| 1 | //go:build linux |
| 2 | |
| 3 | package cgroup |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "io" |
| 10 | "os" |
| 11 | "os/exec" |
| 12 | "path/filepath" |
| 13 | "strings" |
| 14 | "sync" |
| 15 | "sync/atomic" |
| 16 | "syscall" |
| 17 | "time" |
| 18 | |
| 19 | "github.com/carlosframework/platform/internal/controlname" |
| 20 | "golang.org/x/sys/unix" |
| 21 | ) |
| 22 | |
| 23 | type strictIdentity struct{ device, inode uint64 } |
| 24 | |
| 25 | var strictReservations = struct { |
| 26 | sync.Mutex |
| 27 | held map[strictIdentity]bool |
| 28 | }{held: make(map[strictIdentity]bool)} |
| 29 | |
| 30 | // Prepared pins one empty host cgroup. It grants no lease, fence or database |
| 31 | // authority. The caller must also obtain Cmd.Wait evidence after cleanup. |
| 32 | type Prepared struct{ *strictPreparedState } |
| 33 | |
| 34 | type strictPreparedState struct { |
| 35 | mu sync.Mutex |
| 36 | root, dir, kill, events *os.File |
| 37 | limits map[string]*os.File |
| 38 | expected map[string]string |
| 39 | identity strictIdentity |
| 40 | entryParent *os.File |
| 41 | entryName string |
| 42 | parent *Prepared |
| 43 | draining atomic.Bool |
| 44 | host string |
| 45 | attempted, empty, closed bool |
| 46 | start func(*exec.Cmd) error |
| 47 | // genParent marks a generation parent (PrepareStrictParent): it holds |
| 48 | // no process, pins PrepareGen children, and moves its limit only |
| 49 | // through SetParentLimit. tier is the tier it was prepared with, or a |
| 50 | // generation child's own. |
| 51 | genParent bool |
| 52 | tier Limit |
| 53 | } |
| 54 | |
| 55 | func strictOpen(dir *os.File, name string, flags int) (*os.File, error) { |
| 56 | fd, e := unix.Openat(int(dir.Fd()), name, flags|unix.O_CLOEXEC|unix.O_NOFOLLOW, 0) |
| 57 | if e != nil { |
| 58 | return nil, e |
| 59 | } |
| 60 | return os.NewFile(uintptr(fd), name), nil |
| 61 | } |
| 62 | func strictID(f *os.File) (strictIdentity, error) { |
| 63 | var st unix.Stat_t |
| 64 | if e := unix.Fstat(int(f.Fd()), &st); e != nil { |
| 65 | return strictIdentity{}, e |
| 66 | } |
| 67 | return strictIdentity{uint64(st.Dev), st.Ino}, nil |
| 68 | } |
| 69 | func strictRead(f *os.File) (string, error) { |
| 70 | if _, e := f.Seek(0, 0); e != nil { |
| 71 | return "", e |
| 72 | } |
| 73 | b, e := io.ReadAll(io.LimitReader(f, 4097)) |
| 74 | if e != nil || len(b) > 4096 { |
| 75 | return "", ErrStrictRefused |
| 76 | } |
| 77 | return strings.TrimSpace(string(b)), nil |
| 78 | } |
| 79 | func strictWrite(f *os.File, value string) error { |
| 80 | if _, e := f.Seek(0, 0); e != nil { |
| 81 | return e |
| 82 | } |
| 83 | n, e := f.WriteString(value) |
| 84 | if e != nil { |
| 85 | return e |
| 86 | } |
| 87 | if n != len(value) { |
| 88 | return io.ErrShortWrite |
| 89 | } |
| 90 | return nil |
| 91 | } |
| 92 | func strictPopulation(f *os.File) (bool, error) { |
| 93 | raw, e := strictRead(f) |
| 94 | if e != nil { |
| 95 | return false, e |
| 96 | } |
| 97 | seen := false |
| 98 | populated := false |
| 99 | for _, line := range strings.Split(raw, "\n") { |
| 100 | parts := strings.Fields(line) |
| 101 | if len(parts) != 2 { |
| 102 | return false, ErrStrictRefused |
| 103 | } |
| 104 | if parts[0] == "populated" { |
| 105 | if seen || (parts[1] != "0" && parts[1] != "1") { |
