Confused When Boost::Asio::Io_Service Run Method Blocks/Unblocks

Confused when boost::asio::io_service run method blocks/unblocks

Foundation

Lets start with a simplified example and examine the relevant Boost.Asio pieces:

void handle_async_receive(...) { ... }
void print() { ... }

...  

boost::asio::io_service io_service;
boost::asio::ip::tcp::socket socket(io_service);

...

io_service.post(&print);                             // 1
socket.connect(endpoint);                            // 2
socket.async_receive(buffer, &handle_async_receive); // 3
io_service.post(&print);                             // 4
io_service.run();                                    // 5

What Is A Handler?

A handler is nothing more than a callback. In the example code, there are 3 handlers:

The print handler (1).
The handle_async_receive handler (3).
The print handler (4).

Even though the same print() function is used twice, each use is considered to create its own uniquely identifiable handler. Handlers can come in many shapes and sizes, ranging from basic functions like the ones above to more complex constructs such as functors generated from boost::bind() and lambdas. Regardless of the complexity, the handler still remains nothing more than a callback.

What Is Work?

Work is some processing that Boost.Asio has been requested to do on behalf of the application code. Sometimes Boost.Asio may start some of the work as soon as it has been told about it, and other times it may wait to do the work at a later point in time. Once it has finished the work, Boost.Asio will inform the application by invoking the supplied handler.

Boost.Asio guarantees that handlers will only run within a thread that is currently calling run(), run_one(), poll(), or poll_one(). These are the threads that will do work and call handlers. Therefore, in above example, print() is not invoked when it is posted into the io_service (1). Instead, it is added to the io_service and will be invoked at a later point in time. In this case, it within io_service.run() (5).

What Are Asynchronous Operations?

An asynchronous operation creates work and Boost.Asio will invoke a handler to inform the application when the work has completed. Asynchronous operations are created by calling a function that has a name with the prefix async_. These functions are also known as initiating functions.

Asynchronous operations can be decomposed into three unique steps:

Initiating, or informing, the associated io_service that works needs to be done. The async_receive operation (3) informs the io_service that it will need to asynchronously read data from the socket, then async_receive returns immediately.
Doing the actual work. In this case, when socket receives data, bytes will be read and copied into buffer. The actual work will be done in either:
- The initiating function (3), if Boost.Asio can determine that it will not block.
- When the application explicitly run the io_service (5).
Invoking the handle_async_receive ReadHandler. Once again, handlers are only invoked within threads running the io_service. Thus, regardless of when the work is done (3 or 5), it is guaranteed that handle_async_receive() will only be invoked within io_service.run() (5).

The separation in time and space between these three steps is known as control flow inversion. It is one of the complexities that makes asynchronous programming difficult. However, there are techniques that can help mitigate this, such as by using coroutines.

What Does `io_service.run()` Do?

When a thread calls io_service.run(), work and handlers will be invoked from within this thread. In the above example, io_service.run() (5) will block until either:

It has invoked and returned from both print handlers, the receive operation completes with success or failure, and its handle_async_receive handler has been invoked and returned.
The io_service is explicitly stopped via io_service::stop().
An exception is thrown from within a handler.

One potential psuedo-ish flow could be described as the following:


create io_service
create socket
add print handler to io_service (1)
wait for socket to connect (2)
add an asynchronous read work request to the io_service (3)
add print handler to io_service (4)
run the io_service (5)
  is there work or handlers?
    yes, there is 1 work and 2 handlers
      does socket have data? no, do nothing
      run print handler (1)
  is there work or handlers?
    yes, there is 1 work and 1 handler
      does socket have data? no, do nothing
      run print handler (4)
  is there work or handlers?
    yes, there is 1 work
      does socket have data? no, continue waiting
  -- socket receives data --
      socket has data, read it into buffer
      add handle_async_receive handler to io_service
  is there work or handlers?
    yes, there is 1 handler
      run handle_async_receive handler (3)
  is there work or handlers?
    no, set io_service as stopped and return

Notice how when the read finished, it added another handler to the io_service. This subtle detail is an important feature of asynchronous programming. It allows for handlers to be chained together. For instance, if handle_async_receive did not get all the data it expected, then its implementation could post another asynchronous read operation, resulting in io_service having more work, and thus not returning from io_service.run().

Do note that when the io_service has ran out of work, the application must reset() the io_service before running it again.

Example Question and Example 3a code

Now, lets examine the two pieces of code referenced in the question.

Question Code

socket->async_receive adds work to the io_service. Thus, io_service->run() will block until the read operation completes with success or error, and ClientReceiveEvent has either finished running or throws an exception.

Example 3a Code

In hopes of making it easier to understand, here is a smaller annotated Example 3a:

void CalculateFib(std::size_t n);

int main()
{
  boost::asio::io_service io_service;
  boost::optional<boost::asio::io_service::work> work =       // '. 1
      boost::in_place(boost::ref(io_service));                // .'

  boost::thread_group worker_threads;                         // -.
  for(int x = 0; x < 2; ++x)                                  //   :
  {                                                           //   '.
    worker_threads.create_thread(                             //     :- 2
      boost::bind(&boost::asio::io_service::run, &io_service) //   .'
    );                                                        //   :
  }                                                           // -'

  io_service.post(boost::bind(CalculateFib, 3));              // '.
  io_service.post(boost::bind(CalculateFib, 4));              //   :- 3
  io_service.post(boost::bind(CalculateFib, 5));              // .'

  work = boost::none;                                         // 4
  worker_threads.join_all();                                  // 5
}

At a high-level, the program will create 2 threads that will process the io_service's event loop (2). This results in a simple thread pool that will calculate Fibonacci numbers (3).

