bug#49449: 28: TLS connection never gets to "open" stage

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

10 juli 2021 kl. 21.31 skrev Eli Zaretskii <eliz@gnu.org>:

> That's my question: why cannot the sentinel be called in this case?
> what prevents it from being called?

In the failing case, wait_reading_process_output calls gnutls_try_handshake early on, which succeeds and this leads to finish_after_tls_connection being called. Here, we have the condition

  else if ((fd_callback_info[p->outfd].flags & NON_BLOCKING_CONNECT_FD) == 0)

which gates further progress, but this condition is false because the flags have NON_BLOCKING_CONNECT_FD set.

In the successful case, the first call to gnutls_try_handshake from wait_reading_process_output fails because things haven't had the time to be set up yet. This leads to a select being called on the socket for writing (since it's in a nonblocking connect), and when ready, the NON_BLOCKING_CONNECT_FD bit is cleared from the flags.

This is a simplified view. The state is clearly more complex and things need to be done in the proper order.