Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

resolve_host() is called from init() and issue() when transitioning out
of STATE_UNINITIALIZED and from advance_state() right after the call to
shutdown_socket().  In all three cases the socket should get closed, so
drop the redundant call and place asserts in connect() implementations
instead.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

ssl_stream objects can't be reused after shut down: despite
a successful reconnect and handshake, any attempt to read or write
fails with "end of stream" (beast.http:1) or "protocol is shutdown"
(asio.ssl:337690831) error respectively.  This doesn't appear to be
documented, but Beast and ASIO authors both mention that the stream
must be destroyed and recreated [1][2].

This was missed because the only integration test with a big enough
image used http instead of https.

[1] boostorg/beast#821 (comment)
[2] chriskohlhoff/asio#804 (comment)

Fixes: https://tracker.ceph.com/issues/69178
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

…ake()

With m_ssl_enabled set to false, disconnect() is a no-op.  Since
m_ssl_enabled is flipped to true only when the handshake succeeds,
calling disconnect() on "failed to complete handshake" error is bogus
(as would be attempting to shut down SSL there).

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

The remaining callers of disconnect() call it only when m_ssl_enabled
is set to true (i.e. after the handshake is completed):

- shut_down(), in STATE_READY
- maybe_finalize_reset(), very shortly after transitioning out of
  STATE_READY as part of performing a reset
- advance_state(), on a transition to STATE_READY that is intercepted
  by a previously delayed shut down

m_ssl_enabled isn't used outside of disconnect() and on top of that
is never cleared.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

Get rid of get_callback_adapter() which only obfuscates the error:

  handle_handshake: failed to complete SSL handshake: (337047686) Unknown error 337047686

vs

  handle_handshake: failed to complete SSL handshake: certificate verify failed (SSL routines, tls_process_server_certificate)

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

…wn SSL

Propagate ec to handle_disconnect() and use it to suppress
stream_truncated errors.  Here is a quote from Beast documentation [1]:

  // Gracefully shutdown the SSL/TLS connection
  error_code ec;
  stream.shutdown(ec);
  // Non-compliant servers don't participate in the SSL/TLS shutdown process and
  // close the underlying transport layer. This causes the shutdown operation to
  // complete with a `stream_truncated` error. One might decide not to log such
  // errors as there are many non-compliant servers in the wild.
  if(ec != net::ssl::error::stream_truncated)
      log(ec);

... and a commit that made ignoring stream_truncated safe [2]:

  // ssl::error::stream_truncated, also known as an SSL "short read",
  // indicates the peer closed the connection without performing the
  // required closing handshake
  // [...]
  // When a short read would cut off the end of an HTTP message,
  // Beast returns the error beast::http::error::partial_message.
  // Therefore, if we see a short read here, it has occurred
  // after the message has been completed, so it is safe to ignore it.

[1] https://www.boost.org/doc/libs/develop/libs/beast/doc/html/beast/using_io/ssl_tls_shutdown.html
[2] boostorg/beast@094f5ec

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

If the shutdown gets delayed until the state transition from
STATE_RESET_CONNECTING completes and the reconnect is successful
(i.e. next_state is STATE_READY), we eventually hit "unexpected
state transition" assert in advance_state().  The reason is that
advance_state() would update m_state and call disconnect() under
STATE_READY instead of STATE_SHUTTING_DOWN.  After the disconnect
maybe_finalize_shutdown() would enter advance_state() again with
STATE_SHUTDOWN as next_state, but the transition to that from
STATE_READY is invalid.

Plug this by not transitioning to next_state if current_state is
STATE_SHUTTING_DOWN.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

…nsitions

If shut_down() gets delayed until a) the state transition from
STATE_RESET_CONNECTING completes and the reconnect is unsuccessful or
b) the state transition from STATE_RESET_DISCONNECTING completes (i.e.
next_state is STATE_UNINITIALIZED or STATE_RESET_CONNECTING), the
socket needs to be shut down before m_on_shutdown is invoked.  The line
of thought here is the same as for the corresponding state transitions
that don't involve STATE_SHUTTING_DOWN.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>