epoll_reactor.hpp

Boost provides free peer-reviewed portable C++ source libraries. We emphasize libraries that work

  // Cancel any operations that are running against the descriptor and remove
  // its registration from the reactor.
  void close_descriptor(socket_type descriptor, per_descriptor_data&)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);

    // Remove the descriptor from epoll.
    epoll_event ev = { 0, { 0 } };
    epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, descriptor, &ev);

    // Cancel any outstanding operations associated with the descriptor.
    cancel_ops_unlocked(descriptor);
  }

  // Add a new timer queue to the reactor.
  template <typename Time_Traits>
  void add_timer_queue(timer_queue<Time_Traits>& timer_queue)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);
    timer_queues_.push_back(&timer_queue);
  }

  // Remove a timer queue from the reactor.
  template <typename Time_Traits>
  void remove_timer_queue(timer_queue<Time_Traits>& timer_queue)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);
    for (std::size_t i = 0; i < timer_queues_.size(); ++i)
    {
      if (timer_queues_[i] == &timer_queue)
      {
        timer_queues_.erase(timer_queues_.begin() + i);
        return;
      }
    }
  }

  // Schedule a timer in the given timer queue to expire at the specified
  // absolute time. The handler object will be invoked when the timer expires.
  template <typename Time_Traits, typename Handler>
  void schedule_timer(timer_queue<Time_Traits>& timer_queue,
      const typename Time_Traits::time_type& time, Handler handler, void* token)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);
    if (!shutdown_)
      if (timer_queue.enqueue_timer(time, handler, token))
        interrupter_.interrupt();
  }

  // Cancel the timer associated with the given token. Returns the number of
  // handlers that have been posted or dispatched.
  template <typename Time_Traits>
  std::size_t cancel_timer(timer_queue<Time_Traits>& timer_queue, void* token)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);
    std::size_t n = timer_queue.cancel_timer(token);
    if (n > 0)
      interrupter_.interrupt();
    return n;
  }

private:
  friend class task_io_service<epoll_reactor<Own_Thread> >;

  // Run epoll once until interrupted or events are ready to be dispatched.
  void run(bool block)
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);

    // Dispatch any operation cancellations that were made while the select
    // loop was not running.
    read_op_queue_.perform_cancellations();
    write_op_queue_.perform_cancellations();
    except_op_queue_.perform_cancellations();
    for (std::size_t i = 0; i < timer_queues_.size(); ++i)
      timer_queues_[i]->dispatch_cancellations();

    // Check if the thread is supposed to stop.
    if (stop_thread_)
    {
      complete_operations_and_timers(lock);
      return;
    }

    // We can return immediately if there's no work to do and the reactor is
    // not supposed to block.
    if (!block && read_op_queue_.empty() && write_op_queue_.empty()
        && except_op_queue_.empty() && all_timer_queues_are_empty())
    {
      complete_operations_and_timers(lock);
      return;
    }

    int timeout = block ? get_timeout() : 0;
    wait_in_progress_ = true;
    lock.unlock();

    // Block on the epoll descriptor.
    epoll_event events[128];
    int num_events = (block || need_epoll_wait_)
      ? epoll_wait(epoll_fd_, events, 128, timeout)
      : 0;

    lock.lock();
    wait_in_progress_ = false;

    // Block signals while performing operations.
    boost::asio::detail::signal_blocker sb;

    // Dispatch the waiting events.
    for (int i = 0; i < num_events; ++i)
    {
      int descriptor = events[i].data.fd;
      if (descriptor == interrupter_.read_descriptor())
      {
        interrupter_.reset();
      }
      else
      {
        bool more_reads = false;
        bool more_writes = false;
        bool more_except = false;
        boost::system::error_code ec;

