| 1 | //go:build linux |
| 2 | |
| 3 | package hostfiles |
| 4 | |
| 5 | import ( |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "io/fs" |
| 9 | "os" |
| 10 | "path/filepath" |
| 11 | "sync" |
| 12 | "sync/atomic" |
| 13 | "unsafe" |
| 14 | |
| 15 | "golang.org/x/sys/unix" |
| 16 | ) |
| 17 | |
| 18 | const watchMask = unix.IN_MODIFY | unix.IN_CLOSE_WRITE | unix.IN_MOVED_FROM | unix.IN_MOVED_TO | |
| 19 | unix.IN_CREATE | unix.IN_DELETE | unix.IN_ATTRIB |
| 20 | |
| 21 | type inotifyWatcher struct { |
| 22 | root string |
| 23 | skip func(string, bool) bool |
| 24 | f *os.File // the inotify descriptor, non-blocking, so Close unblocks Read |
| 25 | ch chan event |
| 26 | |
| 27 | mu sync.Mutex |
| 28 | wds map[int32]string // watch descriptor -> directory, relative to root ("" is root) |
| 29 | |
| 30 | // Coverage is tracked by counting failures, not with a flag. failures |
| 31 | // goes up whenever a directory may have been left unwatched; repaired |
| 32 | // is the count Repair had seen when it started the walk that |
| 33 | // succeeded. A failure that arrives while Repair is walking is |
| 34 | // therefore never erased by that Repair: the two differ again. |
| 35 | failures atomic.Int64 |
| 36 | repaired atomic.Int64 |
| 37 | dead atomic.Bool // the reader has stopped; no event will ever arrive |
| 38 | } |
| 39 | |
| 40 | // afterRepairWalk runs between Repair's walk and its bookkeeping. A test seam. |
| 41 | var afterRepairWalk = func() {} |
| 42 | |
| 43 | func newWatcher(root string, skip func(string, bool) bool) (watcher, error) { |
| 44 | fd, err := unix.InotifyInit1(unix.IN_CLOEXEC | unix.IN_NONBLOCK) |
| 45 | if err != nil { |
| 46 | return nil, fmt.Errorf("inotify init: %w", err) |
| 47 | } |
| 48 | w := &inotifyWatcher{root: root, skip: skip, f: os.NewFile(uintptr(fd), "inotify"), |
| 49 | ch: make(chan event, 4096), wds: map[int32]string{}} |
| 50 | if err := w.addTree(""); err != nil { |
| 51 | w.failures.Add(1) // partial coverage: Run polls and calls Repair |
| 52 | } |
| 53 | go w.loop() |
| 54 | return w, nil |
| 55 | } |
| 56 | |
| 57 | func (w *inotifyWatcher) Events() <-chan event { return w.ch } |
| 58 | func (w *inotifyWatcher) Degraded() bool { |
| 59 | return w.dead.Load() || w.failures.Load() != w.repaired.Load() |
| 60 | } |
| 61 | func (w *inotifyWatcher) Close() error { return w.f.Close() } |
| 62 | |
| 63 | // Repair re-adds a watch on every directory. InotifyAddWatch on a path |
| 64 | // already watched returns the same descriptor, so this is safe to repeat. |
| 65 | // A watcher whose reader has stopped cannot be repaired. |
| 66 | func (w *inotifyWatcher) Repair() error { |
| 67 | if w.dead.Load() { |
| 68 | return errors.New("the inotify reader has stopped") |
| 69 | } |
| 70 | seen := w.failures.Load() |
| 71 | if err := w.addTree(""); err != nil { |
| 72 | return err |
| 73 | } |
| 74 | afterRepairWalk() |
| 75 | // Only the failures this walk could have covered are cleared. One that |
| 76 | // arrived during the walk may concern a directory the walk had already |
| 77 | // passed, so it leaves the watcher degraded and Repair runs again. |
| 78 | w.repaired.Store(seen) |
| 79 | return nil |
| 80 | } |
| 81 | |
| 82 | // addTree watches rel and every directory under it. It keeps going after a |
| 83 | // failure, so one unreadable directory does not cost the rest their |
| 84 | // watches, and returns the first error. |
| 85 | func (w *inotifyWatcher) addTree(rel string) error { |
| 86 | var first error |
| 87 | start := filepath.Join(w.root, filepath.FromSlash(rel)) |
| 88 | err := filepath.WalkDir(start, func(full string, d fs.DirEntry, err error) error { |
| 89 | if err != nil { |
| 90 | if first == nil { |
| 91 | first = err |
| 92 | } |
| 93 | return nil |
| 94 | } |
| 95 | if !d.IsDir() { |
