diff --git a/.travis/ansible_tests.sh b/.travis/ansible_tests.sh
index a61ed836..e6441343 100755
--- a/.travis/ansible_tests.sh
+++ b/.travis/ansible_tests.sh
@@ -60,5 +60,5 @@ echo travis_fold:end:job_setup
 echo travis_fold:start:ansible
 /usr/bin/time ./run_ansible_playbook.sh \
     all.yml \
-    -i "${TMPDIR}/hosts"
+    -i "${TMPDIR}/hosts" "$@"
 echo travis_fold:end:ansible
diff --git a/ansible_mitogen/plugins/strategy/mitogen.py b/ansible_mitogen/plugins/strategy/mitogen.py
index 3ef522b4..4f595161 100644
--- a/ansible_mitogen/plugins/strategy/mitogen.py
+++ b/ansible_mitogen/plugins/strategy/mitogen.py
@@ -44,12 +44,12 @@ import sys
 # debuggers and isinstance() work predictably.
 #
 
-try:
-    import ansible_mitogen
-except ImportError:
-    base_dir = os.path.dirname(__file__)
-    sys.path.insert(0, os.path.abspath(os.path.join(base_dir, '../../..')))
-    del base_dir
+BASE_DIR = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), '../../..')
+)
+
+if BASE_DIR not in sys.path:
+    sys.path.insert(0, BASE_DIR)
 
 import ansible_mitogen.strategy
 import ansible.plugins.strategy.linear
diff --git a/ansible_mitogen/plugins/strategy/mitogen_free.py b/ansible_mitogen/plugins/strategy/mitogen_free.py
index 34f959ca..8dfaa16e 100644
--- a/ansible_mitogen/plugins/strategy/mitogen_free.py
+++ b/ansible_mitogen/plugins/strategy/mitogen_free.py
@@ -44,12 +44,12 @@ import sys
 # debuggers and isinstance() work predictably.
 #
 
-try:
-    import ansible_mitogen
-except ImportError:
-    base_dir = os.path.dirname(__file__)
-    sys.path.insert(0, os.path.abspath(os.path.join(base_dir, '../../..')))
-    del base_dir
+BASE_DIR = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), '../../..')
+)
+
+if BASE_DIR not in sys.path:
+    sys.path.insert(0, BASE_DIR)
 
 import ansible_mitogen.loaders
 import ansible_mitogen.strategy
diff --git a/ansible_mitogen/plugins/strategy/mitogen_linear.py b/ansible_mitogen/plugins/strategy/mitogen_linear.py
index a5ea2a3d..d995b67b 100644
--- a/ansible_mitogen/plugins/strategy/mitogen_linear.py
+++ b/ansible_mitogen/plugins/strategy/mitogen_linear.py
@@ -44,12 +44,12 @@ import sys
 # debuggers and isinstance() work predictably.
 #
 
-try:
-    import ansible_mitogen
-except ImportError:
-    base_dir = os.path.dirname(__file__)
-    sys.path.insert(0, os.path.abspath(os.path.join(base_dir, '../../..')))
-    del base_dir
+BASE_DIR = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), '../../..')
+)
+
+if BASE_DIR not in sys.path:
+    sys.path.insert(0, BASE_DIR)
 
 import ansible_mitogen.loaders
 import ansible_mitogen.strategy
diff --git a/ansible_mitogen/process.py b/ansible_mitogen/process.py
index f19079ee..4724ca93 100644
--- a/ansible_mitogen/process.py
+++ b/ansible_mitogen/process.py
@@ -33,6 +33,7 @@ import os
 import signal
 import socket
 import sys
+import time
 
 import mitogen
 import mitogen.core
@@ -114,6 +115,9 @@ class MuxProcess(object):
         mitogen.core.set_cloexec(cls.worker_sock.fileno())
         mitogen.core.set_cloexec(cls.child_sock.fileno())
 
