hs-v3: Client request with missing dirinfo will always timeout
When the SOCKS request is handled in
connection_ap_handle_onion(), if we are missing dirinfo (including missing a live consensus), the connection state is put in "AP_CONN_STATE_CIRCUIT_WAIT".
Then it is retried every second through
connection_ap_handshake_attach_circuit() which does a gazillion things but among those it tries to open the IP/RP circuits. Of course, we can't get an IP because we have no descriptor so
circuit_get_open_circ_or_launch(), which tries to get that IP circuit, won't be able so that function will get an intro point from the descriptor (that doesn't exists because we can't even fetch it) but won't work.
It then triggers a descriptor fetch and then puts the connection in
AP_CONN_STATE_RENDDESC_WAIT. Now, because we don't look at the returned code, we don't know what really happened and if tor still doesn't have enough dirinfo to proceed, the connection is still put in "renddesc wait" making it NOT a pending connection anymore thus never retried after that.
The only way for a connection to get out of that "renddesc wait" state is either timing out (which is what happens for v3) or the descriptor arrives and then there is a callback in the HS client subsystem to handle that desc.
The possible solutions are I believe:
Either we keep the connection in "circuit wait" so it gets retried regularly.
We create a new "AP CONN" state that is "waiting for dirinfo" and then when we get the live consensus or/and minimum dirinfo, we callback the HS subsystem and look for SOCKS conn in that state.
This is not a problem v2 HS suffers because it only care about a "reasonably live consensus".