After b9757f5cfd "loplugin:useuniqueptr in
vcl/svdata" ASan/UBSan builds started to fail (like
<https://ci.libreoffice.org//job/lo_ubsan/1025/>) at the end of
PythonTest_dbaccess_python (and probably other PythonTests), when during exit
the static utl::ConfigManager instance already happens to be destroyed by the
time the static ImplSVData's mpSettingsConfigItem is destroyed (which would
normally be cleared during DeInitVCL, if PythonTests would call that, and which
in the past had thus simply been leaked in PythonTests when that
mpSettingsConfigItem was a plain pointer instead of std::unique_ptr).
So ensure that PythonTests that initialize VCL also call DeInitVCL, via a new
private_deinitTestEnvironment, complementing the existing
private_initTestEnvironment.
However, while private_initTestEnvironment is called once (typically via
UnoInProcess.setUp, which internally makes sure to only call it once) as soon as
the first executed test needs it, private_deinitTestEnvironment must be called
once after the lasts test needing it has executed. The only way that I found to
do that is to override unittest.TextTestResult's stopTestRun method, which is
called once after all tests have been executed. Hence a new test runner setup
in unotest/source/python/org/libreoffice/unittest.py that is now called from
solenv/gbuild/PythonTest.mk.
That revealed a few places in PythonTests that didn't yet close/delete documents
that they had opened, which has now been added.
One remaining problem then is that classes like SwXTextDocument and friends call
Application::GetSolarMutex from their dtors, via sw::UnoImplPtrDeleter (a "Smart
pointer class ensuring that the pointed object is deleted with a locked
SolarMutex", sw/inc/unobaseclass.hxx). That means that any PyUNO proxies to
such C++ objects that remain alive after private_deinitTestEnvironment will
cause issues at exit, when Python does a final garbage collection of those
objects. The ultimate fix will be to remove that unhelpful UnoImplPtrDeleter
and its locking of SolarMutex from the dtors of UNO objects; until then, the
Python code is now sprinkled with some HACKs to make sure all those PyUNO
proxies are released in a timely fashion (see the comment in
unotest/source/python/org/libreoffice/unittest.py for details). (Also, it would
probably help if UnoInProcess didn't keep a local self.xDoc around referencing
(just) the last result of calling one of its open* methods, confusingly making
it the responsibility of UnoInProcess to close that one document while making it
the responsibility of the test code making the other UnoInProcess.open* calls to
close any other documents.)
Change-Id: Ief27c81e2b763e9be20cbf3234b68924315f13be
Reviewed-on: https://gerrit.libreoffice.org/60100
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
Stephan Bergmann