GitLab

In bug #366 (moved) we made it so Tor won't open a second dir fetch to an authority if it has one already. Great.

        rs = router_pick_trusteddirserver(type, pds_flags);
        if (rs == NULL && (pds_flags & (PDS_NO_EXISTING_SERVERDESC_FETCH|
                                        PDS_NO_EXISTING_MICRODESC_FETCH))) {
[...]
            log_debug(LD_DIR, "Deferring serverdesc fetch: all authorities "
                      "are in use.");

But we didn't update it to look for begindir conns, so it only applies to direct dir fetches. Ok.

      if (no_microdesc_fetching) {
        if (connection_get_by_type_addr_port_purpose(
             CONN_TYPE_DIR, &addr, d->dir_port, DIR_PURPOSE_FETCH_MICRODESC)) {
          ++n_busy;
          continue;

So it doesn't apply to clients, only relays. That makes sense, because relays are the ones who typically would contact authorities anyway.

  int prefer_authority = directory_fetches_from_authorities(options);

But when a relay starts up and gets a consensus, it has a line like this nowadays:

Sep 04 05:47:40.000 [info] launch_descriptor_downloads(): Launching 33 requests for 3114 routers, 96 at a time

33 requests! Surely that's way more than the 8 or so authorities we have. And relays don't use begindir to talk to authorities, since it slows them down too much:

  int use_begindir = supports_begindir &&
                     directory_command_should_use_begindir(options, _addr,
                       or_port, router_purpose, anonymized_connection);

Doesn't that mean we hit the "one per authority" limit and drop the rest of those requests?

It turns out that directory_fetches_from_authorities() is false for most relays when they start up:

  if (server_mode(options) && router_pick_published_address(options, &addr)<0)
    return 1; /* we don't know our IP address; ask an authority. */
  refuseunknown = ! router_my_exit_policy_is_reject_star() &&
    should_refuse_unknown_exits(options);
  if (options->DirPort == NULL && !refuseunknown)
    return 0;
  if (!server_mode(options) || !advertised_server_mode())
    return 0;
  me = router_get_my_routerinfo();
  if (!me || (!me->dir_port && !refuseunknown))
    return 0; /* if dirport not advertised, return 0 too */
  return 1;

So these relays end up asking arbitrary other relays they found in the consensus! Cue Nick's circus music here. Not the best way to get fresh info.

In my case here (and I expect it's a common case), my relay failed the "!advertised_server_mode" check, since it hadn't done its reachability test yet so it hadn't published a descriptor yet.

Maybe this is actually a feature that just-starting-up relays don't fetch descriptors from authorities. It probably doesn't hurt much, and probably helps authority load a bit.

But I don't think it's a feature that we allow multiple descriptor-fetching dir requests in parallel to an authority iff they're begindir requests.