Naveen Sampath | 04696f7 | 2022-06-13 15:19:14 +0530 | [diff] [blame] | 1 | // Copyright 2016 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 clientv3 |
| 16 | |
| 17 | import ( |
| 18 | "context" |
| 19 | "errors" |
| 20 | "fmt" |
| 21 | "sync" |
| 22 | "time" |
| 23 | |
| 24 | v3rpc "github.com/coreos/etcd/etcdserver/api/v3rpc/rpctypes" |
| 25 | pb "github.com/coreos/etcd/etcdserver/etcdserverpb" |
| 26 | mvccpb "github.com/coreos/etcd/mvcc/mvccpb" |
| 27 | |
| 28 | "google.golang.org/grpc" |
| 29 | "google.golang.org/grpc/codes" |
| 30 | "google.golang.org/grpc/metadata" |
| 31 | "google.golang.org/grpc/status" |
| 32 | ) |
| 33 | |
| 34 | const ( |
| 35 | EventTypeDelete = mvccpb.DELETE |
| 36 | EventTypePut = mvccpb.PUT |
| 37 | |
| 38 | closeSendErrTimeout = 250 * time.Millisecond |
| 39 | ) |
| 40 | |
| 41 | type Event mvccpb.Event |
| 42 | |
| 43 | type WatchChan <-chan WatchResponse |
| 44 | |
| 45 | type Watcher interface { |
| 46 | // Watch watches on a key or prefix. The watched events will be returned |
| 47 | // through the returned channel. If revisions waiting to be sent over the |
| 48 | // watch are compacted, then the watch will be canceled by the server, the |
| 49 | // client will post a compacted error watch response, and the channel will close. |
| 50 | // If the context "ctx" is canceled or timed out, returned "WatchChan" is closed, |
| 51 | // and "WatchResponse" from this closed channel has zero events and nil "Err()". |
| 52 | // The context "ctx" MUST be canceled, as soon as watcher is no longer being used, |
| 53 | // to release the associated resources. |
| 54 | // |
| 55 | // If the context is "context.Background/TODO", returned "WatchChan" will |
| 56 | // not be closed and block until event is triggered, except when server |
| 57 | // returns a non-recoverable error (e.g. ErrCompacted). |
| 58 | // For example, when context passed with "WithRequireLeader" and the |
| 59 | // connected server has no leader (e.g. due to network partition), |
| 60 | // error "etcdserver: no leader" (ErrNoLeader) will be returned, |
| 61 | // and then "WatchChan" is closed with non-nil "Err()". |
| 62 | // In order to prevent a watch stream being stuck in a partitioned node, |
| 63 | // make sure to wrap context with "WithRequireLeader". |
| 64 | // |
| 65 | // Otherwise, as long as the context has not been canceled or timed out, |
| 66 | // watch will retry on other recoverable errors forever until reconnected. |
| 67 | // |
| 68 | // TODO: explicitly set context error in the last "WatchResponse" message and close channel? |
| 69 | // Currently, client contexts are overwritten with "valCtx" that never closes. |
| 70 | // TODO(v3.4): configure watch retry policy, limit maximum retry number |
| 71 | // (see https://github.com/etcd-io/etcd/issues/8980) |
| 72 | Watch(ctx context.Context, key string, opts ...OpOption) WatchChan |
| 73 | |
| 74 | // RequestProgress requests a progress notify response be sent in all watch channels. |
| 75 | RequestProgress(ctx context.Context) error |
| 76 | |
| 77 | // Close closes the watcher and cancels all watch requests. |
| 78 | Close() error |
| 79 | } |
| 80 | |
| 81 | type WatchResponse struct { |
| 82 | Header pb.ResponseHeader |
| 83 | Events []*Event |
| 84 | |
| 85 | // CompactRevision is the minimum revision the watcher may receive. |
| 86 | CompactRevision int64 |
| 87 | |
| 88 | // Canceled is used to indicate watch failure. |
| 89 | // If the watch failed and the stream was about to close, before the channel is closed, |
| 90 | // the channel sends a final response that has Canceled set to true with a non-nil Err(). |
| 91 | Canceled bool |
| 92 | |
| 93 | // Created is used to indicate the creation of the watcher. |
| 94 | Created bool |
| 95 | |
| 96 | closeErr error |
| 97 | |
| 98 | // cancelReason is a reason of canceling watch |
| 99 | cancelReason string |
| 100 | } |
| 101 | |
| 102 | // IsCreate returns true if the event tells that the key is newly created. |
| 103 | func (e *Event) IsCreate() bool { |
| 104 | return e.Type == EventTypePut && e.Kv.CreateRevision == e.Kv.ModRevision |
| 105 | } |
| 106 | |
| 107 | // IsModify returns true if the event tells that a new value is put on existing key. |
| 108 | func (e *Event) IsModify() bool { |
| 109 | return e.Type == EventTypePut && e.Kv.CreateRevision != e.Kv.ModRevision |
| 110 | } |
| 111 | |
| 112 | // Err is the error value if this WatchResponse holds an error. |
| 113 | func (wr *WatchResponse) Err() error { |
| 114 | switch { |
| 115 | case wr.closeErr != nil: |
| 116 | return v3rpc.Error(wr.closeErr) |
| 117 | case wr.CompactRevision != 0: |
| 118 | return v3rpc.ErrCompacted |
| 119 | case wr.Canceled: |
| 120 | if len(wr.cancelReason) != 0 { |
| 121 | return v3rpc.Error(status.Error(codes.FailedPrecondition, wr.cancelReason)) |
| 122 | } |
| 123 | return v3rpc.ErrFutureRev |
| 124 | } |
| 125 | return nil |
| 126 | } |
| 127 | |
| 128 | // IsProgressNotify returns true if the WatchResponse is progress notification. |
| 129 | func (wr *WatchResponse) IsProgressNotify() bool { |
| 130 | return len(wr.Events) == 0 && !wr.Canceled && !wr.Created && wr.CompactRevision == 0 && wr.Header.Revision != 0 |
| 131 | } |
| 132 | |
| 133 | // watcher implements the Watcher interface |
| 134 | type watcher struct { |
| 135 | remote pb.WatchClient |
| 136 | callOpts []grpc.CallOption |
| 137 | |
| 138 | // mu protects the grpc streams map |
| 139 | mu sync.RWMutex |
| 140 | |
| 141 | // streams holds all the active grpc streams keyed by ctx value. |
| 142 | streams map[string]*watchGrpcStream |
| 143 | } |
| 144 | |
| 145 | // watchGrpcStream tracks all watch resources attached to a single grpc stream. |
| 146 | type watchGrpcStream struct { |
| 147 | owner *watcher |
| 148 | remote pb.WatchClient |
| 149 | callOpts []grpc.CallOption |
| 150 | |
| 151 | // ctx controls internal remote.Watch requests |
| 152 | ctx context.Context |
| 153 | // ctxKey is the key used when looking up this stream's context |
| 154 | ctxKey string |
| 155 | cancel context.CancelFunc |
| 156 | |
| 157 | // substreams holds all active watchers on this grpc stream |
| 158 | substreams map[int64]*watcherStream |
| 159 | // resuming holds all resuming watchers on this grpc stream |
| 160 | resuming []*watcherStream |
| 161 | |
| 162 | // reqc sends a watch request from Watch() to the main goroutine |
| 163 | reqc chan watchStreamRequest |
| 164 | // respc receives data from the watch client |
| 165 | respc chan *pb.WatchResponse |
| 166 | // donec closes to broadcast shutdown |
| 167 | donec chan struct{} |
| 168 | // errc transmits errors from grpc Recv to the watch stream reconnect logic |
| 169 | errc chan error |
| 170 | // closingc gets the watcherStream of closing watchers |
| 171 | closingc chan *watcherStream |
| 172 | // wg is Done when all substream goroutines have exited |
| 173 | wg sync.WaitGroup |
| 174 | |
| 175 | // resumec closes to signal that all substreams should begin resuming |
| 176 | resumec chan struct{} |
| 177 | // closeErr is the error that closed the watch stream |
| 178 | closeErr error |
| 179 | } |
| 180 | |
| 181 | // watchStreamRequest is a union of the supported watch request operation types |
| 182 | type watchStreamRequest interface { |
| 183 | toPB() *pb.WatchRequest |
| 184 | } |
| 185 | |
| 186 | // watchRequest is issued by the subscriber to start a new watcher |
| 187 | type watchRequest struct { |
| 188 | ctx context.Context |
| 189 | key string |
| 190 | end string |
| 191 | rev int64 |
| 192 | |
| 193 | // send created notification event if this field is true |
| 194 | createdNotify bool |
| 195 | // progressNotify is for progress updates |
| 196 | progressNotify bool |
| 197 | // fragmentation should be disabled by default |
| 198 | // if true, split watch events when total exceeds |
| 199 | // "--max-request-bytes" flag value + 512-byte |
| 200 | fragment bool |
| 201 | |
| 202 | // filters is the list of events to filter out |
| 203 | filters []pb.WatchCreateRequest_FilterType |
| 204 | // get the previous key-value pair before the event happens |
| 205 | prevKV bool |
| 206 | // retc receives a chan WatchResponse once the watcher is established |
| 207 | retc chan chan WatchResponse |
| 208 | } |
| 209 | |
