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。