| 1 | package fileutil |
| 2 | |
| 3 | import ( |
| 4 | "fmt" |
| 5 | "os" |
| 6 | "path/filepath" |
| 7 | "time" |
| 8 | ) |
| 9 | |
| 10 | var ( |
| 11 | maxReplaceRetries = 12 |
| 12 | replaceRetryBase = 20 * time.Millisecond |
| 13 | |
| 14 | // renameFile is a test seam: the two rename failure classes ReplaceFile |
| 15 | // distinguishes (transient lock vs cross-device) cannot be provoked |
| 16 | // portably on a real filesystem. |
| 17 | renameFile = os.Rename |
| 18 | ) |
| 19 | |
| 20 | // AtomicWriteFile writes data to a sibling temporary file, fsyncs it, then |
| 21 | // publishes it via ReplaceFile. On filesystems that support replacement rename, |
| 22 | // readers see either the old file or the complete new file. ReplaceFile retains |
| 23 | // its compatibility copy fallback for Windows filter drivers that reject a |
| 24 | // same-directory rename as cross-device; callers that cannot tolerate that |
| 25 | // non-atomic fallback must use AtomicWriteFileStrict. |
| 26 | func AtomicWriteFile(path string, data []byte, perm os.FileMode) error { |
| 27 | return atomicWriteFile(path, data, perm, true) |
| 28 | } |
| 29 | |
| 30 | // AtomicWriteFileStrict publishes data only through an atomic rename. Unlike |
| 31 | // AtomicWriteFile, a cross-device/filter-driver error is returned without ever |
| 32 | // truncating path. Use it for commit pointers whose corruption would make the |
| 33 | // surrounding state impossible to recover automatically. |
| 34 | func AtomicWriteFileStrict(path string, data []byte, perm os.FileMode) error { |
| 35 | return atomicWriteFile(path, data, perm, false) |
| 36 | } |
| 37 | |
| 38 | func atomicWriteFile(path string, data []byte, perm os.FileMode, allowCrossDeviceCopy bool) error { |
| 39 | tmpPath, err := writeAtomicTemp(path, data, perm) |
| 40 | if err != nil { |
| 41 | return err |
| 42 | } |
| 43 | if err := replaceFile(tmpPath, path, allowCrossDeviceCopy); err != nil { |
| 44 | os.Remove(tmpPath) |
| 45 | return err |
| 46 | } |
| 47 | return nil |
| 48 | } |
| 49 | |
| 50 | // AtomicCreateFile publishes a complete file only when path is still absent. |
| 51 | // It is the non-overwriting counterpart to AtomicWriteFile: a concurrent writer |
| 52 | // that creates path wins, and its file is never replaced. |
| 53 | func AtomicCreateFile(path string, data []byte, perm os.FileMode) error { |
| 54 | tmpPath, err := writeAtomicTemp(path, data, perm) |
| 55 | if err != nil { |
| 56 | return err |
| 57 | } |
| 58 | defer os.Remove(tmpPath) |
| 59 | if err := os.Link(tmpPath, path); err != nil { |
| 60 | return fmt.Errorf("publish new file %s: %w", path, err) |
| 61 | } |
| 62 | return nil |
| 63 | } |
| 64 | |
| 65 | func writeAtomicTemp(path string, data []byte, perm os.FileMode) (string, error) { |
| 66 | dir := filepath.Dir(path) |
| 67 | dirPerm := os.FileMode(0o755) |
| 68 | if perm&0o077 == 0 { |
| 69 | dirPerm = 0o700 |
| 70 | } |
| 71 | if err := os.MkdirAll(dir, dirPerm); err != nil { |
| 72 | return "", fmt.Errorf("create dir for %s: %w", path, err) |
| 73 | } |
| 74 | tmp, err := os.CreateTemp(dir, ".atomic-*.tmp") |
| 75 | if err != nil { |
| 76 | return "", fmt.Errorf("create tmp for %s: %w", path, err) |
| 77 | } |
| 78 | tmpPath := tmp.Name() |
| 79 | if _, err := tmp.Write(data); err != nil { |
| 80 | tmp.Close() |
| 81 | os.Remove(tmpPath) |
| 82 | return "", fmt.Errorf("write tmp for %s: %w", path, err) |
| 83 | } |
| 84 | if err := tmp.Sync(); err != nil { |
| 85 | tmp.Close() |
| 86 | os.Remove(tmpPath) |
| 87 | return "", fmt.Errorf("fsync tmp for %s: %w", path, err) |
| 88 | } |
| 89 | // Chmod the still-open handle, before Close, so there is no window between |
| 90 | // close and a path-based chmod for another process (Windows AV / search |
| 91 | // indexer) to grab or move the tmp and make the chmod fail with "file not |
