PR #908 change in _get_interrupt_pair may cause a deadlock

[vitaly-krugl]

The original code returned a non-blocking socket pair, but the pair is blocking after PR #908. See https://github.com/pika/pika/pull/908/files#r160844142.

The change from non-blocking to blocking can lead to a deadlock. The implementation was designed for non-blocking model - see

pika/pika/adapters/select_connection.py

Line 539 in 2b145f4

if err.errno != errno.EWOULDBLOCK:

and

pika/pika/adapters/select_connection.py

Line 651 in 2b145f4

if err.errno != errno.EAGAIN:

.

A simple/pathological case that would reproduce the deadlock is calling stop enough times from code executing within the event loop itself to fill the socket's transmit buffer.

A more complex - albeit still possibly pathological - case is where a background thread or many background threads call stop, while the event loop code is blocked waiting for something from the thread that happens to make the stop call that would exceed the _w_interrupt socket's transmit buffer.

Is there any reason for removing the logic that made both sockets non-blocking? See https://github.com/pika/pika/pull/908/files#diff-69e7c8c211c62cdb127e8ab3324ce7c6L649