feature: support graceful shutdown

If connected to Tarantool 2.10 or newer and WatchersFeature is required,
after this patch connection supports server graceful shutdown [1]. In
this case, server will wait until all client requests will be finished
and client disconnects before going down (server also may go down by
timeout). Client reconnect will happen if connection options enable
reconnect.

1. https://www.tarantool.io/en/doc/latest/dev_guide/internals/iproto/graceful_shutdown/

Closes #214

Author: DifferentialOrange

CHANGELOG.md

@@ -15,6 +15,7 @@ Versioning](http://semver.org/spec/v2.0.0.html) except to the first release.
 - Error type support in MessagePack (#209)
 - Event subscription support (#119)
 - Session settings support (#215)
+- Support graceful shutdown (#214)
 
 ### Changed
 

connection.go

@@ -25,14 +25,17 @@ const ignoreStreamId = 0
 const (
 	connDisconnected = 0
 	connConnected    = 1
-	connClosed       = 2
+	connShutdown     = 2
+	connClosed       = 3
 )
 
 const (
 	connTransportNone = ""
 	connTransportSsl  = "ssl"
 )
 
+const shutdownEventKey = "box.shutdown"
+
 type ConnEventKind int
 type ConnLogKind int
 
@@ -45,6 +48,8 @@ const (
 	ReconnectFailed
 	// Either reconnect attempts exhausted, or explicit Close is called.
 	Closed
+	// Shutdown signals that shutdown callback is processing.
+	Shutdown
 
 	// LogReconnectFailed is logged when reconnect attempt failed.
 	LogReconnectFailed ConnLogKind = iota + 1
@@ -134,10 +139,19 @@ func (d defaultLogger) Report(event ConnLogKind, conn *Connection, v ...interfac
 // always returns array of array (array of tuples for space related methods).
 // For Eval* and Call* Tarantool always returns array, but does not forces
 // array of arrays.
+//
+// If connected to Tarantool 2.10 or newer and WatchersFeature is required,
+// connection supports server graceful shutdown. In this case, server will
+// wait until all client requests will be finished and client disconnects
+// before going down (server also may go down by timeout). Client reconnect will
+// happen if connection options enable reconnect.
+//
+// More on graceful shutdown: https://www.tarantool.io/en/doc/latest/dev_guide/internals/iproto/graceful_shutdown/
 type Connection struct {
 	addr  string
 	c     net.Conn
 	mutex sync.Mutex
+	cond  *sync.Cond
 	// Schema contains schema loaded on connection.
 	Schema *Schema
 	// requestId contains the last request ID for requests with nil context.
@@ -162,6 +176,11 @@ type Connection struct {
 	serverProtocolInfo ProtocolInfo
 	// watchMap is a map of key -> chan watchState.
 	watchMap sync.Map
+
+	// shutdownWatcher is the "box.shutdown" event watcher.
+	shutdownWatcher Watcher
+	// requestCnt is a counter of active requests.
+	requestCnt int64
 }
 
 var _ = Connector(&Connection{}) // Check compatibility with connector interface.
@@ -385,6 +404,8 @@ func Connect(addr string, opts Opts) (conn *Connection, err error) {
 		conn.opts.Logger = defaultLogger{}
 	}
 
+	conn.cond = sync.NewCond(&conn.mutex)
+
 	if err = conn.createConnection(false); err != nil {
 		ter, ok := err.(Error)
 		if conn.opts.Reconnect <= 0 {
@@ -421,6 +442,16 @@ func Connect(addr string, opts Opts) (conn *Connection, err error) {
 		}
 	}
 
+	// Subscribe shutdown event to process graceful shutdown.
+	if conn.isWatchersRequired() {
+		watcher, werr := conn.NewWatcher(shutdownEventKey, shutdownEventCallback)
+		if werr != nil {
+			conn.closeConnection(werr, true)
+			return nil, werr
+		}
+		conn.shutdownWatcher = watcher
+	}
+
 	return conn, err
 }
 
