khenaidoo | 59ce9dd | 2019-11-11 13:05:32 -0500 | [diff] [blame] | 1 | // Copyright 2015 The etcd Authors |
| 2 | // |
| 3 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | // you may not use this file except in compliance with the License. |
| 5 | // You may obtain a copy of the License at |
| 6 | // |
| 7 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | // |
| 9 | // Unless required by applicable law or agreed to in writing, software |
| 10 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | // See the License for the specific language governing permissions and |
| 13 | // limitations under the License. |
| 14 | |
| 15 | package mvcc |
| 16 | |
| 17 | import ( |
| 18 | "sync" |
| 19 | "time" |
| 20 | |
| 21 | "go.etcd.io/etcd/lease" |
| 22 | "go.etcd.io/etcd/mvcc/backend" |
| 23 | "go.etcd.io/etcd/mvcc/mvccpb" |
| 24 | "go.uber.org/zap" |
| 25 | ) |
| 26 | |
| 27 | // non-const so modifiable by tests |
| 28 | var ( |
| 29 | // chanBufLen is the length of the buffered chan |
| 30 | // for sending out watched events. |
| 31 | // TODO: find a good buf value. 1024 is just a random one that |
| 32 | // seems to be reasonable. |
| 33 | chanBufLen = 1024 |
| 34 | |
| 35 | // maxWatchersPerSync is the number of watchers to sync in a single batch |
| 36 | maxWatchersPerSync = 512 |
| 37 | ) |
| 38 | |
| 39 | type watchable interface { |
| 40 | watch(key, end []byte, startRev int64, id WatchID, ch chan<- WatchResponse, fcs ...FilterFunc) (*watcher, cancelFunc) |
| 41 | progress(w *watcher) |
| 42 | rev() int64 |
| 43 | } |
| 44 | |
| 45 | type watchableStore struct { |
| 46 | *store |
| 47 | |
| 48 | // mu protects watcher groups and batches. It should never be locked |
| 49 | // before locking store.mu to avoid deadlock. |
| 50 | mu sync.RWMutex |
| 51 | |
| 52 | // victims are watcher batches that were blocked on the watch channel |
| 53 | victims []watcherBatch |
| 54 | victimc chan struct{} |
| 55 | |
| 56 | // contains all unsynced watchers that needs to sync with events that have happened |
| 57 | unsynced watcherGroup |
| 58 | |
| 59 | // contains all synced watchers that are in sync with the progress of the store. |
| 60 | // The key of the map is the key that the watcher watches on. |
| 61 | synced watcherGroup |
| 62 | |
| 63 | stopc chan struct{} |
| 64 | wg sync.WaitGroup |
| 65 | } |
| 66 | |
| 67 | // cancelFunc updates unsynced and synced maps when running |
| 68 | // cancel operations. |
| 69 | type cancelFunc func() |
| 70 | |
| 71 | func New(lg *zap.Logger, b backend.Backend, le lease.Lessor, ig ConsistentIndexGetter, cfg StoreConfig) ConsistentWatchableKV { |
| 72 | return newWatchableStore(lg, b, le, ig, cfg) |
| 73 | } |
| 74 | |
| 75 | func newWatchableStore(lg *zap.Logger, b backend.Backend, le lease.Lessor, ig ConsistentIndexGetter, cfg StoreConfig) *watchableStore { |
| 76 | s := &watchableStore{ |
| 77 | store: NewStore(lg, b, le, ig, cfg), |
| 78 | victimc: make(chan struct{}, 1), |
| 79 | unsynced: newWatcherGroup(), |
| 80 | synced: newWatcherGroup(), |
| 81 | stopc: make(chan struct{}), |
| 82 | } |
| 83 | s.store.ReadView = &readView{s} |
| 84 | s.store.WriteView = &writeView{s} |
| 85 | if s.le != nil { |
| 86 | // use this store as the deleter so revokes trigger watch events |
| 87 | s.le.SetRangeDeleter(func() lease.TxnDelete { return s.Write() }) |
| 88 | } |
| 89 | s.wg.Add(2) |
| 90 | go s.syncWatchersLoop() |
| 91 | go s.syncVictimsLoop() |
| 92 | return s |
| 93 | } |
| 94 | |
| 95 | func (s *watchableStore) Close() error { |
| 96 | close(s.stopc) |
| 97 | s.wg.Wait() |
