Handle more cases, and be more careful with hubs.

Implement timeouts for the no-hub case using spin locks.

docs/changes/1698.bugfix

@@ -2,4 +2,14 @@ Improve the ability to use monkey-patched locks, and
 `gevent.lock.BoundedSemaphore`, across threads, especially when the
 various threads might not have a gevent hub or any other active
 greenlets. In particular, this handles some cases that previously
-raised ``LoopExit``.
+raised ``LoopExit`` or would hang.
+
+The semaphore tries to avoid creating a hub if it seems unnecessary,
+automatically creating one in the single-threaded case when it would
+block, but not in the multi-threaded case. While the differences
+should be correctly detected, it's possible there are corner cases
+where they might not be.
+
+If your application appears to hang acquiring semaphores, but adding a
+call to ``gevent.get_hub()`` in the thread attempting to acquire the
+semaphore before doing so fixes it, please file an issue.

src/gevent/_abstract_linkable.py

@@ -145,11 +145,16 @@ class AbstractLinkable(object):
     def _getcurrent(self):
         return getcurrent() # pylint:disable=undefined-variable
 
+    def _get_thread_ident(self):
+        return _get_thread_ident()
+
     def _capture_hub(self, create):
         # Subclasses should call this as the first action from any
         # public method that could, in theory, block and switch
         # to the hub. This may release the GIL. It may
         # raise InvalidThreadUseError if the result would
+
+        # First, detect a dead hub and drop it.
         while 1:
             my_hub = self.hub
             if my_hub is None:
@@ -178,20 +183,30 @@ class AbstractLinkable(object):
                 get_hub_if_exists(),
                 getcurrent() # pylint:disable=undefined-variable
             )
-        return 1
+        return self.hub
 
     def _check_and_notify(self):
         # If this object is ready to be notified, begin the process.
         if self.ready() and self._links and not self._notifier:
+            hub = None
             try:
-                self._capture_hub(True) # Must create, we need it.
+                hub = self._capture_hub(False) # Must create, we need it.
             except InvalidThreadUseError:
                 # The current hub doesn't match self.hub. That's OK,
                 # we still want to start the notifier in the thread running
                 # self.hub (because the links probably contains greenlet.switch
                 # calls valid only in that hub)
                 pass
-            self._notifier = self.hub.loop.run_callback(self._notify_links, [])
+            if hub is not None:
+                self._notifier = hub.loop.run_callback(self._notify_links, [])
+            else:
+                # Hmm, no hub. We must be the only thing running. Then its OK
+                # to just directly call the callbacks.
+                self._notifier = 1
+                try:
+                    self._notify_links([])
+                finally:
+                    self._notifier = None
 
     def _notify_link_list(self, links):
         # The core of the _notify_links method to notify
@@ -201,6 +216,9 @@ class AbstractLinkable(object):
         only_while_ready = not self._notify_all
         final_link = links[-1]
         done = set() # of ids
+        hub = self.hub
+        if hub is None:
+            hub = get_hub_if_exists()
         while links: # remember this can be mutated
             if only_while_ready and not self.ready():
                 break
@@ -222,7 +240,11 @@ class AbstractLinkable(object):
 
                 except: # pylint:disable=bare-except
                     # We're running in the hub, errors must not escape.
-                    self.hub.handle_error((link, self), *sys.exc_info())
+                    if hub is not None:
+                        hub.handle_error((link, self), *sys.exc_info())
+                    else:
+                        import traceback
+                        traceback.print_exc()
 
             if link is final_link:
                 break

src/gevent/_gevent_c_abstract_linkable.pxd

@@ -51,13 +51,10 @@ cdef class AbstractLinkable(object):
    cpdef unlink(self, callback)
 
    cdef _check_and_notify(self)
-   cdef int _capture_hub(self, bint create) except -1
+   cdef SwitchOutGreenletWithLoop _capture_hub(self, bint create)
    cdef __wait_to_be_notified(self, bint rawlink)
 
    cdef void _quiet_unlink_all(self, obj) # suppress exceptions
-
-   cdef _allocate_lock(self)
-   cdef greenlet _getcurrent(self)
    cdef int _switch_to_hub(self, the_hub) except -1
 
