查看原文
其他

聊聊jstack的工作原理

清泉 IT轻社区 2019-01-12

实现一个jstack

在聊Jstack得工作原理前呢,不如让我们先写一个简单的jstack玩玩。不用怕,很简单的,就几行代码的事,看:

public class MyJstack {
   public static void main(String[] args)throws Exception {
       VirtualMachine virtualMachine = VirtualMachine.attach("6361");
       HotSpotVirtualMachine hotSpotVirtualMachine = (HotSpotVirtualMachine)virtualMachine;
       InputStream inputStream = hotSpotVirtualMachine.remoteDataDump(new String[]{});
       byte[] buff = new byte[256];
       int len;
       do {
           len = inputStream.read(buff);
           if (len > 0) {
               String respone = new String(buff, 0, len, "UTF-8");
               System.out.print(respone);
           }
       } while(len > 0);
       inputStream.close();
       virtualMachine.detach();
   }
}

很简单吧,贴到你的开发环境里,运行就好了,别忘了把6361这个进程号换成你自己的Java进程号哦。

实现原理

jstack有两种实现方式,一种是基于attach api,其实现可以在tools.jar里找到;另一种是基于SA的实现,它被放在了sa-jdi.jar里。如果你通过idea搜索Jstack类,你会看到tools.jar和sa-jdi.jar各有一个Jstack类。

本文呢,就通过分析attch api的源码,来了解jstack的工作原理。

jstack本地源码实现

我们来看一下HotSpotVirtualMachine的remoteDataDump方法:

public InputStream remoteDataDump(Object... var1) throws IOException {
       return this.executeCommand("threaddump", var1);
}

他是在执行一个叫threaddump的命令。沿着这个executeCommand方法继续往里追,会发现他是调用了如下方法:

InputStream execute(String var1, Object... var2) throws AgentLoadException, IOException {
       assert var2.length <= 3;
       String var3;
       synchronized(this) {
           if (this.path == null) {
               throw new IOException("Detached from target VM");
           }
           var3 = this.path;
       }
       int var4 = socket();
       try {
           connect(var4, var3);
       } catch (IOException var9) {
           close(var4);
           throw var9;
       }
       IOException var5 = null;
       try {
           this.writeString(var4, "1");
           this.writeString(var4, var1);

var1参数就是我们的threaddump指令,不难看出,这个方法是建立了一个socket连接,然后将threaddump指令发送给另一端,即我们要检查的jvm进程。

注意:限于篇幅我并没有贴整个方法代码。execute是HotSpotVirtualMachine的抽象方法,不同平台的jdk有不同的execute方法的实现,我这里的代码是mac下的execute实现,位于BsdVirtualMachine类中。

通过jtack本地源代码,我们大致可以粗略的认为:jstack就是通过与指定的jvm进程建立socket连接,然后发送指令,最后将jvm进程返回的内容打印出来。

JVM的源码实现

了解了jstack的本地源码,我们在看看jvm进程是如何处理的。

当我们使用Java命令启动jvm进程时,Java命令会加载虚拟机共享库,然后执行共享库里的JNI_CreateJavaVM方法完成虚拟机的创建,在JNI_CreateJavaVM方法里会调用如下代码,完成具体的一个创建过程:

result = Threads::create_vm((JavaVMInitArgs*) args, &can_try_again);

如果你有心,或许会留意到,在你启动一个jvm进程时,即便你什么线程也没创建,你用jstack查看还是有很多的线程,如:Signal Dispatcher,VM Thread,Attach Listener等等。当过阅读本文,你会了解到这三个线程的作用。

01 VM Thread线程

Threads::create_vm这个方法很长,接下来咱们跳出一些重要的段落,来分析分析。

// Create the VMThread
 { TraceTime timer("Start VMThread", TraceStartupTime);
   VMThread::create();//创建Thread对象
   Thread* vmthread = VMThread::vm_thread();
   if (!os::create_thread(vmthread, os::vm_thread))//调用操作系统api创建线程
     vm_exit_during_initialization("Cannot create VM thread. Out of system resources.");
   // Wait for the VM thread to become ready, and VMThread::run to initialize
   // Monitors can have spurious returns, must always check another state flag
   {
     MutexLocker ml(Notify_lock);
     os::start_thread(vmthread);//启动线程
     while (vmthread->active_handles() == NULL) {
       Notify_lock->wait();
     }
   }
 }

通过注释,你也知道,这一段代码是从来创建VM Thread线程的。VMThread::create()完成了对现成的命名工作,代码如下:

void VMThread::create() {
 assert(vm_thread() == NULL, "we can only allocate one VMThread");
 _vm_thread = new VMThread();
 // Create VM operation queue
 _vm_queue = new VMOperationQueue();
 guarantee(_vm_queue != NULL, "just checking");
 _terminate_lock = new Monitor(Mutex::safepoint, "VMThread::_terminate_lock", true);
 if (UsePerfData) {
   // jvmstat performance counters
   Thread* THREAD = Thread::current();
   _perf_accumulated_vm_operation_time =
                PerfDataManager::create_counter(SUN_THREADS, "vmOperationTime",
                                                PerfData::U_Ticks, CHECK);
 }
}
VMThread::VMThread() : NamedThread() {
 set_name("VM Thread");
}

通过new VMThread()创建线程对象,在VMThread的构造方法里将线程命名成VM Thread,这就是我们jstack看到的VM Thread线程,同时还为这个线程创建了一个叫VMOperationQueue的队列。

至于VM Thread线程的作用,我们留到最后再说。

02 Signal Dispatcher线程

继续沿着 Threads::create_vm方法往下看,我们会看到如下代码:

// Signal Dispatcher needs to be started before VMInit event is posted
 os::signal_init();

这一句代码实现了Signal Dispatcher线程的创建,进入到signal_init()方法看看:

void os::signal_init() {
 if (!ReduceSignalUsage) {
   // Setup JavaThread for processing signals
   EXCEPTION_MARK;
   Klass* k = SystemDictionary::resolve_or_fail(vmSymbols::java_lang_Thread(), true, CHECK);
   instanceKlassHandle klass (THREAD, k);
   instanceHandle thread_oop = klass->allocate_instance_handle(CHECK);
   const char thread_name[] = "Signal Dispatcher";
   Handle string = java_lang_String::create_from_str(thread_name, CHECK);
   // Initialize thread_oop to put it into the system threadGroup
   Handle thread_group (THREAD, Universe::system_thread_group());
   JavaValue result(T_VOID);
   JavaCalls::call_special(&result, thread_oop,
                          klass,
                          vmSymbols::object_initializer_name(),
                          vmSymbols::threadgroup_string_void_signature(),
                          thread_group,
                          string,
                          CHECK);
   KlassHandle group(THREAD, SystemDictionary::ThreadGroup_klass());
   JavaCalls::call_special(&result,
                           thread_group,
                           group,
                           vmSymbols::add_method_name(),
                           vmSymbols::thread_void_signature(),
                           thread_oop,         // ARG 1
                           CHECK);
   os::signal_init_pd();
   { MutexLocker mu(Threads_lock);
     JavaThread* signal_thread = new JavaThread(&signal_thread_entry);
     // At this point it may be possible that no osthread was created for the
     // JavaThread due to lack of memory. We would have to throw an exception
     // in that case. However, since this must work and we do not allow
     // exceptions anyway, check and abort if this fails.
     if (signal_thread == NULL || signal_thread->osthread() == NULL) {
       vm_exit_during_initialization("java.lang.OutOfMemoryError",
                                     "unable to create new native thread");
     }
     java_lang_Thread::set_thread(thread_oop(), signal_thread);
     java_lang_Thread::set_priority(thread_oop(), NearMaxPriority);
     java_lang_Thread::set_daemon(thread_oop());
     signal_thread->set_threadObj(thread_oop());
     Threads::add(signal_thread);
     Thread::start(signal_thread);
   }
   // Handle ^BREAK
   os::signal(SIGBREAK, os::user_handler());
 }
}

在这个方法里,我们可以看到要创建的线程名字:Signal Dispatcher,以及线程启动后调用的方法signal_thread_entry。(方法较长,看重点就好,没必要每句话都扣清楚)。

