http://code.google.com/p/muduo/
在分析muduo之前必须了解一下作者的想法:http://www.cnblogs.com/Solstice/archive/2010/08/29/muduo_net_lib.html
- 线程安全,支持多核多线程
- 不考虑可移植性,不跨平台,只支持 Linux,不支持 Windows。 // 支持Windows有时候代价太大了
- 在不增加复杂度的前提下可以支持 FreeBSD/Darwin,方便将来用 Mac 作为开发用机,但不为它做性能优化。也就是说 IO multiplexing 使用 poll 和 epoll。
- 主要支持 x86-64,兼顾 IA32
- 不支持 UDP,只支持 TCP
- 不支持 IPv6,只支持 IPv4
- 不考虑广域网应用,只考虑局域网 // 不会存在慢连接,所以即使是阻塞读也不会花去太长时间用在阻塞上面
- 只支持一种使用模式:non-blocking IO + one event loop per thread,不考虑阻塞 IO
- API 简单易用,只暴露具体类和标准库里的类,不使用 non-trivial templates,也不使用虚函数 // GP而非OO
- 只满足常用需求的 90%,不面面俱到,必要的时候以 app 来适应 lib
- 只做 library,不做成 framework
- 争取全部代码在 5000 行以内(不含测试)
- 以上条件都满足时,可以考虑搭配 Google Protocol Buffers RPC // RPC可以简化很多东西
muduo使用了很多新的Linux内核特性,包括使用signalfd和timerfd来触发信号以及定时器,所以代码上相对于于hpserver好看很多了。但是里面使用了boost::bind以及boost一些东西, 因为个人对于这个部分不是很清楚,所以很多地方并不是非常理解。muduo和hpserver一样也引入了很多概念,了解这些概念也非常有帮助。muduo做了线程管理,但是仅仅做了event loop 的线程管理,没有做工作线程的管理。所以工作线程还是需要自己管理。异步队列在base目录下面也实现了。所以基本上可以认为muduo里面包含了很多网络编程框架需要的组件。 muduo将hpserver下面的event item和event handler以及handle都在一起,称之为Channel.而Reactor在这里称为EventLoop.所以可以认为相对于hpserver,类层次结构好理解多了。
base下面都是一些关于多线程编程方面需要使用的组件 http://www.cnblogs.com/Solstice/archive/2010/08/21/muduo_thread_lib.html 包括下面这些文件:
- Atomic.h // 原子操作,里面的CAS没有使用汇编而是使用__sync_val_compare_and_swap这个GCC内置函数
- BlockingQueue.h // 异步队列,底层使用std::deque来实现,没有大小限制
- BoundedBlockingQueue.h // 异步队列,但是使用循环数组来实现,有大小限制
- Condition.h // pthread_cond封装
- CountDownLatch.h // 可以用作类似于起跑线机制,值得学习一下
void CountDownLatch::wait() // latchdown之后然后调用wait.等待最后一个线程notifyAll,
// 然后多个线程同时解除锁定就可以同时开始执行了
{
MutexLockGuard lock(mutex_);
while (count_ > 0) {
condition_.wait();
}
}
void CountDownLatch::countDown() // 每个线程在开始都latchdown,最后一个线程会notifyAll
{
MutexLockGuard lock(mutex_);
--count_;
if (count_ == 0) {
condition_.notifyAll();
}
}
- Logging.h // 日志
- Mutex.h // 互斥锁
- ProcessInfo.h // 进程信息
- Singleton.h // 单例模式,实现上比较有特色
static T& instance()
{
pthread_once(&ponce_, &Singleton::init); // 使用pthread_once来进行构造
return *value_;
}
- Thread.h // 线程封装,内部有一个static变量记录当前创建了多少个线程
- ThreadLocal.h // 线程局部变量封装,不用在使用pthread_get/setspecific
- ThreadLocalSignleton.h // 线程局部单例,不用考虑多个线程同时创建
- ThreadPool.h // 线程池包装,内部维护了一个异步队列,多个线程的工作就是取出task来执行
// 外部丢入Task放到线程池内部
void ThreadPool::run(const Task& task)
{
if (threads_.empty()) // 如果没有任何线程
{
task();
}
else
{
MutexLockGuard lock(mutex_);
queue_.push_back(task);
cond_.notify();
}
}
ThreadPool::Task ThreadPool::take()
{
MutexLockGuard lock(mutex_);
while (queue_.empty() && running_)
{
cond_.wait();
}
Task task;
if(!queue_.empty())
{
task = queue_.front();
queue_.pop_front();
}
return task;
}
// 线程回调函数
void ThreadPool::runInThread()
{
try
{
while (running_)
{
Task task(take());
if (task)
{
task();
}
}
}
}
Buffer设计的非常精巧。Buffer内部是一个std::vector<char*>表示的,逻辑上结构是这样的
/// @code /// +-------------------+------------------+------------------+ /// | prependable bytes | readable bytes | writable bytes | /// | | (CONTENT) | | /// +-------------------+------------------+------------------+ /// | | | | /// 0 <= readerIndex <= writerIndex <= size /// @endcode
头部有prependable bytes这个似乎可以不要,内部是占用8个字节,初始化的时候size占用了1024+8个字节。初始的时候readerIndex==writerIndex==8, 就好比现在没有任何数据写入。一旦开始要写入数据的话,那么writerIndex+=size(要写入的字节数)这个buffer会动态地增长。readerIndex标记的就是我们 可以读的下标,如果readerIndex==writerIndex就表示没有数据了。但是很明显这个Buffer并不是无限增长的,在makeSpace函数里面的话就可以看到,实际上 是会进行压缩的。
void makeSpace(size_t len)
{
if (writableBytes() + prependableBytes() < len + kCheapPrepend) // 如果当前不能够通过压缩合并的话
{
buffer_.resize(writerIndex_+len);
}
else
{
// move readable data to the front, make space inside buffer
assert(kCheapPrepend < readerIndex_); // 如果可以压缩的话那么就压缩
size_t readable = readableBytes();
std::copy(begin()+readerIndex_,
begin()+writerIndex_,
begin()+kCheapPrepend);
readerIndex_ = kCheapPrepend;
writerIndex_ = readerIndex_ + readable;
assert(readable == readableBytes());
}
}
class Channel : boost::noncopyable
{
public:
typedef boost::function<void()> EventCallback;
typedef boost::function<void(Timestamp)> ReadEventCallback;
void handleEvent(Timestamp receiveTime);
private:
EventLoop* loop_; // 属于哪一个Reactor
const int fd_; // 关联fd
int events_; // 关注事件
int revents_; // ready事件
int index_; // used by Poller. 在Poller中的编号,实际上没有特别意思
boost::weak_ptr<void> tie_; // 绑定的对象,这个对于boost::weak_ptr不是很了解,但是这个对于理解框架没有用途
bool tied_; // 是否绑定了对象上来
bool eventHandling_; // 当前正在处理event
ReadEventCallback readCallback_;
EventCallback writeCallback_; // 定义如何写数据
EventCallback closeCallback_; // 定义如何关闭连接
EventCallback errorCallback_; // 定义如果出错的话如何处理
};
一旦EventLoop通知Channel触发事件的话那么就会调用handleEvent这个函数。参数receiveTime本身只对ReadEventCallback有效并且作为参数使用, 代表接收超时时间,对于write而言的话没有超时时间。内部的话handleEvent会根据revents触发的事件来分别决定调用哪些回调
void Channel::handleEventWithGuard(Timestamp receiveTime)
{
eventHandling_ = true;
if ((revents_ & POLLHUP) && !(revents_ & POLLIN))
{
LOG_WARN << "Channel::handle_event() POLLHUP";
if (closeCallback_) closeCallback_();
}
if (revents_ & POLLNVAL)
{
LOG_WARN << "Channel::handle_event() POLLNVAL";
}
if (revents_ & (POLLERR | POLLNVAL))
{
if (errorCallback_) errorCallback_();
}
if (revents_ & (POLLIN | POLLPRI | POLLRDHUP))
{
if (readCallback_) readCallback_(receiveTime);
}
if (revents_ & POLLOUT)
{
if (writeCallback_) writeCallback_();
}
eventHandling_ = false;
}
Poller本身也是一个抽象类,然后底层支持poll和epoll.
class Poller : boost::noncopyable
{
public:
typedef std::vector<Channel*> ChannelList;
/// Polls the I/O events.