+        if os.environ.get('MITOGEN_PROFILING', '1'):
+            mitogen.core.enable_profiling()
+
         cls.original_env = dict(os.environ)
         cls.child_pid = os.fork()
         ansible_mitogen.logging.setup()
@@ -199,4 +203,10 @@ class MuxProcess(object):
         ourself. In future this should gracefully join the pool, but TERM is
         fine for now.
         """
+        if os.environ.get('MITOGEN_PROFILING'):
+            # TODO: avoid killing pool threads before they have written their
+            # .pstats. Really shouldn't be using kill() here at all, but hard
+            # to guarantee services can always be unblocked during shutdown.
+            time.sleep(1)
+
         os.kill(os.getpid(), signal.SIGTERM)
diff --git a/ansible_mitogen/runner.py b/ansible_mitogen/runner.py
index ca3928b3..86f7b329 100644
--- a/ansible_mitogen/runner.py
+++ b/ansible_mitogen/runner.py
@@ -264,9 +264,9 @@ class ModuleUtilsImporter(object):
         mod.__loader__ = self
         if is_pkg:
             mod.__path__ = []
-            mod.__package__ = fullname
+            mod.__package__ = str(fullname)
         else:
-            mod.__package__ = fullname.rpartition('.')[0]
+            mod.__package__ = str(fullname.rpartition('.')[0])
         exec(code, mod.__dict__)
         self._loaded.add(fullname)
         return mod
diff --git a/ansible_mitogen/services.py b/ansible_mitogen/services.py
index a7bb7db1..e95fc226 100644
--- a/ansible_mitogen/services.py
+++ b/ansible_mitogen/services.py
@@ -388,6 +388,8 @@ class ModuleDepService(mitogen.service.Service):
     Scan a new-style module and produce a cached mapping of module_utils names
     to their resolved filesystem paths.
     """
+    invoker_class = mitogen.service.SerializedInvoker
+
     def __init__(self, *args, **kwargs):
         super(ModuleDepService, self).__init__(*args, **kwargs)
         self._cache = {}
diff --git a/ansible_mitogen/target.py b/ansible_mitogen/target.py
index e5365dd4..582bf85e 100644
--- a/ansible_mitogen/target.py
+++ b/ansible_mitogen/target.py
@@ -46,16 +46,26 @@ import re
 import signal
 import stat
 import subprocess
+import sys
 import tempfile
 import traceback
+import types
 
-import ansible.module_utils.json_utils
-import ansible_mitogen.runner
 import mitogen.core
 import mitogen.fork
 import mitogen.parent
 import mitogen.service
 
+# Ansible since PR #41749 inserts "import __main__" into
+# ansible.module_utils.basic. Mitogen's importer will refuse such an import, so
+# we must setup a fake "__main__" before that module is ever imported. The
+# str() is to cast Unicode to bytes on Python 2.6.
+if not sys.modules.get(str('__main__')):
+    sys.modules[str('__main__')] = types.ModuleType(str('__main__'))
+
+import ansible.module_utils.json_utils
+import ansible_mitogen.runner
+
 
 LOG = logging.getLogger(__name__)
 
diff --git a/docs/ansible.rst b/docs/ansible.rst
index 6ab6626a..919c6380 100644
--- a/docs/ansible.rst
+++ b/docs/ansible.rst
@@ -118,6 +118,7 @@ Testimonials
   strategy took Clojars' Ansible runs from **14 minutes to 2 minutes**. I still
   can't quite believe it."
 
+* "Enabling the mitogen plugin in ansible feels like switching from floppy to SSD"
 
 
 .. _noteworthy_differences:
diff --git a/docs/api.rst b/docs/api.rst
index 6efca6dd..92abe1ec 100644
--- a/docs/api.rst
+++ b/docs/api.rst
@@ -831,8 +831,8 @@ Context Class
         context's main thread.
 
         :param fn:
-            A free function in module scope, or a classmethod or staticmethod
-            of a class directly reachable from module scope:
+            A free function in module scope or a class method of a class
+            directly reachable from module scope:
 
             .. code-block:: python
 
@@ -842,10 +842,6 @@ Context Class
                     """A free function reachable as mymodule.my_func"""
 
                 class MyClass:
-                    @staticmethod
-                    def my_staticmethod():
-                        """Reachable as mymodule.MyClass.my_staticmethod"""
-
                     @classmethod
                     def my_classmethod(cls):
                         """Reachable as mymodule.MyClass.my_classmethod"""
diff --git a/docs/changelog.rst b/docs/changelog.rst
index 14eeaae6..fd098e6a 100644
--- a/docs/changelog.rst
+++ b/docs/changelog.rst
@@ -24,6 +24,8 @@ Release Notes
       supported under Ansible 2.6. Contributed by `Dan Quackenbush
       <https://github.com/danquack>`_.
 
