Επεξεργασία

Κοινή χρήση μέσω


Advanced concurrency and asynchrony with C++/WinRT

This topic describes advanced scenarios with concurrency and asynchrony in C++/WinRT.

For an introduction to this subject, first read Concurrency and asynchronous operations.

Offloading work onto the Windows thread pool

A coroutine is a function like any other in that a caller is blocked until a function returns execution to it. And, the first opportunity for a coroutine to return is the first co_await, co_return, or co_yield.

So, before you do compute-bound work in a coroutine, you need to return execution to the caller (in other words, introduce a suspension point) so that the caller isn't blocked. If you're not already doing that by co_await-ing some other operation, then you can co_await the winrt::resume_background function. That returns control to the caller, and then immediately resumes execution on a thread pool thread.

The thread pool being used in the implementation is the low-level Windows thread pool, so it's optimally efficient.

IAsyncOperation<uint32_t> DoWorkOnThreadPoolAsync()
{
    co_await winrt::resume_background(); // Return control; resume on thread pool.

    uint32_t result;
    for (uint32_t y = 0; y < height; ++y)
    for (uint32_t x = 0; x < width; ++x)
    {
        // Do compute-bound work here.
    }
    co_return result;
}

Programming with thread affinity in mind

This scenario expands on the previous one. You offload some work onto the thread pool, but then you want to display progress in the user interface (UI).

IAsyncAction DoWorkAsync(TextBlock textblock)
{
    co_await winrt::resume_background();
    // Do compute-bound work here.

    textblock.Text(L"Done!"); // Error: TextBlock has thread affinity.
}

The code above throws a winrt::hresult_wrong_thread exception, because a TextBlock must be updated from the thread that created it, which is the UI thread. One solution is to capture the thread context within which our coroutine was originally called. To do that, instantiate a winrt::apartment_context object, do background work, and then co_await the apartment_context to switch back to the calling context.

IAsyncAction DoWorkAsync(TextBlock textblock)
{
    winrt::apartment_context ui_thread; // Capture calling context.

    co_await winrt::resume_background();
    // Do compute-bound work here.

    co_await ui_thread; // Switch back to calling context.

    textblock.Text(L"Done!"); // Ok if we really were called from the UI thread.
}

As long as the coroutine above is called from the UI thread that created the TextBlock, then this technique works. There will be many cases in your app where you're certain of that.

For a more general solution to updating UI, which covers cases where you're uncertain about the calling thread, you can co_await the winrt::resume_foreground function to switch to a specific foreground thread. In the code example below, we specify the foreground thread by passing the dispatcher object associated with the TextBlock (by accessing its Dispatcher property). The implementation of winrt::resume_foreground calls CoreDispatcher.RunAsync on that dispatcher object to execute the work that comes after it in the coroutine.

IAsyncAction DoWorkAsync(TextBlock textblock)
{
    co_await winrt::resume_background();
    // Do compute-bound work here.

    // Switch to the foreground thread associated with textblock.
    co_await winrt::resume_foreground(textblock.Dispatcher());

    textblock.Text(L"Done!"); // Guaranteed to work.
}

The winrt::resume_foreground function takes an optional priority parameter. If you're using that parameter, then the pattern shown above is appropriate. If not, then you can choose to simplify co_await winrt::resume_foreground(someDispatcherObject); into just co_await someDispatcherObject;.

Execution contexts, resuming, and switching in a coroutine

Broadly speaking, after a suspension point in a coroutine, the original thread of execution may go away and resumption may occur on any thread (in other words, any thread may call the Completed method for the async operation).

But if you co_await any of the four Windows Runtime asynchronous operation types (IAsyncXxx), then C++/WinRT captures the calling context at the point you co_await. And it ensures that you're still on that context when the continuation resumes. C++/WinRT does this by checking whether you're already on the calling context and, if not, switching to it. If you were on a single-threaded apartment (STA) thread before co_await, then you'll be on the same one afterward; if you were on a multi-threaded apartment (MTA) thread before co_await, then you'll be on one afterward.

