Le 26/11/12 22:12, Andrew Harvey a écrit :
Hi,
I am using boost::thread for a plug-in that is loaded by a third-party Linux application. I have no control over the host's threading or when it chooses to unload my code.
Unfortunately I experience a crash every time the host thread unloads the .so and terminates. I get the following backtrace:
(gdb) bt #0 0x00007fffde12c460 in ?? () #1 0x00000032cf607889 in start_thread () from /lib64/libpthread.so.0 #2 0x00000032cf2e5ccd in clone () from /lib64/libc.so.6
Not much to go on...
There are several limitation on Boost.Thread that could make this error. I've not yet the knowledge to fix them.
I am pretty confident the crash is happening in a pthread key destructor registered by Boost. The only parts of the boost::thread library I am currently using are the synchronization classes:
shared_mutex, shared_lock, unique_lock, upgrade_lock, upgrade_to_unique_lock
I'm not directly using any thread-specific storage myself.
What I would like to do is force boost to call its TSS destructor in my plug-in's destroy_dso() function. Is this at all possible?
I don't know which version are you using. For the next release I'm preparing a version that could work without thread interruptions. This should allow to use the synchronization part independently of the boost::thread part, so maybe it could be possible to use the library as a header only one. I will try to test it soon. You could check it on your side, by defining BOOST_THREAD_VERSION to 3 and define BOOST_THREAD_DONT_PROVIDE_INTERRUPTIONS. You will need to link additionally to boost_system and boost_chrono if you are using the Boost.Chrono time related interface. HTH, Vicente