| 210 | // progressRequest is issued by the subscriber to request watch progress |
| 211 | type progressRequest struct { |
| 212 | } |
| 213 | |
| 214 | // watcherStream represents a registered watcher |
| 215 | type watcherStream struct { |
| 216 | // initReq is the request that initiated this request |
| 217 | initReq watchRequest |
| 218 | |
| 219 | // outc publishes watch responses to subscriber |
| 220 | outc chan WatchResponse |
| 221 | // recvc buffers watch responses before publishing |
| 222 | recvc chan *WatchResponse |
| 223 | // donec closes when the watcherStream goroutine stops. |
| 224 | donec chan struct{} |
| 225 | // closing is set to true when stream should be scheduled to shutdown. |
| 226 | closing bool |
| 227 | // id is the registered watch id on the grpc stream |
| 228 | id int64 |
| 229 | |
| 230 | // buf holds all events received from etcd but not yet consumed by the client |
| 231 | buf []*WatchResponse |
| 232 | } |
| 233 | |
| 234 | func NewWatcher(c *Client) Watcher { |
| 235 | return NewWatchFromWatchClient(pb.NewWatchClient(c.conn), c) |
| 236 | } |
| 237 | |
| 238 | func NewWatchFromWatchClient(wc pb.WatchClient, c *Client) Watcher { |
| 239 | w := &watcher{ |
| 240 | remote: wc, |
| 241 | streams: make(map[string]*watchGrpcStream), |
| 242 | } |
| 243 | if c != nil { |
| 244 | w.callOpts = c.callOpts |
| 245 | } |
| 246 | return w |
| 247 | } |
| 248 | |
| 249 | // never closes |
| 250 | var valCtxCh = make(chan struct{}) |
| 251 | var zeroTime = time.Unix(0, 0) |
| 252 | |
| 253 | // ctx with only the values; never Done |
| 254 | type valCtx struct{ context.Context } |
| 255 | |
| 256 | func (vc *valCtx) Deadline() (time.Time, bool) { return zeroTime, false } |
| 257 | func (vc *valCtx) Done() <-chan struct{} { return valCtxCh } |
| 258 | func (vc *valCtx) Err() error { return nil } |
| 259 | |
| 260 | func (w *watcher) newWatcherGrpcStream(inctx context.Context) *watchGrpcStream { |
| 261 | ctx, cancel := context.WithCancel(&valCtx{inctx}) |
| 262 | wgs := &watchGrpcStream{ |
| 263 | owner: w, |
| 264 | remote: w.remote, |
| 265 | callOpts: w.callOpts, |
| 266 | ctx: ctx, |
| 267 | ctxKey: streamKeyFromCtx(inctx), |
| 268 | cancel: cancel, |
| 269 | substreams: make(map[int64]*watcherStream), |
| 270 | respc: make(chan *pb.WatchResponse), |
| 271 | reqc: make(chan watchStreamRequest), |
| 272 | donec: make(chan struct{}), |
| 273 | errc: make(chan error, 1), |
| 274 | closingc: make(chan *watcherStream), |
| 275 | resumec: make(chan struct{}), |
| 276 | } |
| 277 | go wgs.run() |
| 278 | return wgs |
| 279 | } |
| 280 | |
| 281 | // Watch posts a watch request to run() and waits for a new watcher channel |
| 282 | func (w *watcher) Watch(ctx context.Context, key string, opts ...OpOption) WatchChan { |
| 283 | ow := opWatch(key, opts...) |
| 284 | |
| 285 | var filters []pb.WatchCreateRequest_FilterType |
| 286 | if ow.filterPut { |
| 287 | filters = append(filters, pb.WatchCreateRequest_NOPUT) |
| 288 | } |
| 289 | if ow.filterDelete { |
| 290 | filters = append(filters, pb.WatchCreateRequest_NODELETE) |
| 291 | } |
| 292 | |
| 293 | wr := &watchRequest{ |
| 294 | ctx: ctx, |
| 295 | createdNotify: ow.createdNotify, |
| 296 | key: string(ow.key), |
| 297 | end: string(ow.end), |
| 298 | rev: ow.rev, |
| 299 | progressNotify: ow.progressNotify, |
| 300 | fragment: ow.fragment, |
| 301 | filters: filters, |
| 302 | prevKV: ow.prevKV, |
| 303 | retc: make(chan chan WatchResponse, 1), |
| 304 | } |
| 305 | |
| 306 | ok := false |
| 307 | ctxKey := streamKeyFromCtx(ctx) |
| 308 | |
| 309 | // find or allocate appropriate grpc watch stream |
| 310 | w.mu.Lock() |
| 311 | if w.streams == nil { |
| 312 | // closed |
| 313 | w.mu.Unlock() |
| 314 | ch := make(chan WatchResponse) |
| 315 | close(ch) |
| 316 | return ch |
| 317 | } |
| 318 | wgs := w.streams[ctxKey] |
| 319 | if wgs == nil { |
| 320 | wgs = w.newWatcherGrpcStream(ctx) |
| 321 | w.streams[ctxKey] = wgs |
| 322 | } |
| 323 | donec := wgs.donec |
| 324 | reqc := wgs.reqc |
| 325 | w.mu.Unlock() |
| 326 | |
| 327 | // couldn't create channel; return closed channel |
| 328 | closeCh := make(chan WatchResponse, 1) |
| 329 | |
| 330 | // submit request |
| 331 | select { |
| 332 | case reqc <- wr: |
| 333 | ok = true |
| 334 | case <-wr.ctx.Done(): |
| 335 | case <-donec: |
| 336 | if wgs.closeErr != nil { |
| 337 | closeCh <- WatchResponse{Canceled: true, closeErr: wgs.closeErr} |
| 338 | break |
| 339 | } |
| 340 | // retry; may have dropped stream from no ctxs |
| 341 | return w.Watch(ctx, key, opts...) |
| 342 | } |
| 343 | |
| 344 | // receive channel |
| 345 | if ok { |
| 346 | select { |
| 347 | case ret := <-wr.retc: |
| 348 | return ret |
| 349 | case <-ctx.Done(): |
| 350 | case <-donec: |
| 351 | if wgs.closeErr != nil { |
| 352 | closeCh <- WatchResponse{Canceled: true, closeErr: wgs.closeErr} |
| 353 | break |
| 354 | } |
| 355 | // retry; may have dropped stream from no ctxs |
| 356 | return w.Watch(ctx, key, opts...) |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | close(closeCh) |
| 361 | return closeCh |
| 362 | } |
| 363 | |
| 364 | func (w *watcher) Close() (err error) { |
| 365 | w.mu.Lock() |
| 366 | streams := w.streams |
| 367 | w.streams = nil |
| 368 | w.mu.Unlock() |
| 369 | for _, wgs := range streams { |
| 370 | if werr := wgs.close(); werr != nil { |
| 371 | err = werr |
| 372 | } |
| 373 | } |
| 374 | // Consider context.Canceled as a successful close |
| 375 | if err == context.Canceled { |
| 376 | err = nil |
| 377 | } |
| 378 | return err |
| 379 | } |
| 380 | |
| 381 | // RequestProgress requests a progress notify response be sent in all watch channels. |
| 382 | func (w *watcher) RequestProgress(ctx context.Context) (err error) { |
| 383 | ctxKey := streamKeyFromCtx(ctx) |
| 384 | |
| 385 | w.mu.Lock() |
| 386 | if w.streams == nil { |
| 387 | w.mu.Unlock() |
| 388 | return fmt.Errorf("no stream found for context") |
| 389 | } |
| 390 | wgs := w.streams[ctxKey] |
| 391 | if wgs == nil { |
| 392 | wgs = w.newWatcherGrpcStream(ctx) |
| 393 | w.streams[ctxKey] = wgs |
| 394 | } |
| 395 | donec := wgs.donec |
| 396 | reqc := wgs.reqc |
| 397 | w.mu.Unlock() |
| 398 | |
| 399 | pr := &progressRequest{} |
| 400 | |
| 401 | select { |
| 402 | case reqc <- pr: |
| 403 | return nil |
| 404 | case <-ctx.Done(): |
| 405 | if err == nil { |
| 406 | return ctx.Err() |
| 407 | } |
| 408 | return err |
| 409 | case <-donec: |
| 410 | if wgs.closeErr != nil { |
| 411 | return wgs.closeErr |
| 412 | } |
| 413 | // retry; may have dropped stream from no ctxs |
| 414 | return w.RequestProgress(ctx) |
| 415 | } |
| 416 | } |
| 417 | |
| 418 | func (w *watchGrpcStream) close() (err error) { |
| 419 | w.cancel() |
| 420 | <-w.donec |
| 421 | select { |
| 422 | case err = <-w.errc: |
| 423 | default: |
| 424 | } |
| 425 | return toErr(w.ctx, err) |
| 426 | } |
| 427 | |
| 428 | func (w *watcher) closeStream(wgs *watchGrpcStream) { |
| 429 | w.mu.Lock() |
| 430 | close(wgs.donec) |
| 431 | wgs.cancel() |
| 432 | if w.streams != nil { |
| 433 | delete(w.streams, wgs.ctxKey) |
| 434 | } |
| 435 | w.mu.Unlock() |
| 436 | } |
| 437 | |
| 438 | func (w *watchGrpcStream) addSubstream(resp *pb.WatchResponse, ws *watcherStream) { |
| 439 | // check watch ID for backward compatibility (<= v3.3) |
| 440 | if resp.WatchId == -1 || (resp.Canceled && resp.CancelReason != "") { |
| 441 | w.closeErr = v3rpc.Error(errors.New(resp.CancelReason)) |
| 442 | // failed; no channel |
| 443 | close(ws.recvc) |
| 444 | return |
| 445 | } |
| 446 | ws.id = resp.WatchId |
| 447 | w.substreams[ws.id] = ws |
| 448 | } |
| 449 | |
| 450 | func (w *watchGrpcStream) sendCloseSubstream(ws *watcherStream, resp *WatchResponse) { |
| 451 | select { |
| 452 | case ws.outc <- *resp: |
| 453 | case <-ws.initReq.ctx.Done(): |
| 454 | case <-time.After(closeSendErrTimeout): |
| 455 | } |
| 456 | close(ws.outc) |
| 457 | } |
| 458 | |
| 459 | func (w *watchGrpcStream) closeSubstream(ws *watcherStream) { |
| 460 | // send channel response in case stream was never established |
| 461 | select { |
| 462 | case ws.initReq.retc <- ws.outc: |
| 463 | default: |
| 464 | } |
| 465 | // close subscriber's channel |
| 466 | if closeErr := w.closeErr; closeErr != nil && ws.initReq.ctx.Err() == nil { |