| 98 | return s.store.Close() |
| 99 | } |
| 100 | |
| 101 | func (s *watchableStore) NewWatchStream() WatchStream { |
| 102 | watchStreamGauge.Inc() |
| 103 | return &watchStream{ |
| 104 | watchable: s, |
| 105 | ch: make(chan WatchResponse, chanBufLen), |
| 106 | cancels: make(map[WatchID]cancelFunc), |
| 107 | watchers: make(map[WatchID]*watcher), |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | func (s *watchableStore) watch(key, end []byte, startRev int64, id WatchID, ch chan<- WatchResponse, fcs ...FilterFunc) (*watcher, cancelFunc) { |
| 112 | wa := &watcher{ |
| 113 | key: key, |
| 114 | end: end, |
| 115 | minRev: startRev, |
| 116 | id: id, |
| 117 | ch: ch, |
| 118 | fcs: fcs, |
| 119 | } |
| 120 | |
| 121 | s.mu.Lock() |
| 122 | s.revMu.RLock() |
| 123 | synced := startRev > s.store.currentRev || startRev == 0 |
| 124 | if synced { |
| 125 | wa.minRev = s.store.currentRev + 1 |
| 126 | if startRev > wa.minRev { |
| 127 | wa.minRev = startRev |
| 128 | } |
| 129 | } |
| 130 | if synced { |
| 131 | s.synced.add(wa) |
| 132 | } else { |
| 133 | slowWatcherGauge.Inc() |
| 134 | s.unsynced.add(wa) |
| 135 | } |
| 136 | s.revMu.RUnlock() |
| 137 | s.mu.Unlock() |
| 138 | |
| 139 | watcherGauge.Inc() |
| 140 | |
| 141 | return wa, func() { s.cancelWatcher(wa) } |
| 142 | } |
| 143 | |
| 144 | // cancelWatcher removes references of the watcher from the watchableStore |
| 145 | func (s *watchableStore) cancelWatcher(wa *watcher) { |
| 146 | for { |
| 147 | s.mu.Lock() |
| 148 | if s.unsynced.delete(wa) { |
| 149 | slowWatcherGauge.Dec() |
| 150 | break |
| 151 | } else if s.synced.delete(wa) { |
| 152 | break |
| 153 | } else if wa.compacted { |
| 154 | break |
| 155 | } else if wa.ch == nil { |
| 156 | // already canceled (e.g., cancel/close race) |
| 157 | break |
| 158 | } |
| 159 | |
| 160 | if !wa.victim { |
| 161 | panic("watcher not victim but not in watch groups") |
| 162 | } |
| 163 | |
| 164 | var victimBatch watcherBatch |
| 165 | for _, wb := range s.victims { |
| 166 | if wb[wa] != nil { |
| 167 | victimBatch = wb |
| 168 | break |
| 169 | } |
| 170 | } |
| 171 | if victimBatch != nil { |
| 172 | slowWatcherGauge.Dec() |
| 173 | delete(victimBatch, wa) |
| 174 | break |
| 175 | } |
| 176 | |
| 177 | // victim being processed so not accessible; retry |
| 178 | s.mu.Unlock() |
| 179 | time.Sleep(time.Millisecond) |
| 180 | } |
| 181 | |
| 182 | watcherGauge.Dec() |
| 183 | wa.ch = nil |
| 184 | s.mu.Unlock() |
| 185 | } |
| 186 | |
| 187 | func (s *watchableStore) Restore(b backend.Backend) error { |
| 188 | s.mu.Lock() |
| 189 | defer s.mu.Unlock() |
| 190 | err := s.store.Restore(b) |
| 191 | if err != nil { |
| 192 | return err |
| 193 | } |
| 194 | |
| 195 | for wa := range s.synced.watchers { |
| 196 | wa.restore = true |
| 197 | s.unsynced.add(wa) |
| 198 | } |
| 199 | s.synced = newWatcherGroup() |
| 200 | return nil |
| 201 | } |
| 202 | |
| 203 | // syncWatchersLoop syncs the watcher in the unsynced map every 100ms. |
| 204 | func (s *watchableStore) syncWatchersLoop() { |
| 205 | defer s.wg.Done() |
| 206 | |
| 207 | for { |
| 208 | s.mu.RLock() |
| 209 | st := time.Now() |
| 210 | lastUnsyncedWatchers := s.unsynced.size() |
| 211 | s.mu.RUnlock() |
| 212 | |
| 213 | unsyncedWatchers := 0 |
| 214 | if lastUnsyncedWatchers > 0 { |
| 215 | unsyncedWatchers = s.syncWatchers() |
| 216 | } |
| 217 | syncDuration := time.Since(st) |
| 218 | |
| 219 | waitDuration := 100 * time.Millisecond |
| 220 | // more work pending? |
| 221 | if unsyncedWatchers != 0 && lastUnsyncedWatchers > unsyncedWatchers { |
