How to solve the problem that Java System#exit cannot exit the program

Background

A friend encountered a situation: java.lang.System#exit cannot exit the application.

I was surprised when I heard this situation. Can this function still work? This makes me curious

Then, my friend continued to give his scene description: when the Dubbo application connects to the registration center, if the connection (timeout) fails, it is expected to call System#exit to exit the application, but The program does not exit as expected, and the JVM process still exists

At the same time, if the code that executes System#exit is placed in another thread, the program can exit as expected, and the JVM process ends

Use pseudo code to describe it as follows:

Future<Object> future = 连接注册中心的Future;
try {
    Object o = future.get(3, TimeUnit.SECONDS);
} catch (Exception e) {
	log.error("connect failed xxxx");
    System.exit(1); // 程序无法退出
}

-----------

Future<Object> future = 连接注册中心的Future;
try {
    Object o = future.get(3, TimeUnit.SECONDS);
} catch (Exception e) {
	log.error("connect failed xxxx");
    new Thread(() -> System.exit(1)).start(); // 程序能按期望退出
}

The scenario faced by my friend is much more complicated than the situation described in pseudo code, but the essential problems faced are the same.

A more general problem, in Dubbo's org.apache.dubbo.registry.zookeeper.ZookeeperRegistry#ZookeeperRegistry constructor, directly execute System.exit(1);The program cannot exit , executed in an asynchronous thread, it can exit as expected

That is:

// org.apache.dubbo.registry.zookeeper.ZookeeperRegistry#ZookeeperRegistry

public ZookeeperRegistry(URL url, ZookeeperTransporter zookeeperTransporter) {
    super(url);
    System.exit(1); //JVM进程无法退出
    // ...(省略)
}

-----------
// org.apache.dubbo.registry.zookeeper.ZookeeperRegistry#ZookeeperRegistry

public ZookeeperRegistry(URL url, ZookeeperTransporter zookeeperTransporter) {
    super(url);
    new Thread(() -> {System.exit(1);}).start(); //JVM进程正常退出
    // ...(省略)
}

This is even more amazing!

Troubleshooting

To find out the cause of the problem, you must first have some preliminary knowledge, otherwise you will be at a loss and feel like you have no way to start

• The java.lang.System#exit method is a method provided by Java to stop the JVM process

• When this method is triggered, the JVM will call the Shutdown Hook method until The JVM process will not be shut down until all hook methods are executed.

Guess from point 2 above: Is there an infinite loop of hook functions that cannot exit, so that the JVM is not shut down? process?

For example:

public static void main(String[] args) {
    Runtime.getRuntime().addShutdownHook(new Thread(() -> {
        while (true) {
            try {
                System.out.println("closing...");
                TimeUnit.SECONDS.sleep(1);
            } catch (InterruptedException e) {
            }
        }
    }));

    System.out.println("before exit...");
    System.exit(0); 
    System.out.println("after exit..."); //代码不会执行
}

As above, a shutdown hook is registered in the main method. This hook function is an infinite loop and will never Will exit, printing "closing...hellip;" every 3 seconds

Then execute the System.exit(0); method, expecting to exit the JVM process

before exit ...
closing...
closing...
closing...
closing...
closing...

...

The result is that the console keeps printing "closing...", and the JVM process does not exit.

The reason is exactly as mentioned in the second point of reserve knowledge above: the JVM will wait for all hooks to be executed. , before closing the process. The shutdown hook in the example will never be executed, so the JVM process will not be shut down

Despite having the reserve knowledge, I am still confused: if there is an infinite loop shutdown hook, thenSystem .exitWhether it is called in the main thread or in an asynchronous thread, it shouldnotclose the JVM process; conversely, if there is no infinite loop shutdown hook, no matter which thread it is in Calling should will shut down the JVM process. Why in the pseudo code of Background, different calling threads execute System.exit, resulting in different results?

At this time, I had to think of a way to see what the shutdown hooks were doing secretly, and why the execution was not completed, so that the JVM process could not exit.

