Fix SSLError event processing.

[kherrala]

Do not try to use IOLoop to handle socket events.
Poll for events using select.
Catch socket timeout exception.
Remove unused class variable DO_HANDSHAKE.

[coveralls]

Coverage increased (+0.13%) when pulling 8dab34a on kherrala:master into f8c263f on pika:master.

[kherrala]

It seems there's another bug handling ssl.SSL_ERROR_WANT_READ in _handler_error(). Current behaviour is to disconnect immediately. I seem to be having all sorts of strange disconnect behaviour running on weak 3G connection.

[coveralls]

Coverage increased (+0.08%) when pulling c3f390c on kherrala:master into e27b537 on pika:master.

[kherrala]

Sure if you can make it work that way, but the current behaviour is to loop eternally doing the handshake and never return control back to running the IOLoop. The other problem was not catching socket exceptions during handshake, which is not handled well by the caller.

[kherrala]

Any plans getting this to master? It has been working fine for me. Using standard select() during initial handshake doesn't seem like a big issue to me.

[vitaly-krugl]

Using select() here explicitly will interfere with the native framework's (e.g., Tornado) concurrency model. ***But the major issue with forcing the use of select() here is that it has serious limitations with respect to max socket file descriptor number; so, this would cause failures (or require changes) in programs that create many thousands of sockets and other file descriptor-based objects. poll/epoll/etc. don't have that problem.

[gmr]

Closing due to not using native adapters for behavior which seems wrong. Thanks for your patch. If you want to (and it's still needed), please resubmit without the select based implementation, based upon current master.