IAsyncAction ProcessFeedAsync()
{
    Uri rssFeedUri{ L"https://blogs.windows.com/feed" };
    SyndicationClient syndicationClient;

    // The thread context at this point is captured...
    SyndicationFeed syndicationFeed{ co_await syndicationClient.RetrieveFeedAsync(rssFeedUri) };
    // ...and is restored at this point.
}

The reason you can rely on this behavior is because C++/WinRT provides code to adapt those Windows Runtime asynchronous operation types to the C++ coroutine language support (these pieces of code are called wait adapters). The remaining awaitable types in C++/WinRT are simply thread pool wrappers and/or helpers; so they complete on the thread pool.

using namespace std::chrono_literals;
IAsyncOperation<int> return_123_after_5s()
{
    // No matter what the thread context is at this point...
    co_await 5s;
    // ...we're on the thread pool at this point.
    co_return 123;
}

If you co_await some other type—even within a C++/WinRT coroutine implementation—then another library provides the adapters, and you'll need to understand what those adapters do in terms of resumption and contexts.

To keep context switches down to a minimum, you can use some of the techniques that we've already seen in this topic. Let's see some illustrations of doing that. In this next pseudo-code example, we show the outline of an event handler that calls a Windows Runtime API to load an image, drops onto a background thread to process that image, and then returns to the UI thread to display the image in the UI.

IAsyncAction MainPage::ClickHandler(IInspectable /* sender */, RoutedEventArgs /* args */)
{
    // We begin in the UI context.

    // Call StorageFile::OpenAsync to load an image file.

    // The call to OpenAsync occurred on a background thread, but C++/WinRT has restored us to the UI thread by this point.

    co_await winrt::resume_background();

    // We're now on a background thread.

    // Process the image.

    co_await winrt::resume_foreground(this->Dispatcher());

    // We're back on MainPage's UI thread.

    // Display the image in the UI.
}

For this scenario, there's a little bit of inefficiency around the call to StorageFile::OpenAsync. There's a necessary context switch to a background thread (so that the handler can return execution to the caller), on resumption after which C++/WinRT restores the UI thread context. But, in this case, it's not necessary to be on the UI thread until we're about to update the UI. The more Windows Runtime APIs we call before our call to winrt::resume_background, the more unnecessary back-and-forth context switches we incur. The solution is not to call any Windows Runtime APIs before then. Move them all after the winrt::resume_background.

IAsyncAction MainPage::ClickHandler(IInspectable /* sender */, RoutedEventArgs /* args */)
{
    // We begin in the UI context.

    co_await winrt::resume_background();

    // We're now on a background thread.

    // Call StorageFile::OpenAsync to load an image file.

    // Process the image.

    co_await winrt::resume_foreground(this->Dispatcher());

    // We're back on MainPage's UI thread.

    // Display the image in the UI.
}

If you want to do something more advanced, then you could write your own await adapters. For example, if you want a co_await to resume on the same thread that the async action completes on (so, there's no context switch), then you could begin by writing await adapters similar to the ones shown below.

Note

The code example below is provided for educational purposes only; it's to get you started understanding how await adapters work. If you want to use this technique in your own codebase, then we recommend that you develop and test your own await adapter struct(s). For example, you could write complete_on_any, complete_on_current, and complete_on(dispatcher). Also consider making them templates that take the IAsyncXxx type as a template parameter.

struct no_switch
{
    no_switch(Windows::Foundation::IAsyncAction const& async) : m_async(async)
    {
    }

    bool await_ready() const
    {
        return m_async.Status() == Windows::Foundation::AsyncStatus::Completed;
    }

    void await_suspend(std::experimental::coroutine_handle<> handle) const
    {
        m_async.Completed([handle](Windows::Foundation::IAsyncAction const& /* asyncInfo */, Windows::Foundation::AsyncStatus const& /* asyncStatus */)
        {
            handle();
        });
    }

    auto await_resume() const
    {
        return m_async.GetResults();
    }

private:
    Windows::Foundation::IAsyncAction const& m_async;
};

To understand how to use the no_switch await adapters, you'll first need to know that when the C++ compiler encounters a co_await expression it looks for functions called await_ready, await_suspend, and await_resume. The C++/WinRT library provides those functions so that you get reasonable behavior by default, like this.

IAsyncAction async{ ProcessFeedAsync() };
co_await async;

To use the no_switch await adapters, just change the type of that co_await expression from IAsyncXxx to no_switch, like this.

IAsyncAction async{ ProcessFeedAsync() };
co_await static_cast<no_switch>(async);

Then, instead of looking for the three await_xxx functions that match IAsyncXxx, the C++ compiler looks for functions that match no_switch.

A deeper dive into winrt::resume_foreground

As of C++/WinRT 2.0, the winrt::resume_foreground function suspends even if it's called from the dispatcher thread (in previous versions, it could introduce deadlocks in some scenarios because it only suspended if not already on the dispatcher thread).

The current behavior means that you can rely on stack unwinding and re-queuing taking place; and that's important for system stability, especially in low-level systems code. The last code listing in the section Programming with thread affinity in mind, above, illustrates performing some complex calculation on a background thread, and then switching to the appropriate UI thread in order to update the user interface (UI).

Here's how winrt::resume_foreground looks internally.

auto resume_foreground(...) noexcept
{
    struct awaitable
    {
        bool await_ready() const
        {
            return false; // Queue without waiting.
            // return m_dispatcher.HasThreadAccess(); // The C++/WinRT 1.0 implementation.
        }
        void await_resume() const {}
        void await_suspend(coroutine_handle<> handle) const { ... }
    };
    return awaitable{ ... };
};

This current, versus previous, behavior is analogous to the difference between PostMessage and SendMessage in Win32 application development. PostMessage queues the work and then unwinds the stack without waiting for the work to complete. The stack-unwinding can be essential.

The winrt::resume_foreground function also initially only supported the CoreDispatcher (tied to a CoreWindow), which was introduced prior to Windows 10. We've since introduced a more flexible and efficient dispatcher: the DispatcherQueue. You can create a DispatcherQueue for your own purposes. Consider this simple console application.

using namespace Windows::System;

winrt::fire_and_forget RunAsync(DispatcherQueue queue);
 
int main()
{
    auto controller{ DispatcherQueueController::CreateOnDedicatedThread() };
    RunAsync(controller.DispatcherQueue());
    getchar();
}

The example above creates a queue (contained within a controller) on a private thread, and then passes the controller to the coroutine. The coroutine can use the queue to await (suspend and resume) on the private thread. Another common use of DispatcherQueue is to create a queue on the current UI thread for a traditional desktop or Win32 app.

DispatcherQueueController CreateDispatcherQueueController()
{
    DispatcherQueueOptions options
    {
        sizeof(DispatcherQueueOptions),
        DQTYPE_THREAD_CURRENT,
        DQTAT_COM_STA
    };
 
    ABI::Windows::System::IDispatcherQueueController* ptr{};
    winrt::check_hresult(CreateDispatcherQueueController(options, &ptr));
    return { ptr, take_ownership_from_abi };
}

This illustrates how you can call and incorporate Win32 functions into your C++/WinRT projects, by simply calling the Win32-style CreateDispatcherQueueController function to create the controller, and then transfer ownership of the resulting queue controller to the caller as a WinRT object. This is also precisely how you can support efficient and seamless queuing on your existing Petzold-style Win32 desktop application.

winrt::fire_and_forget RunAsync(DispatcherQueue queue);
 
int main()
{
    Window window;
    auto controller{ CreateDispatcherQueueController() };
    RunAsync(controller.DispatcherQueue());
    MSG message;
 
    while (GetMessage(&message, nullptr, 0, 0))
    {
        DispatchMessage(&message);
    }
}

Above, the simple main function begins by creating a window. You can imagine that this registers a window class, and calls CreateWindow to create the top-level desktop window. CreateDispatcherQueueController function is then called to create the queue controller before calling some coroutine with the dispatcher queue owned by this controller. A traditional message pump is then entered where resumption of the coroutine naturally occurs on this thread. Having done that, you can return to the elegant world of coroutines for your async or message-based workflow within your application.