I happened to have done some research on the source code of Dubbo. It is easy to find the constructor of org.apache.dubbo.registry.zookeeper.ZookeeperRegistry#ZookeeperRegistry and add a line of code to it, as shown below. After making the changes, recompile the source code and introduce it into your own project for Debug

Note: The Dubbo version used this time is 2.7.6

// org.apache.dubbo.registry.zookeeper.ZookeeperRegistry#ZookeeperRegistry
public ZookeeperRegistry(URL url, ZookeeperTransporter zookeeperTransporter) {
    super(url);
    System.exit(1); // 新增加的一行代码
    // ...(省略)
}

Start the project. Friends who are familiar with Dubbo should know that the application will go to the registration center during the startup process ( Here is Zookeeper) registration or subscription, because the consumer is started, so the application will try to connect to the registration center Zookeeper, and will go to the constructor of ZookeeperRegistry, because the second line of the constructor is the new codeSystem.exit(1);According to Background, the JVM will not exit and will be stuck. At this time, with the help of IDEA's "snapshot" function, you can "take a picture" of the Java The running status of the thread stack is functionally equivalent to executing the jstack command

to see a thread stack Suspicious thread: DubboShutdownHook

It can be seen from the name that it is a shutdown hook registered by Dubbo. Its main purpose is to close the connection and do some resource recycling.

It can also be seen from the picture that the thread is blocked at org.apache.dubbo.registry.support.AbstractRegistryFactoryLine 83

public static void destroyAll() {
    if (!destroyed.compareAndSet(false, true)) {
        return;
    }

    if (LOGGER.isInfoEnabled()) {
        LOGGER.info("Close all registries " + getRegistries());
    }
    // Lock up the registry shutdown process
    LOCK.lock(); // 83行，DubboShutdownHook线程阻塞在此处
    try {
        for (Registry registry : getRegistries()) {
            try {
                registry.destroy();
            } catch (Throwable e) {
                LOGGER.error(e.getMessage(), e);
            }
        }
        REGISTRIES.clear();
    } finally {
        // Release the lock
        LOCK.unlock();
    }
}

It can be clearly seen from the code, Because the lock cannot be acquired, the thread is blocked at line 83, waiting to acquire the lock. In other words, there are other threads holding the lock, but it has not been released yet, and DubboShutdownHook has to wait.

By IDEA , check where the lock has been obtained, as follows, if

org.apache.dubbo.registry.support.AbstractRegistryFactory#getRegistry(org.apache.dubbo.common.URL)

is found, the lock will be obtained

// org.apache.dubbo.registry.support.AbstractRegistryFactory

public Registry getRegistry(URL url) {
    // ...(省略)
    LOCK.lock(); // 获取锁
    try {
        // ...(省略)
        // 创建Registry，由于我们选用的注册中心是Zookeeper，因此通过SPI选择了ZookeeperRegistryFactory对ZookeeperRegistry进行创建，最终会调用到我们添加过一行System.exit的ZookeeperRegistry构造函数中
        
        registry = createRegistry(url); 
        
        // ...(省略)
    } finally {
        // Release the lock
        LOCK.unlock(); // 创建完registry，与注册中心连上之后，才会释放锁
    }
}

// org.apache.dubbo.registry.zookeeper.ZookeeperRegistryFactory

public Registry createRegistry(URL url) {
	// 调用修改过源码的ZookeeperRegistry构造函数
    return new ZookeeperRegistry(url, zookeeperTransporter);
}

如此，System.exit无法退出JVM进程的问题总算真相大白了：

1.Dubbo启动过程中会先获取锁，然后创建registry与注册中心进行连接，在ZookeeperRegistry中调用了java.lang.System#exit方法，程序转而执行"唤起shutdown hook"的代码并阻塞等待所有勾子函数执行完毕，而此时，之前持有的锁并没有释放

