Golang Stopper.RunWorker方法代码示例

// maybeWarnAboutInit looks for signs indicating a cluster which
// hasn't been initialized and warns. There's no absolutely sure way
// to determine whether the current node is simply waiting to be
// bootstrapped to an existing cluster vs. the operator having failed
// to initialize the cluster via the "cockroach init" command, so
// we can only warn.
//
// This method checks whether all gossip bootstrap hosts are
// connected, and whether the node itself is a bootstrap host, but
// there is still no sentinel gossip.
func (g *Gossip) maybeWarnAboutInit(stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		// Wait 5s before first check.
		select {
		case <-stopper.ShouldStop():
			return
		case <-time.After(5 * time.Second):
		}
		retryOptions := retry.Options{
			InitialBackoff: 5 * time.Second,      // first backoff at 5s
			MaxBackoff:     60 * time.Second,     // max backoff is 60s
			Multiplier:     2,                    // doubles
			Closer:         stopper.ShouldStop(), // stop no matter what on stopper
		}
		// This will never error because of infinite retries.
		for r := retry.Start(retryOptions); r.Next(); {
			g.mu.Lock()
			hasConnections := g.outgoing.len()+g.incoming.len() > 0
			hasSentinel := g.is.getInfo(KeySentinel) != nil
			triedAll := g.triedAll
			g.mu.Unlock()
			// If we have the sentinel, exit the retry loop.
			if hasSentinel {
				break
			}
			if !hasConnections {
				log.Warningf("not connected to gossip; check that gossip flag is set appropriately")
			} else if triedAll {
				log.Warningf("missing gossip sentinel; first range unavailable or cluster not initialized")
			}
		}
	})
}

// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(g *Gossip, disconnected chan *client, ctx *rpc.Context, stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		defer func() {
			disconnected <- c
		}()

		// Note: avoid using `grpc.WithBlock` here. This code is already
		// asynchronous from the caller's perspective, so the only effect of
		// `WithBlock` here is blocking shutdown - at the time of this writing,
		// that ends ups up making `kv` tests take twice as long.
		conn, err := ctx.GRPCDial(c.addr.String())
		if err != nil {
			log.Errorf("failed to dial: %v", err)
			return
		}

		// Start gossiping.
		if err := c.gossip(g, NewGossipClient(conn), stopper); err != nil {
			if !grpcutil.IsClosedConnection(err) {
				g.mu.Lock()
				peerID := c.peerID
				g.mu.Unlock()
				if peerID != 0 {
					log.Infof("closing client to node %d (%s): %s", peerID, c.addr, err)
				} else {
					log.Infof("closing client to %s: %s", c.addr, err)
				}
			}
		}
	})
}

// start will run continuously and mark stores as offline if they haven't been
// heard from in longer than timeUntilStoreDead.
func (sp *StorePool) start(stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		for {
			var timeout time.Duration
			sp.mu.Lock()
			detail := sp.queue.peek()
			if detail == nil {
				// No stores yet, wait the full timeout.
				timeout = sp.timeUntilStoreDead
			} else {
				// Check to see if the store should be marked as dead.
				deadAsOf := detail.lastUpdatedTime.GoTime().Add(sp.timeUntilStoreDead)
				now := sp.clock.Now()
				if now.GoTime().After(deadAsOf) {
					deadDetail := sp.queue.dequeue()
					deadDetail.markDead(now)
					// The next store might be dead as well, set the timeout to
					// 0 to process it immediately.
					timeout = 0
				} else {
					// Store is still alive, schedule the next check for when
					// it should timeout.
					timeout = deadAsOf.Sub(now.GoTime())
				}
			}
			sp.mu.Unlock()
			select {
			case <-time.After(timeout):
			case <-stopper.ShouldStop():
				return
			}
		}
	})
}

// start dials the remote addr and commences gossip once connected.
// Upon exit, signals client is done by pushing it onto the done
// channel. If the client experienced an error, its err field will
// be set. This method starts client processing in a goroutine and
// returns immediately.
func (c *client) start(g *Gossip, done chan *client, context *rpc.Context, stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		var err error

		c.rpcClient = rpc.NewClient(c.addr, context)
		select {
		case <-c.rpcClient.Healthy():
			// Start gossiping and wait for disconnect or error.
			err = c.gossip(g, stopper)
			if context.DisableCache {
				c.rpcClient.Close()
			}
		case <-c.rpcClient.Closed:
			err = util.Errorf("client closed")
		}

		done <- c

		if err != nil {
			if c.peerID != 0 {
				log.Infof("closing client to node %d (%s): %s", c.peerID, c.addr, err)
			} else {
				log.Infof("closing client to %s: %s", c.addr, err)
			}
		}
	})
}

// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(g *Gossip, disconnected chan *client, ctx *rpc.Context, stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		defer func() {
			disconnected <- c
		}()

		conn, err := ctx.GRPCDial(c.addr.String(), grpc.WithBlock())
		if err != nil {
			log.Errorf("failed to dial: %v", err)
			return
		}

		// Start gossiping.
		if err := c.gossip(g, NewGossipClient(conn), stopper); err != nil {
			if !grpcutil.IsClosedConnection(err) {
				g.mu.Lock()
				peerID := c.peerID
				g.mu.Unlock()
				if peerID != 0 {
					log.Infof("closing client to node %d (%s): %s", peerID, c.addr, err)
				} else {
					log.Infof("closing client to %s: %s", c.addr, err)
				}
			}
		}
	})
}

// NewExecutor creates an Executor and registers a callback on the
// system config.
func NewExecutor(db client.DB, gossip *gossip.Gossip, leaseMgr *LeaseManager, metaRegistry *metric.Registry, stopper *stop.Stopper) *Executor {
	exec := &Executor{
		db:       db,
		reCache:  parser.NewRegexpCache(512),
		leaseMgr: leaseMgr,

		latency: metaRegistry.Latency("sql.latency"),
	}
	exec.systemConfigCond = sync.NewCond(&exec.systemConfigMu)

	gossipUpdateC := gossip.RegisterSystemConfigChannel()
	stopper.RunWorker(func() {
		for {
			select {
			case <-gossipUpdateC:
				cfg := gossip.GetSystemConfig()
				exec.updateSystemConfig(cfg)
			case <-stopper.ShouldStop():
				return
			}
		}
	})

	return exec
}

// NewExecutor creates an Executor and registers a callback on the
// system config.
func NewExecutor(db client.DB, gossip *gossip.Gossip, leaseMgr *LeaseManager, stopper *stop.Stopper) *Executor {
	registry := metric.NewRegistry()
	exec := &Executor{
		db:       db,
		reCache:  parser.NewRegexpCache(512),
		leaseMgr: leaseMgr,

		registry:      registry,
		latency:       registry.Latency("latency"),
		txnBeginCount: registry.Counter("transaction.begincount"),
		selectCount:   registry.Counter("select.count"),
		updateCount:   registry.Counter("update.count"),
		insertCount:   registry.Counter("insert.count"),
		deleteCount:   registry.Counter("delete.count"),
		ddlCount:      registry.Counter("ddl.count"),
		miscCount:     registry.Counter("misc.count"),
	}
	exec.systemConfigCond = sync.NewCond(&exec.systemConfigMu)

	gossipUpdateC := gossip.RegisterSystemConfigChannel()
	stopper.RunWorker(func() {
		for {
			select {
			case <-gossipUpdateC:
				cfg := gossip.GetSystemConfig()
				exec.updateSystemConfig(cfg)
			case <-stopper.ShouldStop():
				return
			}
		}
	})

	return exec
}

// bootstrap connects the node to the gossip network. Bootstrapping
// commences in the event there are no connected clients or the
// sentinel gossip info is not available. After a successful bootstrap
// connection, this method will block on the stalled condvar, which
// receives notifications that gossip network connectivity has been
// lost and requires re-bootstrapping.
func (g *Gossip) bootstrap(stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		for {
			g.mu.Lock()
			if g.closed {
				g.mu.Unlock()
				return
			}
			// Check whether or not we need bootstrap.
			haveClients := g.outgoing.len() > 0
			haveSentinel := g.is.getInfo(KeySentinel) != nil
			if !haveClients || !haveSentinel {
				// Try to get another bootstrap address from the resolvers.
				if addr := g.getNextBootstrapAddress(); addr != nil {
					g.startClient(addr, g.bsRPCContext, stopper)
				}
			}
			g.mu.Unlock()

			// Block until we need bootstrapping again.
			select {
			case <-g.stalled:
				// continue
			case <-stopper.ShouldStop():
				return
			}
		}
	})
}

// startGossip loops on a periodic ticker to gossip node-related
// information. Starts a goroutine to loop until the node is closed.
func (n *Node) startGossip(ctx context.Context, stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		gossipStoresInterval := envutil.EnvOrDefaultDuration("gossip_stores_interval",
			gossip.DefaultGossipStoresInterval)
		statusTicker := time.NewTicker(gossipStatusInterval)
		storesTicker := time.NewTicker(gossipStoresInterval)
		nodeTicker := time.NewTicker(gossipNodeDescriptorInterval)
		defer storesTicker.Stop()
		defer nodeTicker.Stop()
		n.gossipStores(ctx) // one-off run before going to sleep
		for {
			select {
			case <-statusTicker.C:
				n.ctx.Gossip.LogStatus()
			case <-storesTicker.C:
				n.gossipStores(ctx)
			case <-nodeTicker.C:
				if err := n.ctx.Gossip.SetNodeDescriptor(&n.Descriptor); err != nil {
					log.Warningf(ctx, "couldn't gossip descriptor for node %d: %s", n.Descriptor.NodeID, err)
				}
			case <-stopper.ShouldStop():
				return
			}
		}
	})
}

