This is the mail archive of the
mailing list for the pthreas-win32 project.
Re: Crash when re-initializing as static library
- From: Klaus Fischer <klaus dot fischer at tara-systems dot de>
- To: pthreads-win32 at sourceware dot org
- Date: Thu, 07 Nov 2013 15:19:22 +0100
- Subject: Re: Crash when re-initializing as static library
- Authentication-results: sourceware.org; auth=none
Let me give you a little background in order to understand my situation:
I'm actually using the Portable SDK for UPnP Devices (libupnp), which in
turn relies on pthreads. I'm developing cross-plattform (Windows/Linux),
so I need Pthread-Win32 (henceforth called PTW32) to enable libupnp to
run under Windows.
libupnp in turn initializes and deinitializes PTW32 every time libupnp
itself gets initialized/deinitialized. De-/initialization of PTW32 is
explicitly done via pthread_win32_process_* when libupnp is compiled for
WIN32 and PTW32 is used as a static library. This is still done that way
in the latest libupnp release (1.6.18).
One of my first tests was to repeatedly call UpnpInit()/UpnpFinish() to
check for general stability, which is also always a good test for
software cleanliness according to my experience; unfortunately, this
immediately resulted in the access violation in PTW32.
It looks like libupnp still has to adapt to the changed calling behavior
of PTW32, although they had plenty of time for it (the latest libupnp
release is from January 2013, while the latest PTW32 is from May 2012).
They probably overlooked the change in calling pthread_win32_process_*;
I will send them a note for it.
Nevertheless, I will continue to use my locally patched version of PTW32
until either libupnp adapted to the changed PTW32 behavior or PTW32 is
available with initialization of globals during the attaching phase.
I've ran into the same problem in some of my libraries too, and decided
to give my globals a proper initialization in an Init()-function to
avoid such issues during repeated de-/initialization. It would be nice
if you could consider doing this in PTW32 too, just for the sake of
backwards-compatibility (if this actually worked in pre-2.9.0 versions).
> Applications statically linked with current versions of the library
no longer need to call those routines explicitly. From README.NONPORTABLE:
> These functions contain the code normally run via DllMain
> when the library is used as a dll. As of version 2.9.0 of the
> library, static builds using either MSC or GCC will call
> pthread_win32_process_* automatically at application startup and
> exit respectively.
> But you are also detaching and reattaching within the same process,
which is unintended use. Is this how you expect to use the library or
are you just analysing?
>> On 7/11/2013 4:53 AM, Klaus Fischer wrote:
>> Dear pthreads-win32 developers,
>> I have experienced a crash when building pthreads-win32 as static
library and re-initializing it using the following sequence:
>> - pthread_win32_process_attach_np()
>> - pthread_win32_process_detach_np()
>> - pthread_win32_process_attach_np()
>> The global variable ptw32_threadReuseTop still points to memory used
between the first attach/detach run, but this memory was already freed
in function ptw32_processTerminate(), which was called during detaching.
>> When using e.g. pthread_self() afterwards, the global
ptw32_threadReuseTop now points to invalid memory, causing an access
violation writing to that memory location.
>> A simple code change fixed that problem by assigning the default
value PTW32_THREAD_REUSE_EMPTY to that global variable at the end of
function ptw32_processTerminate(), after the while loop freeing all
still allocated thread handles.
>> A better way to fix that would probably be to initialize all the
library globals of global.c during the attaching stage. In a static
library, those globals are only initialized once when the process
starts, but _should_ be re-initialized on every attach.
>> I know this is not a concern when using this library as dynamic
library, but since there is an option to use it as static library and
other people also use it that way according to the mailing list, it
would be great if it could survive multiple re-initializations.
>> Thanks in advance,