[置顶] JAVA RMI线程模型及内部实现机制

JAVA RMI线程模型及内部实现机制

1 RMI内部实现

  JAVA RMI是JAVA分布式结构的基础。 远程对象的通信过程中， RMI 使用标准机制： stub 和 skeleton 。远程对象的 stub 担当远程对象的客户本地代表或代理人角色，调用程序将调 用本地 stub 的方法，而本地 stub 将负责执行对远程对象的方法调用。在 RMI 中，远程对象的 stub 与该远程对象所实现的远程接口集相同。调用 stub 的方法时将执行下列操作：
   (1) 初始化与包含远程对象的远程虚拟机的连接；
　 (2) 对远程虚拟机的参数进行编组-传输参数；
　 (3) 等待远程方法调用结果；
　 (4) 解编（读取）返回值或返回的异常；
　 (5) 将值返回给调用程序。

  为了向调用程序展示比较简单的调用机制， stub 将参数的序列化和网络级通信等细节隐藏了起来。在远程虚拟机中，每个远程对象都可以 有相应的 skeleton。s keleton 负责将调用分配给实际的远程对象实现。它在接收方法调用时 执行下列操作：

  (1) 解编（读取）远程方法的参数；

  (2) 调用实际远程对象实现上的方法；

  (3) 将结果（返回值或异常）编组（写入并传输）给调用程序。  

  stub 和 skeleton 由 rmic 编译器生成。在最新的JDK中，不需要手工生产stub和skeleton，用动态代理生成的Proxy代替了stub，而skeleton则取消了。

  我们可以查看源代码来了解RMI的内部实现。Server端调用UnicastRemoteObject的export方法输出远程对象，export方法会在一个线程里监听某个TCP端口上的方法调用请求：

public void exportObject(Target target) throws RemoteException { // other code while (true) { ServerSocket myServer = server; if (myServer == null) return; Throwable acceptFailure = null; final Socket socket; try { socket = myServer.accept(); /* * Find client host name (or "0.0.0.0" if unknown) */ InetAddress clientAddr = socket.getInetAddress(); String clientHost = (clientAddr != null ? clientAddr.getHostAddress() : "0.0.0.0"); /* * Spawn non-system thread to handle the connection */ Thread t = (Thread) java.security.AccessController.doPrivileged ( new NewThreadAction(new ConnectionHandler(socket, clientHost), "TCP Connection(" + ++ threadNum + ")-" + clientHost, true, true)); t.start(); } catch (IOException e) { acceptFailure = e; } catch (RuntimeException e) { acceptFailure = e; } catch (Error e) { acceptFailure = e; } } // other code }

  上面的代码已被修改以展示主要的要点，Server端就是在ServerSocket的accept方法上面监听到来的请求，如果有新的方法调用请求到来，Server产生一个单独的线程来处理新接收的请求：

