| 1 | package archive |
| 2 | |
| 3 | import ( |
| 4 | "archive/tar" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "io" |
| 8 | "io/fs" |
| 9 | "os" |
| 10 | "path/filepath" |
| 11 | "runtime" |
| 12 | "sort" |
| 13 | "strings" |
| 14 | "sync" |
| 15 | ) |
| 16 | |
| 17 | // Extract writes an opened archive into filesDir and envPath, refusing |
| 18 | // anything a tar could smuggle: absolute paths, "..", the protocol marker, |
| 19 | // links, devices and FIFOs, two entries naming one destination, more |
| 20 | // entries than MaxEntries, content over Cap and a decoded stream over |
| 21 | // MaxDecoded. Every directory must appear before anything inside it (Build's |
| 22 | // sorted order always does), so each one carries its archived mode. |
| 23 | // |
| 24 | // Nothing reaches the destination unless the whole archive is accepted: |
| 25 | // entries are written into a private staging directory beside filesDir |
| 26 | // (and env into a temporary file beside envPath), made durable, and moved |
| 27 | // into place only after the stream has been read to its end. filesDir must be absent |
| 28 | // or an empty directory (ErrNotEmpty otherwise; a non-empty directory |
| 29 | // would keep files the archive does not name), and its parent and |
| 30 | // envPath's parent must exist. envPath is replaced when the archive |
| 31 | // carries env and removed when it does not, so afterwards PlainDigest of |
| 32 | // the pair equals the archive's. The files move first and env second; a |
| 33 | // failure between the two renames leaves the files in place and is |
| 34 | // returned. |
| 35 | func Extract(r io.Reader, filesDir, envPath string) error { |
| 36 | if filesDir == "" { |
| 37 | return errors.New("archive: no files directory named") |
| 38 | } |
| 39 | filesDir = filepath.Clean(filesDir) |
| 40 | if envPath != "" { |
| 41 | envPath = filepath.Clean(envPath) |
| 42 | if rel, err := filepath.Rel(filesDir, envPath); err == nil && filepath.IsLocal(rel) { |
| 43 | return fmt.Errorf("%w: the env file %s is inside the files directory", ErrBadPath, envPath) |
| 44 | } |
| 45 | } |
| 46 | rootMode, existed, err := checkDestination(filesDir) |
| 47 | if err != nil { |
| 48 | return err |
| 49 | } |
| 50 | stage, err := os.MkdirTemp(filepath.Dir(filesDir), "."+filepath.Base(filesDir)+".extract-") |
| 51 | if err != nil { |
| 52 | return err |
| 53 | } |
| 54 | x := &extraction{stage: stage, envPath: envPath, files: newSyncPool(syncWorkers)} |
| 55 | defer x.cleanup() |
| 56 | if x.root, err = os.OpenRoot(stage); err != nil { |
| 57 | return err |
| 58 | } |
| 59 | if err := x.read(r); err != nil { |
| 60 | return err |
| 61 | } |
| 62 | return x.commit(filesDir, rootMode, existed) |
| 63 | } |
| 64 | |
| 65 | // checkDestination requires filesDir to be absent or an empty directory |
| 66 | // (not a symlink to one) and returns the mode the extracted root takes: |
| 67 | // the existing directory's, else 0755. |
| 68 | func checkDestination(filesDir string) (fs.FileMode, bool, error) { |
| 69 | info, err := os.Lstat(filesDir) |
| 70 | if errors.Is(err, fs.ErrNotExist) { |
| 71 | return 0o755, false, nil |
| 72 | } |
| 73 | if err != nil { |
| 74 | return 0, false, err |
| 75 | } |
| 76 | if !info.IsDir() { |
| 77 | return 0, false, fmt.Errorf("%w: the destination %s is not a directory", ErrBadPath, filesDir) |
| 78 | } |
| 79 | f, err := os.Open(filesDir) |
| 80 | if err != nil { |
| 81 | return 0, false, err |
| 82 | } |
| 83 | names, err := f.Readdirnames(1) |
| 84 | f.Close() |
| 85 | if len(names) > 0 { |
| 86 | return 0, false, fmt.Errorf("%w: %s", ErrNotEmpty, filesDir) |
| 87 | } |
| 88 | if err != nil && err != io.EOF { |
| 89 | return 0, false, err |
| 90 | } |
| 91 | return info.Mode().Perm(), true, nil |
| 92 | } |
| 93 | |