        // Exception operations must be processed first to ensure that any
        // out-of-band data is read before normal data.
        if (events[i].events & (EPOLLPRI | EPOLLERR | EPOLLHUP))
          more_except = except_op_queue_.perform_operation(descriptor, ec);
        else
          more_except = except_op_queue_.has_operation(descriptor);

        if (events[i].events & (EPOLLIN | EPOLLERR | EPOLLHUP))
          more_reads = read_op_queue_.perform_operation(descriptor, ec);
        else
          more_reads = read_op_queue_.has_operation(descriptor);

        if (events[i].events & (EPOLLOUT | EPOLLERR | EPOLLHUP))
          more_writes = write_op_queue_.perform_operation(descriptor, ec);
        else
          more_writes = write_op_queue_.has_operation(descriptor);

        if ((events[i].events & (EPOLLERR | EPOLLHUP)) != 0
              && (events[i].events & ~(EPOLLERR | EPOLLHUP)) == 0
              && !more_except && !more_reads && !more_writes)
        {
          // If we have an event and no operations associated with the
          // descriptor then we need to delete the descriptor from epoll. The
          // epoll_wait system call can produce EPOLLHUP or EPOLLERR events
          // when there is no operation pending, so if we do not remove the
          // descriptor we can end up in a tight loop of repeated
          // calls to epoll_wait.
          epoll_event ev = { 0, { 0 } };
          epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, descriptor, &ev);
        }
        else
        {
          epoll_event ev = { 0, { 0 } };
          ev.events = EPOLLERR | EPOLLHUP;
          if (more_reads)
            ev.events |= EPOLLIN;
          if (more_writes)
            ev.events |= EPOLLOUT;
          if (more_except)
            ev.events |= EPOLLPRI;
          ev.data.fd = descriptor;
          int result = epoll_ctl(epoll_fd_, EPOLL_CTL_MOD, descriptor, &ev);
          if (result != 0 && errno == ENOENT)
            result = epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, descriptor, &ev);
          if (result != 0)
          {
            ec = boost::system::error_code(errno,
                boost::asio::error::get_system_category());
            read_op_queue_.perform_all_operations(descriptor, ec);
            write_op_queue_.perform_all_operations(descriptor, ec);
            except_op_queue_.perform_all_operations(descriptor, ec);
          }
        }
      }
    }
    read_op_queue_.perform_cancellations();
    write_op_queue_.perform_cancellations();
    except_op_queue_.perform_cancellations();
    for (std::size_t i = 0; i < timer_queues_.size(); ++i)
    {
      timer_queues_[i]->dispatch_timers();
      timer_queues_[i]->dispatch_cancellations();
    }

    // Issue any pending cancellations.
    for (size_t i = 0; i < pending_cancellations_.size(); ++i)
      cancel_ops_unlocked(pending_cancellations_[i]);
    pending_cancellations_.clear();

    // Determine whether epoll_wait should be called when the reactor next runs.
    need_epoll_wait_ = !read_op_queue_.empty()
      || !write_op_queue_.empty() || !except_op_queue_.empty();

    complete_operations_and_timers(lock);
  }

  // Run the select loop in the thread.
  void run_thread()
  {
    boost::asio::detail::mutex::scoped_lock lock(mutex_);
    while (!stop_thread_)
    {
      lock.unlock();
      run(true);
      lock.lock();
    }
  }

  // Entry point for the select loop thread.
  static void call_run_thread(epoll_reactor* reactor)
  {
    reactor->run_thread();
  }

  // Interrupt the select loop.
  void interrupt()
  {
    interrupter_.interrupt();
  }

  // The hint to pass to epoll_create to size its data structures.
  enum { epoll_size = 20000 };

  // Create the epoll file descriptor. Throws an exception if the descriptor
  // cannot be created.
  static int do_epoll_create()
  {
    int fd = epoll_create(epoll_size);
    if (fd == -1)
    {
      boost::throw_exception(
          boost::system::system_error(
            boost::system::error_code(errno,
              boost::asio::error::get_system_category()),
            "epoll"));
    }
    return fd;
  }