    @cython.nonecheck(False)
@@ -72,3 +69,8 @@ cdef class AbstractLinkable(object):
    cdef _wait_core(self, timeout, catch=*)
    cdef _wait_return_value(self, bint waited, bint wait_success)
    cdef _wait(self, timeout=*)
+
+   # Unreleated utilities
+   cdef _allocate_lock(self)
+   cdef greenlet _getcurrent(self)
+   cdef _get_thread_ident(self)

src/gevent/_gevent_c_semaphore.pxd

@@ -5,16 +5,22 @@ from gevent._gevent_c_abstract_linkable cimport AbstractLinkable
 from gevent._gevent_c_hub_local cimport get_hub_if_exists
 from gevent._gevent_c_hub_local cimport get_hub_noargs as get_hub
 
-cdef Timeout
 cdef InvalidThreadUseError
 cdef LoopExit
-cdef spawn_raw
+cdef Timeout
 cdef _native_sleep
+cdef monotonic
+cdef spawn_raw
+
+cdef class _LockReleaseLink(object):
+    cdef object lock
 
 
 cdef class Semaphore(AbstractLinkable):
     cdef public int counter
 
+    cdef long _multithreaded
+
     cpdef bint locked(self)
     cpdef int release(self) except -1000
     # We don't really want this to be public, but
@@ -38,6 +44,15 @@ cdef class Semaphore(AbstractLinkable):
     cdef __acquire_from_other_thread(self, tuple args, bint blocking, timeout)
     cpdef __acquire_from_other_thread_cb(self, list results, bint blocking, timeout, thread_lock)
 
+    cdef __add_link(self, link)
+    cdef __acquire_using_two_hubs(self,
+                                  SwitchOutGreenletWithLoop hub_for_this_thread,
+                                  current_greenlet,
+                                  timeout)
+    cdef __acquire_using_other_hub(self, SwitchOutGreenletWithLoop owning_hub, timeout)
+    cdef bint __acquire_without_hubs(self, timeout)
+    cdef bint __spin_on_native_lock(self, thread_lock, timeout)
+
 cdef class BoundedSemaphore(Semaphore):
     cdef readonly int _initial_value
 

src/gevent/libev/corecffi.py

@@ -294,6 +294,10 @@ class loop(AbstractLoop):
             libev.ev_timer_stop(self._timer0)
 
     def _setup_for_run_callback(self):
+        # XXX: libuv needs to start the callback timer to be sure
+        # that the loop wakes up and calls this. Our C version doesn't
+        # do this.
+        # self._start_callback_timer()
         self.ref() # we should go through the loop now
 
     def destroy(self):

src/gevent/testing/errorhandler.py

@@ -22,31 +22,36 @@ from functools import wraps
 
 
 def wrap_error_fatal(method):
-    import gevent
-    system_error = gevent.get_hub().SYSTEM_ERROR
+    from gevent._hub_local import get_hub_class
+    system_error = get_hub_class().SYSTEM_ERROR
 
     @wraps(method)
     def wrapper(self, *args, **kwargs):
         # XXX should also be able to do gevent.SYSTEM_ERROR = object
         # which is a global default to all hubs
-        gevent.get_hub().SYSTEM_ERROR = object
+        get_hub_class().SYSTEM_ERROR = object
         try:
             return method(self, *args, **kwargs)
         finally:
-            gevent.get_hub().SYSTEM_ERROR = system_error
+            get_hub_class().SYSTEM_ERROR = system_error
     return wrapper
 
 
 def wrap_restore_handle_error(method):
-    import gevent
-    old = gevent.get_hub().handle_error
+    from gevent._hub_local import get_hub_if_exists
+    from gevent import getcurrent
 
     @wraps(method)
     def wrapper(self, *args, **kwargs):
         try:
             return method(self, *args, **kwargs)
         finally:
-            gevent.get_hub().handle_error = old
+            # Remove any customized handle_error, if set on the
+            # instance.
+            try:
+                del get_hub_if_exists().handle_error
+            except AttributeError:
+                pass
         if self.peek_error()[0] is not None:
-            gevent.getcurrent().throw(*self.peek_error()[1:])
+            getcurrent().throw(*self.peek_error()[1:])
     return wrapper

src/gevent/tests/test__semaphore.py

@@ -16,7 +16,6 @@ from gevent.lock import BoundedSemaphore
 
 import gevent.testing as greentest
 from gevent.testing import timing
-from gevent.testing import flaky
 
 class TestSemaphore(greentest.TestCase):
 
