[/ / Copyright (c) 2009 Helge Bahmann / / Distributed under the Boost Software License, Version 1.0. (See accompanying / file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) /] [section:example_reference_counters Reference counting] The purpose of a ['reference counter] is to count the number of pointers to an object. The object can be destroyed as soon as the reference counter reaches zero. [section Implementation] [c++] #include #include class X { public: typedef boost::intrusive_ptr pointer; X() : refcount_(0) {} private: mutable boost::atomic refcount_; friend void intrusive_ptr_add_ref(const X * x) { x->refcount_.fetch_add(1, boost::memory_order_relaxed); } friend void intrusive_ptr_release(const X * x) { if (x->refcount_.fetch_sub(1, boost::memory_order_release) == 1) { boost::atomic_thread_fence(boost::memory_order_acquire); delete x; } } }; [endsect] [section Usage] [c++] X::pointer x = new X; [endsect] [section Discussion] Increasing the reference counter can always be done with [^memory_order_relaxed]: New references to an object can only be formed from an existing reference, and passing an existing reference from one thread to another must already provide any required synchronization. It is important to enforce any possible access to the object in one thread (through an existing reference) to ['happen before] deleting the object in a different thread. This is achieved by a "release" operation after dropping a reference (any access to the object through this reference must obviously happened before), and an "acquire" operation before deleting the object. It would be possible to use [^memory_order_acq_rel] for the [^fetch_sub] operation, but this results in unneeded "acquire" operations when the reference counter does not yet reach zero and may impose a performance penalty. [endsect] [endsect] [section:example_spinlock Spinlock] The purpose of a ['spin lock] is to prevent multiple threads from concurrently accessing a shared data structure. In contrast to a mutex, threads will busy-wait and waste CPU cycles instead of yielding the CPU to another thread. ['Do not use spinlocks unless you are certain that you understand the consequences.] [section Implementation] [c++] #include class spinlock { private: typedef enum {Locked, Unlocked} LockState; boost::atomic state_; public: spinlock() : state_(Unlocked) {} void lock() { while (state_.exchange(Locked, boost::memory_order_acquire) == Locked) { /* busy-wait */ } } void unlock() { state_.store(Unlocked, boost::memory_order_release); } }; [endsect] [section Usage] [c++] spinlock s; s.lock(); // access data structure here s.unlock(); [endsect] [section Discussion] The purpose of the spinlock is to make sure that one access to the shared data structure always strictly "happens before" another. The usage of acquire/release in lock/unlock is required and sufficient to guarantee this ordering. It would be correct to write the "lock" operation in the following way: [c++] lock() { while (state_.exchange(Locked, boost::memory_order_relaxed) == Locked) { /* busy-wait */ } atomic_thread_fence(boost::memory_order_acquire); } This "optimization" is however a) useless and b) may in fact hurt: a) Since the thread will be busily spinning on a blocked spinlock, it does not matter if it will waste the CPU cycles with just "exchange" operations or with both useless "exchange" and "acquire" operations. b) A tight "exchange" loop without any memory-synchronizing instruction introduced through an "acquire" operation will on some systems monopolize the memory subsystem and degrade the performance of other system components. [endsect] [endsect] [section:singleton Singleton with double-checked locking pattern] The purpose of the ['Singleton with double-checked locking pattern] is to ensure that at most one instance of a particular object is created. If one instance has been created already, access to the existing object should be as light-weight as possible. [section Implementation] [c++] #include #include class X { public: static X * instance() { X * tmp = instance_.load(boost::memory_order_consume); if (!tmp) { boost::mutex::scoped_lock guard(instantiation_mutex); tmp = instance_.load(boost::memory_order_consume); if (!tmp) { tmp = new X; instance_.store(tmp, boost::memory_order_release); } } return tmp; } private: static boost::atomic instance_; static boost::mutex instantiation_mutex; }; boost::atomic X::instance_(0); [endsect] [section Usage] [c++] X * x = X::instance(); // dereference x [endsect] [section Discussion] The mutex makes sure that only one