| 106 | return false, ErrStrictRefused |
| 107 | } |
| 108 | seen = true |
| 109 | populated = parts[1] == "1" |
| 110 | } |
| 111 | } |
| 112 | if !seen { |
| 113 | return false, ErrStrictRefused |
| 114 | } |
| 115 | return populated, nil |
| 116 | } |
| 117 | func (p *Prepared) identityOK() error { |
| 118 | if p.closed { |
| 119 | return ErrStrictState |
| 120 | } |
| 121 | id, e := strictID(p.dir) |
| 122 | if e != nil || id != p.identity { |
| 123 | return ErrStrictRefused |
| 124 | } |
| 125 | var st unix.Stat_t |
| 126 | parent, name := p.root, ExecName(p.host) |
| 127 | if p.entryParent != nil { |
| 128 | parent, name = p.entryParent, p.entryName |
| 129 | } |
| 130 | if e = unix.Fstatat(int(parent.Fd()), name, &st, unix.AT_SYMLINK_NOFOLLOW); e != nil || st.Mode&unix.S_IFMT != unix.S_IFDIR || (strictIdentity{uint64(st.Dev), st.Ino}) != p.identity { |
| 131 | return ErrStrictRefused |
| 132 | } |
| 133 | return nil |
| 134 | } |
| 135 | func (p *Prepared) limitsOK() error { |
| 136 | for name, want := range p.expected { |
| 137 | got, e := strictRead(p.limits[name]) |
| 138 | if e != nil || got != want { |
| 139 | return ErrStrictRefused |
| 140 | } |
| 141 | } |
| 142 | return nil |
| 143 | } |
| 144 | func (p *Prepared) closeFiles() { |
| 145 | for _, f := range p.limits { |
| 146 | f.Close() |
| 147 | } |
| 148 | for _, f := range []*os.File{p.events, p.kill, p.dir, p.entryParent, p.root} { |
| 149 | if f != nil { |
| 150 | f.Close() |
| 151 | } |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | // selfCgroupFile is where the strict preflight reads this process's own |
| 156 | // cgroup. A variable so a test can place the agent where it needs. |
| 157 | var selfCgroupFile = "/proc/self/cgroup" |
| 158 | |
| 159 | // strictRoot is the host-independent half of PrepareStrict's runtime |
| 160 | // checks, and the whole of StrictPreflight: a delegated root, this process |
| 161 | // placed in it or in its supervise child, and the root a real cgroup2 |
| 162 | // directory. It returns the root opened without following a symlink. |
| 163 | func (m *Manager) strictRoot() (*os.File, error) { |
| 164 | if m == nil || m.root == "" { |
| 165 | return nil, ErrStrictUnsupported |
| 166 | } |
| 167 | own, e := os.ReadFile(selfCgroupFile) |
| 168 | rel, ok := ParseSelfCgroup(string(own)) |
| 169 | if e != nil || !ok { |
| 170 | return nil, ErrStrictUnsupported |
| 171 | } |
| 172 | current := filepath.Join("/sys/fs/cgroup", rel) |
| 173 | rootPath := filepath.Clean(m.root) |
| 174 | if current != rootPath && current != filepath.Join(rootPath, "supervise") { |
| 175 | return nil, ErrStrictRefused |
| 176 | } |
| 177 | root, e := os.OpenFile(rootPath, os.O_RDONLY|unix.O_DIRECTORY|unix.O_NOFOLLOW, 0) |
| 178 | if e != nil { |
| 179 | return nil, ErrStrictUnsupported |
| 180 | } |
| 181 | var fs unix.Statfs_t |
| 182 | if e = unix.Fstatfs(int(root.Fd()), &fs); e != nil || uint64(fs.Type) != uint64(unix.CGROUP2_SUPER_MAGIC) { |
| 183 | root.Close() |
| 184 | return nil, ErrStrictUnsupported |
| 185 | } |
| 186 | return root, nil |
| 187 | } |
| 188 | |
| 189 | // StrictPreflight reports whether PrepareStrict's runtime checks that do |
| 190 | // not depend on the host pass now: the same function, not a copy, so a |