+    * Compatible with development versions of Ansible post https://github.com/ansible/ansible/pull/41749
+
 
 v0.2.2 (2018-07-26)
 -------------------
diff --git a/docs/compared.rst b/docs/compared.rst
deleted file mode 100644
index b75ae3f2..00000000
--- a/docs/compared.rst
+++ /dev/null
@@ -1,226 +0,0 @@
-
-Mitogen Compared To
--------------------
-
-This provides a little free-text summary of conceptual differences between
-Mitogen and other tools, along with some basic perceptual metrics (project
-maturity/age, quality of tests, function matrix)
-
-
-Ansible
-#######
-
-Ansible_ is a complete provisioning system, Mitogen is a small component of such a system.
-
-You should use Ansible if ...
-
-You should not use Ansible if ...
-
-
-.. _Ansible: https://docs.ansible.com/ansible/latest/index.html
-.. _ansible.src: https://github.com/ansible/ansible/
-
-Baker
-#####
-
-    Baker_ lets you easily add a command line interface to your Python
-    functions using a simple decorator, to create scripts with "sub-commands",
-    similar to Django's ``manage.py``, ``svn``, ``hg``, etc.
-
-- Unmaintained since 2015
-- No obvious remote execution functionality
-
-.. _Baker: https://bitbucket.org/mchaput/baker
-
-Chopsticks
-##########
-
-Chopsticks_ also supports recursion! but the recursively executed instance has no special knowledge of its identity in a tree structure, and little support for functions running in the master to directly invoke functions in a recursive context.. effectively each recursion produces a new master, from which function calls must be made.
-
-executing functions from __main__ entails picking just that function and deps
-out of the main module, not transferring the module intact. that approach works
-but it's much messier than just arranging for __main__ to be imported and
-executed through the import mechanism.
-
-supports sudo but no support for require_tty or typing a sudo password. also supports SSH and Docker.
-
-good set of tests
-
-real PEP-302 module loader, but doesn't try to cope with master also relying on
-a PEP-302 module loader (e.g. py2exe).
-
-Based on the tox configuration Python 2.7, and 3.3 to 3.6 are supported.
-
-I/O multiplexer in the master, but not in children.
-
-As with Execnet it includes its own serialization - pencode_ supports
-
-- most Python primitive types (``bytes``/``str``/``unicode``, ``list``, ``tuple`` ...)
-- identity references
-- self referencing (recursive) data srtuctures
-
-pencode lacks support for arbitrary classes. Byte strings require special
-treatment if they contain non-ascii characters. Some primitive types
-(e.g. ``complex``) are not handled. This would be straightforwar to address.
-Values are length-prefixed with a 32 bit unsigned integer, meaning values
-are limited to 4 billion bytes or items in length.
-
-design is reminiscent of Mitogen in places (Tunnel is practically identical to
-Mitogen's Stream), and closer to Execnet elsewhere (lack of uniformity,
-tendency to prefer logic expressed in if/else special case soup rather than the
-type system, though some of that is due to supporting Python 3, so not judging
-too harshly!)
-
-Chopsticks has its own `Chopsticks vs`_ comparisons.
-
-You should use Chopsticks if you need Python 3 support.
-
-.. _Chopsticks: https://chopsticks.readthedocs.io/en/stable/
-.. _Chopsticks.src: https://github.com/lordmauve/chopsticks/
-.. _Chopsticks vs: https://chopsticks.readthedocs.io/en/stable/intro.html#chopsticks-vs
-.. _pencode: https://github.com/lordmauve/chopsticks/blob/master/doc/pencode.rst
-.. _pencode.src: https://github.com/lordmauve/chopsticks/blob/master/chopsticks/pencode.py
-
-Disco
-#####
-
-    Disco_ is a lightweight, open-source framework for distributed computing
-    based on the MapReduce paradigm.
-
-- An Erlang core, with Python bindings
-- Wire format is pickle, according to `Execnet vs NLTK for distributed NLTK`_
-
-.. _Disco: http://discoproject.org/
-.. _Execnet vs NLTK for distributed NLTK: https://streamhacker.com/2009/12/14/execnet-disco-distributed-nltk/
-
-Execnet
-#######
-
-Execnet_
-
-- Parent and children may use threads, gevent, or eventlet, Mitogen only supports threads.
-- No recursion
-- Similar Channel abstraction but better developed.. includes waiting for remote to close its end
-- Heavier emphasis on passing chunks of Python source code around, modules are loaded one-at-a-time with no dependency resolution mechanism
