| 1 | //go:build linux |
| 2 | |
| 3 | package cgroup |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "errors" |
| 8 | "os" |
| 9 | "os/exec" |
| 10 | "path/filepath" |
| 11 | "testing" |
| 12 | "time" |
| 13 | |
| 14 | "golang.org/x/sys/unix" |
| 15 | ) |
| 16 | |
| 17 | // amadan #272: once cgroup.kill has been written to a cgroup, every later |
| 18 | // clone3(CLONE_INTO_CGROUP) into that same cgroup is SIGKILLed at once. The |
| 19 | // fixture has no kernel behind it, so these tests pin the bookkeeping: a |
| 20 | // cgroup that took a kill is removed and made fresh (its sentinel gone, its limit |
| 21 | // files written again) before the next spawn into its path, and a removal |
| 22 | // that fails refuses the spawn. |
| 23 | |
| 24 | // fakeCgroupRmdir stands in for rmdir on cgroupfs, where the interface files |
| 25 | // do not count against it: the fixture's are ordinary files, so they go |
| 26 | // first. fail, while non-nil, is returned instead. It records each removal. |
| 27 | func fakeCgroupRmdir(t *testing.T, fail *error) *[]string { |
| 28 | t.Helper() |
| 29 | var removed []string |
| 30 | orig := rmdirCgroup |
| 31 | rmdirCgroup = func(path string) error { |
| 32 | if fail != nil && *fail != nil { |
| 33 | return *fail |
| 34 | } |
| 35 | entries, err := os.ReadDir(path) |
| 36 | if err != nil { |
| 37 | return err |
| 38 | } |
| 39 | for _, e := range entries { |
| 40 | if !e.IsDir() { |
| 41 | if err := os.Remove(filepath.Join(path, e.Name())); err != nil { |
| 42 | return err |
| 43 | } |
| 44 | } |
| 45 | } |
| 46 | if err := unix.Rmdir(path); err != nil { |
| 47 | return err |
| 48 | } |
| 49 | removed = append(removed, path) |
| 50 | return nil |
| 51 | } |
| 52 | t.Cleanup(func() { rmdirCgroup = orig }) |
| 53 | return &removed |
| 54 | } |
| 55 | |
| 56 | // forgetKillsAfter clears the kill marks a test leaves, so a later test's |
| 57 | // temp dir can never meet one. |
| 58 | func forgetKillsAfter(t *testing.T) { |
| 59 | t.Helper() |
| 60 | t.Cleanup(forgetKills) |
| 61 | } |
| 62 | |
| 63 | func init() { |
| 64 | forgetKills = func() { |
| 65 | killedCgroups.Lock() |
| 66 | clear(killedCgroups.ids) |
| 67 | killedCgroups.Unlock() |
| 68 | } |
| 69 | } |
| 70 | |
| 71 | // sentinelName is a file production code never creates in a cgroup. A test |
| 72 | // plants it in a directory to learn later whether that very directory was |
| 73 | // removed and made again (the sentinel is gone) or left alone (it is still |
| 74 | // there). Freshness is proved this way, not by a changed inode number: |
| 75 | // ext4 and xfs hand a directory's inode straight back after rmdir+mkdir, |
| 76 | // so an inode comparison passes or fails by filesystem. Real cgroupfs does |
| 77 | // not reuse cgroup inode ids, which is why the production marks can key on |
| 78 | // identity. |
| 79 | const sentinelName = "test-sentinel" |
| 80 | |
| 81 | func plantSentinel(t *testing.T, dir string) { |
| 82 | t.Helper() |
| 83 | writeFixture(t, filepath.Join(dir, sentinelName), "planted") |
| 84 | } |
| 85 | |
| 86 | func hasSentinel(t *testing.T, dir string) bool { |
| 87 | t.Helper() |
| 88 | _, err := os.Lstat(filepath.Join(dir, sentinelName)) |
| 89 | if err != nil && !errors.Is(err, os.ErrNotExist) { |
| 90 | t.Fatalf("stat sentinel in %s: %v", dir, err) |
| 91 | } |
| 92 | return err == nil |
| 93 | } |
| 94 | |
| 95 | // killFixture gives dir the files a cleanup handle reads, empty, and writes |
| 96 | // cgroup.kill through a cleanup handle on host (child "" is the host's own |
| 97 | // cgroup), as KillHost and KillSidecar do. |
| 98 | func killFixture(t *testing.T, m *Manager, host, child, dir string) { |
| 99 | t.Helper() |
| 100 | for name, body := range map[string]string{"cgroup.events": "populated 0\n", "cgroup.procs": "", "cgroup.kill": ""} { |
| 101 | writeFixture(t, filepath.Join(dir, name), body) |
| 102 | } |
| 103 | c, err := m.openCleanupAt(host, child, func(*os.File) error { return nil }) |
| 104 | if err != nil { |
| 105 | t.Fatal(err) |
| 106 | } |
| 107 | ctx, cancel := context.WithTimeout(t.Context(), time.Second) |
| 108 | defer cancel() |
| 109 | if err := c.KillAndEmpty(ctx); err != nil { |
| 110 | t.Fatal(err) |
| 111 | } |
| 112 | if err := c.Close(); err != nil { |
| 113 | t.Fatal(err) |
| 114 | } |
| 115 | if got := readFixture(t, filepath.Join(dir, "cgroup.kill")); got != "1" { |
| 116 | t.Fatalf("cgroup.kill = %q, want 1", got) |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | // A worker's host cgroup (KillHost's) is killed and kept. The next start |
| 121 | // into it removes it and creates it again, with its limit files written as |
| 122 | // a first creation writes them, and a start that follows no kill removes |
| 123 | // nothing. |
| 124 | func TestAKilledHostCgroupIsRecreatedBeforeTheNextStart(t *testing.T) { |
| 125 | forgetKillsAfter(t) |
| 126 | quietOOM(t) |
| 127 | m, err := SetupAt(newDelegatedRoot(t), t.Logf) |
| 128 | if err != nil { |
| 129 | t.Fatal(err) |
| 130 | } |
| 131 | removed := fakeCgroupRmdir(t, nil) |
| 132 | calls := fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 133 | const host = "worker.example.com" |
| 134 | dir := m.hostDir(host) |
| 135 | for range 2 { |
| 136 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 137 | t.Fatal(err) |
| 138 | } |
| 139 | } |
| 140 | if len(*removed) != 0 { |
| 141 | t.Fatalf("starts that followed no kill removed %v", *removed) |
| 142 | } |
| 143 | killFixture(t, m, host, "", dir) |
| 144 | plantSentinel(t, dir) |
| 145 | // A limit file left behind would say nothing about the recreation; |
| 146 | // scribble on it so only a fresh write can restore it. |
| 147 | writeFixture(t, filepath.Join(dir, "memory.max"), "max") |
| 148 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 149 | t.Fatal(err) |
| 150 | } |
| 151 | if len(*calls) != 3 { |
| 152 | t.Fatalf("started %d times, want 3", len(*calls)) |
| 153 | } |
| 154 | if len(*removed) != 1 || (*removed)[0] != dir { |
| 155 | t.Fatalf("removed %v, want only %s", *removed, dir) |
| 156 | } |
| 157 | if hasSentinel(t, dir) { |
| 158 | t.Fatal("the respawn reused the killed cgroup (the sentinel planted in it survived)") |
| 159 | } |
| 160 | for file, want := range map[string]string{ |
| 161 | "memory.max": "536870912", |
| 162 | "memory.high": "483183819", // HighOf: max/10*9 |
| 163 | "memory.swap.max": "536870912", |
| 164 | "memory.oom.group": "1", |
| 165 | } { |
| 166 | if got := readFixture(t, filepath.Join(dir, file)); got != want { |
| 167 | t.Errorf("%s = %q, want %q", file, got, want) |
| 168 | } |