| 96 | return nil // WalkDir does not follow symlinks, so a linked directory is never watched |
| 97 | } |
| 98 | r, rerr := filepath.Rel(w.root, full) |
| 99 | if rerr != nil { |
| 100 | return nil |
| 101 | } |
| 102 | r = filepath.ToSlash(r) |
| 103 | if r == "." { |
| 104 | r = "" |
| 105 | } |
| 106 | if r != "" && w.skip != nil && w.skip(r, true) { |
| 107 | return filepath.SkipDir |
| 108 | } |
| 109 | wd, aerr := unix.InotifyAddWatch(int(w.f.Fd()), full, watchMask|unix.IN_DONT_FOLLOW|unix.IN_ONLYDIR) |
| 110 | if aerr != nil { |
| 111 | if first == nil { |
| 112 | first = fmt.Errorf("watch %s: %w", full, aerr) |
| 113 | } |
| 114 | return nil |
| 115 | } |
| 116 | w.mu.Lock() |
| 117 | w.wds[int32(wd)] = r |
| 118 | w.mu.Unlock() |
| 119 | return nil |
| 120 | }) |
| 121 | if err != nil && first == nil { |
| 122 | first = err |
| 123 | } |
| 124 | return first |
| 125 | } |
| 126 | |
| 127 | func (w *inotifyWatcher) send(ev event) { |
| 128 | select { |
| 129 | case w.ch <- ev: |
| 130 | default: |
| 131 | // Our own buffer is full. What was dropped is a file path (a new |
| 132 | // directory's watch is added before its event is sent), so coverage |
| 133 | // is intact; only a scan can find the dropped path now. |
| 134 | select { |
| 135 | case <-w.ch: |
| 136 | default: |
| 137 | } |
| 138 | select { |
| 139 | case w.ch <- event{rescan: true}: |
| 140 | default: |
| 141 | } |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | func (w *inotifyWatcher) loop() { |
| 146 | defer close(w.ch) |
| 147 | buf := make([]byte, 64<<10) |
| 148 | for { |
| 149 | n, err := w.f.Read(buf) |
| 150 | if err != nil { |
| 151 | // Closed by us, or broken. Either way nothing more will be |
| 152 | // read: the closed channel tells Run to poll from here on. |
| 153 | w.dead.Store(true) |
| 154 | return |
| 155 | } |
| 156 | for off := 0; off+unix.SizeofInotifyEvent <= n; { |
| 157 | raw := (*unix.InotifyEvent)(unsafe.Pointer(&buf[off])) |
| 158 | nameLen := int(raw.Len) |
| 159 | name := "" |
| 160 | if nameLen > 0 { |
| 161 | b := buf[off+unix.SizeofInotifyEvent : off+unix.SizeofInotifyEvent+nameLen] |
| 162 | for i, c := range b { |
| 163 | if c == 0 { |
| 164 | b = b[:i] |
| 165 | break |
| 166 | } |
| 167 | } |
| 168 | name = string(b) |
| 169 | } |
| 170 | off += unix.SizeofInotifyEvent + nameLen |
| 171 | w.handle(raw.Wd, raw.Mask, name) |
| 172 | } |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | func (w *inotifyWatcher) handle(wd int32, mask uint32, name string) { |
| 177 | if mask&unix.IN_Q_OVERFLOW != 0 { |
| 178 | // Events were dropped, and one of them may have been a new |
| 179 | // directory, which is then unwatched. Coverage is uncertain until |
| 180 | // Repair has walked the tree again. |
| 181 | w.failures.Add(1) |
| 182 | w.send(event{rescan: true}) |
| 183 | return |
| 184 | } |
| 185 | w.mu.Lock() |
| 186 | dir, known := w.wds[wd] |
| 187 | if mask&unix.IN_IGNORED != 0 { |
| 188 | delete(w.wds, wd) |
| 189 | } |
| 190 | w.mu.Unlock() |
| 191 | if !known || name == "" { |
| 192 | return |
| 193 | } |
| 194 | rel := name |
| 195 | if dir != "" { |
| 196 | rel = dir + "/" + name |
| 197 | } |
| 198 | if mask&unix.IN_ISDIR != 0 { |
| 199 | if mask&(unix.IN_CREATE|unix.IN_MOVED_TO) != 0 { |
| 200 | if w.skip != nil && w.skip(rel, true) { |
| 201 | return |
| 202 | } |
| 203 | if err := w.addTree(rel); err != nil { |
| 204 | w.failures.Add(1) |
| 205 | } |
| 206 | } |
| 207 | // Files may have appeared in the new directory before its watch |
| 208 | // did, and a directory moved away takes its files with it: only a |
| 209 | // scan finds either. |
| 210 | w.send(event{rescan: true}) |
| 211 | return |
| 212 | } |
| 213 | if w.skip != nil && w.skip(rel, false) { |
| 214 | return |
| 215 | } |
| 216 | w.send(event{rel: rel}) |
| 217 | } |
| 218 | |
amadan