| 191 | // readiness check (moves S2) cannot approve a box PrepareStrict refuses |
| 192 | // for its placement or its mount. Nil means a strict spawn can be |
| 193 | // attempted; the per-host state (a populated group, a held reservation) |
| 194 | // is still PrepareStrict's to judge. |
| 195 | func (m *Manager) StrictPreflight() error { |
| 196 | root, e := m.strictRoot() |
| 197 | if e != nil { |
| 198 | return e |
| 199 | } |
| 200 | return root.Close() |
| 201 | } |
| 202 | |
| 203 | // PrepareStrict never uses accounting's fallback. Only the two existing tiers |
| 204 | // are accepted; a zero or missing limit cannot create an uncapped command. |
| 205 | func (m *Manager) PrepareStrict(host string, limit Limit) (*Prepared, error) { |
| 206 | return m.prepareStrict(host, limit, false) |
| 207 | } |
| 208 | |
| 209 | // prepareStrict is PrepareStrict and PrepareStrictParent: with genParent the |
| 210 | // group is pinned as a generation parent (memory.oom.group=0, controllers |
| 211 | // delegated to its children, Start refused) instead of a leaf. |
| 212 | func (m *Manager) prepareStrict(host string, limit Limit, genParent bool) (out *Prepared, err error) { |
| 213 | if m == nil || m.root == "" { |
| 214 | return nil, ErrStrictUnsupported |
| 215 | } |
| 216 | if !controlname.ValidHostShape(host) || (limit != TierDefault && limit != TierChildren) { |
| 217 | return nil, ErrStrictRefused |
| 218 | } |
| 219 | root, e := m.strictRoot() |
| 220 | if e != nil { |
| 221 | return nil, e |
| 222 | } |
| 223 | p := &Prepared{strictPreparedState: &strictPreparedState{root: root, host: host, limits: map[string]*os.File{}, start: func(cmd *exec.Cmd) error { return cmd.Start() }}} |
| 224 | reserved := false |
| 225 | defer func() { |
| 226 | if err != nil { |
| 227 | p.closeFiles() |
| 228 | if reserved { |
| 229 | strictReservations.Lock() |
| 230 | delete(strictReservations.held, p.identity) |
| 231 | strictReservations.Unlock() |
| 232 | } |
| 233 | } |
| 234 | }() |
| 235 | // A group that took cgroup.kill is removed and made fresh first (amadan |
| 236 | // #272). One a strict handle still holds stays, and the answer is the |
| 237 | // reservation's own ErrStrictState. |
| 238 | if e = m.freshen(host, m.hostDir(host)); e != nil { |
| 239 | if errors.Is(e, ErrStrictState) { |
| 240 | return nil, e |
| 241 | } |
| 242 | return nil, fmt.Errorf("%w: %w", ErrStrictRefused, e) |
| 243 | } |
| 244 | if e = unix.Mkdirat(int(root.Fd()), ExecName(host), 0755); e != nil && !errors.Is(e, unix.EEXIST) { |
| 245 | return nil, ErrStrictRefused |
| 246 | } |
| 247 | p.dir, e = strictOpen(root, ExecName(host), unix.O_RDONLY|unix.O_DIRECTORY) |
| 248 | if e != nil { |
| 249 | return nil, ErrStrictRefused |
| 250 | } |
| 251 | p.identity, e = strictID(p.dir) |
| 252 | if e != nil { |
| 253 | return nil, ErrStrictRefused |
| 254 | } |
| 255 | strictReservations.Lock() |
| 256 | if strictReservations.held[p.identity] { |
| 257 | strictReservations.Unlock() |
| 258 | return nil, ErrStrictState |
| 259 | } |
| 260 | strictReservations.held[p.identity] = true |
| 261 | reserved = true |
| 262 | strictReservations.Unlock() |
| 263 | p.events, e = strictOpen(p.dir, "cgroup.events", unix.O_RDONLY) |
| 264 | if e != nil { |
| 265 | return nil, ErrStrictUnsupported |
| 266 | } |
