zookeeper之服务端接收请求处理流程

服务端有一个 NIOServerCnxn 类，用来处理客户端发送过来的请求

NIOServerCnxn

ZookeeperServer-zks.processPacket(this, bb);

处理客户端传送过来的数据包

public void processPacket(ServerCnxn cnxn, ByteBuffer incomingBuffer) throws IOException {
        // We have the request, now process and setup for next
        InputStream bais = new ByteBufferInputStream(incomingBuffer);
        BinaryInputArchive bia = BinaryInputArchive.getArchive(bais);
        RequestHeader h = new RequestHeader();
        h.deserialize(bia, "header"); // 反序列化客户端 header 头信息
        // Through the magic of byte buffers, txn will not be
        // pointing
        // to the start of the txn
        incomingBuffer = incomingBuffer.slice();
        if (h.getType() == OpCode.auth) { // 判断当前操作类型，如果是 auth操作，则执行下面的代码
            LOG.info("got auth packet " + cnxn.getRemoteSocketAddress());
            AuthPacket authPacket = new AuthPacket();
            ByteBufferInputStream.byteBuffer2Record(incomingBuffer, authPacket);
            String scheme = authPacket.getScheme();
            ServerAuthenticationProvider ap = ProviderRegistry.getServerProvider(scheme);
            Code authReturn = KeeperException.Code.AUTHFAILED;
            if (ap != null) {
                try {
                    authReturn = ap.handleAuthentication(new ServerAuthenticationProvider.ServerObjs(this, cnxn),
                            authPacket.getAuth());
                } catch (RuntimeException e) {
                    LOG.warn("Caught runtime exception from AuthenticationProvider: " + scheme + " due to " + e);
                    authReturn = KeeperException.Code.AUTHFAILED;
                }
            }
            if (authReturn == KeeperException.Code.OK) {
                if (LOG.isDebugEnabled()) {
                    LOG.debug("Authentication succeeded for scheme: " + scheme);
                }
                LOG.info("auth success " + cnxn.getRemoteSocketAddress());
                ReplyHeader rh = new ReplyHeader(h.getXid(), 0, KeeperException.Code.OK.intValue());
                cnxn.sendResponse(rh, null, null);
            } else {
                if (ap == null) {
                    LOG.warn("No authentication provider for scheme: " + scheme + " has "
                            + ProviderRegistry.listProviders());
                } else {
                    LOG.warn("Authentication failed for scheme: " + scheme);
                }
                // send a response...
                ReplyHeader rh = new ReplyHeader(h.getXid(), 0, KeeperException.Code.AUTHFAILED.intValue());
                cnxn.sendResponse(rh, null, null);
                // ... and close connection
                cnxn.sendBuffer(ServerCnxnFactory.closeConn);
                cnxn.disableRecv();
            }
            return;
        } else { // 如果不是授权操作，再判断是否为 sasl 操作
            if (h.getType() == OpCode.sasl) {
                Record rsp = processSasl(incomingBuffer, cnxn);
                ReplyHeader rh = new ReplyHeader(h.getXid(), 0, KeeperException.Code.OK.intValue());
                cnxn.sendResponse(rh, rsp, "response"); // not sure about 3rd
                                                        // arg..what is it?
                return;
            } else {// 最终进入这个代码块进行处理
                // 封装请求对象
                Request si = new Request(cnxn, cnxn.getSessionId(), h.getXid(), h.getType(), incomingBuffer,
                        cnxn.getAuthInfo());
                si.setOwner(ServerCnxn.me);
                // Always treat packet from the client as a possible
                // local request.
                setLocalSessionFlag(si);
                submitRequest(si); // 提交请求
            }
        }
        cnxn.incrOutstandingRequests(h);
    }