2.所有勾子函数(每个勾子函数都对应一个线程)被唤醒并执行，其中有一个Dubbo的勾子函数在执行的过程中，需要获取步骤1中的锁，由于获取锁失败，就阻塞等待着

3.由于1没有释放锁的情况下等待2执行完，而2的执行需要等待1释放锁，这样就形成了一个类似"死锁"的场景，因此也就导致了程序卡死，而JVM进程还存活的现象。之所以称为"类似"死锁，是因为1中执行System.exit的线程，也即持有锁的线程，永远不会走到释放锁的代码：一旦程序进入System.exit的世界里，就像进了一个单向虫洞，只能进不能出，如果勾子函数执行完毕，JVM进程接着就会被关闭，不会有机会再释放锁

那么，为什么在异步线程中执行System.exit，却能够正常退出JVM？

那是因为："唤起shutdown hook"并阻塞等待所有勾子函数执行完毕的线程是其它线程(此处假设是线程A)，该线程在阻塞时并未持有任何锁，而主线程会继续往下执行并接着释放锁。一旦锁释放，Shutdown hook就有机会持有该锁，并且执行其它资源的回收操作，等到所有的shutdown hook执行完毕，A线程就能从阻塞中返回并执行halt方法关闭JVM，因此能够正常退出JVM进程

深入学习

以上是对java.lang.System#exit 无法退出程序问题的分析，来龙去脉已经阐述清楚，受益于对Dubbo源码的了解以及正确的排查思路和排查手段，整个问题排查过程其实并没有花太多时间，但可以趁着这个机会，把java.lang.System#exit系统学习一下，或许会对以后问题排查、基础组件设计提供一些思路

System#exit

// java.lang.System

public static void exit(int status) {
    Runtime.getRuntime().exit(status);
}

Terminates the currently running Java Virtual Machine. The argument serves as a status code; by convention, a nonzero status code indicates abnormal termination.
This method calls the exit method in class Runtime. This method never returns normally.
The call System.exit(n) is effectively equivalent to the call:
Runtime.getRuntime().exit(n)

这个方法实现非常简单，是Runtime#exit的一个简便写法，其作用是用来关闭JVM进程，一旦调用该方法，永远也不会从该方法正常返回：执行完该方法后JVM进程就直接关闭了。

入参status取值分两类：0值与非0值，0值意味着正常关闭，非0值意味着异常关闭。

传入0值[有可能]会去执行所有的finalizer方法，非0值则一定不会执行(都不正常了，还执行啥finalizer呢?)。这儿提及[有可能]是因为，默认并不会执行finalizers，需要调用java.lang.Runtime#runFinalizersOnExit方法开启，而该方法早被JDK标识为Deprecated，因此通常情况下是不会开启的

// java.lang.Runtime

@Deprecated
public static void runFinalizersOnExit(boolean value) {
    SecurityManager security = System.getSecurityManager();
    if (security != null) {
        try {
            security.checkExit(0);
        } catch (SecurityException e) {
            throw new SecurityException("runFinalizersOnExit");
        }
    }
    Shutdown.setRunFinalizersOnExit(value);
}

接着看java.lang.Runtime#exit，可以看到，最终调用的是Shutdown.exit(status);，该方法是个包级别可见的方法，外部不可见

// java.lang.Runtime

public void exit(int status) {
    SecurityManager security = System.getSecurityManager();
    if (security != null) {
        security.checkExit(status);
    }
    Shutdown.exit(status);
}

// java.lang.Shutdown

static void exit(int status) {
    // ...(省略)
    synchronized (Shutdown.class) {
        /* Synchronize on the class object, causing any other thread
         * that attempts to initiate shutdown to stall indefinitely
         */
        // 执行shutdown序列
        sequence();
        // 关闭JVM
        halt(status);
    }
}

// java.lang.Shutdown

private static void sequence() {
    // ...(省略)
    runHooks();
    // ...(省略)
}

// java.lang.Shutdown

private static void runHooks() {
    for (int i=0; i < MAX_SYSTEM_HOOKS; i++) {
        try {
            Runnable hook;
            synchronized (lock) {
                // 这个锁很重要，目的是通过Happens-Before保证内存的可见性
                currentRunningHook = i;
                hook = hooks[i];
            }
            if (hook != null) hook.run(); //执行勾子函数
        } catch(Throwable t) {
            if (t instanceof ThreadDeath) {
                ThreadDeath td = (ThreadDeath)t;
                throw td;
            }
        }
    }
}

java.lang.Shutdown#runHooks有两个点需要注意，第一点MAX_SYSTEM_HOOKS(hooks)这个并不是我们注册的shutdown hooks，而是按顺序预定义的系统关闭勾子，目前JDK源码(JDK8)预定义了三个：

