Boost.Python和Boost.Signals2：分段錯誤

Question

我在將我的boost.python公開的現有C ++庫中集成boost.signals2時遇到問題。

我有一個使用std::shared_ptr公開給python的類。 該類應該能夠在某些事件上發出一些信號。 因此，我公開了一個connect_slot函數，該函數以boost::python::object作為參數。 如果我在連接插槽后立即發出信號，則一切正常，但是如果稍后再上課，則收到分段錯誤。

我認為這可能與c ++ lib中的線程有關（它也使用boos :: asio等）。

以下是一些代碼段：

MyClass.h：

public:
    typedef boost::signals2::signal<void (std::shared_ptr<int>)> signal_my_sig;
    void connect_slot(boost::python::object const & slot);

private:
    signal_my_sig    m_sig;

MyClass.cpp：

void MyClass::connect_slot(boost::python::object const & slot) { 
    std::cout << "register shd" << std::endl;
    m_sig.connect(slot);

    m_sig(12345); // this works
}


void MyClass::some_later_event() {
    m_sig(654321); // this does not work

}

我用這樣的自定義python函數在python中調用MyClass :: connect_slot函數：

def testfunc(some_int):
    print("slot called")

m = myext.MyClass()
m.connect_slot(testfunc)

MyClass::some_later_event引發的分段錯誤的回溯（使用gdb）如下所示：

[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff3c37700 (LWP 20634)]

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7ffff3c37700 (LWP 20634)]
0x00000000004f7480 in PyObject_Call ()
(gdb) 
(gdb) backtrace
#0  0x00000000004f7480 in PyObject_Call ()
#1  0x00000000004f7aa6 in PyEval_CallObjectWithKeywords ()
#2  0x000000000049bd84 in PyEval_CallFunction ()
#3  0x00007ffff5375d9f in boost::python::call<boost::python::api::object, int>
(callable=0x7ffff7ed4578, a0=@0x7ffff3c35b34: 5)
at /usr/local/boost_1_55_0/boost/python/call.hpp:66
#4  0x00007ffff5374b81 in boost::python::api::object_operators<boost::python::api::object>::operator()<int> (this=0x9e3bf0, a0=@0x7ffff3c35b34: 5)
at /usr/local/boost_1_55_0/boost/python/object_call.hpp:19
#5  0x00007ffff5373658 in boost::detail::function::void_function_obj_invoker1<boost::python::api::object, void, int>::invoke (function_obj_ptr=..., a0=5)
at /usr/local/boost_1_55_0/boost/function/function_template.hpp:153
#6  0x00007ffff5378a3c in boost::function1<void, int>::operator() (
this=0x9e3be8, a0=5)
at /usr/local/boost_1_55_0/boost/function/function_template.hpp:767
#7  0x00007ffff53781f9 in boost::signals2::detail::call_with_tuple_args<boost::signals2::detail::void_type>::m_invoke<boost::function<void (int)>, 0u, int&>(void*, boost::function<void (int)>&, boost::signals2::detail::unsigned_meta_array<0u>, std::tuple<int&>) const (this=0x7ffff3c35c7f, func=..., args=...)
at /usr/local/boost_1_55_0/boost/signals2/detail/variadic_slot_invoker.hpp:92

有任何想法嗎？

Answer 1

如果從未顯式管理全局解釋器鎖 （GIL）的C ++線程調用MyClass::some_later_event() ，則可能導致未定義的行為。

---

Python和C ++線程。

讓我們考慮一下C ++線程與Python交互的情況。 例如，可以將C ++線程設置為通過MyClass.event_in(seconds, value)在一段時間后調用MyClass的信號。

這個例子可能會涉及很多，所以讓我們從基礎開始：Python的GIL。 簡而言之，GIL是解釋器周圍的互斥體。 如果線程做的任何事情都會影響python受管理對象的引用計數，則它需要獲取GIL。 在GDB追溯中，Boost.Signals2庫可能試圖在沒有GIL的情況下調用Python對象，從而導致崩潰。 盡管管理GIL非常簡單，但它可能很快變得復雜。