| 467 | go w.sendCloseSubstream(ws, &WatchResponse{Canceled: true, closeErr: w.closeErr}) |
| 468 | } else if ws.outc != nil { |
| 469 | close(ws.outc) |
| 470 | } |
| 471 | if ws.id != -1 { |
| 472 | delete(w.substreams, ws.id) |
| 473 | return |
| 474 | } |
| 475 | for i := range w.resuming { |
| 476 | if w.resuming[i] == ws { |
| 477 | w.resuming[i] = nil |
| 478 | return |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | // run is the root of the goroutines for managing a watcher client |
| 484 | func (w *watchGrpcStream) run() { |
| 485 | var wc pb.Watch_WatchClient |
| 486 | var closeErr error |
| 487 | |
| 488 | // substreams marked to close but goroutine still running; needed for |
| 489 | // avoiding double-closing recvc on grpc stream teardown |
| 490 | closing := make(map[*watcherStream]struct{}) |
| 491 | |
| 492 | defer func() { |
| 493 | w.closeErr = closeErr |
| 494 | // shutdown substreams and resuming substreams |
| 495 | for _, ws := range w.substreams { |
| 496 | if _, ok := closing[ws]; !ok { |
| 497 | close(ws.recvc) |
| 498 | closing[ws] = struct{}{} |
| 499 | } |
| 500 | } |
| 501 | for _, ws := range w.resuming { |
| 502 | if _, ok := closing[ws]; ws != nil && !ok { |
| 503 | close(ws.recvc) |
| 504 | closing[ws] = struct{}{} |
| 505 | } |
| 506 | } |
| 507 | w.joinSubstreams() |
| 508 | for range closing { |
| 509 | w.closeSubstream(<-w.closingc) |
| 510 | } |
| 511 | w.wg.Wait() |
| 512 | w.owner.closeStream(w) |
| 513 | }() |
| 514 | |
| 515 | // start a stream with the etcd grpc server |
| 516 | if wc, closeErr = w.newWatchClient(); closeErr != nil { |
| 517 | return |
| 518 | } |
| 519 | |
| 520 | cancelSet := make(map[int64]struct{}) |
| 521 | |
| 522 | var cur *pb.WatchResponse |
| 523 | for { |
| 524 | select { |
| 525 | // Watch() requested |
| 526 | case req := <-w.reqc: |
| 527 | switch wreq := req.(type) { |
| 528 | case *watchRequest: |
| 529 | outc := make(chan WatchResponse, 1) |
| 530 | // TODO: pass custom watch ID? |
| 531 | ws := &watcherStream{ |
| 532 | initReq: *wreq, |
| 533 | id: -1, |
| 534 | outc: outc, |
| 535 | // unbuffered so resumes won't cause repeat events |
| 536 | recvc: make(chan *WatchResponse), |
| 537 | } |
| 538 | |
| 539 | ws.donec = make(chan struct{}) |
| 540 | w.wg.Add(1) |
| 541 | go w.serveSubstream(ws, w.resumec) |
| 542 | |
| 543 | // queue up for watcher creation/resume |
| 544 | w.resuming = append(w.resuming, ws) |
| 545 | if len(w.resuming) == 1 { |
| 546 | // head of resume queue, can register a new watcher |
| 547 | wc.Send(ws.initReq.toPB()) |
| 548 | } |
| 549 | case *progressRequest: |
| 550 | wc.Send(wreq.toPB()) |
| 551 | } |
| 552 | |
| 553 | // new events from the watch client |
| 554 | case pbresp := <-w.respc: |
| 555 | if cur == nil || pbresp.Created || pbresp.Canceled { |
| 556 | cur = pbresp |
| 557 | } else if cur != nil && cur.WatchId == pbresp.WatchId { |
| 558 | // merge new events |
| 559 | cur.Events = append(cur.Events, pbresp.Events...) |
| 560 | // update "Fragment" field; last response with "Fragment" == false |
| 561 | cur.Fragment = pbresp.Fragment |
| 562 | } |
| 563 | |
| 564 | switch { |
| 565 | case pbresp.Created: |
| 566 | // response to head of queue creation |
| 567 | if ws := w.resuming[0]; ws != nil { |
| 568 | w.addSubstream(pbresp, ws) |
| 569 | w.dispatchEvent(pbresp) |
| 570 | w.resuming[0] = nil |
| 571 | } |
| 572 | |
| 573 | if ws := w.nextResume(); ws != nil { |
| 574 | wc.Send(ws.initReq.toPB()) |
| 575 | } |
| 576 | |
| 577 | // reset for next iteration |
| 578 | cur = nil |
| 579 | |
| 580 | case pbresp.Canceled && pbresp.CompactRevision == 0: |
| 581 | delete(cancelSet, pbresp.WatchId) |
| 582 | if ws, ok := w.substreams[pbresp.WatchId]; ok { |
| 583 | // signal to stream goroutine to update closingc |
| 584 | close(ws.recvc) |
| 585 | closing[ws] = struct{}{} |
| 586 | } |
| 587 | |
| 588 | // reset for next iteration |
| 589 | cur = nil |
| 590 | |
| 591 | case cur.Fragment: |
| 592 | // watch response events are still fragmented |
| 593 | // continue to fetch next fragmented event arrival |
| 594 | continue |
| 595 | |