public void dispatch(Remote obj, RemoteCall call) throws IOException { // positive operation number in 1.1 stubs; // negative version number in 1.2 stubs and beyond... int num; long op; try { // read remote call header ObjectInput in; try { in = call.getInputStream(); num = in.readInt(); if (num >= 0) { if (skel != null) { oldDispatch(obj, call, num); return; } else { throw new UnmarshalException( "skeleton class not found but required " + "for client version"); } } op = in.readLong(); } catch (Exception readEx) { throw new UnmarshalException("error unmarshalling call header", readEx); } /* * Since only system classes (with null class loaders) will be on * the execution stack during parameter unmarshalling for the 1.2 * stub protocol, tell the MarshalInputStream not to bother trying * to resolve classes using its superclasses's default method of * consulting the first non-null class loader on the stack. */ MarshalInputStream marshalStream = (MarshalInputStream) in; marshalStream.skipDefaultResolveClass(); Method method = (Method) hashToMethod_Map.get(new Long(op)); if (method == null) { throw new UnmarshalException("invalid method hash"); } // if calls are being logged, write out object id and operation logCall(obj, method); // unmarshal parameters Class[] types = method.getParameterTypes(); Object[] params = new Object[types.length]; try { unmarshalCustomCallData(in); for (int i = 0; i < types.length; i++) { params[i] = unmarshalValue(types[i], in); } } catch (java.io.IOException e) { throw new UnmarshalException( "error unmarshalling arguments", e); } catch (ClassNotFoundException e) { throw new UnmarshalException( "error unmarshalling arguments", e); } finally { call.releaseInputStream(); } // make upcall on remote object Object result; try { result = method.invoke(obj, params); } catch (InvocationTargetException e) { throw e.getTargetException(); } // marshal return value try { ObjectOutput out = call.getResultStream(true); Class rtype = method.getReturnType(); if (rtype != void.class) { marshalValue(rtype, result, out); } } catch (IOException ex) { throw new MarshalException("error marshalling return", ex); /* * This throw is problematic because when it is caught below, * we attempt to marshal it back to the client, but at this * point, a "normal return" has already been indicated, * so marshalling an exception will corrupt the stream. * This was the case with skeletons as well; there is no * immediately obvious solution without a protocol change. */ } } catch (Throwable e) { logCallException(e); ObjectOutput out = call.getResultStream(false); if (e instanceof Error) { e = new ServerError( "Error occurred in server thread", (Error) e); } else if (e instanceof RemoteException) { e = new ServerException( "RemoteException occurred in server thread", (Exception) e); } if (suppressStackTraces) { clearStackTraces(e); } out.writeObject(e); } finally { call.releaseInputStream(); // in case skeleton doesn't call.releaseOutputStream(); } } protected static void marshalValue(Class type, Object value, ObjectOutput out) throws IOException { if (type.isPrimitive()) { if (type == int.class) { out.writeInt(((Integer) value).intValue()); } else if (type == boolean.class) { out.writeBoolean(((Boolean) value).booleanValue()); } else if (type == byte.class) { out.writeByte(((Byte) value).byteValue()); } else if (type == char.class) { out.writeChar(((Character) value).charValue()); } else if (type == short.class) { out.writeShort(((Short) value).shortValue()); } else if (type == long.class) { out.writeLong(((Long) value).longValue()); } else if (type == float.class) { out.writeFloat(((Float) value).floatValue()); } else if (type == double.class) { out.writeDouble(((Double) value).doubleValue()); } else { throw new Error("Unrecognized primitive type: " + type); } } else { out.writeObject(value); } } protected static Object unmarshalValue(Class type, ObjectInput in) throws IOException, ClassNotFoundException { if (type.isPrimitive()) { if (type == int.class) { return new Integer(in.readInt()); } else if (type == boolean.class) { return new Boolean(in.readBoolean()); } else if (type == byte.class) { return new Byte(in.readByte()); } else if (type == char.class) { return new Character(in.readChar()); } else if (type == short.class) { return new Short(in.readShort()); } else if (type == long.class) { return new Long(in.readLong()); } else if (type == float.class) { return new Float(in.readFloat()); } else if (type == double.class) { return new Double(in.readDouble()); } else { throw new Error("Unrecognized primitive type: " + type); } } else { return in.readObject(); } }

   dispatch方法处理接收的请求，先从输入流中读取方法的编号来获得方法名称：op = in.readLong()，然后读取方法的所有参数：params[i] = unmarshalValue(types[i], in)，接着就可以执行方法调用了：result = method.invoke(obj, params)，最后把方法执行结果写入到输出流中：marshalValue(rtype, result, out)。一个方法调用就执行完成了。

  Client端请求一个远程方法调用时，先建立连接：Connection conn = ref.getChannel().newConnection()，然后发送方法参数： marshalValue(types[i], params[i], out)，再发送执行方法请求：call.executeCall()，最后得到方法的执行结果：Object returnValue = unmarshalValue(rtype, in)，并关闭接：

