ENH: fix Python multi-processing hang on unix

Modules/Core/Common/include/itkThreadPool.h

@@ -124,13 +124,23 @@ auto result = pool->AddWork([](int param) { return param; }, 7);
   SetDoNotWaitForThreads(bool doNotWaitForThreads);
 
 protected:
-  /* We need access to the mutex in AddWork, and the variable is only
+  /** We need access to the mutex in AddWork, and the variable is only
    * visible in .cxx file, so this method returns it. */
   std::mutex &
   GetMutex();
 
   ThreadPool();
-  ~ThreadPool() override;
+
+  /** Stop the pool and release threads. To be called by the destructor and atfork. */
+  void
+  CleanUp();
+
+  ~ThreadPool() override { this->CleanUp(); }
+
+  static void
+  atfork_prepare();
+  static void
+  atfork_resume();
 
 private:
   /** Only used to synchronize the global variable across static libraries.*/

Modules/Core/Common/src/itkThreadPool.cxx

@@ -77,6 +77,9 @@ ThreadPool::GetInstance()
       {
         new ThreadPool(); // constructor sets m_PimplGlobals->m_ThreadPoolInstance
       }
+#if defined(ITK_USE_PTHREADS)
+      pthread_atfork(ThreadPool::atfork_prepare, ThreadPool::atfork_resume, ThreadPool::atfork_resume);
+#endif
     }
   }
   return m_PimplGlobals->m_ThreadPoolInstance;
@@ -132,7 +135,8 @@ ThreadPool::GetNumberOfCurrentlyIdleThreads() const
   return int(m_Threads.size()) - int(m_WorkQueue.size()); // lousy approximation
 }
 
-ThreadPool::~ThreadPool()
+void
+ThreadPool::CleanUp()
 {
   {
     std::unique_lock<std::mutex> mutexHolder(m_PimplGlobals->m_Mutex);
@@ -154,6 +158,23 @@ ThreadPool::~ThreadPool()
   }
 }
 
+void
+ThreadPool::atfork_prepare()
+{
+  printf("In ThreadPool::atfork_prepare");
+  m_PimplGlobals->m_ThreadPoolInstance->CleanUp();
+}
+
+void
+ThreadPool::atfork_resume()
+{
+  printf("In ThreadPool::atfork_resume");
+  ThreadPool * instance = m_PimplGlobals->m_ThreadPoolInstance.GetPointer();
+  ThreadIdType threadCount = instance->m_Threads.size();
+  instance->m_Threads.clear();
+  instance->m_Stopping = false;
+  instance->AddThreads(threadCount);
+}
 
 void
 ThreadPool::ThreadExecute()