How come my custom exception message is lost when it is thrown from a IAsyncAction^?

A customer implemented an IAsyncAction^ using the Parallel Patterns Library (PPL). They had the action throw an exception with a custom message, but found that the custom message was lost when they tried to catch it:

IAsyncAction^ DoSomething()
{
    return concurrency::create_async([] {
        throw ref new Platform::Exception(E_FAIL, "Sorry!");
    });
}

// consumer
concurrency::create_task(DoSomething())
.then([](concurrency::task<void> precedingTask) {
    try {
        precedingTask.get();
    } catch (Platform::Exception^ ex) {
        printf("0x%08x %ls\n",
            ex->HResult,
            ex->Message->Data());
    }
});

This code successfully catches the exception, and the HResult is preserved, but the custom message is lost.

What’s going on?

Recall that at the ABI layer, the only way to report an error from an IAsyncAction is through an HRESULT. You can retrieve it by reading the ErrorCode property, or you can experience it live by calling GetResults() method and receiving the error as the failure code.

Now, there is a side channel mechanism for providing additional information: The Ro­Originate­Error function lets you attach a message to a failure, which is stored in the thread context, and some libraries like C++/WinRT sets and retrieves this context when it generates and reconstructs a hresult_error object.

But let’s see what PPL does when a C++/CX exception occurs:

    // ppltasks.h
    template<typename _Function>
    ref class _AsyncTaskGeneratorThunk sealed :
        _AsyncProgressBase<typename _AsyncLambdaTypeTraits<_Function>::_AsyncAttributes,
        _AsyncLambdaTypeTraits<_Function>::_AsyncAttributes::_TakesProgress>
    {
        ⟦…⟧

        virtual void _OnStart() override
        {
            _M_task = _AsyncAttributesTaskGenerator::
                _Generate_Task<_Function, _AsyncTaskGeneratorThunk<_Function>^,
                    _Attributes>(_M_func, this, _M_cts, _M_creationCallstack);
            _M_task.then([=](task<typename _Attributes::_ReturnType> _Antecedent) {
                try
                {
                    _Antecedent.get();
                }
                catch (task_canceled&)
                {
                    this->_TryTransitionToCancelled();
                }
                catch (::Platform::Exception^ _Ex)            
                {                                             
                    this->_TryTransitionToError(_Ex->HResult);
                }                                             
                catch (...)
                {
                    this->_TryTransitionToError(E_FAIL);
                }
                this->_FireCompletion();
            });
        }
        ⟦…⟧
    };

Observe that when a C++/CX exception is thrown from the lambda, the exception’s HResult is passed to _Try­Transition­To­Error, but all the other details are ignored. The PPL library doesn’t call Ro­Originate­Error, or Get­Error­Info, so it doesn’t use the side channel for conveying additional failure information. All that survives is the HRESULT.

Now, if Do­Something is a function internal to your project, then you can change it to return a concurrency::task<void>. The PPL library preserves exceptions thrown from tasks, and rethrows them when you call .get().

IAsyncAction^ DoSomething()
{
    return concurrency::create_task([] {
        throw ref new Platform::Exception(E_FAIL, "Sorry!");
    });
}

// consumer
DoSomething()
.then([](concurrency::task<void> precedingTask) {
    try {
        precedingTask.get();
    } catch (Platform::Exception^ ex) {
        printf("0x%08x %ls\n",
            ex->HResult,
            ex->Message->Data());
    }
});