| 1 | package fileutil |
| 2 | |
| 3 | import ( |
| 4 | "errors" |
| 5 | "fmt" |
| 6 | "os" |
| 7 | "path/filepath" |
| 8 | "runtime" |
| 9 | "syscall" |
| 10 | "time" |
| 11 | ) |
| 12 | |
| 13 | var ( |
| 14 | maxReplaceRetries = 12 |
| 15 | replaceRetryBase = 20 * time.Millisecond |
| 16 | |
| 17 | // renameFile is a test seam: the two rename failure classes ReplaceFile |
| 18 | // distinguishes (transient lock vs cross-device) cannot be provoked |
| 19 | // portably on a real filesystem. |
| 20 | renameFile = os.Rename |
| 21 | ) |
| 22 | |
| 23 | // CrashPoint, when non-nil, runs before every durable write/replace. Tests use |
| 24 | // it to inject a process-crash panic at persistence boundaries; production |
| 25 | // leaves it nil. |
| 26 | var CrashPoint func(op, path string) |
| 27 | |
| 28 | // Crash invokes the optional crash-consistency fault-injection hook. |
| 29 | func Crash(op, path string) { |
| 30 | if CrashPoint != nil { |
| 31 | CrashPoint(op, path) |
| 32 | } |
| 33 | } |
| 34 | |
| 35 | // AtomicWriteFile writes via temp + fsync + ReplaceFile. On rename-capable |
| 36 | // filesystems readers see only the old or complete new file. ReplaceFile may |
| 37 | // copy on Windows filter-driver EXDEV; callers that cannot tolerate that must |
| 38 | // use AtomicWriteFileStrict. |
| 39 | func AtomicWriteFile(path string, data []byte, perm os.FileMode) error { |
| 40 | return atomicWriteFile(path, data, perm, true) |
| 41 | } |
| 42 | |
| 43 | // AtomicWriteFileStrict publishes only via atomic rename (no EXDEV copy). |
| 44 | // After a successful rename it best-effort fsyncs the parent directory so the |
| 45 | // directory entry can survive power loss. A returned error always means the |
| 46 | // destination was not published; post-rename dir-sync problems are not errors |
| 47 | // (callers that roll back in-memory state on error would otherwise fork from |
| 48 | // the on-disk pointer). |
| 49 | func AtomicWriteFileStrict(path string, data []byte, perm os.FileMode) error { |
| 50 | return atomicWriteFile(path, data, perm, false) |
| 51 | } |
| 52 | |
| 53 | // syncParentDirFn is the post-publish parent-dir fsync implementation. |
| 54 | // Tests replace it via SetSyncParentDirForTest. |
| 55 | var syncParentDirFn = syncParentDir |
| 56 | |
| 57 | // SetSyncParentDirForTest replaces post-rename parent-dir fsync. Restore with |
| 58 | // the returned function. Production must leave the default in place. |
| 59 | func SetSyncParentDirForTest(fn func(path string) error) (restore func()) { |
| 60 | prev := syncParentDirFn |
| 61 | if fn == nil { |
| 62 | syncParentDirFn = syncParentDir |
| 63 | } else { |
| 64 | syncParentDirFn = fn |
| 65 | } |
| 66 | return func() { syncParentDirFn = prev } |
| 67 | } |
| 68 | |
| 69 | func atomicWriteFile(path string, data []byte, perm os.FileMode, allowCrossDeviceCopy bool) error { |
| 70 | Crash("atomic-write", path) |
| 71 | tmpPath, err := writeAtomicTemp(path, data, perm) |
| 72 | if err != nil { |
| 73 | return err |
| 74 | } |
| 75 | if err := replaceFile(tmpPath, path, allowCrossDeviceCopy); err != nil { |
| 76 | os.Remove(tmpPath) |
| 77 | return err |
| 78 | } |
| 79 | // Strict only: parent-dir fsync is power-loss durability after publish. |
| 80 | // Never surface failures here — rename already committed the new file. |
| 81 | if !allowCrossDeviceCopy { |
| 82 | _ = syncParentDirFn(path) |
| 83 | } |
| 84 | return nil |
| 85 | } |
| 86 | |
| 87 | // syncParentDir fsyncs path's parent after rename (including "."). Unsupported |
| 88 | // dir sync on Windows / some network FS is ignored. |
| 89 | func syncParentDir(path string) error { |
| 90 | dirPath := filepath.Dir(path) |
| 91 | if dirPath == "" { |
| 92 | dirPath = "." |
| 93 | } |
| 94 | f, err := os.Open(dirPath) |
| 95 | if err != nil { |
| 96 | return fmt.Errorf("open parent dir for fsync %s: %w", path, err) |
| 97 | } |
| 98 | defer f.Close() |
| 99 | if err := f.Sync(); err != nil { |
| 100 | if runtime.GOOS == "windows" || isDirSyncUnsupported(err) { |
| 101 | return nil |
| 102 | } |
| 103 | return fmt.Errorf("fsync parent dir for %s: %w", path, err) |
| 104 | } |
| 105 | return nil |
| 106 | } |
| 107 | |
| 108 | func isDirSyncUnsupported(err error) bool { |
| 109 | return errors.Is(err, syscall.EINVAL) || |
| 110 | errors.Is(err, syscall.ENOTSUP) || |
| 111 | errors.Is(err, syscall.ENOSYS) |
| 112 | } |
| 113 | |
| 114 | // AtomicCreateFile publishes a complete file only when path is still absent. |
| 115 | // It is the non-overwriting counterpart to AtomicWriteFile: a concurrent writer |
| 116 | // that creates path wins, and its file is never replaced. |
| 117 | func AtomicCreateFile(path string, data []byte, perm os.FileMode) error { |
| 118 | tmpPath, err := writeAtomicTemp(path, data, perm) |
| 119 | if err != nil { |
| 120 | return err |
| 121 | } |
| 122 | defer os.Remove(tmpPath) |
| 123 | if err := os.Link(tmpPath, path); err != nil { |
| 124 | return fmt.Errorf("publish new file %s: %w", path, err) |
| 125 | } |
| 126 | return nil |
| 127 | } |
| 128 | |
| 129 | // AtomicOverwriteFile replaces an existing file's contents atomically while |
| 130 | // keeping the two properties a bare rename drops: the file's current permission |
| 131 | // bits (an executable script must not come back 0644) and the symlink target |
| 132 | // (a link must be written through, not replaced by a regular file). defaultPerm |
| 133 | // applies only when path does not exist yet. |
| 134 | func AtomicOverwriteFile(path string, data []byte, defaultPerm os.FileMode) error { |
| 135 | return atomicOverwriteFile(path, data, defaultPerm, true) |
| 136 | } |
| 137 | |
| 138 | // AtomicOverwriteFileStrict preserves encoding callers' mode and symlink |
| 139 | // semantics without a non-atomic copy fallback on Windows filter drivers. |
| 140 | func AtomicOverwriteFileStrict(path string, data []byte, defaultPerm os.FileMode) error { |
| 141 | return atomicOverwriteFile(path, data, defaultPerm, false) |
| 142 | } |
| 143 | |
| 144 | func atomicOverwriteFile(path string, data []byte, defaultPerm os.FileMode, allowCopy bool) error { |
| 145 | target := path |
| 146 | if resolved, err := filepath.EvalSymlinks(path); err == nil { |
| 147 | target = resolved |
| 148 | } |
| 149 | perm := defaultPerm |
| 150 | if info, err := os.Stat(target); err == nil { |
| 151 | perm = info.Mode().Perm() |
| 152 | } |
| 153 | return atomicWriteFile(target, data, perm, allowCopy) |
| 154 | } |
| 155 | |
| 156 | func writeAtomicTemp(path string, data []byte, perm os.FileMode) (string, error) { |
| 157 | dir := filepath.Dir(path) |
| 158 | dirPerm := os.FileMode(0o755) |
| 159 | if perm&0o077 == 0 { |
| 160 | dirPerm = 0o700 |
| 161 | } |
| 162 | if err := os.MkdirAll(dir, dirPerm); err != nil { |
| 163 | return "", fmt.Errorf("create dir for %s: %w", path, err) |
| 164 | } |
| 165 | tmp, err := os.CreateTemp(dir, ".atomic-*.tmp") |
| 166 | if err != nil { |
| 167 | return "", fmt.Errorf("create tmp for %s: %w", path, err) |
| 168 | } |
| 169 | tmpPath := tmp.Name() |
| 170 | closed := false |
| 171 | closeTmp := func() error { |
| 172 | if closed { |
| 173 | return nil |
| 174 | } |
| 175 | closed = true |
| 176 | return tmp.Close() |
| 177 | } |
| 178 | keep := false |
| 179 | defer func() { |
| 180 | _ = closeTmp() |
| 181 | if !keep { |
| 182 | _ = os.Remove(tmpPath) |
| 183 | } |
| 184 | }() |
| 185 | if _, err := tmp.Write(data); err != nil { |
| 186 | return "", fmt.Errorf("write tmp for %s: %w", path, err) |
| 187 | } |
| 188 | if err := tmp.Sync(); err != nil { |
| 189 | return "", fmt.Errorf("fsync tmp for %s: %w", path, err) |
| 190 | } |
| 191 | // Chmod the still-open handle, before Close, so there is no window between |
| 192 | // close and a path-based chmod for another process (Windows AV / search |
| 193 | // indexer) to grab or move the tmp and make the chmod fail with "file not |
| 194 | // found". CreateTemp makes a 0600 file, so this only widens when perm asks. |
| 195 | if err := tmp.Chmod(perm); err != nil { |
| 196 | return "", fmt.Errorf("chmod tmp for %s: %w", path, err) |
| 197 | } |
| 198 | if err := closeTmp(); err != nil { |