@@ -589,6 +620,7 @@ func (conn *Connection) dial() (err error) {
 	conn.lockShards()
 	conn.c = connection
 	atomic.StoreUint32(&conn.state, connConnected)
+	conn.cond.Broadcast()
 	conn.unlockShards()
 	go conn.writer(w, connection)
 	go conn.reader(r, connection)
@@ -762,10 +794,17 @@ func (conn *Connection) closeConnection(neterr error, forever bool) (err error)
 		if conn.state != connClosed {
 			close(conn.control)
 			atomic.StoreUint32(&conn.state, connClosed)
+			conn.cond.Broadcast()
+			// Free the resources.
+			if conn.shutdownWatcher != nil {
+				go conn.shutdownWatcher.Unregister()
+				conn.shutdownWatcher = nil
+			}
 			conn.notify(Closed)
 		}
 	} else {
 		atomic.StoreUint32(&conn.state, connDisconnected)
+		conn.cond.Broadcast()
 		conn.notify(Disconnected)
 	}
 	if conn.c != nil {
@@ -784,9 +823,7 @@ func (conn *Connection) closeConnection(neterr error, forever bool) (err error)
 	return
 }
 
-func (conn *Connection) reconnect(neterr error, c net.Conn) {
-	conn.mutex.Lock()
-	defer conn.mutex.Unlock()
+func (conn *Connection) reconnectImpl(neterr error, c net.Conn) {
 	if conn.opts.Reconnect > 0 {
 		if c == conn.c {
 			conn.closeConnection(neterr, false)
@@ -799,6 +836,13 @@ func (conn *Connection) reconnect(neterr error, c net.Conn) {
 	}
 }
 
+func (conn *Connection) reconnect(neterr error, c net.Conn) {
+	conn.mutex.Lock()
+	defer conn.mutex.Unlock()
+	conn.reconnectImpl(neterr, c)
+	conn.cond.Broadcast()
+}
+
 func (conn *Connection) lockShards() {
 	for i := range conn.shard {
 		conn.shard[i].rmut.Lock()
@@ -1026,6 +1070,15 @@ func (conn *Connection) newFuture(ctx context.Context) (fut *Future) {
 		fut.done = nil
 		shard.rmut.Unlock()
 		return
+	case connShutdown:
+		fut.err = ClientError{
+			ErrConnectionShutdown,
+			"server shutdown in progress",
+		}
+		fut.ready = nil
+		fut.done = nil
+		shard.rmut.Unlock()
+		return
 	}
 	pos := (fut.requestId / conn.opts.Concurrency) & (requestsMap - 1)
 	if ctx != nil {
@@ -1086,6 +1139,7 @@ func (conn *Connection) send(req Request, streamId uint64) *Future {
 	if fut.ready == nil {
 		return fut
 	}
+
 	if req.Ctx() != nil {
 		select {
 		case <-req.Ctx().Done():
@@ -1094,10 +1148,17 @@ func (conn *Connection) send(req Request, streamId uint64) *Future {
 		default:
 		}
 	}
+
+	if conn.shutdownWatcher != nil {
+		atomic.AddInt64(&(conn.requestCnt), int64(1))
+	}
+
 	conn.putFuture(fut, req, streamId)
+
 	if req.Ctx() != nil {
 		go conn.contextWatchdog(fut, req.Ctx())
 	}
+
 	return fut
 }
 
@@ -1164,6 +1225,12 @@ func (conn *Connection) markDone(fut *Future) {
 	if conn.rlimit != nil {
 		<-conn.rlimit
 	}
+
+	if conn.shutdownWatcher != nil {
+		if atomic.AddInt64(&(conn.requestCnt), int64(-1)) == 0 {
+			conn.cond.Broadcast()
+		}
+	}
 }
 
 func (conn *Connection) peekFuture(reqid uint32) (fut *Future) {
@@ -1458,6 +1525,15 @@ func subscribeWatchChannel(conn *Connection, key string) (chan watchState, error
 	return st, nil
 }
 
+func (conn *Connection) isWatchersRequired() bool {
+	for _, feature := range conn.opts.RequiredProtocolInfo.Features {
+		if feature == WatchersFeature {
+			return true
+		}
+	}
+	return false
+}
+
 // NewWatcher creates a new Watcher object for the connection.
 //
 // You need to require WatchersFeature to use watchers, see examples for the
@@ -1496,15 +1572,7 @@ func (conn *Connection) NewWatcher(key string, callback WatchCallback) (Watcher,
 	// asynchronous. We do not expect any response from a Tarantool instance
 	// That's why we can't just check the Tarantool response for an unsupported
 	// request error.
-	watchersRequired := false
-	for _, feature := range conn.opts.RequiredProtocolInfo.Features {
-		if feature == WatchersFeature {
-			watchersRequired = true
-			break
-		}
-	}
-
-	if !watchersRequired {
+	if !conn.isWatchersRequired() {
 		err := fmt.Errorf("the feature %s must be required by connection "+
 			"options to create a watcher", WatchersFeature)
 		return nil, err
@@ -1563,7 +1631,11 @@ func (conn *Connection) NewWatcher(key string, callback WatchCallback) (Watcher,
 
 				if state.cnt == 0 {
 					// The last one sends IPROTO_UNWATCH.
-					conn.Do(newUnwatchRequest(key)).Get()
+					if !conn.ClosedNow() {
+						// conn.ClosedNow() check is a workaround for calling
+						// Unregister from connectionClose().
+						conn.Do(newUnwatchRequest(key)).Get()
+					}
 					conn.watchMap.Delete(key)
 					close(state.unready)
 				}
@@ -1666,3 +1738,52 @@ func (conn *Connection) ServerProtocolInfo() ProtocolInfo {
 func (conn *Connection) ClientProtocolInfo() ProtocolInfo {
 	return clientProtocolInfo.Clone()
 }
+
+func shutdownEventCallback(event WatchEvent) {
+	// Receives "true" on server shutdown.
+	// See https://www.tarantool.io/en/doc/latest/dev_guide/internals/iproto/graceful_shutdown/
+	// step 2.
+	val, ok := event.Value.(bool)
+	if ok && val {
+		go event.Conn.processShutdown()
+	}
+}
+
+func (conn *Connection) processShutdown() {
+	// Forbid state changes.
+	conn.mutex.Lock()
+	defer conn.mutex.Unlock()
+
+	if !atomic.CompareAndSwapUint32(&(conn.state), connConnected, connShutdown) {
+		return
+	}
+	conn.notify(Shutdown)
+
+	c := conn.c
+	for (atomic.LoadUint32(&(conn.state)) == connShutdown) &&
+		(atomic.LoadInt64(&(conn.requestCnt)) != 0) &&
+		(c == conn.c) {
+		// Use cond var on conn.mutex since request execution may
+		// call reconnect(). It is ok if state changes as part of
+		// reconnect since Tarantool server won't allow to reconnect
+		// in the middle of shutting down.
+		conn.cond.Wait()
+	}
+	// Do not unregister task explicitly here since connection teardown
+	// has the same effect. To clean up connection resources,
+	// unregister on full close.
+
+	if (atomic.LoadUint32(&(conn.state)) == connShutdown) &&
+		(c == conn.c) {
+		// Start to reconnect based on common rules, same as in net.box.
+		// Reconnect also closes the connection: server waits until all
+		// subscribed connections are terminated.
+		// See https://www.tarantool.io/en/doc/latest/dev_guide/internals/iproto/graceful_shutdown/
+		// step 3.
+		conn.reconnectImpl(
+			ClientError{
+				ErrConnectionClosed,
+				"connection closed after server shutdown",
+			}, conn.c)
+	}
+}

errors.go

@@ -55,6 +55,7 @@ const (
 	ErrProtocolError      = 0x4000 + iota
 	ErrTimeouted          = 0x4000 + iota
 	ErrRateLimited        = 0x4000 + iota
+	ErrConnectionShutdown = 0x4000 + iota
 )
 
 // Tarantool server error codes.