标签：Swarmkit

raft.Node有一个memoryStore成员（定义在manager/state/raft/raft.go）：

// Node represents the Raft Node useful
// configuration.
type Node struct {
    ......
    memoryStore         *store.MemoryStore
    ......
}

它非常重要，因为cluster中用来响应swarmctl命令的manager leader中的store成员其实就是指向manager中Node结构体中的memoryStore：

// Server is the Cluster API gRPC server.
type Server struct {
    store  *store.MemoryStore
    raft   *raft.Node
    rootCA *ca.RootCA
}

store.MemoryStore定义在manager/state/store/memory.go：

// MemoryStore is a concurrency-safe, in-memory implementation of the Store
// interface.
type MemoryStore struct {
    // updateLock must be held during an update transaction.
    updateLock sync.Mutex

    memDB *memdb.MemDB
    queue *watch.Queue

    proposer state.Proposer
}

其中实际用来存储的memory database部分使用的是go-memdb项目。初始化store.MemoryStore使用NewMemoryStore()函数：

// NewMemoryStore returns an in-memory store. The argument is an optional
// Proposer which will be used to propagate changes to other members in a
// cluster.
func NewMemoryStore(proposer state.Proposer) *MemoryStore {
    memDB, err := memdb.NewMemDB(schema)
    if err != nil {
        // This shouldn't fail
        panic(err)
    }

    return &MemoryStore{
        memDB:    memDB,
        queue:    watch.NewQueue(0),
        proposer: proposer,
    }
}

其中schema是一个*memdb.DBSchema类型的变量：

schema        = &memdb.DBSchema{
        Tables: map[string]*memdb.TableSchema{},
    }

往schema添加成员使用的是register函数（定义在manager/state/store/memory.go）：

func register(os ObjectStoreConfig) {
    objectStorers = append(objectStorers, os)
    schema.Tables[os.Name] = os.Table
}

register函数在store package里各个文件（分别是cluster.go，networks.go，nodes.go，services.go和tasks.go，正好对应swarmctl的5个子命令。）的init()函数中使用，用来注册如何处理相应的object。

ObjectStoreConfig定义在manager/state/store/object.go：

// ObjectStoreConfig provides the necessary methods to store a particular object
// type inside MemoryStore.
type ObjectStoreConfig struct {
    Name             string
    Table            *memdb.TableSchema
    Save             func(ReadTx, *api.StoreSnapshot) error
    Restore          func(Tx, *api.StoreSnapshot) error
    ApplyStoreAction func(Tx, *api.StoreAction) error
    NewStoreAction   func(state.Event) (api.StoreAction, error)
}

它定义了如何存储一个object。

以services.go为例：

const tableService = "service"

func init() {
    register(ObjectStoreConfig{
        Name: tableService,
        Table: &memdb.TableSchema{
            Name: tableService,
            Indexes: map[string]*memdb.IndexSchema{
                indexID: {
                    Name:    indexID,
                    Unique:  true,
                    Indexer: serviceIndexerByID{},
                },
                indexName: {
                    Name:    indexName,
                    Unique:  true,
                    Indexer: serviceIndexerByName{},
                },
            },
        },
        Save: func(tx ReadTx, snapshot *api.StoreSnapshot) error {
            var err error
            snapshot.Services, err = FindServices(tx, All)
            return err
        },
        Restore: func(tx Tx, snapshot *api.StoreSnapshot) error {
            services, err := FindServices(tx, All)
            if err != nil {
                return err
            }
            for _, s := range services {
                if err := DeleteService(tx, s.ID); err != nil {
                    return err
                }
            }
            for _, s := range snapshot.Services {
                if err := CreateService(tx, s); err != nil {
                    return err
                }
            }
            return nil
        },
        ApplyStoreAction: func(tx Tx, sa *api.StoreAction) error {
            switch v := sa.Target.(type) {
            case *api.StoreAction_Service:
                obj := v.Service
                switch sa.Action {
                case api.StoreActionKindCreate:
                    return CreateService(tx, obj)
                case api.StoreActionKindUpdate:
                    return UpdateService(tx, obj)
                case api.StoreActionKindRemove:
                    return DeleteService(tx, obj.ID)
                }
            }
            return errUnknownStoreAction
        },
        NewStoreAction: func(c state.Event) (api.StoreAction, error) {
            var sa api.StoreAction
            switch v := c.(type) {
            case state.EventCreateService:
                sa.Action = api.StoreActionKindCreate
                sa.Target = &api.StoreAction_Service{
                    Service: v.Service,
                }
            case state.EventUpdateService:
                sa.Action = api.StoreActionKindUpdate
                sa.Target = &api.StoreAction_Service{
                    Service: v.Service,
                }
            case state.EventDeleteService:
                sa.Action = api.StoreActionKindRemove
                sa.Target = &api.StoreAction_Service{
                    Service: v.Service,
                }
            default:
                return api.StoreAction{}, errUnknownStoreAction
            }
            return sa, nil
        },
    })
}