| 596 | default: |
| 597 | // dispatch to appropriate watch stream |
| 598 | ok := w.dispatchEvent(cur) |
| 599 | |
| 600 | // reset for next iteration |
| 601 | cur = nil |
| 602 | |
| 603 | if ok { |
| 604 | break |
| 605 | } |
| 606 | |
| 607 | // watch response on unexpected watch id; cancel id |
| 608 | if _, ok := cancelSet[pbresp.WatchId]; ok { |
| 609 | break |
| 610 | } |
| 611 | |
| 612 | cancelSet[pbresp.WatchId] = struct{}{} |
| 613 | cr := &pb.WatchRequest_CancelRequest{ |
| 614 | CancelRequest: &pb.WatchCancelRequest{ |
| 615 | WatchId: pbresp.WatchId, |
| 616 | }, |
| 617 | } |
| 618 | req := &pb.WatchRequest{RequestUnion: cr} |
| 619 | wc.Send(req) |
| 620 | } |
| 621 | |
| 622 | // watch client failed on Recv; spawn another if possible |
| 623 | case err := <-w.errc: |
| 624 | if isHaltErr(w.ctx, err) || toErr(w.ctx, err) == v3rpc.ErrNoLeader { |
| 625 | closeErr = err |
| 626 | return |
| 627 | } |
| 628 | if wc, closeErr = w.newWatchClient(); closeErr != nil { |
| 629 | return |
| 630 | } |
| 631 | if ws := w.nextResume(); ws != nil { |
| 632 | wc.Send(ws.initReq.toPB()) |
| 633 | } |
| 634 | cancelSet = make(map[int64]struct{}) |
| 635 | |
| 636 | case <-w.ctx.Done(): |
| 637 | return |
| 638 | |
| 639 | case ws := <-w.closingc: |
| 640 | w.closeSubstream(ws) |
| 641 | delete(closing, ws) |
| 642 | // no more watchers on this stream, shutdown |
| 643 | if len(w.substreams)+len(w.resuming) == 0 { |
| 644 | return |
| 645 | } |
| 646 | } |
| 647 | } |
| 648 | } |
| 649 | |
| 650 | // nextResume chooses the next resuming to register with the grpc stream. Abandoned |
| 651 | // streams are marked as nil in the queue since the head must wait for its inflight registration. |
| 652 | func (w *watchGrpcStream) nextResume() *watcherStream { |
| 653 | for len(w.resuming) != 0 { |
| 654 | if w.resuming[0] != nil { |
| 655 | return w.resuming[0] |
| 656 | } |
| 657 | w.resuming = w.resuming[1:len(w.resuming)] |
| 658 | } |
| 659 | return nil |
| 660 | } |
| 661 | |
| 662 | // dispatchEvent sends a WatchResponse to the appropriate watcher stream |
| 663 | func (w *watchGrpcStream) dispatchEvent(pbresp *pb.WatchResponse) bool { |
| 664 | events := make([]*Event, len(pbresp.Events)) |
| 665 | for i, ev := range pbresp.Events { |
| 666 | events[i] = (*Event)(ev) |
| 667 | } |
| 668 | // TODO: return watch ID? |
| 669 | wr := &WatchResponse{ |
| 670 | Header: *pbresp.Header, |
| 671 | Events: events, |
| 672 | CompactRevision: pbresp.CompactRevision, |
| 673 | Created: pbresp.Created, |
| 674 | Canceled: pbresp.Canceled, |
| 675 | cancelReason: pbresp.CancelReason, |
| 676 | } |
| 677 | |
| 678 | // watch IDs are zero indexed, so request notify watch responses are assigned a watch ID of -1 to |
| 679 | // indicate they should be broadcast. |
| 680 | if wr.IsProgressNotify() && pbresp.WatchId == -1 { |
| 681 | return w.broadcastResponse(wr) |
| 682 | } |
| 683 | |
| 684 | return w.unicastResponse(wr, pbresp.WatchId) |
| 685 | |
| 686 | } |
| 687 | |
| 688 | // broadcastResponse send a watch response to all watch substreams. |
| 689 | func (w *watchGrpcStream) broadcastResponse(wr *WatchResponse) bool { |
| 690 | for _, ws := range w.substreams { |
| 691 | select { |
| 692 | case ws.recvc <- wr: |
| 693 | case <-ws.donec: |
| 694 | } |
| 695 | } |
| 696 | return true |
| 697 | } |
| 698 | |
| 699 | // unicastResponse sends a watch response to a specific watch substream. |
| 700 | func (w *watchGrpcStream) unicastResponse(wr *WatchResponse, watchId int64) bool { |
| 701 | ws, ok := w.substreams[watchId] |
| 702 | if !ok { |
| 703 | return false |
| 704 | } |
| 705 | select { |
| 706 | case ws.recvc <- wr: |
| 707 | case <-ws.donec: |
| 708 | return false |
| 709 | } |
| 710 | return true |
| 711 | } |
| 712 | |
| 713 | // serveWatchClient forwards messages from the grpc stream to run() |
| 714 | func (w *watchGrpcStream) serveWatchClient(wc pb.Watch_WatchClient) { |
| 715 | for { |
| 716 | resp, err := wc.Recv() |
| 717 | if err != nil { |
| 718 | select { |
| 719 | case w.errc <- err: |
| 720 | case <-w.donec: |
| 721 | } |
| 722 | return |
| 723 | } |
| 724 | select { |
| 725 | case w.respc <- resp: |
| 726 | case <-w.donec: |
| 727 | return |
| 728 | } |
| 729 | } |
| 730 | } |
| 731 | |
| 732 | // serveSubstream forwards watch responses from run() to the subscriber |
| 733 | func (w *watchGrpcStream) serveSubstream(ws *watcherStream, resumec chan struct{}) { |
| 734 | if ws.closing { |
| 735 | panic("created substream goroutine but substream is closing") |
| 736 | } |
| 737 | |
| 738 | // nextRev is the minimum expected next revision |
| 739 | nextRev := ws.initReq.rev |
| 740 | resuming := false |
| 741 | defer func() { |
| 742 | if !resuming { |
| 743 | ws.closing = true |
| 744 | } |
| 745 | close(ws.donec) |
| 746 | if !resuming { |
| 747 | w.closingc <- ws |
| 748 | } |
| 749 | w.wg.Done() |
| 750 | }() |
| 751 | |
| 752 | emptyWr := &WatchResponse{} |
| 753 | for { |
| 754 | curWr := emptyWr |
| 755 | outc := ws.outc |
| 756 | |
| 757 | if len(ws.buf) > 0 { |
| 758 | curWr = ws.buf[0] |
| 759 | } else { |
| 760 | outc = nil |
| 761 | } |
| 762 | select { |
| 763 | case outc <- *curWr: |
| 764 | if ws.buf[0].Err() != nil { |
| 765 | return |
| 766 | } |
| 767 | ws.buf[0] = nil |
| 768 | ws.buf = ws.buf[1:] |
| 769 | case wr, ok := <-ws.recvc: |
| 770 | if !ok { |
| 771 | // shutdown from closeSubstream |
| 772 | return |
| 773 | } |
| 774 | |
| 775 | if wr.Created { |
| 776 | if ws.initReq.retc != nil { |
| 777 | ws.initReq.retc <- ws.outc |
| 778 | // to prevent next write from taking the slot in buffered channel |
| 779 | // and posting duplicate create events |
| 780 | ws.initReq.retc = nil |
| 781 | |
| 782 | // send first creation event only if requested |
| 783 | if ws.initReq.createdNotify { |
| 784 | ws.outc <- *wr |
| 785 | } |
| 786 | // once the watch channel is returned, a current revision |
| 787 | // watch must resume at the store revision. This is necessary |
| 788 | // for the following case to work as expected: |
| 789 | // wch := m1.Watch("a") |
| 790 | // m2.Put("a", "b") |
| 791 | // <-wch |
| 792 | // If the revision is only bound on the first observed event, |
| 793 | // if wch is disconnected before the Put is issued, then reconnects |
| 794 | // after it is committed, it'll miss the Put. |
| 795 | if ws.initReq.rev == 0 { |
| 796 | nextRev = wr.Header.Revision |
| 797 | } |
| 798 | } |
| 799 | } else { |
| 800 | // current progress of watch; <= store revision |
| 801 | nextRev = wr.Header.Revision |
| 802 | } |
| 803 | |
| 804 | if len(wr.Events) > 0 { |
| 805 | nextRev = wr.Events[len(wr.Events)-1].Kv.ModRevision + 1 |
| 806 | } |
| 807 | ws.initReq.rev = nextRev |
| 808 | |
| 809 | // created event is already sent above, |
| 810 | // watcher should not post duplicate events |
| 811 | if wr.Created { |
| 812 | continue |
| 813 | } |
| 814 | |
| 815 | // TODO pause channel if buffer gets too large |
| 816 | ws.buf = append(ws.buf, wr) |
| 817 | case <-w.ctx.Done(): |
| 818 | return |
| 819 | case <-ws.initReq.ctx.Done(): |
| 820 | return |
| 821 | case <-resumec: |
| 822 | resuming = true |
| 823 | return |
| 824 | } |
| 825 | } |
| 826 | // lazily send cancel message if events on missing id |
| 827 | } |
| 828 | |
| 829 | func (w *watchGrpcStream) newWatchClient() (pb.Watch_WatchClient, error) { |
| 830 | // mark all substreams as resuming |
| 831 | close(w.resumec) |
| 832 | w.resumec = make(chan struct{}) |
| 833 | w.joinSubstreams() |
| 834 | for _, ws := range w.substreams { |
| 835 | ws.id = -1 |
| 836 | w.resuming = append(w.resuming, ws) |
| 837 | } |
| 838 | // strip out nils, if any |
| 839 | var resuming []*watcherStream |
| 840 | for _, ws := range w.resuming { |
| 841 | if ws != nil { |
| 842 | resuming = append(resuming, ws) |
| 843 | } |
| 844 | } |
| 845 | w.resuming = resuming |
| 846 | w.substreams = make(map[int64]*watcherStream) |
| 847 | |
| 848 | // connect to grpc stream while accepting watcher cancelation |
| 849 | stopc := make(chan struct{}) |
| 850 | donec := w.waitCancelSubstreams(stopc) |
| 851 | wc, err := w.openWatchClient() |
| 852 | close(stopc) |
| 853 | <-donec |
| 854 | |
| 855 | // serve all non-closing streams, even if there's a client error |
