转载

JVM源码分析之自定义类加载器如何拉长YGC

概述

本文重点讲述毕玄大师在其公众号上发的一个GC问题 一个jstack/jmap等不能用的case (PS:话说毕大师超级喜欢在题目里用case这个词,我觉得题目还是能尽量做到顾名思义好,不然要找起相关文章来真的好难找),对于毕大师那篇文章,题目上没有提到GC的那个问题,不过进入到文章里可以看到,既然文章提到了jstack/jmap的问题,这里也简单回答下jstack/jmap无法使用的问题,其实最常见的场景是使用jstack/jmap的用户和目标进程不是同一个用户,哪怕你执行jstack/jmap的动作是root用户也无济于事,详情可以参考我的这篇文章,JVM Attach机制实现,主要是讲JVM Attach机制的,不过毕大师这里主要提到的是jmap -heap/histo这两个参数带来的问题,如果使用-heap/histo的参数,其实和大家使用-F参数是一样的,底层都是通过serviceability agent来实现的,并不是jvm attach的方式,通过sa连上去之后会挂起进程,在serviceability agent里存在bug可能导致detach的动作不会被执行,从而会让进程一直挂着,可以通过top命令验证进程是否处于T状态,如果是说明进程被挂起了,如果进程被挂起了,可以通过kill -CONT [pid]来恢复。

再回到那个GC的问题,用的参数如下:

-XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xms512m -Xmx512m -Xmn100m -XX:+UseConcMarkSweepGC

demo程序如下:

import com.thoughtworks.xstream.XStream;   public class XStreamTest {           public static void main(String[] args) throws Exception {         while(true){             XStream xs = new XStream();             xs.toString();             xs = null;         }     }   }

执行效果如下

2016-03-14T22:48:01.502+0800: [GC [ParNew: 327680K->4258K(368640K), 0.0179071 secs] 327680K->4258K(1007616K), 0.0179448 secs] [Times: user=0.06 sys=0.01, real=0.01 secs]  2016-03-14T22:48:05.975+0800: [GC [ParNew: 331938K->10239K(368640K), 0.0336279 secs] 331938K->10239K(1007616K), 0.0336593 secs] [Times: user=0.13 sys=0.02, real=0.03 secs]  2016-03-14T22:48:12.215+0800: [GC [ParNew: 337919K->14444K(368640K), 0.0471005 secs] 337919K->14444K(1007616K), 0.0471257 secs] [Times: user=0.19 sys=0.02, real=0.05 secs]  2016-03-14T22:48:21.768+0800: [GC [ParNew: 342124K->19088K(368640K), 0.0605017 secs] 342124K->19088K(1007616K), 0.0605295 secs] [Times: user=0.26 sys=0.03, real=0.06 secs]  2016-03-14T22:48:35.180+0800: [GC [ParNew: 346768K->20633K(368640K), 0.0993470 secs] 346768K->25248K(1007616K), 0.0993777 secs] [Times: user=0.34 sys=0.04, real=0.09 secs]

发现gc的时间越来越长,但是gc触发的时机以及回收的效果都差不多,那问题究竟在哪里呢?

Demo分析

虽然这个demo代码逻辑很简单,但是其实这是一个特殊的demo,并不简单,如果我们将XStream对象换成Object对象,会发现不存在这个问题,既然如此那有必要进去看看这个XStream的构造函数:

