RPC之美团pigeon源码分析(二)端口绑定和消息处理

上篇我们了解到服务启动过程中,ProviderBootStrap.init()和ProviderBootStrap.startup(providerConfig);分别启动http协议服务和default(TCP)协议的netty服务。本篇我们分析下端口监听和消息处理。

我们以ProviderBootStrap.startup(providerConfig)来切入TCP协议NettyServer实现。
关键代码如下:

    public static ServerConfig startup(ProviderConfig providerConfig) {
            ServerConfig serverConfig = providerConfig.getServerConfig();
            synchronized (ProviderBootStrap.class) {
                //Server接口两个实现:NettyServer和JettyHttpServer
                List servers = ExtensionLoader.newExtensionList(Server.class);
                for (Server s : servers) {
                    if (!s.isStarted()) {
                        //NettyServer实现支持default协议
                        if (s.support(serverConfig)) {
                            //调用AbstractServer.start
                            s.start(serverConfig);
                            s.addService(providerConfig);
                            serversMap.put(s.getProtocol() + serverConfig.getPort(), s);
                            logger.warn("pigeon " + s + "[version:" + VersionUtils.VERSION + "] has been started");
                            break;
                        }
                    }
                }
                server = serversMap.get(serverConfig.getProtocol() + serverConfig.getPort());
                if (server != null) {
                    //预启动请求处理线程池,ThreadPoolExcutor.prestartAllCoreThreads()
                    server.getRequestProcessor().getRequestProcessThreadPool().prestartAllCoreThreads();
                    return server.getServerConfig();
                }
                return null;
            }
     }

跟进AbstractServer.start

    public void start(ServerConfig serverConfig) {
                //请求处理线程池RequestThreadPoolProcessor
        requestProcessor = RequestProcessorFactory.selectProcessor();
                //调用NettyServer.doStart
        doStart(serverConfig);
        if (requestProcessor != null) {
                        //根据上篇ServerConfig的线程池参数配置创建DynamicThreadPool线程池
                        //底层实现为new ThreadPoolExecutor(...)
            requestProcessor.start(serverConfig);
        }
        this.serverConfig = serverConfig;
    }

接下来我们看下NettyServer实现。
该pigeon版本基于Netty3.9.2.Final,本篇不对netty服务做详细说明,感兴趣的朋友可以阅读下netty系列文章。

    public NettyServer() {
        this.bootstrap = new ServerBootstrap(channelFactory);
        this.bootstrap.setPipelineFactory(new NettyServerPipelineFactory(this));
        this.bootstrap.setOption("child.tcpNoDelay", true);
        this.bootstrap.setOption("child.keepAlive", true);
        this.bootstrap.setOption("child.reuseAddress", true);
        this.bootstrap.setOption("child.connectTimeoutMillis", 1000);
    }

    @Override
    public void doStart(ServerConfig serverConfig) {
        if (!started) {
            if (serverConfig.isAutoSelectPort()) {
                int availablePort = getAvailablePort(serverConfig.getPort());
                this.port = availablePort;
            } 
            InetSocketAddress address = null;
            if (this.ip == null) {
                address = new InetSocketAddress(this.port);
            } else {
                address = new InetSocketAddress(this.ip, this.port);
            }
            channel = this.bootstrap.bind(address);
            serverConfig.setActualPort(this.port);
            this.started = true;
        }
    }
    //在NettyServerPipelineFactory中定义了一些列消息处理器
    public ChannelPipeline getPipeline() {
        ChannelPipeline pipeline = pipeline();
        pipeline.addLast("framePrepender", new FramePrepender());
        pipeline.addLast("frameDecoder", new FrameDecoder());
        pipeline.addLast("crc32Handler", new Crc32Handler(codecConfig));
        pipeline.addLast("compressHandler", new CompressHandler(codecConfig));
        pipeline.addLast("providerDecoder", new ProviderDecoder());
        pipeline.addLast("providerEncoder", new ProviderEncoder());
        //NettyServerHandler处理器中处理InvocationRequest
        pipeline.addLast("serverHandler", new NettyServerHandler(server));
        return pipeline;
    }

因为本篇在于介绍消息处理,我们重点看最后一个处理器NettyServerHandler,跟踪代码如下:

    //NettyServerHandler.messageReceived
    @Override
    public void messageReceived(ChannelHandlerContext ctx, MessageEvent message) {
        CodecEvent codecEvent = (CodecEvent) (message.getMessage());
        
        //请求信息
        InvocationRequest request = (InvocationRequest) codecEvent.getInvocation();
        ProviderContext invocationContext = new DefaultProviderContext(request, new NettyServerChannel(ctx.getChannel()));
        try {
            //请求处理
            this.server.processRequest(request, invocationContext);
        } catch (Throwable e) {
            String msg = "process request failed:" + request;
            // 心跳消息只返回正常的, 异常不返回
            if (request.getCallType() == Constants.CALLTYPE_REPLY
                    && request.getMessageType() != Constants.MESSAGE_TYPE_HEART) {
                ctx.getChannel().write(ProviderUtils.createFailResponse(request, e));
            }
        }
    }

    @Override
    public Future processRequest(InvocationRequest request, ProviderContext providerContext) {
                //最终调用RequestThreadPoolProcessor.doProcessRequest
        return requestProcessor.processRequest(request, providerContext);
    }

RequestThreadPoolProcessor.doProcessRequest就是我们所有关注的最终的业务处理逻辑,几点说明如下:
1、将业务处理逻辑封装在Callable任务中;
2、将Callable任务提交到线程池执行;
3、ProviderProcessHandlerFactory中将不同消息类型对应的Filter集合以ServiceInvocationHandler封装成责任链模式的消息处理器链
代码如下:

    public Future doProcessRequest(final InvocationRequest request,
                                                       final ProviderContext providerContext) {
        requestContextMap.put(request, providerContext);

        doMonitorData(request, providerContext);
        
        Callable requestExecutor = new Callable() {

            @Override
            public InvocationResponse call() throws Exception {
                providerContext.getTimeline().add(new TimePoint(TimePhase.T));
                try {
                    //根据消息类型获取对应的Filter责任链
                    ServiceInvocationHandler invocationHandler = ProviderProcessHandlerFactory
                            .selectInvocationHandler(providerContext.getRequest().getMessageType());
                    if (invocationHandler != null) {
                        providerContext.setThread(Thread.currentThread());
                        //执行责任链
                        return invocationHandler.handle(providerContext);
                    }
                } catch (Throwable t) {
                    logger.error("Process request failed with invocation handler, you should never be here.", t);
                } finally {
                    requestContextMap.remove(request);
                }
                return null;
            }
        };
        final ThreadPool pool = selectThreadPool(request);

        try {
            checkRequest(pool, request);
            providerContext.getTimeline().add(new TimePoint(TimePhase.T));
            //提交线程池执行
            return pool.submit(requestExecutor);
        } catch (RejectedExecutionException e) {
            requestContextMap.remove(request);
            throw new RejectedException(getProcessorStatistics(pool), e);
        }

    }

关于ProviderProcessHandlerFactory中不同消息类型对应的Filter集合如下:

    public static void init() {
                //业务消息
        registerBizProcessFilter(new TraceFilter());
        if (Constants.MONITOR_ENABLE) {
            registerBizProcessFilter(new MonitorProcessFilter());
        }
        registerBizProcessFilter(new WriteResponseProcessFilter());
        registerBizProcessFilter(new ContextTransferProcessFilter());
        registerBizProcessFilter(new ExceptionProcessFilter());
        registerBizProcessFilter(new SecurityFilter());
        registerBizProcessFilter(new GatewayProcessFilter());
        registerBizProcessFilter(new BusinessProcessFilter());
        bizInvocationHandler = createInvocationHandler(bizProcessFilters);
                //心跳消息
        registerHeartBeatProcessFilter(new WriteResponseProcessFilter());
        registerHeartBeatProcessFilter(new HeartbeatProcessFilter());
        heartBeatInvocationHandler = createInvocationHandler(heartBeatProcessFilters);
                
        registerHealthCheckProcessFilter(new WriteResponseProcessFilter());
        registerHealthCheckProcessFilter(new HealthCheckProcessFilter());
        healthCheckInvocationHandler = createInvocationHandler(healthCheckProcessFilters);

        registerScannerHeartBeatProcessFilter(new WriteResponseProcessFilter());
        registerScannerHeartBeatProcessFilter(new ScannerHeartBeatProcessFilter());
        scannerHeartBeatInvocationHandler = createInvocationHandler(scannerHeartBeatProcessFilters);
    }

以业务消息对应的Filter集合为例子,来看下责任链的封装以及Filter的执行顺序

        //Filter逆序封装,顺序执行
        //在ServiceInvocationHandler.handle中执行filter.invoke,在filter.invoke中执行next.handle
        private static  ServiceInvocationHandler createInvocationHandler(
            List internalFilters) {
        ServiceInvocationHandler last = null;
        List filterList = new ArrayList();
        filterList.addAll(internalFilters);
        for (int i = filterList.size() - 1; i >= 0; i--) {
            final V filter = filterList.get(i);
            final ServiceInvocationHandler next = last;
            last = new ServiceInvocationHandler() {
                @SuppressWarnings("unchecked")
                @Override
                public InvocationResponse handle(InvocationContext invocationContext) throws Throwable {
                    return filter.invoke(next, invocationContext);
                }
            };
        }
        return last;
    }

关于TraceFilter、MonitorProcessFilter、ExceptionProcessFilter等这里限于篇幅不具体分析代码,这些Filter负责的功能从Filter命名可以猜个七七八八。我们重点看看BusinessProcessFilter

    @Override
    public InvocationResponse invoke(ServiceInvocationHandler handler, ProviderContext invocationContext)
            throws Throwable {
        invocationContext.getTimeline().add(new TimePoint(TimePhase.U));
        InvocationRequest request = invocationContext.getRequest();
        if (request.getMessageType() == Constants.MESSAGE_TYPE_SERVICE) {
            。。。
            InvocationResponse response = null;
            ServiceMethod method = invocationContext.getServiceMethod();
            if (method == null) {
                //ServiceMethodFactory缓存了service bean的方法映射关系
                method = ServiceMethodFactory.getMethod(request);
            }
            。。。
            Object returnObj = null;
            try {
                //基于反射调用
                returnObj = method.invoke(request.getParameters());
            } finally {
                ProviderHelper.clearContext();
            }

            invocationContext.getTimeline().add(new TimePoint(TimePhase.M, System.currentTimeMillis()));
            if (request.getCallType() == Constants.CALLTYPE_REPLY) {
                response = ProviderUtils.createSuccessResponse(request, returnObj);
            }
            return response;
        }
        throw new BadRequestException("message type[" + request.getMessageType() + "] is not supported!");
    }

可以看到我们根据请求信息InvocationRequest从ServiceMethodFactory找到对应的service bean方法,然后基于反射执行。
而ServiceMethodFactory在上篇s服务初始化过程中缓存了每一个service bean的方法映射关系。

至此我们理清了TCP协议NettyServer的消息处理机制。
关于http协议的JettyHttpServer实现不再赘述了,感兴趣的朋友可以自行了解。

转载请备注原文链接。

你可能感兴趣的:(RPC之美团pigeon源码分析(二)端口绑定和消息处理)