// NewExecutor creates an Executor and registers a callback on the
// system config.
func NewExecutor(ctx ExecutorContext, stopper *stop.Stopper, registry *metric.Registry) *Executor {
	exec := &Executor{
		ctx:     ctx,
		reCache: parser.NewRegexpCache(512),

		registry:         registry,
		latency:          registry.Latency("latency"),
		txnBeginCount:    registry.Counter("txn.begin.count"),
		txnCommitCount:   registry.Counter("txn.commit.count"),
		txnAbortCount:    registry.Counter("txn.abort.count"),
		txnRollbackCount: registry.Counter("txn.rollback.count"),
		selectCount:      registry.Counter("select.count"),
		updateCount:      registry.Counter("update.count"),
		insertCount:      registry.Counter("insert.count"),
		deleteCount:      registry.Counter("delete.count"),
		ddlCount:         registry.Counter("ddl.count"),
		miscCount:        registry.Counter("misc.count"),
	}
	exec.systemConfigCond = sync.NewCond(exec.systemConfigMu.RLocker())

	gossipUpdateC := ctx.Gossip.RegisterSystemConfigChannel()
	stopper.RunWorker(func() {
		for {
			select {
			case <-gossipUpdateC:
				cfg, _ := ctx.Gossip.GetSystemConfig()
				exec.updateSystemConfig(cfg)
			case <-stopper.ShouldStop():
				return
			}
		}
	})

	return exec
}

// NewContext creates an rpc Context with the supplied values.
func NewContext(baseCtx *base.Context, clock *hlc.Clock, stopper *stop.Stopper) *Context {
	var ctx *Context
	if baseCtx != nil {
		// TODO(tamird): This form fools `go vet`; `baseCtx` contains several
		// `sync.Mutex`s, and this deference copies them, which is bad. The problem
		// predates this comment, so I'm kicking the can down the road for now, but
		// we should fix this.
		ctx = &Context{
			Context: *baseCtx,
		}
	} else {
		ctx = new(Context)
	}
	if clock != nil {
		ctx.localClock = clock
	} else {
		ctx.localClock = hlc.NewClock(hlc.UnixNano)
	}
	ctx.Stopper = stopper
	ctx.RemoteClocks = newRemoteClockMonitor(clock)
	ctx.HeartbeatInterval = defaultHeartbeatInterval
	ctx.HeartbeatTimeout = 2 * defaultHeartbeatInterval

	stopper.RunWorker(func() {
		<-stopper.ShouldDrain()

		ctx.conns.Lock()
		for key, conn := range ctx.conns.cache {
			ctx.removeConn(key, conn)
		}
		ctx.conns.Unlock()
	})

	return ctx
}

// gossip loops, sending deltas of the infostore and receiving deltas
// in turn. If an alternate is proposed on response, the client addr
// is modified and method returns for forwarding by caller.
func (c *client) gossip(g *Gossip, gossipClient GossipClient, stopper *stop.Stopper) error {
	// For un-bootstrapped node, g.is.NodeID is 0 when client start gossip,
	// so it's better to get nodeID from g.is every time.
	g.mu.Lock()
	addr := g.is.NodeAddr
	g.mu.Unlock()

	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()
	stream, err := gossipClient.Gossip(ctx)
	if err != nil {
		return err
	}

	if err := c.requestGossip(g, addr, stream); err != nil {
		return err
	}

	sendGossipChan := make(chan struct{}, 1)

	// Register a callback for gossip updates.
	updateCallback := func(_ string, _ roachpb.Value) {
		select {
		case sendGossipChan <- struct{}{}:
		default:
		}
	}
	// Defer calling "undoer" callback returned from registration.
	defer g.RegisterCallback(".*", updateCallback)()

	errCh := make(chan error, 1)
	stopper.RunWorker(func() {
		errCh <- func() error {
			for {
				reply, err := stream.Recv()
				if err != nil {
					return err
				}
				if err := c.handleResponse(g, reply); err != nil {
					return err
				}
			}
		}()
	})

	for {
		select {
		case <-c.closer:
			return nil
		case <-stopper.ShouldStop():
			return nil
		case err := <-errCh:
			return err
		case <-sendGossipChan:
			if err := c.sendGossip(g, addr, stream); err != nil {
				return err
			}
		}
	}
}