 public XStream() {         this((ReflectionProvider)null, (Mapper)((Mapper)null), (HierarchicalStreamDriver)(new XppDriver()));     }      /** @deprecated */     public XStream(ReflectionProvider reflectionProvider, Mapper mapper, HierarchicalStreamDriver driver) {         this(reflectionProvider, driver, (ClassLoader)(new CompositeClassLoader()), mapper);     }      /** @deprecated */     public XStream(ReflectionProvider reflectionProvider, HierarchicalStreamDriver driver, ClassLoader classLoader, Mapper mapper) {         this(reflectionProvider, driver, new ClassLoaderReference(classLoader), mapper, new DefaultConverterLookup());     }      public XStream(ReflectionProvider reflectionProvider, HierarchicalStreamDriver driver, ClassLoaderReference classLoader, Mapper mapper, final DefaultConverterLookup defaultConverterLookup) {         this(reflectionProvider, driver, (ClassLoaderReference)classLoader, mapper, new ConverterLookup() {             public Converter lookupConverterForType(Class type) {                 return defaultConverterLookup.lookupConverterForType(type);             }         }, new ConverterRegistry() {             public void registerConverter(Converter converter, int priority) {                 defaultConverterLookup.registerConverter(converter, priority);             }         });     }      /** @deprecated */     public XStream(ReflectionProvider reflectionProvider, HierarchicalStreamDriver driver, ClassLoader classLoader, Mapper mapper, ConverterLookup converterLookup, ConverterRegistry converterRegistry) {         this(reflectionProvider, driver, (ClassLoaderReference)(new ClassLoaderReference(classLoader)), mapper, converterLookup, converterRegistry);     }      public XStream(ReflectionProvider reflectionProvider, HierarchicalStreamDriver driver, ClassLoaderReference classLoaderReference, Mapper mapper, ConverterLookup converterLookup, ConverterRegistry converterRegistry) {         if(reflectionProvider == null) {             reflectionProvider = JVM.newReflectionProvider();         }          this.reflectionProvider = reflectionProvider;         this.hierarchicalStreamDriver = driver;         this.classLoaderReference = classLoaderReference;         this.converterLookup = converterLookup;         this.converterRegistry = converterRegistry;         this.mapper = mapper == null?this.buildMapper():mapper;         this.setupMappers();         this.setupSecurity();         this.setupAliases();         this.setupDefaultImplementations();         this.setupConverters();         this.setupImmutableTypes();         this.setMode(1003);     }

这个构造函数还是很复杂的,里面会创建很多的对象,上面还有一些方法实现我就不贴了,总之都是在不断构建各种大大小小的对象,一个XStream对象构建出来的时候大概好像有12M的样子。

那到底是哪些对象会导致ygc不断增长呢,于是可能想到逐步替换上面这些逻辑,比如将最后一个构造函数里的那些逻辑都禁掉,然后我们再跑测试看看还会不会让ygc不断恶化,最终我们会发现,如果我们直接使用如下构造函数构造对象时,如果传入的classloader是AppClassLoader,那会发现这个问题不再出现了。

 public XStream(ReflectionProvider reflectionProvider, HierarchicalStreamDriver driver, ClassLoader classLoader, Mapper mapper) {         this(reflectionProvider, driver, new ClassLoaderReference(classLoader), mapper, new DefaultConverterLookup());  }

测试代码如下:

 public static void main(String[] args) throws Exception {         int i=0;         while (true) {             XStream xs = new XStream(null,null, new ClassLoaderReference(XStreamTest.class.getClassLoader()),null, new DefaultConverterLookup());             xs.toString();             xs=null;         }   }

gc日志如下:

2016-03-14T23:10:33.537+0800: [GC [ParNew: 327680K->758K(368640K), 0.0019803 secs] 327680K->758K(1007616K), 0.0020182 secs] [Times: user=0.01 sys=0.00, real=0.00 secs]  2016-03-14T23:10:35.189+0800: [GC [ParNew: 328438K->1066K(368640K), 0.0018641 secs] 328438K->1066K(1007616K), 0.0019055 secs] [Times: user=0.01 sys=0.00, real=0.01 secs]  2016-03-14T23:10:36.465+0800: [GC [ParNew: 328746K->1156K(368640K), 0.0010304 secs] 328746K->1156K(1007616K), 0.0010519 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]  2016-03-14T23:10:37.767+0800: [GC [ParNew: 328836K->1065K(368640K), 0.0011329 secs] 328836K->1065K(1007616K), 0.0011543 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]  2016-03-14T23:10:39.035+0800: [GC [ParNew: 328745K->351K(368640K), 0.0043387 secs] 328745K->1127K(1007616K), 0.0043700 secs] [Times: user=0.01 sys=0.00, real=0.01 secs]  2016-03-14T23:10:40.324+0800: [GC [ParNew: 328031K->160K(368640K), 0.0011579 secs] 328807K->936K(1007616K), 0.0011793 secs] [Times: user=0.01 sys=0.00, real=0.00 secs]  2016-03-14T23:10:41.610+0800: [GC [ParNew: 327840K->31K(368640K), 0.0007010 secs] 328616K->826K(1007616K), 0.0007219 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]  2016-03-14T23:10:42.919+0800: [GC [ParNew: 327711K->24K(368640K), 0.0011246 secs] 328506K->819K(1007616K), 0.0011450 secs] [Times: user=0.00 sys=0.00, real=0.01 secs]  2016-03-14T23:10:44.196+0800: [GC [ParNew: 327704K->24K(368640K), 0.0006797 secs] 328499K->819K(1007616K), 0.0007586 secs] [Times: user=0.00 sys=0.00, real=0.00 secs] 