| 169 | } |
| 170 | // The fresh cgroup took no kill: the next start keeps it. |
| 171 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 172 | t.Fatal(err) |
| 173 | } |
| 174 | if len(*removed) != 1 { |
| 175 | t.Fatalf("a start after the recreation removed again: %v", *removed) |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | // A killed cgroup that will not go (EBUSY: something is in it again) is not |
| 180 | // proven fresh, so the start refuses and execs nothing. Once it can be |
| 181 | // removed, the next start proceeds into a fresh one. |
| 182 | func TestAKilledCgroupThatCannotBeRemovedRefusesTheNextStart(t *testing.T) { |
| 183 | forgetKillsAfter(t) |
| 184 | quietOOM(t) |
| 185 | m, err := SetupAt(newDelegatedRoot(t), t.Logf) |
| 186 | if err != nil { |
| 187 | t.Fatal(err) |
| 188 | } |
| 189 | fail := error(unix.EBUSY) |
| 190 | removed := fakeCgroupRmdir(t, &fail) |
| 191 | calls := fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 192 | const host = "worker.example.com" |
| 193 | dir := m.hostDir(host) |
| 194 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 195 | t.Fatal(err) |
| 196 | } |
| 197 | killFixture(t, m, host, "", dir) |
| 198 | plantSentinel(t, dir) |
| 199 | built := 0 |
| 200 | cmd, err := m.Start(host, TierDefault, func() *exec.Cmd { built++; return buildTrue() }) |
| 201 | if !errors.Is(err, ErrKilledCgroupKept) || !errors.Is(err, unix.EBUSY) || cmd != nil { |
| 202 | t.Fatalf("start into a killed cgroup that would not go = %v, %v; want ErrKilledCgroupKept (EBUSY)", cmd, err) |
| 203 | } |
| 204 | if built != 0 || len(*calls) != 1 { |
| 205 | t.Fatalf("the refused start built %d command(s) and started %d time(s) in all, want 0 and 1", built, len(*calls)) |
| 206 | } |
| 207 | if !hasSentinel(t, dir) { |
| 208 | t.Fatal("the refused start replaced the cgroup") |
| 209 | } |
| 210 | // Still refused while it will not go: the mark outlives the failure. |
| 211 | if _, err := m.Start(host, TierDefault, buildTrue); !errors.Is(err, ErrKilledCgroupKept) { |
| 212 | t.Fatalf("second start = %v, want ErrKilledCgroupKept", err) |
| 213 | } |
| 214 | fail = nil |
| 215 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 216 | t.Fatal(err) |
| 217 | } |
| 218 | if len(*removed) != 1 || hasSentinel(t, dir) { |
| 219 | t.Fatalf("removed %v; the start after the removal succeeded must be in a fresh cgroup", *removed) |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | // The legacy sidecar's role child (KillSidecar's) is killed and kept. The |
| 224 | // respawn gets a fresh role-sidecar under the very same parent, whose limit |
| 225 | // and accounting are left as they were. |
| 226 | func TestAKilledRoleSidecarIsRecreatedUnderTheSameParent(t *testing.T) { |
| 227 | forgetKillsAfter(t) |
| 228 | const host = "sc.example.com" |
| 229 | m, parent := genManager(t, host) |
| 230 | removed := fakeCgroupRmdir(t, nil) |
| 231 | fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 232 | child := RoleChild("sidecar") |
| 233 | dir := filepath.Join(parent, child) |
| 234 | if _, err := m.StartChild(host, child, TierDefault, Limit{}, buildTrue); err != nil { |
| 235 | t.Fatal(err) |
| 236 | } |
| 237 | killFixture(t, m, host, child, dir) |
| 238 | plantSentinel(t, dir) |
| 239 | plantSentinel(t, parent) |