/// Must be called in the loop thread.
virtual Timestamp poll(int timeoutMs, ChannelList* activeChannels) = 0; // 进行poll操作,活跃事件放在activeChannels里面
/// Changes the interested I/O events.
/// Must be called in the loop thread.
virtual void updateChannel(Channel* channel) = 0; // 更新channel
/// Remove the channel, when it destructs.
/// Must be called in the loop thread.
virtual void removeChannel(Channel* channel) = 0; // 删除channel
private:
EventLoop* ownerLoop_;
};
在poller目录下面有poll和epoll的对应实现,不过我们这里没有必要仔细阅读。需要注意的是这里的channel处理并没有引入优先级的概念。 poll操作的timeoutMs就是epoll_wait超时时间,而activeChannels就是活跃channel.返回值就是epoll_wait之后的时间戳。
和之前一样,我们还是看看EventLoop有哪些结构。对于EventLoop结构比较复杂,我们列出主要的接口和成员。 首先我们看EventLoop有runInLoop和queueInLoop功能,虽然作者建议event loop和一个线程绑定,但是在其他线程的话依然可以调用runInLoop和 queueInLoop的功能,将一些task加入到这个event loop对应的线程中执行。这样就很地然地引入了pendingFunctors字段。因为需要跨线程激活, 那么就需要线程之间的通知机制,这个使用eventfd来完成,对应字段就是wakeFd并且内部绑定了一个wakeupChannel.如果没有eventfd的话,通常也可以使用 pipe来完成。然后我们还允许向EventLoop里面添加定时器任务,就是runAt,runAfter和runEvery三个函数,我们只需要关注其中一个即可。
class EventLoop : boost::noncopyable
{
public:
typedef boost::function<void()> Functor;
typedef boost::function<void()> TimerCallback;
void loop();
void quit();
/// Runs callback immediately in the loop thread.
/// It wakes up the loop, and run the cb.
/// If in the same loop thread, cb is run within the function.
/// Safe to call from other threads.
void runInLoop(const Functor& cb);
/// Queues callback in the loop thread.
/// Runs after finish pooling.
/// Safe to call from other threads.
void queueInLoop(const Functor& cb);
/// Runs callback at 'time'.
/// Safe to call from other threads.
///
TimerId runAt(const Timestamp& time, const TimerCallback& cb);
///
/// Runs callback after @c delay seconds.
/// Safe to call from other threads.
///
TimerId runAfter(double delay, const TimerCallback& cb);
///
/// Runs callback every @c interval seconds.
/// Safe to call from other threads.
///
TimerId runEvery(double interval, const TimerCallback& cb);
///
/// Cancels the timer.
/// Safe to call from other threads.
///
// void cancel(TimerId timerId);
// internal usage
void wakeup();
void updateChannel(Channel* channel);
void removeChannel(Channel* channel);
bool isInLoopThread() const { return threadId_ == CurrentThread::tid(); }
private:
typedef std::vector<Channel*> ChannelList;
bool looping_; /* atomic */
bool quit_; /* atomic */
bool eventHandling_; /* atomic */
bool callingPendingFunctors_; /* atomic */
const pid_t threadId_;
Timestamp pollReturnTime_;
boost::scoped_ptr<Poller> poller_;
boost::scoped_ptr<TimerQueue> timerQueue_;
int wakeupFd_;
// unlike in TimerQueue, which is an internal class,
// we don't expose Channel to client.
boost::scoped_ptr<Channel> wakeupChannel_;
ChannelList activeChannels_;
MutexLock mutex_;
std::vector<Functor> pendingFunctors_; // @BuardedBy mutex_
};
作者建议一个线程绑定一个EventLoop,这个实现呢?其实还是使用线程局部变量。首先定义线程局部变量
__thread EventLoop* t_loopInThisThread = 0;
然后在EventLoop构造函数的时候判断这个是否==0,如果不为=0的话说明在这个线程已经构造过一个EventLoop了。 直接使用__thread这个关键字,值得学习一下。
首先我们看看跨线程激活是怎么操作的。在EventLoop的初始化函数内部初始化了wakeupFd并且创建了channel.但是如果不仔细阅读, 很可能觉得的这个channel没有注册。而实际上这个channel在enableReading()就会注册的。
EventLoop::EventLoop()
: wakeupFd_(createEventfd()),
wakeupChannel_(new Channel(this, wakeupFd_))
{
wakeupChannel_->setReadCallback(
boost::bind(&EventLoop::handleRead, this)); // 绑定到handleRead上面了
// we are always reading the wakeupfd
wakeupChannel_->enableReading();
}
跨线程激活的函数是wakeUp.我们看看怎么实现
void EventLoop::wakeup()
{
uint64_t one = 1;
ssize_t n = sockets::write(wakeupFd_, &one, sizeof one); // 类似于管道直接写
}
一旦wakeup完成之后那么wakeUpFd_就是可读的,这样EventLoop就会被通知到并且立刻跳出epoll_wait开始处理。当然我们需要将这个wakeupFd_ 上面数据读出来,不然的话下一次又会被通知到,读取函数就是handleRead
void EventLoop::handleRead()
{
uint64_t one = 1;
ssize_t n = sockets::read(wakeupFd_, &one, sizeof one);
}
runInLoop和queueInLoop就是跨线程任务。内容非常简单
void EventLoop::runInLoop(const Functor& cb){
if (isInLoopThread()){ // 如果这个函数在自己的线程调用,那么就可以立即执行
cb();
}else{
queueInLoop(cb); // 如果是其他线程调用,那么加入到pendingFunctors里面去
wakeup(); // 并且通知这个线程,有任务到来
}
}
void EventLoop::queueInLoop(const Functor& cb){
{
MutexLockGuard lock(mutex_);
pendingFunctors_.push_back(cb);
}
if (isInLoopThread() && callingPendingFunctors_){
wakeup(); // 被排上队之后如果是在自己线程并且正在执行pendingFunctors的话,那么就可以激活
// 否则下一轮完全可以被排上,所以没有必要激活
}
}
定时器任务都是交给了TimerQueue来处理的,在TimerQueue这个部分我们会简要地分析一下
TimerId EventLoop::runAt(const Timestamp& time, const TimerCallback& cb)
{
return timerQueue_->addTimer(cb, time, 0.0); // time是在之后什么时候开始,0.0表示以后每次运行时间(0.0表示不会repeat).
}
基本上和hpserver非常相似,不断地调用poller::poll方法,然后在外层不断地查看是否需要quit.poll之后会得到activeChannels.和hpserver不同的是, muduo没有调用器(其实也不需要,本来就没有优先级概念),仅仅遍历这个activeChannels,并且调用内部的handleEvent方法,然后在调用pengdingFunctors 一些跨线程任务。
TimerQueue里面最主要的方法就是addTimer.我们看看addTimer里面做了哪些事情,整个过程有点绕
TimerId TimerQueue::addTimer(const TimerCallback& cb,
Timestamp when,
double interval)
{
Timer* timer = new Timer(cb, when, interval); // 首先创建一个Timer对象,然后将cb放在里面。内部有一个run函数,调用的就是cb
loop_->runInLoop(
boost::bind(&TimerQueue::scheduleInLoop, this, timer)); // 然后将这个timer丢到eventLoop里面去执行
return TimerId(timer, timer->sequence());
}
void TimerQueue::scheduleInLoop(Timer* timer)
{
loop_->assertInLoopThread();
bool earliestChanged = insert(timer); // 将timer插入到内部的链表里面去,按照超时时间顺序插入,并且判断这个插入是否会影响最早时间
if (earliestChanged)
{
resetTimerfd(timerfd_, timer->expiration()); // 如果影响的话,那么要修改这个timerfd超时时间。
}
}
然后一旦timerfd可读的时候,就会调用下面这个函数
void TimerQueue::handleRead()
{
loop_->assertInLoopThread();
Timestamp now(Timestamp::now());
readTimerfd(timerfd_, now);
std::vector<Entry> expired = getExpired(now); // 我们可以知道有哪些计时器超时
// safe to callback outside critical section
for (std::vector<Entry>::iterator it = expired.begin();
it != expired.end(); ++it)
{
it->second->run(); // 对于这些超时的Timer,执行run()函数,对应也就是我们一开始注册的回调函数。
}
reset(expired, now);
}
