Opened 10 years ago

Closed 10 years ago

Last modified 9 years ago

#7552 closed Bugs (fixed)

win_iocp_io_service hangs due to a race condition

Reported by: danielg@… Owned by: chris_kohlhoff
Milestone: To Be Determined Component: asio
Version: Boost 1.49.0 Severity: Regression
Keywords: asio windows hang Cc:

Description

The fix to #6321 introduced a bug that can cause the io_service to hang.

Example code (hangs eventually):

int main()
{
	for (;;)
	{
		boost::asio::io_service ios;
		ios.post([]{ std::cout << '.'; });

		boost::thread_group threads;
		for (size_t i = 0; i < 2; ++i)
			threads.create_thread([&]{ ios.run(); });
		threads.join_all();
	}
}

Suppose that one of the threads read the posted operation, completed it and got to the work_finished function where it decremented outstanding_work to 0 and entered the stop function. Now the second thread starts. In the run function it sees that outstanding_work is 0 so it sets stopped to 1 and returns. Back in the first thread PostQueuedCompletionStatus(iocp_.handle, 0, 0, 0) isn't called since stopped is already 1. do_one returns 1 and is called again. GetQueuedCompletionStatus returns WAIT_TIMEOUT because the operation queue is empty, which causes the loop to continue to the next iteration ad infinitum.

Attachments (1)

asiohang.patch (590 bytes ) - added by danielg@… 10 years ago.
Proposed fix for this bug.

Download all attachments as: .zip

Change History (4)

by danielg@…, 10 years ago

Attachment: asiohang.patch added

Proposed fix for this bug.

comment:1 by chris_kohlhoff, 10 years ago

Resolution: fixed
Status: newclosed

(In [82290]) Merge from trunk:

  • Fix some 64-to-32-bit conversion warnings. Fixes #7459
  • Fix typos in comments. Fixes #7761
  • Fix error in example embedded in basic_socket::get_option's documentation. Fixes #7562
  • Use long rather than int for SSL_CTX options, to match OpenSSL. Fixes #7209
  • Use _snwprintf to address a compile error due to the changed swprintf signature in recent versions of MinGW. Fixes #7373
  • Fix deadlock that can occur on Windows when shutting down a pool of io_service threads due to running out of work. Fixes #7552
  • Enable noexcept qualifier for error categories. Fixes #7797
  • Treat errors from accept as non-fatal. Fixes #7488
  • Add a small block recycling optimisation.
  • Version bump.
  • Regenerate documentation.

comment:2 by anonymous, 9 years ago

Is this fixed in boost 1.53.0 ?

I've noticed there are changes in win_iocp_io_service.ipp but not in the area pointed in the proposed patch, namely inside do_one() to prevent the infinite loop.

comment:3 by chris_kohlhoff, 9 years ago

Note: See TracTickets for help on using tickets.