| 199 | return "", fmt.Errorf("close tmp for %s: %w", path, err) |
| 200 | } |
| 201 | keep = true |
| 202 | return tmpPath, nil |
| 203 | } |
| 204 | |
| 205 | // ReplaceFile renames tmp onto dest, publishing the new content atomically: a |
| 206 | // reader concurrent with the replace sees either the old file or the complete |
| 207 | // new one. The rename can fail in two ways, and they are handled differently: |
| 208 | // |
| 209 | // - A transient lock on dest (antivirus, the search indexer, a concurrent |
| 210 | // reader without delete sharing) fails the rename for a few hundred ms. |
| 211 | // The rename is retried with backoff, and the last error is returned if |
| 212 | // the lock never clears. The failure is loud on purpose: falling back to |
| 213 | // an in-place copy here would truncate dest first, letting a racing |
| 214 | // reader observe an empty or half-written file — exactly the torn state |
| 215 | // AtomicWriteFile promises its callers (session leases, credentials, |
| 216 | // plugin state) can never happen. |
| 217 | // - Windows encryption-software filter drivers report a cross-device link |
| 218 | // (ERROR_NOT_SAME_DEVICE / EXDEV) even for a same-dir rename (#2696), and |
| 219 | // every retry fails identically. Only this class falls back to the |
| 220 | // non-atomic copy, and immediately — retrying a structurally impossible |
| 221 | // rename would only delay it. Torn reads remain possible in that degraded |
| 222 | // mode; it is the only way to write at all on such hosts, and |
| 223 | // rename-capable filesystems never take it. |
| 224 | // |
| 225 | // A missing tmp means the write itself failed and no retry can help. |
| 226 | func ReplaceFile(tmp, dest string) error { |
| 227 | Crash("replace", dest) |
| 228 | return replaceFile(tmp, dest, true) |
| 229 | } |
| 230 | |
| 231 | // ClaimRename renames src to dst for callers that use the rename itself as a |
| 232 | // claim: it retries the same transient locks ReplaceFile does, but never falls |
| 233 | // back to a copy, because a copy would let two claimants both succeed. A src |
| 234 | // that has disappeared ends the retries at once — that is the loser of a race, |
| 235 | // not a fault. |
| 236 | func ClaimRename(src, dst string) error { |
| 237 | return replaceFile(src, dst, false) |
| 238 | } |
| 239 | |
| 240 | func replaceFile(tmp, dest string, allowCrossDeviceCopy bool) error { |
| 241 | var err error |
| 242 | for attempt := 0; ; attempt++ { |
| 243 | if err = renameFile(tmp, dest); err == nil { |
| 244 | return nil |
| 245 | } |
| 246 | if renameCrossesDevice(err) { |
| 247 | if !allowCrossDeviceCopy { |
| 248 | return err |
| 249 | } |
| 250 | if copyOnto(tmp, dest) == nil { |
| 251 | return nil |
| 252 | } |
| 253 | return err |
| 254 | } |
| 255 | if attempt >= maxReplaceRetries || !fileExists(tmp) { |
| 256 | return err |
| 257 | } |
| 258 | time.Sleep(time.Duration(attempt+1) * replaceRetryBase) |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | func fileExists(path string) bool { |
| 263 | _, err := os.Stat(path) |
| 264 | return err == nil |
| 265 | } |
| 266 | |
| 267 | // copyOnto is the non-atomic last resort for hosts whose filesystem cannot |
| 268 | // rename tmp onto dest at all (see ReplaceFile). It truncates dest in place, |
| 269 | // so a concurrent reader can observe an empty or half-written file — it must |
| 270 | // never run for failures a retry could clear. |
| 271 | func copyOnto(tmp, dest string) error { |
| 272 | info, err := os.Stat(tmp) |
| 273 | if err != nil { |
| 274 | return err |
| 275 | } |
| 276 | data, err := os.ReadFile(tmp) |
| 277 | if err != nil { |
| 278 | return err |
| 279 | } |
| 280 | if err := os.WriteFile(dest, data, info.Mode().Perm()); err != nil { |
| 281 | return err |
| 282 | } |
| 283 | // WriteFile keeps an existing dest's mode, so re-apply tmp's mode to match |
| 284 | // what the rename would have done (a 0600 config tmp must not widen to 0644). |
| 285 | _ = os.Chmod(dest, info.Mode().Perm()) |
| 286 | _ = os.Remove(tmp) |
| 287 | return nil |
| 288 | } |
| 289 |