[Boost-users] [thread] Keeping leak detectors still...

From: Daniel Krügler (dsp_at_[hidden])
Date: 2006-07-11 04:43:22

Hello!,

Its a well-known issue, that Boost-Thread users on Windows machines have 
observed that popular leak reporting tools loudly notify a missing 
delete for the dynamically allocated mutex in tss_hooks.cpp. A good 
overview for this discussion can be found in:

http://groups.google.de/group/boost-list/browse_thread/thread/f5788746e458dd86/61b85cdd8384df3f%2361b85cdd8384df3f

While I don't want to heat up the discussion whether this is actually a 
"leak" or not, many code style guides of companies give very bad points
on such behaviour, so I would like to ask whether it is possible to 
ensure a proper clean-up of this by code (before the system does), to
keep those detector programs quiet...

Simple question of a non-expert: Couldn't one simply change the current 
impl of on_process_exit in tss_hooks.cpp to

     extern "C" BOOST_THREAD_DECL void on_process_exit(void)
     {
             { // Wrap into scope
             boost::call_once(init_threadmon_mutex, 
once_init_threadmon_mutex);
                 boost::mutex::scoped_lock lock(*threadmon_mutex);

             BOOST_ASSERT(attached_thread_count == 0);

            //Free the tls slot if one was allocated.
        
                 if (tls_key != invalid_tls_key) {
                     TlsFree(tls_key);
                    tls_key = invalid_tls_key;
                 }
         }

         delete threadmon_mutex; // Shouldn't that be the last point of
access??
         threadmon_mutex = 0; // For those who prefer a defined value...
     }

OK, I see that the docs for this function says: "Must not be omitted; 
may be called multiple times."

Question is: Does there exist at least the guarantee, that there is 
always the same number of on_process_enter and on_process_exit calls?
If yes: Wouldn't a simply reference counting scheme using one of the
atomic integer incrementing/decrementing functions (also provided by 
boost via interlocked.hpp) allow a thread-save deletion of that mutex 
object?

Thanks for your patience,

Daniel Krügler