| 856 | // so that the teardown path can shutdown the streams as expected. |
| 857 | for _, ws := range w.resuming { |
| 858 | if ws.closing { |
| 859 | continue |
| 860 | } |
| 861 | ws.donec = make(chan struct{}) |
| 862 | w.wg.Add(1) |
| 863 | go w.serveSubstream(ws, w.resumec) |
| 864 | } |
| 865 | |
| 866 | if err != nil { |
| 867 | return nil, v3rpc.Error(err) |
| 868 | } |
| 869 | |
| 870 | // receive data from new grpc stream |
| 871 | go w.serveWatchClient(wc) |
| 872 | return wc, nil |
| 873 | } |
| 874 | |
| 875 | func (w *watchGrpcStream) waitCancelSubstreams(stopc <-chan struct{}) <-chan struct{} { |
| 876 | var wg sync.WaitGroup |
| 877 | wg.Add(len(w.resuming)) |
| 878 | donec := make(chan struct{}) |
| 879 | for i := range w.resuming { |
| 880 | go func(ws *watcherStream) { |
| 881 | defer wg.Done() |
| 882 | if ws.closing { |
| 883 | if ws.initReq.ctx.Err() != nil && ws.outc != nil { |
| 884 | close(ws.outc) |
| 885 | ws.outc = nil |
| 886 | } |
| 887 | return |
| 888 | } |
| 889 | select { |
| 890 | case <-ws.initReq.ctx.Done(): |
| 891 | // closed ws will be removed from resuming |
| 892 | ws.closing = true |
| 893 | close(ws.outc) |
| 894 | ws.outc = nil |
| 895 | w.wg.Add(1) |
| 896 | go func() { |
| 897 | defer w.wg.Done() |
| 898 | w.closingc <- ws |
| 899 | }() |
| 900 | case <-stopc: |
| 901 | } |
| 902 | }(w.resuming[i]) |
| 903 | } |
| 904 | go func() { |
| 905 | defer close(donec) |
| 906 | wg.Wait() |
| 907 | }() |
| 908 | return donec |
| 909 | } |
| 910 | |
| 911 | // joinSubstreams waits for all substream goroutines to complete. |
| 912 | func (w *watchGrpcStream) joinSubstreams() { |
| 913 | for _, ws := range w.substreams { |
| 914 | <-ws.donec |
| 915 | } |
| 916 | for _, ws := range w.resuming { |
| 917 | if ws != nil { |
| 918 | <-ws.donec |
| 919 | } |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | var maxBackoff = 100 * time.Millisecond |
| 924 | |
| 925 | // openWatchClient retries opening a watch client until success or halt. |
| 926 | // manually retry in case "ws==nil && err==nil" |
| 927 | // TODO: remove FailFast=false |
| 928 | func (w *watchGrpcStream) openWatchClient() (ws pb.Watch_WatchClient, err error) { |
| 929 | backoff := time.Millisecond |
| 930 | for { |
| 931 | select { |
| 932 | case <-w.ctx.Done(): |
| 933 | if err == nil { |
| 934 | return nil, w.ctx.Err() |
| 935 | } |
| 936 | return nil, err |
| 937 | default: |
| 938 | } |
| 939 | if ws, err = w.remote.Watch(w.ctx, w.callOpts...); ws != nil && err == nil { |
| 940 | break |
| 941 | } |
| 942 | if isHaltErr(w.ctx, err) { |
| 943 | return nil, v3rpc.Error(err) |
| 944 | } |
| 945 | if isUnavailableErr(w.ctx, err) { |
| 946 | // retry, but backoff |
| 947 | if backoff < maxBackoff { |
| 948 | // 25% backoff factor |
| 949 | backoff = backoff + backoff/4 |
| 950 | if backoff > maxBackoff { |
| 951 | backoff = maxBackoff |
| 952 | } |
| 953 | } |
| 954 | time.Sleep(backoff) |
| 955 | } |
| 956 | } |
| 957 | return ws, nil |
| 958 | } |
| 959 | |
| 960 | // toPB converts an internal watch request structure to its protobuf WatchRequest structure. |
| 961 | func (wr *watchRequest) toPB() *pb.WatchRequest { |
| 962 | req := &pb.WatchCreateRequest{ |
| 963 | StartRevision: wr.rev, |
| 964 | Key: []byte(wr.key), |
| 965 | RangeEnd: []byte(wr.end), |
| 966 | ProgressNotify: wr.progressNotify, |
| 967 | Filters: wr.filters, |
| 968 | PrevKv: wr.prevKV, |
| 969 | Fragment: wr.fragment, |
| 970 | } |
| 971 | cr := &pb.WatchRequest_CreateRequest{CreateRequest: req} |
| 972 | return &pb.WatchRequest{RequestUnion: cr} |
| 973 | } |
| 974 | |
| 975 | // toPB converts an internal progress request structure to its protobuf WatchRequest structure. |
| 976 | func (pr *progressRequest) toPB() *pb.WatchRequest { |
| 977 | req := &pb.WatchProgressRequest{} |
| 978 | cr := &pb.WatchRequest_ProgressRequest{ProgressRequest: req} |
| 979 | return &pb.WatchRequest{RequestUnion: cr} |
| 980 | } |
| 981 | |
| 982 | func streamKeyFromCtx(ctx context.Context) string { |
| 983 | if md, ok := metadata.FromOutgoingContext(ctx); ok { |
| 984 | return fmt.Sprintf("%+v", md) |
| 985 | } |
| 986 | return "" |
| 987 | } |