Deadlock in win_iocp_io_context::sutdown() when owned by thread_pool

[glebov-andrey]

Hi.

There appears to be a bug when shutting down a win_iocp_io_context that is owned by a thread_pool, causing a deadlock.

Here is the reproducer (the std::optionals are for explicit destruction and are totally optional):

#include <optional>
#include <asio/steady_timer.hpp>
#include <asio/thread_pool.hpp>

int main() {
    auto thread_pool = std::optional<asio::thread_pool>{std::in_place, 1};
    auto timer = std::optional<asio::steady_timer>{std::in_place, *thread_pool}; // creates a win_iocp_io_context
    thread_pool->join();
    timer.reset();
    thread_pool.reset(); // DEADLOCK!
}

When win_iocp_io_context::shutdown() calls thread_->join()

asio/asio/include/asio/detail/impl/win_iocp_io_context.ipp

Line 134 in 62c4488

thread_->join();

it deadlocks because the thread is blocked in win_iocp_io_context::do_one on a call to GetQueuedCompletionStatus

asio/asio/include/asio/detail/impl/win_iocp_io_context.ipp

Lines 419 to 421 in 62c4488

	BOOL ok = ::GetQueuedCompletionStatus(iocp_.handle,
	&bytes_transferred, &completion_key, &overlapped,
	msec < gqcs_timeout_ ? msec : gqcs_timeout_);

The only way to cause GetQueuedCompletionStatus to return from the interface seems to be win_iocp_io_context::stop(). If I call it directly then that fixes the deadlock:

// ...
asio::use_service<asio::detail::win_iocp_io_context>(*thread_pool).stop();
thread_pool.reset(); // FINE

The same bug arises when creating anything else that uses a win_iocp_io_context, e.g. a socket.
Maybe, there should be a call to stop() (or something equivalent) before joining the thread in shutdown()?

The is reproducible in at least 1.13.0 and 1.14.1 (0a52abc).
The OS is Windows 10 1903, compiled with VS 2019 and clang-cl 9.0.

[helmesjo]

This is sort of a showstopper. Anyone looking into it?

[ihsandemir]

I am having the same problem:

Reproducer:

// register a timer
// cancel the timer
pool.stop();
pool.join();
// the pool does out of scope and destructed

The destructor hangs at:

clientTest_STATIC_32.exe!boost::asio::detail::win_thread::join() Line 46
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_thread.ipp(46)
clientTest_STATIC_32.exe!boost::asio::detail::win_iocp_io_context::shutdown() Line 136
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_iocp_io_context.ipp(136)
clientTest_STATIC_32.exe!boost::asio::detail::service_registry::shutdown_services() Line 44
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\service_registry.ipp(44)
clientTest_STATIC_32.exe!boost::asio::execution_context::shutdown() Line 42
	at C:\Boost\include\boost-1_72\boost\asio\impl\execution_context.ipp(42)
clientTest_STATIC_32.exe!boost::asio::execution_context::~execution_context() Line 35
	at C:\Boost\include\boost-1_72\boost\asio\impl\execution_context.ipp(35)
clientTest_STATIC_32.exe!boost::asio::thread_pool::~thread_pool() Line 61
	at C:\Boost\include\boost-1_72\boost\asio\impl\thread_pool.ipp(61)

The worker thread:
clientTest_STATIC_32.exe!boost::asio::detail::win_iocp_io_context::do_one(unsigned long msec, boost::system::error_code & ec) Line 420
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_iocp_io_context.ipp(420)
clientTest_STATIC_32.exe!boost::asio::detail::win_iocp_io_context::run(boost::system::error_code & ec) Line 202
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_iocp_io_context.ipp(202)
clientTest_STATIC_32.exe!boost::asio::detail::win_iocp_io_context::thread_function::operator()() Line 48
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_iocp_io_context.ipp(48)
clientTest_STATIC_32.exe!boost::asio::detail::win_thread::func<boost::asio::detail::win_iocp_io_context::thread_function>::run() Line 123
	at C:\Boost\include\boost-1_72\boost\asio\detail\win_thread.hpp(123)
clientTest_STATIC_32.exe!boost::asio::detail::win_thread_function(void * arg) Line 119
	at C:\Boost\include\boost-1_72\boost\asio\detail\impl\win_thread.ipp(119)

[chriskohlhoff]

Addressed in asio version 1.16.1.