winrt::fire_and_forget RunAsync(DispatcherQueue queue)
{
    ... // Begin on the calling thread...
 
    co_await winrt::resume_foreground(queue);
 
    ... // ...resume on the dispatcher thread.
}

The call to winrt::resume_foreground will always queue, and then unwind the stack. You can also optionally set the resumption priority.

winrt::fire_and_forget RunAsync(DispatcherQueue queue)
{
    ...
 
    co_await winrt::resume_foreground(queue, DispatcherQueuePriority::High);
 
    ...
}

Or, using the default queuing order.

...
#include <winrt/Windows.System.h>
using namespace Windows::System;
...
winrt::fire_and_forget RunAsync(DispatcherQueue queue)
{
    ...
 
    co_await queue;
 
    ...
}

Note

As shown above, be sure to include the projection header for the namespace of the type you're co_await-ing. For example, Windows::UI::Core::CoreDispatcher, Windows::System::DispatcherQueue, or Microsoft::UI::Dispatching::DispatcherQueue.

Or, in this case detecting queue shutdown, and handling it gracefully.

winrt::fire_and_forget RunAsync(DispatcherQueue queue)
{
    ...
 
    if (co_await queue)
    {
        ... // Resume on dispatcher thread.
    }
    else
    {
        ... // Still on calling thread.
    }
}

The co_await expression returns true, indicating that resumption will occur on the dispatcher thread. In other words, that queuing was successful. Conversely, it returns false to indicate that execution remains on the calling thread because the queue's controller is shutting down and is no longer serving queue requests.

So, you have a great deal of power at your fingertips when you combine C++/WinRT with coroutines; and especially when doing some old-school Petzold-style desktop application development.

Canceling an asynchronous operation, and cancellation callbacks

The Windows Runtime's features for asynchronous programming allow you to cancel an in-flight asynchronous action or operation. Here's an example that calls StorageFolder::GetFilesAsync to retrieve a potentially large collection of files, and it stores the resulting asynchronous operation object in a data member. The user has the option to cancel the operation.

// MainPage.xaml
...
<Button x:Name="workButton" Click="OnWork">Work</Button>
<Button x:Name="cancelButton" Click="OnCancel">Cancel</Button>
...

// MainPage.h
...
#include <winrt/Windows.Foundation.h>
#include <winrt/Windows.Foundation.Collections.h>
#include <winrt/Windows.Storage.Search.h>

using namespace winrt;
using namespace Windows::Foundation;
using namespace Windows::Foundation::Collections;
using namespace Windows::Storage;
using namespace Windows::Storage::Search;
using namespace Windows::UI::Xaml;
...
struct MainPage : MainPageT<MainPage>
{
    MainPage()
    {
        InitializeComponent();
    }

    IAsyncAction OnWork(IInspectable /* sender */, RoutedEventArgs /* args */)
    {
        workButton().Content(winrt::box_value(L"Working..."));

        // Enable the Pictures Library capability in the app manifest file.
        StorageFolder picturesLibrary{ KnownFolders::PicturesLibrary() };

        m_async = picturesLibrary.GetFilesAsync(CommonFileQuery::OrderByDate, 0, 1000);

        IVectorView<StorageFile> filesInFolder{ co_await m_async };

        workButton().Content(box_value(L"Done!"));

        // Process the files in some way.
    }

    void OnCancel(IInspectable const& /* sender */, RoutedEventArgs const& /* args */)
    {
        if (m_async.Status() != AsyncStatus::Completed)
        {
            m_async.Cancel();
            workButton().Content(winrt::box_value(L"Canceled"));
        }
    }

private:
    IAsyncOperation<::IVectorView<StorageFile>> m_async;
};
...

For the implementation side of cancellation, let's begin with a simple example.

// main.cpp
#include <iostream>
#include <winrt/Windows.Foundation.h>

using namespace winrt;
using namespace Windows::Foundation;
using namespace std::chrono_literals;

IAsyncAction ImplicitCancelationAsync()
{
    while (true)
    {
        std::cout << "ImplicitCancelationAsync: do some work for 1 second" << std::endl;
        co_await 1s;
    }
}

IAsyncAction MainCoroutineAsync()
{
    auto implicit_cancelation{ ImplicitCancelationAsync() };
    co_await 3s;
    implicit_cancelation.Cancel();
}

int main()
{
    winrt::init_apartment();
    MainCoroutineAsync().get();
}

If you run the example above, then you'll see ImplicitCancelationAsync print one message per second for three seconds, after which time it automatically terminates as a result of being canceled. This works because, on encountering a co_await expression, a coroutine checks whether it has been canceled. If it has, then it short-circuits out; and if it hasn't, then it suspends as normal.

Cancellation can, of course, happen while the coroutine is suspended. Only when the coroutine resumes, or hits another co_await, will it check for cancellation. The issue is one of potentially too-coarse-grained latency in responding to cancellation.

So, another option is to explicitly poll for cancellation from within your coroutine. Update the example above with the code in the listing below. In this new example, ExplicitCancelationAsync retrieves the object returned by the winrt::get_cancellation_token function, and uses it to periodically check whether the coroutine has been canceled. As long as it's not canceled, the coroutine loops indefinitely; once it is canceled, the loop and the function exit normally. The outcome is the same as the previous example, but here exiting happens explicitly, and under control.

IAsyncAction ExplicitCancelationAsync()
{
    auto cancelation_token{ co_await winrt::get_cancellation_token() };

    while (!cancelation_token())
    {
        std::cout << "ExplicitCancelationAsync: do some work for 1 second" << std::endl;
        co_await 1s;
    }
}

IAsyncAction MainCoroutineAsync()
{
    auto explicit_cancelation{ ExplicitCancelationAsync() };
    co_await 3s;
    explicit_cancelation.Cancel();
}
...

Waiting on winrt::get_cancellation_token retrieves a cancellation token with knowledge of the IAsyncAction that the coroutine is producing on your behalf. You can use the function call operator on that token to query the cancellation state—essentially polling for cancellation. If you're performing some compute-bound operation, or iterating through a large collection, then this is a reasonable technique.

Register a cancellation callback

The Windows Runtime's cancellation doesn't automatically flow to other asynchronous objects. But—introduced in version 10.0.17763.0 (Windows 10, version 1809) of the Windows SDK—you can register a cancellation callback. This is a pre-emptive hook by which cancellation can be propagated, and makes it possible to integrate with existing concurrency libraries.

In this next code example, NestedCoroutineAsync does the work, but it has no special cancellation logic in it. CancelationPropagatorAsync is essentially a wrapper on the nested coroutine; the wrapper forwards cancellation pre-emptively.

// main.cpp
#include <iostream>
#include <winrt/Windows.Foundation.h>

using namespace winrt;
using namespace Windows::Foundation;
using namespace std::chrono_literals;

IAsyncAction NestedCoroutineAsync()
{
    while (true)
    {
        std::cout << "NestedCoroutineAsync: do some work for 1 second" << std::endl;
        co_await 1s;
    }
}

IAsyncAction CancelationPropagatorAsync()
{
    auto cancelation_token{ co_await winrt::get_cancellation_token() };
    auto nested_coroutine{ NestedCoroutineAsync() };

    cancelation_token.callback([=]
    {
        nested_coroutine.Cancel();
    });

    co_await nested_coroutine;
}

IAsyncAction MainCoroutineAsync()
{
    auto cancelation_propagator{ CancelationPropagatorAsync() };
    co_await 3s;
    cancelation_propagator.Cancel();
}

int main()
{
    winrt::init_apartment();
    MainCoroutineAsync().get();
}

CancelationPropagatorAsync registers a lambda function for its own cancellation callback, and then it awaits (it suspends) until the nested work completes. When or if CancellationPropagatorAsync is canceled, it propagates the cancellation to the nested coroutine. There's no need to poll for cancellation; nor is cancellation blocked indefinitely. This mechanism is flexible enough for you to use it to interop with a coroutine or concurrency library that knows nothing of C++/WinRT.

Reporting progress

If your coroutine returns either IAsyncActionWithProgress, or IAsyncOperationWithProgress, then you can retrieve the object returned by the winrt::get_progress_token function, and use it to report progress back to a progress handler. Here's a code example.

// main.cpp
#include <iostream>
#include <winrt/Windows.Foundation.h>

using namespace winrt;
using namespace Windows::Foundation;
using namespace std::chrono_literals;

IAsyncOperationWithProgress<double, double> CalcPiTo5DPs()
{
    auto progress{ co_await winrt::get_progress_token() };

    co_await 1s;
    double pi_so_far{ 3.1 };
    progress.set_result(pi_so_far);
    progress(0.2);

    co_await 1s;
    pi_so_far += 4.e-2;
    progress.set_result(pi_so_far);
    progress(0.4);

    co_await 1s;
    pi_so_far += 1.e-3;
    progress.set_result(pi_so_far);
    progress(0.6);

    co_await 1s;
    pi_so_far += 5.e-4;
    progress.set_result(pi_so_far);
    progress(0.8);

    co_await 1s;
    pi_so_far += 9.e-5;
    progress.set_result(pi_so_far);
    progress(1.0);

    co_return pi_so_far;
}

IAsyncAction DoMath()
{
    auto async_op_with_progress{ CalcPiTo5DPs() };
    async_op_with_progress.Progress([](auto const& sender, double progress)
    {
        std::wcout << L"CalcPiTo5DPs() reports progress: " << progress << L". "
                   << L"Value so far: " << sender.GetResults() << std::endl;
    });
    double pi{ co_await async_op_with_progress };
    std::wcout << L"CalcPiTo5DPs() is complete !" << std::endl;
    std::wcout << L"Pi is approx.: " << pi << std::endl;
}

int main()
{
    winrt::init_apartment();
    DoMath().get();
}

To report progress, invoke the progress token with the progress value as the argument. To set a provisional result, use the set_result() method on the progress token.

Note

Reporting provisional results requires C++/WinRT version 2.0.210309.3 or later.

The above example chooses to set a provisional result for every progress report. You can choose to report provisional results any time, if at all. It need not be coupled with a progress report.

Note

It's not correct to implement more than one completion handler for an asynchronous action or operation. You can have either a single delegate for its completed event, or you can co_await it. If you have both, then the second will fail. Either one of the following two kinds of completion handlers is appropriate; not both for the same async object.

auto async_op_with_progress{ CalcPiTo5DPs() };
async_op_with_progress.Completed([](auto const& sender, AsyncStatus /* status */)
{
    double pi{ sender.GetResults() };
});
auto async_op_with_progress{ CalcPiTo5DPs() };
double pi{ co_await async_op_with_progress };

For more info about completion handlers, see Delegate types for asynchronous actions and operations.

Fire and forget

Sometimes, you have a task that can be done concurrently with other work, and you don't need to wait for that task to complete (no other work depends on it), nor do you need it to return a value. In that case, you can fire off the task and forget it. You can do that by writing a coroutine whose return type is winrt::fire_and_forget (instead of one of the Windows Runtime asynchronous operation types, or concurrency::task).

// main.cpp
#include <winrt/Windows.Foundation.h>

using namespace winrt;
using namespace std::chrono_literals;

winrt::fire_and_forget CompleteInFiveSeconds()
{
    co_await 5s;
}

int main()
{
    winrt::init_apartment();
    CompleteInFiveSeconds();
    // Do other work here.
}

winrt::fire_and_forget is also useful as the return type of your event handler when you need to perform asynchronous operations in it. Here's an example (also see Strong and weak references in C++/WinRT).

winrt::fire_and_forget MyClass::MyMediaBinder_OnBinding(MediaBinder const&, MediaBindingEventArgs args)
{
    auto lifetime{ get_strong() }; // Prevent *this* from prematurely being destructed.
    auto ensure_completion{ unique_deferral(args.GetDeferral()) }; // Take a deferral, and ensure that we complete it.

    auto file{ co_await StorageFile::GetFileFromApplicationUriAsync(Uri(L"ms-appx:///video_file.mp4")) };
    args.SetStorageFile(file);

    // The destructor of unique_deferral completes the deferral here.
}

The first argument (the sender) is left unnamed, because we never use it. For that reason we're safe to leave it as a reference. But observe that args is passed by value. See the Parameter-passing section above.

