This has the different syntax for builder field attributes than what I
originally proposed in my MR, and which therefore is in the pinned
branch.
My upstream MR for the field attributes feature was morged:
https://github.com/colin-kiegel/rust-derive-builder/issues/239
This has the humantime_serde::option module, which we have upstreamed
and are about to switch to.
The remaining dependency with version = "1" is going to be removed
in a moment.
I used
git-grep -P '\#\[serde\((?!default|deny_unknown)'
to find places where I needed to add additional attributes on the
builder method fields.
This is currently a bit duplicative, but when #371 is completely done,
the validated (non-builder) configs won't need to be Deserialize any
more.
This is part of #371 and #372.
We are going to want to specify custom attributes on fields of the
builder struct. This feature was missing from derive_builder.
This commitid is the current head of my MR branch
https://github.com/colin-kiegel/rust-derive-builder/pull/237https://github.com/ijackson/rust-derive-builder/tree/builder-field-attrs
Using the commitid prevents surprises if that branch is updated.
We will require this newer version of derive_builder. The version
will need to be bumped again later, assuming the upstream MR is merged
and upstream do a release containing the needed changes.
When I wrote this, I arranged to skip dumping the field `pending`.
This must have been because I thought that either
(a) PendingEntry couldn't `#[derive(Debug)]` (but it can)
and/or
(b) Some of the fields of PendingEntry ought not to be dumped because
they might contain (eg) packet data. But I think they don't: there's
just the spec, and the Result which is (basically) a Circ.
I tried preseving something closer to the original using educe, but
educe gets somehow tangled up with the generics, and the result fails
to compile. I haven't investigated this further.
Fixes#365
Inspection of the code and logs shows that:
* One of the plan futures' oneshots must be returning Cancelled
* This means that the corresponding sender must have been dropped
* The sender is owned by the task spawned by spawn_launch
Presumably that entire task gets dropped as part of executor shutdown,
or something.
The correct response in this situation is to declare that we are
shutting down, and stop trying to do stuff.
Unfortunately, despite trying quite hard by putting sleeps in various
strategic places, I have not been able to reproduce the problem. So I
can't be 100% sure that the new behaviour is correct.
But I am reasonably confident that this ought not to be able to occur
unless either 1. the task from spawn_launch is dropped, or 2. that
task somehow panics despite its attempts to trap panics and report
them as errors through the oneshot.
So this "burn it all down" action ought only to occur in actually
serious situations.
I observe that
3ff9b187ea
Handle panics from circuit construction.
changed the EK for PendingCanceled to EK::ReactorShuttingDown,
and there's From impl. I think, therefore, that it is right
to reuse this Error variant.
I don't quite understand why when take_action gets an actual error it
doesn't push it, but just logs it. But I am not changing that for
now.
Arguably the two instances of retry_error.push are a sign of an
inferior flow control pattern - maybe the loop body including the code
I am adding ought to be an IEFE returning
`Result<Option<circ>, crate::Error>`.
I wanted this while debugging something.
The ad-hoc impl Debug with f.debug_struct is getting repetitive
and I've already perpetrated one copy-paste mistake.
We should consider using something like the `educe` crate's Clone.
This lint is IMO inherently ill-conceived.
I have looked for the reasons why this might be thought to be a good
idea and there were basically two (and they are sort of contradictory):
I. "Calling ‘.clone()` on an Rc, Arc, or Weak can obscure the fact
that only the pointer is being cloned, not the underlying data."
This is the wording from
https://rust-lang.github.io/rust-clippy/v0.0.212/#clone_on_ref_ptr
It is a bit terse; we are left to infer why it is a bad idea to
obscure this fact. It seems to me that if it is bad to obscure some
fact, that must be because the fact is a hazard. But why would it be
a hazard to not copy the underlying data ?
In other languages, faliing to copy the underlying data is a serious
correctness hazard. There is a whose class of bugs where things were
not copied, and then mutated and/or reused in multiple places in ways
that were not what the programmer intended. In my experience, this is
a very common bug when writing Python and Javascript. I'm told it's
common in golang too.
But in Rust this bug is much much harder to write. The data inside an
Arc is immutable. To have this bug you'd have use interior mutability
- ie mess around with Mutex or RefCell. That provides a good barrier
to these kind of accidents.
II. "The reason for writing Rc::clone and Arc::clone [is] to make it
clear that only the pointer is being cloned, as opposed to the
underlying data. The former is always fast, while the latter can
be very expensive depending on what is being cloned."
This is the reasoning found here
https://github.com/rust-lang/rust-clippy/issues/2048
This is saying that *not* using Arc::clone is hazardous.
Specifically, that a deep clone is a performance hazard.
But for this argument, the lint is precisely backwards. It's linting
the "good" case and asking for it to be written in a more explicit
way; while the supposedly bad case can be written conveniently.
