* src/rtl/filebuf.c
! unlock HVM stack before locking local mutex and calling hb_fs*()
functions which also unlocks HVM. It fixes possible deadlock condition
when hb_gcAll( .T. ) is executed.
* include/hbapi.h
* src/vm/garbage.c
! changed mark function semantic.
Adding support for user defined mark function I created race condition
in MT GC code. It happens because blocks marked as deleted were not
scanned by GC mark code so their subitems where not accessible.
To fix it we have to change mark function semantic. Now mark function
can be executed also for blocks currently deleted. It means that GC
block destructor should clean references to just removed items and
subblocks. The best place to make it is clearing pointers to GC blocks
just after hb_itemRelease() or hb_gcRefFree().
It is save to clean the reference just before hb_itemRelease() or
hb_gcRefFree() but only for the single block passed to these functions.
It is not save to clear reference to more then one block and then
execute above functions.
+ check reference count after destructor execution for all blocks
not only arrays - warning it may exploit some wrong C code.
- removed not longer used hb_gcRefCheck() function.
* include/hbvm.h
* src/vm/hvm.c
+ added new internal function hb_vmLockForce()
! fixed to mark GC blocks with active threads
* src/vm/thread.c
* include/hbthread.h
% modified sync mutexes used by xBase++ signal class emulation
to not use item array internally
* updated mutex destructor to new GC mark semantic
! mark GT items if GT is bound with thread item
! fixed mutex notify/subscribe code to not change GC items
when HVM stack is unlocked
! use hb_vmLockForce() to eliminate potential deadlock
+ added new C function hb_threadEnterCriticalSectionGC().
It should be used instead of hb_threadEnterCriticalSection()
in code which manipulates GC items inside critical section.
It's slower but eliminates possible deadlock condition.
Please remember that remember that both functions cannot
be used for the same mutex. So if it's necessary to use
hb_threadEnterCriticalSectionGC() in one place then it
has to be used in all others when the same mutex is locked.
* src/vm/arrays.c
* src/vm/codebloc.c
* src/vm/hashes.c
* src/rtl/hbgtcore.c
* contrib/hbcurl/core.c
* contrib/hbexpat/core.c
* updated destructors to new GC mark semantic
* contrib/hbxpp/dllx.c
! fixed destructor for pointer item created by DllPrepareCall()
Now it respects reference counter.
; added note about real behavior of library handle destructor,
I haven't changed existing code behavior but maybe it should
be done.
* contrib/xhb/hbserv.c
* src/debug/dbgentry.c
* src/rdd/wacore.c
! fixed possible deadlocks by using hb_threadEnterCriticalSectionGC()
instead of hb_threadEnterCriticalSection().
* contrib/hbmemio/memio.c
! slightly modified hb_memfsDirectory() code to avoid race condition
without using hb_threadEnterCriticalSectionGC()
* src/vm/fm.c
% reduce the lock range in HB_FM_STAT builds
* src/rtl/hbsocket.c
! added missing HVM stack unlocking in hb_socketSelect() function
* src/rtl/gtxwc/gtxwc.c
! added few missing locks for version compiled with HB_XWC_XLIB_NEEDLOCKS
* src/rdd/dbtotal.prg
* allow to use symbols instead of codeblocks
* modified Harbour extension which uses ordKey() as default group
signature to work also without index. In such case all records
are summarized into single one.
; Most of above modifications were critical for stability of MT programs.
They should allow to activate GC in any place.
7349602fcb