JVM volatile语义：
mark：
(1)Java 内存模型不会对valatile指令的操作进行重排序：这个保证对volatile变量的操作时按照指令的出现顺序执行的。
(2)volatile变量不会被缓存在寄存器中(只有拥有线程可见)或者其他对CPU不可见的地方，每次总是从主存中读取volatile变量的结果。也就是说对于volatile变量的修改，其它线程总是可见的，并且不是使用自己线程栈内部的变量。也就是在happens-before法则中，对一个valatile变量的写操作后，其后的任何读操作理解可见此写操作的结果。
Volatile变量的读写
读：置本地数据无效，总是从主存中读取数据
写：更新本地数据，然后刷新主存中的数据
CAS操作：利用CPU中的特殊指令cmpxchg (intel)
CAS有3个操作数，内存值V，旧的预期值A，要修改的新值B。当且仅当预期值A和内存值V相同时，将内存值V修改为B，否则什么都不做。
原子变量的实现：
首先变量是volatile的保证可见性；
对变量的更新在一个循环中实现，例如：
for (;;) {
int current = get();
int next = current + 1;
if (compareAndSet(current, next))
return next;
}
最少执行一次方法就可以返回，否则一直执行；
CAS的ABA问题
比如说一个线程one从内存位置V中取出A，这时候另一个线程two也从内存中取出A，并且two进行了一些操作变成了B，然后two又将V位置的数据变成A，这时候线程one进行CAS操作发现内存中仍然是A，然后one操作成功。尽管线程one的CAS操作成功，但是不代表这个过程就是没有问题的。如果链表的头在变化了两次后恢复了原值，但是不代表链表就没有变化。因此前面提到的原子操作AtomicStampedReference/AtomicMarkableReference就很有
用了。这允许一对变化的元素进行原子操作。
Java通过Unsafe类通过JNI结合CPU的CAS指令来完成
sun.misc.natUnsafe.cc 的源码：
#include <gcj/cni.h>
#include <gcj/field.h>
#include <gcj/javaprims.h>
#include <jvm.h>
#include <sun/misc/Unsafe.h>
#include <java/lang/System.h>
#include <java/lang/InterruptedException.h>
#include <java/lang/Thread.h>
#include <java/lang/Long.h>
#include "sysdep/locks.h"
// Use a spinlock for multi-word accesses
class spinlock
{
static volatile obj_addr_t lock;
public:
spinlock ()
{
while (! compare_and_swap (&lock, 0, 1))
_Jv_ThreadYield ();
}
~spinlock ()
{
release_set (&lock, 0);
}
};
// This is a single lock that is used for all synchronized accesses if
// the compiler can't generate inline compare-and-swap operations. In
// most cases it'll never be used, but the i386 needs it for 64-bit
// locked accesses and so does PPC32. It's worth building libgcj with
// target=i486 (or above) to get the inlines.
volatile obj_addr_t spinlock::lock;
static inline bool
compareAndSwap (volatile jint *addr, jint old, jint new_val)
{
jboolean result = false;
spinlock lock;
if ((result = (*addr == old)))
*addr = new_val;
return result;
}
static inline bool
compareAndSwap (volatile jlong *addr, jlong old, jlong new_val)
{
jboolean result = false;
spinlock lock;
if ((result = (*addr == old)))
*addr = new_val;
return result;
}
static inline bool
compareAndSwap (volatile jobject *addr, jobject old, jobject new_val)
{
jboolean result = false;
spinlock lock;
if ((result = (*addr == old)))
*addr = new_val;
return result;
}
jlong
sun::misc::Unsafe::objectFieldOffset (::java::lang::reflect::Field *field)
{
_Jv_Field *fld = _Jv_FromReflectedField (field);
// FIXME: what if it is not an instance field?
return fld->getOffset();
}
jint
sun::misc::Unsafe::arrayBaseOffset (jclass arrayClass)
{
// FIXME: assert that arrayClass is array.
jclass eltClass = arrayClass->getComponentType();
return (jint)(jlong) _Jv_GetArrayElementFromElementType (NULL, eltClass);
}
jint
sun::misc::Unsafe::arrayIndexScale (jclass arrayClass)
{
// FIXME: assert that arrayClass is array.
