Everything I've read in the Boost ASIO docs and here on StackOverflow suggests I can stop an async_accept operation by calling close on the acceptor socket. However, I get an intermittent not_socket error in the async_accept handler when I try to do this. Am I doing something wrong or does Boost ASIO not support this?
(Related questions: here and here.)
(Note: I'm running on Windows 7 and using the Visual Studio 2015 compiler.)
The core problem I face is a race condition between the async_accept operation accepting an incoming connection and my call to close. This happens even when using a strand, explicit or implicit.
Note my call to async_accept strictly happens before my call to close. I conclude the race condition is between my call to close and the under-the-hood code in Boost ASIO that accepts the incoming connection.
I've included code demonstrating the problem. The program repeatedly creates an acceptor, connects to it, and immediately closes the acceptor. It expects the async_accept operation to either complete successfully or else be canceled. Any other error causes the program to abort, which is what I'm seeing intermittently.
For synchronization the program uses an explicit strand. Nevertheless, the call to close is unsynchronized with the effect of the async_accept operation, so sometimes the acceptor closes before it accepts the incoming connection, sometimes it closes afterward, sometimes neither—hence the problem.
Here's the code:
#include <algorithm>
#include <boost/asio.hpp>
#include <cstdlib>
#include <future>
#include <iostream>
#include <memory>
#include <thread>
int main()
{
boost::asio::io_service ios;
auto work = std::make_unique<boost::asio::io_service::work>(ios);
const auto ios_runner = [&ios]()
{
boost::system::error_code ec;
ios.run(ec);
if (ec)
{
std::cerr << "io_service runner failed: " << ec.message() << '\n';
abort();
}
};
auto thread = std::thread{ios_runner};
const auto make_acceptor = [&ios]()
{
boost::asio::ip::tcp::resolver resolver{ios};
boost::asio::ip::tcp::resolver::query query{
"localhost",
"",
boost::asio::ip::resolver_query_base::passive |
boost::asio::ip::resolver_query_base::address_configured};
const auto itr = std::find_if(
resolver.resolve(query),
boost::asio::ip::tcp::resolver::iterator{},
[](const boost::asio::ip::tcp::endpoint& ep) { return true; });
assert(itr != boost::asio::ip::tcp::resolver::iterator{});
return boost::asio::ip::tcp::acceptor{ios, *itr};
};
for (auto i = 0; i < 1000; ++i)
{
auto acceptor = make_acceptor();
const auto saddr = acceptor.local_endpoint();
boost::asio::io_service::strand strand{ios};
boost::asio::ip::tcp::socket server_conn{ios};
// Start accepting.
std::promise<void> accept_promise;
strand.post(
[&]()
{
acceptor.async_accept(
server_conn,
strand.wrap(
[&](const boost::system::error_code& ec)
{
accept_promise.set_value();
if (ec.category() == boost::asio::error::get_system_category()
&& ec.value() == boost::asio::error::operation_aborted)
return;
if (ec)
{
std::cerr << "async_accept failed (" << i << "): " << ec.message() << '\n';
abort();
}
}));
});
// Connect to the acceptor.
std::promise<void> connect_promise;
strand.post(
[&]()
{
boost::asio::ip::tcp::socket client_conn{ios};
{
boost::system::error_code ec;
client_conn.connect(saddr, ec);
if (ec)
{
std::cerr << "connect failed: " << ec.message() << '\n';
abort();
}
connect_promise.set_value();
}
});
connect_promise.get_future().get(); // wait for connect to finish
// Close the acceptor.
std::promise<void> stop_promise;
strand.post([&acceptor, &stop_promise]()
{
acceptor.close();
stop_promise.set_value();
});
stop_promise.get_future().get(); // wait for close to finish
accept_promise.get_future().get(); // wait for async_accept to finish
}
work.reset();
thread.join();
}
Here's the output from a sample run:
async_accept failed (5): An operation was attempted on something that is not a socket
The number in parentheses denotes how many successfully iterations the program ran.
UPDATE #1: Based on Tanner Sansbury's answer, I've added a std::promise for signaling the completion of the async_accept handler. This has no effect on the behavior I'm seeing.
UPDATE #2: The not_socket error originates from a call to setsockopt, from call_setsockopt, from socket_ops::setsockopt in the file boost\asio\detail\impl\socket_ops.ipp (Boost version 1.59). Here's the full call:
socket_ops::setsockopt(new_socket, state,
SOL_SOCKET, SO_UPDATE_ACCEPT_CONTEXT,
&update_ctx_param, sizeof(SOCKET), ec);
Microsoft's documentation for setsockopt says about SO_UPDATE_ACCEPT_CONTEXT:
Updates the accepting socket with the context of the listening socket.
I'm not sure what exactly this means, but it sounds like something that fails if the listening socket is closed. This suggests that, on Windows, one cannot safely close an acceptor that is currently running a completion handler for an async_accept operation.
I hope someone can tell me I'm wrong and that there is a way to safely close a busy acceptor.