| 222 | // be fair to other store operations by yielding time taken |
| 223 | waitDuration = syncDuration |
| 224 | } |
| 225 | |
| 226 | select { |
| 227 | case <-time.After(waitDuration): |
| 228 | case <-s.stopc: |
| 229 | return |
| 230 | } |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | // syncVictimsLoop tries to write precomputed watcher responses to |
| 235 | // watchers that had a blocked watcher channel |
| 236 | func (s *watchableStore) syncVictimsLoop() { |
| 237 | defer s.wg.Done() |
| 238 | |
| 239 | for { |
| 240 | for s.moveVictims() != 0 { |
| 241 | // try to update all victim watchers |
| 242 | } |
| 243 | s.mu.RLock() |
| 244 | isEmpty := len(s.victims) == 0 |
| 245 | s.mu.RUnlock() |
| 246 | |
| 247 | var tickc <-chan time.Time |
| 248 | if !isEmpty { |
| 249 | tickc = time.After(10 * time.Millisecond) |
| 250 | } |
| 251 | |
| 252 | select { |
| 253 | case <-tickc: |
| 254 | case <-s.victimc: |
| 255 | case <-s.stopc: |
| 256 | return |
| 257 | } |
| 258 | } |
| 259 | } |
| 260 | |
| 261 | // moveVictims tries to update watches with already pending event data |
| 262 | func (s *watchableStore) moveVictims() (moved int) { |
| 263 | s.mu.Lock() |
| 264 | victims := s.victims |
| 265 | s.victims = nil |
| 266 | s.mu.Unlock() |
| 267 | |
| 268 | var newVictim watcherBatch |
| 269 | for _, wb := range victims { |
| 270 | // try to send responses again |
| 271 | for w, eb := range wb { |
| 272 | // watcher has observed the store up to, but not including, w.minRev |
| 273 | rev := w.minRev - 1 |
| 274 | if w.send(WatchResponse{WatchID: w.id, Events: eb.evs, Revision: rev}) { |
| 275 | pendingEventsGauge.Add(float64(len(eb.evs))) |
| 276 | } else { |
| 277 | if newVictim == nil { |
| 278 | newVictim = make(watcherBatch) |
| 279 | } |
| 280 | newVictim[w] = eb |
| 281 | continue |
| 282 | } |
| 283 | moved++ |
| 284 | } |
| 285 | |
| 286 | // assign completed victim watchers to unsync/sync |
| 287 | s.mu.Lock() |
| 288 | s.store.revMu.RLock() |
| 289 | curRev := s.store.currentRev |
| 290 | for w, eb := range wb { |
| 291 | if newVictim != nil && newVictim[w] != nil { |
| 292 | // couldn't send watch response; stays victim |
| 293 | continue |
| 294 | } |
| 295 | w.victim = false |
| 296 | if eb.moreRev != 0 { |
| 297 | w.minRev = eb.moreRev |
| 298 | } |
| 299 | if w.minRev <= curRev { |
| 300 | s.unsynced.add(w) |
| 301 | } else { |
| 302 | slowWatcherGauge.Dec() |
| 303 | s.synced.add(w) |
| 304 | } |
| 305 | } |
| 306 | s.store.revMu.RUnlock() |
| 307 | s.mu.Unlock() |
| 308 | } |
| 309 | |
| 310 | if len(newVictim) > 0 { |
| 311 | s.mu.Lock() |
| 312 | s.victims = append(s.victims, newVictim) |
| 313 | s.mu.Unlock() |
| 314 | } |
| 315 | |
| 316 | return moved |
| 317 | } |
| 318 | |
| 319 | // syncWatchers syncs unsynced watchers by: |
| 320 | // 1. choose a set of watchers from the unsynced watcher group |
| 321 | // 2. iterate over the set to get the minimum revision and remove compacted watchers |
| 322 | // 3. use minimum revision to get all key-value pairs and send those events to watchers |
| 323 | // 4. remove synced watchers in set from unsynced group and move to synced group |
| 324 | func (s *watchableStore) syncWatchers() int { |
| 325 | s.mu.Lock() |
| 326 | defer s.mu.Unlock() |
| 327 | |
| 328 | if s.unsynced.size() == 0 { |
| 329 | return 0 |
| 330 | } |
| 331 | |
| 332 | s.store.revMu.RLock() |
| 333 | defer s.store.revMu.RUnlock() |
| 334 | |
| 335 | // in order to find key-value pairs from unsynced watchers, we need to |
| 336 | // find min revision index, and these revisions can be used to |