• Console restore hook

• Application hooks

• DeleteOnExit hook

其中，Application hooks才是我们应用程序中主动注册的shutdown hook。

在java.lang.ApplicationShutdownHooks类初始化时，会执行static代码块，并在其中注册了Application hooks

// java.lang.ApplicationShutdownHooks

class ApplicationShutdownHooks {
    /* The set of registered hooks */
    // 这个才是我们应用程序代码中注册的shutdown hook
    private static IdentityHashMap<Thread, Thread> hooks; 
    static {
        try {
            Shutdown.add(1 /* shutdown hook invocation order */,
                false /* not registered if shutdown in progress */,
                new Runnable() {
                    public void run() {
                        runHooks();
                    }
                }
            );
            hooks = new IdentityHashMap<>();
        } catch (IllegalStateException e) {
            // application shutdown hooks cannot be added if
            // shutdown is in progress.
            hooks = null;
        }
    }

其次要注意的点是，给hook变量赋值的时候进行了加锁

Runnable hook;
synchronized (lock) {
    currentRunningHook = i;
    hook = hooks[i];
}

一般而言，给局部变量赋值是不需要加锁的，因为局部变量是栈上变量，而线程栈之间数据是隔离的，不会出现线程安全的问题，因此不需要靠加锁来保证数据并发访问的安全性。

而此处加锁也并非为了解决线程安全问题，其真正的目的在于，通过Happens-Before规则来保证hooks的内存可见性：An unlock on a monitor happens-before every subsequent lock on that monitor。

如果不加锁，有可能导致从hooks数组中读取到的值并不是内存中最新的变量值，而是一个旧值

上面是读取hooks数组给hook变量赋值，为了满足HB(Happens-Before)原则，需要确保写操作中同样对hooks变量进行了加锁，因此我们看一下写hooks数组的地方，如下：

// java.lang.Shutdown

static void add(int slot, boolean registerShutdownInProgress, Runnable hook) {
    synchronized (lock) {
    		// ...(省略)
        hooks[slot] = hook;
    }
}

写 操作确实加了锁，这样才能让接下来的 读 操作的加锁行为满足HB原则

由于篇幅原因，就不展开具体的HB介绍，相信了解过HB原则的朋友一下就能明白其中的原理

这个点个人感觉很有意思，因为锁的作用不单是为了保证线程安全，还可以用来做为内存通信、保证内存可见性的手段，因此可以当作面试的一个点，当下次面试官问到：你写的代码中用过锁（synchronized）吗？什么场景用到锁？都集群部署了，单机锁还有意义吗？ 我们就可以回答：为了保证内存的可见性，balabalaba

所以你瞧，这个点其实也给我们设计基础组件带来很大的启发，synchronized在当今集群、分布式环境下并非一无是处，总有合适的地方在等待着它发挥光和热

注：JDK源码中真处处是宝藏，很多地方隐藏着巧妙而不可缺少的设计

在给hook变量赋值之后，就执行 if (hook != null) hook.run();，其中会执行到Application hooks，即上面提到的在ApplicationShutdownHooks类初始化时注册的勾子，勾子内部调用了java.lang.ApplicationShutdownHooks#runHooks方法

// java.lang.ApplicationShutdownHooks

Shutdown.add(1 /* shutdown hook invocation order */,
    false /* not registered if shutdown in progress */,
    new Runnable() {
        public void run() {
            runHooks();
        }
    }
);

// java.lang.ApplicationShutdownHooks

static void runHooks() {
    Collection<Thread> threads;
    synchronized(ApplicationShutdownHooks.class) {
        threads = hooks.keySet(); // hooks才是应用程序真正注册的shutdown hook
        hooks = null;
    }
		// 每一个shutdown hook都对应一个thread，由此可见是并发执行关闭勾子函数
    for (Thread hook : threads) {
        hook.start();
    }
    for (Thread hook : threads) {
        while (true) {
            try {
                hook.join(); // 死等到hook执行完毕
                break;
            } catch (InterruptedException ignored) {
                // 即便被唤醒都不搭理，接着进行下一轮循环，继续死等
            }
        }
    }
}

上面的hooks才是应用程序真正注册的shutdown hook，由源码可以看出，每一个hook都对应着一个thread，且调用了它们的start方法，即开启thread，意味着shutdown hook是并发、无序地执行

接着，唤起shutdown hook的线程，会通过死循环和join死等到所有关闭勾子都执行完毕，且忽略任何 唤醒异常。也即是说，如果勾子们不执行完，唤醒线程是不会离开的

等所有的Application hooks执行完毕，接下来会执行DeleteOnExit hook(如果存在)，等所有system hooks执行完毕，也基本意味着sequence方法执行完毕，接下来就执行halt方法关闭JVM虚拟机

synchronized (Shutdown.class) {
    sequence();
    halt(status);
}

这里额外还有一个知识点，上文只是提了一嘴，可能会容易忽略，此处拿出来解释一下：执行java.lang.System#exit永远也不会从该方法正常返回，也即是说，即便System#exit后边跟着的是finally，也不会执行 。一不注意就容易掉坑里

try {
    // ...
    System.exit(0);
} finally {
    // 这里的代码永远执行不到
}

java.lang.Runtime#addShutdownHook

聊完System#exit方法，接着来聊聊注册shutdown hook的方法。该方法本身实现上很简单，如下示：

// java.lang.Runtime
public void addShutdownHook(Thread hook) {
    // ...(省略)
    ApplicationShutdownHooks.add(hook);
}

// java.lang.ApplicationShutdownHooks
static synchronized void add(Thread hook) {
    // ...(省略)
    hooks.put(hook, hook);
}

需要注意的是，注册的关闭勾子会在以下几种时机被调用到

程序正常退出

• 最后一个非守护线程执行完毕退出时

• System.exit方法被调用时

程序响应外部事件

• 程序响应用户输入事件，例如在控制台按ctrl+c(^+c)

• 程序响应系统事件，如用户注销、系统关机等

除此之外，shutdown hook是不会被执行的

Shutdown hook存在的意义之一，是能够帮助我们实现优雅停机，而优雅停机的意义是：应用的重启、停机等操作，不影响业务的连续性

以Dubbo Provider的视角为例，优雅停机需要满足两点基本诉求：

• Consumer不应该请求到已经下线的Provider

• 在途请求需要处理完毕，不能被停机指令中断

Dubbo注册了Shutdown hook，JVM在收到操作系统发来的关闭指令时，会执行关闭勾子

• 在勾子中停止与注册中心的连接，注册中心会通知Consumer某个Provider已下线，后续不应该再调用该Provider进行服务。此行为是断掉上游流量，满足第一点诉求

• 接着，勾子执行Protocol(Dubbo相关概念)的注销逻辑，在其中判断server(Dubbo相关概念)是否还在处理请求，在超时时间内等待所有任务处理完毕，则关闭server。此行为是处理在途请求，满足第二点述求

因此，一种优雅停机的整体方案如下：

$pid = ps | grep xxx // 查找要关闭的应用
kill $pid // 发出关闭应用指令
sleep for a period of time // 等待一段时间，让应用程序执行shutdown hook进行现场的保留跟资源的清理工作

$pid = ps | grep xxx // 再次查找要关闭的应用，如果还存在，就需要强行关闭应用
if($pid){kill -9 $pid} // 等待一段时间之后，应用程序仍然没有正常停止，则需要强行关闭应用

The above is the detailed content of How to solve the problem that Java System#exit cannot exit the program. For more information, please follow other related articles on the PHP Chinese website!