| 92 | // found". CreateTemp makes a 0600 file, so this only widens when perm asks. |
| 93 | if err := tmp.Chmod(perm); err != nil { |
| 94 | tmp.Close() |
| 95 | os.Remove(tmpPath) |
| 96 | return "", fmt.Errorf("chmod tmp for %s: %w", path, err) |
| 97 | } |
| 98 | if err := tmp.Close(); err != nil { |
| 99 | os.Remove(tmpPath) |
| 100 | return "", fmt.Errorf("close tmp for %s: %w", path, err) |
| 101 | } |
| 102 | return tmpPath, nil |
| 103 | } |
| 104 | |
| 105 | // ReplaceFile renames tmp onto dest, publishing the new content atomically: a |
| 106 | // reader concurrent with the replace sees either the old file or the complete |
| 107 | // new one. The rename can fail in two ways, and they are handled differently: |
| 108 | // |
| 109 | // - A transient lock on dest (antivirus, the search indexer, a concurrent |
| 110 | // reader without delete sharing) fails the rename for a few hundred ms. |
| 111 | // The rename is retried with backoff, and the last error is returned if |
| 112 | // the lock never clears. The failure is loud on purpose: falling back to |
| 113 | // an in-place copy here would truncate dest first, letting a racing |
| 114 | // reader observe an empty or half-written file — exactly the torn state |
| 115 | // AtomicWriteFile promises its callers (session leases, credentials, |
| 116 | // plugin state) can never happen. |
| 117 | // - Windows encryption-software filter drivers report a cross-device link |
| 118 | // (ERROR_NOT_SAME_DEVICE / EXDEV) even for a same-dir rename (#2696), and |
| 119 | // every retry fails identically. Only this class falls back to the |
| 120 | // non-atomic copy, and immediately — retrying a structurally impossible |
| 121 | // rename would only delay it. Torn reads remain possible in that degraded |
| 122 | // mode; it is the only way to write at all on such hosts, and |
| 123 | // rename-capable filesystems never take it. |
| 124 | // |
| 125 | // A missing tmp means the write itself failed and no retry can help. |
| 126 | func ReplaceFile(tmp, dest string) error { |
| 127 | return replaceFile(tmp, dest, true) |
| 128 | } |
| 129 | |
| 130 | func replaceFile(tmp, dest string, allowCrossDeviceCopy bool) error { |
| 131 | var err error |
| 132 | for attempt := 0; ; attempt++ { |
| 133 | if err = renameFile(tmp, dest); err == nil { |
| 134 | return nil |
| 135 | } |
| 136 | if renameCrossesDevice(err) { |
| 137 | if !allowCrossDeviceCopy { |
| 138 | return err |
| 139 | } |
| 140 | if copyOnto(tmp, dest) == nil { |
| 141 | return nil |
| 142 | } |
| 143 | return err |
| 144 | } |
| 145 | if attempt >= maxReplaceRetries || !fileExists(tmp) { |
| 146 | return err |
| 147 | } |
| 148 | time.Sleep(time.Duration(attempt+1) * replaceRetryBase) |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | func fileExists(path string) bool { |
| 153 | _, err := os.Stat(path) |
| 154 | return err == nil |
| 155 | } |
| 156 | |
| 157 | // copyOnto is the non-atomic last resort for hosts whose filesystem cannot |
| 158 | // rename tmp onto dest at all (see ReplaceFile). It truncates dest in place, |
| 159 | // so a concurrent reader can observe an empty or half-written file — it must |
| 160 | // never run for failures a retry could clear. |
| 161 | func copyOnto(tmp, dest string) error { |
| 162 | info, err := os.Stat(tmp) |
| 163 | if err != nil { |
| 164 | return err |
| 165 | } |
| 166 | data, err := os.ReadFile(tmp) |
| 167 | if err != nil { |
| 168 | return err |
| 169 | } |
| 170 | if err := os.WriteFile(dest, data, info.Mode().Perm()); err != nil { |
| 171 | return err |
| 172 | } |
| 173 | // WriteFile keeps an existing dest's mode, so re-apply tmp's mode to match |
| 174 | // what the rename would have done (a 0600 config tmp must not widen to 0644). |
| 175 | _ = os.Chmod(dest, info.Mode().Perm()) |
| 176 | _ = os.Remove(tmp) |
| 177 | return nil |
| 178 | } |
| 179 |