Document Number: P0055R00
Date: 2015-09-12
Project: Programming Language C++, LEWG
Revises: none
Reply to:

P0055R00: On Interactions Between Coroutines and Networking Library


Proposed Networking Library (N4478) uses the callback based asynchronous model described in N4045 which is shown to have lower overhead than the asynchronous I/O abstractions based on future.then ([4399]). The overhead of the Networking Library abstractions can be made even lower if it can take advantage of coroutines N4499. This paper suggests altering completion token transformation class templates described in N4478/[async.reqmts.async] to achieve near zero-overhead efficiency when used with coroutines. These changes do not alter the interfaces to asynchronous functions and do not change the performance characteristics of the Networking Library when used with callbacks.


Networking Library asynchronous functions uses class templates completion_handler_type_t and async_result to transform CompletionToken passed as a parameter to the interface functions starting with prefix async_ into a callable function object to be submitted to unspecified underlying implementation functions. This transformation allows to use the same set of functions whether using a callback model or relying on future based continuation mechanism. For the latter, an object of type use_future_t is provided in place of the callback parameter (ex: async_xyz(buf, len, use_future)).

template<class CompletionToken>
auto async_xyz(T1 t1, T2 t2, CompletionToken&& token)
  completion_handler_type_t<decay_t<CompletionToken>, void(R1 r1, R2 r2)>

  async_result<decltype(completion_handler)> result(completion_handler);

  async_xyz_impl(t1, t2, completion_handler); // do the work

  return result.get();

We propose to use a single completion_token_transform function to perform transformation currently done via completion_handler_type_t and async_result. Not only this results in less boilerplate code for the user/library developer to write, but also enables zero-overhead mode when working with coroutines as described in the next section.

template<class CompletionToken>
auto async_xyz(T1 t1, T2 t2, CompletionToken&& token) noexcept(auto)
  return completion_token_transform<void(R1 r1, R2 r2)>(
       [=](auto typeErasedHandler) { async_xyz_impl_raw(t1, t2, typeErasedHandler); });


Let's explore how a high level asynchronous function async_xyz can be built on top of a low level os_xyz supplied by the platform. At first, we will write both callback and coroutine based solutions separately. Then, we will show how utilizing completion_token_transform as shown in the previous section allows the same API to handle efficiently both cases.

Let ParamType be the type representing all the input parameters to an asynchronous call, ResultType be the type of the result provided asynchronously and OsContext* is a pointer to a context structure OsContext that os_xyz requires to remain valid until the asynchronous operation is complete. The general shape of the low level API is assumed to be as shown below.

using CallbackFnPtr = void(*)(OsResultType r, OsContext*); // os wants this signature
void os_associate_completion_callback(CallbackFnPtr cb); // usually per handle or per threadpool
void os_xyz(ParamType p, OsContext* o); // initiating routine (per operation)

To transform a call to async_xyz(P, CompletionHandler) into a call to os_xyz, we need to type erase the completion handler and pass it to the os_xyz as OsContext* parameter. In the completion callback, given an OsContext*, the callback will downcast it to the type containing the actual handler class and invoke it. In a simplified form it can look like:

template <typename CompletionHandler>
void async_xyz(ParamType p, CompletionHandler && cb) {
    auto o = make_unique<Handler<decay_t<CompletionHandler>>>(forward<CompletionHandler>(cb));
    os_xyz(p, o.get());

where Handler and HandlerBase defined as follows

struct HandlerBase : OsContext {
    CallbackFnPtr cb;
    explicit HandlerBase(CallbackFnPtr cb) : cb(cb) {}
    static void callback(ResultType r, OsContext* o) { // register this with OS
        static_cast<HandlerBase*>(o)->cb(r, o);

template <typename CompletionHandler>
struct Handler : HandlerBase, CompletionHandler {
    template <typename CompletionHandlerFwd>
    explicit Handler(CompletionHandlerFwd&& h)
        : CompletionHandler(forward<CompletionHandlerFwd>(h))
        , HandlerBase(&Handler::callback)
    static void callback(ResultType r, OsContext* o) {
        auto me = static_cast<Handler*>(o);
        auto handler = move(*static_cast<CompletionHandler*>(me));
        delete me;  // deleting it prior to invoke improves allocator behavior
        handler(r); // as handle is likely to request a similar block which can be immediately reused

While sophisticated implementations may utilize specialized allocation / deallocation functions to lessen the overhead of type erasure and memory allocations, the overhead cannot be eliminated completely in a callback model.

However, when asynchronous API is used in a coroutine, no type erasure or memory allocation needs to be performed at all. No only this results in less code and faster execution, it also eliminates the sole source of failure mode of async APIs allowing the library to mark async_xxx functions as noexcept.

Let's compare mapping async_xyz to an os_xyz when used in a coroutine. To be usable in an await expression (N4499/[expr.await]), async_xyz(P, use_await_t) function needs to return an object with member functions await_ready, await_suspend and await_resume defined as follows:

auto async_xyz(ParamType p, use_await_t = use_await_t{}) {
    struct Awaiter : AwaitableBase {
        ParamType p;
        explicit Awaiter(ParamType & p) : p(move(p)) {}

        bool await_ready() { return false; } // the operation has not started yet
        auto await_resume() { return move(this->result); } // unpack the result when done
        void await_suspend(coroutine_handle<> h) { // call the OS and setup completion
            this->resume = h;
            os_xyz(p, this);
    return Awaiter{ p };

where AwaitableBase defined as follows

struct AwaitableBase : HandlerBase {
    coroutine_handle<> resume;
    ResultType result;

    AwaitableBase() : HandlerBase(&AwaitableBase::Callback) {}

    static void Callback(ResultType r, OsContext* o) {
        auto me = static_cast<AwaitableBase*>(o);
        me->result = r;

The following example illustrates how a compiler transforms expression await async_xyz(p).
Note the absence of memory allocations / deallocations and type erasure of any kind.