NewStoreAction是创建针对service这张table的api.StoreAction；而ApplyStoreAction则是根据具体情况，使用相应的action（create，update还是delete，等等）；Save是从数据库读取所有的service并保存到一个snapshot中；Restore则是用snapshot中的值更新数据库中相应的service。

// CreateService creates and return a Service based on the provided ServiceSpec.
// - Returns `InvalidArgument` if the ServiceSpec is malformed.
// - Returns `Unimplemented` if the ServiceSpec references unimplemented features.
// - Returns `AlreadyExists` if the ServiceID conflicts.
// - Returns an error if the creation fails.
func (s *Server) CreateService(ctx context.Context, request *api.CreateServiceRequest) (*api.CreateServiceResponse, error) {
    ......
    err := s.store.Update(func(tx store.Tx) error {
        return store.CreateService(tx, service)
    })
    if err != nil {
        return nil, err
    }
    ......
}

// Update executes a read/write transaction.
func (s *MemoryStore) Update(cb func(Tx) error) error {
    return s.update(s.proposer, cb)
} 

func (s *MemoryStore) update(proposer state.Proposer, cb func(Tx) error) error {
    s.updateLock.Lock()
    memDBTx := s.memDB.Txn(true)

    var curVersion *api.Version

    if proposer != nil {
        curVersion = proposer.GetVersion()
    }

    var tx tx
    tx.init(memDBTx, curVersion)

    err := cb(&tx)

    if err == nil {
        if proposer == nil {
            memDBTx.Commit()
        } else {
            var sa []*api.StoreAction
            sa, err = tx.changelistStoreActions()

            if err == nil {
                if sa != nil {
                    err = proposer.ProposeValue(context.Background(), sa, func() {
                        memDBTx.Commit()
                    })
                } else {
                    memDBTx.Commit()
                }
            }
        }
    }

    if err == nil {
        for _, c := range tx.changelist {
            s.queue.Publish(c)
        }
        if len(tx.changelist) != 0 {
            s.queue.Publish(state.EventCommit{})
        }
    } else {
        memDBTx.Abort()
    }
    s.updateLock.Unlock()
    return err

}

memDBTx := s.memDB.Txn(true)

if proposer != nil {
    curVersion = proposer.GetVersion()
}

// ProposeValue calls Propose on the raft and waits
// on the commit log action before returning a result
func (n *Node) ProposeValue(ctx context.Context, storeAction []*api.StoreAction, cb func()) error {
    _, err := n.processInternalRaftRequest(ctx, &api.InternalRaftRequest{Action: storeAction}, cb)
    if err != nil {
        return err
    }
    return nil
}

// GetVersion returns the sequence information for the current raft round.
func (n *Node) GetVersion() *api.Version {
    n.stopMu.RLock()
    defer n.stopMu.RUnlock()

    if !n.IsMember() {
        return nil
    }

    status := n.Node.Status()
    return &api.Version{Index: status.Commit}
}

var tx tx
tx.init(memDBTx, curVersion)

err := cb(&tx)

// Tx is a read/write transaction. Note that transaction does not imply
// any internal batching. The purpose of this transaction is to give the
// user a guarantee that its changes won't be visible to other transactions
// until the transaction is over.
type Tx interface {
    ReadTx
    create(table string, o Object) error
    update(table string, o Object) error
    delete(table, id string) error
}

type tx struct {
    readTx
    curVersion *api.Version
    changelist []state.Event
}

func (tx *tx) init(memDBTx *memdb.Txn, curVersion *api.Version) {
    tx.memDBTx = memDBTx
    tx.curVersion = curVersion
    tx.changelist = nil
}

func(tx store.Tx) error {
        return store.CreateService(tx, service)
}

// CreateService adds a new service to the store.
// Returns ErrExist if the ID is already taken.
func CreateService(tx Tx, s *api.Service) error {
    // Ensure the name is not already in use.
    if tx.lookup(tableService, indexName, strings.ToLower(s.Spec.Annotations.Name)) != nil {
        return ErrNameConflict
    }

