c++ - std::shared_ptr 在线程中使用时崩溃

Question

在线程 1 中(释义代码):

std::vector<std::shared_ptr<Object>> list;

// Initialization
list.reserve(prop_count);

for (size_t i = 0; i < count; ++i)
{
    list.push_back(std::shared_ptr<Object>());
}

// Looped code
for (auto iter = indexes.begin(); iter != indexes.end(); ++iter)
{
    uint32_t i = *iter;

    std::shared_ptr<Object> item = make_object(table->data[i]);  // returns a shared_ptr of Object
    list[i].swap(item);
}

在线程 2 中(释义代码):

for(auto iter = list.begin(); iter != list.end(); ++iter)
{
    shared_ptr<Property> o(*iter);

    if(o)
    {
         // some work with casting it
         // dynamic_pointer_cast
    }
}  // <--- crashes here (after o is out of scope)

这是调用堆栈:

0x006ea218  C/C++
std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)1>::_M_release(this = 0x505240)  C/C++
std::__shared_count<(__gnu_cxx::_Lock_policy)1>::~__shared_count(this = 0xb637dc94) C/C++
std::__shared_ptr<Property, (__gnu_cxx::_Lock_policy)1>::~__shared_ptr(this = 0xb637dc90)   C/C++
std::shared_ptr<Property>::~shared_ptr(this = 0xb637dc90)   C/C++
startSending()  C/C++
libpthread.so.0!start_thread()  C/C++
libc.so.6 + 0xb52b8 C/C++

查看shared_ptr_base.h，好像这里崩溃了:

if (__gnu_cxx::__exchange_and_add_dispatch(&_M_use_count, -1) == 1)
  {
        _GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER(&_M_use_count);
    _M_dispose();  // <--- HERE

我不确定如何解决这个问题。任何帮助表示赞赏。谢谢!

Answer 1

来自 http://en.cppreference.com/w/cpp/memory/shared_ptr我强调的是:

If multiple threads of execution access the same shared_ptr without synchronization and any of those accesses uses a non-const member function of shared_ptr then a data race will occur; the shared_ptr overloads of atomic functions can be used to prevent the data race.

在这种情况下，list[i]和 *iter是相同的实例。

对于线程1，推荐std::atomic_store(&list[i], item)而不是 list[i].swap(item)

对于线程2，推荐std::shared_ptr<Property> o(std::atomic_load(&*iter))而不是 std::shared_ptr<Property> o(*iter);

这一切都假设 vector 的大小不会改变，并引入容器的线程安全、迭代器失效等问题。不过，这超出了这个问题的范围，并在别处涵盖。