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。
原文:https://www.cnblogs.com/albizzia/p/9000364.html