    return tx.create(tableService, serviceEntry{s})
}

// create adds a new object to the store.
// Returns ErrExist if the ID is already taken.
func (tx *tx) create(table string, o Object) error {
    if tx.lookup(table, indexID, o.ID()) != nil {
        return ErrExist
    }

    copy := o.Copy()
    meta := copy.Meta()
    if err := touchMeta(&meta, tx.curVersion); err != nil {
        return err
    }
    copy.SetMeta(meta)

    err := tx.memDBTx.Insert(table, copy)
    if err == nil {
        tx.changelist = append(tx.changelist, copy.EventCreate())
        o.SetMeta(meta)
    }
    return err
}

    if err == nil {
        if proposer == nil {
            memDBTx.Commit()
        } else {
            var sa []*api.StoreAction
            sa, err = tx.changelistStoreActions()

            if err == nil {
                if sa != nil {
                    err = proposer.ProposeValue(context.Background(), sa, func() {
                        memDBTx.Commit()
                    })
                } else {
                    memDBTx.Commit()
                }
            }
        }
    }

    if err == nil {
        for _, c := range tx.changelist {
            s.queue.Publish(c)
        }
        if len(tx.changelist) != 0 {
            s.queue.Publish(state.EventCommit{})
        }
    } else {
        memDBTx.Abort()
    }

此外，MemoryStore还提供了View函数，用来完成read transaction：

// ReadTx is a read transaction. Note that transaction does not imply
// any internal batching. It only means that the transaction presents a
// consistent view of the data that cannot be affected by other
// transactions.
type ReadTx interface {
    lookup(table, index, id string) Object
    get(table, id string) Object
    find(table string, by By, checkType func(By) error, appendResult func(Object)) error
}

type readTx struct {
    memDBTx *memdb.Txn
}

// View executes a read transaction.
func (s *MemoryStore) View(cb func(ReadTx)) {
    memDBTx := s.memDB.Txn(false)

    readTx := readTx{
        memDBTx: memDBTx,
    }
    cb(readTx)
    memDBTx.Commit()
}

Manager结构体（定义在manager/manager.go）包含一个*raft.Node成员：

// Manager is the cluster manager for Swarm.
// This is the high-level object holding and initializing all the manager
// subsystems.
type Manager struct {
    ......
    RaftNode   *raft.Node
    ......
}

而raft.Node（定义在manager/state/raft/raft.go）则包含一个*events.Broadcaster成员，用来接收改变manager role的消息（变成leader还是follower）：

// Node represents the Raft Node useful
// configuration.
type Node struct {
    ......
    leadershipBroadcast *events.Broadcaster
    ......
}   

发送改变当前manager role的代码位于manager/state/raft/raft.go：

// Run is the main loop for a Raft node, it goes along the state machine,
// acting on the messages received from other Raft nodes in the cluster.
//
// Before running the main loop, it first starts the raft node based on saved
// cluster state. If no saved state exists, it starts a single-node cluster.
func (n *Node) Run(ctx context.Context) error {
    ......
            // If we cease to be the leader, we must cancel
            // any proposals that are currently waiting for
            // a quorum to acknowledge them. It is still
            // possible for these to become committed, but
            // if that happens we will apply them as any
            // follower would.
            if rd.SoftState != nil {
                if wasLeader && rd.SoftState.RaftState != raft.StateLeader {
                    wasLeader = false
                    n.wait.cancelAll()
                    if atomic.LoadUint32(&n.signalledLeadership) == 1 {
                        atomic.StoreUint32(&n.signalledLeadership, 0)
                        n.leadershipBroadcast.Write(IsFollower)
                    }
                } else if !wasLeader && rd.SoftState.RaftState == raft.StateLeader {
                    wasLeader = true
                }
            }

            if wasLeader && atomic.LoadUint32(&n.signalledLeadership) != 1 {
                // If all the entries in the log have become
                // committed, broadcast our leadership status.
                if n.caughtUp() {
                    atomic.StoreUint32(&n.signalledLeadership, 1)
                    n.leadershipBroadcast.Write(IsLeader)
                }
            }
    ......
}

接收消息的代码在Manager.Run()函数（manager/manager.go）:

// Run starts all manager sub-systems and the gRPC server at the configured
// address.
// The call never returns unless an error occurs or `Stop()` is called.
func (m *Manager) Run(parent context.Context) error {
    ......
    leadershipCh, cancel := m.RaftNode.SubscribeLeadership()
    defer cancel()

    go m.handleLeadershipEvents(ctx, leadershipCh)

    ......

