bug#36017: 27.0.50; TLS 1.3 on macOS exhibits similar issue to #34341

["Mattias Engdegård" <mattiase@acm.org>]

There seems to be a race of some kind.

On macOS 10.14 and Emacs master, NS build, `list-packages' (with only elpa.gnu.org configured) fails while fetching archive-contents asynchronously. However, running

(url-retrieve
 "https://elpa.gnu.org/packages/archive-contents"
 (lambda (status) (message "callback: status = %S" status))
 nil
 'silent)

works (albeit slowly); this is (modulo the callback) what list-packages does behind the curtains. The only difference is that list-packages then immediately proceeds with some other work (package-menu--refresh). This can be simulated with

(progn
 (url-retrieve ...)  ; like above
 (busy-wait 1.0))    ; busy-wait 1 s

which fails in the same way. As reported by Joseph Mingrone, the process sentinel, `url-http-async-sentinel', is called with a WHY argument of "connect", which makes little sense; if the process is still trying to connect, why is the sentinel called?

In any case, `url-http-async-sentinel' interprets the "connect" status as an error and fails, but this is just a symptom of something earlier; changing the sentinel to ignore "connect" doesn't help.

If run in a TTY (with -nw), everything works, further indicating that it's may be a problem with handling of asynchronous processes in the NS build.

However, this seems insufficient by itself: is gnutls (or TLS 1.3) really so brittle that it cannot cope with a mere delay of 1 s during the initial handshake?

More details:

* In all runs, gnutls_handshake fails when called right after the initial connect() in connect_network_socket; this is expected since the socket is nonblocking and the TCP connection hasn't been set up yet.

* In failed runs, gnutls_handshake succeeds immediately in wait_reading_process_output.

* In successful runs, gnutls_handshake initially fails (with GNUTLS_E_AGAIN) in wait_reading_process_output. Only after several tries does it return success.

I didn't make much progress beyond this point. Could someone more familiar with gnutls in Emacs help me debug it?