  // Check if all timer queues are empty.
  bool all_timer_queues_are_empty() const
  {
    for (std::size_t i = 0; i < timer_queues_.size(); ++i)
      if (!timer_queues_[i]->empty())
        return false;
    return true;
  }

  // Get the timeout value for the epoll_wait call. The timeout value is
  // returned as a number of milliseconds. A return value of -1 indicates
  // that epoll_wait should block indefinitely.
  int get_timeout()
  {
    if (all_timer_queues_are_empty())
      return -1;

    // By default we will wait no longer than 5 minutes. This will ensure that
    // any changes to the system clock are detected after no longer than this.
    boost::posix_time::time_duration minimum_wait_duration
      = boost::posix_time::minutes(5);

    for (std::size_t i = 0; i < timer_queues_.size(); ++i)
    {
      boost::posix_time::time_duration wait_duration
        = timer_queues_[i]->wait_duration();
      if (wait_duration < minimum_wait_duration)
        minimum_wait_duration = wait_duration;
    }

    if (minimum_wait_duration > boost::posix_time::time_duration())
    {
      int milliseconds = minimum_wait_duration.total_milliseconds();
      return milliseconds > 0 ? milliseconds : 1;
    }
    else
    {
      return 0;
    }
  }

  // Cancel all operations associated with the given descriptor. The do_cancel
  // function of the handler objects will be invoked. This function does not
  // acquire the epoll_reactor's mutex.
  void cancel_ops_unlocked(socket_type descriptor)
  {
    bool interrupt = read_op_queue_.cancel_operations(descriptor);
    interrupt = write_op_queue_.cancel_operations(descriptor) || interrupt;
    interrupt = except_op_queue_.cancel_operations(descriptor) || interrupt;
    if (interrupt)
      interrupter_.interrupt();
  }

  // Clean up operations and timers. We must not hold the lock since the
  // destructors may make calls back into this reactor. We make a copy of the
  // vector of timer queues since the original may be modified while the lock
  // is not held.
  void complete_operations_and_timers(
      boost::asio::detail::mutex::scoped_lock& lock)
  {
    timer_queues_for_cleanup_ = timer_queues_;
    lock.unlock();
    read_op_queue_.complete_operations();
    write_op_queue_.complete_operations();
    except_op_queue_.complete_operations();
    for (std::size_t i = 0; i < timer_queues_for_cleanup_.size(); ++i)
      timer_queues_for_cleanup_[i]->complete_timers();
  }

  // Mutex to protect access to internal data.
  boost::asio::detail::mutex mutex_;

  // The epoll file descriptor.
  int epoll_fd_;

  // Whether the epoll_wait call is currently in progress
  bool wait_in_progress_;

  // The interrupter is used to break a blocking epoll_wait call.
  select_interrupter interrupter_;

  // The queue of read operations.
  reactor_op_queue<socket_type> read_op_queue_;

  // The queue of write operations.
  reactor_op_queue<socket_type> write_op_queue_;

  // The queue of except operations.
  reactor_op_queue<socket_type> except_op_queue_;

  // The timer queues.
  std::vector<timer_queue_base*> timer_queues_;

  // A copy of the timer queues, used when cleaning up timers. The copy is
  // stored as a class data member to avoid unnecessary memory allocation.
  std::vector<timer_queue_base*> timer_queues_for_cleanup_;

  // The descriptors that are pending cancellation.
  std::vector<socket_type> pending_cancellations_;

  // Does the reactor loop thread need to stop.
  bool stop_thread_;

  // The thread that is running the reactor loop.
  boost::asio::detail::thread* thread_;

  // Whether the service has been shut down.
  bool shutdown_;

  // Whether we need to call epoll_wait the next time the reactor is run.
  bool need_epoll_wait_;
};

} // namespace detail
} // namespace asio
} // namespace boost

#endif // defined(BOOST_ASIO_HAS_EPOLL)

#include <boost/asio/detail/pop_options.hpp>

#endif // BOOST_ASIO_DETAIL_EPOLL_REACTOR_HPP