Boost::asio - how to interrupt a blocked tcp server thread?

Question

I'm working on a multithreaded application in which one thread acts as a tcp server which receives commands from a client. The thread uses a Boost socket and acceptor to wait for a client to connect, receives a command from the client, passes the command to the rest of the application, then waits again. Here's the code:

void ServerThreadFunc()
{
    using boost::asio::ip::tcp;
    boost::asio::io_service io_service;
    tcp::acceptor acceptor(io_service, tcp::endpoint(tcp::v4(), port_no));

    for (;;)
    {
        //  listen for command connection
        tcp::socket socket(io_service);
        acceptor.accept(socket);

        //  connected; receive command
        boost::array<char,256> msg_buf;
        socket.receive(boost::asio::buffer(msg_buf));

        //  do something with received bytes here
    }
}

This thread spends most of its time blocked on the call to acceptor.accept(). At the moment, the thread only gets terminated when the application exits. Unfortunately, this causes a crash after main() returns - I believe because the thread tries to access the app's logging singleton after the singleton has been destroyed. (It was like that when I got here, honest guv.)

How can I shut this thread down cleanly when it's time for the application to exit? I've read that a blocking accept() call on a raw socket can be interrupted by closing the socket from another thread, but this doesn't appear to work on a Boost socket. I've tried converting the server logic to asynchronous i/o using the Boost asynchronous tcp echo server example, but that just seems to exchange a blocking call to acceptor::accept() for a blocking call to io_service::run(), so I'm left with the same problem: a blocked call which I can't interrupt. Any ideas?

Answer 1

In short, there are two options:

• Change code to be asynchronous (acceptor::async_accept() and async_read), run within the event loop via io_service::run(), and cancel via io_service::stop().
• Force blocking calls to interrupt with lower level mechanics, such as signals.

I would recommend the first option, as it is more likely to be the portable and easier to maintain. The important concept to understand is that the io_service::run() only blocks as long as there is pending work. When io_service::stop() is invoked, it will try to cause all threads blocked on io_service::run() to return as soon as possible; it will not interrupt synchronous operations, such as acceptor::accept() and socket::receive(), even if the synchronous operations are invoked within the event loop. It is important to note that io_service::stop() is a non-blocking call, so synchronization with threads that were blocked on io_service::run() must use another mechanic, such as thread::join().

Here is an example that will run for 10 seconds and listens to port 8080:

#include <boost/asio.hpp>
#include <boost/bind.hpp>
#include <boost/shared_ptr.hpp>
#include <boost/thread.hpp>
#include <iostream>

void StartAccept( boost::asio::ip::tcp::acceptor& );

void ServerThreadFunc( boost::asio::io_service& io_service )
{
  using boost::asio::ip::tcp;
  tcp::acceptor acceptor( io_service, tcp::endpoint( tcp::v4(), 8080 ) );

  // Add a job to start accepting connections.
  StartAccept( acceptor );

  // Process event loop.
  io_service.run();

  std::cout << "Server thread exiting." << std::endl;
}

void HandleAccept( const boost::system::error_code& error,
                   boost::shared_ptr< boost::asio::ip::tcp::socket > socket,
                   boost::asio::ip::tcp::acceptor& acceptor )
{
  // If there was an error, then do not add any more jobs to the service.
  if ( error )
  {
    std::cout << "Error accepting connection: " << error.message() 
              << std::endl;
    return;
  }

  // Otherwise, the socket is good to use.
  std::cout << "Doing things with socket..." << std::endl;

  // Perform async operations on the socket.

  // Done using the socket, so start accepting another connection.  This
  // will add a job to the service, preventing io_service::run() from
  // returning.
  std::cout << "Done using socket, ready for another connection." 
            << std::endl;
  StartAccept( acceptor );
};

void StartAccept( boost::asio::ip::tcp::acceptor& acceptor )
{
  using boost::asio::ip::tcp;
  boost::shared_ptr< tcp::socket > socket(
                                new tcp::socket( acceptor.get_io_service() ) );

  // Add an accept call to the service.  This will prevent io_service::run()
  // from returning.
  std::cout << "Waiting on connection" << std::endl;
  acceptor.async_accept( *socket,
    boost::bind( HandleAccept,
      boost::asio::placeholders::error,
      socket,
      boost::ref( acceptor ) ) );
}

int main()
{
  using boost::asio::ip::tcp;

  // Create io service.
  boost::asio::io_service io_service;

  // Create server thread that will start accepting connections.
  boost::thread server_thread( ServerThreadFunc, boost::ref( io_service ) );

  // Sleep for 10 seconds, then shutdown the server.
  std::cout << "Stopping service in 10 seconds..." << std::endl;
  boost::this_thread::sleep( boost::posix_time::seconds( 10 ) );
  std::cout << "Stopping service now!" << std::endl;

  // Stopping the io_service is a non-blocking call.  The threads that are
  // blocked on io_service::run() will try to return as soon as possible, but
  // they may still be in the middle of a handler.  Thus, perform a join on 
  // the server thread to guarantee a block occurs.
  io_service.stop();

  std::cout << "Waiting on server thread..." << std::endl;
  server_thread.join();
  std::cout << "Done waiting on server thread." << std::endl;

  return 0;
}

While running, I opened two connections. Here is the output:

Stopping service in 10 seconds...
Waiting on connection
Doing things with socket...
Done using socket, ready for another connection.
Waiting on connection
Doing things with socket...
Done using socket, ready for another connection.
Waiting on connection
Stopping service now!
Waiting on server thread...
Server thread exiting.
Done waiting on server thread.

Answer 2

When you receive an event that it's time to exit, you can call acceptor.cancel(), which will cancel the pending accept (with an error code of operation_canceled). On some systems, you might also have to close() the acceptor as well to be safe.

Answer 3

If it comes to it, you could open a temporary client connection to it on localhost - that will wake it up. You could even send it a special message so that you can shut down your server from the pub - there should be an app for that:)

Answer 4

Simply call shutdown with native handle and the SHUT_RD option, to cancel the existing receive(accept) operation.