boost::asio bug? task_io_service перед уничтожением io_service

Я обнаружил странную ошибку в своем коде. Вот автономный тестовый пример, к которому мне удалось его получить.

#include <memory>
#include <thread>
#include <stack>
#include <system_error>

#include <boost/asio.hpp>

using boost::asio::io_service;
using std::placeholders::_1;

class async_service
{
public:
    async_service();
    async_service(size_t number_threads);
    ~async_service();

    async_service(const async_service&) = delete;
    void operator=(const async_service&) = delete;

    void spawn();
    void shutdown();

    io_service& get_service();
    const io_service& get_service() const;

private:
    io_service service_;
    io_service::work* work_;
    std::vector<std::thread> threads_;
};

async_service::async_service()
  : work_(nullptr)
{
}

async_service::async_service(size_t number_threads)
  : work_(nullptr)
{
    for (size_t i = 0; i < number_threads; ++i)
        spawn();
}

async_service::~async_service()
{
    std::cout << __PRETTY_FUNCTION__ << std::endl;
    service_.stop();
    for (std::thread& t: threads_)
        t.join();
}

void run_service(io_service* service)
{
    service->run();
}

void async_service::spawn()
{
    if (!work_)
        work_ = new io_service::work(service_);
    threads_.push_back(std::thread(run_service, &service_));
}
void async_service::shutdown()
{
    delete work_;
    work_ = nullptr;
}

io_service& async_service::get_service()
{
    return service_;
}
const io_service& async_service::get_service() const
{
    return service_;
}

// --------------------------------------------------------------

template <typename... Args>
class subscriber
  : public std::enable_shared_from_this<subscriber<Args...>>
{
public:
    typedef std::function<void (Args...)> handler_type;
    typedef std::shared_ptr<subscriber<Args...>> ptr;

    subscriber(async_service& service)
      : strand_(service.get_service())
    {
    }

    void subscribe(handler_type handle)
    {
        strand_.dispatch(
            std::bind(&subscriber<Args...>::do_subscribe,
                this->shared_from_this(), handle));
    }

    void relay(Args... params)
    {
        strand_.dispatch(
            std::bind(&subscriber<Args...>::do_relay,
                this->shared_from_this(), std::forward<Args>(params)...));
    }

private:
    typedef std::stack<handler_type> registry_stack;

    void do_subscribe(handler_type handle)
    {
        registry_.push(handle);
    }

    void do_relay(Args... params)
    {
        registry_stack notify_copy = std::move(registry_);
        registry_ = registry_stack();
        while (!notify_copy.empty())
        {
            notify_copy.top()(params...);
            notify_copy.pop();
        }
        assert(notify_copy.empty());
    }

    io_service::strand strand_;
    registry_stack registry_;
};

// --------------------------------------------------------

class lala_channel_proxy
  : public std::enable_shared_from_this<lala_channel_proxy>
{
public:
    typedef std::function<void (const std::error_code&)> receive_inventory_handler;

    lala_channel_proxy(async_service& service)
      : strand_(service.get_service())
    {
        inventory_subscriber_ =
            std::make_shared<inventory_subscriber_type>(service);
    }

    void start()
    {
        read_header();
    }

    void subscribe_inventory(receive_inventory_handler handle_receive)
    {
        inventory_subscriber_->subscribe(handle_receive);
    }

    typedef subscriber<const std::error_code&> inventory_subscriber_type;

    void read_header()
    {
        strand_.post(
            std::bind(&lala_channel_proxy::handle_read_header,
                shared_from_this(), boost::system::error_code(), 0));
    }

    void handle_read_header(const boost::system::error_code& ec,
        size_t bytes_transferred)
    {
        std::cout << "inventory ----------" << std::endl;
        inventory_subscriber_->relay(std::error_code());
        sleep(1.0);
        read_header();
    }

    io_service::strand strand_;
    inventory_subscriber_type::ptr inventory_subscriber_;
};

typedef std::shared_ptr<lala_channel_proxy> lala_channel_proxy_ptr;

class lala_channel
{
public:
    lala_channel(async_service& service)
    {
        lala_channel_proxy_ptr proxy =
            std::make_shared<lala_channel_proxy>(service);
        proxy->start();
        //weak_proxy_ = proxy;
        strong_proxy_ = proxy;
    }
    void subscribe_inventory(
        lala_channel_proxy::receive_inventory_handler handle_receive)
    {
        lala_channel_proxy_ptr proxy = strong_proxy_;
        proxy->subscribe_inventory(handle_receive);
    }
    lala_channel_proxy_ptr strong_proxy_;
    // Normally this has a weak pointer to the channel pimpl to allow
    // it to die, but whether it uses a weak_ptr or shared_ptr makes
    // no difference.
    //std::weak_ptr<channel_proxy> weak_proxy_;
};