| 337 | // query the backend store of key-value pairs |
| 338 | curRev := s.store.currentRev |
| 339 | compactionRev := s.store.compactMainRev |
| 340 | |
| 341 | wg, minRev := s.unsynced.choose(maxWatchersPerSync, curRev, compactionRev) |
| 342 | minBytes, maxBytes := newRevBytes(), newRevBytes() |
| 343 | revToBytes(revision{main: minRev}, minBytes) |
| 344 | revToBytes(revision{main: curRev + 1}, maxBytes) |
| 345 | |
| 346 | // UnsafeRange returns keys and values. And in boltdb, keys are revisions. |
| 347 | // values are actual key-value pairs in backend. |
| 348 | tx := s.store.b.ReadTx() |
| 349 | tx.RLock() |
| 350 | revs, vs := tx.UnsafeRange(keyBucketName, minBytes, maxBytes, 0) |
| 351 | var evs []mvccpb.Event |
| 352 | if s.store != nil && s.store.lg != nil { |
| 353 | evs = kvsToEvents(s.store.lg, wg, revs, vs) |
| 354 | } else { |
| 355 | // TODO: remove this in v3.5 |
| 356 | evs = kvsToEvents(nil, wg, revs, vs) |
| 357 | } |
| 358 | tx.RUnlock() |
| 359 | |
| 360 | var victims watcherBatch |
| 361 | wb := newWatcherBatch(wg, evs) |
| 362 | for w := range wg.watchers { |
| 363 | w.minRev = curRev + 1 |
| 364 | |
| 365 | eb, ok := wb[w] |
| 366 | if !ok { |
| 367 | // bring un-notified watcher to synced |
| 368 | s.synced.add(w) |
| 369 | s.unsynced.delete(w) |
| 370 | continue |
| 371 | } |
| 372 | |
| 373 | if eb.moreRev != 0 { |
| 374 | w.minRev = eb.moreRev |
| 375 | } |
| 376 | |
| 377 | if w.send(WatchResponse{WatchID: w.id, Events: eb.evs, Revision: curRev}) { |
| 378 | pendingEventsGauge.Add(float64(len(eb.evs))) |
| 379 | } else { |
| 380 | if victims == nil { |
| 381 | victims = make(watcherBatch) |
| 382 | } |
| 383 | w.victim = true |
| 384 | } |
| 385 | |
| 386 | if w.victim { |
| 387 | victims[w] = eb |
| 388 | } else { |
| 389 | if eb.moreRev != 0 { |
| 390 | // stay unsynced; more to read |
| 391 | continue |
| 392 | } |
| 393 | s.synced.add(w) |
| 394 | } |
| 395 | s.unsynced.delete(w) |
| 396 | } |
| 397 | s.addVictim(victims) |
| 398 | |
| 399 | vsz := 0 |
| 400 | for _, v := range s.victims { |
| 401 | vsz += len(v) |
| 402 | } |
| 403 | slowWatcherGauge.Set(float64(s.unsynced.size() + vsz)) |
| 404 | |
| 405 | return s.unsynced.size() |
| 406 | } |
| 407 | |
| 408 | // kvsToEvents gets all events for the watchers from all key-value pairs |
| 409 | func kvsToEvents(lg *zap.Logger, wg *watcherGroup, revs, vals [][]byte) (evs []mvccpb.Event) { |
| 410 | for i, v := range vals { |
| 411 | var kv mvccpb.KeyValue |
| 412 | if err := kv.Unmarshal(v); err != nil { |
| 413 | if lg != nil { |
| 414 | lg.Panic("failed to unmarshal mvccpb.KeyValue", zap.Error(err)) |
| 415 | } else { |
| 416 | plog.Panicf("cannot unmarshal event: %v", err) |
| 417 | } |
| 418 | } |
| 419 | |
| 420 | if !wg.contains(string(kv.Key)) { |
| 421 | continue |
| 422 | } |
| 423 | |
| 424 | ty := mvccpb.PUT |
| 425 | if isTombstone(revs[i]) { |
| 426 | ty = mvccpb.DELETE |
| 427 | // patch in mod revision so watchers won't skip |
| 428 | kv.ModRevision = bytesToRev(revs[i]).main |
| 429 | } |
| 430 | evs = append(evs, mvccpb.Event{Kv: &kv, Type: ty}) |
| 431 | } |
| 432 | return evs |
| 433 | } |
| 434 | |
| 435 | // notify notifies the fact that given event at the given rev just happened to |
| 436 | // watchers that watch on the key of the event. |
| 437 | func (s *watchableStore) notify(rev int64, evs []mvccpb.Event) { |
| 438 | var victim watcherBatch |
| 439 | for w, eb := range newWatcherBatch(&s.synced, evs) { |
| 440 | if eb.revs != 1 { |
| 441 | if s.store != nil && s.store.lg != nil { |
| 442 | s.store.lg.Panic( |