首先，模塊需要讓Python初始化GIL以進行線程化。

BOOST_PYTHON_MODULE(example)
{
  PyEval_InitThreads(); // Initialize GIL to support non-python threads.
  // ...
}

為了方便起見，讓我們創建一個簡單的類來幫助通過范圍管理GIL：

/// @brief RAII class used to lock and unlock the GIL.
class gil_lock
{
public:
  gil_lock()  { state_ = PyGILState_Ensure(); }
  ~gil_lock() { PyGILState_Release(state_);   }
private:
  PyGILState_STATE state_;
};

讓我們確定C ++線程何時需要GIL：

• boost::signals2::signal可以創建連接對象的其他副本，就像在同時調用信號時所做的那樣。
• 調用通過boost::signals2::signal連接的Python對象。 回調肯定會影響python對象。 例如，提供給__call__方法的self參數將增加和減少對象的引用計數。

MyClass類。

這是一個基於原始代碼的基本模型類：

/// @brief Mockup class.
class MyClass
{
public:
  /// @brief Connect a slot to the signal.
  template <typename Slot>
  void connect_slot(const Slot& slot)
  {
    signal_.connect(slot);
  }

  /// @brief Send an event to the signal.
  void event(int value)
  {
    signal_(value);
  }

private:
  boost::signals2::signal<void(int)> signal_;
};

由於C ++線程可能正在調用MyClass的信號，因此MyClass的生存期必須至少與線程一樣長。 一個很好的候選方法是讓Boost.Python使用boost::shared_ptr管理MyClass 。

BOOST_PYTHON_MODULE(example)
{
  PyEval_InitThreads(); // Initialize GIL to support non-python threads.

  namespace python = boost::python;
  python::class_<MyClass, boost::shared_ptr<MyClass>,
                 boost::noncopyable>("MyClass")
    .def("event", &MyClass::event)
    // ...
    ;
}

boost::signals2::signal與python對象進行交互。

boost::signals2::signal在被調用時可能會復制。 另外，可能有C ++插槽連接到信號，因此，理想的是僅在調用Python插槽時鎖定GIL。 但是， signal沒有提供鈎子來允許我們在創建插槽副本或調用插槽之前獲取GIL。

為了避免產生signal創建boost::python::object插槽的副本，可以使用包裝類來創建boost::python::object的副本，以便引用計數保持准確，並通過shared_ptr管理副本。 這允許signal自由地創建shared_ptr副本，而不是在沒有GIL的情況下復制boost::python::object 。

該GIL安全插槽可以封裝在幫助程序類中。

/// @brief Helepr type that will manage the GIL for a python slot.
///
/// @detail GIL management:
///           * Caller must own GIL when constructing py_slot, as 
///             the python::object will be copy-constructed (increment
///             reference to the object)
///           * The newly constructed python::object will be managed
///             by a shared_ptr.  Thus, it may be copied without owning
///             the GIL.  However, a custom deleter will acquire the
///             GIL during deletion.
///           * When py_slot is invoked (operator()), it will acquire
///             the GIL then delegate to the managed python::object.
struct py_slot
{
public:

  /// @brief Constructor that assumes the caller has the GIL locked.
  py_slot(const boost::python::object& object)
    : object_(
        new boost::python::object(object),  // GIL locked, so copy.
        [](boost::python::object* object)   // Delete needs GIL.
        {
          gil_lock lock;
          delete object;
        }
      )
  {}

  // Use default copy-constructor and assignment-operator.
  py_slot(const py_slot&) = default;
  py_slot& operator=(const py_slot&) = default;

  template <typename ...Args>
  void operator()(Args... args)
  {
    // Lock the GIL as the python object is going to be invoked.
    gil_lock lock;
    (*object_)(args...); 
  }

private:
  boost::shared_ptr<boost::python::object> object_;
};

Python將公開一個輔助函數，以幫助修改類型。

/// @brief MyClass::connect_slot helper.
template <typename ...Args>
void MyClass_connect_slot(
  MyClass& self,
  boost::python::object object)
{
  py_slot slot(object); // Adapt object to a py_slot for GIL management.