有了对上边代码的分析,我们只需要看看signal_thread_entry方法,就知道Signal Dispatcher线程的作用了。

static void signal_thread_entry(JavaThread* thread, TRAPS) {
 os::set_priority(thread, NearMaxPriority);
 while (true) {
   int sig;
   {
     // FIXME : Currently we have not decieded what should be the status
     //         for this java thread blocked here. Once we decide about
     //         that we should fix this.
     sig = os::signal_wait();//等待获取信号
   }
   if (sig == os::sigexitnum_pd()) {
      // Terminate the signal thread
      return;
   }
   switch (sig) {
     case SIGBREAK: {
       // Check if the signal is a trigger to start the Attach Listener - in that
       // case don't print stack traces.
       if (!DisableAttachMechanism && AttachListener::is_init_trigger()) {
         continue;
       }
       // Print stack traces
       // Any SIGBREAK operations added here should make sure to flush
       // the output stream (e.g. tty->flush()) after output.  See 4803766.
       // Each module also prints an extra carriage return after its output.
       VM_PrintThreads op;
       VMThread::execute(&op);
       VM_PrintJNI jni_op;
       VMThread::execute(&jni_op);
       VM_FindDeadlocks op1(tty);
       VMThread::execute(&op1);
       Universe::print_heap_at_SIGBREAK();
       if (PrintClassHistogram) {
         VM_GC_HeapInspection op1(gclog_or_tty, true /* force full GC before heap inspection */);
         VMThread::execute(&op1);
       }
       if (JvmtiExport::should_post_data_dump()) {
         JvmtiExport::post_data_dump();
       }
       break;

这个方法里调用os::signal_wait()获取传给该jvm进程的信号,然后对信号进行处理。

说下case SIGBREAK里的处理逻辑,当接收到SIGBREAK信号时,会先判断是否禁止Attach机制,如果没有禁止,会调用AttachListener::is_init_trigger()方法触发Attach Listener线程的初始化.如果attach机制被禁用,则会创建VM_PrintThreads、VM_PrintJNI、VM_FindDeadlocks等代表某一个操作的对象,通过VMThread::execute()方法扔到VM Thread线程的VMOperationQueue队列。

03 Attach Listener线程

继续沿着 Threads::create_vm方法往下看,在紧挨着启动Signal Dispatcher线程的下边,就是启动Attach Listener线程的语句:

// Start Attach Listener if +StartAttachListener or it can't be started lazily
 if (!DisableAttachMechanism) {
   AttachListener::vm_start();
   if (StartAttachListener || AttachListener::init_at_startup()) {
     AttachListener::init();
   }
 }

重点就在AttachListener::init()方法里:

// Starts the Attach Listener thread
void AttachListener::init() {
 EXCEPTION_MARK;
 Klass* k = SystemDictionary::resolve_or_fail(vmSymbols::java_lang_Thread(), true, CHECK);
 instanceKlassHandle klass (THREAD, k);
 instanceHandle thread_oop = klass->allocate_instance_handle(CHECK);
 const char thread_name[] = "Attach Listener";
 Handle string = java_lang_String::create_from_str(thread_name, CHECK);
 // Initialize thread_oop to put it into the system threadGroup
 Handle thread_group (THREAD, Universe::system_thread_group());
 JavaValue result(T_VOID);
 JavaCalls::call_special(&result, thread_oop,
                      klass,
                      vmSymbols::object_initializer_name(),
                      vmSymbols::threadgroup_string_void_signature(),
                      thread_group,
                      string,
                      THREAD);
 if (HAS_PENDING_EXCEPTION) {
   tty->print_cr("Exception in VM (AttachListener::init) : ");
   java_lang_Throwable::print(PENDING_EXCEPTION, tty);
   tty->cr();
   CLEAR_PENDING_EXCEPTION;
   return;
 }
 KlassHandle group(THREAD, SystemDictionary::ThreadGroup_klass());
 JavaCalls::call_special(&result,
                       thread_group,
                       group,
                       vmSymbols::add_method_name(),
                       vmSymbols::thread_void_signature(),
                       thread_oop,             // ARG 1
                       THREAD);
 if (HAS_PENDING_EXCEPTION) {
   tty->print_cr("Exception in VM (AttachListener::init) : ");
   java_lang_Throwable::print(PENDING_EXCEPTION, tty);
   tty->cr();
   CLEAR_PENDING_EXCEPTION;
   return;
 }
 { MutexLocker mu(Threads_lock);
   JavaThread* listener_thread = new JavaThread(&attach_listener_thread_entry);
   // Check that thread and osthread were created
   if (listener_thread == NULL || listener_thread->osthread() == NULL) {
     vm_exit_during_initialization("java.lang.OutOfMemoryError",
                                   "unable to create new native thread");
   }
   java_lang_Thread::set_thread(thread_oop(), listener_thread);
   java_lang_Thread::set_daemon(thread_oop());
   listener_thread->set_threadObj(thread_oop());
   Threads::add(listener_thread);
   Thread::start(listener_thread);
 }
}

我们可以通过代码看出其创建了一个叫Attach Listener的线程,线程执行的逻辑封装在了attach_listener_thread_entry方法里。

Attach Listener线程的作用,我们看看attach_listener_thread_entry方法便知:

static void attach_listener_thread_entry(JavaThread* thread, TRAPS) {
 os::set_priority(thread, NearMaxPriority);
 thread->record_stack_base_and_size();
 if (AttachListener::pd_init() != 0) {
   return;
 }
 AttachListener::set_initialized();
 for (;;) {
   AttachOperation* op = AttachListener::dequeue();//从队列里获取操作对象
   if (op == NULL) {
     return;   // dequeue failed or shutdown
   }
   ResourceMark rm;
   bufferedStream st;
   jint res = JNI_OK;
   // handle special detachall operation
   if (strcmp(op->name(), AttachOperation::detachall_operation_name()) == 0) {
     AttachListener::detachall();
   } else {
     // find the function to dispatch too
     AttachOperationFunctionInfo* info = NULL;
     for (int i=0; funcs[i].name != NULL; i++) {
       const char* name = funcs[i].name;
       assert(strlen(name) <= AttachOperation::name_length_max, "operation <= name_length_max");
       if (strcmp(op->name(), name) == 0) {
         info = &(funcs[i]);
         break;
       }
     }
     // check for platform dependent attach operation
     if (info == NULL) {
       info = AttachListener::pd_find_operation(op->name());
     }
     if (info != NULL) {
       // dispatch to the function that implements this operation
       res = (info->func)(op, &st);//执行操作对象
     } else {
       st.print("Operation %s not recognized!", op->name());
       res = JNI_ERR;
     }
   }
   // operation complete - send result and output to client
   op->complete(res, &st);
 }
}

方法很长,我把重点挑出来分析。

首先我们看看调用AttachListener::pd_init()完了什么:

int AttachListener::pd_init() {
 JavaThread* thread = JavaThread::current();
 ThreadBlockInVM tbivm(thread);
 thread->set_suspend_equivalent();
 // cleared by handle_special_suspend_equivalent_condition() or
 // java_suspend_self() via check_and_wait_while_suspended()
 int ret_code = LinuxAttachListener::init();
 // were we externally suspended while we were waiting?
 thread->check_and_wait_while_suspended();
 return ret_code;
}
int LinuxAttachListener::init() {
 char path[UNIX_PATH_MAX];          // socket file
 char initial_path[UNIX_PATH_MAX];  // socket file during setup
 int listener;                      // listener socket (file descriptor)
 // register function to cleanup
 ::atexit(listener_cleanup);
 int n = snprintf(path, UNIX_PATH_MAX, "%s/.java_pid%d",
                  os::get_temp_directory(), os::current_process_id());
 if (n < (int)UNIX_PATH_MAX) {
   n = snprintf(initial_path, UNIX_PATH_MAX, "%s.tmp", path);
 }
 if (n >= (int)UNIX_PATH_MAX) {
   return -1;
 }
 // create the listener socket
 listener = ::socket(PF_UNIX, SOCK_STREAM, 0);//创建套接字
 if (listener == -1) {
   return -1;
 }
 // bind socket
 struct sockaddr_un addr;
 addr.sun_family = AF_UNIX;
 strcpy(addr.sun_path, initial_path);
 ::unlink(initial_path);
 int res = ::bind(listener, (struct sockaddr*)&addr, sizeof(addr));//绑定地址
 if (res == -1) {
   ::close(listener);
   return -1;
 }
 // put in listen mode, set permissions, and rename into place
 res = ::listen(listener, 5);//发起监听
 if (res == 0) {
     RESTARTABLE(::chmod(initial_path, S_IREAD|S_IWRITE), res);
     if (res == 0) {
         res = ::rename(initial_path, path);
     }
 }
 if (res == -1) {
   ::close(listener);
   ::unlink(initial_path);
   return -1;
 }
 set_path(path);
 set_listener(listener);
 return 0;
}

不难发现,AttachListener::pd_init()方法又调用了LinuxAttachListener::init()方法,完成了对套接字的创建和监听。这与jstack本地代码建立socket连接发送命令,不谋而合。

