Cancelling callbacks in Boost ASIO - multithreading

Question

I've been trying to switch my code from using one io_service per network connection to using shared ones, and I'm seeing some very odd behaviour on server sockets (client ones seem to work OK).
In order to try to work out what's going on I've re-started building up a simple example that will allow me to check my assumptions about everything that ought to happen. The first problem I've hit is that io_service::run doesn't exit when there are no handlers left, and as far as I can tell the handlers aren't removed from the work queue.
I have one thread that does an async_accept followed by an async_read. There is a separate client thread (which has its own io_service). The client thread's io_service is never run, and the server's one is run in yet another thread.
I'm using a condition variable to wait in the server thread for the read to complete (which will never happen as the client never writes). This times out just fine and then I call socket.cancel(). I would expect this to remove the read handler and run to exit as the work queue is now empty.
I do see the read handler get called (with a cancel error), but run never exits. When I tie the socket lifetime to the handler lifetime (by lambda capturing a shared_ptr to the socket) the memory isn't freed either.
The server is set up like this:
std::mutex mutex;
std::unique_lock<std::mutex> lock(mutex);
std::condition_variable signal;
boost::asio::io_service server_service;
boost::asio::ip::tcp::acceptor listener(server_service);
std::mutex read_mutex;
std::unique_lock<std::mutex> read_lock(read_mutex);
std::condition_variable read_done;
std::thread server([&]() {
std::unique_lock<std::mutex> lock(mutex);
listener.open(boost::asio::ip::tcp::v4());
listener.set_option(boost::asio::socket_base::enable_connection_aborted(true));
listener.bind(boost::asio::ip::tcp::endpoint(boost::asio::ip::tcp::v4(), 4567));
listener.listen();
std::shared_ptr<connection> server_cnx(new connection(server_service));
listener.async_accept(server_cnx->socket,
[&, server_cnx](const boost::system::error_code& error) {
log_thread() << "Server got a connection " << error << std::endl;
boost::asio::async_read_until(server_cnx->socket, server_cnx->buffer, '\n',
[&, server_cnx](const boost::system::error_code& error, std::size_t bytes) {
log_thread() << "Got " << bytes << ", " << error << std::endl;
std::unique_lock<std::mutex> lock(read_mutex);
lock.unlock();
read_done.notify_one();
});
});
lock.unlock();
signal.notify_one();
if ( read_done.wait_for(read_lock, std::chrono::seconds(1)) == std::cv_status::timeout ) {
log_thread() << "Server read timed out -- cancelling socket jobs" << std::endl;
server_cnx->socket.cancel();
server_cnx->socket.close();
} else {
log_thread() << "Server data read" << std::endl;
}
log_thread() << "Exiting server thread" << std::endl;
});
signal.wait(lock);
log_thread() << "Server set up" << std::endl;
The io_service thread is set up like this:
std::thread server_io([&]() {
log_thread() << "About to service server IO requests" << std::endl;
try {
server_service.run();
} catch ( ... ) {
log_thread() << "Exception caught" << std::endl;
}
log_thread() << "**** Service jobs all run" << std::endl;
signal.notify_one();
});
The output is as follows:
10.0002 139992957945728 Server set up
10.0005 139992957945728 Client set up
10.0006 139992848398080 About to service server IO requests
10.0006 139992848398080 Server got a connection system:0
11.0003 139992934819584 Server read timed out -- cancelling socket jobs
11.0004 139992934819584 Exiting server thread
11.0004 139992848398080 Got 0, system:125
20.0006 139992957945728 IO thread timed out servicing requests -- stopping it
^^^ This should not happen because the server service should have run out of work
20.0006 139992957945728 Waiting for things to close....
22.0008 139992957945728 Wait over, exiting
(Columns are time + 10s, thread ID, log message)
At the 11 second mark you can see that the async_read_until is called. This is the last handler in the server's io_service and yet run doesn't exit.
Even after the time out waiting for run to exit fires and the waiting thread does io_service::stop(), still run doesn't exit (there's another 2 second wait there).
The full code is on github

Answer 1

The program is invoking undefined behavior when the server thread attempts to unlock the read_lock that it does not own.
int main()
{
...
std::mutex read_mutex;
std::unique_lock<std::mutex> read_lock(read_mutex); // Acquired by main.
std::condition_variable read_done;
std::thread server([&]() { // Capture lock reference.
std::unique_lock<std::mutex> lock(mutex);
...
// The next line invokes undefined behavior as this thread does did
// not acquire read_lock.mutex().
if (read_done.wait_for(read_lock, ...)
// ^^^^^^^^^ caller does not own.
{
...
}
});
signal.wait(lock);
...
}
In particular, when invoking condition_variable::wait_for(lock), the standard requires that lock.owns_lock() is true and lock.mutex() is locked by the calling thread.
Mixing synchronous and asynchronous flows often add complexity. In this particular case, where the synchronous calls are intertwined throughout each layer using lower-level constructs for event/signal notification without a persisted state, I think that it adds unnecessary complexity and overcomplicates the flow. Furthermore, the broad scope of variables can add complexity. If read_lock had never been captured by the lambdas, then a compiler error would have occurred.
Consider the separation in space when trying to observe two events:
// I will eventually be interested when the server starts
// accepting connections, so start setting up now.
std::mutex server_mutex;
std::unique_lock<std::mutex> server_lock(server_mutex);
std::condition_variable server_started;
std::thread server([&]()
{
// I will eventually be interested when the server reads
// data, so start setting up now.
std::mutex read_mutex;
std::unique_lock<std::mutex> read_lock(read_mutex);
std::condition_variable read_done;
listener.async_accept(...,
[&](...)
{
// Got connection.
async_read_until(...,
[&](...)
{
// Someone may be interested that data has been read,
// so use the correct mutex and condition_variable
// pair.
std::unique_lock<std::mutex> read_lock(read_mutex);
read_lock.unlock();
read_done.notify_one();
});
}); // async_accept
// Someone may be interested that I am accepting connections,
// so use the correct mutex and condition_variable pair.
std::unique_lock<std::mutex> server_lock(server_mutex);
server_lock.unlock();
server_done.notify_one();
// I am now interested in if data has been read.
read_done.wait_for(read_lock);
}); // server thread
// I am now interested in if the server has started.
server_started.wait(server_lock);
The caller has to prepare to handle an event, start an operation, then wait for the event, and the operation must know the event the caller is interested in. To worsen the situation, one must now consider lock ordering to prevent deadlocks. Note how in the above example, the server thread acquires the read_mutex and then the server_mutex. Another thread cannot acquire the mutexes in a difference order without introducing the chance of a deadlock. In terms of complexity, this approach scales poorly with the number of events.
It may be worth considering re-examining the program's flow and control structure. If it can be written to be primarily asynchronous, then callback chains, continuations, or a signal-and-slot system (Boost.Signals) may uncomplicate the solution. If one prefers to have asynchronous code read as if it was synchronous, then Boost.Asio's support for coroutines can provide a clean solution. Finally, if one needs to synchronously wait on an asynchronous operation's result or timeout, then consider using Boost.Asio's support for std::future or using them directly.
// Use an asynchronous operation so that it can be cancelled on timeout.
std::future<std::size_t> on_read = boost::asio::async_read_until(
socket, buffer, '\n',boost::asio::use_future);
// If timeout occurs, then cancel the operation.
if (on_read.wait_for(std::chrono::seconds(1)) == std::future_status::timeout)
{
socket.cancel();
}
// Otherwise, the operation completed (with success or error).
else
{
// If the operation failed, then on_read.get() will throw a
// boost::system::system_error.
auto bytes_transferred = on_read.get();
}
While I would strongly advocate re-examining the overall control structure and reducing variable scope, the following sample is roughly equivalent to the above example, but may be slightly easier to maintain with its use of std::future:
// I will eventually be interested when the server starts
// accepting connections, so start setting up now.
std::promise<void> server_started_promise;
auto server_started = server_started_promise.get_future();
std::thread server([&]()
{
// I will eventually be interested when the server reads
// data, so start setting up now.
std::promise<void> read_done_promise;