| 94 | type modeFix struct { |
| 95 | rel string |
| 96 | mode fs.FileMode |
| 97 | } |
| 98 | |
| 99 | // extraction is one Extract in progress. |
| 100 | type extraction struct { |
| 101 | stage string // the staging directory, removed unless committed |
| 102 | root *os.Root // stage, for every write |
| 103 | envPath string |
| 104 | envTmp string // the staged env file, if the archive carried one |
| 105 | dirs []modeFix |
| 106 | files *syncPool // fsyncs and closes each staged file as it is written |
| 107 | } |
| 108 | |
| 109 | func (x *extraction) cleanup() { |
| 110 | x.files.wait() |
| 111 | if x.root != nil { |
| 112 | x.root.Close() |
| 113 | } |
| 114 | if x.stage != "" { |
| 115 | removeStage(x.stage) |
| 116 | } |
| 117 | if x.envTmp != "" { |
| 118 | os.Remove(x.envTmp) |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | // removeStage removes a staging directory whose directories may already |
| 123 | // carry their final, restrictive modes. It holds only what Extract |
| 124 | // created: regular files and directories. |
| 125 | func removeStage(stage string) { |
| 126 | filepath.WalkDir(stage, func(p string, d fs.DirEntry, err error) error { |
| 127 | if err == nil && d.IsDir() { |
| 128 | os.Chmod(p, 0o700) |
| 129 | } |
| 130 | return nil |
| 131 | }) |
| 132 | os.RemoveAll(stage) |
| 133 | } |
| 134 | |
| 135 | // read writes every entry into the staging area and reads the stream to |
| 136 | // its end. |
| 137 | func (x *extraction) read(r io.Reader) error { |
| 138 | zr, err := newDecoder(r) |
| 139 | if err != nil { |
| 140 | return err |
| 141 | } |
| 142 | defer zr.Close() |
| 143 | dec := &boundedReader{r: zr, left: decodedLimit} |
| 144 | tr := tar.NewReader(dec) |
| 145 | rules := newStreamRules() |
| 146 | for { |
| 147 | hdr, err := tr.Next() |
| 148 | if err == io.EOF { |
| 149 | break |
| 150 | } |
| 151 | if err != nil { |
| 152 | return err |
| 153 | } |
| 154 | _, rel, isEnv, err := rules.admit(hdr) |
| 155 | if err != nil { |
| 156 | return err |
| 157 | } |
| 158 | mode := fs.FileMode(hdr.Mode).Perm() |
| 159 | if hdr.Typeflag == tar.TypeDir { |
| 160 | // Created writable so children can land; the archived mode is |
| 161 | // set in commit, before the directory is synced. |
| 162 | if err := x.root.Mkdir(rel, 0o700); err != nil { |
| 163 | return err |
| 164 | } |
| 165 | x.dirs = append(x.dirs, modeFix{rel, mode}) |
| 166 | continue |
| 167 | } |
| 168 | if isEnv { |
| 169 | err = x.writeEnv(tr, hdr.Size, mode) |
| 170 | } else { |
| 171 | err = x.writeFile(tr, rel, hdr.Size, mode) |
| 172 | } |
| 173 | if err != nil { |
| 174 | return err |
| 175 | } |
| 176 | } |
| 177 | if _, err := io.Copy(io.Discard, dec); err != nil { |
| 178 | return err |
| 179 | } |
| 180 | return requireEOF(r) |
| 181 | } |
| 182 | |
| 183 | func (x *extraction) writeFile(src io.Reader, rel string, size int64, mode fs.FileMode) error { |
| 184 | // O_EXCL: a name that already exists in the fresh staging tree (on a |
| 185 | // case-insensitive filesystem, "A" after "a") is refused, not merged. |
| 186 | f, err := x.root.OpenFile(rel, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o600) |
| 187 | if err != nil { |
| 188 | return err |
| 189 | } |
| 190 | return x.finishFile(f, src, size, mode) |
| 191 | } |
| 192 | |
| 193 | func (x *extraction) writeEnv(src io.Reader, size int64, mode fs.FileMode) error { |
| 194 | if x.envPath == "" { |
| 195 | return fmt.Errorf("%w: the archive carries env and no env path was given", ErrBadPath) |
| 196 | } |
| 197 | f, err := os.CreateTemp(filepath.Dir(x.envPath), "."+filepath.Base(x.envPath)+".extract-") |
| 198 | if err != nil { |
| 199 | return err |
| 200 | } |
| 201 | x.envTmp = f.Name() |
| 202 | return x.finishFile(f, src, size, mode) |
| 203 | } |
| 204 | |
