Chapter 45. Boost.Atomic

Boost.Atomic provides the class boost::atomic, which can be used to create atomic variables. They are called atomic variables because all access is atomic. Boost.Atomic is used in multithreaded programs when access to a variable in one thread shouldn’t be interrupted by another thread accessing the same variable. Without boost::atomic, attempts to access shared variables from multiple threads would need to be synchronized with locks.

boost::atomic depends on the target platform supporting atomic variable access. Otherwise, boost::atomic uses locks. The library allows you to detect whether a target platform supports atomic variable access.

If your development environment supports C++11, you don’t need Boost.Atomic. The C++11 standard library provides a header file atomic that defines the same functionality as Boost.Atomic. For example, you will find a class named std::atomic.

Boost.Atomic supports more or less the same functionality as the standard library. While a few functions are overloaded in Boost.Atomic, they may have different names in the standard library. The standard library also provides some functions, such as std::atomic_init() and std::kill_dependency(), which are missing in Boost.Atomic.

#include <boost/atomic.hpp>
#include <thread>
#include <iostream>

boost::atomic<int> a{0};

void thread()
{
  ++a;
}

int main()
{
  std::thread t1{thread};
  std::thread t2{thread};
  t1.join();
  t2.join();
  std::cout << a << '\n';
}

Example 45.1 uses two threads to increment the int variable a. Instead of a lock, the example uses boost::atomic for atomic access to a. The example writes 2 to standard output.

boost::atomic works because some processors support atomic access on variables. If incrementing an int variable is an atomic operation, a lock isn’t required. If this example is run on a platform that cannot increment a variable as an atomic operation, boost::atomic uses a lock.

#include <boost/atomic.hpp>
#include <iostream>

int main()
{
  std::cout.setf(std::ios::boolalpha);

  boost::atomic<short> s;
  std::cout << s.is_lock_free() << '\n';

  boost::atomic<int> i;
  std::cout << i.is_lock_free() << '\n';

  boost::atomic<long> l;
  std::cout << l.is_lock_free() << '\n';
}

You can call is_lock_free() on an atomic variable to check whether accessing the variable is done without a lock. If you run the example on an Intel x86 processor, it displays true three times. If you run it on a processor without lock-free access on short, int and long variables, false is displayed.

Boost.Atomic provides the BOOST_ATOMIC_INT_LOCK_FREE and BOOST_ATOMIC_LONG_LOCK_FREE macros to check, at compile time, which data types support lock-free access.

Example 45.2 uses integral data types only. You should not use boost::atomic with classes like std::string or std::vector. Boost.Atomic supports integers, pointers, booleans (bool), and trivial classes. Examples of integral types include short, int and long. Trivial classes define objects that can be copied with std::memcpy().

#include <boost/atomic.hpp>
#include <thread>
#include <iostream>

boost::atomic<int> a{0};

void thread()
{
  a.fetch_add(1, boost::memory_order_seq_cst);
}

int main()
{
  std::thread t1{thread};
  std::thread t2{thread};
  t1.join();
  t2.join();
  std::cout << a << '\n';
}

Example 45.3 increments a twice – this time not with operator++ but with a call to fetch_add(). The member function fetch_add() can take two parameters: the number by which a should be incremented and the memory order.

The memory order specifies the order in which access operations on memory must occur. By default, this order is undetermined and does not depend on the order of the lines of code. Compilers and processors are allowed to change the order as long as a program behaves as if memory access operations were executed in source code order. This rule only applies to a thread. If more than one thread is used, variations in the order of memory accesses can lead to a program acting erroneously. Boost.Atomic supports specifying a memory order when accessing variables to make sure memory accesses occur in the desired order in a multithreaded program.

Example 45.3 uses the memory order boost::memory_order_seq_cst to increment a by 1. The memory order stands for sequential consistency. This is the most restrictive memory order. All memory accesses that appear before the fetch_add() call must occur before this member function is executed. All memory accesses that appear after the fetch_add() call must occur after this member function is executed. Compilers and processors may reorder memory accesses before and after the call to fetch_add(), but they must not move a memory access from before to after the call to fetch_add() or vice versa. boost::memory_order_seq_cst is a strict boundary for memory accesses in both directions.

boost::memory_order_seq_cst is the most restrictive memory order. It is used by default when memory order is not set. Therefore, in Example 45.1, when a is incremented with operator++, boost::memory_order_seq_cst will be used.

boost::memory_order_seq_cst isn’t always necessary. For example, in Example 45.3 there is no need to synchronize memory accesses for other variables because a is the only variable the threads use. a is written to standard output in main(), but only after both threads have terminated. The call to join() guarantees that a is only read after both threads have finished.

#include <boost/atomic.hpp>
#include <thread>
#include <iostream>

boost::atomic<int> a{0};

void thread()
{
  a.fetch_add(1, boost::memory_order_relaxed);
}

int main()
{
  std::thread t1{thread};
  std::thread t2{thread};
  t1.join();
  t2.join();
  std::cout << a << '\n';
}

Example 45.4 sets the memory order to boost::memory_order_relaxed. This is the least restrictive memory order: it allows arbitrary reordering of memory accesses. This example works with this memory order because the threads access no variables except a. Therefore, no specific order is required.

#include <boost/atomic.hpp>
#include <thread>
#include <iostream>

boost::atomic<int> a{0};
int b = 0;

void thread1()
{
  b = 1;
  a.store(1, boost::memory_order_release);
}

void thread2()
{
  while (a.load(boost::memory_order_acquire) != 1)
    ;
  std::cout << b << '\n';
}

int main()
{
  std::thread t1{thread1};
  std::thread t2{thread2};
  t1.join();
  t2.join();
}

There are choices between the most restrictive memory order, boost::memory_order_seq_cst, and the least restrictive one, boost::memory_order_relaxed. Example 45.5 introduces the memory orders boost::memory_order_release and boost::memory_order_acquire.

Memory accesses that appear in the code before the boost::memory_order_release statement are executed before the boost::memory_order_release statement is executed. Compilers and processors must not move memory accesses from before to after boost::memory_order_release. However, they may move memory accesses from after to before boost::memory_order_release.

boost::memory_order_acquire works like boost::memory_order_release, but refers to memory accesses after boost::memory_order_acquire. Compilers and processors must not move memory accesses from after the boost::memory_order_acquire statement to before it. However, they may move memory accesses from before to after boost::memory_order_acquire.

Example 45.5 uses boost::memory_order_release in the first thread to make sure that b is set to 1 before a is set to 1. boost::memory_order_release guarantees that the memory access on b occurs before the memory access on a.

To specify a memory order when accessing a, store() is called. This member function corresponds to an assignment with operator=.

The second thread reads a in a loop. This is done with the member function load(). Again, no assignment operator is used.

In the second thread, boost::memory_order_acquire makes sure that the memory access on b doesn’t occur before the memory access on a. The second thread waits in the loop for a to be set to 1 by the first thread. Once this happens, b is read.

The example writes 1 to standard output. The memory orders ensure that all memory accesses occur in the right order. The first thread always writes 1 to b first before the second thread accesses and reads b.