[英]C++11, std::atomic broken in clang 3.2 and libc++?
我在Mac OS X 10.8上遇到了一个奇怪的问题:XCode版本4.6.3(4H1503)和clang:Apple LLVM版本4.2(clang-425.0.28)(基于LLVM 3.2svn)目标:x86_64-apple-darwin12.4.0线程模型:posix
我创建了一个非常基本的thread_pool类,并希望对它进行压力测试。
struct last_item_exception {};
template<class Task>
struct thread_reclaimer
{
explicit
thread_reclaimer(impl::basic_task_queue<Task>& q)noexcept
: q_(q)
{}
void operator()()
{
try
{
while(true)
{
using namespace boost;
q_.wait_and_pop()();
}
}
catch(last_item_exception)
{
TRACE_LOG << "caught last item exception";
}
}
private:
impl::basic_task_queue<Task>& q_;
};
template<class Task>
thread_reclaimer<Task> make_thread_reclaimer
(impl::basic_task_queue<Task>& queue)noexcept
{
return std::move(thread_reclaimer<Task>(queue));
}
class thread_pool : public boost::noncopyable
{
typedef std::function<void()> task_type;
public:
explicit thread_pool(size_t number_of_threads)
: threads_created_()
, pool_()
, queue_()
, done_()
{
try
{
TRACE_LOG << "thread_pool: creating " <<number_of_threads << " threads";
while(threads_created_<number_of_threads)
{
++threads_created_;
pool_.create_thread(make_thread_reclaimer(queue_));
TRACE_LOG << "thread_pool: created thread number: "
<< threads_created_
;
}
TRACE_LOG << "thread_pool: all threads started";
}
catch(...)
{
TRACE_LOG << "thread_pool: exception occured";
finish_and_join_all();
throw;
}
TRACE_LOG << "thread_pool: constructor finished";
}
~thread_pool()
{
finish_and_join_all();
}
bool enqueue(task_type t)
{
if(done_) return false;
queue_.push(t, done_);
return !done_; //if done was set inbetween, no push occured!
}
void finish_and_join_all()
{
TRACE_LOG << "entered: finish_and_join_all & done is: " << done_;
if(done_) return;
done_ = true;
std::vector<task_type> cancellation_tasks
( threads_created_
, []
{
TRACE_LOG << "throwing last item exception";
throw impl::last_item_exception();
}
)
;
TRACE_LOG << "pushing last item to the queue";
// atomically pushes all cancellation tasks
queue_.push(cancellation_tasks.begin(), cancellation_tasks.end());
threads_created_ = 0;
TRACE_LOG << "waiting for all threads to finish";
pool_.join_all();
}
private:
size_t threads_created_;
boost::thread_group pool_;
//thread safe producer consumer queue, which uses condition variables for sync
impl::basic_task_queue<std::function<void()>> queue_;
std::atomic<bool> done_;
};
我的测试用例创建并销毁一个带有1个线程的thread_pool对象,即
for(size_t i = 0; i<100; ++i)
thread_pool pool(1);
上第二循环,因为迭代测试用例失败done_
在进入第一次finish_and_join_all()
是true
。 将std::atomic<bool>
类型更改为volatile bool
解决问题。
测试的输出低于(导致condition_variable)失败,因为析构函数没有等待所有线程完成:
Entering test case "ctor_stress_test"
Assertion failed: (!ret), function ~condition_variable, file /usr/local/include/boost/thread/pthread/condition_variable_fwd.hpp, line 86.
[2013-07-03 22:17:32.462265] [0x76f73180] [trace] thread_pool: creating 1 threads
[2013-07-03 22:17:32.462303] [0x76f73180] [trace] thread_pool: created thread number: 1
[2013-07-03 22:17:32.462312] [0x76f73180] [trace] thread_pool: all threads started
[2013-07-03 22:17:32.462320] [0x76f73180] [trace] thread_pool: constructor finished
[2013-07-03 22:17:32.462326] [0x76f73180] [trace] entered: finish_and_join_all & done is: false
[2013-07-03 22:17:32.462332] [0x76f73180] [trace] pushing last item to the queue
[2013-07-03 22:17:32.462357] [0x76f73180] [trace] waiting for all threads to finish
[2013-07-03 22:17:32.462383] [0x10d9e000] [trace] throwing last item exception
[2013-07-03 22:17:32.462404] [0x10d9e000] [trace] caught last item exception
[2013-07-03 22:17:32.462434] [0x76f73180] [trace] executing ~basic_task_queue
[2013-07-03 22:17:32.462447] [0x76f73180] [trace] thread_pool: creating 1 threads
[2013-07-03 22:17:32.462474] [0x76f73180] [trace] thread_pool: created thread number: 1
[2013-07-03 22:17:32.462480] [0x76f73180] [trace] thread_pool: all threads started
[2013-07-03 22:17:32.462485] [0x76f73180] [trace] thread_pool: constructor finished
[2013-07-03 22:17:32.462490] [0x76f73180] [trace] entered: finish_and_join_all & done is: true
[2013-07-03 22:17:32.462495] [0x76f73180] [trace] executing ~basic_task_queue
unknown location:0: fatal error in "ctor_stress_test": signal: SIGABRT (application abort requested)
Leaving test case "ctor_stress_test"; testing time: 448mks
这是std :: atomic的预期行为吗?
原子对象的初始化不是原子操作,而且,默认初始化使其处于未初始化状态(根据§29.6.5[atomics.types.operations.req]/4
)。
使用构造函数开头的原子或至少定义良好的操作将其设置为false
。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.