Ticket27247

[nmathewson]

No description provided.

[coveralls]

Coverage decreased (-0.0009%) to 59.831% when pulling 33a0c61 on nmathewson:ticket27247 into 9f0e8d8 on torproject:master.

[teor2345]

Hi, this branch looks good, and I think it's fine to revert parts of e5c608e - clients shouldn't need to cache consensuses in RAM, because they'll never serve them. I also checked that routerstatus_format_entry() only uses parsed descriptors and routerstatuses.

I would happily merge this as-is, but I have some questions.

Microdesc Consensus GETINFO

Is there a microdesc-consensus equivalent for GETINFO dir/status-vote/current/consensus?
(I couldn't find one.)

Because this branch drops all flavours of cached consensus, but only refactors the GETINFO code for the ns consensus.

Consensus Request Headers

When we are setting the consensus request headers, we use the parsed consensus, then fall back to the RAM cached consensus:

tor/src/feature/dircache/directory.c

Lines 560 to 593 in 5a2374b

	if (flav != -1) {
	/* IF we have a parsed consensus of this type, we can do an
	* if-modified-time based on it. */
	networkstatus_t *v;
	v = networkstatus_get_latest_consensus_by_flavor(flav);
	if (v) {
	/* In networks with particularly short V3AuthVotingIntervals,
	* ask for the consensus if it's been modified since half the
	* V3AuthVotingInterval of the most recent consensus. */
	time_t ims_delay = DEFAULT_IF_MODIFIED_SINCE_DELAY;
	if (v->fresh_until > v->valid_after
	&& ims_delay > (v->fresh_until - v->valid_after)/2) {
	ims_delay = (v->fresh_until - v->valid_after)/2;
	}
	if_modified_since = v->valid_after + ims_delay;
	if (v->valid_after >= approx_time() - max_age_for_diff) {
	memcpy(or_diff_from, v->digest_sha3_as_signed, DIGEST256_LEN);
	or_diff_from_is_set = 1;
	}
	}
	} else {
	/* Otherwise it might be a consensus we don't parse, but which we
	* do cache. Look at the cached copy, perhaps. */
	cached_dir_t *cd = dirserv_get_consensus(resource);
	/* We have no method of determining the voting interval from an
	* unparsed consensus, so we use the default. */
	if (cd) {
	if_modified_since = cd->published + DEFAULT_IF_MODIFIED_SINCE_DELAY;
	if (cd->published >= approx_time() - max_age_for_diff) {
	memcpy(or_diff_from, cd->digest_sha3_as_signed, DIGEST256_LEN);
	or_diff_from_is_set = 1;
	}
	}
	}

Should we add a further fall back to load and parse the consensus file before every consensus request? The fallback would only happen on clients with FetchUselessDescriptors 1.

If we do add this fallback, it will save bandwidth, but cost CPU and RAM.

(How many clients set FetchUselessDescriptors 1? Are there a significant number? If so, we should try to save this bandwidth. If not, it probably doesn't matter.)

Or is there some sensible default published time we could use for a consensus file?

[teor2345]

We could add dir_server_mode() to the conditions in if (we_want_to_fetch_flavor(get_options(), FLAV_NS)) {.

But I think the current code works as-is, and it is is more future-proof.