// maybeWarnAboutInit looks for signs indicating a cluster which
// hasn't been initialized and warns. There's no absolutely sure way
// to determine whether the current node is simply waiting to be
// bootstrapped to an existing cluster vs. the operator having failed
// to initialize the cluster via the "cockroach init" command, so
// we can only warn.
//
// This method checks whether all gossip bootstrap hosts are
// connected, and whether the node itself is a bootstrap host, but
// there is still no sentinel gossip.
func (g *Gossip) maybeWarnAboutInit(stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		// Wait 5s before first check.
		select {
		case <-stopper.ShouldStop():
			return
		case <-time.After(5 * time.Second):
		}
		retryOptions := retry.Options{
			InitialBackoff: 5 * time.Second,  // first backoff at 5s
			MaxBackoff:     60 * time.Second, // max backoff is 60s
			Multiplier:     2,                // doubles
			Stopper:        stopper,          // stop no matter what on stopper
		}
		// This will never error because of infinite retries.
		for r := retry.Start(retryOptions); r.Next(); {
			g.mu.Lock()
			hasSentinel := g.is.getInfo(KeySentinel) != nil
			triedAll := g.triedAll
			g.mu.Unlock()
			// If we have the sentinel, exit the retry loop.
			if hasSentinel {
				break
			}
			// Otherwise, if all bootstrap hosts are connected, warn.
			if triedAll {
				log.Warningf("connected to gossip but missing sentinel. Has the cluster been initialized? " +
					"Use \"cockroach init\" to initialize.")
			}
		}
	})
}

// start initializes the infostore with the rpc server address and
// then begins processing connecting clients in an infinite select
// loop via goroutine. Periodically, clients connected and awaiting
// the next round of gossip are awoken via the conditional variable.
func (s *server) start(rpcServer *rpc.Server, stopper *stop.Stopper) {
	addr := rpcServer.Addr()
	s.is.NodeAddr = util.MakeUnresolvedAddr(addr.Network(), addr.String())
	if err := rpcServer.Register("Gossip.Gossip", s.Gossip, &Request{}); err != nil {
		log.Fatalf("unable to register gossip service with RPC server: %s", err)
	}
	rpcServer.AddCloseCallback(s.onClose)

	updateCallback := func(_ string, _ roachpb.Value) {
		// Wakeup all pending clients.
		s.ready.Broadcast()
	}
	unregister := s.is.registerCallback(".*", updateCallback)

	stopper.RunWorker(func() {
		// Periodically wakeup blocked client gossip requests.
		for {
			select {
			case <-stopper.ShouldStop():
				s.stop(unregister)
				return
			}
		}
	})
}

// RefreshLeases starts a goroutine that refreshes the lease manager
// leases for tables received in the latest system configuration via gossip.
func (m *LeaseManager) RefreshLeases(s *stop.Stopper, db *client.DB, gossip *gossip.Gossip) {
	s.RunWorker(func() {
		descKeyPrefix := keys.MakeTablePrefix(uint32(sqlbase.DescriptorTable.ID))
		gossipUpdateC := gossip.RegisterSystemConfigChannel()
		for {
			select {
			case <-gossipUpdateC:
				cfg, _ := gossip.GetSystemConfig()
				if m.testingKnobs.GossipUpdateEvent != nil {
					m.testingKnobs.GossipUpdateEvent(cfg)
				}
				// Read all tables and their versions
				if log.V(2) {
					log.Info("received a new config; will refresh leases")
				}

				// Loop through the configuration to find all the tables.
				for _, kv := range cfg.Values {
					if !bytes.HasPrefix(kv.Key, descKeyPrefix) {
						continue
					}
					// Attempt to unmarshal config into a table/database descriptor.
					var descriptor sqlbase.Descriptor
					if err := kv.Value.GetProto(&descriptor); err != nil {
						log.Warningf("%s: unable to unmarshal descriptor %v", kv.Key, kv.Value)
						continue
					}
					switch union := descriptor.Union.(type) {
					case *sqlbase.Descriptor_Table:
						table := union.Table
						if err := table.Validate(); err != nil {
							log.Errorf("%s: received invalid table descriptor: %v", kv.Key, table)
							continue
						}
						if log.V(2) {
							log.Infof("%s: refreshing lease table: %d (%s), version: %d",
								kv.Key, table.ID, table.Name, table.Version)
						}
						// Try to refresh the table lease to one >= this version.
						if t := m.findTableState(table.ID, false /* create */, nil); t != nil {
							if err := t.purgeOldLeases(
								db, table.Deleted(), table.Version, m.LeaseStore); err != nil {
								log.Warningf("error purging leases for table %d(%s): %s",
									table.ID, table.Name, err)
							}
						}
					case *sqlbase.Descriptor_Database:
						// Ignore.
					}
				}
				if m.testingKnobs.TestingLeasesRefreshedEvent != nil {
					m.testingKnobs.TestingLeasesRefreshedEvent(cfg)
				}

			case <-s.ShouldStop():
				return
			}
		}
	})
}