| 240 | writeFixture(t, filepath.Join(parent, "memory.max"), "1073741824") // a raise the respawn must not undo |
| 241 | writeFixture(t, filepath.Join(parent, "cgroup.events"), "populated 1\n") |
| 242 | if _, err := m.StartChild(host, child, TierDefault, Limit{}, buildTrue); err != nil { |
| 243 | t.Fatal(err) |
| 244 | } |
| 245 | if len(*removed) != 1 || (*removed)[0] != dir { |
| 246 | t.Fatalf("removed %v, want only %s", *removed, dir) |
| 247 | } |
| 248 | if hasSentinel(t, dir) { |
| 249 | t.Fatal("the sidecar respawn reused the killed role child") |
| 250 | } |
| 251 | if !hasSentinel(t, parent) { |
| 252 | t.Fatal("the sidecar respawn replaced the parent") |
| 253 | } |
| 254 | if got := readFixture(t, filepath.Join(parent, "memory.max")); got != "1073741824" { |
| 255 | t.Fatalf("parent memory.max = %q, the respawn touched it", got) |
| 256 | } |
| 257 | if _, err := os.Stat(filepath.Join(dir, "memory.max")); !errors.Is(err, os.ErrNotExist) { |
| 258 | t.Fatalf("a role child carries no limit of its own: %v", err) |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | // KillHost on a generation parent kills everything under it. The next start |
| 263 | // of any child removes the whole killed tree, children first, and builds |
| 264 | // the parent again as a cold start does: controllers delegated, the tier |
| 265 | // written with memory.oom.group=0. |
| 266 | func TestAKilledParentIsRecreatedWithItsChildren(t *testing.T) { |
| 267 | forgetKillsAfter(t) |
| 268 | const host = "gp.example.com" |
| 269 | m, parent := genManager(t, host) |
| 270 | removed := fakeCgroupRmdir(t, nil) |
| 271 | fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 272 | sidecar := RoleChild("sidecar") |
| 273 | if _, err := m.StartChild(host, sidecar, TierDefault, Limit{}, buildTrue); err != nil { |
| 274 | t.Fatal(err) |
| 275 | } |
| 276 | if _, err := m.StartChild(host, "g4", TierDefault, TierDefault, buildTrue); err != nil { |
| 277 | t.Fatal(err) |
| 278 | } |
| 279 | killFixture(t, m, host, "", parent) |
| 280 | plantSentinel(t, parent) |
| 281 | plantSentinel(t, filepath.Join(parent, sidecar)) |
| 282 | plantSentinel(t, filepath.Join(parent, "g4")) |
| 283 | // The fixture's parent would lose its controller list with it; the |
| 284 | // kernel lists them on every new cgroup, so the fake rmdir's recreation |
| 285 | // needs it back. |
| 286 | orig := rmdirCgroup |
| 287 | rmdirCgroup = func(path string) error { |
| 288 | if err := orig(path); err != nil { |
| 289 | return err |
| 290 | } |
| 291 | if path == parent { |
| 292 | if err := os.Mkdir(parent, 0o755); err != nil { |
| 293 | return err |
| 294 | } |
| 295 | writeFixture(t, filepath.Join(parent, "cgroup.controllers"), "cpu io memory pids\n") |
| 296 | } |
| 297 | return nil |
| 298 | } |
| 299 | t.Cleanup(func() { rmdirCgroup = orig }) |
| 300 | if _, err := m.StartChild(host, sidecar, TierDefault, Limit{}, buildTrue); err != nil { |
| 301 | t.Fatal(err) |
| 302 | } |
| 303 | want := map[string]bool{filepath.Join(parent, sidecar): true, filepath.Join(parent, "g4"): true} |
| 304 | if len(*removed) != 3 || !want[(*removed)[0]] || !want[(*removed)[1]] || (*removed)[2] != parent { |
| 305 | t.Fatalf("removed %v, want both children then %s", *removed, parent) |
| 306 | } |
| 307 | if hasSentinel(t, parent) { |