auto read_done = read_done_promise.get_future();
listener.async_accept(...,
[&](...)
{
// Got connection.
async_read_until(...,
[&](...)
{
// Someone may be interested that data has been read.
read_done_promise.set_value();
});
}); // async_accept
// Someone may be interested that I am accepting connections.
server_started_promise.set_value();
// I am now interested in if data has been read.
read_done.wait_for(...);
}); // server thread
// I am now interested in if the server has started.
server_started.wait();
Here is a complete example based on the original code that demonstrates using std::future to control flow and timeout asynchronous operations in a synchronous manner:
#include <future>
#include <iostream>
#include <thread>
#include <boost/asio.hpp>
#include <boost/asio/use_future.hpp>
#include <boost/optional.hpp>
#include <boost/utility/in_place_factory.hpp>
int main()
{
using boost::asio::ip::tcp;
// Setup server thread.
boost::asio::io_service server_io_service;
std::promise<tcp::endpoint> server_promise;
auto server_future = server_promise.get_future();
// Start server thread.
std::thread server_thread(
[&server_io_service, &server_promise]
{
tcp::acceptor acceptor(server_io_service);
acceptor.open(tcp::v4());
acceptor.set_option(
boost::asio::socket_base::enable_connection_aborted(true));
acceptor.bind(tcp::endpoint(tcp::v4(), 0));
acceptor.listen();
// Handlers will not chain work, so control the io_service with a work
// object.
boost::optional<boost::asio::io_service::work> work(
boost::in_place(std::ref(server_io_service)));
// Accept a connection.
tcp::socket server_socket(server_io_service);
auto on_accept = acceptor.async_accept(server_socket,
boost::asio::use_future);
// Server has started, so notify caller.
server_promise.set_value(acceptor.local_endpoint());
// Wait for connection or error.
boost::system::system_error error =
make_error_code(boost::system::errc::success);
try
{
on_accept.get();
}
catch (const boost::system::system_error& e)
{
error = e;
}
std::cout << "Server got a connection " << error.code() << std::endl;
// Read from connection.
boost::asio::streambuf buffer;
auto on_read = boost::asio::async_read_until(
server_socket, buffer, '\n', boost::asio::use_future);
// The async_read operation is work, so destroy the work object allowing
// run() to exit.
work = boost::none;
// Timeout the async read operation.
if (on_read.wait_for(std::chrono::seconds(1)) ==
std::future_status::timeout)
{
std::cout << "Server read timed out -- cancelling socket jobs"
<< std::endl;
server_socket.close();
}
else
{
error = make_error_code(boost::system::errc::success);
std::size_t bytes_transferred = 0;
try
{
bytes_transferred = on_read.get();
}
catch (const boost::system::system_error& e)
{
error = e;
}
std::cout << "Got " << bytes_transferred << ", "
<< error.code() << std::endl;
}
std::cout << "Exiting server thread" << std::endl;
});
// Wait for server to start accepting connections.
auto server_endpoint = server_future.get();
std::cout << "Server set up" << std::endl;
// Client thread.
std::promise<void> promise;
auto future = promise.get_future();
std::thread client_thread(
[&server_endpoint, &promise]
{
boost::asio::io_service io_service;
tcp::socket client_socket(io_service);
boost::system::error_code error;
client_socket.connect(server_endpoint, error);
std::cout << "Connected " << error << std::endl;
promise.set_value();
// Keep client socket alive, allowing server to timeout.
std::this_thread::sleep_for(std::chrono::seconds(2));
std::cout << "Exiting client thread" << std::endl;
});
// Wait for client to connect.
future.get();
std::cout << "Client set up" << std::endl;
// Reset generic promise and future.
promise = std::promise<void>();
future = promise.get_future();
// Run server's io_service.
std::thread server_io_thread(
[&server_io_service, &promise]
{
std::cout << "About to service server IO requests" << std::endl;
try
{
server_io_service.run();
}
catch (const std::exception& e)
{
std::cout << "Exception caught: " << e.what() << std::endl;
}
std::cout << "Service jobs all run" << std::endl;
promise.set_value();
});
if (future.wait_for(std::chrono::seconds(3)) ==
std::future_status::timeout)
{
std::cout << "IO thread timed out servicing requests -- stopping it"
<< std::endl;
server_io_service.stop();
}
// Join all threads.
server_io_thread.join();
server_thread.join();
client_thread.join();
}

Answer 2

So of course this multi-threading is a tricky business. Turns out that in this case the read lock is acquired in the wrong place so the handler is being blocked by the thread waiting for it to complete.
I guess the lesson here is to never handle thread locks without some sort of time out.