We used to not set the guard state in launch_direct_bridge_descriptor_fetch().
So when a bridge descriptor fetch failed, the guard subsystem would never
learn about the fail (and hence the guard's reachability state would not
be updated).
When calculating max sampled size, Tor would only count the number of
bridges in torrc, without considering that our state file might already
have sampled bridges in it. This caused problems when people swap
bridges, since the following error would trigger:
[warn] Not expanding the guard sample any further; just hit the
maximum sample threshold of 1
In that chutney test, the bridge client is configured to connect to
the same bridge at 127.0.0.1:5003 _and_ at [::1]:5003, with no
change in transports.
That meant, I think, that the descriptor is only assigned to the
first bridge when it arrives, and never the second.
- Make sure we check at least two guards for descriptor before making
circuits. We typically use the first primary guard for circuits, but
it can also happen that we use the second primary guard (e.g. if we
pick our first primary guard as an exit), so we should make sure we
have descriptors for both of them.
- Remove BUG() from the guard_has_descriptor() check since we now know
that this can happen in rare but legitimate situations as well, and we
should just move to the next guard in that case.
Previously I'd made a bad assumption in the implementation of
prop271 in 0.3.0.1-alpha: I'd assumed that there couldn't be two
guards with the same identity. That's true for non-bridges, but in
the bridge case, we allow two bridges to have the same ID if they
have different addr:port combinations -- in order to have the same
bridge ID running multiple PTs.
Fortunately, this assumption wasn't deeply ingrained: we stop
enforcing the "one guard per ID" rule in the bridge case, and
instead enforce "one guard per <id,addr,port>".
We also needed to tweak our implementation of
get_bridge_info_for_guard, since it made the same incorrect
assumption.
Fixes bug 21027; bugfix on 0.3.0.1-alpha.
Since we can call this function more than once before we update all
the confirmed_idx fields, we can't rely on all the relays having an
accurate confirmed_idx.
Fixes bug 21129; bugfix on 0.3.0.1-alpha
In addition to not wanting to build circuits until we can see most
of the paths in the network, and in addition to not wanting to build
circuits until we have a consensus ... we shouldn't build circuits
till all of our (in-use) primary guards have descriptors that we can
use for them.
This is another bug 21242 fix.
Actually, it's _fine_ to use a descriptorless guard for fetching
directory info -- we just shouldn't use it when building circuits.
Fortunately, we already have a "usage" flag that we can use here.
Partial fix for bug 21242.
This relates to the 21242 fix -- entry_guard_pick_for_circuit()
should never yield nodes without descriptors when the node is going
to be used for traffic, since we won't be able to extend through
them.
In the past, when we exhausted all guards in our sampled set, we just
waited there till we mark a guard for retry again (usually takes 10 mins
for a primary guard, 1 hour for a non-primary guard). This patch marks
all guards as maybe-reachable when we exhaust all guards (this can
happen when network is down for some time).
Previously, we had NumEntryGuards kind of hardwired to 1. Now we
have the code (but not the configuarability) to choose randomly from
among the first N primary guards that would work, where N defaults
to 1.
Part of 20831 support for making NumEntryGuards work again.
Since we already had a separate function for getting the universe of
possible guards, all we had to do was tweak it to handle very the
GS_TYPE_RESTRICTED case.
asn found while testing that this function can be reached with
GUARD_STATE_COMPLETE circuits; I believe this happens when
cannibalization occurs.
The added complexity of handling one more state made it reasonable
to turn the main logic here into a switch statement.
Letting the maximum sample size grow proportionally to the number of
guards defeats its purpose to a certain extent. Noted by asn during
code review.
Fixes bug 20920; bug not in any released (or merged) version of Tor.
- Correctly maintain the previous guard selection in choose_guard_selection().
- Print bridge identifier instead of nothing in entry_guard_describe()._
If a complete circuit C2 doesn't obey the restrictions of C1, then
C2 cannot block C1.
The patch here is a little big-ish, since we can no longer look
through all the complete circuits and all the waiting circuits on a
single pass: we have to find the best waiting circuit first.
This is an important thing I hadn't considered when writing prop271:
sometimes you have to restrict what guard you use for a particular
circuit. Most frequently, that would be because you plan to use a
certain node as your exit, and so you can't choose that for your
guard.
