欢迎您访问程序员文章站本站旨在为大家提供分享程序员计算机编程知识!
您现在的位置是: 首页  >  IT编程

folly无锁队列正确性说明

程序员文章站 2022-09-07 16:39:27
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::mem ......

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。