再就是有一个for死循环,不停地调用AttachOperation* op = AttachListener::dequeue();获取操作对象。如果进入到AttachListener::dequeue()方法看一看,其实就是在读上边监听的套接字,我这里就不贴源码了。

在这个死循环里,我们重点看看如下代码:

   // find the function to dispatch too
     AttachOperationFunctionInfo* info = NULL;
     for (int i=0; funcs[i].name != NULL; i++) {
       const char* name = funcs[i].name;
       assert(strlen(name) <= AttachOperation::name_length_max, "operation <= name_length_max");
       if (strcmp(op->name(), name) == 0) {
         info = &(funcs[i]);
         break;
       }
     }
     // check for platform dependent attach operation
     if (info == NULL) {
       info = AttachListener::pd_find_operation(op->name());
     }
     if (info != NULL) {
       // dispatch to the function that implements this operation
       res = (info->func)(op, &st);//调动方法
     } else {
       st.print("Operation %s not recognized!", op->name());
       res = JNI_ERR;
     }
   }
   // operation complete - send result and output to client
   op->complete(res, &st);

这个for循环会遍历funcs数组,然后根据从队列里拿到的AttachOperation对象的name来找到一个匹配的AttachOperationFunctionInfo对象,然后调用其func方法。

看到这里你或许很多疑惑,当然看看funcs数组里的东西,就开朗了:

static AttachOperationFunctionInfo funcs[] = {
 { "agentProperties",  get_agent_properties },
 { "datadump",         data_dump },
 { "dumpheap",         dump_heap },
 { "load",             JvmtiExport::load_agent_library },
 { "properties",       get_system_properties },
 { "threaddump",       thread_dump },
 { "inspectheap",      heap_inspection },
 { "setflag",          set_flag },
 { "printflag",        print_flag },
 { "jcmd",             jcmd },
 { NULL,               NULL }
};

有没有看到上文中我们提到的threaddump命令。jstack通过与jvm进程建立socket连接,然后向jvm进程发送threaddump指令。上文说道调用AttachOperationFunctionInfo对象的func方法处理指令,其实就是调用了thread_dump方法,针对threaddump命令来说。

坚持,马上就要说完了。来看看thread_dump方法干了些啥吧:

// Implementation of "threaddump" command - essentially a remote ctrl-break
// See also: ThreadDumpDCmd class
//
static jint thread_dump(AttachOperation* op, outputStream* out) {
 bool print_concurrent_locks = false;
 if (op->arg(0) != NULL && strcmp(op->arg(0), "-l") == 0) {
   print_concurrent_locks = true;
 }
 // thread stacks
 VM_PrintThreads op1(out, print_concurrent_locks);
 VMThread::execute(&op1);
 // JNI global handles
 VM_PrintJNI op2(out);
 VMThread::execute(&op2);
 // Deadlock detection
 VM_FindDeadlocks op3(out);
 VMThread::execute(&op3);
 return JNI_OK;
}

很简单,创建了VM_PrintThreads、VM_PrintJNI、VM_FindDeadlocks三个对象,扔给了VM Thread线程的队列。

说到这里,VM Thread线程的作用,应该真相大白了,就是读取队列,然后执行相应的操作。有兴趣你可以继续追进去看看源代码,我这里就不追下去了。

总结

看了这么多代码,确实很头疼,总结下吧。

jstack是通过与jvm进程建立socket连接,然后发送指令来实现相关操作。

jvm的Attach Listener线程监听套接字,读取jstack发来的指令,然后将相关的操作扔给VM Thread线程来执行,最后返回给jstack。

在jvm启动的时候,如果没有指定StartAttachListener,Attach Listener线程是不会启动的,在Signal Dispatcher线程收到SIGBREAK信号时,会调用 AttachListener::is_init_trigger()通过调用用AttachListener::init()启动了Attach Listener 线程。

END



往期作者热门文章


注:作者赞赏码,请大家支持鼓励!

作者:清泉

介绍:多年从事互联网开发,以Java后端开发为主,其它为辅。先后在某上市游戏公司做页游服务端开发,也曾在创业公司做移动APP后端架构设计,目前在阿里巴巴从事资深研发的工作。

点击“阅读原文”,加入作者知识星球深度交流!

    您可能也对以下帖子感兴趣

    文章有问题?点此查看未经处理的缓存