Application boost::thread stuck on mutex lock when compiled for ARM running on BeagleBone Black

Question

I am currently working on a C++ app that runs on my PC just fine and I want to make it work on a BeagleBone Black on which I have already installed a Debian Wheezy.

I'm cross compiling using ELDK v5.3 and qmake from my PC (Intel Celeron, Debian Wheezy, boost v1.49.0-3.2) to the BeagleBone Black (ARM Cortex A8, Debian Wheezy, boost v1.49.0-3.2).

Everything works like it should, but every now and then, my app freezes (is stuck on a __pthread_mutex_lock) when a boost::thread is created or the app is waiting for a boost::thread::join(). I just want to understand what's going on.

If you have some tips that can help me secure these calls to boost::thread (try/catch, error status check ...) please share them :)

Thank you !!

Here is a small source code that freezes in a similar fashion when executed on the BeagleBone Black followed by the backtrace printed when the app freezes (GDB) and the .pro file used when executing the qmake command :

#include <stdio.h>
#include <boost/thread.hpp>

#define NB_THREADS              20
#define THREAD_LIFE_DURATION    5

int g_nb_thread = 0;

boost::thread * subRegisterThread(boost::thread * pthread)
{
    printf("Register thread #%d  %p created\n",
        g_nb_thread++, pthread);

    return pthread;
}

/////////////////// Sub dummy class

class sub_dummy_class
{
public:
    boost::thread *mThread;

    sub_dummy_class();
    ~sub_dummy_class();
    void loop();
    void start();
};



sub_dummy_class::sub_dummy_class()
{
    mThread = NULL;
}

sub_dummy_class::~sub_dummy_class()
{
    if(mThread)
    {
        mThread->join();
    }
}

void sub_dummy_class::start()
{
    mThread = subRegisterThread(new boost::thread(boost::bind(&sub_dummy_class::loop, this)));
}

void sub_dummy_class::loop()
{
    int life_duration = THREAD_LIFE_DURATION;
    while(life_duration > 0)
    {
        life_duration--;
        printf("Got %d seconds to live !\n", life_duration);
        usleep(1000000);
    }
    return;
}


////////////////////////// Dummy class

class dummy_class
{
public:

    sub_dummy_class dummies[NB_THREADS];

    dummy_class();
    ~dummy_class();
    void start();


};

dummy_class::dummy_class()
{

}

dummy_class::~dummy_class()
{

}

void dummy_class::start()
{
    for(int i = 0 ; i < NB_THREADS ; i++)
    {
        dummies[i].start();
    }
}


int main(int argc, char* argv[])
{
    printf("Starting program\n");

    printf("Creating dummy class\n");
    dummy_class *DC = new dummy_class();

    printf("Starting dummy class\n");
    DC->start();

    printf("Deleting dummy class\n");
    delete DC;

    return 0;
}

Stack trace :

(gdb) bt
#0  __libc_do_syscall () at ../ports/sysdeps/unix/sysv/linux/arm/eabi/libc-do-syscall.S:44
#1  0xb6756f92 in __lll_lock_wait (futex=0x1881c, private=0) at ../ports/sysdeps/unix/sysv/linux/arm/nptl/lowlevellock.c:47
#2  0xb675357a in __pthread_mutex_lock (mutex=0x1881c) at pthread_mutex_lock.c:61
#3  0xb68029dc in pthread_mutex_lock (mutex=<optimized out>) at forward.c:182
#4  0xb6991710 in lock (this=0xb69a2c68) at ./boost/smart_ptr/detail/spinlock_pt.hpp:41
#5  scoped_lock (pv=0x1f714, this=<synthetic pointer>) at ./boost/smart_ptr/detail/spinlock_pool.hpp:65
#6  atomic_increment (pw=0x1f714) at ./boost/smart_ptr/detail/sp_counted_base_spin.hpp:41
#7  add_ref_copy (this=0x1f710) at ./boost/smart_ptr/detail/sp_counted_base_spin.hpp:90
#8  shared_count (r=..., this=<optimized out>) at ./boost/smart_ptr/detail/shared_count.hpp:316
#9  shared_ptr (this=<optimized out>) at ./boost/smart_ptr/shared_ptr.hpp:164
#10 operator= (r=..., this=0x1f604) at ./boost/smart_ptr/shared_ptr.hpp:311
#11 boost::thread::start_thread (this=0x1f5e8) at libs/thread/src/pthread/thread.cpp:185
#12 0x0000cf7c in boost::thread::thread<boost::_bi::bind_t<void, boost::_mfi::mf0<void, sub_dummy_class>, boost::_bi::list1<boost::_bi::value<sub_dummy_class*> > > > (this=0x1f5e8, f=...)
    at /opt/eldk-5.3/armv7a-hf/sysroots/armv7ahf-vfp-neon-linux-gnueabi/usr/include/boost/thread/detail/thread.hpp:205
#13 0x0000aa44 in sub_dummy_class::start (this=0x1cfdc) at main.cpp:51
#14 0x0000abc4 in dummy_class::start (this=0x1cf90) at main.cpp:96
#15 0x0000ac44 in main (argc=1, argv=0xbefff834) at main.cpp:109

EDIT : Here is the .pro file I'm using for the qmake command (qmake -spec linux-armv7a-hf-g++) :

# .pro for test_boost
TEMPLATE = app
CONFIG =

QT -= qt webkit

TARGET = test_boost_ARM
QMAKE_LFLAGS += -Wl,-rpath ./
QMAKE_LFLAGS_RPATH =

HEADERS =
SOURCES = main.cpp

OTHER_FILES =

DESTDIR = .

# external libs
LIBS += -lboost_thread

I tested these two parameters a bit. When NB_THREADS=15, it seems to go just fine every time but when it's equal to 16, it often freezes. THREAD_LIFE_DURATION doesn't seem to have any effect on the program's behaviour. I hope it will help someone understand what's going on.

Answer 1

Looks like an upgrade of boost to version 1.54 fixes the issue. Took some time to get the right packages because Debian Wheezy's boost stable release is 1.49, so I had to manually upgrade my cross-compilation platform.

The app can now run up to 380 threads without any problems, this is more than enough for me.