| 308 | t.Fatal("the respawn reused the killed parent") |
| 309 | } |
| 310 | if _, err := os.Stat(filepath.Join(parent, "g4")); !errors.Is(err, os.ErrNotExist) { |
| 311 | t.Fatalf("the killed generation child survived: %v", err) |
| 312 | } |
| 313 | for file, want := range map[string]string{ |
| 314 | "cgroup.subtree_control": "+cpu +memory +io +pids", |
| 315 | "memory.max": "536870912", |
| 316 | "memory.oom.group": "0", |
| 317 | } { |
| 318 | if got := readFixture(t, filepath.Join(parent, file)); got != want { |
| 319 | t.Errorf("parent %s = %q, want %q", file, got, want) |
| 320 | } |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | // A killed cgroup that a strict handle still pins (its reservation is |
| 325 | // held) is not removed under it: removing it would strand the handle, whose |
| 326 | // identity is that directory. The start refuses until the handle lets go. |
| 327 | func TestAKilledCgroupAStrictHandleHoldsIsNotRemoved(t *testing.T) { |
| 328 | forgetKillsAfter(t) |
| 329 | quietOOM(t) |
| 330 | m, err := SetupAt(newDelegatedRoot(t), t.Logf) |
| 331 | if err != nil { |
| 332 | t.Fatal(err) |
| 333 | } |
| 334 | removed := fakeCgroupRmdir(t, nil) |
| 335 | fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 336 | const host = "pinned.example.com" |
| 337 | dir := m.hostDir(host) |
| 338 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 339 | t.Fatal(err) |
| 340 | } |
| 341 | killFixture(t, m, host, "", dir) |
| 342 | var st unix.Stat_t |
| 343 | if err := unix.Lstat(dir, &st); err != nil { |
| 344 | t.Fatal(err) |
| 345 | } |
| 346 | id := strictIdentity{uint64(st.Dev), st.Ino} |
| 347 | strictReservations.Lock() |
| 348 | strictReservations.held[id] = true |
| 349 | strictReservations.Unlock() |
| 350 | release := func() { |
| 351 | strictReservations.Lock() |
| 352 | delete(strictReservations.held, id) |
| 353 | strictReservations.Unlock() |
| 354 | } |
| 355 | t.Cleanup(release) |
| 356 | if _, err := m.Start(host, TierDefault, buildTrue); !errors.Is(err, ErrStrictState) || !errors.Is(err, ErrKilledCgroupKept) { |
| 357 | t.Fatalf("start into a pinned killed cgroup = %v, want ErrKilledCgroupKept and ErrStrictState", err) |
| 358 | } |
| 359 | if len(*removed) != 0 { |
| 360 | t.Fatalf("removed %v under a strict handle", *removed) |
| 361 | } |
| 362 | release() |
| 363 | if _, err := m.Start(host, TierDefault, buildTrue); err != nil { |
| 364 | t.Fatal(err) |
| 365 | } |
| 366 | if len(*removed) != 1 { |
| 367 | t.Fatalf("removed %v once the handle let go, want the killed cgroup", *removed) |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | // A strict role child that took a kill is never pinned again for a spawn. |
| 372 | func TestAKilledStrictRoleChildIsNotPinnedAgain(t *testing.T) { |
| 373 | forgetKillsAfter(t) |
| 374 | p := strictFixture(t) |
| 375 | makeRoleFixture(t, p, "seed", "populated 0\n") |
| 376 | child, err := p.PrepareRoleChild("seed") |
| 377 | if err != nil { |
| 378 | t.Fatal(err) |
| 379 | } |
| 380 | ctx, cancel := context.WithTimeout(t.Context(), time.Second) |
| 381 | defer cancel() |
| 382 | if err := child.KillAndEmpty(ctx); err != nil { |
| 383 | t.Fatal(err) |
| 384 | } |
| 385 | if err := child.Close(); err != nil { |
| 386 | t.Fatal(err) |
| 387 | } |