The one major difference between the Question Code and this code is that this code invokes io_service::run() (2) before actual work and handlers are added to the io_service (3). To prevent the io_service::run() from returning immediately, an io_service::work object is created (1). This object prevents the io_service from running out of work; therefore, io_service::run() will not return as a result of no work.

The overall flow is as follows:

Create and add the io_service::work object added to the io_service.
Thread pool created that invokes io_service::run(). These worker threads will not return from io_service because of the io_service::work object.
Add 3 handlers that calculate Fibonacci numbers to the io_service, and return immediately. The worker threads, not the main thread, may start running these handlers immediately.
Delete the io_service::work object.
Wait for worker threads to finish running. This will only occur once all 3 handlers have finished execution, as the io_service neither has handlers nor work.

The code could be written differently, in the same manner as the Original Code, where handlers are added to the io_service, and then the io_service event loop is processed. This removes the need to use io_service::work, and results in the following code:

int main()
{
  boost::asio::io_service io_service;

  io_service.post(boost::bind(CalculateFib, 3));              // '.
  io_service.post(boost::bind(CalculateFib, 4));              //   :- 3
  io_service.post(boost::bind(CalculateFib, 5));              // .'

  boost::thread_group worker_threads;                         // -.
  for(int x = 0; x < 2; ++x)                                  //   :
  {                                                           //   '.
    worker_threads.create_thread(                             //     :- 2
      boost::bind(&boost::asio::io_service::run, &io_service) //   .'
    );                                                        //   :
  }                                                           // -'
  worker_threads.join_all();                                  // 5
}

Synchronous vs. Asynchronous

Although the code in the question is using an asynchronous operation, it is effectively functioning synchronously, as it is waiting for the asynchronous operation to complete:

socket.async_receive(buffer, handler)
io_service.run();

is equivalent to:

boost::asio::error_code error;
std::size_t bytes_transferred = socket.receive(buffer, 0, error);
handler(error, bytes_transferred);

As a general rule of thumb, try to avoid mixing synchronous and asynchronous operations. Often times, it can turn a complex system into a complicated system. This answer highlights advantages of asynchronous programming, some of which are also covered in the Boost.Asio documentation.

boost asio io_service.run()

"until all work has finished and there are no more handlers to be dispatched, or until the io_service has been stopped"

Notice that you DO install a handler, named handle_accept, that reinstalls itself at each execution. Hence, the io_service.run will never return, at least until you quit it manually.

Basically, at the moment you run io_service.run in a thread, io_services proactor takes over program flow, using the handler's you installed. From that point on, you handle the program based on events (like the handle_accept) instead of normal procedural program flow. The loop you're mentioning is somewhere deep in the scary depths of the asio's proactor ;-).

Calling boost::asio::io_service run function from multiple threads

To restart the event handler I need to call run again but the documentation stated that restart() has to be called first.

No the documentation does not say that. You need to reset once the service had run out of work/been stopped. You did neither, so you do not need reset there.

Simply do as explained in this post Should the exception thrown by boost::asio::io_service::run() be caught? (which links to the docs)

Run boost asio io_service forever

The io_service::work object exists for this purpose.

Boost::asio::async_write, handler called only once

I do not understand the aversion to calling io_service::reset(). In this case, it is necessary to invoke prior to any subsequent calls to io_service::run():

reset() must be called prior to any second or later set of invocations of the run(), run_one(), poll() or poll_one() functions when a previous invocation of these functions returned due to the io_service being stopped or running out of work.

It is possible that a thread returns from run() as a result of an exception being thrown, yet the io_service has neither been stopped nor ran out of work. In this case, the thread can invoke run() without calling reset().

The current Client::sendCommand() is synchronous. It is an implementation detail that it initiates an asynchronous operation, then blocks in io_service::run() waiting for the operation to complete. Unless there are multiple threads invoking commands on socket, multiple threads running the io_service, or the write operation needs to be cancellable, such as from a timeout, then it would be functionally equivalent and possible easier to implement Client::sendCommand() with a synchronous write().

void Client::sendCommand(const string & p_command)
{
  boost::system::error_code ec;
  std::size_t bytes_transferred =
      boost::asio::write(socket, boost::asio::buffer(p_command), ec);
  onSendingFinished(ec, bytes_transferred);
}

If Client::sendCommand() needs to be asynchronous, then:

The io_service should be ran from outside of Client::sendCommand(). If the io_service does not always have outstanding work, then io_service::work can be used control when run() returns. See this answer for more details as to when io_service::run() blocks and unblocks.
The underlying memory provided to async_write() as the buffer (p_command) needs to remain valid until the operation's handler, Client::onSendingFinished(), has been called. In this case, it may require making a copy of p_command in Client::sendCommand(), writing the copy to the socket, then deleting the copy from within the handler.

[...] ownership of the underlying memory blocks is retained by the caller, which must guarantee that they remain valid until the handler is called.

boost::asio::io_service::run does not return while having no work

Apparently if you choose to define BOOST_ASIO_ENABLE_HANDLER_TRACKING then you must do so in all boost::asio translation units. I don't see this mentioned in the documentation, but I did find it on the Boost mailing list.

When I add

add_definitions(-DBOOST_ASIO_ENABLE_HANDLER_TRACKING)

to your CMakeLists.txt so it is applied globally then I don't see the hang.

Confused When Boost::Asio::Io_Service Run Method Blocks/Unblocks