  // Using a lambda here allows for the args to be expanded automatically.
  // If bind was used, the placeholders would need to be explicitly added.
  self.connect_slot([slot](Args... args) mutable { slot(args...); });
}

更新的綁定公開了輔助函數：

python::class_<MyClass, boost::shared_ptr<MyClass>,
               boost::noncopyable>("MyClass")
  .def("connect_slot", &MyClass_connect_slot<int>)
  .def("event",        &MyClass::event)
  // ...
  ;

線程本身。

線程的功能是相當基本的：它先休眠然后調用信號。 但是，了解GIL的上下文很重要。

/// @brief Sleep then invoke an event on MyClass.
template <typename ...Args>
void MyClass_event_in_thread(
  boost::shared_ptr<MyClass> self,
  unsigned int seconds,
  Args... args)
{
  // Sleep without the GIl.
  std::this_thread::sleep_for(std::chrono::seconds(seconds));

  // We do not want to hold the GIL while invoking or copying 
  // C++-specific slots connected to the signal.  Thus, it is the 
  // responsibility of python slots to manage the GIL via the 
  // py_slot wrapper class.
  self->event(args...);
}

/// @brief Function that will be exposed to python that will create
///        a thread to call the signal.
template <typename ...Args>
void MyClass_event_in(
  boost::shared_ptr<MyClass> self,
  unsigned int seconds,
  Args... args)
{
  // The caller may or may not have the GIL.  Regardless, spawn off a 
  // thread that will sleep and then invoke an event on MyClass.  The
  // thread will not be joined so detach from it.  Additionally, as
  // shared_ptr is thread safe, copies of it can be made without the
  // GIL.
  std::thread(&MyClass_event_in_thread<Args...>, self, seconds, args...)
      .detach();
}

_{請注意， MyClass_event_in_thread可以表示為lambda，但是在lambda中解MyClass_event_in_thread模板包在某些編譯器上不起作用。}

並更新了MyClass綁定。

python::class_<MyClass, boost::shared_ptr<MyClass>,
               boost::noncopyable>("MyClass")
  .def("connect_slot", &MyClass_connect_slot<int>)
  .def("event",        &MyClass::event)
  .def("event_in",     &MyClass_event_in<int>)
  ;

---

最終的解決方案如下所示：

#include <thread> // std::thread, std::chrono
#include <boost/python.hpp>
#include <boost/shared_ptr.hpp>
#include <boost/signals2/signal.hpp>

/// @brief Mockup class.
class MyClass
{
public:
  /// @brief Connect a slot to the signal.
  template <typename Slot>
  void connect_slot(const Slot& slot)
  {
    signal_.connect(slot);
  }

  /// @brief Send an event to the signal.
  void event(int value)
  {
    signal_(value);
  }

private:
  boost::signals2::signal<void(int)> signal_;
};

/// @brief RAII class used to lock and unlock the GIL.
class gil_lock
{
public:
  gil_lock()  { state_ = PyGILState_Ensure(); }
  ~gil_lock() { PyGILState_Release(state_);   }
private:
  PyGILState_STATE state_;
};    

/// @brief Helepr type that will manage the GIL for a python slot.
///
/// @detail GIL management:
///           * Caller must own GIL when constructing py_slot, as 
///             the python::object will be copy-constructed (increment
///             reference to the object)
///           * The newly constructed python::object will be managed
///             by a shared_ptr.  Thus, it may be copied without owning
///             the GIL.  However, a custom deleter will acquire the
///             GIL during deletion.
///           * When py_slot is invoked (operator()), it will acquire
///             the GIL then delegate to the managed python::object.
struct py_slot
{
public:

  /// @brief Constructor that assumes the caller has the GIL locked.
  py_slot(const boost::python::object& object)
    : object_(
        new boost::python::object(object),  // GIL locked, so copy.
        [](boost::python::object* object)   // Delete needs GIL.
        {
          gil_lock lock;
          delete object;
        }
      )
  {}

  // Use default copy-constructor and assignment-operator.
  py_slot(const py_slot&) = default;
  py_slot& operator=(const py_slot&) = default;

  template <typename ...Args>
  void operator()(Args... args)
  {
    // Lock the GIL as the python object is going to be invoked.
    gil_lock lock;
    (*object_)(args...); 
  }

private:
  boost::shared_ptr<boost::python::object> object_;
};

/// @brief MyClass::connect_slot helper.
template <typename ...Args>
void MyClass_connect_slot(
  MyClass& self,
  boost::python::object object)
{
  py_slot slot(object); // Adapt object to a py_slot for GIL management.

  // Using a lambda here allows for the args to be expanded automatically.
  // If bind was used, the placeholders would need to be explicitly added.
  self.connect_slot([slot](Args... args) mutable { slot(args...); });
}

/// @brief Sleep then invoke an event on MyClass.
template <typename ...Args>
void MyClass_event_in_thread(
  boost::shared_ptr<MyClass> self,
  unsigned int seconds,
  Args... args)
{
  // Sleep without the GIL.
  std::this_thread::sleep_for(std::chrono::seconds(seconds));

  // We do not want to hold the GIL while invoking or copying 
  // C++-specific slots connected to the signal.  Thus, it is the 
  // responsibility of python slots to manage the GIL via the 
  // py_slot wrapper class.
  self->event(args...);
}

/// @brief Function that will be exposed to python that will create
///        a thread to call the signal.
template <typename ...Args>
void MyClass_event_in(
  boost::shared_ptr<MyClass> self,
  unsigned int seconds,
  Args... args)
{
  // The caller may or may not have the GIL.  Regardless, spawn off a 
  // thread that will sleep and then invoke an event on MyClass.  The
  // thread will not be joined so detach from it.  Additionally, as
  // shared_ptr is thread safe, copies of it can be made without the
  // GIL.
  // Note: MyClass_event_in_thread could be expressed as a lambda,
  //       but unpacking a template pack within a lambda does not work
  //       on some compilers.
  std::thread(&MyClass_event_in_thread<Args...>, self, seconds, args...)
      .detach();
}

BOOST_PYTHON_MODULE(example)
{
  PyEval_InitThreads(); // Initialize GIL to support non-python threads.

  namespace python = boost::python;
  python::class_<MyClass, boost::shared_ptr<MyClass>,
                 boost::noncopyable>("MyClass")
    .def("connect_slot", &MyClass_connect_slot<int>)
    .def("event",        &MyClass::event)
    .def("event_in",     &MyClass_event_in<int>)
    ;
}

還有一個測試腳本：

from time import sleep
import example

def spam1(x):
  print "spam1: ", x

def spam2(x):
  print "spam2: ", x

c = example.MyClass()
c.connect_slot(spam1)
c.connect_slot(spam2)
c.event(123)
print "Sleeping"
c.event_in(3, 321)
sleep(5)
print "Done sleeping"

結果如下：

spam1:  123
spam2:  123
Sleeping
spam1:  321
spam2:  321
Done sleeping

Answer 2

感謝Tanner Sansbury鏈接到他對此帖子的回答。 這解決了我的問題，除了我無法調用接受參數的信號。

我通過編輯py_slot類解決了這個問題：

struct py_slot {
    public:
        /// @brief Constructor that assumes the caller has the GIL locked.
        py_slot(const boost::python::object& object)
            : object_(new boost::python::object(object),   // GIL locked, so     copy.
            py_deleter<boost::python::object>()) // Delete needs GIL.
            {}

        void operator()(SomeParamClass param) {
            // Lock the gil as the python object is going to be invoked.
            gil_lock lock;

            (*object_)(param);

    private:
        boost::shared_ptr<boost::python::object> object_;
};

boost :: bind調用看起來像這樣：

self->connect_client_ready(boost::bind(&py_slot<SomeParamClass>::operator(), py_slot<SomeParamClass>(object), _1)); // note the _1