Also, many objects (in our codebase, and in all the libraries we use)
that are Clone are in fact simply handles. They contain Arc(s) (or
similar) and are cheap to clone. Indeed, that is the usual case.
It does not make sense to distinguish in the syntax we use to clone
such a handle, whether the handle is a transparent Arc, or an opaque
struct containing one or more other handles.
Forcing Arc::clone to be written as such makes for code churn when a
type is changed from Arc<Something> to Something: Clone, or vice
versa.
To implement this, we had to refactor the tor_circmgr api for
flushing state changes to disk, so that it checks if it has the lock,
and only then tries to store.
We handle them by reporting them to task that's waiting for the
circuit, then relaying the panic.
Doing so allows the waiting task to distinguish panics
(EK::Internal) from cases where the reactor dropped the task
entirely (EK::ReactorShuttingDown). And doing _that_ removes one
case of EK::Canceled, which helps us on our goals towards #348.
Closes#347.
From its old name, this error had implied that we were giving no
useful information when we were waiting on a pending cirucit request
that failed. In fact, this error would only happen if we dropped the
`mpsc::Sender` for a circuit attempt without reporting success or
failure.
These errors should almost never be seen by the user; we should instead
retry the circuit. But they _can_ be seen by the use if selecting a
guard takes too long, or too many attempts. (Therefore, they aren't true
"internal" errors.)
I suspect that we might not want to keep this TransientFailure kind, but
I'm not sure what else to do here for now.
In one of the two places, nightly no longer warns. In the other
place, it's fine for nightly to warn: I just fixed the code to take
a slice instead.
Partial revert of 856aca8791.
Resolves part of #310.
Provide an enum variant to contain the SpawnError and a From impl.
We use `#[from]` here because it doesn't really make sense to attach
any context, as it's not likely to be very relevant.
The type annotation may not be necessary for inference, but as a
comment it risks becoming false. So it should be uncommented, or
deleted.
Error types round here are not entirely trivial so uncomment it.
Prompted by clippy::needless_question_mark. Sometimes Ok(r?) is
needed to do automatic error conversion. I assume the lint checks for
that. Anyway, in these cases it's not needed.
tor-netdir needs to bump because tor-netdoc bumped, even though
there were no other changes in tor-netdir. Whoops.
tor-guardmgr needs to bump because it already published, with the
older tor-netdir.
This commit puts the native-tls crate behind a feature. The feature
is off-by-default in the tor-rtcompat crate, but can be enabled
either from arti or arti-client.
There is an included script that I used to test that tor-rtcompat
could build and run its tests with all subsets of its features.
Closes#300
The docs even say this is about stream.
As @nickm writes in
https://gitlab.torproject.org/tpo/core/arti/-/merge_requests/252#note_2771289
we generally call end-to-end connections that are tunneled over Tor
"Streams" to distinguish them from everything else in the Tor
protocols that could possibly be called a "Connection".
That seems to apply here too.
This is totally not just an exercise to get combined test coverage
for tor-circmgr over 90% because I needed something to do that
wouldn't distract anybody else. :)
Since it implements a "<=" type relationship, it should be called
"at_least_as_permissive_as()." Since it's a crate-private function,
the long name isn't too bad.
I found these versions empirically, by using the following process:
First, I used `cargo tree --depth 1 --kind all` to get a list of
every immediate dependency we had.
Then, I used `cargo upgrade --workspace package@version` to change
each dependency to the earliest version with which (in theory) the
current version is semver-compatible. IOW, if the current version
was 3.2.3, I picked "3". If the current version was 0.12.8, I
picked "0.12".
Then, I used `cargo +nightly upgrade -Z minimal-versions` to
downgrade Cargo.lock to the minimal listed version for each
dependency. (I had to override a few packages; see .gitlab-ci.yml
for details).
Finally, I repeatedly increased the version of each of our
dependencies until our code compiled and the tests passed. Here's
what I found that we need:
anyhow >= 1.0.5: Earlier versions break our hyper example.
async-broadcast >= 0.3.2: Earlier versions fail our tests.
async-compression 0.3.5: Earlier versions handled futures and tokio
differently.
async-trait >= 0.1.2: Earlier versions are too buggy to compile our
code.
clap 2.33.0: For Arg::default_value_os().
coarsetime >= 0.1.20: exposed as_ticks() function.
curve25519-dalek >= 3.2: For is_identity().
generic-array 0.14.3: Earlier versions don't implement
From<&[T; 32]>
httparse >= 1.2: Earlier versions didn't implement Error.
itertools at 0.10.1: For at_most_once.
rusqlite >= 0.26.3: for backward compatibility with older rustc.
serde 1.0.103: Older versions break our code.
serde_json >= 1.0.50: Since we need its Value type to implement Eq.
shellexpand >= 2.1: To avoid a broken dirs crate version.
tokio >= 1.4: For Handle::block_on().
tracing >= 0.1.18: Previously, tracing_core and tracing had separate
LevelFilter types.
typenum >= 1.12: Compatibility with rust-crypto crates
x25519-dalek >= 1.2.0: For was_contributory().
