
| 727 | if err != nil { |
| 728 | http.Error(w, "list-units: "+err.Error(), http.StatusBadGateway) |
| 729 | return |
| 730 | } |
| 731 | for _, line := range strings.Split(out, "\n") { |
| 732 | f := strings.Fields(line) |
| 733 | if len(f) < 3 { |
| 734 | continue |
| 735 | } |
| 736 | if _, _, ok := ParseUnitName(f[0]); !ok { |
| 737 | continue |
| 738 | } |
| 739 | if f[2] != "inactive" { |
| 740 | targets[f[0]] = true |
| 741 | } |
| 742 | } |
| 743 | jobs, err := s.Systemctl(ctx, "list-jobs", "--plain", "--no-legend", "--no-pager") |
| 744 | if err != nil { |
| 745 | http.Error(w, "list-jobs: "+err.Error(), http.StatusBadGateway) |
| 746 | return |
| 747 | } |
| 748 | for _, line := range strings.Split(jobs, "\n") { |
| 749 | f := strings.Fields(line) |
| 750 | if len(f) < 2 { |
| 751 | continue |
| 752 | } |
| 753 | if _, _, ok := ParseUnitName(f[1]); ok { |
| 754 | targets[f[1]] = true |
| 755 | } |
| 756 | } |
| 757 | // Every target gets stopped and waited on; a stuck unit is recorded and |
| 758 | // the sweep continues to the rest rather than aborting on the first one |
| 759 | // (design §6.7, codex review fix round 2 item 3): a sweep exists to |
| 760 | // clear the box for the next thing, and giving up after unit one would |
| 761 | // leave every unit after it running. |
| 762 | var stopped []string |
| 763 | failed := map[string]string{} |
| 764 | for unit := range targets { |
| 765 | wctx, cancel := context.WithTimeout(ctx, s.waitBudget()) |
| 766 | if _, err := s.Systemctl(ctx, "stop", unit); err != nil { |
| 767 | failed[unit] = "stop: " + err.Error() |
| 768 | cancel() |
| 769 | continue |
| 770 | } |
| 771 | // A unit that is still activating has no cgroup yet; wait for |
| 772 | // systemd to settle it to inactive or failed before trusting an |
| 773 | // absent cgroup as "swept" (codex review fix round 2 item 4). |
| 774 | if err := s.waitInactiveOrFailed(wctx, unit, 50*time.Millisecond); err != nil { |
| 775 | failed[unit] = err.Error() |
| 776 | cancel() |
| 777 | continue |
| 778 | } |
| 779 | _, esc, _ := ParseUnitName(unit) |
| 780 | err := WaitCgroupEmpty(wctx, s.CgroupRoot, "carlos-instances-"+esc+".slice", unit, 50*time.Millisecond) |
| 781 | cancel() |
| 782 | if err != nil { |
| 783 | failed[unit] = err.Error() |
| 784 | continue |
| 785 | } |
| 786 | stopped = append(stopped, unit) |
| 787 | s.removeGenFiles(unit) |
| 788 | if role, esc, _ := ParseUnitName(unit); role == RoleCmd { |
| 789 | if iid, ok := Unescape(esc); ok { |
| 790 | s.removeCmdExe(iid) |
| 791 | } |
| 792 | } |
| 793 | } |
| 794 | s.sweepGenLeftovers(ctx, failed) |
| 795 | if len(failed) > 0 { |
| 796 | writeJSON(w, http.StatusBadGateway, SweepResponse{Units: stopped, Failed: failed}) |
| 797 | return |
| 798 | } |
| 799 | writeJSON(w, http.StatusOK, SweepResponse{Units: stopped}) |
| 800 | } |
| 801 | |
| 802 | // waitInactiveOrFailed polls a unit's ActiveState until systemd reports it |
| 803 | // settled (inactive after a clean stop, failed after one that wasn't), |
| 804 | // bounded by ctx. |
| 805 | func (s *Server) waitInactiveOrFailed(ctx context.Context, unit string, every time.Duration) error { |
| 806 | var last string |
| 807 | for { |
| 808 | if out, err := s.Systemctl(ctx, "show", "-p", "ActiveState", "--value", unit); err == nil { |
| 809 | last = strings.TrimSpace(out) |
| 810 | if last == "inactive" || last == "failed" { |
| 811 | return nil |
| 812 | } |
| 813 | } |
| 814 | select { |
| 815 | case <-ctx.Done(): |
| 816 | return fmt.Errorf("%s: ActiveState still %q", unit, last) |
| 817 | case <-time.After(every): |
| 818 | } |
| 819 | } |
| 820 | } |
| 821 | |
| 822 | func (s *Server) handleReady(w http.ResponseWriter, r *http.Request) { |
| 823 | if s.Build != "" { |
| 824 | w.Header().Set(BuildHeader, s.Build) |
| 825 | } |
| 826 | rep, ready := s.Ready() |
| 827 | if ready { |
| 828 | w.WriteHeader(http.StatusNoContent) |
| 829 | return |
| 830 | } |
| 831 | writeJSON(w, http.StatusServiceUnavailable, rep) |
| 832 | } |
| 833 | |
| 834 | // Ready returns the last probe verdict, plus the one check that cannot |
| 835 | // wait for the next probe: that this process still runs the installed |
| 836 | // binary (ExeCurrent). Before the first probe has ever run, there is no |
| 837 | // verdict to report, so a synthetic line names that reason explicitly |
| 838 | // rather than answering with an empty Lines (final review, plan 2, item |
| 839 | // 6): a caller reading /ready during startup should see why, not a report |
| 840 | // that looks like every check silently passed. |
| 841 | func (s *Server) Ready() (ProbeReport, bool) { |
| 842 | s.mu.Lock() |
| 843 | rep, probed := s.report, s.probed |
| 844 | s.mu.Unlock() |
| 845 | if !probed { |
| 846 | rep = ProbeReport{Lines: []VerdictLine{{Check: "probe", Verdict: "fail", Detail: "not yet run"}}} |
| 847 | } |
| 848 | if s.ExeCurrent != nil { |
| 849 | if err := s.ExeCurrent(); err != nil { |
| 850 | // A copy: the stored report must not grow a line per call. |
| 851 | rep = ProbeReport{Ready: false, Lines: append(append([]VerdictLine(nil), rep.Lines...), |
| 852 | VerdictLine{Check: "binary", Verdict: "fail", Detail: err.Error()})} |
| 853 | } |
| 854 | } |
| 855 | return rep, rep.Ready |
| 856 | } |
| 857 | |
| 858 | // installedBinaryCurrent is ExeCurrent in production: /proc/self/exe |
| 859 | // resolves to the file this process was started from, even after that |
| 860 | // file was renamed over (the kernel reports it as deleted), and |
| 861 | // os.SameFile compares device and inode, so a service started from the |
| 862 | // previous release's binary answers a mismatch the moment carlos system |
| 863 | // apply renames the new one into place, and keeps answering it until it |
| 864 | // is restarted (#256). |
| 865 | func installedBinaryCurrent() error { |
| 866 | running, err := os.Stat("/proc/self/exe") |
| 867 | if err != nil { |
| 868 | return fmt.Errorf("stat /proc/self/exe: %v", err) |
| 869 | } |
| 870 | installed, err := os.Stat(SystemBinary) |
| 871 | if err != nil { |
| 872 | return fmt.Errorf("stat %s: %v", SystemBinary, err) |
| 873 | } |
| 874 | if !os.SameFile(running, installed) { |
| 875 | return fmt.Errorf("instance control is running a binary that is no longer %s; restart carlos-instance-control.service", SystemBinary) |
| 876 | } |
| 877 | return nil |
| 878 | } |
| 879 | |
| 880 | func (s *Server) waitBudget() time.Duration { |
| 881 | if s.cgroupWaitBudget > 0 { |
| 882 | return s.cgroupWaitBudget |
| 883 | } |
| 884 | return 30 * time.Second |
| 885 | } |
| 886 | |
| 887 | // WaitCgroupEmpty waits until the unit's cgroup under the host slice is |
| 888 | // absent or reports "populated 0" in cgroup.events, polling every `every`, |
| 889 | // bounded by ctx and the caller's budget (design §6.5, codex R2-5). |
| 890 | func WaitCgroupEmpty(ctx context.Context, root, slice, unit string, every time.Duration) error { |
| 891 | events := filepath.Join(root, parentSlices, slice, unit, "cgroup.events") |
| 892 | for { |
| 893 | b, err := os.ReadFile(events) |
| 894 | switch { |
| 895 | case errors.Is(err, os.ErrNotExist): |
| 896 | return nil |
| 897 | case err != nil: |
| 898 | return fmt.Errorf("%s: %w", unit, err) |
| 899 | case strings.Contains(string(b), "populated 0"): |
| 900 | return nil |
| 901 | } |
| 902 | select { |
| 903 | case <-ctx.Done(): |
| 904 | return fmt.Errorf("processes remain in %s", unit) |
| 905 | case <-time.After(every): |
| 906 | } |
| 907 | } |
| 908 | } |
| 909 | |