| 443 | "unexpected multiple revisions in watch notification", |
| 444 | zap.Int("number-of-revisions", eb.revs), |
| 445 | ) |
| 446 | } else { |
| 447 | plog.Panicf("unexpected multiple revisions in notification") |
| 448 | } |
| 449 | } |
| 450 | if w.send(WatchResponse{WatchID: w.id, Events: eb.evs, Revision: rev}) { |
| 451 | pendingEventsGauge.Add(float64(len(eb.evs))) |
| 452 | } else { |
| 453 | // move slow watcher to victims |
| 454 | w.minRev = rev + 1 |
| 455 | if victim == nil { |
| 456 | victim = make(watcherBatch) |
| 457 | } |
| 458 | w.victim = true |
| 459 | victim[w] = eb |
| 460 | s.synced.delete(w) |
| 461 | slowWatcherGauge.Inc() |
| 462 | } |
| 463 | } |
| 464 | s.addVictim(victim) |
| 465 | } |
| 466 | |
| 467 | func (s *watchableStore) addVictim(victim watcherBatch) { |
| 468 | if victim == nil { |
| 469 | return |
| 470 | } |
| 471 | s.victims = append(s.victims, victim) |
| 472 | select { |
| 473 | case s.victimc <- struct{}{}: |
| 474 | default: |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | func (s *watchableStore) rev() int64 { return s.store.Rev() } |
| 479 | |
| 480 | func (s *watchableStore) progress(w *watcher) { |
| 481 | s.mu.RLock() |
| 482 | defer s.mu.RUnlock() |
| 483 | |
| 484 | if _, ok := s.synced.watchers[w]; ok { |
| 485 | w.send(WatchResponse{WatchID: w.id, Revision: s.rev()}) |
| 486 | // If the ch is full, this watcher is receiving events. |
| 487 | // We do not need to send progress at all. |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | type watcher struct { |
| 492 | // the watcher key |
| 493 | key []byte |
| 494 | // end indicates the end of the range to watch. |
| 495 | // If end is set, the watcher is on a range. |
| 496 | end []byte |
| 497 | |
| 498 | // victim is set when ch is blocked and undergoing victim processing |
| 499 | victim bool |
| 500 | |
| 501 | // compacted is set when the watcher is removed because of compaction |
| 502 | compacted bool |
| 503 | |
| 504 | // restore is true when the watcher is being restored from leader snapshot |
| 505 | // which means that this watcher has just been moved from "synced" to "unsynced" |
| 506 | // watcher group, possibly with a future revision when it was first added |
| 507 | // to the synced watcher |
| 508 | // "unsynced" watcher revision must always be <= current revision, |
| 509 | // except when the watcher were to be moved from "synced" watcher group |
| 510 | restore bool |
| 511 | |
| 512 | // minRev is the minimum revision update the watcher will accept |
| 513 | minRev int64 |
| 514 | id WatchID |
| 515 | |
| 516 | fcs []FilterFunc |
| 517 | // a chan to send out the watch response. |
| 518 | // The chan might be shared with other watchers. |
| 519 | ch chan<- WatchResponse |
| 520 | } |
| 521 | |
| 522 | func (w *watcher) send(wr WatchResponse) bool { |
| 523 | progressEvent := len(wr.Events) == 0 |
| 524 | |
| 525 | if len(w.fcs) != 0 { |
| 526 | ne := make([]mvccpb.Event, 0, len(wr.Events)) |
| 527 | for i := range wr.Events { |
| 528 | filtered := false |
| 529 | for _, filter := range w.fcs { |
| 530 | if filter(wr.Events[i]) { |
| 531 | filtered = true |
| 532 | break |
| 533 | } |
| 534 | } |
| 535 | if !filtered { |
| 536 | ne = append(ne, wr.Events[i]) |
| 537 | } |
| 538 | } |
| 539 | wr.Events = ne |
| 540 | } |
| 541 | |
| 542 | // if all events are filtered out, we should send nothing. |
| 543 | if !progressEvent && len(wr.Events) == 0 { |
| 544 | return true |
| 545 | } |
| 546 | select { |
| 547 | case w.ch <- wr: |
| 548 | return true |
| 549 | default: |
| 550 | return false |
| 551 | } |
| 552 | } |