Jdk1.6 JUC源码解析(1)-atomic-AtomicXXX

Jdk1.6 JUC源码解析(1)-atomic-AtomicXXX

作者:大飞

 

功能简介:

  • 原子量和普通变量相比,主要体现在读写的线程安全上。对原子量的是原子的(比如多线程下的共享变量i++就不是原子的),由CAS操作保证原子性。对原子量的读可以读到最新值,由volatile关键字来保证可见性
  • 原子量多用于数据统计(如接口调用次数)、一些序列生成(多线程环境下)以及一些同步数据结构中。

源码分析:

  • 首先,原子量的一些较底层的操作都是来自sun.misc.Unsafe类,所以原子量内部有一个Unsafe的静态引用。
    private static final Unsafe unsafe = Unsafe.getUnsafe();

         在openJdk代码中可以找到这个类,目录openJdk的jdk/share/classes/sun/misc/。        这个类里面大多数方法都是native的,方法实现可以在openJdk的hotspot/share/vm/prims/unsafe.cpp里面找到。        

  • 接下来,先看下AtomicInteger的源码。

       在AtomicInteger源码中,由内部的一个int域来保存值:

    private volatile int value;

       注意到这个int域由volatile关键字修饰,可以保证可见性。        
细节:volatile怎么保证可见性呢?对于被volatile修饰的域来说,对域进行的写入操作,在指令层面会在必要的时候(多核CPU)加入内存屏障(如:lock addl $0x0),这个内存屏障的作用是令本次写操作刷回主存,同时使其他CPU的cacheline中相应数据失效。所以当其他CPU需要访问相应数据的时候,会到主存中访问,从而保证了多线程环境下相应域的可见性。          接下来看一下CAS操作,AtomicInteger中的CAS操作体现在方法compareAndSet。它的实现在unsafe.cpp里面:

/*
 *      Implementation of class sun.misc.Unsafe
 */
...
UNSAFE_ENTRY(jboolean, Unsafe_CompareAndSwapInt(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jint e, jint x))
  UnsafeWrapper("Unsafe_CompareAndSwapInt");
  oop p = JNIHandles::resolve(obj);
  jint* addr = (jint *) index_oop_from_field_offset_long(p, offset);
  return (jint)(Atomic::cmpxchg(x, addr, e)) == e;
UNSAFE_END

        这里调用了Atomic的cmpxchg方法,继续找一下。这个方法定义在hotspot/share/vm/runtime/atomic.hpp中,实现在hotspot/share/vm/runtime/atomic.cpp中,最终实现取决于底层OS,比如linux x86,实现内联在hotspot部分代码os_cpu/linux_x86/vm/atomic_linux_x86.inline.hpp:

// Adding a lock prefix to an instruction on MP machine
#define LOCK_IF_MP(mp) "cmp $0, " #mp "; je 1f; lock; 1: "
...
inline jint     Atomic::cmpxchg    (jint     exchange_value, volatile jint*     dest, jint     compare_value) {
  int mp = os::is_MP();
  __asm__ volatile (LOCK_IF_MP(%4) "cmpxchgl %1,(%3)"
                    : "=a" (exchange_value)
                    : "r" (exchange_value), "a" (compare_value), "r" (dest), "r" (mp)
                    : "cc", "memory");
  return exchange_value;
}

       从上面的代码中可以看到,如果是CPU是多核(multi processors)的话,会添加一个lock;前缀,这个lock;前缀也是内存屏障,它的作用是在执行后面指令的过程中锁总线(或者是锁cacheline),保证一致性。后面的指令cmpxchgl就是x86的比较并交换指令了。          接下来有一个和compareAndSet类似的方法,weakCompareAndSet。        从注释看,这个方法会发生fail spuriously(伪失败),而且不保证(指令)顺序,只能在一些特性场景(一些计数和统计)下替换compareAndSet。但从方法实现上看和compareAndSet没什么区别:

    /**
     * Atomically sets the value to the given updated value
     * if the current value {@code ==} the expected value.
     *
     * <p>May <a href="package-summary.html#Spurious">fail spuriously</a>
     * and does not provide ordering guarantees, so is only rarely an
     * appropriate alternative to {@code compareAndSet}.
     *
     * @param expect the expected value
     * @param update the new value
     * @return true if successful.
     */
    public final boolean weakCompareAndSet(int expect, int update) {
        return unsafe.compareAndSwapInt(this, valueOffset, expect, update);
    }

       但还是应该按照API的说明来使用这两个方法,以防未来方法内部(实现或者底层内部机制)发生变化。          其余的大多数方法都是基于compareAndSet方法来实现的,来看其中一个,incrementAndGet方法:

    /**
     * Atomically increments by one the current value.
     *
     * @return the updated value
     */
    public final int incrementAndGet() {
        for (;;) {
            int current = get();
            int next = current + 1;
            if (compareAndSet(current, next))
                return next;
        }
    }

         这个方法也体现了CAS一般是使用风格-CAS Loop,不断重试,直到成功。        当然,这也不是绝对的,JVM底层完全可以用更好的方式也替换这些方法,比如使用内联的lock;xadd就会比cmpxchg指令更好一些。          最后,AtomicInteger还有一个方法,lazySet:

    /**
     * Eventually sets to the given value.
     *
     * @param newValue the new value
     * @since 1.6
     */
    public final void lazySet(int newValue) {
        unsafe.putOrderedInt(this, valueOffset, newValue);
    }

 
       看下unsafe.cpp中putOrderedInt方法的实现:

    {CC"putOrderedInt",      CC"("OBJ"JI)V",             FN_PTR(Unsafe_SetOrderedInt)},
    ...
// The non-intrinsified versions of setOrdered just use setVolatile
UNSAFE_ENTRY(void, Unsafe_SetOrderedInt(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jint x))
  UnsafeWrapper("Unsafe_SetOrderedInt");
  SET_FIELD_VOLATILE(obj, offset, jint, x);
UNSAFE_END

       注:上面有句注释,说明这只是一个非内联的setOrdered方法的实现,使用了setVolatile(和setVolatile一样的效果)。        其中,SET_FIELD_VOLATILE的定义如下:

#define SET_FIELD_VOLATILE(obj, offset, type_name, x) \
  oop p = JNIHandles::resolve(obj); \
  OrderAccess::release_store_fence((volatile type_name*)index_oop_from_field_offset_long(p, offset), x);

       在hotspot/src/os_cpu/linux_x86/vm/orderAccess_linux_x86.inline.hpp找到了这个方法的内联实现。(hotspot/src/share/vm/runtime/orderAccess.hpp这个头文件里面的注释值得留意一下):

inline void     OrderAccess::release_store_fence(volatile jint*   p, jint   v) {
  __asm__ volatile (  "xchgl (%2),%0"
                    : "=r" (v)
                    : "0" (v), "r" (p)
                    : "memory");
}

       可见,这里是通过xchgl这个指令来实现的setOrdered。          但是,上面只是非内联的实现,我们看下内联的实现是什么样的。        在hotspot/src/share/vm/classfile/vmSymbols.hpp中有如下代码:

  do_intrinsic(_putOrderedInt,            sun_misc_Unsafe,        putOrderedInt_name, putOrderedInt_signature,   F_RN)

        然后找到hotspot/src/share/vm/opto/library_call.cpp中找到相应实现:

  case vmIntrinsics::_putOrderedInt:
    return inline_unsafe_ordered_store(T_INT);
  ...
bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
  // This is another variant of inline_unsafe_access, differing in
  // that it always issues store-store ("release") barrier and ensures
  // store-atomicity (which only matters for "long").
  if (callee()->is_static())  return false;  // caller must have the capability!
  ...//省略不重要的部分
  insert_mem_bar(Op_MemBarRelease);
  insert_mem_bar(Op_MemBarCPUOrder);
  // Ensure that the store is atomic for longs:
  bool require_atomic_access = true;
  Node* store;
  if (type == T_OBJECT) // reference stores need a store barrier.
    store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
  else {
    store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
  }
  insert_mem_bar(Op_MemBarCPUOrder);
  return true;
}

        我们看到这个方法里在保存动作的前后,有3个地方插入了内存屏障。        再看下hotspot/src/cpu/x86/vm/x86_64.ad:

instruct membar_release() %{
  match(MemBarRelease);
  ins_cost(400);
  size(0);
  format %{ "MEMBAR-release ! (empty encoding)" %}
  ins_encode( );
  ins_pipe(empty);
%}
...
instruct membar_volatile(eFlagsReg cr) %{
  match(MemBarVolatile);
  effect(KILL cr);
  ins_cost(400);
  format %{ 
    $$template
    if (os::is_MP()) {
      $$emit$$"LOCK ADDL [ESP + #0], 0\t! membar_volatile"
    } else {
      $$emit$$"MEMBAR-volatile ! (empty encoding)"
    }
  %}
  ins_encode %{
    __ membar(Assembler::StoreLoad);
  %}
  ins_pipe(pipe_slow);
%}

