Boost C++ Libraries: Ticket #7571: Mutex fails to unlock and causes deadlock

attachment set

yalon-l@… — Thu, 25 Oct 2012 14:06:00 GMT

attachment → Main.cpp

Bug demonstrator

yalon-l@… — Thu, 25 Oct 2012 14:16:21 GMT

Results when running the attached demo file:

D:\Projects\BoostMutexBug\BoostMutexBug\Tests\Debug>Tests 5000000
Thread 1, iteration 0
Thread 2, iteration 0
Thread 1, iteration 1
Thread 1, iteration 2
Thread 1, iteration 3
Thread 1, iteration 4
Deadlock occurred

D:\Projects\BoostMutexBug\BoostMutexBug\Tests\Debug>Tests 100
Thread 1, iteration 0
Thread 1, iteration 1
Thread 2, iteration 0
Thread 1, iteration 2
Thread 2, iteration 1
Thread 2, iteration 2
Thread 2, iteration 3
Thread 1, iteration 3
Thread 1, iteration 4
Thread 2, iteration 4

status changed; resolution set

viboes — Sun, 28 Oct 2012 17:24:59 GMT

status new → closed
resolution → invalid

The problem is in your example. The timeout should depend on MAX_COUNTS. When MAX_COUNTS 5000000, it seems that 10000 milliseconds is not enough to do the loop :(

attachment set

yalon-l@… — Mon, 29 Oct 2012 08:34:07 GMT

attachment → Tests.exe

Executable of the bug demo

anonymous — Mon, 29 Oct 2012 08:44:48 GMT

I don't think that the problem is in the example. The 10000 milliseconds is about 10 times the needed duration. I'm attacing the executable so one can see the messages written and then a long pause (of about 9 sec) after which deadlock is detected. I can increase of course the 10000, but it does nothing except for a longer wait till the program detects the deadlock.

I forgot to mention that I'm running on Windows 7.

In my experiments, I replace boost::mutex with a mutex of my own, based upon Windows native CreateMutex. With it the problem doesn't show and the deadlock never occurs.

status changed; resolution deleted

yalon-l@… — Mon, 29 Oct 2012 08:49:43 GMT

status closed → reopened
resolution invalid

viboes — Mon, 29 Oct 2012 16:49:57 GMT

I don't reach to reproduce the deadlock on MacOs/Ubuntu/Windows XP.

Could you replace the timed join by join and report what happens? Please, could you try to debug this issue on your side?

Wed, 31 Oct 2012 13:38:49 GMT

Replacing the timed join by join didn't change a thing, except that the program now hangs forever instead of announcing that it had detected a deadlock... I'm surprised that you didn't manage to reproduce the bug on Windows XP. Did you try building it or using the attached exe? Running the exe on XP reproduced the bug for me. The same applies to building and running it on XP.

If you build it, pay attention to use Visual Studio 2003. I gave the code to a colleague with Visual Studio 2005 and with that compiler that problem didn't show.

I can debug the problem on my side. What information would be useful for you?

status, severity changed; resolution set

viboes — Wed, 31 Oct 2012 21:11:12 GMT

status reopened → closed
resolution → wontfix
severity Showstopper → Problem

I have added the test to the regression test. For the time been no compiler is signaling a failure.

VS 2003 is too old to maintain it now, in 2012.

For the debug, I'm not waiting for any specific information other than, Hey, I have found what is happening? or better yet, And I've found a workaround ;-)

No seriously, I don't have the possibility and the time to debug this issue, but any patch is welcome.