LockSupport类是Java6(JSR166-JUC)引入的一个类,提供了基本的线程同步原语。LockSupport实际上是调用了Unsafe类里的函数,归结到Unsafe里,只有两个函数:
public native void unpark(Thread jthread);
public native void park(boolean isAbsolute, long time);
isAbsolute参数是指明时间是绝对的,还是相对的。
仅仅两个简单的接口,就为上层提供了强大的同步原语。
先来解析下两个函数是做什么的。
unpark函数为线程提供“许可(permit)”,线程调用park函数则等待“许可”。这个有点像信号量,但是这个“许可”是不能叠加的,“许可”是一次性的。
比如线程B连续调用了三次unpark函数,当线程A调用park函数就使用掉这个“许可”,如果线程A再次调用park,则进入等待状态。
注意,unpark函数可以先于park调用。比如线程B调用unpark函数,给线程A发了一个“许可”,那么当线程A调用park时,它发现已经有“许可”了,那么它会马上再继续运行。
实际上,park函数即使没有“许可”,有时也会无理由地返回,这点等下再解析。
park和unpark的灵活之处
上面已经提到,unpark函数可以先于park调用,这个正是它们的灵活之处。
一个线程它有可能在别的线程unPark之前,或者之后,或者同时调用了park,那么因为park的特性,它可以不用担心自己的park的时序问题,否则,如果park必须要在unpark之前,那么给编程带来很大的麻烦!!
考虑一下,两个线程同步,要如何处理?
在Java5里是用wait/notify/notifyAll来同步的。wait/notify机制有个很蛋疼的地方是,比如线程B要用notify通知线程A,那么线程B要确保线程A已经在wait调用上等待了,否则线程A可能永远都在等待。编程的时候就会很蛋疼。
另外,是调用notify,还是notifyAll?
notify只会唤醒一个线程,如果错误地有两个线程在同一个对象上wait等待,那么又悲剧了。为了安全起见,貌似只能调用notifyAll了。
park/unpark模型真正解耦了线程之间的同步,线程之间不再需要一个Object或者其它变量来存储状态,不再需要关心对方的状态。
HotSpot里park/unpark的实现
每个java线程都有一个Parker实例,Parker类是这样定义的:
class Parker : public os::PlatformParker {可以看到Parker类实际上用Posix的mutex,condition来实现的。
private:
volatile int _counter ;
...
public:
void park(bool isAbsolute, jlong time);
void unpark();
...
}
class PlatformParker : public CHeapObj<mtInternal> {
protected:
pthread_mutex_t _mutex [1] ;
pthread_cond_t _cond [1] ;
...
}
在Parker类里的_counter字段,就是用来记录所谓的“许可”的。
当调用park时,先尝试直接能否直接拿到“许可”,即_counter>0时,如果成功,则把_counter设置为0,并返回:
void Parker::park(bool isAbsolute, jlong time) {
// Ideally we'd do something useful while spinning, such
// as calling unpackTime().
// Optional fast-path check:
// Return immediately if a permit is available.
// We depend on Atomic::xchg() having full barrier semantics
// since we are doing a lock-free update to _counter.
if (Atomic::xchg(0, &_counter) > 0) return;
如果不成功,则构造一个ThreadBlockInVM,然后检查_counter是不是>0,如果是,则把_counter设置为0,unlock mutex并返回:
ThreadBlockInVM tbivm(jt);
if (_counter > 0) { // no wait needed
_counter = 0;
status = pthread_mutex_unlock(_mutex);
否则,再判断等待的时间,然后再调用pthread_cond_wait函数等待,如果等待返回,则把_counter设置为0,unlock mutex并返回:
if (time == 0) {当unpark时,则简单多了,直接设置_counter为1,再unlock mutext返回。如果_counter之前的值是0,则还要调用pthread_cond_signal唤醒在park中等待的线程:
status = pthread_cond_wait (_cond, _mutex) ;
}
_counter = 0 ;
status = pthread_mutex_unlock(_mutex) ;
assert_status(status == 0, status, "invariant") ;
OrderAccess::fence();
void Parker::unpark() {简而言之,是用mutex和condition保护了一个_counter的变量,当park时,这个变量置为了0,当unpark时,这个变量置为1。
int s, status ;
status = pthread_mutex_lock(_mutex);
assert (status == 0, "invariant") ;
s = _counter;
_counter = 1;
if (s < 1) {
if (WorkAroundNPTLTimedWaitHang) {
status = pthread_cond_signal (_cond) ;
assert (status == 0, "invariant") ;
status = pthread_mutex_unlock(_mutex);
assert (status == 0, "invariant") ;
} else {
status = pthread_mutex_unlock(_mutex);
assert (status == 0, "invariant") ;
status = pthread_cond_signal (_cond) ;
assert (status == 0, "invariant") ;
}
} else {
pthread_mutex_unlock(_mutex);
assert (status == 0, "invariant") ;
}
}
值得注意的是在park函数里,调用pthread_cond_wait时,并没有用while来判断,所以posix condition里的"Spurious wakeup"一样会传递到上层Java的代码里。
关于"Spurious wakeup",参考上一篇blog:http://blog.csdn.net/hengyunabc/article/details/27969613
if (time == 0) {
status = pthread_cond_wait (_cond, _mutex) ;
}
这也就是为什么Java dos里提到,当下面三种情况下park函数会返回:
- Some other thread invokes unpark with the current thread as the target; or
- Some other thread interrupts the current thread; or
- The call spuriously (that is, for no reason) returns.
相关的实现代码在:
http://hg.openjdk.java.net/jdk7/jdk7/hotspot/file/81d815b05abb/src/share/vm/runtime/park.hpp
http://hg.openjdk.java.net/jdk7/jdk7/hotspot/file/81d815b05abb/src/share/vm/runtime/park.cpp
http://hg.openjdk.java.net/jdk7/jdk7/hotspot/file/81d815b05abb/src/os/linux/vm/os_linux.hpp
http://hg.openjdk.java.net/jdk7/jdk7/hotspot/file/81d815b05abb/src/os/linux/vm/os_linux.cpp
其它的一些东东:
Parker类在分配内存时,使用了一个技巧,重载了new函数来实现了cache line对齐。
// We use placement-new to force ParkEvent instances to beParker里使用了一个无锁的队列在分配释放Parker实例:
// aligned on 256-byte address boundaries. This ensures that the least
// significant byte of a ParkEvent address is always 0.
void * operator new (size_t sz) ;
volatile int Parker::ListLock = 0 ;
Parker * volatile Parker::FreeList = NULL ;
Parker * Parker::Allocate (JavaThread * t) {
guarantee (t != NULL, "invariant") ;
Parker * p ;
// Start by trying to recycle an existing but unassociated
// Parker from the global free list.
for (;;) {
p = FreeList ;
if (p == NULL) break ;
// 1: Detach
// Tantamount to p = Swap (&FreeList, NULL)
if (Atomic::cmpxchg_ptr (NULL, &FreeList, p) != p) {
continue ;
}
// We've detached the list. The list in-hand is now
// local to this thread. This thread can operate on the
// list without risk of interference from other threads.
// 2: Extract -- pop the 1st element from the list.
Parker * List = p->FreeNext ;
if (List == NULL) break ;
for (;;) {
// 3: Try to reattach the residual list
guarantee (List != NULL, "invariant") ;
Parker * Arv = (Parker *) Atomic::cmpxchg_ptr (List, &FreeList, NULL) ;
if (Arv == NULL) break ;
// New nodes arrived. Try to detach the recent arrivals.
if (Atomic::cmpxchg_ptr (NULL, &FreeList, Arv) != Arv) {
continue ;
}
guarantee (Arv != NULL, "invariant") ;
// 4: Merge Arv into List
Parker * Tail = List ;
while (Tail->FreeNext != NULL) Tail = Tail->FreeNext ;
Tail->FreeNext = Arv ;
}
break ;
}
if (p != NULL) {
guarantee (p->AssociatedWith == NULL, "invariant") ;
} else {
// Do this the hard way -- materialize a new Parker..
// In rare cases an allocating thread might detach
// a long list -- installing null into FreeList --and
// then stall. Another thread calling Allocate() would see
// FreeList == null and then invoke the ctor. In this case we
// end up with more Parkers in circulation than we need, but
// the race is rare and the outcome is benign.
// Ideally, the # of extant Parkers is equal to the
// maximum # of threads that existed at any one time.
// Because of the race mentioned above, segments of the
// freelist can be transiently inaccessible. At worst
// we may end up with the # of Parkers in circulation
// slightly above the ideal.
p = new Parker() ;
}
p->AssociatedWith = t ; // Associate p with t
p->FreeNext = NULL ;
return p ;
}
void Parker::Release (Parker * p) {
if (p == NULL) return ;
guarantee (p->AssociatedWith != NULL, "invariant") ;
guarantee (p->FreeNext == NULL , "invariant") ;
p->AssociatedWith = NULL ;
for (;;) {
// Push p onto FreeList
Parker * List = FreeList ;
p->FreeNext = List ;
if (Atomic::cmpxchg_ptr (p, &FreeList, List) == List) break ;
}
}
总结与扯谈
JUC(Java Util Concurrency)仅用简单的park, unpark和CAS指令就实现了各种高级同步数据结构,而且效率很高,令人惊叹。
在C++程序员各种自制*的时候,Java程序员则有很丰富的并发数据结构,如lock,latch,queue,map等信手拈来。
要知道像C++直到C++11才有标准的线程库,同步原语,但离高级的并发数据结构还有很远。boost库有提供一些线程,同步相关的类,但也是很简单的。Intel的tbb有一些高级的并发数据结构,但是国内boost都用得少,更别说tbb了。
最开始研究无锁算法的是C/C++程序员,但是后来很多Java程序员,或者类库开始自制各种高级的并发数据结构,经常可以看到有分析Java并发包的文章。反而C/C++程序员总是在分析无锁的队列算法。高级的并发数据结构,比如并发的HashMap,没有看到有相关的实现或者分析的文章。在C++11之后,这种情况才有好转。
因为正确高效实现一个Concurrent Hash Map是很困难的,要对内存CPU有深刻的认识,而且还要面对CPU不断升级带来的各种坑。
我认为真正值得信赖的C++并发库,只有Intel的tbb和微软的PPL。
https://software.intel.com/en-us/node/506042 Intel® Threading Building Blocks
http://msdn.microsoft.com/en-us/library/dd492418.aspx Parallel Patterns Library (PPL)
另外FaceBook也开源了一个C++的类库,里面也有并发数据结构。