strepsirrhini/office-gobmx
9
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Do not try to notify, if the thread is already killed

Browse source 
On a particular system, I see unit tests reliably hung at the exit time
(at a DLL unload), when it waits indefinitely in notify_all. It is the
only thread in the process left; TimerManager thread was likely killed
by that time.

This is the call stack of the hung state:

 ntdll.dll!ZwReleaseKeyedEvent()
 ntdll.dll!string "Enabling heap debug options\n"()
 msvcp140d.dll!Concurrency::details::stl_condition_variable_win7::notify_all() Line 178
 msvcp140d.dll!_Cnd_broadcast(_Cnd_internal_imp_t * cond) Line 93
 salhelper3MSC.dll!std::condition_variable::notify_all() Line 590
 salhelper3MSC.dll!salhelper::TimerManager::~TimerManager() Line 221
 salhelper3MSC.dll!``anonymous namespace'::getTimerManager'::`2'::`dynamic atexit destructor for 'aManager''()
 ucrtbased.dll!_execute_onexit_table::__l2::<lambda>() Line 206
 ucrtbased.dll!__crt_seh_guarded_call<int>::operator()<void <lambda>(void),int <lambda>(void) &,void <lambda>(void)>(__acrt_lock_and_call::__l2::void <lambda>(void) && setup, _execute_onexit_table::__l2::int <lambda>(void) & action, __acrt_lock_and_call::__l2::void <lambda>(void) && cleanup) Line 204
 ucrtbased.dll!__acrt_lock_and_call<int <lambda>(void)>(const __acrt_lock_id lock_id, _execute_onexit_table::__l2::int <lambda>(void) && action) Line 974
 ucrtbased.dll!_execute_onexit_table(_onexit_table_t * table) Line 231
 salhelper3MSC.dll!__scrt_dllmain_uninitialize_c() Line 399
 salhelper3MSC.dll!dllmain_crt_process_detach(const bool is_terminating) Line 182
 salhelper3MSC.dll!dllmain_crt_dispatch(HINSTANCE__ * const instance, const unsigned long reason, void * const reserved) Line 220
 salhelper3MSC.dll!dllmain_dispatch(HINSTANCE__ * const instance, const unsigned long reason, void * const reserved) Line 293
 salhelper3MSC.dll!_DllMainCRTStartup(HINSTANCE__ * const instance, const unsigned long reason, void * const reserved) Line 335
 ntdll.dll!LdrShutdownProcess()
 ntdll.dll!RtlExitUserProcess()
 ucrtbased.dll!exit_or_terminate_process(const unsigned int return_code) Line 144
 ucrtbased.dll!common_exit(const int return_code, const _crt_exit_cleanup_mode cleanup_mode, const _crt_exit_return_mode return_mode) Line 280
 ucrtbased.dll!exit(int return_code) Line 294
 cppunittester.exe!__scrt_common_main_seh() Line 297
 cppunittester.exe!__scrt_common_main() Line 331
 cppunittester.exe!mainCRTStartup(void * __formal) Line 17
 kernel32.dll!BaseThreadInitThunk()
 ntdll.dll!RtlUserThreadStart()

This check prevents that hang.
The behavior is *possibly* related to my commit
8a0c43fa86 (Use _beginthreadex instead of
CreateThread, 2023-08-09), which put the threads under control of CRT;
and also to the specific version of CRT on the affected system, which
might affect the order of statics destruction and thread termination.

Change-Id: I6da95ea369ac9433a426a12d62cbd2a09cb4ce4a
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/172093
Reviewed-by: Mike Kaganski <mike.kaganski@collabora.com>
Tested-by: Jenkins
This commit is contained in:
Mike Kaganski 2024-08-20 11:32:03 +05:00
parent b95059acaf
commit a27a5dda1c
1 changed files with 4 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
salhelper/source/timer.cxx
View file

@ -215,6 +215,10 @@ TimerManager::TimerManager() :
TimerManager::~TimerManager() { TimerManager::~TimerManager() {
{ {
std::scoped_lock g(m_Lock); std::scoped_lock g(m_Lock);
// Sometimes, the TimerManager thread gets killed before the static's destruction;
// in that case, notify_all could hang in unit tests
if (!isRunning())
return;
m_terminate = true; m_terminate = true;
} }
m_notEmpty.notify_all(); m_notEmpty.notify_all();