instance of the object is ever created. The [^instance] method must make sure that any dereference of the object strictly "happens after" creating the instance in another thread. The use of [^memory_order_release] after creating and initializing the object and [^memory_order_consume] before dereferencing the object provides this guarantee. It would be permissible to use [^memory_order_acquire] instead of [^memory_order_consume], but this provides a stronger guarantee than is required since only operations depending on the value of the pointer need to be ordered. [endsect] [endsect] [section:example_ringbuffer Wait-free ring buffer] A ['wait-free ring buffer] provides a mechanism for relaying objects from one single "producer" thread to one single "consumer" thread without any locks. The operations on this data structure are "wait-free" which means that each operation finishes within a constant number of steps. This makes this data structure suitable for use in hard real-time systems or for communication with interrupt/signal handlers. [section Implementation] [c++] #include template class ringbuffer { public: ringbuffer() : head_(0), tail_(0) {} bool push(const T & value) { size_t head = head_.load(boost::memory_order_relaxed); size_t next_head = next(head); if (next_head == tail_.load(boost::memory_order_acquire)) return false; ring_[head] = value; head_.store(next_head, boost::memory_order_release); return true; } bool pop(T & value) { size_t tail = tail_.load(boost::memory_order_relaxed); if (tail == head_.load(boost::memory_order_acquire)) return false; value = ring_[tail]; tail_.store(next(tail), boost::memory_order_release); return true; } private: size_t next(size_t current) { return (current + 1) % Size; } T ring_[Size]; boost::atomic head_, tail_; }; [endsect] [section Usage] [c++] ringbuffer r; // try to insert an element if (r.push(42)) { /* succeeded */ } else { /* buffer full */ } // try to retrieve an element int value; if (r.pop(value)) { /* succeeded */ } else { /* buffer empty */ } [endsect] [section Discussion] The implementation makes sure that the ring indices do not "lap-around" each other to ensure that no elements are either lost or read twice. Furthermore it must guarantee that read-access to a particular object in [^pop] "happens after" it has been written in [^push]. This is achieved by writing [^head_ ] with "release" and reading it with "acquire". Conversely the implementation also ensures that read access to a particular ring element "happens before" before rewriting this element with a new value by accessing [^tail_] with appropriate ordering constraints. [endsect] [endsect] [section:mp_queue Wait-free multi-producer queue] The purpose of the ['wait-free multi-producer queue] is to allow an arbitrary number of producers to enqueue objects which are retrieved and processed in FIFO order by a single consumer. [section Implementation] [c++] template class waitfree_queue { public: struct node { T data; node * next; }; void push(const T &data) { node * n = new node; n->data = data; node * stale_head = head_.load(boost::memory_order_relaxed); do { n->next = stale_head; } while (!head_.compare_exchange_weak(stale_head, n, boost::memory_order_release)); } node * pop_all(void) { T * last = pop_all_reverse(), * first = 0; while(last) { T * tmp = last; last = last->next; tmp->next = first; first = tmp; } return first; } waitfree_queue() : head_(0) {} // alternative interface if ordering is of no importance node * pop_all_reverse(void) { return head_.exchange(0, boost::memory_order_consume); } private: boost::atomic head_; }; [endsect] [section Usage] [c++] waitfree_queue q; // insert elements q.push(42); q.push(2); // pop elements waitfree_queue::node * x = q.pop_all() while(x) { X * tmp = x; x = x->next; // process tmp->data, probably delete it afterwards delete tmp; } [endsect] [section Discussion] The implementation guarantees that all objects enqueued are processed in the order they were enqueued by building a singly-linked list of object in reverse processing order. The queue is atomically emptied by the consumer and brought into correct order. It must be guaranteed that any access to an object to be enqueued by the producer "happens before" any access by the consumer. This is assured by inserting objects into the list with ['release] and dequeuing them with ['consume] memory order. It is not necessary to use ['acquire] memory order in [^waitfree_queue::pop_all] because all operations involved depend on the value of the atomic pointer through dereference [endsect] [endsect]