jclass eltClass = arrayClass->getComponentType();
if (eltClass->isPrimitive())
return eltClass->size();
return sizeof (void *);
}
// These methods are used when the compiler fails to generate inline
// versions of the compare-and-swap primitives.
jboolean
sun::misc::Unsafe::compareAndSwapInt (jobject obj, jlong offset,
jint expect, jint update)
{
jint *addr = (jint *)((char *)obj + offset);
return compareAndSwap (addr, expect, update);
}
jboolean
sun::misc::Unsafe::compareAndSwapLong (jobject obj, jlong offset,
jlong expect, jlong update)
{
volatile jlong *addr = (jlong*)((char *) obj + offset);
return compareAndSwap (addr, expect, update);
}
jboolean
sun::misc::Unsafe::compareAndSwapObject (jobject obj, jlong offset,
jobject expect, jobject update)
{
jobject *addr = (jobject*)((char *) obj + offset);
return compareAndSwap (addr, expect, update);
}
void
sun::misc::Unsafe::putOrderedInt (jobject obj, jlong offset, jint value)
{
volatile jint *addr = (jint *) ((char *) obj + offset);
*addr = value;
}
void
sun::misc::Unsafe::putOrderedLong (jobject obj, jlong offset, jlong value)
{
volatile jlong *addr = (jlong *) ((char *) obj + offset);
spinlock lock;
*addr = value;
}
void
sun::misc::Unsafe::putOrderedObject (jobject obj, jlong offset, jobject value)
{
volatile jobject *addr = (jobject *) ((char *) obj + offset);
*addr = value;
}
void
sun::misc::Unsafe::putIntVolatile (jobject obj, jlong offset, jint value)
{
write_barrier ();
volatile jint *addr = (jint *) ((char *) obj + offset);
*addr = value;
}
void
sun::misc::Unsafe::putLongVolatile (jobject obj, jlong offset, jlong value)
{
volatile jlong *addr = (jlong *) ((char *) obj + offset);
spinlock lock;
*addr = value;
}
void
sun::misc::Unsafe::putObjectVolatile (jobject obj, jlong offset, jobject value)
{
write_barrier ();
volatile jobject *addr = (jobject *) ((char *) obj + offset);
*addr = value;
}
#if 0 // FIXME
void
sun::misc::Unsafe::putInt (jobject obj, jlong offset, jint value)
{
jint *addr = (jint *) ((char *) obj + offset);
*addr = value;
}
#endif
void
sun::misc::Unsafe::putLong (jobject obj, jlong offset, jlong value)
{
jlong *addr = (jlong *) ((char *) obj + offset);
spinlock lock;
*addr = value;
}
void
sun::misc::Unsafe::putObject (jobject obj, jlong offset, jobject value)
{
jobject *addr = (jobject *) ((char *) obj + offset);
*addr = value;
}
jint
sun::misc::Unsafe::getIntVolatile (jobject obj, jlong offset)
{
volatile jint *addr = (jint *) ((char *) obj + offset);
jint result = *addr;
read_barrier ();
return result;
}
jobject
sun::misc::Unsafe::getObjectVolatile (jobject obj, jlong offset)
{
volatile jobject *addr = (jobject *) ((char *) obj + offset);
jobject result = *addr;
read_barrier ();
return result;
}
jlong
sun::misc::Unsafe::getLong (jobject obj, jlong offset)
{
jlong *addr = (jlong *) ((char *) obj + offset);
spinlock lock;
return *addr;
}
jlong
sun::misc::Unsafe::getLongVolatile (jobject obj, jlong offset)
{
volatile jlong *addr = (jlong *) ((char *) obj + offset);
spinlock lock;
return *addr;
}
void
sun::misc::Unsafe::unpark (::java::lang::Thread *thread)
{
natThread *nt = (natThread *) thread->data;
nt->park_helper.unpark ();
}
void
sun::misc::Unsafe::park (jboolean isAbsolute, jlong time)
{
using namespace ::java::lang;
Thread *thread = Thread::currentThread();
natThread *nt = (natThread *) thread->data;
nt->park_helper.park (isAbsolute, time);
}