| 1 | //go:build linux && rootrunner |
| 2 | |
| 3 | package host |
| 4 | |
| 5 | import ( |
| 6 | "bufio" |
| 7 | "bytes" |
| 8 | "context" |
| 9 | "crypto/ecdh" |
| 10 | "crypto/ed25519" |
| 11 | "crypto/rand" |
| 12 | "crypto/sha256" |
| 13 | "database/sql" |
| 14 | "encoding/base64" |
| 15 | "encoding/hex" |
| 16 | "encoding/json" |
| 17 | "fmt" |
| 18 | "io" |
| 19 | "os" |
| 20 | "os/exec" |
| 21 | "os/user" |
| 22 | "path/filepath" |
| 23 | "strconv" |
| 24 | "strings" |
| 25 | "syscall" |
| 26 | "testing" |
| 27 | "time" |
| 28 | |
| 29 | lsfile "github.com/benbjohnson/litestream/file" |
| 30 | "github.com/superfly/ltx" |
| 31 | |
| 32 | "github.com/carlosframework/platform/internal/dbdoor" |
| 33 | "github.com/carlosframework/platform/internal/dbdoor/dbdoortest" |
| 34 | "github.com/carlosframework/platform/internal/dbreads" |
| 35 | "github.com/carlosframework/platform/internal/hostclaim" |
| 36 | "github.com/carlosframework/platform/internal/provision" |
| 37 | "github.com/carlosframework/platform/internal/release" |
| 38 | ) |
| 39 | |
| 40 | const ( |
| 41 | dbreadRunnerHost = "shop.example.com" |
| 42 | // dbreadReportPrefix marks the payload's one report line. m.Run prints |
| 43 | // PASS after the payload test returns, so the report is never the last |
| 44 | // line (plan review finding 3). |
| 45 | dbreadReportPrefix = "DBREAD-REPORT " |
| 46 | // dbreadServiceFile is where ensureSystemUnits("/etc", ...) installs the |
| 47 | // worker's service (plan review finding 4). |
| 48 | dbreadServiceFile = "/etc/systemd/system/carlos-dbread@.service" |
| 49 | ) |
| 50 | |
| 51 | // dbreadRunner is the production wiring on a throwaway root runner: the real |
| 52 | // units installed by the real ensureSystemUnits into /etc, as the agent's |
| 53 | // adopt pass installs them (agent.go), this test binary at /opt/carlos/carlos |
| 54 | // where the unit's ExecStart finds it, and the agent half run as carlos in a |
| 55 | // child process, because the worker serves only the carlos user's |
| 56 | // connections. /var/tmp is a tmpfs of tmpSize: the runner's small root |
| 57 | // volume would refuse every read's 5 GiB floor, and tmpfs only spends memory |
| 58 | // on what is written. |
| 59 | type dbreadRunner struct { |
| 60 | t *testing.T |
| 61 | dir string // owned by carlos: the store and the replica |
| 62 | uid, gid int |
| 63 | rootPub ed25519.PublicKey |
| 64 | rootPriv ed25519.PrivateKey |
| 65 | } |
| 66 | |
| 67 | func runSystemctl(args ...string) (string, error) { |
| 68 | out, err := exec.Command("systemctl", args...).CombinedOutput() |
| 69 | return string(out), err |
| 70 | } |
| 71 | |
| 72 | func mustSystemctl(t *testing.T, args ...string) { |
| 73 | t.Helper() |
| 74 | if out, err := runSystemctl(args...); err != nil { |
| 75 | t.Fatalf("systemctl %s: %v: %s", strings.Join(args, " "), err, out) |
| 76 | } |
| 77 | } |
| 78 | |
| 79 | // sqliteFixture writes a real SQLite database at path, so an ATTACH of it |
| 80 | // fails only when the file cannot be reached (plan review finding 17). |
| 81 | func sqliteFixture(t *testing.T, path string) { |
| 82 | t.Helper() |
| 83 | db, err := sql.Open("sqlite", path) |
| 84 | if err != nil { |
| 85 | t.Fatal(err) |
| 86 | } |
| 87 | defer db.Close() |
| 88 | if _, err := db.Exec(`CREATE TABLE IF NOT EXISTS secret (v TEXT); INSERT INTO secret VALUES ('not for tenants')`); err != nil { |
| 89 | t.Fatal(err) |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | func newDBReadRunner(t *testing.T, tmpSize string) *dbreadRunner { |
| 94 | t.Helper() |
| 95 | if os.Geteuid() != 0 { |
| 96 | t.Fatal("run on the root runner, as root") |
| 97 | } |
| 98 | r := &dbreadRunner{t: t} |
| 99 | if out, err := exec.Command("mount", "-t", "tmpfs", "-o", "size="+tmpSize+",mode=1777", "tmpfs", "/var/tmp").CombinedOutput(); err != nil { |
| 100 | t.Fatalf("mount /var/tmp: %v: %s", err, out) |
| 101 | } |
| 102 | t.Cleanup(func() { exec.Command("umount", "/var/tmp").Run() }) |
| 103 | u, err := user.Lookup("carlos") |
| 104 | if err != nil { |
| 105 | if out, err := exec.Command("useradd", "--system", "--user-group", "--no-create-home", "--shell", "/usr/sbin/nologin", "carlos").CombinedOutput(); err != nil { |
| 106 | t.Fatalf("useradd: %v: %s", err, out) |
| 107 | } |
| 108 | t.Cleanup(func() { exec.Command("userdel", "carlos").Run() }) |
| 109 | if u, err = user.Lookup("carlos"); err != nil { |
| 110 | t.Fatal(err) |
| 111 | } |
| 112 | } |
| 113 | r.uid, _ = strconv.Atoi(u.Uid) |
| 114 | r.gid, _ = strconv.Atoi(u.Gid) |
| 115 | if _, err := os.Stat("/opt/carlos/carlos"); err == nil { |
| 116 | t.Fatal("/opt/carlos/carlos exists: this test runs only on a fresh runner") |
| 117 | } |
| 118 | self, err := os.ReadFile(os.Args[0]) |
| 119 | if err != nil { |
| 120 | t.Fatal(err) |
| 121 | } |
| 122 | for _, d := range []string{"/opt/carlos", "/etc/carlos", "/data/instances/other.example.com", "/run/carlos", dbreadTestMarkers} { |
| 123 | if err := os.MkdirAll(d, 0o755); err != nil { |
| 124 | t.Fatal(err) |
| 125 | } |
| 126 | } |
| 127 | // Things the witness must not reach, present on the host: real SQLite |
| 128 | // databases where the witness ATTACHes, a host.env it opens. |
| 129 | sqliteFixture(t, "/data/registry.db") |
| 130 | sqliteFixture(t, "/etc/carlos/witness.db") |
| 131 | sqliteFixture(t, "/data/instances/other.example.com/other.db") |
| 132 | // host.env arms db reads (dbReadsSettings), so the adopt pass's |
| 133 | // ensureSystemUnits arms the socket: it stays dark otherwise. |
| 134 | r.rootPub, r.rootPriv, _ = ed25519.GenerateKey(rand.Reader) |
| 135 | if err := os.WriteFile("/etc/carlos/host.env", []byte(dbreadRunnerHostEnv(r.rootPub)), 0o640); err != nil { |
| 136 | t.Fatal(err) |
| 137 | } |
| 138 | if err := os.WriteFile("/opt/carlos/carlos", self, 0o755); err != nil { |
| 139 | t.Fatal(err) |
| 140 | } |
| 141 | t.Cleanup(func() { |
| 142 | for _, p := range []string{"/opt/carlos/carlos", "/data/registry.db", "/etc/carlos/witness.db", "/etc/carlos/host.env"} { |
| 143 | os.Remove(p) |
| 144 | } |
| 145 | os.RemoveAll("/data/instances/other.example.com") |
| 146 | os.RemoveAll(dbreadTestMarkers) |
| 147 | }) |
| 148 | ensureSystemUnits("/etc", filepath.Join(t.TempDir(), "lock"), runSystemctl, systemctlProbe, t.Logf) |
| 149 | t.Cleanup(func() { |
| 150 | runSystemctl("stop", "carlos-dbread.socket", "carlos-instance-control.socket", "carlos-system-update.path") |
| 151 | runSystemctl("disable", "carlos-dbread.socket", "carlos-instance-control.socket", "carlos-system-update.path", "carlos-system-recover.service") |
| 152 | for _, u := range systemUnitFiles { |
| 153 | os.Remove(filepath.Join("/etc/systemd/system", u)) |
| 154 | } |
| 155 | runSystemctl("daemon-reload") |
| 156 | }) |
| 157 | if _, err := os.Stat(dbreadServiceFile); err != nil { |
| 158 | t.Fatalf("ensureSystemUnits installed no %s: %v", dbreadServiceFile, err) |
| 159 | } |
| 160 | if out, _ := systemctlProbe("is-active", dbReadSocketUnit); strings.TrimSpace(out) != "active" { |
| 161 | t.Fatalf("carlos-dbread.socket is %q after ensureSystemUnits", out) |
| 162 | } |
| 163 | r.dir, err = os.MkdirTemp("/var/lib", "carlos-dbread-runner-") |
| 164 | if err != nil { |
| 165 | t.Fatal(err) |
| 166 | } |
| 167 | t.Cleanup(func() { os.RemoveAll(r.dir) }) |
| 168 | replica := filepath.Join(r.dir, "replica", "instances", dbreadRunnerHost) |
| 169 | dbdoortest.Replicate(t, replica, 3) |
| 170 | yes := true |
| 171 | rec := provision.Record{Host: dbreadRunnerHost, App: "shop", Account: "bab", Region: "uk", Hibernate: &yes, |
| 172 | CreatedAt: time.Date(2026, 9, 1, 0, 0, 0, 0, time.UTC)} |
| 173 | body, _ := json.Marshal(rec) |
| 174 | if _, err := r.store().Put(context.Background(), provision.RecordKey("bab", "shop", dbreadRunnerHost), body, ""); err != nil { |
| 175 | t.Fatal(err) |
| 176 | } |
| 177 | r.chown() |
| 178 | return r |
| 179 | } |
| 180 | |
| 181 | // dbreadRunnerHostEnv is a host.env that arms db reads, plus extra lines. |
| 182 | func dbreadRunnerHostEnv(pub ed25519.PublicKey, extra ...string) string { |
| 183 | lines := append([]string{"CARLOS_REGION=uk", "CARLOS_HIBERNATE_BUCKET=data-uk", "CARLOS_RELEASE_PUBKEY=" + base64.StdEncoding.EncodeToString(pub)}, extra...) |
| 184 | return strings.Join(lines, "\n") + "\n" |
| 185 | } |
| 186 | |
| 187 | func (r *dbreadRunner) store() release.Store { |
| 188 | return release.NewDirStore(filepath.Join(r.dir, "store")) |
| 189 | } |
| 190 | |
| 191 | func (r *dbreadRunner) chown() { |
| 192 | filepath.Walk(r.dir, func(p string, _ os.FileInfo, _ error) error { return os.Lchown(p, r.uid, r.gid) }) |
| 193 | } |
| 194 | |
| 195 | func (r *dbreadRunner) marker(name string, on bool) { |
| 196 | r.t.Helper() |
| 197 | p := filepath.Join(dbreadTestMarkers, name) |
| 198 | var err error |
| 199 | if on { |
| 200 | err = os.WriteFile(p, nil, 0o644) |
| 201 | } else if err = os.Remove(p); os.IsNotExist(err) { |
| 202 | err = nil |
| 203 | } |
| 204 | if err != nil { |
| 205 | r.t.Fatal(err) |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | // submit is the console's two creates for a read of op; it returns the |
| 210 | // read's id and reply key. |
| 211 | func (r *dbreadRunner) submit(op, sqlText string) (string, *ecdh.PrivateKey) { |
| 212 | r.t.Helper() |
| 213 | reply, _ := dbreads.NewReplyKey() |
| 214 | id, _ := dbreads.MintID() |
| 215 | inc, _ := hostclaim.Incarnation("bab", "shop", dbreadRunnerHost, time.Date(2026, 9, 1, 0, 0, 0, 0, time.UTC)) |
| 216 | req := dbreads.Request{ID: id, Account: "bab", App: "shop", Host: dbreadRunnerHost, Incarnation: inc, Region: "uk", |
| 217 | Op: op, SQL: sqlText, ReplyKey: dbreads.EncodeReplyKey(reply.PublicKey()), |
| 218 | By: "member@example.com", IssuedAt: time.Now().UTC().Format(time.RFC3339)} |
| 219 | wire, err := dbreads.Sign(release.RootSigner(r.rootPriv), req) |
| 220 | if err != nil { |
| 221 | r.t.Fatal(err) |
| 222 | } |
| 223 | s := r.store() |
| 224 | if _, err := s.Put(context.Background(), dbreads.RequestKey(id), wire, ""); err != nil { |
| 225 | r.t.Fatal(err) |
| 226 | } |
| 227 | issued, _ := req.Issued() |
| 228 | if _, err := s.Put(context.Background(), dbreads.QueueKey("uk", issued, id), nil, ""); err != nil { |
| 229 | r.t.Fatal(err) |
| 230 | } |
| 231 | r.chown() |
| 232 | return id, reply |
| 233 | } |
| 234 | |
| 235 | // payload runs the agent half as carlos. With wait it returns the payload's |
| 236 | // report, the line marked dbreadReportPrefix; without, the running command. |
| 237 | // id, when set, is the read the scenario waits on or reports about. |
| 238 | func (r *dbreadRunner) payload(scenario string, wait bool, id string) (map[string]any, *exec.Cmd) { |
| 239 | r.t.Helper() |
| 240 | // The copy at /opt/carlos/carlos, not os.Args[0]: the runner keeps its |
| 241 | // binary under /root, which carlos cannot enter. |
| 242 | cmd := exec.Command("/opt/carlos/carlos") |
| 243 | cmd.Env = append(os.Environ(), "DBREAD_RUNNER_PAYLOAD="+scenario, "DBREAD_RUNNER_DIR="+r.dir, |
| 244 | "DBREAD_RUNNER_ID="+id, "DBREAD_RUNNER_PUB="+base64.StdEncoding.EncodeToString(r.rootPub)) |
| 245 | cmd.SysProcAttr = &syscall.SysProcAttr{Credential: &syscall.Credential{Uid: uint32(r.uid), Gid: uint32(r.gid)}} |
| 246 | var out bytes.Buffer |
| 247 | cmd.Stdout, cmd.Stderr = &out, os.Stderr |
| 248 | if !wait { |
| 249 | if err := cmd.Start(); err != nil { |
| 250 | r.t.Fatal(err) |
| 251 | } |
| 252 | return nil, cmd |
| 253 | } |
| 254 | if err := cmd.Run(); err != nil { |
| 255 | r.t.Fatalf("payload %s: %v: %s", scenario, err, out.String()) |
| 256 | } |
| 257 | sc := bufio.NewScanner(&out) |
| 258 | for sc.Scan() { |
| 259 | if line, ok := strings.CutPrefix(sc.Text(), dbreadReportPrefix); ok { |
| 260 | var rep map[string]any |
| 261 | if err := json.Unmarshal([]byte(line), &rep); err != nil { |
| 262 | r.t.Fatalf("payload %s: unreadable report %q", scenario, line) |
| 263 | } |
| 264 | return rep, cmd |
| 265 | } |
| 266 | } |
| 267 | r.t.Fatalf("payload %s gave no report: %s", scenario, out.String()) |
| 268 | return nil, cmd |
| 269 | } |
| 270 | |
| 271 | func (r *dbreadRunner) answer(id string, reply *ecdh.PrivateKey, stream string) ([]byte, dbreads.Result) { |
| 272 | r.t.Helper() |
| 273 | s := r.store() |
| 274 | raw, _, err := s.Get(context.Background(), dbreads.ResultKey(id)) |
| 275 | if err != nil { |
| 276 | r.t.Fatalf("no result: %v", err) |
| 277 | } |
| 278 | var res dbreads.Result |
| 279 | json.Unmarshal(raw, &res) |
| 280 | var out bytes.Buffer |
| 281 | for seq := int64(1); seq <= res.LastSeq; seq++ { |
| 282 | sealed, _, _ := s.Get(context.Background(), dbreads.ChunkKey(id, seq)) |
| 283 | plain, err := dbreads.OpenChunk(reply, id, seq, stream, seq == res.LastSeq, sealed) |
| 284 | if err != nil { |
| 285 | r.t.Fatal(err) |
| 286 | } |
| 287 | out.Write(plain) |
| 288 | } |
| 289 | return out.Bytes(), res |
| 290 | } |
| 291 | |
| 292 | // privateTmpGone waits up to limit for every worker's private /var/tmp to |
| 293 | // be removed. |
| 294 | func privateTmpGone(t *testing.T, limit time.Duration) { |
| 295 | t.Helper() |
| 296 | deadline := time.Now().Add(limit) |
| 297 | for { |
| 298 | left, _ := filepath.Glob("/var/tmp/systemd-private-*-carlos-dbread@*") |
| 299 | if len(left) == 0 { |
| 300 | return |
| 301 | } |
| 302 | if time.Now().After(deadline) { |
| 303 | t.Fatalf("a worker's private tmp outlived it: %v", left) |
| 304 | } |
| 305 | time.Sleep(200 * time.Millisecond) |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | func unitsActive(t *testing.T) int { |
| 310 | n, err := dbReadUnits(context.Background()) |
| 311 | if err != nil { |
| 312 | t.Fatal(err) |
| 313 | } |
| 314 | return n |
| 315 | } |
| 316 | |
| 317 | // dbreadWaitFor polls ok until it holds or limit passes. (waitFor is |
| 318 | // housekeeper_test.go's, with another signature: plan review finding 2.) |
| 319 | func dbreadWaitFor(t *testing.T, limit time.Duration, ok func() bool) { |
| 320 | t.Helper() |
| 321 | deadline := time.Now().Add(limit) |
| 322 | for !ok() { |
| 323 | if time.Now().After(deadline) { |
| 324 | t.Fatal("timed out") |
| 325 | } |
| 326 | time.Sleep(200 * time.Millisecond) |
| 327 | } |
| 328 | } |
| 329 | |
| 330 | // T2: a forwarded query through the real socket and the real service, |
| 331 | // installed by the real ensureSystemUnits; the socket is the worker's stdin |
| 332 | // and stdout (codex 2). The probe that makes the box capable runs the |
| 333 | // default SelfView inside the unit (plan review finding 1). |
| 334 | func TestRootDBReadServesThroughTheInstalledUnits(t *testing.T) { |
| 335 | r := newDBReadRunner(t, "40g") |
| 336 | id, reply := r.submit(dbreads.OpQuery, "SELECT count(*) AS n FROM t") |
| 337 | rep, _ := r.payload("serve", true, id) |
| 338 | if rep["capable"] != true { |
| 339 | t.Fatalf("the installed worker failed its probe: %v", rep) |
| 340 | } |
| 341 | rows, res := r.answer(id, reply, dbreads.StreamRows) |
| 342 | if res.State != dbreads.StateDone || string(rows) != "n\n30\n" { |
| 343 | t.Fatalf("result %+v rows %q", res, rows) |
| 344 | } |
| 345 | privateTmpGone(t, 10*time.Second) |
| 346 | } |
| 347 | |
| 348 | // A forwarded pull through the real socket and the real service (plan 3 |
| 349 | // Task 5): the installed worker verifies and holds the planned files and |
| 350 | // sends them back after check B; the answer, opened as the ltx stream, is |
| 351 | // the planned files byte for byte in plan order, and adds up to the sealed |
| 352 | // summary. The worker's private tmp is gone afterwards. |
| 353 | func TestRootDBReadPullThroughTheInstalledUnits(t *testing.T) { |
| 354 | r := newDBReadRunner(t, "40g") |
| 355 | id, reply := r.submit(dbreads.OpPull, "") |
| 356 | rep, _ := r.payload("serve", true, id) |
| 357 | if rep["capable"] != true { |
| 358 | t.Fatalf("the installed worker failed its probe: %v", rep) |
| 359 | } |
| 360 | got, res := r.answer(id, reply, dbreads.StreamLTX) |
| 361 | if res.State != dbreads.StateDone || res.Code != "" { |
| 362 | t.Fatalf("result %+v", res) |
| 363 | } |
| 364 | sum, err := dbreads.OpenSummary(reply, id, res.Sealed) |
| 365 | if err != nil { |
| 366 | t.Fatal(err) |
| 367 | } |
| 368 | h := sha256.Sum256(got) |
| 369 | if hex.EncodeToString(h[:]) != sum.SHA256 || int64(len(got)) != sum.Bytes { |
| 370 | t.Fatalf("%d bytes do not add up to the summary %+v", len(got), sum) |
| 371 | } |
| 372 | rd := lsfile.NewReplicaClient(filepath.Join(r.dir, "replica", "instances", dbreadRunnerHost)) |
| 373 | plan, err := dbdoor.PlanFrom(context.Background(), rd, time.Time{}) |
| 374 | if err != nil { |
| 375 | t.Fatal(err) |
| 376 | } |
| 377 | var want bytes.Buffer |
| 378 | for _, f := range plan.Planned { |
| 379 | rc, err := rd.OpenLTXFile(context.Background(), f.Level, ltx.TXID(f.MinTXID), ltx.TXID(f.MaxTXID), 0, 0) |
| 380 | if err != nil { |
| 381 | t.Fatal(err) |
| 382 | } |
| 383 | io.Copy(&want, rc) |
| 384 | rc.Close() |
| 385 | } |
| 386 | if !bytes.Equal(got, want.Bytes()) || sum.TXID != plan.TXID || len(sum.Files) != len(plan.Planned) { |
| 387 | t.Fatalf("pulled %d bytes at %d in %d files; planned %d bytes at %d in %d files", |
| 388 | len(got), sum.TXID, len(sum.Files), want.Len(), plan.TXID, len(plan.Planned)) |
| 389 | } |
| 390 | privateTmpGone(t, 10*time.Second) |
| 391 | } |
| 392 | |
| 393 | // T3: from inside the worker, nothing outside its scratch is reachable, |
| 394 | // including any network interface but loopback; the private tmp is gone |
| 395 | // after a normal end, a cancel and kill -9 of the agent. Mutation: remove |
| 396 | // PrivateNetwork=yes from the installed unit; "see an interface other than |
| 397 | // loopback" succeeds and this test is red (plan review finding 17). |
| 398 | func TestRootDBReadWorkerSandbox(t *testing.T) { |
| 399 | r := newDBReadRunner(t, "40g") |
| 400 | r.marker("witness", true) |
| 401 | rep, _ := r.payload("witness", true, "") |
| 402 | r.marker("witness", false) |
| 403 | report, _ := rep["report"].(map[string]any) |
| 404 | if len(report) != 14 { |
| 405 | t.Fatalf("an incomplete witness report: %v", rep) |
| 406 | } |
| 407 | for name, v := range report { |
| 408 | failed := v.(string) != "" |
| 409 | if name == "write /var/tmp" { |
| 410 | if failed { |
| 411 | t.Errorf("the positive control failed: %s", v) |
| 412 | } |
| 413 | continue |
| 414 | } |
| 415 | if !failed { |
| 416 | t.Errorf("%s succeeded inside the worker", name) |
| 417 | } |
| 418 | } |
| 419 | // A cancel, then kill -9 of the agent, mid-read. |
| 420 | r.marker("hang-restore", true) |
| 421 | defer r.marker("hang-restore", false) |
| 422 | id, _ := r.submit(dbreads.OpQuery, "SELECT 1") |
| 423 | _, cmd := r.payload("serve-and-hang", false, "") |
| 424 | dbreadWaitFor(t, 30*time.Second, func() bool { return unitsActive(t) >= 1 }) |
| 425 | if _, err := r.store().Put(context.Background(), dbreads.CancelKey(id), nil, ""); err != nil { |
| 426 | t.Fatal(err) |
| 427 | } |
| 428 | dbreadWaitFor(t, 20*time.Second, func() bool { return unitsActive(t) == 0 }) |
| 429 | privateTmpGone(t, 10*time.Second) |
| 430 | cmd.Process.Kill() |
| 431 | cmd.Wait() |
| 432 | r.submit(dbreads.OpQuery, "SELECT 1") |
| 433 | _, cmd = r.payload("serve-and-hang", false, "") |
| 434 | dbreadWaitFor(t, 30*time.Second, func() bool { return unitsActive(t) >= 1 }) |
| 435 | cmd.Process.Signal(syscall.SIGKILL) |
| 436 | cmd.Wait() |
| 437 | dbreadWaitFor(t, 20*time.Second, func() bool { return unitsActive(t) == 0 }) |
| 438 | privateTmpGone(t, 10*time.Second) |
| 439 | } |
| 440 | |
| 441 | // T12 with real units. The disk floor: /var/tmp on a 12 GiB tmpfs fits one |
| 442 | // read's floor reservation (5 GiB and 10%) and not two. The first read is |
| 443 | // running (its unit observed) before the second is submitted, so which read |
| 444 | // waits never depends on queue order (plan review finding 16); the agent |
| 445 | // passes every second and still never picks the second while the first |
| 446 | // holds its floor. A fixture database small enough for the runner cannot |
| 447 | // make the plan's growth bind, so the no-room result itself is pinned by |
| 448 | // TestReadTooLargeAndNoRoom. Then the agent is killed mid-read: the |
| 449 | // restarted agent takes no read and no probe while the old unit is active, |
| 450 | // the old worker ends within 5 s of losing the connection though its |
| 451 | // restore ignores the context, and the restarted agent admits the held-back |
| 452 | // read once the unit and its scratch are gone. Mutations: drop the orphan |
| 453 | // rule in reconcileLocked (the restarted agent picks beside the old unit), |
| 454 | // or drop the worker's 5 s exit (the old unit outlives the 15 s bound). |
| 455 | func TestRootDBReadAdmissionWithRealUnits(t *testing.T) { |
| 456 | r := newDBReadRunner(t, "12g") |
| 457 | r.marker("hang-restore", true) |
| 458 | r.submit(dbreads.OpQuery, "SELECT 1") |
| 459 | _, hung := r.payload("serve-and-hang", false, "") |
| 460 | dbreadWaitFor(t, 30*time.Second, func() bool { return unitsActive(t) == 1 }) |
| 461 | second, reply := r.submit(dbreads.OpQuery, "SELECT count(*) AS n FROM t") |
| 462 | time.Sleep(5 * time.Second) // five passes |
| 463 | if _, _, err := r.store().Get(context.Background(), dbreads.PickedKey(second)); err == nil { |
| 464 | t.Fatal("a second read was picked with room for one") |
| 465 | } |
| 466 | hung.Process.Signal(syscall.SIGKILL) |
| 467 | hung.Wait() |
| 468 | killedAt := time.Now() |
| 469 | rep, _ := r.payload("restarted", true, second) |
| 470 | if rep["units_at_start"].(float64) < 1 { |
| 471 | t.Fatalf("the old unit was gone before the restarted agent looked: %v", rep) |
| 472 | } |
| 473 | if rep["capable"] == true || rep["picked"] == true { |
| 474 | t.Fatalf("a restarted agent admitted beside the old unit: %v", rep) |
| 475 | } |
| 476 | dbreadWaitFor(t, 20*time.Second, func() bool { return unitsActive(t) == 0 }) |
| 477 | if gone := time.Since(killedAt); gone > 15*time.Second { |
| 478 | t.Fatalf("the old worker took %s to stop; its 5 s exit did not hold", gone) |
| 479 | } |
| 480 | privateTmpGone(t, 10*time.Second) |
| 481 | r.marker("hang-restore", false) |
| 482 | rep, _ = r.payload("serve", true, second) |
| 483 | if rep["capable"] != true { |
| 484 | t.Fatalf("the agent did not admit once the unit was gone: %v", rep) |
| 485 | } |
| 486 | if rows, res := r.answer(second, reply, dbreads.StreamRows); res.State != dbreads.StateDone || string(rows) != "n\n30\n" { |
| 487 | t.Fatalf("the held-back read: %+v %q", res, rows) |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | // T13: a socket whose service is missing, masked, failing to start or |
| 492 | // speaking another protocol version withdraws db-read-1 within one probe and |
| 493 | // edges pass; a read whose worker fails to start withdraws it at once |
| 494 | // through the real loop (plan review finding 15); a box running two reads |
| 495 | // keeps it (codex 9). |
| 496 | func TestRootDBReadReadiness(t *testing.T) { |
| 497 | r := newDBReadRunner(t, "40g") |
| 498 | capable := func() bool { |
| 499 | rep, _ := r.payload("probe", true, "") |
| 500 | return rep["capable"] == true && rep["published"] == true |
| 501 | } |
| 502 | socketState := func(when string) string { |
| 503 | out, _ := runSystemctl("show", "-p", "ActiveState,SubState,Result", "carlos-dbread.socket") |
| 504 | t.Logf("carlos-dbread.socket %s: %s", when, strings.Join(strings.Fields(out), " ")) |
| 505 | return out |
| 506 | } |
| 507 | // repair is the box's own way back once the service is restored: a |
| 508 | // connection to a socket whose service cannot load fails the socket unit |
| 509 | // (systemd stops listening), and the next adopt pass's ensureSystemUnits |
| 510 | // sees it inactive and arms it again. Nothing else restarts it. |
| 511 | repair := func(when string) { |
| 512 | t.Helper() |
| 513 | socketState(when + ", before the adopt pass") |
| 514 | ensureSystemUnits("/etc", filepath.Join(t.TempDir(), "lock"), runSystemctl, systemctlProbe, t.Logf) |
| 515 | if out, _ := systemctlProbe("is-active", dbReadSocketUnit); strings.TrimSpace(out) != "active" { |
| 516 | t.Fatalf("carlos-dbread.socket is %q after the adopt pass", out) |
| 517 | } |
| 518 | } |
| 519 | if !capable() { |
| 520 | t.Fatal("baseline: not capable") |
| 521 | } |
| 522 | body, err := os.ReadFile(dbreadServiceFile) |
| 523 | if err != nil { |
| 524 | t.Fatal(err) |
| 525 | } |
| 526 | // Cleanup removes first: during the mask the path is a symlink to |
| 527 | // /dev/null, and writing through it would write nothing. |
| 528 | t.Cleanup(func() { |
| 529 | os.Remove(dbreadServiceFile) |
| 530 | os.WriteFile(dbreadServiceFile, body, 0o644) |
| 531 | runSystemctl("daemon-reload") |
| 532 | }) |
| 533 | if err := os.Remove(dbreadServiceFile); err != nil { |
| 534 | t.Fatal(err) |
| 535 | } |
| 536 | mustSystemctl(t, "daemon-reload") |
| 537 | if capable() { |
| 538 | t.Error("capable with the service missing") |
| 539 | } |
| 540 | if err := os.WriteFile(dbreadServiceFile, body, 0o644); err != nil { |
| 541 | t.Fatal(err) |
| 542 | } |
| 543 | mustSystemctl(t, "daemon-reload") |
| 544 | repair("with the service missing") |
| 545 | if !capable() { |
| 546 | t.Fatal("not capable once the service file is back") |
| 547 | } |
| 548 | // Masked where the unit is installed (plan review finding 27): a mask is |
| 549 | // a /dev/null symlink, and only one in /etc/systemd/system itself |
| 550 | // outranks the unit file ensureSystemUnits writes there. `systemctl |
| 551 | // mask` refuses while the file exists, and `--runtime` would land in |
| 552 | // /run/systemd/system, which /etc outranks; so swap the file for the |
| 553 | // symlink by hand, and prove the mask took before asking. |
| 554 | if err := os.Remove(dbreadServiceFile); err != nil { |
| 555 | t.Fatal(err) |
| 556 | } |
| 557 | if err := os.Symlink("/dev/null", dbreadServiceFile); err != nil { |
| 558 | t.Fatal(err) |
| 559 | } |
| 560 | mustSystemctl(t, "daemon-reload") |
| 561 | if out, _ := runSystemctl("show", "-p", "LoadState", "--value", "carlos-dbread@probe.service"); strings.TrimSpace(out) != "masked" { |
| 562 | t.Fatalf("the mask did not take: LoadState %q", out) |
| 563 | } |
| 564 | if capable() { |
| 565 | t.Error("capable with the service masked") |
| 566 | } |
| 567 | if err := os.Remove(dbreadServiceFile); err != nil { |
| 568 | t.Fatal(err) |
| 569 | } |
| 570 | if err := os.WriteFile(dbreadServiceFile, body, 0o644); err != nil { |
| 571 | t.Fatal(err) |
| 572 | } |
| 573 | mustSystemctl(t, "daemon-reload") |
| 574 | repair("with the service masked") |
| 575 | for _, m := range []string{"exit-1", "version-2"} { |
| 576 | r.marker(m, true) |
| 577 | if capable() { |
| 578 | t.Errorf("capable with %s", m) |
| 579 | } |
| 580 | socketState("with " + m) |
| 581 | r.marker(m, false) |
| 582 | } |
| 583 | if !capable() { |
| 584 | t.Fatal("not capable again once restored") |
| 585 | } |
| 586 | |
| 587 | // A read whose worker fails to start, through the real loop. |
| 588 | _, loop := r.payload("fail-start", false, "") |
| 589 | dbreadWaitFor(t, 30*time.Second, func() bool { _, err := os.Stat(filepath.Join(r.dir, "probed")); return err == nil }) |
| 590 | r.marker("exit-1", true) |
| 591 | r.submit(dbreads.OpQuery, "SELECT 1") |
| 592 | if err := os.WriteFile(filepath.Join(r.dir, "go"), nil, 0o644); err != nil { |
| 593 | t.Fatal(err) |
| 594 | } |
| 595 | if err := loop.Wait(); err != nil { |
| 596 | t.Fatalf("fail-start payload: %v", err) |
| 597 | } |
| 598 | if _, err := os.Stat(filepath.Join(r.dir, "withdrawn")); err != nil { |
| 599 | t.Fatal("the loop kept db-read-1 after a worker failed to start") |
| 600 | } |
| 601 | r.marker("exit-1", false) |
| 602 | |
| 603 | r.marker("hang-restore", true) |
| 604 | defer r.marker("hang-restore", false) |
| 605 | r.submit(dbreads.OpQuery, "SELECT 1") |
| 606 | r.submit(dbreads.OpQuery, "SELECT 1") |
| 607 | rep, _ := r.payload("busy-probe", true, "") |
| 608 | if rep["capable"] != true { |
| 609 | t.Fatalf("a box running two reads lost the capability: %v", rep) |
| 610 | } |
| 611 | } |
| 612 | |
| 613 | // TestRootDBReadPayload is the agent half, run as carlos by payload(); |
| 614 | // never run it by name. |
| 615 | func TestRootDBReadPayload(t *testing.T) { |
| 616 | scenario := os.Getenv("DBREAD_RUNNER_PAYLOAD") |
| 617 | if scenario == "" { |
| 618 | t.Skip("run by the db-read root-runner tests as their carlos child") |
| 619 | } |
| 620 | dir := os.Getenv("DBREAD_RUNNER_DIR") |
| 621 | pubRaw, _ := base64.StdEncoding.DecodeString(os.Getenv("DBREAD_RUNNER_PUB")) |
| 622 | store := release.NewDirStore(filepath.Join(dir, "store")) |
| 623 | a := dbreads.NewAgent(dbreads.AgentConfig{ |
| 624 | Store: store, Region: "uk", Box: "runner", Boot: newBootID(), Pub: ed25519.PublicKey(pubRaw), |
| 625 | HibernatePrefix: "instances", |
| 626 | OpenReplica: func(_ context.Context, prefix string) (dbdoor.ReplicaReader, error) { |
| 627 | return lsfile.NewReplicaClient(filepath.Join(dir, "replica", prefix)), nil |
| 628 | }, |
| 629 | DialWorker: dialDBReadWorker, Units: dbReadUnits, Free: varTmpFree, Logf: t.Logf, |
| 630 | }) |
| 631 | ctx := context.Background() |
| 632 | report := map[string]any{} |
| 633 | say := func() { |
| 634 | b, _ := json.Marshal(report) |
| 635 | fmt.Println(dbreadReportPrefix + string(b)) |
| 636 | } |
| 637 | switch scenario { |
| 638 | case "probe": |
| 639 | a.Probe(ctx) |
| 640 | report["capable"] = a.Capable() |
| 641 | report["published"] = len(withDBReadCapability(nil, a.Capable())) == 1 |
| 642 | case "serve": |
| 643 | a.Probe(ctx) |
| 644 | report["capable"] = a.Capable() |
| 645 | id := os.Getenv("DBREAD_RUNNER_ID") |
| 646 | deadline := time.Now().Add(2 * time.Minute) |
| 647 | for time.Now().Before(deadline) { |
| 648 | a.Pass(ctx) |
| 649 | a.Wait() |
| 650 | if _, _, err := store.Get(ctx, dbreads.ResultKey(id)); err == nil { |
| 651 | break |
| 652 | } |
| 653 | time.Sleep(time.Second) |
| 654 | } |
| 655 | case "serve-and-hang": |
| 656 | a.Probe(ctx) |
| 657 | for { // the parent kills this process |
| 658 | a.Pass(ctx) |
| 659 | time.Sleep(time.Second) |
| 660 | } |
| 661 | case "restarted": |
| 662 | units, _ := dbReadUnits(ctx) |
| 663 | report["units_at_start"] = units |
| 664 | a.Probe(ctx) |
| 665 | a.Pass(ctx) |
| 666 | report["capable"] = a.Capable() |
| 667 | _, _, err := store.Get(ctx, dbreads.PickedKey(os.Getenv("DBREAD_RUNNER_ID"))) |
| 668 | report["picked"] = err == nil |
| 669 | case "fail-start": |
| 670 | // The production loop: probe at start, a pass on a nudge, a probe |
| 671 | // as soon as a read's worker fails to start. |
| 672 | nudge := make(chan struct{}, 1) |
| 673 | rctx, cancel := context.WithCancel(ctx) |
| 674 | defer cancel() |
| 675 | go a.Run(rctx, nudge) |
| 676 | for !a.Capable() { |
| 677 | time.Sleep(100 * time.Millisecond) |
| 678 | } |
| 679 | os.WriteFile(filepath.Join(dir, "probed"), nil, 0o644) |
| 680 | for { |
| 681 | if _, err := os.Stat(filepath.Join(dir, "go")); err == nil { |
| 682 | break |
| 683 | } |
| 684 | time.Sleep(100 * time.Millisecond) |
| 685 | } |
| 686 | nudge <- struct{}{} |
| 687 | deadline := time.Now().Add(30 * time.Second) |
| 688 | for a.Capable() && time.Now().Before(deadline) { |
| 689 | time.Sleep(100 * time.Millisecond) |
| 690 | } |
| 691 | if !a.Capable() { |
| 692 | os.WriteFile(filepath.Join(dir, "withdrawn"), nil, 0o644) |
| 693 | } |
| 694 | case "busy-probe": |
| 695 | a.Probe(ctx) |
| 696 | a.Pass(ctx) |
| 697 | deadline := time.Now().Add(30 * time.Second) |
| 698 | for n, _ := dbReadUnits(ctx); n < 2 && time.Now().Before(deadline); n, _ = dbReadUnits(ctx) { |
| 699 | time.Sleep(200 * time.Millisecond) |
| 700 | } |
| 701 | a.Probe(ctx) |
| 702 | report["capable"] = a.Capable() |
| 703 | case "witness": |
| 704 | c, err := dialDBReadWorker(ctx) |
| 705 | if err != nil { |
| 706 | t.Fatal(err) |
| 707 | } |
| 708 | var w map[string]string |
| 709 | json.NewDecoder(io.LimitReader(c, 1<<20)).Decode(&w) |
| 710 | c.Close() |
| 711 | report["report"] = w |
| 712 | } |
| 713 | say() |
| 714 | } |
| 715 | |
| 716 | // dbreadTenantUnit is the real tenant template, carlos-instance@.service, |
| 717 | // installed under its own name with only what a runner cannot provide taken |
| 718 | // out: the edge and mount dependencies, the per-instance slice, the stop |
| 719 | // hook and the log socket. Its sandbox (User=carlos, the hidden /run/carlos, |
| 720 | // /data, /opt/carlos and /etc/carlos, the address families) is the |
| 721 | // production one, line for line. ExecStart is this binary as a tenant that |
| 722 | // tries to reach the db-read worker. |
| 723 | const dbreadTenantUnit = "carlos-dbread-tenant@.service" |
| 724 | |
| 725 | func installDBReadTenantUnit(t *testing.T, instance string) { |
| 726 | t.Helper() |
| 727 | raw, err := unitsFS.ReadFile("units/carlos-instance@.service") |
| 728 | if err != nil { |
| 729 | t.Fatal(err) |
| 730 | } |
| 731 | var out []string |
| 732 | for _, line := range strings.Split(string(raw), "\n") { |
| 733 | switch { |
| 734 | case strings.HasPrefix(line, "BindsTo="), strings.HasPrefix(line, "Requires="), strings.HasPrefix(line, "After="), |
| 735 | strings.HasPrefix(line, "ExecStopPost="), strings.HasPrefix(line, "Slice="): |
| 736 | continue |
| 737 | case strings.HasPrefix(line, "ExecStart="): |
| 738 | line = "ExecStart=/opt/carlos/carlos dbread-tenant-witness" |
| 739 | case line == "Type=simple": |
| 740 | line = "Type=oneshot" |
| 741 | case strings.HasPrefix(line, "StandardOutput=file:"): |
| 742 | line = "StandardOutput=journal" |
| 743 | } |
| 744 | out = append(out, line) |
| 745 | } |
| 746 | unit := strings.Join(out, "\n") |
| 747 | if !strings.Contains(unit, "TemporaryFileSystem=/run/carlos:ro") || !strings.Contains(unit, "User=carlos") { |
| 748 | t.Fatal("the tenant template no longer hides /run/carlos as carlos; the witness would prove nothing") |
| 749 | } |
| 750 | path := filepath.Join("/etc/systemd/system", dbreadTenantUnit) |
| 751 | if err := os.WriteFile(path, []byte(unit), 0o644); err != nil { |
| 752 | t.Fatal(err) |
| 753 | } |
| 754 | t.Cleanup(func() { os.Remove(path); runSystemctl("daemon-reload") }) |
| 755 | mustSystemctl(t, "daemon-reload") |
| 756 | uid, gid := carlosIDs(t) |
| 757 | for _, d := range []string{"/data/instances/" + instance, "/run/carlos/instances/" + instance} { |
| 758 | if err := os.MkdirAll(d, 0o750); err != nil { |
| 759 | t.Fatal(err) |
| 760 | } |
| 761 | os.Chown(d, uid, gid) |
| 762 | } |
| 763 | for _, d := range []string{"/run/carlos-control/exe/" + instance + "/app", "/run/carlos-control/exe/" + instance + "/live"} { |
| 764 | if err := os.MkdirAll(d, 0o755); err != nil { |
| 765 | t.Fatal(err) |
| 766 | } |
| 767 | } |
| 768 | t.Cleanup(func() { |
| 769 | os.RemoveAll("/data/instances/" + instance) |
| 770 | os.RemoveAll("/run/carlos/instances/" + instance) |
| 771 | os.RemoveAll("/run/carlos-control/exe/" + instance) |
| 772 | }) |
| 773 | } |
| 774 | |
| 775 | func carlosIDs(t *testing.T) (int, int) { |
| 776 | t.Helper() |
| 777 | u, err := user.Lookup("carlos") |
| 778 | if err != nil { |
| 779 | t.Fatal(err) |
| 780 | } |
| 781 | uid, _ := strconv.Atoi(u.Uid) |
| 782 | gid, _ := strconv.Atoi(u.Gid) |
| 783 | return uid, gid |
| 784 | } |
| 785 | |
| 786 | // tenantReach runs the installed tenant once and returns its report, after |
| 787 | // checking that no connection reached the socket and no worker started. |
| 788 | func tenantReach(t *testing.T, instance string) map[string]string { |
| 789 | t.Helper() |
| 790 | accepted := func() string { |
| 791 | out, _ := runSystemctl("show", "-p", "NAccepted", "--value", dbReadSocketUnit) |
| 792 | return strings.TrimSpace(out) |
| 793 | } |
| 794 | before := accepted() |
| 795 | report := filepath.Join("/data/instances", instance, "dbread-witness.json") |
| 796 | os.Remove(report) |
| 797 | if out, err := runSystemctl("start", strings.Replace(dbreadTenantUnit, "@", "@"+instance, 1)); err != nil { |
| 798 | jout, _ := exec.Command("journalctl", "-u", strings.Replace(dbreadTenantUnit, "@", "@"+instance, 1), "-n", "30", "--no-pager").CombinedOutput() |
| 799 | t.Fatalf("the tenant unit did not run: %v: %s\n%s", err, out, jout) |
| 800 | } |
| 801 | raw, err := os.ReadFile(report) |
| 802 | if err != nil { |
| 803 | t.Fatalf("the tenant wrote no report: %v", err) |
| 804 | } |
| 805 | var rep map[string]string |
| 806 | if err := json.Unmarshal(raw, &rep); err != nil { |
| 807 | t.Fatal(err) |
| 808 | } |
| 809 | t.Logf("tenant report: %v", rep) |
| 810 | if rep["uid"] == "" || rep["uid"] == "0" { |
| 811 | t.Fatalf("the witness did not run as the tenant user: %v", rep) |
| 812 | } |
| 813 | for _, k := range []string{"connect " + dbReadSocketPath, "connect /run/carlos-dbread/worker.sock"} { |
| 814 | if v, ok := rep[k]; !ok || v == "" { |
| 815 | t.Errorf("a tenant could %s: %v", k, rep) |
| 816 | } |
| 817 | } |
| 818 | if after := accepted(); after != before { |
| 819 | t.Errorf("the socket accepted a tenant connection: NAccepted %s -> %s", before, after) |
| 820 | } |
| 821 | if n := unitsActive(t); n != 0 { |
| 822 | t.Errorf("%d db-read workers running after the tenant's attempt", n) |
| 823 | } |
| 824 | return rep |
| 825 | } |
| 826 | |
| 827 | // Astra round 1, finding 1, on real units: a real tenant unit (the |
| 828 | // production carlos-instance@ sandbox, running as carlos) can neither |
| 829 | // connect to nor activate a db-read worker, with db reads armed on the box |
| 830 | // and with them off; and with them off nothing listens, for the agent |
| 831 | // either. Mutations: put the socket back at /run/carlos-dbread (the tenant |
| 832 | // connects, NAccepted moves); drop the gate in ensureSystemUnits (the |
| 833 | // socket listens with CARLOS_DB_READS=0). |
| 834 | func TestRootDBReadTenantCannotReachTheWorker(t *testing.T) { |
| 835 | r := newDBReadRunner(t, "40g") |
| 836 | const instance = "tenant.example.com" |
| 837 | installDBReadTenantUnit(t, instance) |
| 838 | |
| 839 | // Armed: the socket listens, the agent's probe passes, the tenant |
| 840 | // reaches nothing. |
| 841 | if out, _ := systemctlProbe("is-active", dbReadSocketUnit); strings.TrimSpace(out) != "active" { |
| 842 | t.Fatalf("armed box: carlos-dbread.socket is %q", out) |
| 843 | } |
| 844 | if rep, _ := r.payload("probe", true, ""); rep["capable"] != true { |
| 845 | t.Fatalf("armed box: the agent's probe failed: %v", rep) |
| 846 | } |
| 847 | tenantReach(t, instance) |
| 848 | |
| 849 | for _, off := range [][]string{{"CARLOS_DB_READS=0"}, {"CARLOS_HIBERNATE_BUCKET="}} { |
| 850 | if err := os.WriteFile("/etc/carlos/host.env", []byte(dbreadRunnerHostEnv(r.rootPub, off...)), 0o640); err != nil { |
| 851 | t.Fatal(err) |
| 852 | } |
| 853 | ensureSystemUnits("/etc", filepath.Join(t.TempDir(), "lock"), runSystemctl, systemctlProbe, t.Logf) |
| 854 | active, _ := systemctlProbe("is-active", dbReadSocketUnit) |
| 855 | enabled, _ := systemctlProbe("is-enabled", dbReadSocketUnit) |
| 856 | if strings.TrimSpace(active) == "active" || strings.TrimSpace(enabled) == "enabled" { |
| 857 | t.Fatalf("%v: carlos-dbread.socket is %s and %s", off, strings.TrimSpace(active), strings.TrimSpace(enabled)) |
| 858 | } |
| 859 | if rep, _ := r.payload("probe", true, ""); rep["capable"] == true { |
| 860 | t.Fatalf("%v: something still answers on the worker socket: %v", off, rep) |
| 861 | } |
| 862 | tenantReach(t, instance) |
| 863 | } |
| 864 | } |
| 865 | |
amadan