ResultType r = await async_xyz(p);


     async_xyz`Awaiter __tmp{p}; 
     $promise.resume_addr = &__resume_label;   // save the resumption point of the coroutine
     __tmp.resume = $RBP;                      // inlined await_suspend
     os_xyz(p,&OsContextBase::Invoke, &__tmp); // inlined await_suspend
     jmp Epilogue; // suspends the coroutine
__resume_label:    // will be resumed at this point once the operation is finished
     R r = move(__tmp.result); // inlined await_resume

Now with completion_token_transform

Given the public async function async_xyz defined as described in the Overview section (and repeated below for readers convenience)

template<class CompletionToken>
auto async_xyz(T1 t1, T2 t2, CompletionToken&& token) noexcept(auto)
  return completion_token_transform<void(R1 r1, R2 r2)>(
       [=](auto typeErasedHandler) { async_xyz_impl_raw(t1, t2, typeErasedHandler); });

with the completion_token_transform defined as follows, we can achieve the same efficient implementation of asynchronous function when using callbacks:

template <typename Signature, typename CompletionHandler, typename Invoker>
void completion_token_transform(CompletionHandler && fn, Invoker invoker)
    auto p = make_unique<Handler<decay_t<CompletionHandler>>>(forward<CompletionHandler>(fn));
    p.release(); // if we reached this point, handler is owned by async activity and unique_ptr can relinquish the ownership

By defining overload for use_await_t, we can get efficient implementation of async_xyz when used in coroutines.

template <typename Signature, typename Invoker>
auto completion_token_transform(use_await_t, Invoker invoker)
    struct Awaiter : AwaiterBase, Invoker {
        bool await_ready() { return false; }
        ResultType await_resume() { return move(this->result); }
        void await_suspend(coroutine_handle<> h) {
            this->resume = h;
        Awaiter(Invoker& invoker) : Invoker(move(invoker)) {}
    return Awaiter{ invoker };

And finally, for completeness, here is how completion_token_transform overload for use_future_t will look like:

template <typename Signature, typename Invoker>
auto completion_token_transform(use_future_t, Invoker invoker) {
    struct FutHandler {
        promise<ResultType> p;
        void operator()(ResultType r) { p.set_value(move(r)); }
    auto p = make_unique<Handler<FutHandler>>(FutHandler{});
    auto f = p->p.get_future();
    return f;


Proposed changes improve efficiency of the networking library by altering the mechanism how high-level public API interprets CompletionToken when invoking unspecified internal implementation. If this direction has support, the author of this article will gladly help the author of Networking Library proposal to flesh out the relevant details and provide testing of proposed changes using coroutines available in MSVC compiler.

Future Work / Musing

There is an upcoming proposal (see [c++std-ext-17433]) to add [[nodiscard]] attribute/context-sensitive keyword to be applicable to classes and functions. If that attribute is applied to an awaiter class returned from the completion_token_transform, it will make it safe to add a default CompletionToken use_await_t to all async_xyz APIs.

template<class CompletionToken = use_await_t>
auto async_xyz(T1 t1, T2 t2, CompletionToken&& token  = use_await_t{}) noexcept(auto)
  return completion_token_transform<void(R1 r1, R2 r2)>(
       [=](auto typeErasedHandler) { async_xyz_impl_raw(t1, t2, typeErasedHandler); });

If a user accidentally writes async_xyz(t1,t2) instead of await async_xyz(t1,t2), the mistake will be caught at compile time due to nodiscard tag on the awaitable class.

Moreover, given that coroutines enable coding simplicity of synchronous functions combined with efficiency and scalability of asynchronous I/O, we may chose to use the nicest names, namely (send, receive, accept) to asynchronous functions and use CompletionToken form of the API to deal with all cases. A single API function async_xyz can be utilized for all flavors of operations. This shrinks required API surface by two thirds.

Instead of 3 forms of every API:

   void send(T1,T2);
   void send(T1,T2,error_code&);
   void async_send(T1,T2, CompletionToken);

We can use a single form

   auto send(T1,T2,CompletionToken);

To be used as follows:

   await send(t1,t2); // CompletionToken defaults to use_await_t as being the most efficient and convenient way of using the async API
   send(t1,t2,block); // synchronous version throwing an exception
   send(t1,t2,block[ec]); // synchronous version reporting an error by setting error code into ec
   send(t1,t2,[]{ completion }); // asynchronous call using callback model
   auto fut = send(t1,t2,use_future); // completion via future

Benefit of this approach extends beyond the networking library to other future standard or non-standard libraries modeling their APIs on the CompletionToken/completion_token_transform.


Great thanks to Christopher Kohlhoff whose N4045 provided the inspiration for this work.


N4045: Library Foundations for Asynchronous Operations, Revision 2
N4399: Technical Specification for C++ Extensions for Concurrency
N4478: Networking Library Proposal (Revision 5)
N4499: Draft Wording For Coroutines (Revision 2)
[c++std-ext-17433] Andrew Tomazos: Draft proposal of [[unused]], [[nodiscard]] and [[fallthrough]] attributes.