This change means that the upgrade-waiting-circuits algorithm needs
a slight tweak too: circuit A cannot block circuit B from upgrading
if circuit B needs to follow a restriction that circuit A does not
follow.
I had been asking myself, "hey, doesn't the new code need to look at
this "info" parameter? The old code did!" But it turns out that the
old code hasn't, since 05f7336624.
So instead of "support this!" the comment now says "we can remove
this!"
George pointed out that (-1,0,1) for (never usable, maybe usable
later, usable right now) was a pretty rotten convention that made
the code harder to read.
Here we handle most (all?) of the remaining tasks, and fix some
bugs, in the prop271 bridge implementation.
* We record bridge identities as we learn them.
* We only call deprecated functions from bridges.c when the
deprecated guard algorithm is in use.
* We update any_bridge_descriptors_known() and
num_bridges_usable() to work correctly with the new backend
code. (Previously, they called into the guard selection logic.
* We update bridge directory fetches to work with the new
guard code.
* We remove some erroneous assertions where we assumed that we'd
never load a guard that wasn't for the current selection.
Also, we fix a couple of typos.
Still missing is functionality for picking bridges when we don't
know a descriptor for them yet, and functionality for learning a
bridge ID.
Everything else remains (basically) the same. Neat!
This includes:
* making bridge_info_t exposed but opaque
* allowing guards where we don't know an identity
* making it possible to learn the identity of a guard
* creating a guard that lacks a node_t
* remembering a guard's address and port.
* Looking up a guard by address and port.
* Only enforcing the rule that we need a live consensus to update
the "listed" status for guards when we are not using bridges.
This is safe, because no entry_guard_t ever outlives its
guard_selection_t.
I want this because now that multiple guard selections can be active
during one tor session, we should make sure that any information we
register about guards is with respect to the selection that they came
from.
Currently, this code doesn't actually have the contexts behave
differently, (except for the legacy context), but it does switch
back and forth between them nicely.
If a guard becomes primary as a result of confirming it, consider
the circuit through that guard as a primary circuit.
Also, note open questions on behavior when confirming nonprimary guards
Some of these will get torrc options to override them too; this
is just the mechanical conversion.
Also, add documentation for a couple of undocumented (but now used)
parameters.
To do this, it makes sense to treat legacy guards as a separate
guard_selection_t *, and handle them separately. This also means we
add support here for having multiple guard selections.
Note that we don't persist pathbias information yet; that will take
some refactoring.
This patch doesn't cover every case; omitted cases are marked with
"XXXX prop271", as usual. It leaves both the old interface and the
new interface for guard status notification, since they don't
actually work in the same way: the new API wants to be told when a
circuit has failed or succeeded, whereas the old API wants to know
when a channel has failed or succeeded.
I ran into some trouble with directory guard stuff, since when we
pick the directory guard, we don't actually have a circuit to
associate it with. I solved that by allowing guard states to be
associated with directory connections, not just circuits.
I expect we'll be ripping this out somewhere in 0.3.0, but let's
keep it around for a little while in case it turns out to be the
only way to avert disaster?
When a nonprimary guard's circuit is complete, we don't call it
actually usable until we are pretty sure that every better guard
is indeed not going to give us a working circuit.
Here we add a little bit of state to origin circuits, and set up
the necessary functions for the circuit code to call in order to
find guards, use guards, and decide when circuits can be used.
There's also an incomplete function for the hard part of the
circuit-maintenance code, where we figure out whether any waiting
guards are ready to become usable.
(This patch finally uses the handle.c code to make safe handles to
entry_guard_t objects, so that we are allowed to free an
entry_guard_t without checking whether any origin_circuit_t is
holding a reference to it.)
This code handles:
* Maintaining the sampled set, the filtered set, and the
usable_filtered set.
* Maintaining the confirmed and primary guard lists.
* Picking guards for circuits, and updating guard state when
circuit state changes.
Additionally, I've done code structure movement: even more constants
and structures from entrynodes.c have become ENTRYNODES_PRIVATE
fields of entrynodes.h.
I've also included a bunch of documentation and a bunch of unit
tests. Coverage on the new code is pretty high.
I've noted important things to resolve before this branch is done
with the /XXXX.*prop271/ regex.