EventLoopThread就是将一个EventLoop和Thread包装在一起的对象。这个内容到没有什么,不过觉得代码方面有点技巧。 我们在启动startLoop这个样就会执行线程threadFunc,但是我们必须等待threadFunc将栈上面的EventLoop绑定之后才可以返回,所以这里用到了条件变量。
EventLoop* EventLoopThread::startLoop(){
thread_.start();
{
MutexLockGuard lock(mutex_);
while (loop_ == NULL)
{
cond_.wait();
}
}
return loop_;
}
void EventLoopThread::threadFunc(){
EventLoop loop;
{
MutexLockGuard lock(mutex_);
loop_ = &loop;
cond_.notify();
}
loop.loop();
}
而EventLoopThreadPool就是维持一个EventLoopThread线程池,所以没有什么特别好说的。我们只需要setThreadNum告诉开多少个线程,然后调用start即可。
Acceptor帮助简化了搭建服务器accept这个部分的逻辑。通常这个逻辑是在单个线程里面完成的,所以抽取出来蛮有必要的。 代码不是很麻烦,用户要做的就是编写一个回调,这个回调在新建立连接时候出发,参数分别是链接fd和连接地址。
typedef boost::function<void (int sockfd,const InetAddress&)> NewConnectionCallback;
原理很简单,初始化socket和对应的channel并且监听READ事件,然后开始进行listen.一旦触发read事件的话那么就证明我们无阻塞 地进行accept,然后在READ事件回调里面进行accept。一旦accept成功的话就调用这个回调函数即可。
Connector也是为了简化客户端编写,用户只需要提供这个逻辑即可,这个回调函数在建立链接成功的时候使用
typedef boost::function<void (int sockfd)> NewConnectionCallback;
Connector初始化以EventLoop和服务器地址初始化,然后在start的时候的话开始尝试进行connect.如果返回非阻塞的错误码的话,那么 创建一个channel并且监视WRITE和ERROR事件,否则就会尝试重连(按照一定时间间隔).在Connector::handleWrite里面的话会将这个channel 移除,然后调用NewConnectionCallback来处理连接建立的事件。
TcpConnection完成的工作就是当TCP连接建立之后处理socket的读写以及关闭。同样我们看看TcpConnection的结构
class TcpConnection : boost::noncopyable,
public boost::enable_shared_from_this<TcpConnection>
{
public:
/// Constructs a TcpConnection with a connected sockfd
///
/// User should not create this object.
TcpConnection(EventLoop* loop, // 建立连接需要一个Reactor
const string& name, // 连接名称
int sockfd, // 连接fd.
const InetAddress& localAddr, // 连接的address.
const InetAddress& peerAddr);
// called when TcpServer accepts a new connection
void connectEstablished(); // should be called only once
// called when TcpServer has removed me from its map
void connectDestroyed(); // should be called only once
private:
enum StateE { kDisconnected, kConnecting, kConnected, kDisconnecting };
void sendInLoop(const void* message, size_t len); // 发送消息
void setState(StateE s) { state_ = s; }
EventLoop* loop_;
string name_;
StateE state_; // FIXME: use atomic variable
// we don't expose those classes to client.
boost::scoped_ptr<Socket> socket_; // socket.
boost::scoped_ptr<Channel> channel_; // 连接channel
InetAddress localAddr_;
InetAddress peerAddr_;
ConnectionCallback connectionCallback_; // 连接回调,这个触发包括在连接建立和断开都会触发
MessageCallback messageCallback_; // 有数据可读的回调
WriteCompleteCallback writeCompleteCallback_; // 写完毕的回调
CloseCallback closeCallback_; // 连接关闭回调
Buffer inputBuffer_; // 数据读取buffer.
Buffer outputBuffer_; // FIXME: use list<Buffer> as output buffer.
boost::any context_; // 上下文环境
// FIXME: creationTime_, lastReceiveTime_
// bytesReceived_, bytesSent_
};
首先TcpConnection在初始化的时候会建立好channel.然后一旦TcpClient或者是TcpServer建立连接之后的话,那么调用TcpConnection::connectEstablished. 这个函数内部的话就会将channel设置成为可读。一旦可读的话那么TcpConnection内部就会调用handleRead这个动作,内部托管了读取数据这个操作。 读取完毕之后然后交给MessageBack这个回调进行操作。如果需要写的话调用sendInLoop,那么会将message放在outputBuffer里面,并且设置可写。 后当可写的话TcpConnection内部就托管写,然后写完之后的话会发生writeCompleteCallback这个回调。托管的读写操作都是非阻塞的。如果希望断开的话调用 shutdown。解除这个连接的话那么可以调用TcpConnection::connectDestroyed,内部大致操作就是从reactor移除这个channel.
