曹工杂谈：Java 类加载还会死锁？这是什么情况？

一、前言

今天事不是很多，正好在Java交流群里，看到一个比较有意思的问题，于是花了点时间研究了一下，这里做个简单的分享。

先贴一份测试代码，大家可以先猜测一下，执行结果会是怎样的：

 import java.util.concurrent.TimeUnit;

 public class TestClassLoading {
     public static class A{
         static {
             System.out.println("class A init");
             try {
                 TimeUnit.SECONDS.sleep(1);
             } catch (InterruptedException e) {
                 e.printStackTrace();
             }
             new B();
         }

         public static void test() {
             System.out.println("aaa");
         }
     }

     public static class B{
         static {
             System.out.println("class B init");
             new A();
         }

         public static void test() {
             System.out.println("bbb");
         }
     }
     public static void main(String[] args) {
         new Thread(() -> A.test()).start();
         new Thread(() -> B.test()).start();
     }
 }

不知道，你猜对了没有呢，实际的执行结果会是下面这样的：

二、原因分析

这里，一开始大家分析的是，和new有关系；但下面的代码和上面的结果完全一致，基本可以排除 new 的嫌疑：

 public class TestClassLoadingNew {
     public static class A{
         static {
             System.out.println("class A init");
             try {
                 TimeUnit.SECONDS.sleep(1);
             } catch (InterruptedException e) {
                 e.printStackTrace();
             }
10             B.test();
         }

         public static void test() {
             System.out.println("aaa");
         }
     }

     public static class B{
         static {
             System.out.println("class B init");
21             A.test();
         }

         public static void test() {
             System.out.println("bbb");
         }
     }
     public static void main(String[] args) {
         new Thread(() -> A.test()).start();
         new Thread(() -> B.test()).start();
     }
 }

这里，问题的根本原因，其实是：

classloader在初始化一个类的时候，会对当前类加锁，加锁后，再执行类的静态初始化块。

所以，上面会发生：

1、线程1：类A对class A加锁，加锁后，执行类的静态初始化块（在堆栈里体现为<clinit>函数），发现用到了class B，于是去加载B；

2、线程2：类B对class B加锁，加锁后，执行类的静态初始化块（在堆栈里体现为<clinit>函数），发现用到了class A，于是去加载A；

3、死锁发生。

有经验的同学，对于死锁是毫无畏惧的，因为我们有神器，jstack。 jstack 加上 -l 参数，即可打印出各个线程持有的锁的信息。（windows上直接jconsole就行，还能死锁检测）：

"Thread-1" #15 prio=5 os_prio=0 tid=0x000000002178a000 nid=0x2df8 in Object.wait() [0x0000000021f4e000]
   java.lang.Thread.State: RUNNABLE
        at com.dmtest.netty_learn.TestClassLoading$B.<clinit>(TestClassLoading.java:32)
        at com.dmtest.netty_learn.TestClassLoading.lambda$main$1(TestClassLoading.java:42)
        at com.dmtest.netty_learn.TestClassLoading$$Lambda$2/736709391.run(Unknown Source)
        at java.lang.Thread.run(Thread.java:748)

   Locked ownable synchronizers:
        - None

"Thread-0" #14 prio=5 os_prio=0 tid=0x0000000021787800 nid=0x2618 in Object.wait() [0x00000000213be000]
   java.lang.Thread.State: RUNNABLE
        at com.dmtest.netty_learn.TestClassLoading$A.<clinit>(TestClassLoading.java:21)
        at com.dmtest.netty_learn.TestClassLoading.lambda$main$0(TestClassLoading.java:41)
        at com.dmtest.netty_learn.TestClassLoading$$Lambda$1/611437735.run(Unknown Source)
        at java.lang.Thread.run(Thread.java:748)

   Locked ownable synchronizers:
        - None

这里，很奇怪的一个原因是，明明这两个线程发生了死锁，为什么没有显示呢？

因为，这是 jvm 内部加了锁，所以，jconsole、jstack都失效了。

三、一起深入JVM，探个究竟

1、单步跟踪

class 的加载都是由 classloader 来完成的，而且部分工作是在 jvm 层面完成，我们可以看到，在 java.lang.ClassLoader#defineClass1 的定义中：