       可见,除了membar_volatile中会添加LOCK ADDL这些指令,其他的貌似没什么卵用。 所以上面那3个insert_mem_bar也相当于没有加任何内存屏障。在这种情况下,lazySet就相当于对一个普通域的写操作喽。  

  • 再看下AtomicBoolean的源码。

       AtomicBoolean内部是用一个int域来表示布尔状态,1表示true;0表示false:

    private volatile int value;
    /**
     * Creates a new {@code AtomicBoolean} with the given initial value.
     *
     * @param initialValue the initial value
     */
    public AtomicBoolean(boolean initialValue) {
        value = initialValue ? 1 : 0;
    }

        CAS、lazySet方法也都分别调用unsafe的compareAndSwapInt和putOrderedInt,上面分析过了。    

  • 继续看下AtomicLong的源码。

       AtomicLong内部是用一个long域来保存值:

    /**
     * Records whether the underlying JVM supports lockless
     * compareAndSwap for longs. While the Unsafe.compareAndSwapLong
     * method works in either case, some constructions should be
     * handled at Java level to avoid locking user-visible locks.
     */
    static final boolean VM_SUPPORTS_LONG_CAS = VMSupportsCS8();
    /**
     * Returns whether underlying JVM supports lockless CompareAndSet
     * for longs. Called only once and cached in VM_SUPPORTS_LONG_CAS.
     */
    private static native boolean VMSupportsCS8();
    static {
      try {
        valueOffset = unsafe.objectFieldOffset
            (AtomicLong.class.getDeclaredField("value"));
      } catch (Exception ex) { throw new Error(ex); }
    }
    private volatile long value;

       这里注意到,AtomicLong中还提供了一个包内可见的静态域VM_SUPPORTS_LONG_CAS来表示底层是否支持Long类型(8字节)的lockless CAS操作。          AtomicLong内部整体结构和AtomicInteger类似,主要来看下内部使用的unsafe的方法有什么不同,首先CAS操作使用了unsafe的compareAndSwapLong方法:

    /**
     * Atomically sets the value to the given updated value
     * if the current value {@code ==} the expected value.
     *
     * @param expect the expected value
     * @param update the new value
     * @return true if successful. False return indicates that
     * the actual value was not equal to the expected value.
     */
    public final boolean compareAndSet(long expect, long update) {
        return unsafe.compareAndSwapLong(this, valueOffset, expect, update);
    }

       在unsafe.cpp中找到实现:

    {CC"compareAndSwapLong", CC"("OBJ"J""J""J"")Z",      FN_PTR(Unsafe_CompareAndSwapLong)},
    
    ...
UNSAFE_ENTRY(jboolean, Unsafe_CompareAndSwapLong(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jlong e, jlong x))
  UnsafeWrapper("Unsafe_CompareAndSwapLong");
  Handle p (THREAD, JNIHandles::resolve(obj));
  jlong* addr = (jlong*)(index_oop_from_field_offset_long(p(), offset));
  if (VM_Version::supports_cx8())
    return (jlong)(Atomic::cmpxchg(x, addr, e)) == e;
  else {
    jboolean success = false;
    ObjectLocker ol(p, THREAD);
    if (*addr == e) { *addr = x; success = true; }
    return success;
  }
UNSAFE_END

       从实现中可以看到,如果平台不支持8字节的CAS操作,就会加锁然后进行设置操作;如果支持,就会调用Atomic::cmpxchg方法,方法实现可以参考具体平台内联代码hotspot/os_cpu/linux_x86/vm/atomic_linux_x86.inline.hpp:

// Adding a lock prefix to an instruction on MP machine
#define LOCK_IF_MP(mp) "cmp $0, " #mp "; je 1f; lock; 1: "
inline jlong    Atomic::cmpxchg    (jlong    exchange_value, volatile jlong*    dest, jlong    compare_value) {
  bool mp = os::is_MP();
  __asm__ __volatile__ (LOCK_IF_MP(%4) "cmpxchgq %1,(%3)"
                        : "=a" (exchange_value)
                        : "r" (exchange_value), "a" (compare_value), "r" (dest), "r" (mp)
                        : "cc", "memory");
  return exchange_value;
}

