Long story short, if you have a TCP asyncio-server (implementing asyncio.Protocol) that sends something to socket in connection_made callback, it will leak fds and won't be able to handle some consequent requests, if you will send to it RST right after TCP-handshake.

In my case, these requests are sent by keepalived demon (it makes TCP-checks that way), but they can be easily hand-crafted.

Example server:

import asyncio

class EchoServerClientProtocol(asyncio.Protocol):
    def connection_made(self, transport):
        self.transport = transport
        self.transport.write(b'Say something')

    def data_received(self, data):
        self.transport.write(data)
        self.transport.close()

loop = asyncio.get_event_loop()
coro = loop.create_server(EchoServerClientProtocol, '127.0.0.1', 8888)
server = loop.run_until_complete(coro)
loop.run_forever()

Example client:

import socket
import struct
import time

# first request
sock = socket.socket()
l_onoff = 1
l_linger = 0
sock.setsockopt(socket.SOL_SOCKET, socket.SO_LINGER, struct.pack('ii', l_onoff, l_linger))
sock.connect(('127.0.0.1', 8022))
sock.close()

time.sleep(1)

# second request
sock = socket.socket()
sock.connect(('127.0.0.1', 8888))
print('Got', sock.recv(1024))
sock.sendall(b'Hi there')
print('Got', sock.recv(1024))  # <-- Will hang here
sock.close()

Why is this happening?

1. First request
   1.1. _SelectorSocketTransport.__init__ schedules connection_made and loop.add_reader.
   1.2. connection_made tries to send data to the socket, it fails, _SelectorTransport._fatal_error is called, then _SelectorTransport._force_close is called.
   1.3. _SelectorTransport._force_close removes the reader from the loop (there is none, it's ok), and schedules closing the socket in _SelectorTransport._call_connection_lost.
   1.4. loop.add_reader is called (scheduled on step 1), it calls epoll_ctl, no error is returned because the socket isn't closed yet.
   1.5. the socket is closed by _SelectorTransport._call_connection_lost (scheduled on step 3). The corresponding entry in _SelectorMapping remains and isn't cleaned up.

2. Second request
   2.1. FD from the first request is reused by the OS
   2.2. BaseSelectorEventLoop._add_reader sees that the FD is already in BaseSelectorEventLoop._selector, so it calls _selector.modify
   2.3 _BaseSelectorImpl.modify sees that the events mask is the same, so it uses a shortcut, changing only the callback, but without calling epoll_ctl. The socket won't be polled.

I think the source of the error is in step 1.4, we shouldn't add the reader if we are already closing the transport. I've made a simple PR fixing this, but maybe there are other options here.

New changeset a84d0b3 by Andrew Svetlov (Vlad Starostin) in branch 'master':
bpo-33263: Fix FD leak in _SelectorSocketTransport (GH-6450)
a84d0b3

New changeset b8b8000 by Andrew Svetlov (Miss Islington (bot)) in branch '3.7':
bpo-33263: Fix FD leak in _SelectorSocketTransport (GH-6450) (bpo-7022)
b8b8000

New changeset 7208bfb by Andrew Svetlov in branch '3.6':
[3.6] bpo-33263: Fix FD leak in _SelectorSocketTransport (GH-6450) (bpo-7025)
7208bfb