是不是觉得很神奇,由此可见,这个classloader至关重要。

不得不说的类加载器

这里着重要说的两个概念是 初始类加载器定义类加载器 。举个栗子说吧,AClassLoader->BClassLoader->CClassLoader,表示AClassLoader在加载类的时候会委托BClassLoader类加载器来加载,BClassLoader加载类的时候会委托CClassLoader来加载,假如我们使用AClassLoader来加载X这个类,而X这个类最终是被CClassLoader来加载的,那么我们称CClassLoader为X类的定义类加载器,而AClassLoader和BClassLoader分别为X类的初始类加载器,JVM在加载某个类的时候对这三种类加载器都会记录,记录的数据结构是一个叫做SystemDictionary的hashtable,其key是根据ClassLoader对象和类名算出来的hash值,而value是真正的由定义类加载器加载的Klass对象,因为初始类加载器和定义类加载器是不同的classloader,因此算出来的hash值也是不同的,因此在SystemDictionary里会有多项值的value都是指向同一个Klass对象。

那么JVM为什么要分这两种类加载器呢,其实主要是为了快速找到已经加载的类,比如我们已经通过AClassLoader来触发了对X类的加载,当我们再次使用AClassLoader这个类加载器来加载X这个类的时候就不需要再委托给BClassLoader去找了,因为加载过的类在JVM里有这个类加载器的直接加载的记录,只需要直接返回对应的Klass对象即可。

Demo中的类加载器是否会加载类

我们的demo里发现构建了一个CompositeClassLoader的类加载器,那到底有没有用这个类加载器加载类呢,我们可以设置一个断点在CompositeClassLoader的loadClass方法上,于是看到下面的堆栈:

main@1, prio=5, in group 'main', status: 'RUNNING'     at com.thoughtworks.xstream.core.util.CompositeClassLoader.loadClass(CompositeClassLoader.java:53)     at java.lang.Class.forName0(Class.java:-1)     at java.lang.Class.forName(Class.java:249)     at com.thoughtworks.xstream.XStream.buildMapperDynamically(XStream.java:191)     at com.thoughtworks.xstream.XStream.buildMapper(XStream.java:170)     at com.thoughtworks.xstream.XStream.<init>(XStream.java:142)     at com.thoughtworks.xstream.XStream.<init>(XStream.java:116)     at com.BBBB.main(BBBB.java:15)

可见确实有类加载的动作,根据类加载委托机制,在这个demo中我们能肯定类是交给AppClassLoader来加载的,这样一来CompositeClassLoader就变成了初始类加载器,而AppClassLoader会是定义类加载器,都会在SystemDictionary里存在,因此当我们不断new XStream的时候会不断new CompositeClassLoader对象,加载类的时候会不断往SystemDictionary里插入记录,从而使SystemDictionary越来越膨胀,那自然而然会想到如果GC过程不断去扫描这个SystemDictionary的话,那随着SystemDictionary不断膨胀,那么GC的效率也就越低,抱着验证下猜想的方式我们可以使用perf工具来看看,如果发现cpu占比排前的函数如果都是操作SystemDictionary的,那就基本验证了我们的说法,下面是perf工具的截图,基本证实了这一点。

