folly无锁队列正确性说明
folly无锁队列是facebook开源的一个无所队列，使用的是单向链表，通过compare_exchange语句实现的多生产多消费的队列，我曾经花了比较多的时间学习memory_order的说明，对release-acquire语义，自认为还是比较了解。如果一个atomic对象使用std::memory_order_release进行写操作，而另外一个线程使用std::memory_order_acquire进行读操作，那么这两个线程之间形成同步关系。std::memory_order_release之前写的效果，在std::memory_order_acquire之后可见。不过对于多生产多消费模型，存在多个生产者的情况，在有多个生产者的情况下，结果正确吗？
这里给出folly的源代码，这里请重点关注insertHead函数和sweepOnce函数。
/*
* Copyright 2014-present Facebook, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma once
#include <atomic>
#include <cassert>
#include <utility>
namespace folly {
/**
* A very simple atomic single-linked list primitive.
*
* Usage:
*
* class MyClass {
* AtomicIntrusiveLinkedListHook<MyClass> hook_;
* }
*
* AtomicIntrusiveLinkedList<MyClass, &MyClass::hook_> list;
* list.insert(&a);
* list.sweep([] (MyClass* c) { doSomething(c); }
*/
template <class T>
struct AtomicIntrusiveLinkedListHook {
T* next{ nullptr };
};
template <class T, AtomicIntrusiveLinkedListHook<T> T::*HookMember>
class AtomicIntrusiveLinkedList {
public:
AtomicIntrusiveLinkedList() {}
AtomicIntrusiveLinkedList(const AtomicIntrusiveLinkedList&) = delete;
AtomicIntrusiveLinkedList& operator=(const AtomicIntrusiveLinkedList&) =
delete;
AtomicIntrusiveLinkedList(AtomicIntrusiveLinkedList&& other) noexcept {
auto tmp = other.head_.load();
other.head_ = head_.load();
head_ = tmp;
}
AtomicIntrusiveLinkedList& operator=(
AtomicIntrusiveLinkedList&& other) noexcept {
auto tmp = other.head_.load();
other.head_ = head_.load();
head_ = tmp;
return *this;
}
/**
* Note: list must be empty on destruction.
*/
~AtomicIntrusiveLinkedList() {
assert(empty());
}
bool empty() const {
return head_.load() == nullptr;
}
/**
* Atomically insert t at the head of the list.
* @return True if the inserted element is the only one in the list
* after the call.
*/
bool insertHead(T* t) {
assert(next(t) == nullptr);
auto oldHead = head_.load(std::memory_order_relaxed);
do {
next(t) = oldHead;
/* oldHead is updated by the call below.
NOTE: we don't use next(t) instead of oldHead directly due to
compiler bugs (GCC prior to 4.8.3 (bug 60272), clang (bug 18899),
MSVC (bug 819819); source:
http://en.cppreference.com/w/cpp/atomic/atomic/compare_exchange */
} while (!head_.compare_exchange_weak(oldHead, t,
std::memory_order_release,
std::memory_order_relaxed));
return oldHead == nullptr;
}
/**
* Replaces the head with nullptr,
* and calls func() on the removed elements in the order from tail to head.
* Returns false if the list was empty.
*/
template <typename F>
bool sweepOnce(F&& func) {
if (auto head = head_.exchange(nullptr)) {
auto rhead = reverse(head);
unlinkAll(rhead, std::forward<F>(func));
return true;
}
return false;
}/**
* Repeatedly replaces the head with nullptr,
* and calls func() on the removed elements in the order from tail to head.
* Stops when the list is empty.
*/
template <typename F>
void sweep(F&& func) {
while (sweepOnce(func)) {
}
}
/**
* Similar to sweep() but calls func() on elements in LIFO order.
*
* func() is called for all elements in the list at the moment
* reverseSweep() is called. Unlike sweep() it does not loop to ensure the
* list is empty at some point after the last invocation. This way callers
* can reason about the ordering: elements inserted since the last call to
* reverseSweep() will be provided in LIFO order.
*
* Example: if elements are inserted in the order 1-2-3, the callback is
* invoked 3-2-1. If the callback moves elements onto a stack, popping off
* the stack will produce the original insertion order 1-2-3.
*/
template <typename F>
void reverseSweep(F&& func) {
// We don't loop like sweep() does because the overall order of callbacks
// would be strand-wise LIFO which is meaningless to callers.
auto head = head_.exchange(nullptr);
unlinkAll(head, std::forward<F>(func));
}
private:
std::atomic<T*> head_{ nullptr };
static T*& next(T* t) {
return (t->*HookMember).next;
}
/* Reverses a linked list, returning the pointer to the new head
(old tail) */
static T* reverse(T* head) {
T* rhead = nullptr;
while (head != nullptr) {
auto t = head;
head = next(t);
next(t) = rhead;
rhead = t;
}
return rhead;
}
/* Unlinks all elements in the linked list fragment pointed to by `head',
* calling func() on every element */
template <typename F>
void unlinkAll(T* head, F&& func) {
while (head != nullptr) {
auto t = head;
head = next(t);
next(t) = nullptr;
func(t);
}
}
};
} // namespace folly
如果存在两个线程先后向同一个队列中插入节点，由于两个线程中没有一个使用acquire，如果仅按照release-acquire语义，显然，正确性无法保证，后一个insertHead函数中，无论是auto oldHead = head_.load(std::memory_order_relaxed);，还是while (!head_.compare_exchange_weak(oldHead, t, std::memory_order_release,std::memory_order_relaxed));都可能读取的是前一个线程插入前的数据。那么，还有什么C++语义，可以保证folly队列的正确性？那就是release sequence。release sequence其中的一部分说的是：
如果一个存储使用memory_order_release或更严格的内存序，后面跟着若干读-改-写（read-modify-write）（可以是同一个线程，也可以是不同的线程）操作的话。
（1）那么中间的读-改-写操作 读取的要么是前一次读-改-写的结果，要么是存储的数据。
那么，如果存在一个release操作，后面跟着一个读改写操作的话，这个读改写操作肯定会得到之前release操作写入的效果。我们可以观察到insertHead中的compare_exchange_weak为一个release操作，同时也是一个读改写操作，那么前面一个线程的修改，一定会在后面一个compare_exchange_weak中可见，无论是同一个线程调用，还是不同线程调用。注意到auto oldHead = head_.load(std::memory_order_relaxed);得到的结果的正确性与否，不影响compare_exchange_weak的正确性，因为如果前一个读取的结果是旧值，这个操作就会失败，而且将oldHead的值更新为最新值，这点对于理解folly的正确性很重要。其他的情况应该根据类似的原理得到正确的解答，这里就不详细说明了。