以上几个方法都是本地方法。

其实际的实现在：/home/ckl/openjdk-jdk8u/jdk/src/share/native/java/lang/ClassLoader.c，

 JNIEXPORT jclass JNICALL
 Java_java_lang_ClassLoader_defineClass1(JNIEnv *env,
                                         jobject loader,
                                         jstring name,
                                         jbyteArray data,
                                         jint offset,
                                         jint length,
                                         jobject pd,
                                         jstring source)
 {
     jbyte *body;
     char *utfName;
     jclass result = ;
     char buf[];
     char* utfSource;
     char sourceBuf[];

     if (data == NULL) {
         JNU_ThrowNullPointerException(env, );
         return ;
     }

     /* Work around 4153825. malloc crashes on Solaris when passed a
      * negative size.
      */
     if (length < ) {
         JNU_ThrowArrayIndexOutOfBoundsException(env, );
         return ;
     }

     body = (jbyte *)malloc(length);

     if (body == ) {
         JNU_ThrowOutOfMemoryError(env, );
         return ;
     }

     (*env)->GetByteArrayRegion(env, data, offset, length, body);

     if ((*env)->ExceptionOccurred(env))
         goto free_body;

     if (name != NULL) {
         utfName = getUTF(env, name, buf, sizeof(buf));
         if (utfName == NULL) {
             goto free_body;
         }
         VerifyFixClassname(utfName);
     } else {
         utfName = NULL;
     }

     if (source != NULL) {
         utfSource = getUTF(env, source, sourceBuf, sizeof(sourceBuf));
         if (utfSource == NULL) {
             goto free_utfName;
         }
     } else {
         utfSource = NULL;
     }
61     result = JVM_DefineClassWithSource(env, utfName, loader, body, length, pd, utfSource);

     if (utfSource && utfSource != sourceBuf)
         free(utfSource);

  free_utfName:
     if (utfName && utfName != buf)
         free(utfName);

  free_body:
     free(body);
     return result;
 }

大家可以跟着标红的代码，我们一起大概看一下，这个方法的实现在/home/ckl/openjdk-jdk8u/hotspot/src/share/vm/prims/jvm.cpp 中，

 JVM_ENTRY(jclass, JVM_DefineClassWithSource(JNIEnv *env, const char *name, jobject loader, const jbyte *buf, jsize len, jobject pd, const char *source))
   JVMWrapper2("JVM_DefineClassWithSource %s", name);

4   return jvm_define_class_common(env, name, loader, buf, len, pd, source, true, THREAD);
 JVM_END

jvm_define_class_common 的实现，还是在  jvm.cpp 中，

 // common code for JVM_DefineClass() and JVM_DefineClassWithSource()
 // and JVM_DefineClassWithSourceCond()
 static jclass jvm_define_class_common(JNIEnv *env, const char *name,
                                       jobject loader, const jbyte *buf,
                                       jsize len, jobject pd, const char *source,
                                       jboolean verify, TRAPS) {
   if (source == NULL)  source = "__JVM_DefineClass__";

   assert(THREAD->is_Java_thread(), "must be a JavaThread");
   JavaThread* jt = (JavaThread*) THREAD;

   PerfClassTraceTime vmtimer(ClassLoader::perf_define_appclass_time(),
                              ClassLoader::perf_define_appclass_selftime(),
                              ClassLoader::perf_define_appclasses(),
                              jt->get_thread_stat()->perf_recursion_counts_addr(),
                              jt->get_thread_stat()->perf_timers_addr(),
                              PerfClassTraceTime::DEFINE_CLASS);

   if (UsePerfData) {
     ClassLoader::perf_app_classfile_bytes_read()->inc(len);
   }

   // Since exceptions can be thrown, class initialization can take place
   // if name is NULL no check for class name in .class stream has to be made.
   TempNewSymbol class_name = NULL;
   if (name != NULL) {
     const int str_len = (int)strlen(name);
     if (str_len > Symbol::max_length()) {
       // It's impossible to create this class;  the name cannot fit
       // into the constant pool.
       THROW_MSG_0(vmSymbols::java_lang_NoClassDefFoundError(), name);
     }
     class_name = SymbolTable::new_symbol(name, str_len, CHECK_NULL);
   }

   ResourceMark rm(THREAD);
   ClassFileStream st((u1*) buf, len, (char *)source);
   Handle class_loader (THREAD, JNIHandles::resolve(loader));
   if (UsePerfData) {
     is_lock_held_by_thread(class_loader,
                            ClassLoader::sync_JVMDefineClassLockFreeCounter(),
                            THREAD);
   }
   Handle protection_domain (THREAD, JNIHandles::resolve(pd));
45   Klass* k = SystemDictionary::resolve_from_stream(class_name, class_loader,
                                                      protection_domain, &st,
                                                      verify != ,
                                                      CHECK_NULL);

   if (TraceClassResolution && k != NULL) {
     trace_class_resolution(k);
   }

   return (jclass) JNIHandles::make_local(env, k->java_mirror());
 }

resolve_from_stream 的实现在 SystemDictionary 类中，下面我们看下：

 Klass* SystemDictionary::resolve_from_stream(Symbol* class_name,
                                              Handle class_loader,
                                              Handle protection_domain,
                                              ClassFileStream* st,
                                              bool verify,
                                              TRAPS) {

   // Classloaders that support parallelism, e.g. bootstrap classloader,
   // or all classloaders with UnsyncloadClass do not acquire lock here
   bool DoObjectLock = true;
   if (is_parallelCapable(class_loader)) {
     DoObjectLock = false;
   }

   ClassLoaderData* loader_data = register_loader(class_loader, CHECK_NULL);

   // Make sure we are synchronized on the class loader before we proceed
18   Handle lockObject = compute_loader_lock_object(class_loader, THREAD);
19   check_loader_lock_contention(lockObject, THREAD);
20   ObjectLocker ol(lockObject, THREAD, DoObjectLock);

   TempNewSymbol parsed_name = NULL;

   // Parse the stream. Note that we do this even though this klass might
   // already be present in the SystemDictionary, otherwise we would not
   // throw potential ClassFormatErrors.
   //
   // Note: "name" is updated.

   instanceKlassHandle k = ClassFileParser(st).parseClassFile(class_name,
                                                              loader_data,
                                                              protection_domain,
                                                              parsed_name,
                                                              verify,
                                                              THREAD);

   const char* pkg = "java/";
   size_t pkglen = strlen(pkg);
   if (!HAS_PENDING_EXCEPTION &&
       !class_loader.is_null() &&
       parsed_name != NULL &&
       parsed_name->utf8_length() >= (int)pkglen &&
       !strncmp((const char*)parsed_name->bytes(), pkg, pkglen)) {
44     // It is illegal to define classes in the "java." package from
45     // JVM_DefineClass or jni_DefineClass unless you're the bootclassloader
     ResourceMark rm(THREAD);
     char* name = parsed_name->as_C_string();
     char* index = strrchr(name, '/');
     assert(index != NULL, "must be");
     *index = '\0'; // chop to just the package name
     while ((index = strchr(name, '/')) != NULL) {
       *index = '.'; // replace '/' with '.' in package name
     }
     const char* fmt = "Prohibited package name: %s";
     size_t len = strlen(fmt) + strlen(name);
     char* message = NEW_RESOURCE_ARRAY(char, len);
     jio_snprintf(message, len, fmt, name);
     Exceptions::_throw_msg(THREAD_AND_LOCATION,
       vmSymbols::java_lang_SecurityException(), message);
   }

   if (!HAS_PENDING_EXCEPTION) {
     assert(parsed_name != NULL, "Sanity");
     assert(class_name == NULL || class_name == parsed_name, "name mismatch");
     // Verification prevents us from creating names with dots in them, this
     // asserts that that's the case.
     assert(is_internal_format(parsed_name),
            "external class name format used internally");

     // Add class just loaded
     // If a class loader supports parallel classloading handle parallel define requests
     // find_or_define_instance_class may return a different InstanceKlass
     if (is_parallelCapable(class_loader)) {
       k = find_or_define_instance_class(class_name, class_loader, k, THREAD);
     } else {
76       define_instance_class(k, THREAD);
     }
   }

   return k();
 }

上面的方法里，有几处值得注意的：

1：18-20行，进行了加锁，18行获取锁对象，这里是当前类加载器（从注释可以看出），20行就是加锁的语法

2：37-60行，这里是判断要加载的类的包名是否以 java 开头，以 java 开头的类是非法的，不能加载

3：第76行， define_instance_class(k, THREAD); 进行后续操作

接下来，我们看看 define_instance_class 的实现：

 void SystemDictionary::define_instance_class(instanceKlassHandle k, TRAPS) {

   ClassLoaderData* loader_data = k->class_loader_data();
   Handle class_loader_h(THREAD, loader_data->class_loader());

   for (uintx it = ; it < GCExpandToAllocateDelayMillis; it++){}

  // for bootstrap and other parallel classloaders don't acquire lock,
  // use placeholder token
  // If a parallelCapable class loader calls define_instance_class instead of
  // find_or_define_instance_class to get here, we have a timing
  // hole with systemDictionary updates and check_constraints
  if (!class_loader_h.is_null() && !is_parallelCapable(class_loader_h)) {
     assert(ObjectSynchronizer::current_thread_holds_lock((JavaThread*)THREAD,
          compute_loader_lock_object(class_loader_h, THREAD)),
          "define called without lock");
   }