JVM源码分析之自定义类加载器如何拉长YGC

SystemDictionary为什么会影响GC过程

想象一下这么个情况,我们加载了一个类,然后构建了一个对象(这个对象在eden里构建)当一个属性设置到这个类里,如果gc发生的时候,这个对象是不是要被找出来标活才行,那么自然而然我们加载的类肯定是我们一项重要的gc root,这样SystemDictionary就成为了gc过程中的被扫描对象了,事实也是如此,可以看vm的具体代码:

void SharedHeap::process_strong_roots(bool activate_scope,                                       bool collecting_perm_gen,                                       ScanningOption so,                                       OopClosure* roots,                                       CodeBlobClosure* code_roots,                                       OopsInGenClosure* perm_blk) {   StrongRootsScope srs(this, activate_scope);   // General strong roots.   assert(_strong_roots_parity != 0, "must have called prologue code");   // _n_termination for _process_strong_tasks should be set up stream   // in a method not running in a GC worker.  Otherwise the GC worker   // could be trying to change the termination condition while the task   // is executing in another GC worker.   if (!_process_strong_tasks->is_task_claimed(SH_PS_Universe_oops_do)) {     Universe::oops_do(roots);     // Consider perm-gen discovered lists to be strong.     //将perm gen的非强引用标记为root的一部分     perm_gen()->ref_processor()->weak_oops_do(roots);   }   // Global (strong) JNI handles   if (!_process_strong_tasks->is_task_claimed(SH_PS_JNIHandles_oops_do))     JNIHandles::oops_do(roots);   // All threads execute this; the individual threads are task groups.   if (ParallelGCThreads > 0) {     Threads::possibly_parallel_oops_do(roots, code_roots);   } else {     Threads::oops_do(roots, code_roots);   }   if (!_process_strong_tasks-> is_task_claimed(SH_PS_ObjectSynchronizer_oops_do))     ObjectSynchronizer::oops_do(roots);   if (!_process_strong_tasks->is_task_claimed(SH_PS_FlatProfiler_oops_do))     FlatProfiler::oops_do(roots);   if (!_process_strong_tasks->is_task_claimed(SH_PS_Management_oops_do))     Management::oops_do(roots);   if (!_process_strong_tasks->is_task_claimed(SH_PS_jvmti_oops_do))     JvmtiExport::oops_do(roots);    if (!_process_strong_tasks->is_task_claimed(SH_PS_SystemDictionary_oops_do)) {     if (so & SO_AllClasses) {       SystemDictionary::oops_do(roots);     } else if (so & SO_SystemClasses) {       SystemDictionary::always_strong_oops_do(roots);     }   }    if (!_process_strong_tasks->is_task_claimed(SH_PS_StringTable_oops_do)) {     //JavaObjectsInPerm为false,那么String intern的对象已经class对象都是存在heap里的,否则都存在perm里       if (so & SO_Strings || (!collecting_perm_gen && !JavaObjectsInPerm)) {       //虽然不回收perm,但是interned的String对象不在perm里,那么还是需要遍历下StringTable里的String对象,因为这些对象在heap里       StringTable::oops_do(roots);     }     if (JavaObjectsInPerm) {       // Verify the string table contents are in the perm gen       NOT_PRODUCT(StringTable::oops_do(&assert_is_perm_closure));     }   }    if (!_process_strong_tasks->is_task_claimed(SH_PS_CodeCache_oops_do)) {     if (so & SO_CodeCache) {       // (Currently, CMSCollector uses this to do intermediate-strength collections.)       assert(collecting_perm_gen, "scanning all of code cache");       assert(code_roots != NULL, "must supply closure for code cache");       if (code_roots != NULL) {         CodeCache::blobs_do(code_roots);       }     } else if (so & (SO_SystemClasses|SO_AllClasses)) {       if (!collecting_perm_gen) {         // If we are collecting from class statics, but we are not going to         // visit all of the CodeCache, collect from the non-perm roots if any.         // This makes the code cache function temporarily as a source of strong         // roots for oops, until the next major collection.         //         // If collecting_perm_gen is true, we require that this phase will call         // CodeCache::do_unloading.  This will kill off nmethods with expired         // weak references, such as stale invokedynamic targets.         CodeCache::scavenge_root_nmethods_do(code_roots);       }     }     // Verify that the code cache contents are not subject to     // movement by a scavenging collection.     DEBUG_ONLY(CodeBlobToOopClosure assert_code_is_non_scavengable(&assert_is_non_scavengable_closure, /*do_marking=*/ false));     DEBUG_ONLY(CodeCache::asserted_non_scavengable_nmethods_do(&assert_code_is_non_scavengable));   }    if (!collecting_perm_gen) {     //如果是不回收perm,那找出所有perm指向new的对象       // All threads perform this; coordination is handled internally.     rem_set()->younger_refs_iterate(perm_gen(), perm_blk);//perm的level是-1   }   _process_strong_tasks->all_tasks_completed(); } 

