With the latest change on how we use the HSDir index, the client and service
need to pick their responsible HSDir differently that is depending on if they
are before or after a new time period.
The overlap mode is active function has been renamed for this and test added.
Signed-off-by: David Goulet <dgoulet@torproject.org>
Because of #23387, we've realized that there is one scenario that makes
the client unable to reach the service because of a desynch in the time
period used. The scenario is as follows:
+------------------------------------------------------------------+
| |
| 00:00 12:00 00:00 12:00 00:00 12:00 |
| SRV#1 TP#1 SRV#2 TP#2 SRV#3 TP#3 |
| |
| $==========|-----------$===========|-----------$===========| |
| ^ ^ |
| C S |
+------------------------------------------------------------------+
In this scenario the HS has a newer consensus than the client, and the
HS just moved to the next TP but the client is still stuck on the old
one. However, the service is not in any sort of overlap mode so it
doesn't cover the old TP anymore, so the client is unable to fetch a
descriptor.
We've decided to solve this by extending the concept of overlap period
to be permanent so that the service always publishes two descriptors and
aims to cover clients with both older and newer consensuses. See the
spec patch in #23387 for more details.
Based on our #23387 findings, it seems like to maintain 24/7
reachability we need to employ different logic when computing hsdir
indices for fetching vs storing. That's to guarantee that the client
will always fetch the current descriptor, while the service will always
publish two descriptors aiming to cover all possible edge cases.
For more details see the next commit and the spec branch.
Signed-off-by: David Goulet <dgoulet@torproject.org>
Use the valid_after time from the consensus to get the time period number else
we might get out of sync with the overlap period that uses valid_after.
Make it an optional feature since some functions require passing a
specific time (like hs_get_start_time_of_next_time_period()).
Signed-off-by: David Goulet <dgoulet@torproject.org>
When merging #20657, somehow hs_service_dir_info_changed() became unused
leading to not use the re-upload to HSDir when we were missing information
feature.
Turns out that it is not possible to pick an HSDir with a missing descriptor
because in order to compute the HSDir index, the descriptor is mandatory to
have so we can know its position on the hashring.
This commit removes that dead feature and fix the
hs_service_dir_info_changed() not being used.
Signed-off-by: David Goulet <dgoulet@torproject.org>
We used to check if it was set to 0 which is what unused circuit have but when
the rendezvous circuit was cannibalized, the timestamp_dirty is not 0 but we
still need to reset it so we can actually use it without having the chance of
expiring the next second (or very soon).
Fixes#23123
Signed-off-by: David Goulet <dgoulet@torproject.org>
This fixes a serious bug in our hsdir set change logic:
We used to add nodes in the list of previous hsdirs everytime we
uploaded to a new hsdir and we only cleared the list when we built a new
descriptor. This means that our prev_hsdirs list could end up with 7
hsdirs, if for some reason we ended up uploading our desc to 7 hsdirs
before rebuilding our descriptor (e.g. this can happen if the set of
hsdirs changed).
After our previous hdsir set had 7 nodes, then our old algorithm would
always think that the set has changed since it was comparing a smartlist
with 7 elements against a smartlist with 6 elements.
This commit fixes this bug, by clearning the prev_hsdirs list before we
upload to all hsdirs. This makes sure that our prev_hsdirs list always
contains the latest hsdirs!
Our logic for detecting hsdir set changes was needlessly compicated: we
had to sort smartlists and compare them.
Instead, we can simplify things by employing the following logic:
"We should reupload our descriptor if the latest HSDir set contains
nodes that were not previously there"
We used to not copy the state which means that after HUP we would forget
if we are in overlap mode or not. That caused bugs where the service
would enter overlap mode twice, and rotate its descs twice, causing all
sorts of bugs.
Apart from the fact that a newly allocated service doesn't have descriptors
thus the move condition can never be true, the service needs the descriptor
signing key to cross-certify the authentication key of each intro point so we
need to move the descriptors between services and not only the intro points.
Fixes#23056
Signed-off-by: David Goulet <dgoulet@torproject.org>
We refactor the descriptor reupload logic to be similar to the v2 logic
where we update a global 'consider_republishing_rend_descriptors' flag
and then we use that to check for hash ring changes during the global
hidden service callbacks.
This fixes bugs where we would inspect the hash ring immediately as we
receive new dirinfo (e.g. consensus) but before running the hidden
service housekeeping events. That was leaving us in an inconsistent
state wrt hsdir indices and causing bugs all around.
The problem was that when we went from overlap mode to non-overlap mode,
we were not wiping the 'desc_next' descriptor and instead we left it on
the service. This meant that all functions that iterated service
descriptors were also inspecting the useless 'desc_next' descriptor that
should have been deleted.
This commit refactors rotate_all_descriptors() so that it rotates
descriptor both when entering overlap mode and also when leaving it.
When we enter overlap mode we start using the next hsdir index of
relays. However, we only compute the next hsdir index of relays when we
receive a consensus or their descriptor. This means that there is a
window of time between entering the overlap period and fetching the
consensus where relays have their next hsdir index uninitialized. This
patch fixes this by recomputing all hsdir indices when we first enter
the overlap period.
We want to reupload our descriptor if its set of responsible HSDirs
changed to minimize reachability issues.
This patch adds a callback everytime we get new dirinfo which checks if
the hash ring changed and reuploads descriptor if needed.
Our Windows compiler treats "time_t" as long long int but Linux likes it
long int so cast those to make Windows happy.
Signed-off-by: David Goulet <dgoulet@torproject.org>
- Fix various ssize_t/size_t confusions in the tests.
- Fix a weird memset argument:
"bad_memset: Argument -16 in memset loses precision in
memset(&desc_two->blinded_kp.pubkey.pubkey, -16, 32UL)."
- Fix check_after_deref instance in check_state_line_for_service_rev_counter():
"check_after_deref: Null-checking items suggests that it may be null,
but it has already been dereferenced on all paths leading to the
check."
I repurposed the old directory_request_set_hs_ident() into a new
directory_request_upload_set_hs_ident() which is only used for the
upload purpose and so it can assert on the dir_purpose.
When coding the client-side we can make a second function for fetch.
- Fix log message format string.
- Do extra circuit purpose check.
- wipe memory in a clear function
- Make sure we don't double add intro points in our list
- Make sure we don't double close intro circuits.
- s/tt_u64_op/tt_i64_op/
Make this function public so we can use it both in hs_circuit.c and
hs_service.c to avoid code duplication.
Signed-off-by: David Goulet <dgoulet@torproject.org>
Change the timing for intro point's lifetime and maximum amount of circuit we
are allowed to launch in a TestingNetwork. This is particurlarly useful for
chutney testing to test intro point rotation.
Signed-off-by: David Goulet <dgoulet@torproject.org>
When the circuit is about to be freed which has been marked close before, for
introduction circuit we now call this has_closed() callback so we can cleanup
any introduction point that have retried to many times or at least flag them
that their circuit is not established anymore.
Signed-off-by: David Goulet <dgoulet@torproject.org>
