TLS support and event loops

[Eric Wong <e@80x24.org>]

Eric Wong <e@80x24.org> wrote:
> "Eric W. Biederman" <ebiederm@xmission.com> wrote:
> > Oh that and IO::Socket::SSL.  Something that is currently missing  (or
> > at least missing until recently I haven't check in the last couple of
> > months) is ssl support for our nntp sockets.  So do imap or prevent
> > mischief for our nntp streams we need to use tls.  Which IO::Socket::SSL
> > seems to do.
> 
> Yup, IO::Socket::SSL will be what we use for STARTTLS support,
> not sure when I'll get around to it.
> 
> Fwiw, it's also used by Perlbal for non-blocking HTTPS with
> Danga::Socket.  I plan on keeping PublicInbox::DS close to
> Danga::Socket, for now[1])...
> 
> Using the TLS support in the kernel might be a fun option to
> add, too; but portable stuff, first.

> [1] I've also been thinking about making PublicInbox::DS switch
>     to EPOLLONESHOT/EV_ONESHOT to map better to my mind when
>     distinguishing between SSL_ERROR_WANT_READ/WRITE..  And it
>     would give me an excuse to pick up and dogfood an epoll API
>     to reduce syscall overhead:
>     https://lore.kernel.org/lkml/1393206162-18151-1-git-send-email-n1ght.4nd.d4y@gmail.com/

Or the userspace pollable epoll stuff:
     https://lore.kernel.org/lkml/20190516085810.31077-1-rpenyaev@suse.de/

But yeah, I think the first step would be to convert to
EPOLLONESHOT or EPOLLET to simplify socket state management.
That should make dealing with non-blocking OpenSSL easier.

Fwiw, I never understood the appeal of level-triggering or the
"reactor pattern", but that's what everybody else seems to
pick for event loops.