Add kludge to make the default threadpool and threaded resolver work in forked...

Add kludge to make the default threadpool and threaded resolver work in forked children. The real issue seems to be that fork watchers aren't firing; there doesn't seem to be a test case for that. Fixes #230.

changelog.rst

@@ -45,6 +45,10 @@ Unreleased
 - ``gevent.iwait`` no longer throws ``LoopExit`` if the caller
   switches greenlets between return values. Reported and initial patch
   in :pr:`467` by Alexey Borzenkov.
+- The default threadpool and default threaded resolver work in a
+  forked child process, such as with ``multiprocessing.Process``.
+  Previously the child process would hang indefinitely. Reported in
+  :issue:`230` by Lx Yu.
 
 1.1a1 (Jun 29, 2015)
 ====================

gevent/hub.py

@@ -153,6 +153,21 @@ def reinit():
     hub = _get_hub()
     if hub is not None:
         hub.loop.reinit()
+        # XXX: libev's fork watchers seem not to be firing for some reason
+        # in both the cython (core.ppyx) and CFFI (corecffi.py) implementations
+        # (at least on OS X; confirm on other platforms)
+        # This breaks the threadpool and anything that uses it, including
+        # resolver_thread in the forked process (if there was already one thread
+        # in the pool before fork, adding an additional task will hang forever post-fork)
+        # The below is a kludge. The correct fix is to figure out why the fork watchers
+        # don't work. Fortunately, both of these methods are idempotent and can be called
+        # multiple times following a fork if the suddenly started working, or were already
+        # working on some platforms.
+        if hasattr(hub.threadpool, '_on_fork'):
+            hub.threadpool._on_fork()
+        # resolver_ares also has a fork watcher that's not firing
+        if hasattr(hub.resolver, '_on_fork'):
+            hub.resolver._on_fork()
 
 
 def get_hub_class():

gevent/resolver_ares.py

@@ -36,6 +36,7 @@ class Resolver(object):
         return '<gevent.resolver_ares.Resolver at 0x%x ares=%r>' % (id(self), self.ares)
 
     def _on_fork(self):
+        # NOTE: See comment in gevent.hub.reinit.
         pid = os.getpid()
         if pid != self.pid:
             self.hub.loop.run_callback(self.ares.destroy)

gevent/threadpool.py

@@ -86,7 +86,8 @@ class ThreadPool(GroupMappingMixin):
 
     def _on_fork(self):
         # fork() only leaves one thread; also screws up locks;
-        # let's re-create locks and threads
+        # let's re-create locks and threads.
+        # NOTE: See comment in gevent.hub.reinit.
         pid = os.getpid()
         if pid != self.pid:
             self.pid = pid

greentest/test__issue230.py

+import gevent.monkey
+gevent.monkey.patch_all()
+
+import socket
+import multiprocessing
+
+# Make sure that using the resolver in a forked process
+# doesn't hang forever.
+
+def block():
+	socket.getaddrinfo('localhost', 8001)
+
+def main():
+    socket.getaddrinfo('localhost', 8001)
+
+    p = multiprocessing.Process(target=block)
+    p.start()
+    p.join()
+
+if __name__ == '__main__':
+    main()