    go func() {
            err := m.RaftNode.Run(ctx)
            if err != nil {
                log.G(ctx).Error(err)
                m.Stop(ctx)
            }
        }()

    ......
}

Node.SubscribeLeadership()和Manager.handleLeadershipEvents()代码比较简单，不再赘述。

swarmctl实质上是通过controlClient向swarm cluster发命令。controlClient定义在api/control.pb.go：

// Client API for Control service

type ControlClient interface {
    GetNode(ctx context.Context, in *GetNodeRequest, opts ...grpc.CallOption) (*GetNodeResponse, error)
    ListNodes(ctx context.Context, in *ListNodesRequest, opts ...grpc.CallOption) (*ListNodesResponse, error)
    ......
}

type controlClient struct {
    cc *grpc.ClientConn
}

func NewControlClient(cc *grpc.ClientConn) ControlClient {
    return &controlClient{cc}
}

func (c *controlClient) GetNode(ctx context.Context, in *GetNodeRequest, opts ...grpc.CallOption) (*GetNodeResponse, error) {
    out := new(GetNodeResponse)
    err := grpc.Invoke(ctx, "/docker.swarmkit.v1.Control/GetNode", in, out, c.cc, opts...)
    if err != nil {
        return nil, err
    }
    return out, nil
}

......

Manager处理session请求是通过_Dispatcher_Session_Handler这个函数（./api/dispatcher.pb.go）：

func _Dispatcher_Session_Handler(srv interface{}, stream grpc.ServerStream) error {
    m := new(SessionRequest)
    if err := stream.RecvMsg(m); err != nil {
        return err
    }
    return srv.(DispatcherServer).Session(m, &dispatcherSessionServer{stream})
}

实际函数调用栈如下：

0  0x0000000000b65cbf in github.com/docker/swarmkit/manager/dispatcher.(*Dispatcher).Session
   at /go/src/github.com/docker/swarmkit/manager/dispatcher/dispatcher.go:768
1  0x0000000000782aa5 in github.com/docker/swarmkit/api.(*authenticatedWrapperDispatcherServer).Session
   at /go/src/github.com/docker/swarmkit/api/dispatcher.pb.go:207
2  0x000000000078e505 in github.com/docker/swarmkit/api.(*raftProxyDispatcherServer).Session
   at /go/src/github.com/docker/swarmkit/api/dispatcher.pb.go:1121
3  0x0000000000789c2a in github.com/docker/swarmkit/api._Dispatcher_Session_Handler
   at /go/src/github.com/docker/swarmkit/api/dispatcher.pb.go:667
4  0x0000000000909646 in github.com/docker/swarmkit/vendor/google.golang.org/grpc.(*Server).processStreamingRPC
   at /go/src/github.com/docker/swarmkit/vendor/google.golang.org/grpc/server.go:602
5  0x000000000090b002 in github.com/docker/swarmkit/vendor/google.golang.org/grpc.(*Server).handleStream
   at /go/src/github.com/docker/swarmkit/vendor/google.golang.org/grpc/server.go:686
6  0x000000000090fcbe in github.com/docker/swarmkit/vendor/google.golang.org/grpc.(*Server).serveStreams.func1.1
   at /go/src/github.com/docker/swarmkit/vendor/google.golang.org/grpc/server.go:348
7  0x0000000000462bf0 in runtime.goexit
   at /usr/local/go/src/runtime/asm_amd64.s:1998

Dispatcher.Session()函数代码如下：

// Session is a stream which controls agent connection.
// Each message contains list of backup Managers with weights. Also there is
// a special boolean field Disconnect which if true indicates that node should
// reconnect to another Manager immediately.
func (d *Dispatcher) Session(r *api.SessionRequest, stream api.Dispatcher_SessionServer) error {
    ctx := stream.Context()
    nodeInfo, err := ca.RemoteNode(ctx)
    if err != nil {
        return err
    }
    nodeID := nodeInfo.NodeID

    if err := d.isRunningLocked(); err != nil {
        return err
    }

    // register the node.
    sessionID, err := d.register(stream.Context(), nodeID, r.Description)
    if err != nil {
        return err
    }

    fields := logrus.Fields{
        "node.id":      nodeID,
        "node.session": sessionID,
        "method":       "(*Dispatcher).Session",
    }
    if nodeInfo.ForwardedBy != nil {
        fields["forwarder.id"] = nodeInfo.ForwardedBy.NodeID
    }
    log := log.G(ctx).WithFields(fields)

    var nodeObj *api.Node
    nodeUpdates, cancel, err := store.ViewAndWatch(d.store, func(readTx store.ReadTx) error {
        nodeObj = store.GetNode(readTx, nodeID)
        return nil
    }, state.EventUpdateNode{Node: &api.Node{ID: nodeID},
        Checks: []state.NodeCheckFunc{state.NodeCheckID}},
    )
    if cancel != nil {
        defer cancel()
    }

    if err != nil {
        log.WithError(err).Error("ViewAndWatch Node failed")
    }

    if _, err = d.nodes.GetWithSession(nodeID, sessionID); err != nil {
        return err
    }

    if err := stream.Send(&api.SessionMessage{
        SessionID:            sessionID,
        Node:                 nodeObj,
        Managers:             d.getManagers(),
        NetworkBootstrapKeys: d.networkBootstrapKeys,
    }); err != nil {
        return err
    }

    managerUpdates, mgrCancel := d.mgrQueue.Watch()
    defer mgrCancel()
    keyMgrUpdates, keyMgrCancel := d.keyMgrQueue.Watch()
    defer keyMgrCancel()

    // disconnectNode is a helper forcibly shutdown connection
    disconnectNode := func() error {
        // force disconnect by shutting down the stream.
        transportStream, ok := transport.StreamFromContext(stream.Context())
        if ok {
            // if we have the transport stream, we can signal a disconnect
            // in the client.
            if err := transportStream.ServerTransport().Close(); err != nil {
                log.WithError(err).Error("session end")
            }
        }

        nodeStatus := api.NodeStatus{State: api.NodeStatus_DISCONNECTED, Message: "node is currently trying to find new manager"}
        if err := d.nodeRemove(nodeID, nodeStatus); err != nil {
            log.WithError(err).Error("failed to remove node")
        }
        // still return an abort if the transport closure was ineffective.
        return grpc.Errorf(codes.Aborted, "node must disconnect")
    }

    for {
        // After each message send, we need to check the nodes sessionID hasn't
        // changed. If it has, we will the stream and make the node
        // re-register.
        node, err := d.nodes.GetWithSession(nodeID, sessionID)
        if err != nil {
            return err
        }

        var mgrs []*api.WeightedPeer

        var disconnect bool

        select {
        case ev := <-managerUpdates:
            mgrs = ev.([]*api.WeightedPeer)
        case ev := <-nodeUpdates:
            nodeObj = ev.(state.EventUpdateNode).Node
        case <-stream.Context().Done():
            return stream.Context().Err()
        case <-node.Disconnect:
            disconnect = true
        case <-d.ctx.Done():
            disconnect = true
        case <-keyMgrUpdates:
        }
        if mgrs == nil {
            mgrs = d.getManagers()
        }

        if err := stream.Send(&api.SessionMessage{
            SessionID:            sessionID,
            Node:                 nodeObj,
            Managers:             mgrs,
            NetworkBootstrapKeys: d.networkBootstrapKeys,
        }); err != nil {
            return err
        }
        if disconnect {
            return disconnectNode()
        }
    }
}

这个stream是处理agent连接的。前半部分是把连接的agent记录下来；后半部分是如果cluster信息发送变化，比如manager的leader发生变化，需要通知agent重新连接。disconnectNode()函数则是需要同agent node断开连接时的处理：包括断开连接，agent node信息删除，等等。

Node.runManager()函数会启动一个manager：

func (n *Node) runManager(ctx context.Context, securityConfig *ca.SecurityConfig, ready chan struct{}) error {
    for {
        n.waitRole(ctx, ca.ManagerRole)
        if ctx.Err() != nil {
            return ctx.Err()
        }
        remoteAddr, _ := n.remotes.Select(n.nodeID)
        m, err := manager.New(&manager.Config{
            ForceNewCluster: n.config.ForceNewCluster,
            ProtoAddr: map[string]string{
                "tcp":  n.config.ListenRemoteAPI,
                "unix": n.config.ListenControlAPI,
            },
            AdvertiseAddr:  n.config.AdvertiseRemoteAPI,
            SecurityConfig: securityConfig,
            ExternalCAs:    n.config.ExternalCAs,
            JoinRaft:       remoteAddr.Addr,
            StateDir:       n.config.StateDir,
            HeartbeatTick:  n.config.HeartbeatTick,
            ElectionTick:   n.config.ElectionTick,
        })
        if err != nil {
            return err
        }
        done := make(chan struct{})
        go func() {
            m.Run(context.Background()) // todo: store error
            close(done)
        }()

        n.Lock()
        n.manager = m
        n.Unlock()

        connCtx, connCancel := context.WithCancel(ctx)
        go n.initManagerConnection(connCtx, ready)