-- Built-in unidirectional rsync-alike, compared to Mitogen's SSH emulation which allows use of real rsync in any supported mode
-- no support for sudo, but supports connecting to vagrant
-- works with read-only filesystem
-- includes its own serialization_ independent of the standard library
-
-      The obj and all contained objects must be of a builtin python type
-      (so nested dicts, sets, etc. are all ok but not user-level instances).
-
-- Known uses include `pytest-xdist`_, and `Distributed NLTK`_
-
-You should use Execnet if you value code maturity more than featureset.
-
-.. _Execnet: https://codespeak.net/execnet/
-.. _serialization: https://codespeak.net/execnet/basics.html#dumps-loads
-.. _pytest-xdist: https://pypi.python.org/pypi/pytest-xdist
-.. _Distributed NLTK: https://streamhacker.com/2009/12/14/execnet-disco-distributed-nltk/
-
-Fabric
-######
-
-Fabric_ allows execution of shell snippets on remote machines, Python functions run
-locally, any remote interaction is fundamentally done via shell, with all the
-limitations that entails. prefers to depend on SSH features (e.g. tunnelling)
-than reinvent them
-
-You should use Fabric if you enjoy being woken at 4am to pages about broken
-shell snippets.
-
-.. _fabric: http://www.fabfile.org/
-
-Invoke
-######
-
-Invoke_
-
-Python 2.6+, 3.3+
-
-Basically a Fabric-alike
-
-.. _invoke: http://www.pyinvoke.org/
-
-Multiprocessing
-###############
-
-multiprocessing_ was added to the stdlib in Python 2.6.
-
-    multiprocessing is a package that supports spawning processes using an
-    API similar to the threading module. The multiprocessing package offers
-    both local and remote concurrency
-
-There is a backport_ for Python 2.4 & 2.5, but it is not pure Python.
-pymultiprocessing_ appears to be a pure Python implementation.
-An ecosystem_ of packages has built up around multiprocessing.
-
-The `programming guidelines`_ section notes
-
-- Arguments to proxies must be picklable. On Windows this also applies to
-  ``multiprocessing.Process.__init__()`` arguments.
-- Callers should beware replacing ``sys.stdin``, because
-  ``multiprocessing.Process._bootstrap()``
-  will close it and open /dev/null instead
-
-.. _programming guidelines: https://docs.python.org/2/library/multiprocessing.html#programming-guidelines
-.. _backport: https://pypi.python.org/pypi/multiprocessing
-.. _pymultiprocessing: https://pypi.python.org/pypi/pymultiprocessing
-.. _ecosystem: https://pypi.python.org/pypi?%3Aaction=search&term=multiprocessing&submit=search
-
-Paver
-#####
-
-Paver_
-
-More or less another task execution framework / make-alike, doesn't really deal
-with remote execution at all.
-
-.. _Paver: https://github.com/paver/paver/
-
-Plumbum
-#######
-
-Plumbum_
-
-Shell-only
-
-Basically syntax sugar for running shell commands. Nicer than raw shell
-(depending on your opinions of operating overloading), but it's still shell.
-
-.. _Plumbum: https://pypi.python.org/pypi/plumbum
-
-Pyro4
-#####
-
-Pyro4_
-...
-
-.. _Pyro4: https://pythonhosted.org/Pyro4/
-
-RPyC
-####
-
-RPyC_
-
-- supports transparent object proxies similar to Pyro (with all the pain and suffering hidden network IO entails)
-- significantly more 'frameworkey' feel
-- runs multiplexer in a thread too?
-- bootstrap over SSH only, no recursion and no sudo
-- requires a writable filesystem
-
-.. _RPyC: https://rpyc.readthedocs.io/en/latest/
-
-Salt
-####
-
-Salt_
-
-- no crappy deps
-
-You should use Salt if you enjoy firefighting endless implementation bugs,
-otherwise you should prefer Ansible.
-
-.. _Salt: https://docs.saltstack.com/en/latest/topics/
-.. _Salt.src: https://github.com/saltstack/salt
diff --git a/mitogen/core.py b/mitogen/core.py
index dd706311..cfcd70d2 100644
--- a/mitogen/core.py
+++ b/mitogen/core.py
@@ -350,6 +350,7 @@ def enable_profiling():
         try:
             return func(*args)
         finally:
+            profiler.dump_stats('/tmp/mitogen.%d.%s.pstat' % (os.getpid(), name))
             profiler.create_stats()
             fp = open('/tmp/mitogen.stats.%d.%s.log' % (os.getpid(), name), 'w')
             try:
diff --git a/mitogen/fork.py b/mitogen/fork.py
index 12bb7dfa..cf769788 100644
--- a/mitogen/fork.py
+++ b/mitogen/fork.py
@@ -75,6 +75,17 @@ def reset_logging_framework():
     ]
 
