Fix: pthread_create self-recursion on glibc >= 2.34 (source/real.cpp)

Bug

real.cpp::initializer() resolves the real pthread_create via:

void *pthread_handle = dlopen("libpthread.so.0", RTLD_NOW | RTLD_GLOBAL | RTLD_NOLOAD);
INIT_WRAPPER(pthread_create, pthread_handle);

This assumes pthread functions live in a separate libpthread.so.0, which was true before glibc 2.34 (2021). On any system with glibc >= 2.34 (e.g. Ubuntu with glibc 2.42), pthread functions are merged into libc.so.6, so libpthread.so.0 is never loaded into the process. RTLD_NOLOAD then makes dlopen return NULL.

In dlsym(NULL, "pthread_create"), NULL means "search the global default scope." Since this library is loaded via LD_PRELOAD, its own interposed pthread_create is first in that scope, so Real::pthread_create ends up pointing at itself. Every call to xthread::thread_create() then recurses through its own pthread_create wrapper infinitely, never reaching a real OS thread, until _alives blows past MAX_ALIVE_THREADS and trips the assertion in allocThreadIndex():

threadspawn_test: ./xthread.hh:213: int xthread::allocThreadIndex(): Assertion `_alives < _max' failed.

Confirmed with a gdb backtrace on the crash - repeating pattern all the way down:

xthread::thread_create (xthread.cpp:62)
  -> pthread_create (libnumalloc.cpp:343)   <- this library's own wrapper
    -> xthread::thread_create (xthread.cpp:62)
      -> pthread_create (libnumalloc.cpp:343)
        -> ... (repeats)

Also confirmed on this system:

• ldd <any binary preloading libnumalloc.so> | grep pthread → empty (libpthread.so.0 never loaded)
• nm -D libnumalloc.so | grep pthread_create → T pthread_create (confirms it's first in global scope)

Reproduces on an unmodified build (NUMA_NODES=2, default config) - unrelated to node count or topology.

Fix

Resolve pthread_create/pthread_join/pthread_kill via RTLD_NEXT instead - the same mechanism this file already uses correctly for malloc/free/calloc/realloc two lines above. RTLD_NEXT finds the next definition after the calling library in the normal search order, regardless of whether pthread symbols live in a separate libpthread.so.0 or in libc.so.6.

 void initializer() {
   INIT_WRAPPER(malloc, RTLD_NEXT);
   INIT_WRAPPER(free, RTLD_NEXT);
   INIT_WRAPPER(calloc, RTLD_NEXT);
   INIT_WRAPPER(realloc, RTLD_NEXT);

-  void *pthread_handle = dlopen("libpthread.so.0", RTLD_NOW | RTLD_GLOBAL | RTLD_NOLOAD);
-  INIT_WRAPPER(pthread_create, pthread_handle);
-  INIT_WRAPPER(pthread_join, pthread_handle);
-  INIT_WRAPPER(pthread_kill, pthread_handle);
+  INIT_WRAPPER(pthread_create, RTLD_NEXT);
+  INIT_WRAPPER(pthread_join, RTLD_NEXT);
+  INIT_WRAPPER(pthread_kill, RTLD_NEXT);
 }

Testing

Multithreaded smoke test (6 threads, malloc/free per thread) via LD_PRELOAD:

• Before: crashes on the 2nd thread creation, every time, regardless of NUMA_NODES.
• After: all 6 threads complete, exit code 0.

Tested on Ubuntu, glibc 2.42, clang++ build per the existing Makefile.

Scope

This PR only touches source/, since that's the copy matching the published ISMM'23 paper. The identical dlopen("libpthread.so.0", ...) pattern also exists in source-0104, source-48-class-size, source-fasterthantcmalloc-raytrace, source-fullinterleaved, source-linklist, source-reducememory, source-remmove-jumplist, source-scalability, source-smallbag, and (with minor unrelated differences) binding/ and try/ - all 12 copies in the repo are affected. Happy to open follow-up PRs for those if useful.