Victor, thanks for the message.
I forgot mentioned changes.
You are right, the issue is "theoretical", I didn't see problems with current implementation.
The self-pipe buffer is pretty big, a chance to reach its limit is relatively low.
But a change to break asyncio on some OS with the patch is much higher.
Moreover, the patch assumes that python signal handler will be called *before* reading from self-pipe. 
Otherwise, a signal callback will be postponed up to next writing to the pipe, which looks like a hidden bug.

Let's close the issue.