Question

I have programmed a boost::thread application, where I might have some race conditions. I want to debug this program. Therefore I used the following valgrind tools:

1. halgrind
2. drd

unfortunately they have a very false positive rate. So with the really simple program below valgrind --tool=drd complains about 94 errors, where no should be. So with my complex program I get about 15000 errors. So it is really hard to find the actual error.

I could reproduce this behavior with the following boost libraries 1.46.0 and 1.47.0. And with valgrind 3.7.0 SVN and valgrind 3.8.0 SVN. The operating systems I tried it out where Ubuntu 11.10 and Mac OS X 10.7. The compiler where gcc 4.2.1 and gcc 4.6.1.

#include <iostream>
#include <boost/thread.hpp>

void run()
{
    //do some stuff here
}

int main(int argc, char* argv[])
{
    boost::thread thread(run);
    thread.join();
    std::cerr << "main: done" << std::endl;
    return 0;
}
;

How do you debug your boost thread programs? Are there other tools which might be better suited?

Solution

It seems that valgrind after version 3.6.1 is broken. If I use valgrind 3.6.1 everything works fine.

Here on error report from valgrind --tool=drd:

==60767== Thread 1:
==60767== Conflicting store by thread 1 at 0x100026ec0 size 8
==60767==    at 0x2A316E: pthread_mutex_lock (in /usr/lib/system/libsystem_c.dylib)
==60767==    by 0x2A82FA: _pthread_cond_wait (in /usr/lib/system/libsystem_c.dylib)
==60767==    by 0x32A4E: boost::condition_variable::wait(boost::unique_lock<boost::mutex>&) (in /usr/local/lib/libboost_thread.dylib)
==60767==    by 0x2BE5A: boost::thread::join() (in /usr/local/lib/libboost_thread.dylib)
==60767==    by 0x10000195C: main (in ./playgroudThreads)
==60767== Address 0x100026ec0 is at offset 144 from 0x100026e30. Allocation context:
==60767==    at 0xC5B3: malloc (vg_replace_malloc.c:266)
==60767==    by 0x9968D: operator new(unsigned long) (in /usr/lib/libstdc++.6.0.9.dylib)
==60767==    by 0x1000069ED: boost::detail::thread_data<void (*)()>* boost::detail::heap_new_impl<boost::detail::thread_data<void (*)()>, void (*&)()>(void (*&)()) (in ./playgroudThreads)
==60767==    by 0x100006A87: boost::detail::thread_data<void (*)()>* boost::detail::heap_new<boost::detail::thread_data<void (*)()>, void (*)()>(void (*&)()) (in ./playgroudThreads)
==60767==    by 0x100006ACA: boost::shared_ptr<boost::detail::thread_data_base> boost::thread::make_thread_info<void (*)()>(void (*)()) (in ./playgroudThreads)
==60767==    by 0x100006B08: boost::thread::thread<void (*)()>(void (*)(), boost::disable_if<boost::is_convertible<void (*&)(), boost::detail::thread_move_t<void (*)()> >, boost::thread::dummy*>::type) (in ./playgroudThreads)
==60767==    by 0x100001950: main (in ./playgroudThreads)
==60767== Other segment start (thread 2)
==60767==    at 0x2A7B68: thread_start (in /usr/lib/system/libsystem_c.dylib)
==60767== Other segment end (thread 2)
==60767==    at 0x3E667A: mach_msg_trap (in /usr/lib/system/libsystem_kernel.dylib)
==60767==    by 0x3DED38: semaphore_create (in /usr/lib/system/libsystem_kernel.dylib)
==60767==    by 0x2A50F7: new_sem_from_pool (in /usr/lib/system/libsystem_c.dylib)
==60767==    by 0x2A6199: _pthread_exit (in /usr/lib/system/libsystem_c.dylib)
==60767==    by 0x2A48C9: _pthread_start (in /usr/lib/system/libsystem_c.dylib)
==60767==    by 0x2A7B74: thread_start (in /usr/lib/system/libsystem_c.dylib)

Answer 1

From DRD manual

An important advantage of happens-before data race detectors is that these do not report any false positives.
DRD is based on the happens-before algorithm.

So DRD does not have any false positive if the implementation uses just POSIX threads and not non-POSIX stuff like Linux’s futex.

Therefore, if DRD reports a false positive, on a race-free program like yours, it has a bug. It is not supposed to have false positives.

However, I could not reproduce the errors you reported, on my machine (Archlinux/ gcc 4.6.2 / valgrind 3.6.1 / boost 1.47).