submitRequest

负责在服务端提交当前请求

public void submitRequest(Request si) {
        if (firstProcessor == null) { // processor 处理器，request
                                        // 过来以后会经历一系列处理器的处理过程
            synchronized (this) {
                try {
                    // Since all requests are passed to the request
                    // processor it should wait for setting up the request
                    // processor chain. The state will be updated to RUNNING
                    // after the setup.
                    while (state == State.INITIAL) {
                        wait(1000);
                    }
                } catch (InterruptedException e) {
                    LOG.warn("Unexpected interruption", e);
                }
                if (firstProcessor == null || state != State.RUNNING) {
                    throw new RuntimeException("Not started");
                }
            }
        }
        try {
            touch(si.cnxn);
            boolean validpacket = Request.isValid(si.type); // 判断是否合法
            if (validpacket) {
                firstProcessor.processRequest(si); // 调用 firstProcessor发起请求，而这个
                                                    // firstProcess
                                                    // 是一个接口，有多个实现类，具体的调用链是怎么样的？往下看吧
                if (si.cnxn != null) {
                    incInProcess();
                }
            } else {
                LOG.warn("Received packet at server of unknown type " + si.type);
                new UnimplementedRequestProcessor().processRequest(si);
            }
        } catch (MissingSessionException e) {
            if (LOG.isDebugEnabled()) {
                LOG.debug("Dropping request: " + e.getMessage());
            }
        } catch (RequestProcessorException e) {
            LOG.error("Unable to process request:" + e.getMessage(), e);
        }
    }

firstProcessor 的请求链组成

　　1. firstProcessor 的初始化是在 ZookeeperServer 的 setupRequestProcessor 中完成的，代码如下：

protected void setupRequestProcessors() {
    RequestProcessor finalProcessor = new FinalRequestProcessor(this);
    RequestProcessor syncProcessor = new SyncRequestProcessor(this, finalProcessor);
    ((SyncRequestProcessor) syncProcessor).start();
    firstProcessor = new PrepRequestProcessor(this, syncProcessor);
    // 需要注意的是，PrepRequestProcessor 中传递的是一个 syncProcessor
    ((PrepRequestProcessor) firstProcessor).start();
} 

从上面我们可以看到 firstProcessor 的实例是一个 PrepRequestProcessor，而这个

构造方法中又传递了一个 Processor 构成了一个调用链。

1	`RequestProcessor syncProcessor =` `new` `SyncRequestProcessor(this, finalProcessor);`

而 syncProcessor 的构造方法传递的又是一个 Processor，对应的是FinalRequestProcessor

　　1. 所以整个调用链是 PrepRequestProcessor -> SyncRequestProcessor ->FinalRequestProcessor

PredRequestProcessor.processRequest(si);

通过上面了解到调用链关系以后，我们继续再看 firstProcessor.processRequest(si)；会调用到 PrepRequestProcessor

public void processRequest(Request request) {
     submittedRequests.add(request);
}

唉，很奇怪，processRequest 只是把 request 添加到 submittedRequests 中，根据前面的经验，很自然的想到这里又是一个异步操作。而 subittedRequests 又是一个阻塞队列

1	`LinkedBlockingQueue submittedRequests =` `new` `LinkedBlockingQueue();`

而 PrepRequestProcessor 这个类又继承了线程类，因此我们直接找到当前类中的run 方法如下

public void run() {
        try {
            while (true) {
                Request request = submittedRequests.take(); // ok，从队列中拿到请求进行处理
                long traceMask = ZooTrace.CLIENT_REQUEST_TRACE_MASK;
                if (request.type == OpCode.ping) {
                    traceMask = ZooTrace.CLIENT_PING_TRACE_MASK;
                }
                if (LOG.isTraceEnabled()) {
                    ZooTrace.logRequest(LOG, traceMask, 'P', request, " ");
                }
                if (Request.requestOfDeath == request) {
                    break;
                }
                pRequest(request); // 调用 pRequest 进行预处理
            }
        } catch (RequestProcessorException e) {
            if (e.getCause() instanceof XidRolloverException) {
                LOG.info(e.getCause().getMessage());
            }
            handleException(this.getName(), e);
        } catch (Exception e) {
            handleException(this.getName(), e);
        }
        LOG.info("PrepRequestProcessor exited loop!");
    }

pRequest

预处理这块的代码太长，就不好贴了。前面的 N 行代码都是根据当前的 OP 类型进行判断和做相应的处理，在这个方法中的最后一行中，我们会看到如下代码

nextProcessor.processRequest(request);

