Bug throwing exceptions with std::call_once

Question

I tried to run the example from the docs page https://en.cppreference.com/w/cpp/thread/call_once but it doesn't work as spected. It gets stuck infinitely. I'd like to know why this happens or if it's just a bug related to some specific compiler version. This is what I'm using to run the program https://repl.it/repls/UtterJubilantArchitects

#include <iostream>
#include <thread>
#include <mutex>

std::once_flag flag1, flag2;

void simple_do_once()
{
    std::call_once(flag1, [](){ std::cout << "Simple example: called once\n"; });
}

void may_throw_function(bool do_throw)
{
  if (do_throw) {
    std::cout << "throw: call_once will retry\n"; // this may appear more than once
    throw std::exception();
  }
  std::cout << "Didn't throw, call_once will not attempt again\n"; // guaranteed once
}

void do_once(bool do_throw)
{
  try {
    std::call_once(flag2, may_throw_function, do_throw);
  }
  catch (...) {
  }
}

int main()
{
    std::thread st1(simple_do_once);
    std::thread st2(simple_do_once);
    std::thread st3(simple_do_once);
    std::thread st4(simple_do_once);
    st1.join();
    st2.join();
    st3.join();
    st4.join();

    std::thread t1(do_once, true);
    std::thread t2(do_once, true);
    std::thread t3(do_once, false);
    std::thread t4(do_once, true);
    t1.join();
    t2.join();
    t3.join();
    t4.join();
}

Yonder

The issue is libstdc++-specific (note that repl.it uses clang with libstdc++ as can be verified by checking the __GLIBCXX__ macro). It is not present with libc++, see e.g. wandbox.org/permlink/7rPCL57q8E55Eq6m. See also answer below and my comments on it. – Bettinabettine

Thx, this is solution for me) – Scan

Answer 1

This behavior is an implementation bug. call_onse (via pthread_once) supposedly uses int pthread_mutex_lock(pthread_mutex_t)* and int pthread_mutex_unlock(pthread_mutex_t)*, as well as your code between them, they are not exception-safe.

Links to some related bugs:https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66146, sourceware.org/bugzilla/show_bug.cgi?id=18435, austingroupbugs.net/view.php?id=863#c2619

Details of implementation:

this is pthread implementation from gnu.org

int
__pthread_once (pthread_once_t *once_control, void (*init_routine) (void))
{
  __memory_barrier ();
  if (once_control->__run == 0)
    {
      __pthread_spin_lock (&once_control->__lock);

      if (once_control->__run == 0)
      {
        init_routine ();
        __memory_barrier ();
        once_control->__run = 1;
      }

      __pthread_spin_unlock (&once_control->__lock);
    }

  return 0;
}

this is call_once implementation from g++ 7.4.0 package (ubuntu 18.4.0): (we have active _GLIBCXX_HAVE_TLS branch)

  /// call_once
  template<typename _Callable, typename... _Args>
    void
    call_once(once_flag& __once, _Callable&& __f, _Args&&... __args)
    {
      // _GLIBCXX_RESOLVE_LIB_DEFECTS
      // 2442. call_once() shouldn't DECAY_COPY()
      auto __callable = [&] {
      std::__invoke(std::forward<_Callable>(__f),
            std::forward<_Args>(__args)...);
      };
#ifdef _GLIBCXX_HAVE_TLS
      __once_callable = std::__addressof(__callable);
      __once_call = []{ (*(decltype(__callable)*)__once_callable)(); };
#else
      unique_lock<mutex> __functor_lock(__get_once_mutex());
      __once_functor = __callable;
      __set_once_functor_lock_ptr(&__functor_lock);
#endif

      int __e = __gthread_once(&__once._M_once, &__once_proxy);

#ifndef _GLIBCXX_HAVE_TLS
      if (__functor_lock)
        __set_once_functor_lock_ptr(0);
#endif

#ifdef __clang_analyzer__
      // PR libstdc++/82481
      __once_callable = nullptr;
      __once_call = nullptr;
#endif

      if (__e)
          __throw_system_error(__e);
    }

__once_proxy is:

extern "C"
  {
    void __once_proxy()
    {
#ifndef _GLIBCXX_HAVE_TLS
      function<void()> __once_call = std::move(__once_functor);
      if (unique_lock<mutex>* __lock = __get_once_functor_lock_ptr())
      {
        // caller is using new ABI and provided lock ptr
        __get_once_functor_lock_ptr() = 0;
        __lock->unlock();
      }
      else
        __get_once_functor_lock().unlock();  // global lock
#endif
      __once_call();
    }
  }

_GLIBCXX_END_NAMESPACE_VERSION
} // namespace std

this is implementation of pthread_once:

struct __pthread_once
{
  int __run;
  __pthread_spinlock_t __lock;
};

and __pthread_spinlock_t:

typedef __volatile int __pthread_spinlock_t;

__pthread_spin_lock is typedef for

void
__spin_lock_solid (spin_lock_t *lock)
{
  while (__spin_lock_locked (lock) || ! __spin_try_lock (lock))
    /* Yield to another thread (system call).  */
    __swtch_pri (0);
}

Personally, I do not see in these areas when the mutex unlock protection is against exceptions. We have double checking lock and call of Callable object inside.

I think, that noexcept Callable or a double-check lock with lock_guard/unique_lock(if you are sure of the atomic nature of reading a variable), may be used as a solution. I'm followed the solution of walnut, installed libc++-dev, libc++abi-dev and compiled this code using

clang++ -stdlib=libc++ -lc++abi 

and it works well.

Scan