Boost C++ Libraries Home Libraries People FAQ More

PrevUpHomeNext

Resumable C++ 20 Coroutines

[Note] Note

This is an experimental feature.

The experimental::coro class provides support for a universal C++20 coroutine. These coroutines can be used as tasks, generators and transfomers, depending on their signature.

coro<std::string_view> line_reader(tcp::socket stream)
{
   while (stream.is_open())
   {
     std::array<char, 4096> buf;

     auto read = co_await stream.async_read_some(
         boost::asio::buffer(buf), experimental::use_coro);

     if (read == 0u)
       continue;

     co_yield std::string_view { buf.data(), read };
   }
}

coro<void, std::size_t> line_logger(tcp::socket stream)
{
  std::size_t lines_read = 0u;
  auto reader = line_reader(std::move(stream));
  while (auto l = co_await reader)
  {
    std::cout << "Read: '" << *l << "'" << std::endl;
    lines_read++;
  }
  co_return lines_read;
}

void read_lines(tcp::socket sock)
{
  co_spawn(line_logger(std::move(sock),
      [](std::exception_ptr, std::size_t lines)
      {
        std::clog << "Read " << lines << " lines" << std::endl;
      }));
}

A coro is highly configurable, so that it can cover a set of different use cases.

template<
    typename Yield,
    typename Return = void,
    typename Executor = any_io_executor>
struct coro;
Yield

The Yield parameter designates how a co_yield statement behaves. It can either be a type, like int or a signature with zero or one types:

coro<void> // A coroutine with no yield
coro<int> // A coroutine that can yield int

coro<void()> // A coroutine with no yield
coro<int()> // A coroutine that can yield int

coro<int(double)> // A coroutine that can yield int and receive double

Receiving a value means that the co_yield statement returns a value.

coro<int(int)> my_sum(any_io_executor)
{
  int value = 0;
  while (true)
    value += co_yield value; //sum up all values
}

Putting values into a coroutine can be done it two ways: either by direct resumption (from another coro) or through async_resume. The first value gets ignored because the coroutines are lazy.

coro<void> c(any_io_executor exec)
{
  auto sum = my_sum(exec);
  assert(0  == co_await sum(-1));
  assert(0  == co_await sum(10));
  assert(10 == co_await sum(15));
  assert(25 == co_await sum(0));
}

awaitable<void> a()
{
  auto sum = my_sum(co_await this_coro::executor);
  assert(0  == co_await sum.async_resume(-1, use_awaitable));
  assert(0  == co_await sum.async_resume(10, use_awaitable));
  assert(10 == co_await sum.async_resume(15, use_awaitable));
  assert(25 == co_await sum.async_resume(0, use_awaitable));
}
noexcept

A coro may be noexcept:

coro<void() noexcept> c;
coro<int() noexcept> c;
coro<int(double) noexcept> c;

This will change its @c async_resume signature, from void(std::exception_ptr) to void() or void(std::exception_ptr, T) to void(T). A noexcept coro that ends with an exception will cause std::terminate to be called.

Furthermore, calls of async_resume and co_await of an expired noexcept coro will cause undefined behaviour.

Return

A coro can also define a type that can be used with co_return:

coro<void() noexcept, int> c(any_io_executor)
{
  co_return 42;
}

A coro can have both a Yield and Return that are non void at the same time.

Result

The result type of a coroutine is dermined by both Yield and Return. Note that in the follwing table only the yield output value is considered, i.e. T(U) means T.

Table 2. Result type deduction

Yield

Return

noexcept

result_type

completion_signature

T

U

false

variant<T, U>

void(std::exception_ptr, variant<T, U>)

T

U

true

variant<T, U>

void(variant<T, U>)

T

void

false

optional<T>

void(std::exception_ptr, optional<T>)

T

void

true

optional<T>

void(optional<T>)

void

void

false

optional<T>

void(std::exception_ptr)

void

void

true

optional<T>

void()

void

T

false

optional<T>

void(std::exception_ptr, T)

void

T

true

optional<T>

void(T)


Executor

Every coroutine needs to have its own executor. Since the coroutine gets called multiple times, it cannot take the executor from the caller like an awaitable. Therefore a coro requires to get an executor or an execution_context passed in as the first parameter.

coro<int> with_executor(any_io_executor);
coro<int> with_context(io_context &);

It is to note, that an execution_context is defined as loosely as possible. An execution context is any object that has a get_executor() function, which returns an executor that can be transformed into the executor_type of the coroutine. This allows most io_objects to be used as the source of the executor:

coro<int> with_socket(tcp::socket);

Additionally, a coro that is a member function will check the this pointer as well, either if it's an executor or an execution context:

struct my_io_object
{
  any_io_executor get_executor();

  coro<int> my_coro();
};

Finally, a member coro can be given an explicit executor or execution context, to override the one of the object:

struct my_io_object
{
  any_io_executor get_executor();

  coro<int> my_coro(any_io_executor exec); // it will use exec
};
co_await

The @c co_await within a coro is not the same as async_resume(use_coro), unless both coros use different executors. If they use the same, the coro will direclty suspend and resume the executor, without any usage of the executor.

co_await this_coro:: behaves the same as coroutines that use @c boost::asio::awaitable.

Integrating with awaitable

As the coro member function async_resume is an asynchronous operation, it may also be used in conjunction with awaitable coroutines in a single control flow. For example:

#include <asio.hpp>
#include <boost/asio/experimental/coro.hpp>

using boost::asio::ip::tcp;

boost::asio::experimental::coro<std::string> reader(tcp::socket& sock)
{
  std::string buf;
  while (sock.is_open())
  {
    std::size_t n = co_await boost::asio::async_read_until(
        sock, boost::asio::dynamic_buffer(buf), '\n',
        boost::asio::experimental::use_coro);
    co_yield buf.substr(0, n);
    buf.erase(0, n);
  }
}

boost::asio::awaitable<void> consumer(tcp::socket sock)
{
  auto r = reader(sock);
  auto msg1 = co_await r.async_resume(boost::asio::use_awaitable);
  std::cout << "Message 1: " << msg1.value_or("\n");
  auto msg2 = co_await r.async_resume(boost::asio::use_awaitable);
  std::cout << "Message 2: " << msg2.value_or("\n");
}

boost::asio::awaitable<void> listen(tcp::acceptor& acceptor)
{
  for (;;)
  {
    co_spawn(
        acceptor.get_executor(),
        consumer(co_await acceptor.async_accept(boost::asio::use_awaitable)),
        boost::asio::detached);
  }
}

int main()
{
  boost::asio::io_context ctx;
  tcp::acceptor acceptor(ctx, {tcp::v4(), 54321});
  co_spawn(ctx, listen(acceptor), boost::asio::detached);
  ctx.run();
}
See Also

co_spawn, experimental::coro, C++20 Coroutines, Stackful Coroutines, Stackless Coroutines.


PrevUpHomeNext