   // Check class-loading constraints. Throw exception if violation is detected.
   // Grabs and releases SystemDictionary_lock
   // The check_constraints/find_class call and update_dictionary sequence
   // must be "atomic" for a specific class/classloader pair so we never
   // define two different instanceKlasses for that class/classloader pair.
   // Existing classloaders will call define_instance_class with the
   // classloader lock held
   // Parallel classloaders will call find_or_define_instance_class
   // which will require a token to perform the define class
   Symbol*  name_h = k->name();
   unsigned int d_hash = dictionary()->compute_hash(name_h, loader_data);
   int d_index = dictionary()->hash_to_index(d_hash);
   check_constraints(d_index, d_hash, k, class_loader_h, true, CHECK);

   // Register class just loaded with class loader (placed in Vector)
   // Note we do this before updating the dictionary, as this can
   // fail with an OutOfMemoryError (if it does, we will *not* put this
   // class in the dictionary and will not update the class hierarchy).
   // JVMTI FollowReferences needs to find the classes this way.
   if (k->class_loader() != NULL) {
     methodHandle m(THREAD, Universe::loader_addClass_method());
     JavaValue result(T_VOID);
     JavaCallArguments args(class_loader_h);
     args.push_oop(Handle(THREAD, k->java_mirror()));
     JavaCalls::call(&result, m, &args, CHECK);
   }

   // Add the new class. We need recompile lock during update of CHA.
   {
     unsigned int p_hash = placeholders()->compute_hash(name_h, loader_data);
     int p_index = placeholders()->hash_to_index(p_hash);

     MutexLocker mu_r(Compile_lock, THREAD);

     // Add to class hierarchy, initialize vtables, and do possible
     // deoptimizations.
     add_to_hierarchy(k, CHECK); // No exception, but can block

     // Add to systemDictionary - so other classes can see it.
     // Grabs and releases SystemDictionary_lock
     update_dictionary(d_index, d_hash, p_index, p_hash,
                       k, class_loader_h, THREAD);
   }
62   k->eager_initialize(THREAD);

   // notify jvmti
   if (JvmtiExport::should_post_class_load()) {
       assert(THREAD->is_Java_thread(), "thread->is_Java_thread()");
       JvmtiExport::post_class_load((JavaThread *) THREAD, k());

   }

 }

这里，由于我们的案例中，是class A 在初始化过程中出现死锁，所以我们关注第62行，eager_initialize：

 void InstanceKlass::eager_initialize(Thread *thread) {
   if (!EagerInitialization) return;

   if (this->is_not_initialized()) {
     // abort if the the class has a class initializer
     if (this->class_initializer() != NULL) return;

     // abort if it is java.lang.Object (initialization is handled in genesis)
     Klass* super = this->super();
     if (super == NULL) return;

     // abort if the super class should be initialized
     if (!InstanceKlass::cast(super)->is_initialized()) return;

     // call body to expose the this pointer
     instanceKlassHandle this_oop(thread, this);
17     eager_initialize_impl(this_oop);
   }
 }

我们接着进入 eager_initialize_impl，该方法进入到了 InstanceKlass：

 void InstanceKlass::eager_initialize_impl(instanceKlassHandle this_oop) {
   EXCEPTION_MARK;
 3   oop init_lock = this_oop->init_lock();
 4   ObjectLocker ol(init_lock, THREAD, init_lock != NULL);

   // abort if someone beat us to the initialization
   if (!this_oop->is_not_initialized()) return;  // note: not equivalent to is_initialized()

   ClassState old_state = this_oop->init_state();
   link_class_impl(this_oop, true, THREAD);
   if (HAS_PENDING_EXCEPTION) {
     CLEAR_PENDING_EXCEPTION;
     // Abort if linking the class throws an exception.

     // Use a test to avoid redundantly resetting the state if there's
     // no change.  Set_init_state() asserts that state changes make
     // progress, whereas here we might just be spinning in place.
     if( old_state != this_oop->_init_state )
       this_oop->set_init_state (old_state);
   } else {
     // linking successfull, mark class as initialized
     this_oop->set_init_state (fully_initialized);
     this_oop->fence_and_clear_init_lock();
     // trace
     if (TraceClassInitialization) {
       ResourceMark rm(THREAD);
       tty->print_cr("[Initialized %s without side effects]", this_oop->external_name());
     }
   }
 }

这里，我们重点关注第3,4行：

1、第3行，获取初始化锁；

2、第4行，加锁

2、获取初始化锁并加锁

这里，我们首先获取锁的操作，

 oop InstanceKlass::init_lock() const {
   // return the init lock from the mirror
   oop lock = java_lang_Class::init_lock(java_mirror());
   // Prevent reordering with any access of initialization state
   OrderAccess::loadload();
   assert((oop)lock != NULL || !is_not_initialized(), // initialized or in_error state
          "only fully initialized state can have a null lock");
   return lock;
 }

其中，java_mirror() 方法就是返回 Klass 类中的以下字段：

   // java/lang/Class instance mirroring this class
   oop       _java_mirror;

再看 init_lock 方法：

 oop java_lang_Class::init_lock(oop java_class) {
   assert(_init_lock_offset != , "must be set");
   return java_class->obj_field(_init_lock_offset);
 }

这里呢，应该就是获取 我们传入的 java_class 中的某个字段，该字段就是充当 init_lock。（个人水平有限，还请指正）

下面为加锁操作的语句：

   ObjectLocker ol(init_lock, THREAD, init_lock != NULL);

 / ObjectLocker enforced balanced locking and can never thrown an
 // IllegalMonitorStateException. However, a pending exception may
 // have to pass through, and we must also be able to deal with
 // asynchronous exceptions. The caller is responsible for checking
 // the threads pending exception if needed.
 // doLock was added to support classloading with UnsyncloadClass which
 // requires flag based choice of locking the classloader lock.
 class ObjectLocker : public StackObj {
  private:
   Thread*   _thread;
   Handle    _obj;
   BasicLock _lock;
   bool      _dolock;   // default true
  public:
15   ObjectLocker(Handle obj, Thread* thread, bool doLock = true);

 // -----------------------------------------------------------------------------
 // Internal VM locks on java objects
 // standard constructor, allows locking failures
 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
   _dolock = doLock;
   _thread = thread;
   _obj = obj;

10   if (_dolock) {
     TEVENT (ObjectLocker) ;

     ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
   }
 }

接下来会进入到 synchronizer.cpp，

 // -----------------------------------------------------------------------------
 //  Fast Monitor Enter/Exit
 // This the fast monitor enter. The interpreter and compiler use
 // some assembly copies of this code. Make sure update those code
 // if the following function is changed. The implementation is
 // extremely sensitive to race condition. Be careful.

 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
  if (UseBiasedLocking) {
     if (!SafepointSynchronize::is_at_safepoint()) {
       BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
       if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
         return;
       }
     } else {
       assert(!attempt_rebias, "can not rebias toward VM thread");
       BiasedLocking::revoke_at_safepoint(obj);
     }
     assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
  }

22  slow_enter (obj, lock, THREAD) ;
 }

上面会判断，是否使用偏向锁，如果不使用，则走 slow_enter 。

 // -----------------------------------------------------------------------------
 // Interpreter/Compiler Slow Case
 // This routine is used to handle interpreter/compiler slow case
 // We don't need to use fast path here, because it must have been
 // failed in the interpreter/compiler code.
 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
   markOop mark = obj->mark();
   assert(!mark->has_bias_pattern(), "should not see bias pattern here");

   if (mark->is_neutral()) {
     // Anticipate successful CAS -- the ST of the displaced mark must
     // be visible <= the ST performed by the CAS.
     lock->set_displaced_header(mark);
     if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
       TEVENT (slow_enter: release stacklock) ;
       return ;
     }
     // Fall through to inflate() ...
   } else
   if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
     assert(lock != mark->locker(), "must not re-lock the same lock");
     assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
     lock->set_displaced_header(NULL);
     return;
   }

   // The object header will never be displaced to this lock,
   // so it does not matter what the value is, except that it
   // must be non-zero to avoid looking like a re-entrant lock,
   // and must not look locked either.
   lock->set_displaced_header(markOopDesc::unused_mark());
   ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
 }

这里的代码结合注释，能大概看出来是，前面部分为轻量级锁，这里先不展开了，锁这块都可以单独写了。有兴趣的读者可以自行阅读。

四、总结

这里再说下结论吧，类初始化的过程，会对class加锁，再执行class的初始化，如果这时候发生了循环依赖，就会导致死锁。

如果有读者对上面的c++代码感兴趣，可以参考下面的文章，搭建调试环境：

源码编译OpenJdk 8，Netbeans调试Java原子类在JVM中的实现（Ubuntu 16.04）