| 205 | // finishFile copies exactly size bytes and sets the archived mode (fchmod, |
| 206 | // so the umask does not apply; §8.1), then hands the still-open file to |
| 207 | // the sync pool, which fsyncs and closes it. Syncing the open descriptor |
| 208 | // means a mode that refuses every open (0000) cannot stop the sync, and |
| 209 | // the pool overlaps the syncs with the rest of the extraction. |
| 210 | func (x *extraction) finishFile(f *os.File, src io.Reader, size int64, mode fs.FileMode) error { |
| 211 | if _, err := io.CopyN(f, src, size); err != nil { |
| 212 | f.Close() |
| 213 | return err |
| 214 | } |
| 215 | if err := f.Chmod(mode); err != nil { |
| 216 | f.Close() |
| 217 | return err |
| 218 | } |
| 219 | x.files.submit(func() error { return syncAndClose(f, fsyncFile) }) |
| 220 | return nil |
| 221 | } |
| 222 | |
| 223 | // commit makes the staged tree durable with its archived modes, then |
| 224 | // publishes it. |
| 225 | // |
| 226 | // Crash ordering. Every staged file is fsynced with its archived mode |
| 227 | // (finishFile) and every staged directory is fsynced after its archived |
| 228 | // mode is set, deepest level first, then the staged root with its final |
| 229 | // mode: all before the rename that publishes the tree. The rename is made |
| 230 | // durable by fsyncing the destination's parent, then env is published the |
| 231 | // same way. So a crash at any point leaves either no published tree (a |
| 232 | // hidden .extract- directory to sweep) or a complete one carrying the |
| 233 | // archived modes, never a published tree whose directories still have |
| 234 | // their staging modes. The landing writes its marker only after Extract |
| 235 | // returns nil. |
| 236 | // |
| 237 | // Cost. Scoped fsyncs only, never a filesystem-wide flush: other tenants' |
| 238 | // dirty data on the box is not this landing's to wait for. They run on |
| 239 | // syncWorkers workers, since concurrent fsyncs on ext4 and xfs share |
| 240 | // journal commits. The cut budget prices this time under its Archive and |
| 241 | // RestoreStaged terms (moves.Estimate). |
| 242 | func (x *extraction) commit(filesDir string, rootMode fs.FileMode, existed bool) error { |
| 243 | if err := x.files.wait(); err != nil { |
| 244 | return err |
| 245 | } |
| 246 | if err := x.settleDirs(); err != nil { |
| 247 | return err |
| 248 | } |
| 249 | if err := settleDir(func() (*os.File, error) { return os.Open(x.stage) }, rootMode); err != nil { |
| 250 | return err |
| 251 | } |
| 252 | if existed { |
| 253 | // os.Remove refuses a directory that is no longer empty. |
| 254 | if err := os.Remove(filesDir); err != nil { |
| 255 | return err |
| 256 | } |
| 257 | } |
| 258 | if err := os.Rename(x.stage, filesDir); err != nil { |
| 259 | return err |
| 260 | } |
| 261 | x.stage = "" |
| 262 | if err := syncDir(filepath.Dir(filesDir)); err != nil { |
| 263 | return err |
| 264 | } |
| 265 | if x.envPath == "" { |
| 266 | return nil |
| 267 | } |
| 268 | if x.envTmp == "" { |
| 269 | if err := os.Remove(x.envPath); err != nil && !errors.Is(err, fs.ErrNotExist) { |
| 270 | return err |
| 271 | } |
| 272 | } else { |
| 273 | if err := os.Rename(x.envTmp, x.envPath); err != nil { |
| 274 | return err |
| 275 | } |
| 276 | x.envTmp = "" |
| 277 | } |
| 278 | return syncDir(filepath.Dir(x.envPath)) |
| 279 | } |
| 280 | |
| 281 | // settleDirs sets every staged directory's archived mode through an open |
| 282 | // descriptor and fsyncs it, one depth level at a time, deepest first: a |
| 283 | // directory is opened by a path through its ancestors, which are still |
| 284 | // 0700 until their own level, so a restrictive mode (0300, 0000) never |
| 285 | // blocks a deeper directory, and the descriptor syncs whatever the mode. |
| 286 | func (x *extraction) settleDirs() error { |
