Boost C++ Libraries: Ticket #13283: boost log crash during thread termination

version changed

Marco Strohner — Thu, 22 Mar 2018 13:19:04 GMT

version Boost 1.64.0 → Boost 1.66.0

I've tested it again with boost 1.66.0 and the behaviour is the same. I also tried to compile boost with BOOST_LOG_USE_COMPILER_TLS but the behaviour is still the same.

My results after more debugging:

the logging core creates a thread_data object for the first log that is executed in a thread (core.cpp:open_record). This is triggered by the line BOOST_LOG_TRIVIAL(warning) << "Destroy event #" << i;
This created thread_data object ist added to the tss_node map for automatic cleanup after thread execution is finished.
After the thread execution is finished the tls_destructor is called (thread.cpp:thread_proxy). This tls_destructor destroys the in step 1 created thread_data object
When the method thread_proxy is finished the reference counter of the lambda parameter event is decreased and the destructor is called.
The destructor tries to log one line (BOOST_LOG_TRIVIAL(warning) << "crashes here";). But the method get_thread_data still returns the already destroyed pointer to the thread_data object.

component changed; owner set

Kohei Takahashi — Thu, 22 Mar 2018 15:20:43 GMT

owner set to Andrey Semashev
component None → log

status changed; resolution set

Andrey Semashev — Thu, 22 Mar 2018 17:55:45 GMT

status new → closed
resolution → invalid

So the problem seems to be that the code attempts to write a log record after the thread-local data has been destroyed, am I correct? If so, I'm sorry, but this use case is not going to be supported. Like with global destructors on program termination, there's generally no way to make sure that a thread-local object is still alive while the thread terminates and attempting to log at this point will always be unreliable. My advice is to avoid this practice and finish logging before the thread terminates.