看上面的 SH_PS_SystemDictionary_oops_do task就知道了,这个就是对SystemDictionary进行扫描。

但是这里要说的是虽然有对SystemDictionary进行扫描,但是ygc的过程并不会对SystemDictionary进行处理,如果要对它进行处理需要开启类卸载的vm参数,CMS算法下,CMS GC和Full GC在开启CMSClassUnloadingEnabled的情况下是可能对类做卸载动作的,此时会对SystemDictionary进行清理,所以当我们在跑上面demo的时候,通过 jmap -dump:live,format=b,file=heap.bin <pid> 命令执行完之后,ygc的时间瞬间降下来了,不过又会慢慢回去,这是因为jmap的这个命令会做一次gc,这个gc过程会对SystemDictionary进行清理。

修改VM代码验证

很遗憾hotspot目前没有对ygc的每个task做一个时间的统计,因此无法直接知道是不是 SH_PS_SystemDictionary_oops_do 这个task导致了ygc的时间变长,为了证明这个结论,我特地修改了一下代码,在上面的代码上加了一行:

if (!_process_strong_tasks->is_task_claimed(SH_PS_SystemDictionary_oops_do)) {   GCTraceTime t("SystemDictionary_OOPS_DO",PrintGCDetails,true,NULL);     if (so & SO_AllClasses) {       SystemDictionary::oops_do(roots);     } else if (so & SO_SystemClasses) {       SystemDictionary::always_strong_oops_do(roots);     }   }

然后重新编译,跑我们的demo,测试结果如下:

2016-03-14T23:57:24.293+0800: [GC2016-03-14T23:57:24.294+0800: [ParNew2016-03-14T23:57:24.296+0800: [SystemDictionary_OOPS_DO, 0.0578430 secs] : 81920K->3184K(92160K), 0.0889740 secs] 81920K->3184K(514048K), 0.0900970 secs] [Times: user=0.27 sys=0.00, real=0.09 secs] 2016-03-14T23:57:28.467+0800: [GC2016-03-14T23:57:28.468+0800: [ParNew2016-03-14T23:57:28.468+0800: [SystemDictionary_OOPS_DO, 0.0779210 secs] : 85104K->5175K(92160K), 0.1071520 secs] 85104K->5175K(514048K), 0.1080490 secs] [Times: user=0.65 sys=0.00, real=0.11 secs] 2016-03-14T23:57:32.984+0800: [GC2016-03-14T23:57:32.984+0800: [ParNew2016-03-14T23:57:32.984+0800: [SystemDictionary_OOPS_DO, 0.1075680 secs] : 87095K->8188K(92160K), 0.1434270 secs] 87095K->8188K(514048K), 0.1439870 secs] [Times: user=0.90 sys=0.01, real=0.14 secs] 2016-03-14T23:57:37.900+0800: [GC2016-03-14T23:57:37.900+0800: [ParNew2016-03-14T23:57:37.901+0800: [SystemDictionary_OOPS_DO, 0.1745390 secs] : 90108K->7093K(92160K), 0.2876260 secs] 90108K->9992K(514048K), 0.2884150 secs] [Times: user=1.44 sys=0.02, real=0.29 secs]

我们会发现YGC的时间变长的时候,SystemDictionary_OOPS_DO的时间也会相应变长多少,因此验证了我们的说法。

原文  http://lovestblog.cn/blog/2016/03/15/ygc-classloader/
正文到此结束
Loading...