        // this happens only on initial start
        if ready != nil {
            go func(ready chan struct{}) {
                select {
                case <-ready:
                    n.remotes.Observe(api.Peer{NodeID: n.nodeID, Addr: n.config.ListenRemoteAPI}, picker.DefaultObservationWeight)
                case <-connCtx.Done():
                }
            }(ready)
            ready = nil
        }

        n.waitRole(ctx, ca.AgentRole)

        n.Lock()
        n.manager = nil
        n.Unlock()

        select {
        case <-done:
        case <-ctx.Done():
            err = ctx.Err()
            m.Stop(context.Background())
            <-done
        }
        connCancel()

        if err != nil {
            return err
        }
    }
}

（1）

        n.waitRole(ctx, ca.ManagerRole)
        if ctx.Err() != nil {
            return ctx.Err()
        }

首先runManager()函数会阻塞在waitRole()函数。一旦获得manager角色，就会往下执行。

（2）

        remoteAddr, _ := n.remotes.Select(n.nodeID)
        m, err := manager.New(&manager.Config{
            ForceNewCluster: n.config.ForceNewCluster,
            ProtoAddr: map[string]string{
                "tcp":  n.config.ListenRemoteAPI,
                "unix": n.config.ListenControlAPI,
            },
            AdvertiseAddr:  n.config.AdvertiseRemoteAPI,
            SecurityConfig: securityConfig,
            ExternalCAs:    n.config.ExternalCAs,
            JoinRaft:       remoteAddr.Addr,
            StateDir:       n.config.StateDir,
            HeartbeatTick:  n.config.HeartbeatTick,
            ElectionTick:   n.config.ElectionTick,
        })
        if err != nil {
            return err
        }
        done := make(chan struct{})
        go func() {
            m.Run(context.Background()) // todo: store error
            close(done)
        }()

        n.Lock()
        n.manager = m
        n.Unlock()

a）remoteAddr, _ := n.remotes.Select(n.nodeID)作用是从当前cluster的manager中（当然需要排除掉当前node）选出一个leader，赋给remoteAddr。如果当前node是cluster中的第一个manager，则remoteAddr就是一个“空的”值：{NodeID: "", Addr: ""}；
b）在使用manager.New()函数创建manager时，要注意n.config.AdvertiseRemoteAPI是一直为""的。 manager.New()最后会返回一个Manager结构体：

func New(config *Config) (*Manager, error) {
    ......
    m := &Manager{
        config:      config,
        listeners:   listeners,
        caserver:    ca.NewServer(RaftNode.MemoryStore(), config.SecurityConfig),
        Dispatcher:  dispatcher.New(RaftNode, dispatcherConfig),
        server:      grpc.NewServer(opts...),
        localserver: grpc.NewServer(opts...),
        RaftNode:    RaftNode,
        started:     make(chan struct{}),
        stopped:     make(chan struct{}),
    }

    return m, nil
}

其中的listeners包含监听listen-remote-api(tcp)和listen-control-api(unix)的两个socket。

（3）

        connCtx, connCancel := context.WithCancel(ctx)
        go n.initManagerConnection(connCtx, ready)

其中Node.initManagerConnection()实现如下：

func (n *Node) initManagerConnection(ctx context.Context, ready chan<- struct{}) error {
    opts := []grpc.DialOption{}
    insecureCreds := credentials.NewTLS(&tls.Config{InsecureSkipVerify: true})
    opts = append(opts, grpc.WithTransportCredentials(insecureCreds))
    // Using listen address instead of advertised address because this is a
    // local connection.
    addr := n.config.ListenControlAPI
    opts = append(opts, grpc.WithDialer(
        func(addr string, timeout time.Duration) (net.Conn, error) {
            return net.DialTimeout("unix", addr, timeout)
        }))
    conn, err := grpc.Dial(addr, opts...)
    if err != nil {
        return err
    }
    state := grpc.Idle
    for {
        s, err := conn.WaitForStateChange(ctx, state)
        if err != nil {
            n.setControlSocket(nil)
            return err
        }
        if s == grpc.Ready {
            n.setControlSocket(conn)
            if ready != nil {
                close(ready)
                ready = nil
            }
        } else if state == grpc.Shutdown {
            n.setControlSocket(nil)
        }
        state = s
    }
}

功能就是建立一个同本地listen-control-api(unix) socket的一个连接，用来监控node的状态。

（4）把当前node也加入remotes的监控列表中：

    // this happens only on initial start
        if ready != nil {
            go func(ready chan struct{}) {
                select {
                case <-ready:
                    n.remotes.Observe(api.Peer{NodeID: n.nodeID, Addr: n.config.ListenRemoteAPI}, picker.DefaultObservationWeight)
                case <-connCtx.Done():
                }
            }(ready)
            ready = nil
        }

（5）阻塞在下列代码，等待角色变化：

n.waitRole(ctx, ca.AgentRole)

// session encapsulates one round of registration with the manager. session
// starts the registration and heartbeat control cycle. Any failure will result
// in a complete shutdown of the session and it must be reestablished.
//
// All communication with the master is done through session.  Changes that
// flow into the agent, such as task assignment, are called back into the
// agent through errs, messages and tasks.
type session struct {
    agent     *Agent
    sessionID string
    session   api.Dispatcher_SessionClient
    errs      chan error
    messages  chan *api.SessionMessage
    tasks     chan *api.TasksMessage

    registered chan struct{} // closed registration
    closed     chan struct{}
}

（1）registered channel是用来通知agent已经向manager注册成功了：

func (s *session) run(ctx context.Context, delay time.Duration) {
    time.Sleep(delay) // delay before registering.

    if err := s.start(ctx); err != nil {
        select {
        case s.errs <- err:
        case <-s.closed:
        case <-ctx.Done():
        }
        return
    }

    ctx = log.WithLogger(ctx, log.G(ctx).WithField("session.id", s.sessionID))

    go runctx(ctx, s.closed, s.errs, s.heartbeat)
    go runctx(ctx, s.closed, s.errs, s.watch)
    go runctx(ctx, s.closed, s.errs, s.listen)

    close(s.registered)
}

session.run函数中，如果session.start()运行没有问题，则会在最后close registered这个channel。而在Agent.Run()中：

func (a *Agent) run(ctx context.Context) {
        .....
        session    = newSession(ctx, a, backoff) // start the initial session
        registered = session.registered
for {
        select {
            ......
        case <-registered:
            log.G(ctx).Debugln("agent: registered")
            if ready != nil {
                close(ready)
            }
            ready = nil
            registered = nil // we only care about this once per session
            backoff = 0      // reset backoff
            sessionq = a.sessionq
            ......
    }
}

一旦registered被close，<-registered这个case则会马上被执行。

（2）当session运行出现错误时，会把error发到errs channel。在Agent.Run()中：

case err := <-session.errs:
        // TODO(stevvooe): This may actually block if a session is closed
        // but no error was sent. Session.close must only be called here
        // for this to work.
        if err != nil {
            log.G(ctx).WithError(err).Error("agent: session failed")
            backoff = initialSessionFailureBackoff + 2*backoff
            if backoff > maxSessionFailureBackoff {
                backoff = maxSessionFailureBackoff
            }
        }

        if err := session.close(); err != nil {
            log.G(ctx).WithError(err).Error("agent: closing session failed")
        }
        sessionq = nil
        // if we're here before <-registered, do nothing for that event
        registered = nil

        // Bounce the connection.
        if a.config.Picker != nil {
            a.config.Picker.Reset()
        }

收到error后，会关闭这个session并做一些扫尾工作。

（3）messages channel用来接收manager发送给agent的消息，并转给Agent.run()函数进行处理：

case msg := <-session.messages:
        if err := a.handleSessionMessage(ctx, msg); err != nil {
            log.G(ctx).WithError(err).Error("session message handler failed")
        }

（4）tasks channel用来接收manager发送给agent的需要在这个node上运行的task信息，同样需要转给Agent.run()函数进行处理：

case msg := <-session.tasks:
        if err := a.worker.Assign(ctx, msg.Tasks); err != nil {
            log.G(ctx).WithError(err).Error("task assignment failed")
        }

（5）closed channel在session.close()函数中被关闭。也就是在case err := <-session.errs:这个分支中才会执行。一旦closed channel被关闭后，会重新建立连接：

case <-session.closed:
        log.G(ctx).Debugf("agent: rebuild session")

        // select a session registration delay from backoff range.
        delay := time.Duration(rand.Int63n(int64(backoff)))
        session = newSession(ctx, a, delay)
        registered = session.registered
        sessionq = a.sessionq  