ref.getChannel().free(conn, true)。

public Object invoke(Remote obj, java.lang.reflect.Method method, Object[] params, long opnum) throws Exception { if (clientRefLog.isLoggable(Log.VERBOSE)) { clientRefLog.log(Log.VERBOSE, "method: " + method); } if (clientCallLog.isLoggable(Log.VERBOSE)) { logClientCall(obj, method); } Connection conn = ref.getChannel().newConnection(); RemoteCall call = null; boolean reuse = true; /* If the call connection is "reused" early, remember not to * reuse again. */ boolean alreadyFreed = false; try { if (clientRefLog.isLoggable(Log.VERBOSE)) { clientRefLog.log(Log.VERBOSE, "opnum = " + opnum); } // create call context call = new StreamRemoteCall(conn, ref.getObjID(), -1, opnum); // marshal parameters try { ObjectOutput out = call.getOutputStream(); marshalCustomCallData(out); Class[] types = method.getParameterTypes(); for (int i = 0; i < types.length; i++) { marshalValue(types[i], params[i], out); } } catch (IOException e) { clientRefLog.log(Log.BRIEF, "IOException marshalling arguments: ", e); throw new MarshalException("error marshalling arguments", e); } // unmarshal return call.executeCall(); try { Class rtype = method.getReturnType(); if (rtype == void.class) return null; ObjectInput in = call.getInputStream(); /* StreamRemoteCall.done() does not actually make use * of conn, therefore it is safe to reuse this * connection before the dirty call is sent for * registered refs. */ Object returnValue = unmarshalValue(rtype, in); /* we are freeing the connection now, do not free * again or reuse. */ alreadyFreed = true; /* if we got to this point, reuse must have been true. */ clientRefLog.log(Log.BRIEF, "free connection (reuse = true)"); /* Free the call's connection early. */ ref.getChannel().free(conn, true); return returnValue; } catch (IOException e) { clientRefLog.log(Log.BRIEF, "IOException unmarshalling return: ", e); throw new UnmarshalException("error unmarshalling return", e); } catch (ClassNotFoundException e) { clientRefLog.log(Log.BRIEF, "ClassNotFoundException unmarshalling return: ", e); throw new UnmarshalException("error unmarshalling return", e); } finally { try { call.done(); } catch (IOException e) { /* WARNING: If the conn has been reused early, * then it is too late to recover from thrown * IOExceptions caught here. This code is relying * on StreamRemoteCall.done() not actually * throwing IOExceptions. */ reuse = false; } } } catch (RuntimeException e) { /* * Need to distinguish between client (generated by the * invoke method itself) and server RuntimeExceptions. * Client side RuntimeExceptions are likely to have * corrupted the call connection and those from the server * are not likely to have done so. If the exception came * from the server the call connection should be reused. */ if ((call == null) || (((StreamRemoteCall) call).getServerException() != e)) { reuse = false; } throw e; } catch (RemoteException e) { /* * Some failure during call; assume connection cannot * be reused. Must assume failure even if ServerException * or ServerError occurs since these failures can happen * during parameter deserialization which would leave * the connection in a corrupted state. */ reuse = false; throw e; } catch (Error e) { /* If errors occurred, the connection is most likely not * reusable. */ reuse = false; throw e; } finally { /* alreadyFreed ensures that we do not log a reuse that * may have already happened. */ if (!alreadyFreed) { if (clientRefLog.isLoggable(Log.BRIEF)) { clientRefLog.log(Log.BRIEF, "free connection (reuse = " + reuse + ")"); } ref.getChannel().free(conn, reuse); } } }

2 RMI线程模型

  在JDK1.5及以前版本中，RMI每接收一个远程方法调用就生成一个单独的线程来处理这个请求，请求处理完成后，这个线程就会释放：

Thread t = (Thread) java.security.AccessController.doPrivileged ( new NewThreadAction(new ConnectionHandler(socket, clientHost), "TCP Connection(" + ++ threadNum + ")-" + clientHost, true, true));

   在JDK1.6之后，RMI使用线程池来处理新接收的远程方法调用请求-ThreadPoolExecutor。

  下面是一个简单的RMI程序的执行线程抓图，我们可以更好的了解RMI的线程机制。这个简单的RMI程序是服务端有一个远程方法实现，一个客户端同时请求执行这个远程方法100次。在JDK1.5中执行时生成的线程如下图所示，每个方法调用请求都是在一个单独的线程里执行，即 A Thread per Request。

在JDK1.6中执行时生成的线程如下图所示，这些线程都是在ThreadPoolExecutor线程池中执行的。

3 RMI线程池参数

  在JDK1.6中，RMI提供了可配置的线程池参数属性：

sun.rmi.transport.tcp.maxConnectionThread - 线程池中的最大线程数量

sun.rmi.transport.tcp.threadKeepAliveTime - 线程池中空闲的线程存活时间