| 388 | again, err := p.PrepareRoleChild("seed") |
| 389 | if err == nil { |
| 390 | again.Close() |
| 391 | t.Fatal("a killed role child was pinned again for a spawn") |
| 392 | } |
| 393 | if !errors.Is(err, ErrStrictRefused) { |
| 394 | t.Fatalf("err = %v, want ErrStrictRefused", err) |
| 395 | } |
| 396 | } |
| 397 | |
| 398 | // A killed parent that a strict handle still pins keeps everything below it |
| 399 | // too: the reservation is checked before the walk touches a child (Task 19 |
| 400 | // fix round 1). |
| 401 | func TestAPinnedKilledParentLosesNoChild(t *testing.T) { |
| 402 | forgetKillsAfter(t) |
| 403 | const host = "pinnedparent.example.com" |
| 404 | m, parent := genManager(t, host) |
| 405 | removed := fakeCgroupRmdir(t, nil) |
| 406 | fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 407 | sidecar := RoleChild("sidecar") |
| 408 | if _, err := m.StartChild(host, sidecar, TierDefault, Limit{}, buildTrue); err != nil { |
| 409 | t.Fatal(err) |
| 410 | } |
| 411 | killFixture(t, m, host, "", parent) |
| 412 | var st unix.Stat_t |
| 413 | if err := unix.Lstat(parent, &st); err != nil { |
| 414 | t.Fatal(err) |
| 415 | } |
| 416 | id := strictIdentity{uint64(st.Dev), st.Ino} |
| 417 | strictReservations.Lock() |
| 418 | strictReservations.held[id] = true |
| 419 | strictReservations.Unlock() |
| 420 | t.Cleanup(func() { |
| 421 | strictReservations.Lock() |
| 422 | delete(strictReservations.held, id) |
| 423 | strictReservations.Unlock() |
| 424 | }) |
| 425 | if _, err := m.StartChild(host, sidecar, TierDefault, Limit{}, buildTrue); !errors.Is(err, ErrStrictState) { |
| 426 | t.Fatalf("start under a pinned killed parent = %v, want ErrStrictState", err) |
| 427 | } |
| 428 | if len(*removed) != 0 { |
| 429 | t.Fatalf("removed %v under a pinned parent", *removed) |
| 430 | } |
| 431 | if _, err := os.Stat(filepath.Join(parent, sidecar)); err != nil { |
| 432 | t.Fatalf("the pinned parent's child went: %v", err) |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | // retainedTree builds what an earlier run left under a delegated root: |
| 437 | // carlos-exec@<host> with cgroup.events and cgroup.procs, and children |
| 438 | // (name to events body) with the same files, empty. |
| 439 | func retainedTree(t *testing.T, root, host, events, procs string, kids map[string]string) string { |
| 440 | t.Helper() |
| 441 | dir := filepath.Join(root, ExecName(host)) |
| 442 | if err := os.Mkdir(dir, 0o755); err != nil { |
| 443 | t.Fatal(err) |
| 444 | } |
| 445 | writeFixture(t, filepath.Join(dir, "cgroup.events"), events) |
| 446 | writeFixture(t, filepath.Join(dir, "cgroup.procs"), procs) |
| 447 | writeFixture(t, filepath.Join(dir, "cgroup.controllers"), "cpu io memory pids\n") |
| 448 | for kid, body := range kids { |
| 449 | k := filepath.Join(dir, kid) |
| 450 | if err := os.Mkdir(k, 0o755); err != nil { |
| 451 | t.Fatal(err) |
| 452 | } |
| 453 | writeFixture(t, filepath.Join(k, "cgroup.events"), body) |
| 454 | writeFixture(t, filepath.Join(k, "cgroup.procs"), "") |
| 455 | } |
| 456 | return dir |
| 457 | } |
| 458 | |
| 459 | // Claim-entry mode (SetupRetaining) skips the sweep, and the kill marks die |
| 460 | // with the process that made them, so a cgroup an earlier run killed and |
| 461 | // kept would be cloned into again. The retaining setup marks every empty |
| 462 | // exec cgroup (and every empty child of one) instead: the first start into |
| 463 | // each makes it fresh with its limits written again. One that holds a |
| 464 | // process, and a claim containment group, are left exactly as they are |
| 465 | // (Task 19 fix round 1). |
| 466 | func TestARetainingSetupMakesEmptyLeftoversFresh(t *testing.T) { |
| 467 | forgetKillsAfter(t) |
| 468 | quietOOM(t) |
| 469 | root := newDelegatedRoot(t) |
| 470 | empty := retainedTree(t, root, "worker.example.com", "populated 0\n", "", nil) |
| 471 | busy := retainedTree(t, root, "busy.example.com", "populated 1\n", "4242\n", nil) |
| 472 | // A generation parent whose live g3 survived, beside an empty sidecar |
| 473 | // child a KillSidecar left. |
| 474 | live := retainedTree(t, root, "live.example.com", "populated 1\n", "", map[string]string{"g3": "populated 1\n", "role-sidecar": "populated 0\n"}) |
| 475 | writeFixture(t, filepath.Join(live, "cgroup.subtree_control"), "cpu memory io pids\n") |
| 476 | claim := retainedTree(t, root, "claim-0123abcd.local", "populated 0\n", "", nil) |
| 477 | for _, d := range []string{empty, busy, live, filepath.Join(live, "g3"), filepath.Join(live, "role-sidecar")} { |
| 478 | plantSentinel(t, d) |
| 479 | } |
| 480 | m, err := SetupAtRetaining(root, t.Logf) |
| 481 | if err != nil { |
| 482 | t.Fatal(err) |
| 483 | } |
| 484 | removed := fakeCgroupRmdir(t, nil) |
| 485 | fakeStart(t, func(*exec.Cmd) error { return nil }) |
| 486 | FakePlacementForTest(t) // a temp dir is no cgroupfs, and a role child is never moved by PID |
| 487 | |
| 488 | if _, err := m.Start("worker.example.com", TierDefault, buildTrue); err != nil { |
| 489 | t.Fatal(err) |
| 490 | } |
| 491 | if hasSentinel(t, empty) { |
| 492 | t.Fatal("the first start reused an empty retained host cgroup") |
| 493 | } |
| 494 | if got := readFixture(t, filepath.Join(empty, "memory.max")); got != "536870912" { |
| 495 | t.Fatalf("recreated memory.max = %q, want the tier", got) |
| 496 | } |
| 497 | |
| 498 | if _, err := m.Start("busy.example.com", TierDefault, buildTrue); err != nil { |
| 499 | t.Fatal(err) |
| 500 | } |
| 501 | if !hasSentinel(t, busy) { |
| 502 | t.Fatal("a retained cgroup holding a process was replaced") |
| 503 | } |
| 504 | |
| 505 | if _, err := m.StartChild("live.example.com", RoleChild("sidecar"), TierDefault, Limit{}, buildTrue); err != nil { |
| 506 | t.Fatal(err) |
| 507 | } |
| 508 | if hasSentinel(t, filepath.Join(live, "role-sidecar")) { |
| 509 | t.Fatal("the sidecar respawn reused an empty retained role child") |
| 510 | } |
| 511 | if !hasSentinel(t, live) || !hasSentinel(t, filepath.Join(live, "g3")) { |
| 512 | t.Fatal("a populated parent or its live generation was replaced") |
| 513 | } |
| 514 | |
| 515 | if id, ok, err := pathIdentity(claim); err != nil || !ok || wasKilled(id) { |
| 516 | t.Fatalf("a claim containment group was marked (ok=%v err=%v)", ok, err) |
| 517 | } |
| 518 | want := map[string]bool{empty: true, filepath.Join(live, "role-sidecar"): true} |
| 519 | if len(*removed) != 2 || !want[(*removed)[0]] || !want[(*removed)[1]] { |
| 520 | t.Fatalf("removed %v, want only the empty host cgroup and the empty role child", *removed) |
| 521 | } |
| 522 | } |
| 523 | |
amadan