在TcpConnection这层并不知道一次需要读取多少个字节,这个是在上层进行消息拆分的,这点可以仔细阅读一下Httpserver这个example. TcpConnection一次最多读取64K字节的内容,然后交给上层。上层决定这些内容是否足够,如果不够的话那么直接返回让Reactor继续等待读。 同样写的话内部也是会分多次写。这样就要求reactor内部必须使用水平触发而不是边缘触发。
一旦我们了解了TcpConnection之后的话,这个托管了建立好连接之后所需要的处理的所有事情,那么我们对于client关心的重点就是如果触发连接的建立以及连接是如何断开的。
TcpClient::TcpClient(EventLoop* loop,
const InetAddress& serverAddr,
const string& name)
: loop_(CHECK_NOTNULL(loop)),
connector_(new Connector(loop, serverAddr)),
name_(name),
connectionCallback_(defaultConnectionCallback),
messageCallback_(defaultMessageCallback),
retry_(false),
connect_(true),
nextConnId_(1)
{
connector_->setNewConnectionCallback(
boost::bind(&TcpClient::newConnection, this, _1));
// FIXME setConnectFailedCallback
}
可以看到初始化了connector这个对象并且设置了connector的连接建立的回调。我们需要设置一下TcpConnection所需要设置的回调之后,然后调用connect()这个方法。 内部会调用connector::start方法,一旦连接建立成功的话那么会调用TcpClient::newConnection这个函数。在这个函数内部会建立TcpConnection,并且调用 TcpConnection::connectEstablished,之后的所有操作都交给TcpConnection了。如果需要断开连接的话调用disconnect,内部会调用TcpConnection::shutdown.在析构 函数里面会调用TcpConneciton::connectDestroyed来移除连接。
从分析上我们和TcpClient一样只是关心连接是如何建立这个过程。
TcpServer::TcpServer(EventLoop* loop,
const InetAddress& listenAddr,
const string& nameArg)
: loop_(CHECK_NOTNULL(loop)),
hostport_(listenAddr.toHostPort()),
name_(nameArg),
acceptor_(new Acceptor(loop, listenAddr)),
threadPool_(new EventLoopThreadPool(loop)),
connectionCallback_(defaultConnectionCallback),
messageCallback_(defaultMessageCallback),
started_(false),
nextConnId_(1)
{
acceptor_->setNewConnectionCallback(
boost::bind(&TcpServer::newConnection, this, _1, _2));
}
同样是建立好acceptor这个对象然后设置好回调为TcpServer::newConnection,同时在外部设置好TcpConnection的各个回调。然后调用start来启动服务器,start 会调用acceptor::listen这个方法,一旦有连接建立的话那么会调用newConnection.下面是newConnection代码
void TcpServer::newConnection(int sockfd, const InetAddress& peerAddr)
{
loop_->assertInLoopThread();
EventLoop* ioLoop = threadPool_->getNextLoop();
char buf[32];
snprintf(buf, sizeof buf, ":%s#%d", hostport_.c_str(), nextConnId_);
++nextConnId_;
string connName = name_ + buf;
// FIXME poll with zero timeout to double confirm the new connection
TcpConnectionPtr conn(
new TcpConnection(ioLoop, connName, sockfd, localAddr, peerAddr));
connections_[connName] = conn;
conn->setConnectionCallback(connectionCallback_);
conn->setMessageCallback(messageCallback_);
conn->setWriteCompleteCallback(writeCompleteCallback_);
conn->setCloseCallback(
boost::bind(&TcpServer::removeConnection, this, _1)); // FIXME: unsafe
ioLoop->runInLoop(boost::bind(&TcpConnection::connectEstablished, conn));
}
对于服务端来说连接都被唯一化了然后映射称为字符串放在connections_这个容器内部。threadPool_->getNextLoop()可以轮询地将取出么一个线程然后将 TcpConnection::connectEstablished轮询地丢到每个线程里面去完成。存放在connections_是有原因了,每个TcpConnection有唯一一个名字,这样Server 就可以根据TcpConnection来从自己内部移除链接了。在析构函数里面可以遍历connections_内容得到所有建立的连接并且逐一释放。