 
+def on_fork():
+    """
+    Should be called by any program integrating Mitogen each time the process
+    is forked, in the context of the new child.
+    """
+    reset_logging_framework()  # Must be first!
+    fixup_prngs()
+    mitogen.core.Latch._on_fork()
+    mitogen.core.Side._on_fork()
+
+
 def handle_child_crash():
     """
     Respond to _child_main() crashing by ensuring the relevant exception is
@@ -134,10 +145,7 @@ class Stream(mitogen.parent.Stream):
             handle_child_crash()
 
     def _child_main(self, childfp):
-        reset_logging_framework()  # Must be first!
-        fixup_prngs()
-        mitogen.core.Latch._on_fork()
-        mitogen.core.Side._on_fork()
+        on_fork()
         if self.on_fork:
             self.on_fork()
         mitogen.core.set_block(childfp.fileno())
diff --git a/tests/ansible/ansible.cfg b/tests/ansible/ansible.cfg
index 7bf849d5..d9224ab7 100644
--- a/tests/ansible/ansible.cfg
+++ b/tests/ansible/ansible.cfg
@@ -7,7 +7,7 @@ callback_plugins = lib/callback
 stdout_callback = nice_stdout
 vars_plugins = lib/vars
 library = lib/modules
-# module_utils = lib/module_utils
+module_utils = lib/module_utils
 retry_files_enabled = False
 forks = 50
 
diff --git a/tests/ansible/gcloud/controller.yml b/tests/ansible/gcloud/controller.yml
index 48f233d9..b4ce3fcf 100644
--- a/tests/ansible/gcloud/controller.yml
+++ b/tests/ansible/gcloud/controller.yml
@@ -56,6 +56,10 @@
         editable: true
         name: ~/ansible
 
+    - pip:
+        virtualenv: ~/venv
+        name: debops
+
     - lineinfile:
         line: "source $HOME/venv/bin/activate"
         path: ~/.profile
diff --git a/tests/ansible/integration/all.yml b/tests/ansible/integration/all.yml
index 264ae716..bf534aed 100644
--- a/tests/ansible/integration/all.yml
+++ b/tests/ansible/integration/all.yml
@@ -9,7 +9,7 @@
 - import_playbook: connection_loader/all.yml
 - import_playbook: context_service/all.yml
 - import_playbook: local/all.yml
-#- import_playbook: module_utils/all.yml
+- import_playbook: module_utils/all.yml
 - import_playbook: playbook_semantics/all.yml
 - import_playbook: remote_tmp/all.yml
 - import_playbook: runner/all.yml
diff --git a/tests/ansible/integration/module_utils/all.yml b/tests/ansible/integration/module_utils/all.yml
index 920b5d1c..c8b8f2fb 100644
--- a/tests/ansible/integration/module_utils/all.yml
+++ b/tests/ansible/integration/module_utils/all.yml
@@ -1,6 +1,6 @@
 
-- import_playbook: from_config_path.yml
-- import_playbook: from_config_path_pkg.yml
-- import_playbook: adjacent_to_playbook.yml
+#- import_playbook: from_config_path.yml
+#- import_playbook: from_config_path_pkg.yml
+#- import_playbook: adjacent_to_playbook.yml
 - import_playbook: adjacent_to_role.yml
-- import_playbook: overrides_builtin.yml
+#- import_playbook: overrides_builtin.yml
diff --git a/tests/bench/roundtrip.py b/tests/bench/roundtrip.py
index 40582d46..13b9413d 100644
--- a/tests/bench/roundtrip.py
+++ b/tests/bench/roundtrip.py
@@ -12,6 +12,6 @@ def do_nothing():
 def main(router):
     f = router.fork()
     t0 = time.time()
-    for x in xrange(1000):
+    for x in xrange(10000):
         f.call(do_nothing)
     print '++', int(1e6 * ((time.time() - t0) / (1.0+x))), 'usec'