| 267 | p.kill, e = strictOpen(p.dir, "cgroup.kill", unix.O_WRONLY) |
| 268 | if e != nil { |
| 269 | return nil, ErrStrictUnsupported |
| 270 | } |
| 271 | if populated, e := strictPopulation(p.events); e != nil || populated { |
| 272 | return nil, ErrStrictRefused |
| 273 | } |
| 274 | oomGroup := "1" |
| 275 | if genParent { |
| 276 | oomGroup = "0" |
| 277 | } |
| 278 | p.expected = tierExpected(limit, oomGroup) |
| 279 | for name := range p.expected { |
| 280 | p.limits[name], e = strictOpen(p.dir, name, unix.O_RDWR) |
| 281 | if e != nil { |
| 282 | return nil, ErrStrictUnsupported |
| 283 | } |
| 284 | } |
| 285 | for _, name := range []string{"memory.max", "memory.high", "memory.swap.max", "memory.oom.group"} { |
| 286 | if e = strictWrite(p.limits[name], p.expected[name]); e != nil { |
| 287 | return nil, ErrStrictRefused |
| 288 | } |
| 289 | } |
| 290 | if e = p.limitsOK(); e != nil { |
| 291 | return nil, e |
| 292 | } |
| 293 | if genParent { |
| 294 | if e = delegateStrict(p.dir); e != nil { |
| 295 | return nil, e |
| 296 | } |
| 297 | p.genParent, p.tier = true, limit |
| 298 | } |
| 299 | if e = p.identityOK(); e != nil { |
| 300 | return nil, e |
| 301 | } |
| 302 | return p, nil |
| 303 | } |
| 304 | |
| 305 | // Start consumes one attempt even if exec or containment fails. The caller |
| 306 | // supplies the complete command, including its approved executable descriptor. |
| 307 | func (p *Prepared) Start(cmd *exec.Cmd) error { |
| 308 | if p == nil || p.strictPreparedState == nil { |
| 309 | return ErrStrictState |
| 310 | } |
| 311 | if p.parent != nil { |
| 312 | p.parent.mu.Lock() |
| 313 | defer p.parent.mu.Unlock() |
| 314 | if p.parent.closed || p.parent.empty || p.parent.draining.Load() { |
| 315 | return ErrStrictState |
| 316 | } |
| 317 | if err := p.parent.identityOK(); err != nil { |
| 318 | return err |
| 319 | } |
| 320 | if err := p.parent.limitsOK(); err != nil { |
| 321 | return err |
| 322 | } |
| 323 | } |
| 324 | p.mu.Lock() |
| 325 | defer p.mu.Unlock() |
| 326 | if p.genParent { |
| 327 | // A generation parent holds no process (zero-downtime spec §7.3). |
| 328 | return ErrStrictState |
| 329 | } |
| 330 | if p.closed || p.attempted || p.empty || p.draining.Load() { |
| 331 | return ErrStrictState |
| 332 | } |
| 333 | p.attempted = true |
| 334 | if cmd == nil || cmd.Process != nil || cmd.ProcessState != nil { |
| 335 | return ErrStrictState |
| 336 | } |
| 337 | if e := p.identityOK(); e != nil { |
| 338 | return e |
| 339 | } |
| 340 | if e := p.limitsOK(); e != nil { |
| 341 | return e |
| 342 | } |
| 343 | if populated, e := strictPopulation(p.events); e != nil || populated { |
| 344 | return ErrStrictRefused |
| 345 | } |
| 346 | setInto(cmd, int(p.dir.Fd())) |
| 347 | cmd.SysProcAttr.Pdeathsig = syscall.SIGKILL |
| 348 | if e := p.start(cmd); e != nil { |
| 349 | if shouldLatch(e) { |
| 350 | return ErrStrictUnsupported |
| 351 | } |
| 352 | return ErrStrictRefused |
| 353 | } |
| 354 | if cmd.Process == nil { |
| 355 | return ErrStrictRefused |
| 356 | } |
| 357 | return nil |
| 358 | } |
| 359 | |
| 360 | // Populated checks the retained group after the direct child has been reaped. |
| 361 | func (p *Prepared) Populated() (bool, error) { |
| 362 | if p == nil || p.strictPreparedState == nil { |
| 363 | return false, ErrStrictState |
| 364 | } |
| 365 | p.mu.Lock() |
| 366 | defer p.mu.Unlock() |
| 367 | if e := p.identityOK(); e != nil { |
| 368 | return false, e |
| 369 | } |
| 370 | return strictPopulation(p.events) |
| 371 | } |
| 372 | |
| 373 | // KillAndEmpty proves kernel population is zero before the caller's deadline. |
| 374 | // Failure retains the pinned cleanup handle and permits another bounded attempt. |
| 375 | func (p *Prepared) KillAndEmpty(ctx context.Context) error { |
| 376 | if p == nil || p.strictPreparedState == nil { |
| 377 | return ErrStrictState |
| 378 | } |
| 379 | p.draining.Store(true) |
| 380 | if ctx == nil { |
| 381 | return ErrStrictDeadline |
| 382 | } |
| 383 | deadline, ok := ctx.Deadline() |
| 384 | if !ok || ctx.Err() != nil || !time.Now().Before(deadline) { |
| 385 | return ErrStrictDeadline |
| 386 | } |
| 387 | lockTick := time.NewTicker(time.Millisecond) |
| 388 | defer lockTick.Stop() |
| 389 | for !p.mu.TryLock() { |
| 390 | select { |
| 391 | case <-ctx.Done(): |
| 392 | return ErrStrictDeadline |
| 393 | case <-lockTick.C: |
| 394 | } |
| 395 | } |
| 396 | defer p.mu.Unlock() |
| 397 | if p.closed || p.empty { |
| 398 | return ErrStrictState |
| 399 | } |
| 400 | if ctx.Err() != nil { |
| 401 | return ErrStrictDeadline |
| 402 | } |
| 403 | if e := p.identityOK(); e != nil { |
| 404 | return e |
| 405 | } |
| 406 | p.attempted = true |
| 407 | // Marked first, so no later start reuses this directory (amadan #272, |
| 408 | // recreate_linux.go). |
| 409 | markKilled(p.identity) |
| 410 | if e := strictWrite(p.kill, "1"); e != nil { |
| 411 | return ErrStrictRefused |
| 412 | } |
| 413 | timer := time.NewTicker(5 * time.Millisecond) |
| 414 | defer timer.Stop() |
| 415 | for { |
| 416 | if ctx.Err() != nil || !time.Now().Before(deadline) { |
| 417 | return ErrStrictDeadline |
| 418 | } |
| 419 | if e := p.identityOK(); e != nil { |
| 420 | return e |
| 421 | } |
| 422 | populated, e := strictPopulation(p.events) |
| 423 | if e != nil { |
| 424 | return ErrStrictRefused |
| 425 | } |
| 426 | if !populated { |
| 427 | if ctx.Err() != nil || !time.Now().Before(deadline) { |
| 428 | return ErrStrictDeadline |
| 429 | } |
| 430 | p.empty = true |
| 431 | return nil |
| 432 | } |
| 433 | select { |
| 434 | case <-ctx.Done(): |
| 435 | return ErrStrictDeadline |
| 436 | case <-timer.C: |
| 437 | } |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | // Close never removes a path. A populated or unverifiable group retains its |
| 442 | // descriptors and reservation so the same handle can still attempt cleanup. |
| 443 | func (p *Prepared) Close() error { |
| 444 | if p == nil || p.strictPreparedState == nil { |
| 445 | return ErrStrictState |
| 446 | } |
| 447 | p.mu.Lock() |
| 448 | defer p.mu.Unlock() |
| 449 | p.draining.Store(true) |
| 450 | if p.closed { |
| 451 | return ErrStrictState |
| 452 | } |
| 453 | if e := p.identityOK(); e != nil { |
| 454 | return e |
| 455 | } |
| 456 | if populated, e := strictPopulation(p.events); e != nil || populated { |
| 457 | return ErrStrictRefused |
| 458 | } |
| 459 | p.closed = true |
| 460 | p.closeFiles() |
| 461 | strictReservations.Lock() |
| 462 | delete(strictReservations.held, p.identity) |
| 463 | strictReservations.Unlock() |
| 464 | return nil |
| 465 | } |
| 466 | |
amadan