2014-03-26 14:10 UTC+0100 Przemyslaw Czerpak (druzus/at/poczta.onet.pl)

* src/vm/garbage.c
    ! added protection for activating GC from thread with unlocked HVM stack
      when other thread is already executing GC passes. Such situation does
      not happen in core code and the protections is only just in case.

  * src/vm/hvm.c
    ! do not release HVM thread STOP signal when GC mark pass is activated.
      It was very serious bug which could cause memory corruption when
      thread unlocked HVM stack what allow to start GC mark pass by other
      thread and then locked HVM and begin to change complex variables
      still scanned by the GC thread. This problem have existed from the
      beginning when I committed initial version of Harbour MT support.
      I do not remember why I added code to remove the STOP signal - probably
      I was experimenting with some new things working on Harbour MT code
      and I forgot to deactivate it.
      NOTE: 3-rd party developers who use GC collectible pointer items
            must not activate any PRG code execution inside their mark
            functions.

ChangeLog.txt

@@ -10,6 +10,26 @@
      * Change, ! Fix, % Optimization, + Addition, - Removal, ; Comment
  */
 
+2014-03-26 14:10 UTC+0100 Przemyslaw Czerpak (druzus/at/poczta.onet.pl)
+  * src/vm/garbage.c
+    ! added protection for activating GC from thread with unlocked HVM stack
+      when other thread is already executing GC passes. Such situation does
+      not happen in core code and the protections is only just in case.
+
+  * src/vm/hvm.c
+    ! do not release HVM thread STOP signal when GC mark pass is activated.
+      It was very serious bug which could cause memory corruption when
+      thread unlocked HVM stack what allow to start GC mark pass by other
+      thread and then locked HVM and begin to change complex variables
+      still scanned by the GC thread. This problem have existed from the
+      beginning when I committed initial version of Harbour MT support.
+      I do not remember why I added code to remove the STOP signal - probably
+      I was experimenting with some new things working on Harbour MT code
+      and I forgot to deactivate it.
+      NOTE: 3-rd party developers who use GC collectible pointer items
+            must not activate any PRG code execution inside their mark
+            functions.
+
 2014-03-25 00:38 UTC+0100 Przemyslaw Czerpak (druzus/at/poczta.onet.pl)
   * tests/iotcp.prg
     ! fixed typo in macro name - thanks to Itamar for the info

src/vm/garbage.c

@@ -598,7 +598,7 @@ void hb_gcCollectAll( HB_BOOL fForce )
    {
       PHB_GARBAGE pAlloc, pDelete;
 
-      if( ! s_pCurrBlock )
+      if( ! s_pCurrBlock || s_bCollecting )
       {
          hb_vmResumeThreads();
          return;

src/vm/hvm.c

@@ -533,7 +533,6 @@ HB_BOOL hb_vmSuspendThreads( HB_BOOL fWait )
       {
          if( s_iRunningCount <= 0 )
          {
-            hb_vmThreadRequest &= ~HB_THREQUEST_STOP;
             ++s_iRunningCount;
             return HB_TRUE;
          }