SyncRequestProcessor. processRequest

public void processRequest(Request request) {
 // request.addRQRec(">sync");
     queuedRequests.add(request);
}

这个方法的代码也是一样，基于异步化的操作，把请求添加到 queuedRequets 中，那么我们继续在当前类找到 run 方法

public void run() {
        try {
            int logCount = 0;
            // we do this in an attempt to ensure that not all of the servers
            // in the ensemble take a snapshot at the same time
            int randRoll = r.nextInt(snapCount / 2);
            while (true) {
                Request si = null;
                // 从阻塞队列中获取请求
                if (toFlush.isEmpty()) {
                    si = queuedRequests.take();
                } else {
                    si = queuedRequests.poll();
                    if (si == null) {
                        flush(toFlush);
                        continue;
                    }
                }
                if (si == requestOfDeath) {
                    break;
                }
                if (si != null) {
                    // track the number of records written to the log
                    // 下面这块代码，粗略看来是触发快照操作，启动一个处理快照的线程
                    if (zks.getZKDatabase().append(si)) {
                        logCount++;
                        if (logCount > (snapCount / 2 + randRoll)) {
                            randRoll = r.nextInt(snapCount / 2);
                            // roll the log
                            zks.getZKDatabase().rollLog();
                            // take a snapshot
                            if (snapInProcess != null && snapInProcess.isAlive()) {
                                LOG.warn("Too busy to snap, skipping");
                            } else {
                                snapInProcess = new ZooKeeperThread("Snapshot Thread") {
                                    public void run() {
                                        try {
                                            zks.takeSnapshot();
                                        } catch (Exception e) {
                                            LOG.warn("Unexpected exception", e);
                                        }
                                    }
                                };
                                snapInProcess.start();
                            }
                            logCount = 0;
                        }
                    } else if (toFlush.isEmpty()) {
                        // optimization for read heavy workloads
                        // iff this is a read, and there are no pending
                        // flushes (writes), then just pass this to the next
                        // processor
                        if (nextProcessor != null) {
                            nextProcessor.processRequest(si); // 继续调用下一个处理器来处理请求
                            if (nextProcessor instanceof Flushable) {
                                ((Flushable) nextProcessor).flush();
                            }
                        }
                        continue;
                    }
                    toFlush.add(si);
                    if (toFlush.size() > 1000) {
                        flush(toFlush);
                    }
                }
            }
        } catch (Throwable t) {
            handleException(this.getName(), t);
        } finally {
            running = false;
        }
        LOG.info("SyncRequestProcessor exited!");
    }

FinalRequestProcessor. processRequest

这个方法就是我们在课堂上分析到的方法了，FinalRequestProcessor.processRequest 方法并根据 Request 对象中的操作更新内存中 Session 信息或者 znode 数据。这块代码有小 300 多行，就不全部贴出来了，我们直接定位到关键代码，根据客户端的 OP 类型找到如下的代码：

case OpCode.exists:
    {
    　　　　　lastOp = "EXIS";
    　　　　　// TODO we need to figure out the security requirement for this!
         ExistsRequest existsRequest = new ExistsRequest();
         //反序列化 (将 ByteBuffer 反序列化成为 ExitsRequest.这个就是我们在客户端发起请求的时候传递过来的 Request 对象
         ByteBufferInputStream.byteBuffer2Record(request.request,existsRequest);
         String path = existsRequest.getPath(); //得到请求的路径
         if (path.indexOf('\0') != -1) {
             throw new KeeperException.BadArgumentsException();
         }
    　　　　 　//终于找到一个很关键的代码，判断请求的 getWatch 是否存在，如果存在，则传递 cnxn（servercnxn）
    　　　　　 //对于 exists 请求，需要监听 data 变化事件，添加 watcher 
         Stat stat = zks.getZKDatabase().statNode(path, existsRe
         quest.getWatch() ? cnxn : null);
         rsp = new ExistsResponse(stat); //在服务端内存数据库中根据路径得到结果进行组装，设置为 ExistsResponse
         break;
     }