// start begins the session and returns the first SessionMessage.
func (s *session) start(ctx context.Context) error {
    log.G(ctx).Debugf("(*session).start")

    client := api.NewDispatcherClient(s.agent.config.Conn)

    description, err := s.agent.config.Executor.Describe(ctx)
    if err != nil {
        log.G(ctx).WithError(err).WithField("executor", s.agent.config.Executor).
            Errorf("node description unavailable")
        return err
    }
    // Override hostname
    if s.agent.config.Hostname != "" {
        description.Hostname = s.agent.config.Hostname
    }

    errChan := make(chan error, 1)
    var (
        msg*api.SessionMessage
        stream api.Dispatcher_SessionClient
    )
    // Note: we don't defer cancellation of this context, because the
    // streaming RPC is used after this function returned. We only cancel
    // it in the timeout case to make sure the goroutine completes.
    sessionCtx, cancelSession := context.WithCancel(ctx)

    // Need to run Session in a goroutine since there's no way to set a
    // timeout for an individual Recv call in a stream.
    go func() {
        stream, err = client.Session(sessionCtx, &api.SessionRequest{
            Description: description,
        })
        if err != nil {
            errChan <- err
            return
        }

        msg, err = stream.Recv()
        errChan <- err
    }()

    select {
    case err := <-errChan:
        if err != nil {
            return err
        }
    case <-time.After(dispatcherRPCTimeout):
        cancelSession()
        return errors.New("session initiation timed out")
    }

    s.sessionID = msg.SessionID
    s.session = stream

    return s.handleSessionMessage(ctx, msg)
}

    client := api.NewDispatcherClient(s.agent.config.Conn)

    description, err := s.agent.config.Executor.Describe(ctx)
    if err != nil {
        log.G(ctx).WithError(err).WithField("executor", s.agent.config.Executor).
            Errorf("node description unavailable")
        return err
    }
    // Override hostname
    if s.agent.config.Hostname != "" {
        description.Hostname = s.agent.config.Hostname
    }

    type dispatcherClient struct {
        cc *grpc.ClientConn
    }

    func NewDispatcherClient(cc *grpc.ClientConn) DispatcherClient {
        return &dispatcherClient{cc}
    }

conn, err := grpc.Dial(manager.Addr,
        grpc.WithPicker(picker),
        grpc.WithTransportCredentials(creds),
        grpc.WithBackoffMaxDelay(maxSessionFailureBackoff))

    errChan := make(chan error, 1)
    var (
        msg*api.SessionMessage
        stream api.Dispatcher_SessionClient
    )
    // Note: we don't defer cancellation of this context, because the
    // streaming RPC is used after this function returned. We only cancel
    // it in the timeout case to make sure the goroutine completes.
    sessionCtx, cancelSession := context.WithCancel(ctx)

    // Need to run Session in a goroutine since there's no way to set a
    // timeout for an individual Recv call in a stream.
    go func() {
        stream, err = client.Session(sessionCtx, &api.SessionRequest{
            Description: description,
        })
        if err != nil {
            errChan <- err
            return
        }

        msg, err = stream.Recv()
        errChan <- err
    }()

func (c *dispatcherClient) Session(ctx context.Context, in *SessionRequest, opts ...grpc.CallOption) (Dispatcher_SessionClient, error) {
    stream, err := grpc.NewClientStream(ctx, &_Dispatcher_serviceDesc.Streams[0], c.cc, "/docker.swarmkit.v1.Dispatcher/Session", opts...)
    if err != nil {
        return nil, err
    }
    x := &dispatcherSessionClient{stream}
    if err := x.ClientStream.SendMsg(in); err != nil {
        return nil, err
    }
    if err := x.ClientStream.CloseSend(); err != nil {
        return nil, err
    }
    return x, nil
}

type Dispatcher_SessionClient interface {
    Recv() (*SessionMessage, error)
    grpc.ClientStream
}

type dispatcherSessionClient struct {
    grpc.ClientStream
}  

func (x *dispatcherSessionClient) Recv() (*SessionMessage, error) {
    m := new(SessionMessage)
    if err := x.ClientStream.RecvMsg(m); err != nil {
        return nil, err
    }
    return m, nil
}

// SessionRequest starts a session.
type SessionRequest struct {
    Description *NodeDescription `protobuf:"bytes,1,opt,name=description" json:"description,omitempty"`
}

msg, err = stream.Recv()
errChan <- err

    select {
    case err := <-errChan:
        if err != nil {
            return err
        }
    case <-time.After(dispatcherRPCTimeout):
        cancelSession()
        return errors.New("session initiation timed out")
    }

    s.sessionID = msg.SessionID
    s.session = stream

    return s.handleSessionMessage(ctx, msg)

go runctx(ctx, s.closed, s.errs, s.heartbeat)
go runctx(ctx, s.closed, s.errs, s.watch)
go runctx(ctx, s.closed, s.errs, s.listen)