       其实就是(多核情况下带lock前缀的)cmpxchgq指令。          然后看一下lazySet中使用到的unsafe的putOrderedLong方法。

    /**
     * Eventually sets to the given value.
     *
     * @param newValue the new value
     * @since 1.6
     */
    public final void lazySet(long newValue) {
        unsafe.putOrderedLong(this, valueOffset, newValue);
    }

       同样在unsafe.cpp中可以找到该方法的实现:

    {CC"putOrderedLong",     CC"("OBJ"JJ)V",             FN_PTR(Unsafe_SetOrderedLong)},
    ...
UNSAFE_ENTRY(void, Unsafe_SetOrderedLong(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jlong x))
  UnsafeWrapper("Unsafe_SetOrderedLong");
#if defined(SPARC) || defined(X86)
  // Sparc and X86 have atomic jlong (8 bytes) instructions
  SET_FIELD_VOLATILE(obj, offset, jlong, x);
#else
  // Keep old code for platforms which may not have atomic long (8 bytes) instructions
  {
    if (VM_Version::supports_cx8()) {
      SET_FIELD_VOLATILE(obj, offset, jlong, x);
    }
    else {
      Handle p (THREAD, JNIHandles::resolve(obj));
      jlong* addr = (jlong*)(index_oop_from_field_offset_long(p(), offset));
      ObjectLocker ol(p, THREAD);
      *addr = x;
    }
  }
#endif
UNSAFE_END

       从实现上看,如果平台是SPARC或者X86或者平台支持8字节CAS,就相当于执行了一个volatile write;否则,加锁写。        当然还要看一下内联方法,在hotspot/src/share/vm/classfile/vmSymbols.hpp中有如下代码:

  do_intrinsic(_putOrderedLong,           sun_misc_Unsafe,        putOrderedLong_name, putOrderedLong_signature, F_RN)  \

       然后找到hotspot/src/share/vm/opto/library_call.cpp中找到相应实现。

  case vmIntrinsics::_putOrderedLong:
    return inline_unsafe_ordered_store(T_LONG);
  ...
bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
  // This is another variant of inline_unsafe_access, differing in
  // that it always issues store-store ("release") barrier and ensures
  // store-atomicity (which only matters for "long").
  if (callee()->is_static())  return false;  // caller must have the capability!
  ...//忽略不重要部分
  // Ensure that the store is atomic for longs:
  bool require_atomic_access = true;
  Node* store;
  if (type == T_OBJECT) // reference stores need a store barrier.
    store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
  else {
    store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
  }
  insert_mem_bar(Op_MemBarCPUOrder);
  return true;
}

       从代码的注释上可以发现,require_atomic_access设置为true,为了保证long写操作的原子性。继续跟代码,找到hotspot/src/share/vm/opto/graphKit.cpp:

Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
                                int adr_idx,
                                bool require_atomic_access) {
  assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
  const TypePtr* adr_type = NULL;
  debug_only(adr_type = C->get_adr_type(adr_idx));
  Node *mem = memory(adr_idx);
  Node* st;
  if (require_atomic_access && bt == T_LONG) {
    st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
  } else {
    st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
  }
  st = _gvn.transform(st);
  set_memory(st, adr_idx);
  // Back-to-back stores can only remove intermediate store with DU info
  // so push on worklist for optimizer.
  if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
    record_for_igvn(st);
  return st;
}

       可见,这里会针对long做原子写操作(这里的原子操作应该指的是将long的高4字节和低4字节的操作合并成一个原子操作,比如某些平台不支持非volatile的long/double域的原子操作)。    

  • 最后看下AtomicReference的源码。

       AtomicReference内部构成和其他原子量基本一致,区别只是这个类内部保存一个对象引用。

public class AtomicReference<V>  implements java.io.Serializable {
    private static final long serialVersionUID = -1848883965231344442L;
    private static final Unsafe unsafe = Unsafe.getUnsafe();
    private static final long valueOffset;
    static {
      try {
        valueOffset = unsafe.objectFieldOffset
            (AtomicReference.class.getDeclaredField("value"));
      } catch (Exception ex) { throw new Error(ex); }
    }
    private volatile V value;

       重点看一下内部调用的unsafe的compareAndSwapObject和putOrderedObject方法,先看一下compareAndSwapObject:

    /**
     * Atomically sets the value to the given updated value
     * if the current value {@code ==} the expected value.
     * @param expect the expected value
     * @param update the new value
     * @return true if successful. False return indicates that
     * the actual value was not equal to the expected value.
     */
    public final boolean compareAndSet(V expect, V update) {
        return unsafe.compareAndSwapObject(this, valueOffset, expect, update);
    }

       在unsafe.cpp中可以找到实现:

    {CC"compareAndSwapObject", CC"("OBJ"J"OBJ""OBJ")Z",  FN_PTR(Unsafe_CompareAndSwapObject)},
    ...
UNSAFE_ENTRY(jboolean, Unsafe_CompareAndSwapObject(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jobject e_h, jobject x_h))
  UnsafeWrapper("Unsafe_CompareAndSwapObject");
  oop x = JNIHandles::resolve(x_h);
  oop e = JNIHandles::resolve(e_h);
  oop p = JNIHandles::resolve(obj);
  HeapWord* addr = (HeapWord *)index_oop_from_field_offset_long(p, offset);
  if (UseCompressedOops) {
    update_barrier_set_pre((narrowOop*)addr, e);
  } else {
    update_barrier_set_pre((oop*)addr, e);
  }
  oop res = oopDesc::atomic_compare_exchange_oop(x, addr, e);
  jboolean success  = (res == e);
  if (success)
    update_barrier_set((void*)addr, x);
  return success;
UNSAFE_END

       可以看到里面实际上是执行oopDesc::atomic_compare_exchange_oop这个方法。找到hotspot/src/share/vm/oops/oop.inline.hpp中该方法实现:

inline oop oopDesc::atomic_compare_exchange_oop(oop exchange_value,
                                                volatile HeapWord *dest,
                                                oop compare_value) {
  if (UseCompressedOops) {
    // encode exchange and compare value from oop to T
    narrowOop val = encode_heap_oop(exchange_value);
    narrowOop cmp = encode_heap_oop(compare_value);
    narrowOop old = (narrowOop) Atomic::cmpxchg(val, (narrowOop*)dest, cmp);
    // decode old from T to oop
    return decode_heap_oop(old);
  } else {
    return (oop)Atomic::cmpxchg_ptr(exchange_value, (oop*)dest, compare_value);
  }
}

       cmpxchg之前分析过,看看cmpxche_ptr,找到hotspot/os_cpu/linux_x86/vm/atomic_linux_x86.inline.hpp中实现:

inline intptr_t Atomic::cmpxchg_ptr(intptr_t exchange_value, volatile intptr_t* dest, intptr_t compare_value) {
  return (intptr_t)cmpxchg((jlong)exchange_value, (volatile jlong*)dest, (jlong)compare_value);
}
inline jlong    Atomic::cmpxchg    (jlong    exchange_value, volatile jlong*    dest, jlong    compare_value) {
  bool mp = os::is_MP();
  __asm__ __volatile__ (LOCK_IF_MP(%4) "cmpxchgq %1,(%3)"
                        : "=a" (exchange_value)
                        : "r" (exchange_value), "a" (compare_value), "r" (dest), "r" (mp)
                        : "cc", "memory");
  return exchange_value;
}

       就是(多核下带lock前缀的)cmpxchgq命令了。        putOrderedObject方法按之前几篇的查找方法,会发现内联之后,相当于一个普通写操作了。          OK,源码分析到此结束!                 参考资料:        
http://gee.cs.oswego.edu/dl/jmm/cookbook.html        
http://x86.renejeschke.de/html/file_module_x86_id_41.html        
http://stackoverflow.com/questions/2443239/how-can-weakcompareandset-fail-spuriously-if-it-is-implemented-exactly-like-comp        
http://faydoc.tripod.com/cpu/lock.htm        
http://ifeve.com/juc-atomic-class-lazyset-que/        
http://bugs.java.com/bugdatabase/view_bug.do?bug_id=6275329        
http://stackoverflow.com/questions/4232660/which-is-a-better-write-barrier-on-x86-lockaddl-or-xchgl        
http://bugs.java.com/bugdatabase/view_bug.do?bug_id=7023898      

 

 

    原文作者:JUC
    原文地址: https://blog.csdn.net/iteye_11160/article/details/82642189
    本文转自网络文章,转载此文章仅为分享知识,如有侵权,请联系博主进行删除。
点赞