// NewSender returns an implementation of Sender which exposes the Key-Value
// database provided by a Cockroach cluster by connecting via RPC to a
// Cockroach node.
//
// This must not be used by server.Server or any of its components, only by
// clients talking to a Cockroach cluster through the external interface.
func NewSender(ctx *rpc.Context, target string) (Sender, error) {
	conn, err := ctx.GRPCDial(target)
	if err != nil {
		return nil, err
	}
	return sender{roachpb.NewExternalClient(conn)}, nil
}

// grpcTransportFactoryImpl is the default TransportFactory, using GRPC.
// Do not use this directly - use grpcTransportFactory instead.
//
// During race builds, we wrap this to hold on to and read all obtained
// requests in a tight loop, exposing data races; see transport_race.go.
func grpcTransportFactoryImpl(
	opts SendOptions, rpcContext *rpc.Context, replicas ReplicaSlice, args roachpb.BatchRequest,
) (Transport, error) {
	clients := make([]batchClient, 0, len(replicas))
	for _, replica := range replicas {
		conn, err := rpcContext.GRPCDial(replica.NodeDesc.Address.String())
		if err != nil {
			return nil, err
		}
		argsCopy := args
		argsCopy.Replica = replica.ReplicaDescriptor
		remoteAddr := replica.NodeDesc.Address.String()
		clients = append(clients, batchClient{
			remoteAddr: remoteAddr,
			conn:       conn,
			client:     roachpb.NewInternalClient(conn),
			args:       argsCopy,
			healthy:    rpcContext.IsConnHealthy(remoteAddr),
		})
	}

	// Put known-unhealthy clients last.
	splitHealthy(clients)

	return &grpcTransport{
		opts:           opts,
		rpcContext:     rpcContext,
		orderedClients: clients,
	}, nil
}

// 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,
	rpcCtx *rpc.Context,
	stopper *stop.Stopper,
	nodeID roachpb.NodeID,
	breaker *circuit.Breaker,
) {
	stopper.RunWorker(func() {
		ctx, cancel := context.WithCancel(c.AnnotateCtx(context.Background()))
		var wg sync.WaitGroup
		defer func() {
			// This closes the outgoing stream, causing any attempt to send or
			// receive to return an error.
			//
			// Note: it is still possible for incoming gossip to be processed after
			// this point.
			cancel()

			// The stream is closed, but there may still be some incoming gossip
			// being processed. Wait until that is complete to avoid racing the
			// client's removal against the discovery of its remote's node ID.
			wg.Wait()
			disconnected <- c
		}()

		consecFailures := breaker.ConsecFailures()
		var stream Gossip_GossipClient
		if err := breaker.Call(func() error {
			// 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 := rpcCtx.GRPCDial(c.addr.String())
			if err != nil {
				return err
			}
			if stream, err = NewGossipClient(conn).Gossip(ctx); err != nil {
				return err
			}
			return c.requestGossip(g, stream)
		}, 0); err != nil {
			if consecFailures == 0 {
				log.Warningf(ctx, "node %d: failed to start gossip client: %s", nodeID, err)
			}
			return
		}

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