Block TLS handshake until TCP connection established

If a TLS handshake is attempted before the completion of an
asynchronous TCP connection has been ascertained, our local state will
not be set up correctly for further progress and the sentinel "open"
event will never be sent.  This can occur if sufficient time passes
after the initiation of an async TCP connection so that by the time
`wait_reading_process_output` is called, the connection has already
been established on the TCP level.

This somewhat timing-sensitive bug has plagued HTTPS connections on
some platforms, notably macOS, for a long time (bug#49449).

* src/process.c (wait_reading_process_output): Gate the TLS handshake
by the NON_BLOCKING_CONNECT_FD flag.  The flag will be cleared as soon
as the TCP socket is found to be writable.
* test/src/process-tests.el (process-async-https-with-delay):
New test.

1 files changed, 4 insertions, 1 deletions

diff --git a/src/process.c b/src/process.c
index b8c3e4ecfbc..c3186eed750 100644
--- a/src/process.c
+++ b/src/process.c
@@ -5232,7 +5232,10 @@ wait_reading_process_output (intmax_t time_limit, int nsecs, int read_kbd,
 #ifdef HAVE_GNUTLS
                 /* Continue TLS negotiation. */
                 if (p->gnutls_initstage == GNUTLS_STAGE_HANDSHAKE_TRIED
-                    && p->is_non_blocking_client)
+                    && p->is_non_blocking_client
+                    /* Don't proceed until we have established a connection. */
+                    && !(fd_callback_info[p->outfd].flags
+                         & NON_BLOCKING_CONNECT_FD))
                   {
                     gnutls_try_handshake (p);
                     p->gnutls_handshakes_tried++;