Closes#275.
We are going to get rid of the Arc. Happily there is an id which is
always constructed uniquely and preserved by clone.
(auto-deref lets us make the function take &Self instead of &Arc)
Signed-off-by: Ian Jackson <ijackson@chiark.greenend.org.uk>
This is a method, so the resolution is automatic. It's not clear to
me why this was written out this way, given that extend_ntor is right
above.
Signed-off-by: Ian Jackson <ijackson@chiark.greenend.org.uk>
This closes arti#256. It makes our behavior match Tor's more closely,
though it has a simpler implementation than Tor. I think that the extra
complexity in Tor's logic is because we used to record timeouts in
the histogram as well as in the success/failure log.
The implementations of `Add` / `Sub` (et al.) on `std::time::Instant`
can panic if the underlying OS structure can't represent the result
(like arti#266). Use Instant::checked_add and print a warning instead,
to prevent panicking.
Also, we now add instead of subtracting; I suspect it's reasonable that
you might not be able to go backward past the first `Instant` created on
some platforms, but going *forward* should probably work?
IIUC, these anticipatd a need to store min_exit_circs_per_port in
CircMgr. But the current design, where it goes into preemptive.rs and
thence to usage, seems to work fine.
This required re-centralizing the configuration object for preemptive
circuits, since previously the settings from it were a bit spread out
over the crate.
And now the complexity begins: when the user changes the path_rules,
they not only want new circuits to obey those rules: they want
_all new requests_ to be put onto circuits that obey those rules.
That means that when the path rules become more restrictive, we need
to retire all the circuits, and make sure that currently pending
circuits aren't used for any requests.
If it's any comfort, doing this was even more complicated in C tor. ;)
This patch doesn't actually make anything reconfigurable, but it
does create an API that will tell you "you can't change the value of
that!" If the API looks reasonable, I can start making it possible
to change the values of individual items.
Previously we used Duration::mul_f64, which panics if its output is
out-of-range. That shouldn't actually be possible for the values
we're giving it, but probably it's better to just multiply in a safe
way.
This resolves a couple of XXXXs and therefore relates to #231.
We want to only use TODO in the codebase for non-blockers, and open
tickets for anything that is a bigger blocker than a TODO. These
XXXXs seem like definite non-blockers to me.
Part of arti#231.
Just as `in_same_family` is a member of Relay, so the function for
getting all the real family members of a relay should belong in the
same crate.
This change also removes the `family()` accessor: it gives the _claimed_ family rather
than the _acknlowedged_ family, and is therefore a bit dangerous.
There's still a hole in this logic; I've noted it in the Limitations
section. If we get a microdescriptor for a relay in between creating
and using the guard restriction, it might be omitted from the family
list.
There's not much reason to use a HashSet here, since we're just
going over the whole list.
This reverts commit 16e8489abb and does a little more
refactoring.
The new CircMgr::build_circuits_preemptively function actually causes
preemptive circuits to be built; it gets called from arti-client, like
the other daemon tasks the CircMgr has.
In preparation for making Arti build circuits preemptively, this commit
introduces `TargetCircUsage::Preemptive`, a circuit usage that works
somewhat differently from other ones: it requires at least 2 circuits to
exist that can exit the port it contains in order for an existing
circuit to match against it (path-spec.txt § 2.1.1); if that's not the
case, that usage will require building new circuits (in order that we
build enough to have 2 available).
This required refactoring how circuit reuse worked; now,
`CircList::find_open` uses the new `AbstractSpec::find_supported` trait
method, which we customize to implement the above check in the case of
`Preemptive` circuit usages. To make that work, `OpenEntry` now takes
two type parameters (the spec and circuit types), instead of taking a
builder type parameter and using its associated types. (We also got rid
of type constraints on that struct, yay!)
A WIP implementation of a preemptive circuit predictor that implements
path-spec.txt § 2.1.1 is also included, but this will require additional
effort to wire it up with the `CircMgr` properly.
This patch makes sure that for every* config type we have, the defaults
you get from a Builder match those you get from Serde, and that both
match the value that you get from arti_defaults.toml. Later down the
line I'll be adding some tests to keep these in sync.
* StorageConfig still has no defaults of its own, since we aren't so
sure we want other applications to use Arti's directories by default.
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)
Ian Jackson
trinity-1686a
Trinity Pointard
Ian Jackson
Nick Mathewson
eta
Yuan Lyu
Dimitris Apostolou
Neel Chauhan
Daniel Eades
Yuan Lyu
Jani Monoses