It seems that the crash occurs due to a stack overflow in startWaiter lambda function. Coroutines awaiting on the async_auto_reset_event are sometimes resumed inside a call to set(), which is a documented behavior. There are multiple coroutines created by startWaiter, so the chain of nested continuations may become too long.

cppcoro/test/async_auto_reset_event_tests.cpp

Lines 100 to 106 in b9c28b1

    
           auto startWaiter = [&]() -> cppcoro::task<> 
        
           { 
        
           	co_await tp.schedule(); 
        
           	co_await event; 
        
           	++value; 
        
           	event.set(); 
        
           };

With a slightly modified startWaiter it is possible to detect nested continuations.

(see code)

	std::atomic<int> depth = 0;

	auto startWaiter = [&]() -> cppcoro::task<>
	{
		co_await tp.schedule();
		co_await event;
		++value;
		int d = depth++;
		assert(d < 3); // if depth is greater than number of threads in the pool,
		               // this is executed as a nested continuation
		event.set();
		depth -= 1;
	};

Interestingly, when there is only one thread in the thread pool, the test completes successfully.

	auto startWaiter = [&]() -> cppcoro::task<>
	{
	co_await tp.schedule();
	co_await event;
	++value;
	event.set();
	};

On MSVC 2017, segfaults occur in test 37 #25

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions