Corking + starttls

[daurnimator]

If a tcp socket is corked, starttls will block for an extra 200ms (until the cork times out).

The starttls lua wrapper should probably gain an :uncork call.

[daurnimator]

Possible fix for this:

diff --git a/src/socket.lua b/src/socket.lua
index 9513c57..84cbc3a 100644
--- a/src/socket.lua
+++ b/src/socket.lua
@@ -18,6 +18,7 @@ local EPIPE = errno.EPIPE
 local ETIMEDOUT = errno.ETIMEDOUT
 local ENOTCONN = errno.ENOTCONN
 local ENOTSOCK = errno.ENOTSOCK
+local EOPNOTSUPP = errno.EOPNOTSUPP
 local strerror = errno.strerror

 local format = string.format
@@ -288,6 +289,13 @@ local _starttls; _starttls = socket.interpose("starttls", function(self, arg1, a
        local deadline = timeout and monotime() + timeout
        local ok, why = _starttls(self, ctx)

+       do -- May need to uncork so that TLS data is flushed
+               local ok, why = self:uncork()
+               if not ok and why ~= EOPNOTSUPP then
+                       return nil, oops(self, "starttls", why)
+               end
+       end
+
        while not ok do
                if why == EAGAIN then
                        if not timed_poll(self, deadline) then

This does result in unconditional uncorks, which may impact performance. The patch in #21 (comment) helps for non-TCP sockets.

It also makes the relatively wild assumption that the first starttls is going to write something, and will need uncorking. This doesn't harm (as uncorking never really can), but a better solution might be possible.

This doesn't solve the similar corner case of a corked socket during renegotiations; but that may need to be visited separately.

[daurnimator]

It also makes the relatively wild assumption that the first starttls is going to write something, and will need uncorking. This doesn't harm (as uncorking never really can), but a better solution might be possible.

I spent a while trying to find a way to only do this when openssl had written a 'full' record. In theory, we could do it in BIO_flush/BIO_CTRL_FLUSH, but openssl doesn't call flush for the important case of handshakes! (this seems like a major breakage in openssl, and all BIOs that rely on BIO_flush are fatally flawed)

[daurnimator]

Also, instead of using :uncork, we should call so_nodelay on linux, as that triggers a full flush, without uncorking.

[daurnimator]

BIO_flush now gets called correctly in openssl 1.1.0