FW: Very serious problem with Cyg_Mutex::trylock(), when eCos is built with CYGDBG_USE_ASSERTS

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

FW: Very serious problem with Cyg_Mutex::trylock(), when eCos is built with CYGDBG_USE_ASSERTS

Hans Peter Jepsen
Hi

After a few days of serious debugging, we think, that we have found, that eCos has a very serious problem with Cyg_Mutex::trylock(), when eCos is built with CYGDBG_USE_ASSERTS.

The problem has the impact, that when using cyg_mutex_trylock(), the system might die with an assert failed, although there is no reason for it.

Here is, what we have found:

When build with CYGDBG_USE_ASSERTS the Cyg_Mutex::check_this() method will check, that there is consistency between the instance variables "locked" and "owner". See code below (from file packages/kernel/.../src/sync/mutex.cxx). However, since this consistency check is not atomic, is must only be done, when thread switching is prohibited.

#ifdef CYGDBG_USE_ASSERTS

cyg_bool
Cyg_Mutex::check_this( cyg_assert_class_zeal zeal) const
{
//    CYG_REPORT_FUNCTION();

    // check that we have a non-NULL pointer first
    if( this == NULL ) return false;

    switch( zeal )
    {
    case cyg_system_test:
    case cyg_extreme:
    case cyg_thorough:
    case cyg_quick:
    case cyg_trivial:
        if(  locked && owner == NULL ) return false;
        if( !locked && owner != NULL ) return false;
    case cyg_none:
    default:
        break;
    };

    return true;
}

#endif


This checking is used in the macro call CYG_ASSERTCLASS( this, "Bad this pointer") .
In Cyg_Mutex::trylock() the checking is done, before launching Cyg_Scheduler::lock() :

cyg_bool
Cyg_Mutex::trylock(void)
{
    CYG_REPORT_FUNCTYPE("returning %d");

    cyg_bool result = true;

    CYG_ASSERTCLASS( this, "Bad this pointer");

    // Prevent preemption
    Cyg_Scheduler::lock();


We ended up using Cyg_Mutex::lock() instead, where the order of Cyg_Scheduler::lock() and CYG_ASSERTCLASS( this, "Bad this pointer") in our eyes are correct:

cyg_bool
Cyg_Mutex::lock(void)
{
    CYG_REPORT_FUNCTYPE("returning %d");

    cyg_bool result = true;
    Cyg_Thread *self = Cyg_Thread::self();

    // Prevent preemption
    Cyg_Scheduler::lock();

    CYG_ASSERTCLASS( this, "Bad this pointer");


Kind regards

Hans Peter Jepsen
Lodam Electronics, S√łnderborg, Denmark

--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss