{"id":107182,"date":"2022-09-15T07:00:00","date_gmt":"2022-09-15T14:00:00","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/oldnewthing\/?p=107182"},"modified":"2022-09-15T07:14:16","modified_gmt":"2022-09-15T14:14:16","slug":"20220915-00","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/oldnewthing\/20220915-00\/?p=107182","title":{"rendered":"Serializing asynchronous operations in C++\/WinRT"},"content":{"rendered":"

Armed with some insight into C++ coroutines and lazy-start coroutines, we return to the issue of making asynchronous operations run one after the other instead of concurrently.<\/p>\n

Adapting the solution from C# comes with a frustrating complication: Unlike C# Task<\/code>s, C++\/WinRT IAsyncAction<\/code> and IAsyncOperation<\/code> support only one completion callback, so we cannot co_await<\/code> them more than once. That’s too bad, because the C# version relied on having two awaiters: One is the ultimate caller of the asynchronous method, and the other is the internal awaiter that we use for sequencing.<\/p>\n

Since the caller is going to perform a co_await<\/code>, our internal awaiter will have to use some other mechanism for sequencing. We’ll do that by hooking up the continuations manually into our own data structures.<\/p>\n

struct task_sequencer\r\n{\r\n    task_sequencer() = default;\r\n    task_sequencer(const task_sequencer&) = delete;\r\n    void operator=(const task_sequencer&) = delete;\r\n\r\nprivate:\r\n    using coro_handle = std::experimental::coroutine_handle<>;\r\n\r\n    struct suspender\r\n    {\r\n        bool await_ready() const noexcept { return false; }\r\n        void await_suspend(coro_handle h)\r\n            noexcept { handle = h; }\r\n        void await_resume() const noexcept { }\r\n\r\n        coro_handle handle;\r\n    };\r\n\r\n    static void* completed()\r\n    { return reinterpret_cast<void*>(1); }\r\n\r\n    struct chained_task\r\n    {\r\n        chained_task(void* state = nullptr) : next(state) {}\r\n\r\n        void continue_with(coro_handle h) {\r\n            if (next.exchange(h.address(),\r\n                        std::memory_order_acquire) != nullptr) {\r\n                h();\r\n            }\r\n        }\r\n\r\n        void complete() {\r\n            auto resume = next.exchange(completed());\r\n            if (resume) {\r\n                coro_handle::from_address(resume).resume();\r\n            }\r\n        }\r\n\r\n        std::atomic<void*> next;\r\n    };\r\n\r\n    struct completer\r\n    {\r\n        ~completer()\r\n        {\r\n            chain->complete();\r\n        }\r\n        std::shared_ptr<chained_task> chain;\r\n    };\r\n\r\n    winrt::slim_mutex m_mutex;\r\n    std::shared_ptr<chained_task> m_latest =\r\n        std::make_shared<chained_task>(completed());\r\n\r\npublic:\r\n    template<typename Maker>\r\n    auto QueueTaskAsync(Maker&& maker) ->decltype(maker())\r\n    {\r\n        auto current = std::make_shared<chained_task>();\r\n        auto previous = [&]\r\n        {\r\n            winrt::slim_lock_guard guard(m_mutex);\r\n            return std::exchange(m_latest, current);\r\n        }();\r\n\r\n        suspender suspend;\r\n\r\n        using Async = decltype(maker());\r\n        auto task = [](auto&& current, auto&& makerParam,\r\n                       auto&& contextParam, auto& suspend)\r\n                    -> Async\r\n        {\r\n            completer completer{ std::move(current) };\r\n            auto maker = std::move(makerParam);\r\n            auto context = std::move(contextParam);\r\n\r\n            co_await suspend;\r\n            co_await context;\r\n            co_return co_await maker();\r\n        }(current, std::forward<Maker>(maker),\r\n          winrt::apartment_context(), suspend);\r\n\r\n        previous->continue_with(suspend.handle);\r\n\r\n        return task;\r\n    }\r\n};\r\n<\/pre>\n
There’s a lot going on here, so let’s take it bit by bit.<\/p>\n
struct task_sequencer\r\n{\r\n    task_sequencer() = default;\r\n    task_sequencer(const task_sequencer&) = delete;\r\n    void operator=(const task_sequencer&) = delete;\r\n<\/pre>\n
Our task_sequencer<\/code> is default-constructible but is not copyable or assignable.<\/p>\n
Inside the class, we start with the suspender<\/code>, which  we saw last time. We use this to force the lambda coroutine (coming later) to suspend and capture the coroutine handle that lets us resume it.<\/p>\n
Next, we have the chained_task<\/code>. This class connects the coroutines that want to run in sequence.<\/p>\n
\u00a0<\/div>\n\n\n\n\n\n\n\n\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
\u2192<\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n\u2199\ufe0e<\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n\u2191
\nm_latest<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
Each coroutine has a local variable called current<\/code> which is a shared pointer to a chained_task<\/code>. That chained_task<\/code> has a member called next<\/code> which points to the coroutine to run after the current coroutine has completed. The chain ends with a chained_task<\/code> (also known as m_latest<\/code>) whose next<\/code> is null.<\/p>\n
Each chained_task<\/code> remembers the coroutine that needs to run next in its next<\/code> member. The most recently-queued one is remembered in m_latest<\/code>, and its next<\/code> is nullptr<\/code> if the coroutine is still running, or is completed<\/code> if the coroutine has completed (and is waiting for somebody to run next).<\/p>\n
We set our initial condition by initializing m_latest<\/code> to a chained_task<\/code> that has already completed. That way, the next coroutine to be queued will run immediately.<\/p>\n
First, we create a new chained_task<\/code> node and make it the m_latest<\/code>, while saving the previous value of m_latest<\/code> in previous<\/code>.<\/p>\n
    auto current = std::make_shared<chained_task>();\r\n    auto previous = [&] {\r\n        winrt::slim_lock_guard guard(m_mutex);\r\n        return std::exchange(m_latest, current);\r\n    }();\r\n<\/pre>\n
In pictures, we’ve created this:<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
\u2192<\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n\u2199\ufe0e<\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n\u2191
\nm_latest<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
Once we’ve set up the new node, we capture it into a coroutine which represents the queued task, and then use the suspender<\/code> in order to suspend the coroutine immediately and obtain its coroutine handle.<\/p>\n
    suspender suspend;\r\n\r\n    using Async = decltype(maker());\r\n    auto task = [](auto&& current, ..., auto& suspend)\r\n                -> Async\r\n    {\r\n        chained_task_completer completer{ std::move(current) };\r\n        ...\r\n\r\n        co_await suspend;\r\n        ...\r\n    }(current, ..., suspend);\r\n<\/pre>\n
This fills in another par of the diagram:<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
\u2192<\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n\u2199\ufe0e<\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n\u2191
\nm_latest<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
And then we take that coroutine handle and hook it up to the previous chained_task<\/code>:<\/p>\n
    previous->continue_with(suspend.handle);\r\n<\/pre>\n
That last step links everything together again:<\/p>\n
The linked list now looks like this:<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
\u2192<\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n\u2199\ufe0e<\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n\u2199\ufe0e<\/td>\n<\/tr>\n
 <\/td>\ncoroutine<\/code><\/td>\n <\/td>\nchained_task<\/code><\/td>\n<\/tr>\n
 <\/td>\ncurrent<\/code><\/td>\n\u2192<\/td>\nnext<\/code><\/td>\n\u2192<\/td>\nnullptr<\/code><\/td>\n<\/tr>\n
 <\/td>\n <\/td>\n <\/td>\n\u2191
\nm_latest<\/code><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
As a special case, if the previously-final chained_task<\/code> has already completed, then instead of hooking up the arrow from that chained_task<\/code> to the latest coroutine, we just resume the latest coroutine immediately.<\/p>\n
That’s how nodes get added to the list. But how are they removed?<\/p>\n
The nodes disappear from the front of the list when coroutines complete. When a coroutine completes, its completer<\/code> destructs, which destructs the shared pointer to the chained_task<\/code>. If the chained_task<\/code> is not the m_latest<\/code>, then this destroys the last shared pointer to the chained_task<\/code>, so it too destructs. As coroutines complete, the head of the linked list gets gobbled up until only m_latest<\/code> remains.<\/p>\n
Now let’s look inside the coroutine we created.<\/p>\n
    using Async = decltype(maker());\r\n    auto task = [](auto&& current, auto&& makerParam,\r\n                   auto&& contextParam, auto& suspend)\r\n                -> Async\r\n    {\r\n        completer completer{ std::move(current) };\r\n        auto maker = std::move(makerParam);\r\n        auto context = std::move(contextParam);\r\n\r\n        co_await suspend;\r\n        co_await context;\r\n        co_return co_await maker();\r\n    }(current, std::forward<Maker>(maker),\r\n      winrt::apartment_context(), suspend);\r\n<\/pre>\n
Again, there are a few things going on here.<\/p>\n
This is a captureless lambda because coroutine lambdas with captures are scary. The captures are instead passed as explicit parameters so they go into the coroutine frame.<\/p>\n
The first thing we do is move the objects out of the parameters into locals, so that they destruct as soon as the coroutine completes. We saw some time ago that  coroutine parameters do not destruct until the coroutine is destroyed, but we want to resume the next coroutine in the chain as soon as the previous one completes.<\/p>\n
We create the completer<\/code> as the first local variable so that it destructs last. That way, we resume the next coroutine only after the current one has destructed everything it had captured. We use an object with a destructor to ensure that chaining to the next coroutine occurs even if the current coroutine exits with an exception.<\/p>\n
The maker<\/code> is moved into a local variable so that it (and all of its own captures) destructs as soon as the coroutine completes, rather than lingering until the coroutine is destroyed.<\/p>\n
We also move the apartment_context<\/code> into a local variable so that we can switch back to that context once we are resumed. The previous coroutine may have completed on a different COM context, and we need to start the next one in the original context.<\/p>\n
When the coroutine completes (either normally or via an exception), the completer<\/code> destructor resume the next coroutine in the chain, if one exists. If not, it just marks itself as complete so that when the next coroutine shows up, it knows it should run immediately.<\/p>\n
The atomic operation for publishing the coroutine handle to next<\/code> uses release semantics so that all of the coroutine state generated by the current thread are made visible before we publish the coroutine handle. Conversely, the exchange operation that obtains the coroutine handle uses acquire semantics to ensure that the processor uses the published values instead of locally-cached ones.<\/p>\n
Note that if your object is single-threaded, and tasks can be queued only from a single thread, then you don’t need the m_mutex<\/code>, which also simplifies the updating of m_latest<\/code>:<\/p>\n
    auto current = std::make_shared<chained_task>();\r\n    auto previous = std::exchange(m_latest, current);\r\n<\/pre>\n
But wait, we’re not done yet. There are some exceptional conditions we’ll look at next time.<\/p>\n
\n