@@ -132,6 +131,7 @@ class TestSemaphoreMultiThread(greentest.TestCase):
             acquired, exc_info,
             **thread_acquire_kwargs
         ))
+        t.daemon = True
         t.start()
         thread_running.wait(10) # implausibly large time
         if release:
@@ -151,6 +151,14 @@ class TestSemaphoreMultiThread(greentest.TestCase):
 
             self.assertEqual(acquired, [True])
 
+        if not release and thread_acquire_kwargs.get("timeout"):
+            # Spin the loop to be sure that the timeout has a chance to
+            # process. Interleave this with something that drops the GIL
+            # so the background thread has a chance to notice that.
+            for _ in range(3):
+                gevent.idle()
+                if thread_acquired.wait(timing.LARGE_TICK):
+                    break
         thread_acquired.wait(timing.LARGE_TICK * 5)
         if require_thread_acquired_to_finish:
             self.assertTrue(thread_acquired.is_set())
@@ -210,9 +218,12 @@ class TestSemaphoreMultiThread(greentest.TestCase):
             acquired, exc_info,
             timeout=timing.LARGE_TICK
         ))
+        thread.daemon = True
         gevent.idle()
         sem.release()
         glet.join()
+        for _ in range(3):
+            gevent.idle()
             thread.join(timing.LARGE_TICK)
 
         self.assertEqual(glet.value, True)
@@ -260,13 +271,20 @@ class TestSemaphoreMultiThread(greentest.TestCase):
                     sem.acquire(*acquire_args)
                     sem.release()
                     results[ix] = i
+                    if not create_hub:
+                        # We don't artificially create the hub.
+                        self.assertIsNone(
+                            get_hub_if_exists(),
+                            (get_hub_if_exists(), ix, i)
+                        )
                     if create_hub and i % 10 == 0:
                         gevent.sleep(timing.SMALLEST_RELIABLE_DELAY)
                     elif i % 100 == 0:
                         native_sleep(timing.SMALLEST_RELIABLE_DELAY)
             except Exception as ex: # pylint:disable=broad-except
                 import traceback; traceback.print_exc()
-                results[ix] = ex
+                results[ix] = str(ex)
+                ex = None
             finally:
                 hub = get_hub_if_exists()
                 if hub is not None:
@@ -285,23 +303,14 @@ class TestSemaphoreMultiThread(greentest.TestCase):
 
         while t1.is_alive() or t2.is_alive():
             cur = list(results)
-            t1.join(2)
-            t2.join(2)
+            t1.join(7)
+            t2.join(7)
             if cur == results:
                 # Hmm, after two seconds, no progress
-                print("No progress!", cur, results, t1, t2)
-                from gevent.util import print_run_info
-                print_run_info()
                 run = False
                 break
 
-        try:
         self.assertEqual(results, [count - 1, count - 1])
-        except AssertionError:
-            if greentest.PY2:
-                flaky.reraiseFlakyTestRaceCondition()
-            else:
-                raise
 
     def test_dueling_threads_timeout(self):
         self.test_dueling_threads((True, 4))

src/gevent/thread.py

@@ -116,14 +116,7 @@ class LockType(BoundedSemaphore):
             if timeout > self._TIMEOUT_MAX:
                 raise OverflowError('timeout value is too large')
 
-        acquired = BoundedSemaphore.acquire(self, 0)
-        if not acquired and getcurrent() is not get_hub_if_exists() and blocking and not timeout:
-            # If we would block forever, and we're not in the hub, and a trivial non-blocking
-            # check didn't get us the lock, then try to run pending callbacks that might
-            # release the lock.
-            sleep()
 
-        if not acquired:
         try:
             acquired = BoundedSemaphore.acquire(self, blocking, timeout)
         except LoopExit: