Tcl Source Code

Version: 8.5.11 (original and current core-8-5-branch)
Platform: Windows (linux not yet tested)

After exiting of threads, they leave a memory leak. 
Consequence of that: we cannot use techniques such as 'idletime' (to release some system resources, for example large tcl codes are no more used, etc, etc.).

After run of following script I have lost 100MB of memory. In fact, nothing will be done here, except threads are created and exited.

Test script:

puts [thread::names]
time {
  thread::create {::thread::release}
} 500
puts [thread::names]

That solved it.  Thanks!

Merging now...

*nix segfault cleared: we should reset a thread key after freeing of alloc cache (in tclUnixThrd.c);
Fixed within [1dc6ea5ab2];

@Don: could you please test it on CentOS and OSX.

Same troubles on Mavericks OSX.

Just try to run the test suite on a

./configure --disable-shared --enable-threads --enable-symbols

build.

Same troubles on Mavericks OSX.

Just try to run the test suite on a

./configure --disable-shared --enable-threads --enable-symbols

build.

I can't test CentOS right now, but it work on all my linux systems, for ex.: debian / kernel 3.13.0-23-generic. I think, possible in tclUnixThrd.c TclpFreeAllocCache may miss a something like "pthread_setspecific(key, NULL);" after call of "TclFreeAllocCache(ptr);" (in comparison to tclWinThrd.c). But [c251d61424] alone, without this, works fine on my linux mashines.

void
TclpFreeAllocCache(
    void *ptr)
{
    if (ptr != NULL) {
	/*
	 * Called by the pthread lib when a thread exits
	 */

	TclFreeAllocCache(ptr);
	pthread_setspecific(key, NULL);

    } else if (initialized) {
	/*
	 * Called by us in TclFinalizeThreadAlloc() during the library
	 * finalization initiated from Tcl_Finalize()
	 */

	pthread_key_delete(key);
	initialized = 0;
    }
}

Reviews and tests are welcome

Leak found and fixed within [c251d61424]
TclpFreeAllocCache (if called with ptr != NULL) has a comment: "Called by us in TclpFinalizeThreadData when a thread exits ..."
But until now it was really never called.
After fixing, no byte is lost during test.

much short variant of leak test :

puts [thread::names]
time {
  thread::create {::thread::release}
} 500
puts [thread::names]

It seems to be a problem with tcl core (threaded memory subsystem etc.) - I made a test without thread.dll, where I create threads with WinAPI, call inits and hereafter finalize - leak remains.

Following singlethreaded script runs without leak also:

time {
  try {
    set ntp [interp create]
    $ntp eval "load thread27"
  } finally {
    interp delete $ntp
  }
} 500

tested with thread2.6.5, thread2.6.7 and current thread-2-7-branch