【Dubbo】默认hession2反序列化机制导致dubbo接口返回HashMap

秒速五厘米 2024-02-05 17:11 39阅读 0赞

## 问题描述 ##

*  在使用`dubbo`调用接口的时候，莫名其妙出现`java.lang.ClassCastException: java.util.HashMap cannot be cast to xxxx`异常
 *  经过排查发现，是因为`dubbo`接口返回的不是`xxxx`对象，而是`HashMap`

## 源码分析 ##

*  `dubbo`的反序列化机制默认是`hessian2`
 *  首先定位到`SerializerFactory`类的`getDeserializer()`方法

try {
    //  Class cl = Class.forName(type, false, _loader);
        Class cl = loadSerializedClass(type);
        deserializer = getDeserializer(cl);
    } catch (Exception e) {
        log.warning("Hessian/Burlap: '" + type + "' is an unknown class in " + _loader + ":\n" + e);
        _typeNotFoundDeserializerMap.put(type, PRESENT);
        log.log(Level.FINER, e.toString(), e);
        _unrecognizedTypeCache.put(type, new AtomicLong(1L));
    }

*  断点发现`deserializer = getDeserializer(cl);`这一行代码抛了异常，所以能看到日志输出

log.warning("Hessian/Burlap: '" + type + "' is an unknown class in " + _loader + ":\n" + e);

*  然后手工试了`Class.forName`，确实会报`java.lang.ClassNotFoundException`异常
 *  那问题就与类加载器有关了，看看`hessian2`的类加载器是哪个
 *  打开`org.apache.dubbo.common.serialize.hessian2.Hessian2SerializerFactory`这个类

public class Hessian2SerializerFactory extends SerializerFactory {
        
    
        @Override
        public ClassLoader getClassLoader() {
        
            return Thread.currentThread().getContextClassLoader();
        }
    }

*  `Hessian2`重写了类加载器，为了观察这是哪个类加载器，增加日志输出

public class Hessian2SerializerFactory extends SerializerFactory {
        
    
        @Override
        public ClassLoader getClassLoader() {
        
            ClassLoader classLoaderTest = Thread.currentThread().getContextClassLoader();
            loadClass(classLoaderTest, "com.xxxx");
            ClassLoader parent = classLoaderTest.getParent();
            while(parent!=null) {
        
                loadClass(classLoaderTest, "com.xxxx");
                parent = parent.getParent();
            }
            return classLoaderTest;
        }
    
        private void loadClass(ClassLoader classLoader, String className) {
        
            try {
        
                classLoader.loadClass(className);
                log.info("load success: {}", classLoader.getClass().getName());
            } catch (ClassNotFoundException e) {
        
                log.info("load fail: {}", classLoader.getClass().getName());
            }
        }
    }

*  经过测试发现，能正常加载类的类加载器是`org.springframework.boot.web.embedded.tomcat.TomcatEmbeddedWebappClassLoader`和`org.springframework.boot.loader.LaunchedURLClassLoader`，两个类加载器都是`Spring`的
 *  而不可以加载到类的类加载器是`jdk.internal.loader.ClassLoaders.AppClassLoader`
 *  所以，到了这里，问题就很明显了，`dubbo`消费者在反序列化的时候，由于加载不到这个类，所以返回的对象按`HashMap`处理

## 解决方案 ##

*  参考`org.springframework.util.ClassUtils`工具类的`getDefaultClassLoader()`方法

@Nullable
    public static ClassLoader getDefaultClassLoader() {
      ClassLoader cl = null;
      try {
        cl = Thread.currentThread().getContextClassLoader();
      }
      catch (Throwable ex) {
        // Cannot access thread context ClassLoader - falling back...
      }
      if (cl == null) {
        // No thread context class loader -> use class loader of this class.
        cl = ClassUtils.class.getClassLoader();
        if (cl == null) {
          // getClassLoader() returning null indicates the bootstrap ClassLoader
          try {
            cl = ClassLoader.getSystemClassLoader();
          }
          catch (Throwable ex) {
            // Cannot access system ClassLoader - oh well, maybe the caller can live with null...
          }
        }
      }
      return cl;
    }

*  如果`Thread.currentThread().getContextClassLoader()`返回的是JDK类加载器
 *  那么就使用`ClassUtils.class.getClassLoader()`代替
 *  修改后的代码如下

@Override
        public ClassLoader getClassLoader() {
            ClassLoader classLoader = ClassUtils.getDefaultClassLoader();
            if(Objects.nonNull(classLoader) && classLoader.getClass().getName().contains("spring")) {
                return classLoader;
            }
            ClassLoader classLoaderNew = ClassUtils.class.getClassLoader();
            log.info("classLoader is wrong: {}, change to: {}", classLoader, classLoaderNew);
            return classLoaderNew;
        }

*  可以看到日志输出

[2023-12-20 18:56:48.683] [INFO] [ForkJoinPool.commonPool-worker-84]c.m.v.i.d.h.xxxxHessian2SerializerFactory[56][]- classLoader is wrong: jdk.internal.loader.ClassLoaders$AppClassLoader@73d16e93, change to: org.springframework.boot.loader.LaunchedURLClassLoader@14514713

*  最终`dubbo`接口可以正常返回`Java`对象，而不是`HashMap`了

## 利用dubbo的SPI机制 ##

*  在`resources`目录下新建文件`META-INF/dubbo/org.apache.dubbo.common.serialize.hessian2.dubbo.Hessian2FactoryInitializer`
 *  文件内容如下

default=xxxx.xxxxHessian2FactoryInitializer

*  这样可以指定自定义的Hessian2FactoryInitializer
 *  在这个Hessian2FactoryInitializer里面创建自定义的Hessian2SerializerFactory

@Slf4j
    public class xxxxHessian2FactoryInitializer extends DefaultHessian2FactoryInitializer {
        
    
        @Override
        public SerializerFactory getSerializerFactory() {
        
            Hessian2SerializerFactory hessian2SerializerFactory = new xxxxHessian2SerializerFactory();
            hessian2SerializerFactory.getClassFactory().allow(RuntimeException.class.getName());
            hessian2SerializerFactory.setAllowNonSerializable(Boolean.parseBoolean(System.getProperty("dubbo.hessian.allowNonSerializable", "false")));
            return hessian2SerializerFactory;
        }
    }

*  剩下的就以前面的逻辑全部串起来了

## 总结 ##

*  虽然找到了问题的原因，也解决了问题，但依然有疑问，如果启动多次，有时候能正常转成`Java`对象，是一个随机的状态，感觉启动的时候决定了这个`hessian2`的类加载器
 *  遇到一些莫名其妙的问题时，不要慌，耐心的断点，分析代码查找问题的原因
 *  解决问题时，最好深刻理解源码的设计，利用原有框架的扩展性，进行问题的修复，这样的代码是最优雅的