Reputation: 9494
So I'm using the Python asyncio
module (on Linux) to launch a child process and then asynchronously monitor it. My code works fine... when run on the main thread. But when I run it on a worker thread, it hangs, and the process_exited
callback is never invoked.
I suspect this may actually be some kind of undocumented defect or issue with running subprocess_exec
on a worker thread, likely having to do with how the implementation handles signals in a background thread. But it could also just be me screwing things up.
A simple, reproducible example is as follows:
class MyProtocol(asyncio.SubprocessProtocol):
def __init__(self, done_future):
super().__init__()
self._done_future = done_future
def pipe_data_received(self, fd, data):
print("Received:", len(data))
def process_exited(self):
print("PROCESS EXITED!")
self._done_future.set_result(None)
def run(loop):
done_future = asyncio.Future(loop = loop)
transport = None
try:
transport, protocol = yield from loop.subprocess_exec(
lambda : MyProtocol(done_future),
"ls",
"-lh",
stdin = None
)
yield from done_future
finally:
if transport: transport.close()
return done_future.result()
def run_loop():
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop) # bind event loop to current thread
try:
return loop.run_until_complete(run(loop))
finally:
loop.close()
So here, I setup an asyncio
event loop to execute the shell command ls -lh
, and then trigger a callback for when data is received from the subprocess, and another callback for when the subprocess exits.
If I just call run_loop()
directly in the main thread of a Python program, everything goes fine. But if I say:
t = threading.Thread(target = run_loop)
t.start()
t.join()
Then what happens is that the pipe_data_received()
callback is invoked successfully, but process_exited()
is never invoked, and the program just hangs.
After Googling around and looking at the asyncio
source code for the implementation of unix_events.py
, I discovered it might be necessary to manually attach my event loop to the global "child watcher" object, as follows:
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop) # bind event loop to current thread
asyncio.get_child_watcher().attach_loop(loop)
Apparently, the child watcher is an (undocumented) object that is responsible for calling waitpid
under the hood (or something like that). But when I tried this, and ran run_event_loop()
in a background thread, I got the error:
File "/usr/lib/python3.4/asyncio/unix_events.py", line 77, in add_signal_handler
raise RuntimeError(str(exc))
RuntimeError: set_wakeup_fd only works in main thread
So here it looks like the implementation actually does a check to make sure that signal handlers can only be used on the main thread, leading me to believe that in the current implementation, using subprocess_exec
on a background thread is in fact, simply impossible without changing the Python source code itself.
Am I correct about this? Sadly, the asyncio
module is very under-documented, so it's hard for me to be confident about my conclusion here. I may simply be doing something wrong.
Upvotes: 8
Views: 4014
Reputation: 13415
Handling subprocesses in a worker thread is fine as long as an asyncio loop is running in the main thread with its child watcher instanciated:
asyncio.get_child_watcher()
loop = asyncio.get_event_loop()
coro = loop.run_in_executor(None, run_loop)
loop.run_until_complete(coro)
See this post and the documentation.
Upvotes: 8