| 910 | // probeLoop runs the probe at start and every ProbeEvery. |
| 911 | func (s *Server) probeLoop(ctx context.Context) { |
| 912 | s.RunProbe(ctx) |
| 913 | t := time.NewTicker(s.ProbeEvery) |
| 914 | defer t.Stop() |
| 915 | for { |
| 916 | select { |
| 917 | case <-ctx.Done(): |
| 918 | return |
| 919 | case <-t.C: |
| 920 | s.RunProbe(ctx) |
| 921 | } |
| 922 | } |
| 923 | } |
| 924 | |
| 925 | // RunProbe starts the probe unit (a oneshot: start blocks until it exits), |
| 926 | // reads the report it left, and stores the verdict. A fixture that is not |
| 927 | // root:carlos 0640, a probe that could not be started, or a report that |
| 928 | // could not be read all leave the box not ready, with the reason on a |
| 929 | // verdict line; readiness is never the default. |
| 930 | func (s *Server) RunProbe(ctx context.Context) { |
| 931 | set := func(r ProbeReport) { |
| 932 | s.mu.Lock() |
| 933 | s.report, s.probed = r, true |
| 934 | s.mu.Unlock() |
| 935 | } |
| 936 | if err := s.FixtureOK(); err != nil { |
| 937 | set(ProbeReport{Lines: []VerdictLine{{Check: "fixture", Verdict: "fail", Detail: err.Error()}}}) |
| 938 | return |
| 939 | } |
| 940 | // The probe binds the same /opt view an instance does, so its sources |
| 941 | // must exist (#214 item 2), and the resolver allow it checks must be |
| 942 | // in place (#214 item 3). |
| 943 | if s.ExeRoot != "" { |
| 944 | if err := ensureExe(s.ExeRoot, ProbeDirName, SystemBinary, ProbeDirName); err != nil { |
| 945 | set(ProbeReport{Lines: []VerdictLine{{Check: "probe-exe", Verdict: "fail", Detail: err.Error()}}}) |
| 946 | return |
| 947 | } |
| 948 | } |
| 949 | var resolverLines []VerdictLine |
| 950 | if s.ResolvConf != "" { |
| 951 | refused, err := s.allowResolvers(ctx) |
| 952 | if err != nil { |
| 953 | set(ProbeReport{Lines: []VerdictLine{{Check: "resolver-allow", Verdict: "fail", Detail: err.Error()}}}) |
| 954 | return |
| 955 | } |
| 956 | resolverLines = append(resolverLines, resolverLocalLine(refused)) |
| 957 | } |
| 958 | if out, err := s.Systemctl(ctx, "start", probeUnit); err != nil { |
| 959 | set(ProbeReport{Lines: []VerdictLine{{Check: "probe-unit", Verdict: "fail", Detail: strings.TrimSpace(out + " " + err.Error())}}}) |
| 960 | return |
| 961 | } |
| 962 | b, err := os.ReadFile(s.ProbeOut) |
| 963 | if err != nil { |
| 964 | set(ProbeReport{Lines: []VerdictLine{{Check: "probe-output", Verdict: "fail", Detail: err.Error()}}}) |
| 965 | return |
| 966 | } |
| 967 | var r ProbeReport |
| 968 | if err := json.Unmarshal(b, &r); err != nil { |
| 969 | set(ProbeReport{Lines: []VerdictLine{{Check: "probe-output", Verdict: "fail", Detail: err.Error()}}}) |
| 970 | return |
| 971 | } |
| 972 | for _, l := range resolverLines { |
| 973 | r.Lines = append(r.Lines, l) |
| 974 | if l.Verdict == "fail" { |
| 975 | r.Ready = false |
| 976 | } |
| 977 | } |
| 978 | for _, l := range r.Lines { |
| 979 | if l.Verdict != "ok" { |
| 980 | s.logf("instance-control: probe: %s %s %s", l.Check, l.Verdict, l.Detail) |
| 981 | } |
| 982 | } |
| 983 | set(s.withStartLines(r)) |
| 984 | } |
| 985 | |
| 986 | // mountpoints are the two bind mounts whose underlying directories |
| 987 | // inspectMountpoints looks under (design §8.2, "Recovery for a layout that |
| 988 | // was unmounted anyway"). |
| 989 | var mountpoints = []struct{ name, where string }{ |
| 990 | {"data-instances", "/data/instances"}, |
| 991 | {"run-carlos-instances", "/run/carlos/instances"}, |
| 992 | } |
| 993 | |
| 994 | // InspectMountpoints runs inspectMountpoints once and keeps its lines for |
| 995 | // every later report. Run calls it before the probe loop starts. |
| 996 | func (s *Server) InspectMountpoints(ctx context.Context) { |
| 997 | lines := s.inspectMountpoints(ctx) |
| 998 | for _, l := range lines { |
| 999 | if l.Verdict != "ok" { |
| 1000 | s.logf("instance-control: %s %s %s", l.Check, l.Verdict, l.Detail) |
| 1001 | } |
| 1002 | } |
| 1003 | s.mu.Lock() |
| 1004 | s.startLines = lines |
| 1005 | s.mu.Unlock() |
| 1006 | } |
| 1007 | |
| 1008 | // inspectMountpoints looks UNDER the two bind mounts: in a private mount |
| 1009 | // namespace it unmounts each bind and lists the raw mountpoint directory. |
| 1010 | // Anything there was written while the layout was unmounted (an operator |
| 1011 | // removing units by hand, a boot without the mounts), and the layout |
| 1012 | // pass's destination check would otherwise never see it: the mount hides |
| 1013 | // it, and the old binary would restore a second database over it on the |
| 1014 | // next unmount. The verdict is taken once at start, as root, because |
| 1015 | // unmounting in a private namespace needs CAP_SYS_ADMIN; it clears only |
| 1016 | // when an operator moves or removes the files and restarts the service. |
| 1017 | // An inspection that cannot run (no unshare, no permission) is |
| 1018 | // unverified, and does not block readiness on its own: the probe's own |
| 1019 | // checks stand, and the unverified line is in every report for the |
| 1020 | // operator to see. |
| 1021 | func (s *Server) inspectMountpoints(ctx context.Context) []VerdictLine { |
| 1022 | if s.Run == nil { |
| 1023 | return nil |
| 1024 | } |
| 1025 | var lines []VerdictLine |
| 1026 | for _, m := range mountpoints { |
| 1027 | // The listing runs after the umount so it reads the underlying |
| 1028 | // directory; a umount that fails because nothing is mounted there |
| 1029 | // is fine (the listing is then of the directory itself, which is |
| 1030 | // what an unmounted layout looks like), so its error is dropped. |
| 1031 | out, err := s.Run(ctx, "unshare", "-m", "--propagation", "private", "sh", "-c", |
| 1032 | "umount "+m.where+" 2>/dev/null; ls -A "+m.where) |
| 1033 | if err != nil { |
| 1034 | lines = append(lines, VerdictLine{Check: "mountpoint-" + m.name, Verdict: "unverified", Detail: strings.TrimSpace(out + " " + err.Error())}) |
| 1035 | continue |
| 1036 | } |
| 1037 | if n := len(strings.Fields(out)); n > 0 { |
| 1038 | lines = append(lines, VerdictLine{Check: "mountpoint-" + m.name, Verdict: "fail", |
| 1039 | Detail: fmt.Sprintf("mountpoint held files while unmounted: %d entries under %s; move or remove them, then restart carlos-instance-control", n, m.where)}) |
| 1040 | continue |
| 1041 | } |
| 1042 | lines = append(lines, VerdictLine{Check: "mountpoint-" + m.name, Verdict: "ok"}) |
| 1043 | } |
| 1044 | return lines |
| 1045 | } |
| 1046 | |
| 1047 | // withStartLines prepends the start inspection to a probe report and |
| 1048 | // recomputes Ready: the probe's own verdict, and no start line failing. |
| 1049 | // A copy, so the stored report never aliases the probe's slice. |
| 1050 | func (s *Server) withStartLines(r ProbeReport) ProbeReport { |
| 1051 | s.mu.Lock() |
| 1052 | start := append([]VerdictLine(nil), s.startLines...) |
| 1053 | s.mu.Unlock() |
| 1054 | if len(start) == 0 { |
| 1055 | return r |
| 1056 | } |
| 1057 | out := ProbeReport{Ready: r.Ready, Lines: append(start, r.Lines...)} |
| 1058 | for _, l := range start { |
| 1059 | if l.Verdict == "fail" { |
| 1060 | out.Ready = false |
| 1061 | } |
| 1062 | } |
| 1063 | return out |
| 1064 | } |
| 1065 | |
| 1066 | func (s *Server) claimAwareLegacy(next http.HandlerFunc) http.HandlerFunc { |
| 1067 | return func(w http.ResponseWriter, r *http.Request) { |
| 1068 | if s.ClaimStop != nil { |
| 1069 | w.WriteHeader(http.StatusServiceUnavailable) |
| 1070 | return |
| 1071 | } |
| 1072 | next(w, r) |
| 1073 | } |
| 1074 | } |
| 1075 | |
amadan