uvicorn reload fails on Windows and Python 3.8

[f0ff886f]

I know this was referred to in #477 and #509 but it still isn't solved (making a new issue as I hijacked a closed one before).

Using uvicorn checked out 996aa2d, running with or without --loop asyncio changes nothing (that code was removed):

ERROR:    Accept failed on a socket
socket: <asyncio.TransportSocket fd=764, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=6, laddr=('127.0.0.1', 8000)>
Traceback (most recent call last):
  File "C:\Users\f\scoop\apps\python\current\lib\asyncio\proactor_events.py", line 817, in loop
    f = self._proactor.accept(sock)
  File "C:\Users\f\scoop\apps\python\current\lib\asyncio\windows_events.py", line 545, in accept
    self._register_with_iocp(listener)
  File "C:\Users\f\scoop\apps\python\current\lib\asyncio\windows_events.py", line 714, in _register_with_iocp
    _overlapped.CreateIoCompletionPort(obj.fileno(), self._iocp, 0, 0)
OSError: [WinError 87] The parameter is incorrect

Python 3.8.0 (tags/v3.8.0:fa919fd, Oct 14 2019, 19:37:50) [MSC v.1916 64 bit (AMD64)] on win32

[f0ff886f]

This does not repro on py37, reloading works fine there.

Python 3.7.5 (tags/v3.7.5:5c02a39a0b, Oct 15 2019, 00:11:34) [MSC v.1916 64 bit (AMD64)] on win32

[tomchristie]

Right, so the problem here is to do with IOCP.

In Python 3.8 the default loop policy changes on Windows to use IOCP by default. We probably want to explcitly turn this off (and remove our loop="iocp" support completly, since it appears to not work.)

Does this only occur when using multiple workers / reload, or does it always occur on Python 3.8 with Windows?

[euri10]

I can trigger the same on windows and 3.7.5
passing --loop="asyncio" is fine on 3.7.5
passing --loop="asyncio" on 3.8.0 the bug still here

[f0ff886f]

Does this only occur when using multiple workers / reload, or does it always occur on Python 3.8 with Windows?

I'm not sure I understand the question. If I launch uvicorn with --reload and just use it (without changing code), it works OK with no errors. As soon as the reload is requested all hell breaks loose.

[euri10]

feels like it's more about the way sockets are passed around in child processes since it happens also on asyncio loop and only with reload
sorry cant trigger on 3.7.5

[euri10]

summary:
3.7.5
no bug with
C:/Users/benoi/.virtualenvs/uvicorn/Scripts/uvicorn-script.py toto:app --reload
C:/Users/benoi/.virtualenvs/uvicorn/Scripts/uvicorn-script.py toto:app --reload --loop=asyncio
bug with
C:/Users/benoi/.virtualenvs/uvicorn/Scripts/uvicorn-script.py toto:app --reload --loop=iocp

3.8
bug with all
C:/Users/benoi/.virtualenvs/uvicorn3.8/Scripts/uvicorn-script.py toto:app --reload
C:/Users/benoi/.virtualenvs/uvicorn3.8/Scripts/uvicorn-script.py toto:app --reload --loop=asyncio
C:/Users/benoi/.virtualenvs/uvicorn3.8/Scripts/uvicorn-script.py toto:app --reload --loop=iocp

[euri10]

note also that

uvicorn/uvicorn/subprocess.py

Lines 33 to 36 in a75fe13

	if platform.system() == "Windows":
	from multiprocessing.resource_sharer import DupSocket
	sockets = [DupSocket(s) for s in sockets]

and

uvicorn/uvicorn/subprocess.py

Lines 65 to 66 in a75fe13

	if platform.system() == "Windows":
	sockets = [s.detach() for s in sockets]

can be replaced by multiprocessing.allow_connection_pickling() which doesnt solve the 3.8 bug but simplifies things

[tomchristie]

3.8 - bug with all

Indeed. The problem is that loop=asyncio will use the default loop policy for asyncio.
On windows that's IOCP under 3.8, which appears to have issues with sharing sockets between processes.

To resolve this we should:

• Drop our IOCP support completely.
• Force the loop policy to SelectorEventLoop on Windows with 3.8+. https://docs.python.org/3/library/asyncio-eventloop.html?highlight=iocp#asyncio.SelectorEventLoop

[euri10]

seems like that did the trick

C:\Users\benoi\.virtualenvs\uvicorn3.8\Scripts\python.exe C:/Users/benoi/.virtualenvs/uvicorn3.8/Scripts/uvicorn-script.py toto:app --reload --log-level=debug
INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)
INFO:     Started reloader process [6740]
DEBUG:    Using selector: SelectSelector
INFO:     Started server process [8380]
INFO:     <Server sockets=(<asyncio.TransportSocket fd=712, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=6, laddr=('127.0.0.1', 8000)>,)> is serving
INFO:     Waiting for application startup.
INFO:     ASGI 'lifespan' protocol appears unsupported.
INFO:     Application startup complete.
WARNING:  Detected file change in 'toto.py'. Reloading...
DEBUG:    Using selector: SelectSelector
INFO:     Started server process [2684]
INFO:     <Server sockets=(<asyncio.TransportSocket fd=540, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=6, laddr=('127.0.0.1', 8000)>,)> is serving
INFO:     Waiting for application startup.
INFO:     ASGI 'lifespan' protocol appears unsupported.
INFO:     Application startup complete.