| 287 | levels := map[int][]modeFix{} |
| 288 | for _, d := range x.dirs { |
| 289 | depth := strings.Count(d.rel, "/") |
| 290 | levels[depth] = append(levels[depth], d) |
| 291 | } |
| 292 | depths := make([]int, 0, len(levels)) |
| 293 | for depth := range levels { |
| 294 | depths = append(depths, depth) |
| 295 | } |
| 296 | sort.Sort(sort.Reverse(sort.IntSlice(depths))) |
| 297 | for _, depth := range depths { |
| 298 | p := newSyncPool(syncWorkers) |
| 299 | for _, d := range levels[depth] { |
| 300 | p.submit(func() error { |
| 301 | return settleDir(func() (*os.File, error) { return x.root.Open(d.rel) }, d.mode) |
| 302 | }) |
| 303 | } |
| 304 | if err := p.wait(); err != nil { |
| 305 | return err |
| 306 | } |
| 307 | } |
| 308 | return nil |
| 309 | } |
| 310 | |
| 311 | func settleDir(open func() (*os.File, error), mode fs.FileMode) error { |
| 312 | f, err := open() |
| 313 | if err != nil { |
| 314 | return err |
| 315 | } |
| 316 | if err := f.Chmod(mode); err != nil { |
| 317 | f.Close() |
| 318 | return err |
| 319 | } |
| 320 | return syncAndClose(f, syncDirFile) |
| 321 | } |
| 322 | |
| 323 | // fsyncFile is the one fsync every staged file and directory goes |
| 324 | // through. A variable so a test can see each one and the mode it carried. |
| 325 | var fsyncFile = func(f *os.File) error { return f.Sync() } |
| 326 | |
| 327 | // syncDirFile fsyncs a directory. Windows cannot sync a directory handle; |
| 328 | // the platform's boxes are Linux, and this only keeps that build whole. |
| 329 | func syncDirFile(f *os.File) error { |
| 330 | if runtime.GOOS == "windows" { |
| 331 | return nil |
| 332 | } |
| 333 | return fsyncFile(f) |
| 334 | } |
| 335 | |
| 336 | func syncAndClose(f *os.File, sync func(*os.File) error) error { |
| 337 | err := sync(f) |
| 338 | if cerr := f.Close(); err == nil { |
| 339 | err = cerr |
| 340 | } |
| 341 | return err |
| 342 | } |
| 343 | |
| 344 | // syncDir makes a rename in dir durable. |
| 345 | func syncDir(dir string) error { |
| 346 | d, err := os.Open(dir) |
| 347 | if err != nil { |
| 348 | return err |
| 349 | } |
| 350 | return syncAndClose(d, syncDirFile) |
| 351 | } |
| 352 | |
| 353 | // syncWorkers is how many fsyncs run at once (defaultSyncWorkers, per |
| 354 | // platform). A variable so the timing test can compare serial and parallel. |
| 355 | var syncWorkers = defaultSyncWorkers |
| 356 | |
| 357 | // syncPool runs jobs on a bounded set of workers and keeps the first |
| 358 | // error. At most workers jobs run and workers more wait, so the staged |
| 359 | // files held open for syncing stay bounded. |
| 360 | type syncPool struct { |
| 361 | jobs chan func() error |
| 362 | wg sync.WaitGroup |
| 363 | once sync.Once |
| 364 | mu sync.Mutex |
| 365 | err error |
| 366 | } |
| 367 | |
| 368 | func newSyncPool(workers int) *syncPool { |
| 369 | workers = max(workers, 1) |
| 370 | p := &syncPool{jobs: make(chan func() error, workers)} |
| 371 | p.wg.Add(workers) |
| 372 | for range workers { |
| 373 | go func() { |
| 374 | defer p.wg.Done() |
| 375 | for job := range p.jobs { |
| 376 | if err := job(); err != nil { |
| 377 | p.mu.Lock() |
| 378 | if p.err == nil { |
| 379 | p.err = err |
| 380 | } |
| 381 | p.mu.Unlock() |
| 382 | } |
| 383 | } |
| 384 | }() |
| 385 | } |
| 386 | return p |
| 387 | } |
| 388 | |
| 389 | func (p *syncPool) submit(job func() error) { p.jobs <- job } |
| 390 | |
| 391 | // wait closes the pool, waits for every job and returns the first error. |
| 392 | // Safe to call more than once. |
| 393 | func (p *syncPool) wait() error { |
| 394 | p.once.Do(func() { |
| 395 | close(p.jobs) |
| 396 | p.wg.Wait() |
| 397 | }) |
| 398 | p.mu.Lock() |
| 399 | defer p.mu.Unlock() |
| 400 | return p.err |
| 401 | } |
| 402 | |
amadan