To do this at all neatly, I had to split out `tor-config` from
`arti-config` again, and putting the lower level stuff (paths,
builder errors) into tor-config. I also changed our use of
derive_builder to always use a common error type, to avoid
error type proliferation.
Rather like e8e9699c3c ("Get rid of
tor-proto's ChannelImpl, and use the reactor more instead"), this
admittedly rather large commit refactors the way circuits in `tor-proto`
work, centralising all of the logic in one large nonblocking reactor
which other things send messages into and out of, instead of having a
bunch of `-Impl` types that are protected by mutexes.
Congestion control becomes a lot simpler with this refactor, since the
reactor can manage both stream- and circuit-level congestion control
unilaterally without having to share this information with consumers,
meaning we can get rid of some locks.
The way streams work also changes, in order to facilitate better
handling of backpressure / fairness between streams: each stream now has
a set of channels to send and receive messages over, instead of sending
relay cells directly onto the channel (now, the reactor pulls messages
off each stream in each map, and tries to avoid doing so if it won't be
able to forward them yet).
Additionally, a lot of "close this circuit / stream" messages aren't
required any more, since that state is simply indicated by one end of a
channel going away. This should make cleanup a lot less brittle.
Getting all of this to work involved writing a fair deal of intricate
nonblocking code in Reactor::run_once that tries very hard to be mindful
of making backpressure work correctly (and congestion control); the old
code could get away with having tasks .await on things, but the new
reactor can't really do this (as it'd lock the reactor up), so has to do
everything in a nonblocking manner.
Instead of awkwardly sharing the internals of a `tor-proto` `Channel`
between the reactor task and any other tasks, move most of the internals
into the reactor and have other tasks communicate with the reactor via
message-passing to allocate circuits and send cells.
This makes a lot of things simple, and has convenient properties like
not needing to wrap the `Channel` in an `Arc` (though some places in the
code still do this for now).
A lot of test code required tweaking in order to deal with the refactor;
in fact, fixing the tests probably took longer than writing the mainline
code (!). Importantly, we now use `tokio`'s `tokio::test` annotation
instead of `async_test`, so that we can run things in the background
(which is required to have reactors running for the circuit tests).
This is an instance of #205, and also kind of #217.
Most of the structs in `arti-client` have example code now, to give a
clearer idea of how they're used.
Annoyingly, a lot of the types exposed in `arti-client` are actually
re-exports, which makes documentation a bit harder: example code that
references other parts of `arti-client` can't actually be run as a
doctest, since the crate it's in is a dependency of `arti-client`.
We might be able to fix this in future by doing the documentation in
`arti-client` itself, but rustdoc seems to have some weird behaviours
there that need to be investigated first (for example, it seems to merge
the re-export and original documentation, and also put the re-export
documentation on the `impl` block for some reason).
For now, though, this commit just writes the docs from the point of view
of an `arti-client` consumer, removing notes specific to the crate in
which they're defined. It's not ideal, but at least the end user
experience is decent.
Instead of putting a fully qualified name in the text, in most cases
we should just use the short name of the type or function we're
referring to.
In other words, instead of saying [`crate::module::Foo`], we should
typically say [`Foo`](crate::module::Foo).
The previous version of this test used the old, racy version of
wait_for (see #149). The new version is refactored so that
simulated time is only allowed to advance after each step is done,
so that we can actually be sure that each step in the process will
happen as it should.
In order to get the time-advances to proceed properly, and avoid
polluting state between tests, I've had to introduce some machinery
to encode the proper amount for time to advance. It isn't something
I'd want to use for a whole bunch of tests, but for just one set,
it's fine.
These tests now pass reliably for me.
I wonder if a discrete-event-simulation approach (hello, Shadow)
would let us write tests like these to our hearts' content?
As per arti#175, we'd like to be able to handle newer Arti versions
storing additional state in the persisted state files, without dropping
this data on the floor when we write out changes to these files.
Use the #[serde(flatten)] mechanism to achieve this, by adding catch-all
HashMap<String, JsonValue> fields to all structs that are at risk of
this happening to them.
There seems to be some issue here with the new WaitFor code,
where using the same MockSleepProvider with both of these wait_for()
calls gives questionable behavior under some circumstances (like
when running under Tarpaulin with the wrong set of flags).
We must not apply our new path-bias behavior (where we blame a guard
if it gives us too many indeterminate circuit failures) if the path
was not chosen at random. If too many random paths fail, we know
that's suspicious, since the other relays are a random sample. But
if a bunch of user-provided paths fail, that could simply be because
the user's chosen exit is down.
Instead of racily advancing time forward, this commit attempts to rework
how WaitFor works, such that it makes advances when all sleeper futures
that have been created have been polled (by handing the MockSleepRuntime
a Waker with which to wake up the WaitFor).
The above described mechanics work well enough for the double timeout
test, but fail in the presence of code that spawns asynchronous /
background tasks that must make progress before time is advanced for the
test to work properly. In order to deal with these cases, a set of APIs
are introduced in order to block time from being advanced until some
code has run, and a carveout added in order to permit small advances in
time where required.
(In some cases, code needed to be hacked up a bit in order to be made
properly testable using these APIs; the `MockablePlan` trait included in
here is somewhat unfortunate.)
This should fix arti#149.
Now that we have two kinds of isolation tokens (those set on a
stream, and those set by the stream's associated TorClient), we
need a more sophisticated kind of isolation.
This fixes the bug introduced with the previous commit, where
per-stream tokens would override per-TorClient tokens.
The three arguments TorClient::bootstrap requires by way of configuration
have been factored into a new TorClientConfig object.
This object gains two associated functions: one which uses `tor_config`'s
`CfgPath` machinery to generate sane defaults for the state and cache
directories, and one that accepts said directories in order to create a
config object with those inserted.
(this commit was inspired by trying to use arti as a library and being somewhat
overwhelmed by the amount of config stuff there was to do :p)
The previous code would report all failures to build a circuit as
failures of the guard. But of course that's not right: If we
fail to extend to the second or third hop, that might or might not
be the guard's fault.
Now we use the "pending status" feature of the GuardMonitor type so
that an early failure is attributed to the guard, but a later
failure is attributed as "Indeterminate". Only a complete circuit
is called a success. We use a new "GuardStatusHandle" type here so
that we can report the status early if there is a timeout.
(When we're building a path with a guard, we need to tell the guard
manager whether the path succeeded, and we need to wait to hear
whether the guard is usable.)
There are some missing parts here (like persistence and tests)
and some incorrect parts (I am 90% sure that the "exploratory
circuit" flag is bogus). Also it is not integrated with the circuit
manager code.
On torspec!40, Mike says:
I don't think there is a practical difference here. As per
Section 2.4.5, if 60 seconds is not enough and causes the
liveness test to fail due to too many timeouts, we will double
the initial timeout.
This makes our behavior the same as C tor.
The C Tor implementation doesn't do this, and Mike says:
I think it is a reasonable enough assumption that if Tor has
restarted, this kind data is no longer fresh enough to be
accurate for this purpose. This is also only 20 circuits here,
and typical timeouts are now around 1-2 seconds or less.. So a
restarted client with a timeout that is too low for a new
internet connection will figure this out pretty quickly. I think
that is OK.
(from torspec!40)