amqp_ssl_socket_close intermittent crash

[korymiller1489]

We are using the 0.10 build compiled on SCO Unix OS5 with gcc 4.2.4.

We have encountered a crashing issue with our build intermittently during the amqp_destroy_connection -> amqp_ssl_socket_delete ->amqp_ssl_socket_close call. Specifically, the program crashes on the SSL_shutdown call.

static int amqp_ssl_socket_close(void *base, amqp_socket_close_enum force) {
  struct amqp_ssl_socket_t *self = (struct amqp_ssl_socket_t *)base;

  if (-1 == self->sockfd) {
    return AMQP_STATUS_SOCKET_CLOSED;
  }

  if (AMQP_SC_NONE == force) {
    /* don't try too hard to shutdown the connection */
    SSL_shutdown(self->ssl);
  }

  SSL_free(self->ssl);
  self->ssl = NULL;

  if (amqp_os_socket_close(self->sockfd)) {
    return AMQP_STATUS_SOCKET_ERROR;
  }
  self->sockfd = -1;

  return AMQP_STATUS_OK;
}

Admittedly, our env has some network issues with latency and dropped/out of order packets, and the crashes are likely related to corrupted/malformed packets, which we then drop the connection (and thus crash sometimes) and reconnect and continue running. But we do not have a way around the network issues at this time.

Modifying the amqp_ssl_socket_delete call to use AMQP_SC_FORCE (which bypasses the SSL_shutdown and skips to SSL_free) alleviates the problem, but does not seem like a good long term solution.

I noticed that there is no real library mechanism to call an amqp_socket_close_force type of operation. Would it be agreeable to add this logic? I am happy to provide a PR to this end.

[alanxz]

Can you describe what sort of crash you're getting?

[korymiller1489]

We're getting segfaults in the SSL_shutdown function periodically.

We see in the Rabbit logs a corresponding frame header error.

[error] closing AMQP connection {handshake_timeout,frame_header}

we don't crash everytime a frame header error is seen, but everytime the library crashes, we see a frame header error.

Basically, when we detect that the connection has a problem (connection closed etc, network issues) and we need to reconnect, we call amqp_destroy_connection to clean up resources before the next attempt. That call will sometimes segfault, at the chain mentioned above. The crash is always associated with the frame header error in the Rabbit log.