Turns out people use transports' sockets with add_reader / add_writer low level APIs. I've discovered this while debugging a crash of a webapp deployed with uvloop. The reason of that crash is how aiohttp implements sendfile: they use transport.get_extra_info('socket') to get the underlying socket, and then they use loop.add_writer and loop.remove_writer on that socket.

The crash in uvloop is caused by how file descriptors are stored internally by libuv. I'll make a workaround for that by using os.dup() to return a duplicate socket from transport.get_extra_info('socket').

However, I think we should do this in asyncio too. The thing is that when you use add_writer and remove_writer on the transport's socket, you're messing with the internal state of the transport. For instance, a transport might be in the middle of writing data, and calling remove_writer will cause the whole program to hang.

I see two options:

1. We always return duplicate sockets from transport.get_extra_info(), hiding the actual socket that transport is attached to.
2. We find a way to raise an exception if add|remove_witer|reader is used on a transport's socket.

I'm more inclined to do the 1.

/cc @asvetlov