Awaiting a kernel handle

C++/WinRT provides a winrt::resume_on_signal function, which you can use to suspend until a kernel event is signaled. You're responsible for ensuring that the handle remains valid until your co_await resume_on_signal(h) returns. resume_on_signal itself can't do that for you, because you may have lost the handle even before the resume_on_signal starts, as in this first example.

IAsyncAction Async(HANDLE event)
{
    co_await DoWorkAsync();
    co_await resume_on_signal(event); // The incoming handle is not valid here.
}

The incoming HANDLE is valid only until the function returns, and this function (which is a coroutine) returns at the first suspension point (the first co_await in this case). While awaiting DoWorkAsync, control has returned to the caller, the calling frame has gone out of scope, and you no longer know whether the handle will be valid when your coroutine resumes.

Technically, our coroutine is receiving its parameters by value, as it should (see Parameter-passing above). But in this case we need to go a step further so that we're following the spirit of that guidance (rather than just the letter). We need to pass a strong reference (in other words, ownership) along with the handle. Here's how.

IAsyncAction Async(winrt::handle event)
{
    co_await DoWorkAsync();
    co_await resume_on_signal(event); // The incoming handle *is* valid here.
}

Passing a winrt::handle by value provides ownership semantics, which ensures that the kernel handle remains valid for the lifetime of the coroutine.

Here's how you might call that coroutine.

namespace
{
    winrt::handle duplicate(winrt::handle const& other, DWORD access)
    {
        winrt::handle result;
        if (other)
        {
            winrt::check_bool(::DuplicateHandle(::GetCurrentProcess(),
		        other.get(), ::GetCurrentProcess(), result.put(), access, FALSE, 0));
        }
        return result;
    }

    winrt::handle make_manual_reset_event(bool initialState = false)
    {
        winrt::handle event{ ::CreateEvent(nullptr, true, initialState, nullptr) };
        winrt::check_bool(static_cast<bool>(event));
        return event;
    }
}

IAsyncAction SampleCaller()
{
    handle event{ make_manual_reset_event() };
    auto async{ Async(duplicate(event)) };

    ::SetEvent(event.get());
    event.close(); // Our handle is closed, but Async still has a valid handle.

    co_await async; // Will wake up when *event* is signaled.
}

You can pass a timeout value to resume_on_signal, as in this example.

winrt::handle event = ...

if (co_await winrt::resume_on_signal(event.get(), std::literals::2s))
{
    puts("signaled");
}
else
{
    puts("timed out");
}

Asynchronous timeouts made easy

C++/WinRT is invested heavily in C++ coroutines. Their effect on writing concurrency code is transformational. This section discusses cases where details of asynchrony are not important, and all you want is the result there and then. For that reason, C++/WinRT's implementation of the IAsyncAction Windows Runtime asynchronous operation interface has a get function, similar to that provided by std::future.

using namespace winrt::Windows::Foundation;
int main()
{
    IAsyncAction async = ...
    async.get();
    puts("Done!");
}

The get function blocks indefinitely, while the async object completes. Async objects tend to be very short-lived, so this is often all you need.

But there are cases where that's not sufficient, and you need to abandon the wait after some time has elapsed. Writing that code has always been possible, thanks to the building blocks provided by the Windows Runtime. But now C++/WinRT makes it a lot easier by providing the wait_for function. It's also implemented on IAsyncAction, and again it's similar to that provided by std::future.

using namespace std::chrono_literals;
int main()
{
    IAsyncAction async = ...
 
    if (async.wait_for(5s) == AsyncStatus::Completed)
    {
        puts("done");
    }
}

Note

wait_for uses std::chrono::duration at the interface, but it is limited to some range smaller than what std::chrono::duration provides (roughly 49.7 days).

The wait_for in this next example waits for around five seconds and then it checks completion. If the comparison is favorable, then you know that the async object completed successfully, and you're done. If you're waiting for some result, then you can simply follow that with a call to the GetResults method to retrieve the result.

Note

wait_for and get are mutually exclusive (you can't call both of them). They each count as a waiter, and Windows Runtime asynchronous actions/operations support only a single waiter.

int main()
{
    IAsyncOperation<int> async = ...
 
    if (async.wait_for(5s) == AsyncStatus::Completed)
    {
        printf("result %d\n", async.GetResults());
    }
}

Because the async object has completed by then, the GetResults method returns the result immediately, without any further wait. As you can see, wait_for returns the state of the async object. So, you can use it for more fine-grained control, like this.

switch (async.wait_for(5s))
{
case AsyncStatus::Completed:
    printf("result %d\n", async.GetResults());
    break;
case AsyncStatus::Canceled:
    puts("canceled");
    break;
case AsyncStatus::Error:
    puts("failed");
    break;
case AsyncStatus::Started:
    puts("still running");
    break;
}
  • Remember that AsyncStatus::Completed means that the async object completed successfully, and you may call the GetResults method to retrieve any result.
  • AsyncStatus::Canceled means that the async object was canceled. A cancellation is typically requested by the caller, so it would be rare to handle this state. Typically, a canceled async object is simply discarded. You can call the GetResults method to rethrow the cancellation exception if you wish.
  • AsyncStatus::Error means that the async object has failed in some way. You can call the GetResults method to rethrow the exception if you wish.
  • AsyncStatus::Started means that the async object is still running. The Windows Runtime async pattern doesn't allow multiple waits, nor waiters. That means that you can't call wait_for in a loop. If the wait has effectively timed-out, then you're left with a few choices. You can abandon the object, or you can poll its status before calling the GetResults method to retrieve any result. But it's best just to discard the object at this point.

An alternative pattern is to check only for Started, and let GetResults deal with the other cases.

if (async.wait_for(5s) == AsyncStatus::Started)
{
    puts("timed out");
}
else
{
    // will throw appropriate exception if in canceled or error state
    auto results = async.GetResults();
}

Returning an array asynchronously

Below is an example of MIDL 3.0 that produces error MIDL2025: [msg]syntax error [context]: expecting > or, near "[".

Windows.Foundation.IAsyncOperation<Int32[]> RetrieveArrayAsync();

The reason is that it's invalid to use an array as a parameter type argument to a parameterized interface. So we need a less obvious way to achieve the aim of asynchronously passing an array back from a runtime class method.

You can return the array boxed into a PropertyValue object. The calling code then unboxes it. Here's a code example, which you can try out by adding the SampleComponent runtime class to a Windows Runtime Component (C++/WinRT) project, and then consuming that from (for example) a Core App (C++/WinRT) project.

// SampleComponent.idl
namespace MyComponentProject
{
    runtimeclass SampleComponent
    {
        Windows.Foundation.IAsyncOperation<IInspectable> RetrieveCollectionAsync();
    };
}

// SampleComponent.h
...
struct SampleComponent : SampleComponentT<SampleComponent>
{
    ...
    Windows::Foundation::IAsyncOperation<Windows::Foundation::IInspectable> RetrieveCollectionAsync()
    {
        co_return Windows::Foundation::PropertyValue::CreateInt32Array({ 99, 101 }); // Box an array into a PropertyValue.
    }
}
...

// SampleCoreApp.cpp
...
MyComponentProject::SampleComponent m_sample_component;
...
auto boxed_array{ co_await m_sample_component.RetrieveCollectionAsync() };
auto property_value{ boxed_array.as<winrt::Windows::Foundation::IPropertyValue>() };
winrt::com_array<int32_t> my_array;
property_value.GetInt32Array(my_array); // Unbox back into an array.
...

Important APIs