boost::asio::windows::random_access_handle is seriously broken

[Niall Douglas]

This year a new library called Boost.AFIO extending Boost.ASIO with async file i/o was given funding for Google Summer of Code. This library has found some serious problems with Boost.ASIO's Windows IOCP boost::asio::windows::random_access_handle implementation:

1. It appears to limit buffer sizes to 64Kb which corrupts output. This may be appropriate for non-random access e.g. network use (see ​http://comments.gmane.org/gmane.comp.lib.boost.asio.user/3676), but it is a showstopper for file i/o. The windows::random_access_handle should have no such limitation.

2. It appears to corrupt output when many buffers are passed, even when each of these is under 64Kb.

3. It also doesn't appear to like doing scatter/gather file i/o from multiple threads.

You will note that the POSIX (Linux) implementation doesn't have any of these problems (see ​https://travis-ci.org/BoostGSoC/boost.afio/builds/7882496). This is something wrong with the Windows IOCP implementation in Boost.ASIO.

To illustrate the problem, I have created a special branch for this bug report at ​https://github.com/ned14/TripleGit/tree/boost_asio_bug1. If you examine ​https://github.com/ned14/TripleGit/blob/boost_asio_bug1/unittests/main.cpp#L11 you'll see that I've created a macro BOOST_ASIO_BUG_WORKAROUND which when defined gets the torture test to pass. If you run through main.cpp looking for BOOST_ASIO_BUG_WORKAROUND you'll see the changes I had to make to get things working - these are a good indicator of what's going wrong in Boost.ASIO I think.

To run just the failing unit test which exhibits the problem, use "x64/Debug/tests async_io/torture". You'll need the experimental VS2012 Nov 2012 CTP compiler supporting variadic templates to compile the code successfully in Visual Studio 2012.

My thanks in advance for your time.

Niall

[Niall Douglas]

Sorry I was unclear with problem no 2: I suspect Boost.ASIO isn't incrementing in an atomic fashion (or at all) the Offset member of the OVERLAPPED structure during multi-buffer reads and writes. With sockets and pipes of course this would never have been necessary.

Niall

[Niall Douglas]

Bug #8935 may be related.

[chris_kohlhoff]

In r85750, added a note to clarify that programs must not issue concurrent async_write_at operations for overlapping regions. Included in boost 1.55.