Running Coz with strace shows that the child process terminates with SIGSEGV. If you capture a core dump, you can see that there is a stack overflow due to infinite recursion in pthread_cond_broadcast. This is because RTLD_DEFAULT is defined as NULL in include/dlfcn.h, so dlsym interprets the NULL handle passed to it as RTLD_DEFAULT. This causes the lookup for the "real" function to use the default search order, hence returning the overridden symbol from libcoz.so instead of the actual pthreads implementation.
In
get_pthread_handle(real.cpp),dlopencan fail to load "libpthread.so.0" and return a NULL pointer. This condition is not properly handled by Coz.How to reproduce the problem:
Build https://github.com/ClickHouse/ClickHouse. Try profiling the
clickhouse_serverprogram with Coz. It fails silently (no output).Debugging:
Running Coz with
straceshows that the child process terminates with SIGSEGV. If you capture a core dump, you can see that there is a stack overflow due to infinite recursion inpthread_cond_broadcast. This is becauseRTLD_DEFAULTis defined asNULLininclude/dlfcn.h, sodlsyminterprets theNULLhandle passed to it asRTLD_DEFAULT. This causes the lookup for the "real" function to use the default search order, hence returning the overridden symbol fromlibcoz.soinstead of the actual pthreads implementation.