typedef std::shared_ptr<lala_channel> lala_channel_ptr;
//typedef lala_channel_proxy_ptr lala_channel_ptr;

class session
  : public std::enable_shared_from_this<session>
{
public:
    typedef std::function<void (const std::error_code&)> completion_handler;

    session(async_service& service, async_service& mempool_service,
            async_service& disk_service)
      : strand_(service.get_service()),
        txpool_strand_(mempool_service.get_service()),
        chain_strand_(disk_service.get_service()), service_(service)
    {
    }

    void start()
    {
        auto this_ptr = shared_from_this();
        lala_channel_ptr node =
            std::make_shared<lala_channel>(service_);
        node->subscribe_inventory(
            std::bind(&session::inventory, this_ptr, _1, node));
        for (size_t i = 0; i < 500; ++i)
        {
            chain_strand_.post(
                []()
                {
                    std::cout << "HERE!" << std::endl;
                    sleep(2);
                });
        }
    }

private:
    void inventory(const std::error_code& ec, lala_channel_ptr node)
    {
        if (ec)
        {
            std::cerr << ec.message() << std::endl;
            return;
        }
        auto this_ptr = shared_from_this();
        txpool_strand_.post([]() {});
        node->subscribe_inventory(
            std::bind(&session::inventory, this_ptr, _1, node));
    }

    async_service& service_;
    io_service::strand txpool_strand_, strand_, chain_strand_;
};

int main()
{
    // First level
    {
        // Bug only happens for this ordering of async_service's
        // That means it is only triggered when they are destroyed in
        // this reverse order.
        async_service network_service(1), disk_service(1), mempool_service(1);
        //async_service network_service(1), mempool_service(1), disk_service(1);
        //async_service disk_service(1), mempool_service(1), network_service(1);
        //async_service disk_service(1), network_service(1), mempool_service(1);
        //async_service mempool_service(1), disk_service(1), network_service(1);
        //async_service mempool_service(1), network_service(1), disk_service(1);

        // Second level
        {
            // Should be kept alive by io_service
            auto s = std::make_shared<session>(network_service, mempool_service, disk_service);
            s->start();
        }
        //network_service.shutdown();
        //disk_service.shutdown();
        //mempool_service.shutdown();
        sleep(3);
    // Never gets past here
    }
    std::cout << "Exiting..." << std::endl;
    return 0;
}

Когда я запускаю его, я получаю следующее:

$ g++ -std=c++0x /tmp/ideone_y6OlI.cpp -lboost_system -pthread -ggdb
$ gdb a.out 
GNU gdb (Ubuntu/Linaro 7.4-2012.04-0ubuntu2) 7.4-2012.04
Copyright (C) 2012 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-linux-gnu".
For bug reporting instructions, please see:
<http://bugs.launchpad.net/gdb-linaro/>...
Reading symbols from /home/genjix/src/brokenlibbtc/a.out...done.
(gdb) r
Starting program: /home/genjix/src/brokenlibbtc/a.out 
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff6deb700 (LWP 28098)]
[New Thread 0x7ffff65ea700 (LWP 28099)]
[New Thread 0x7ffff5de9700 (LWP 28100)]
inventory ----------
HERE!
inventory ----------
HERE!
inventory ----------
async_service::~async_service()
async_service::~async_service()
[Thread 0x7ffff5de9700 (LWP 28100) exited]

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7ffff6deb700 (LWP 28098)]
0x0000000000405873 in boost::asio::detail::task_io_service::wake_one_idle_thread_and_unlock (this=0x6255e0, lock=...) at /usr/include/boost/asio/detail/impl/task_io_service.ipp:461
461     first_idle_thread_ = idle_thread->next;

То же самое с бустами 1.48 и 1.49.

Интересно, почему это происходит. Это происходит только с этой очень конкретной конфигурацией. Если я что-то меняю, то ошибка не возникает.

async_service — это удобная оболочка для io_service. Как ни странно, если я изменю io_service на *io_service и не удалю io_service, то ошибки не произойдет... но, конечно, это не имеет значения?

Если вы посмотрите на исходный код в main(), там создано 3 объекта async_service. Каждый из них управляет временем жизни одного io_service.

        // Bug only happens for this ordering of async_service's
        // That means it is only triggered when they are destroyed in
        // this reverse order.
        async_service network_service(1), disk_service(1), mempool_service(1);
        //async_service network_service(1), mempool_service(1), disk_service(1);
        //async_service disk_service(1), mempool_service(1), network_service(1);
        //async_service disk_service(1), network_service(1), mempool_service(1);
        //async_service mempool_service(1), disk_service(1), network_service(1);
        //async_service mempool_service(1), network_service(1), disk_service(1);

Класс абонента представляет объект подписки... вызова для конкретного события. Сеанс и канал адаптированы из более крупной программы, поэтому они могут показаться запутанными/запутанными.