After failed tests fixing, of course.

Thank you Andrew! @gvanrossum would be cool if you can take a quick look over this PR, I want to merge it before beta2.

I'm confused. Where is self._transports ever updated? It seems it's always empty, so the check will never fail. What am I missing?

@gvanrossum Transports register when they are being constructed: https://github.com/python/asyncio/pull/420/files#diff-ddadd6cf041d620407e07c7c4f1b0149R572

I see. What about transports that don't inherit from _SelectorTransport? I guess those don't get this error when misused in debug mode but otherwise no harm is done right?

If that's so I think this is fine for b2.

Random thought.
Let's assume I want to make my own transport for some reason.
Honestly I did it only in aiozmq (ZMQ sockets are very different from regular ones, you know) only but maybe somebody will need it for TCP sockets for some reason.
Who knows, in aiohttp we have own StreamReader for working with HTTP chunks etc.

It would be cool to have loop methods for fd protecting by protect/unprotect calls.
Not in 3.6, sure -- but maybe in 3.7?

@gvanrossum I think we cover all transports except SubprocessTransport, but those don't expose the socket via get_extra_info. BTW, the check is always enabled, even in production mode (keeping a WeakValueDict for transports is cheap).

Honestly I did it only in aiozmq (ZMQ sockets are very different from regular ones, you know) only but maybe somebody will need it for TCP sockets for some reason.

@asvetlov Let's think about that after 3.6 :)

@asvetlov Let's think about that after 3.6 :)

Agree

Committed by hand, closing this PR now.