Merge pull request #121 from njsmith/fix-120

docs-requirements.in

@@ -5,8 +5,9 @@ sphinxcontrib-trio
 # Workaround for this weird issue:
 # https://travis-ci.org/python-trio/pytest-trio/jobs/407495415
 attrs >= 17.4.0
-# != 19.9.0rc1 for https://github.com/twisted/towncrier/issues/180
-towncrier != 19.9.0rc1
+# != 19.9.0 for https://github.com/twisted/towncrier/issues/180
+# != 21.3.0 for https://github.com/twisted/towncrier/issues/346
+towncrier != 19.9.0,!= 21.3.0
 
 # pytest-trio's own dependencies
 trio >= 0.15.0

docs-requirements.txt

@@ -2,7 +2,7 @@
 # This file is autogenerated by pip-compile
 # To update, run:
 #
-#    pip-compile --output-file=docs-requirements.txt docs-requirements.in
+#    pip-compile docs-requirements.in
 #
 alabaster==0.7.12
     # via sphinx
@@ -22,12 +22,8 @@ certifi==2020.12.5
     # via requests
 chardet==4.0.0
     # via requests
-click-default-group==1.2.2
-    # via towncrier
 click==7.1.2
-    # via
-    #   click-default-group
-    #   towncrier
+    # via towncrier
 docutils==0.16
     # via
     #   sphinx
@@ -101,7 +97,7 @@ toml==0.10.2
     # via
     #   pytest
     #   towncrier
-towncrier==21.3.0
+towncrier==19.2.0
     # via -r docs-requirements.in
 trio==0.18.0
     # via -r docs-requirements.in

newsfragments/120.bugfix.rst

@@ -0,0 +1,3 @@
+Fix an issue where if two fixtures are being set up concurrently, and
+one crashes and the other hangs, then the test as a whole would hang,
+rather than being cancelled and unwound after the crash.

pytest_trio/_tests/test_fixture_ordering.py

@@ -154,13 +154,15 @@ def test_error_collection(testdir):
 
         @trio_fixture
         async def crash_nongen():
-            await trio.sleep(2)
+            with trio.CancelScope(shield=True):
+                await trio.sleep(2)
             raise RuntimeError("crash_nongen".upper())
 
         @trio_fixture
         @async_generator
         async def crash_early_agen():
-            await trio.sleep(2)
+            with trio.CancelScope(shield=True):
+                await trio.sleep(2)
             raise RuntimeError("crash_early_agen".upper())
             await yield_()
 
@@ -330,3 +332,43 @@ async def test_try(fixture):
     result.stdout.fnmatch_lines_random([
         "*OOPS*",
     ])
+
+
+# Makes sure that
+# See https://github.com/python-trio/pytest-trio/issues/120
+def test_fixtures_crash_and_hang_concurrently(testdir):
+    testdir.makepyfile(
+        """
+        import trio
+        import pytest
+
+
+        @pytest.fixture
+        async def hanging_fixture():
+            print("hanging_fixture:start")
+            await trio.Event().wait()
+            yield
+            print("hanging_fixture:end")
+
+
+        @pytest.fixture
+        async def exploding_fixture():
+            print("exploding_fixture:start")
+            raise Exception
+            yield
+            print("exploding_fixture:end")
+
+
+        @pytest.mark.trio
+        async def test_fails_right_away(exploding_fixture):
+            ...
+
+
+        @pytest.mark.trio
+        async def test_fails_needs_some_scopes(exploding_fixture, hanging_fixture):
+            ...
+        """
+    )
+
+    result = testdir.runpytest()
+    result.assert_outcomes(passed=0, failed=2)

pytest_trio/plugin.py

@@ -121,19 +121,25 @@ def pytest_exception_interact(node, call, report):
 # If a fixture crashes, whether during setup, teardown, or in a background
 # task at any other point, then we mark the whole test run as "crashed". When
 # a run is "crashed", two things happen: (1) if any fixtures or the test
-# itself haven't started yet, then we don't start them. (2) if the test is
-# running, we cancel it. That's all. In particular, if a fixture has a
-# background crash, we don't propagate that to any other fixtures, we still
-# follow the normal teardown sequence, and so on – but since the test is
-# cancelled, the teardown sequence should start immediately.
+# itself haven't started yet, then we don't start them, and treat them as if
+# they've already exited. (2) if the test is running, we cancel it. That's
+# all. In particular, if a fixture has a background crash, we don't propagate
+# that to any other fixtures, we still follow the normal teardown sequence,
+# and so on – but since the test is cancelled, the teardown sequence should
+# start immediately.
 
 canary = contextvars.ContextVar("pytest-trio canary")
 
 
 class TrioTestContext:
     def __init__(self):
         self.crashed = False
-        self.test_cancel_scope = None
+        # This holds cancel scopes for whatever setup steps are currently
+        # running -- initially it's the fixtures that are in the middle of
+        # evaluating themselves, and then once fixtures are set up it's the
+        # test itself. Basically, at any given moment, it's the stuff we need
+        # to cancel if we want to start tearing down our fixture DAG.
+        self.active_cancel_scopes = set()
         self.fixtures_with_errors = set()
         self.fixtures_with_cancel = set()
         self.error_list = []
@@ -145,8 +151,8 @@ def crash(self, fixture, exc):
             self.error_list.append(exc)
             self.fixtures_with_errors.add(fixture)
         self.crashed = True
-        if self.test_cancel_scope is not None:
-            self.test_cancel_scope.cancel()
+        for cscope in self.active_cancel_scopes:
+            cscope.cancel()
 
 
 class TrioFixture:
@@ -240,16 +246,17 @@ async def run(self, test_ctx, contextvars_ctx):
                 return
 
             # Run actual fixture setup step
+            # If another fixture crashes while we're in the middle of setting
+            # up, we want to be cancelled immediately, so we'll save an
+            # encompassing cancel scope where self._crash can find it.
+            test_ctx.active_cancel_scopes.add(nursery_fixture.cancel_scope)
             if self._is_test:
-                # Tests are exactly like fixtures, except that they (1) have
-                # to be regular async functions, (2) if there's a crash, we
-                # should cancel them.
+                # Tests are exactly like fixtures, except that they to be
+                # regular async functions.
                 assert not self.user_done_events
                 func_value = None
-                with trio.CancelScope() as cancel_scope:
-                    test_ctx.test_cancel_scope = cancel_scope
-                    assert not test_ctx.crashed
-                    await self._func(**resolved_kwargs)
+                assert not test_ctx.crashed
+                await self._func(**resolved_kwargs)
             else:
                 func_value = self._func(**resolved_kwargs)
                 if isinstance(func_value, Coroutine):
@@ -261,8 +268,16 @@ async def run(self, test_ctx, contextvars_ctx):
                 else:
                     # Regular synchronous function
                     self.fixture_value = func_value
-
-            # Notify our users that self.fixture_value is ready
+            # Now that we're done setting up, we don't want crashes to cancel
+            # us immediately; instead we want them to cancel our downstream
+            # dependents, and then eventually let us clean up normally. So
+            # remove this from the set of cancel scopes affected by self._crash.
+            test_ctx.active_cancel_scopes.remove(nursery_fixture.cancel_scope)
+
+            # self.fixture_value is ready, so notify users that they can
+            # continue. (Or, maybe we crashed and were cancelled, in which
+            # case our users will check test_ctx.crashed and immediately exit,
+            # which is fine too.)
             self.setup_done.set()
 
             # Wait for users to be finished