在学习C++11多线程的时候,会碰到一大堆概念,mutex, lock, atomic, memory model, memory barrier, lock-free等。要更好的理解,可以先了解下CPU的Memory Barriers机制(可读Paul McKenny的Memory Barriers: a Hardware View for Software Hackers),然后看Jeff Preshing的blog, 他的帖子深入浅出,写得非常好。再看看Herb Sutter,Hans Boehm等人的文章。 Bartosz Milewski的博客也值得看看.
double-checked locking是一个用来学习的好例子。 Scott Meyers和Andrei Alexandrescu两位大牛写过一篇paper,讨论了double-checked实现的困难(Java的memeory model没有完善之前有同样的问题). Jeff Preshing写篇文章讨论这个问题.
看完大牛们的文章, 一步一步来动手实现一个Singleton,最后用template泛化.
1. 单线程
Singleton的单线程实现很简单:
//header file
class Singleton {
public:
static Singleton* getInstance ();
private:
static Singleton* m_Instance;
};
//implementation. Version 0.1
Singleton* Singleton::m_Instance = nullptr;
Singleton* Singleton::getInstance () {
if (m_Instance == nullptr) {
m_Instance = new Singleton;
}
return Singleton;
}
在单线程环境,这个版本工作的很好。 但在多线程线环境下有data race了, 关键在那个if判断, 多个线程可能同时进入if里面。
2.多线程的尝试实现
C++11已经支持多线程,无需调用库,用std::mutex加个锁, 把if判断放到临界区里保护起来:
//header file
class Singleton {
public:
static Singleton* getInstance ();
private:
static Singleton* m_Instance;
static mutex m_mutex;
};
//implementation. Version 0.2 it is ok, but too expensive
Singleton* Singleton::m_Instance = nullptr;
//oops, high cost.
Singleton* Singleton::getInstance () {
lock_guard<mutex> lock(m_mutex);
if (m_Instance == nullptr) {
m_Instance = new Singleton;
}
return Singleton;
}
这个版本有什么问题呢?成本太高,每个调用都去获取锁,单例创建好之后,其实已经没有必要获取锁了,并发情况下会导致其他线程因等待锁而被系统休眠,成本太高了。 那么,每次调用都加锁, 在获取锁之前再加一个if(m_Instance == nullptr)判断, 是否可行?
//implementation. Version 0.3
Singleton* Singleton::m_Instance = nullptr;
//oops! it does NOT work
Singleton* Singleton::getInstance () {
if (m_Instance == nullptr) {
lock_guard<mutex> lock(m_mutex);
if (m_Instance == nullptr) {
m_Instance = new Singleton;
}
}
return Singleton;
}
想法很好,但是有严重的缺陷,来看看 m_Instance = new Singleton, 这个new操作是先分配一块空间,然后执行构造函数,相当于:
pInstance = operator new(sizeof(Singleton)); // Step 1
new (pInstance) Singleton; // Step 2
如果一个线程执行到step 1时, 另一个线程发现 m_Instance != nullptr, 直接把 m_Instance 返回,而Step 2 还没来得及执行,返回的指针指向一块并没有构造好的空间...
那么,来加一个临时变量,思路是让allocator和constructor都做完之后,再把指针赋给m_Instance,这样可行么?
//implementation. Version 0.4
Singleton* Singleton::getInstance () {
Singleton * tmp = m_Instance;
if (m_Instance == nullptr) {
lock_guard<mutex> lock(m_mutex);
tmp = m_Instance;
if (m_Instance == nullptr) {
tmp = new Singleton; //oops
m_Instance = tmp;
}
}
return Singleton;
}
但是,我们知道,编译器优化和CPU流水线执行都有可能对代码执行顺序进行re-order.(参考Memory Model), 这样:
//re-order之后,不能保证 step 2一定在step 3之前执行完毕。
tmp = operator new(sizeof(Singleton)); // Step 1
new (pInstance) Singleton; // Step 2
m_Instance = tmp; //Step 3
3.C++11 Sequentially Consistent Atomics
要保证step 3在step 2之后执行,可以用Sequential ordering实现(即使用默认的memory_order_seq_cst),编译器会插入memery barrier来保证。
#include <mutex>
#include <atomic>
using namespace std;
class Singleton {
public:
static Singleton* getInstance ();
private:
static std::atomic<Singleton*> m_Instance;
static mutex m_mutex;
};
std::atomic<Singleton*> Singleton::m_Instance;
std::mutex Singleton::m_mutex;
Singleton* Singleton::getInstance () {
Singleton* tmp = m_Instance.load (); //memory_order_seq_cst, Sequential ordering
if (tmp == nullptr) {
std::lock_guard<std::mutex> lock(m_mutex);
//maybe many threads are locked here by the mutex at the same time.
//After the lock is free(the singleton is created by one thread), load again
tmp = m_Instance.load(std::memory_order_relaxed);
if (tmp == nullptr) {
tmp = new Singleton;
m_Instance.store (tmp); //memory_order_seq_cst
}
}
return tmp;
}
那么,我们来看看, atomic的load和store是怎样保证re-order之后语义还是正确的呢?
用gcc生成汇编代码(默认是AT&T风格汇编, 我习惯看intel风格的,加个masm=intel参数):
# intel i5, ubuntu14.04, gcc 4.8.2
g++ -O2 -S -masm=intel -pthread -std=c++11 Singleton.cpp -o asm.s
call _Znwm ; call new
.LEHE0:
mov QWORD PTR _ZN9Singleton10m_instanceE[rip], rax ;return value into rax
test rbp, rbp ;
mov rbx, rax ; rbx is the var tmp
mfence ;memory fence!
je .L11
可以看到编译器在x86平台上为store()生成了mfence指令。 那load()为什么没有memory fence呢,是因为x86/64是"Strong"类型的CPU(细节可参考weak vs strong memory models和Paul McKenny的文章Memory Barriers).
4. Low-Level Ordering Constraints
一般来说,用默认的memory_order_seq_cst已经够用了,代码也简单一些。不过mfence指令的成本较高(几十倍于register to register指令,因为需要在CPU各个core和cache里进行复杂的通讯,同步cache line等等),如果是在高并发情景下,可以考虑进一步优化。可以用low-level的acquire/release operation. 有点晦涩,可以参考acquire and release fences和acquire and release semantics
std::atomic<Singleton*> Singleton::m_Instance;
std::mutex Singleton::m_mutex;
Singleton* Singleton::getInstance () {
Singleton* tmp = m_Instance.load (std::memory_order_acquire);
if (tmp == nullptr) {
std::lock_guard<std::mutex> lock (m_mutex);
tmp = m_Instance.load (std::memory_order_relaxed);
if (tmp == nullptr) {
tmp = new Singleton;
m_Instance.store (tmp, std::memory_order_release);
}
}
return tmp;
}
再生成汇编代码,在x86/64平台上,可以看到,memory_order_release没有生成mfence指令。
那,为什么还要memory_order_relaxed、memory_order_release呢? 因为它们是语言层次上的抽象,可以阻止编译器的指令re-order. 同时保证了不同CPU平台的可移植性,在ARM, PowerPC等平台会生成对应的指令。
5. 用Template泛化
#include <mutex>
#include <atomic>
using namespace std;
template<typename T> class Singleton {
private:
static atomic<T*> m_instance;
static mutex m_mutex;
public:
static T* getInstance () ;
};
template<typename T> atomic<T*> Singleton<T>::m_instance;
template<typename T> mutex Singleton<T>::m_mutex;
template<typename T> T* Singleton<T>::getInstance () {
T* tmp = m_instance.load(std::memory_order_acquire);
//atomic_thread_fence (std::memory_order_acquire);
if (tmp == nullptr) {
lock_guard<mutex> lock(m_mutex);
tmp = m_Instance.load(std::memory_order_relaxed);
if (tmp == nullptr) {
tmp = new T;
//atomic_thread_fence (std::memory_order_release);
m_instance.store (tmp, std::memory_order_release);
}
}
return tmp;
}
class Foo {};
int main()
{
Foo* inst = Singleton<Foo>::getInstance ();
return 0;
}