At the site of modify_hop, we now have a comment explaining the
internal-error issue.
To make the internal error less likely, we lower the modify_hop call
in lib.rs into GuardSet, where it can make sure it's looking at the
same filter as was used to select the guard.
The function name "pick_guard_ext" is not permanent; I'm going to
rename it in the next commit.
Previously, the API said "you need to call this in a loop till it
returns false". We did that in one place, but not another.
With the introduction of filters, forgetting to loop here becomes a
bug: so instead, change the behavior of extend_sample_as_needed so
it handles looping itself.
The fake network we construct for these tests is small in ways that
can sometimes cause weird behavior with guard filters. We fix this
by adjusting the parameters of the guard selection algorithm
accordingly in the tests.
With these new parameters, #491 no longer occurs.
This commit also adds comments to explain why the parameters are set
as they are.
Closes#491.
guard-spec.txt specifies that we have multiple separate samples of
guards that we can use depending on whether the filter is
restrictive or not. Here we implement the rules for switching
between samples.
When we're filtering guards, we have to check whether the filter is
"restrictive": if it forbids most of the guards (by bandwidth), we
keep its guards separated from the main set. If it is
super-restrictive, we also warn.
This functionality is specified in guard-spec.txt.
Since a guard can have a bunch of addresses, and the guard is
permitted if any one of those addresses is permitted, then we might
decide to use a guard with some non-permitted addresses. Thus, we
need to filter those addresses before returning the view of the
guard as a FirstHop.
Some of our existing code optionally takes a netdir from the
caller. When it doesn't give us one, use the netdir from the
installed NetDirProvider.
(Possibly someday we should remove the NetDir arguments
entirely. I'm deferring that because there are only two APIs
affected, and because making this change would force us to rewrite a
pretty large mess of unit tests.)
Previously it was the job of a task in CircMgr to do this; but we're
going to want to give GuardMgr full access to the latest NetDir for
this, and for other code-simplification reasons.
With this change I'm deprecating a couple of functions in
tor-circmgr. It's no longer necessary for us to have an artificial
external way for you to feed new NetDirs to a circmgr. (I could
just remove them, but I want practice deprecating.)
This only affects uses of thread_rng(), and affects them all more or
less indiscriminately. One test does not work with
ARTI_TEST_PRNG=deterministic; the next commit will fix it.
This commit was made by reverting the previous commit, then
re-running the script I used to generate it. In theory there should
be no semantic changes: only changes due to improved formatting from
cargo edit.
I followed the following procedure to make these changes:
* I used maint/changed_crates to find out which crates had changed
since 0.3.0.
* I used grep and maint/list_crates to sort those crates in
topological (dependency) order.
* I looked through semver_status to find which crates were listed as
having semver-relevant changes (new APIs and breaking changes).
* I scanned through the git logs of the crates with no
semver-relevant changes listed to confirm that, indeed, they had
no changes. For those crates, I incremented their patch-level
version _without_ changing the version that other crates depend on.
* I scanned through the git logs of the crates with no
semver-relevant changes listed to confirm that, indeed, they had
no obvious breaking changes.
* I treated all crates that depend on `arti` and/or `arti-client` as
having breaking changes.
* I identified crates that depend on crates that have changed, even
if they have not changed themselves, and identified them as having
a non-breaking change.
* For all of the crates, I used `cargo set-version -p $CRATE --bump
$STATUS` (where `STATUS` is `patch` or `minor`) to update the
versions, and the depended-upon versions.
Test the Deserialize impl of every config struct.
This detects bugs like the one fixed in !502.
The macro now becomes more complex because it needs to take options.
Right now this tt-munching option parser is overkill, but this
leave space for further options in the future.
I have Plans for this macro. In particular:
* I have a wip branch which tests that the Builder can be
deserialised from an empty config (ie, that config reading
of a config with a blank section for this item works).
* I think we should autogenerate $Config::builder(),
and promote that, rather than $ConfigBuilder::default().
This macro could do that.
For reference, the git source for this crate (and the others in its
workspace) currently lives in my personal github account (ijackson).
If this fork turns out to be long-lived and gains features and/or
users, it would be good to move it to a gitlab somewhere.
I have granted Nick crate ownership on the crates.io system.
* Builders additionally derive: Debug, Serialize, Deserialize.
* Validated structs no longer derive: Serialize, Deserialize
and all related attributes deleted.
* As a consequence, all the `#[serde(deny_unknown_fields)]`
are gone. That means that right now unknown fields are totally
ignored. This is good for compatibility but poor for useability.
Doing something better here is arti#417, in progress.
* As a consequence, delete tor_dirmgr::retry::default_parallelism.
(The default value was already duplicated into a builder attr.)
And drop the ad-hoc orport() method. This brings FallbackDir's
orports field in line with our list builder API.
The general semver note in "configuation" seems to cover most of this.
Although these do not appear in the config, it does have a builder.
It seems sensible to get rid of this ad-hoc list manipulation site,
and replace it with our standard list builder API.
define_list_builder_helper requires that the builder element type be
Deserialize. Currently GuardUsageRestriction is a transparent, public
enum, so we aren't really exposing anything.
We could introduce GuardUsageRestrictionBuilder now, but
since it's not in the config and thereofore only in the public API of
the lower crates, we can definitely put that off.
The new API is (roughly) as discussed in
https://gitlab.torproject.org/tpo/core/arti/-/issues/451
This is quite a large commit and it is not convenient to split it up.
It contains the following changes:
* Redo the list builder and accessor macros implemnetation,
including docs and tests.
* Change uses of define_list_config_builder. In each case:
- Move the docs about the default value to the containing field.
- Remove the other docs (which were just recapitulations, and
are now not needed since the ListBuilder is no longer public).
- Rewmove or replace `pub` in the define_list_builder_helper call,
so that the builder is no longer public.
- Change the main macro call site to use define_list_builder_helper.
- Add a call to define_list_builder_accessors.
* Make the module `list_builder` pub so that we have somewhere to
put the overview documentation.
* Consequential changes:
- Change `outer.inner().replace(X)` to `outer.set_inner(X)`
- Consequential changes to imports (`use` statements).
The `[patch]` approach causes the tree not to build when used as a
dependency, unless the `[patch]` is replicated into the depending
project.
Instead, replace our `derive_builer =` dependencies with a reference
to a specific git commit:
perl -i~ -pe 'next unless m/^derive_builder/; s#"(0\.11\.2)"#{ version = "$1", git = "https://github.com/ijackson/rust-derive-builder", rev = "ba0c1a5311bd9f93ddf5f5b8ec2a5f6f03b22fbe" }#' crates/*/Cargo.toml
Note that the commitid has changed. This is because derive_builder is
in fact a workspace of 4 crates. 3 of them are of interest to arti
itself (the 4th exists only for testing). So the same "add git
revision" treatment had to be done to the `derive_builder` and
`derive_builder_macro` crates. Each dependency edge involves a new
commit in the derive_builder workspace, since we can't create a git
commit containing its own commitid. (We want to use commits, rather
than a branch, so that what we are depending on is actually properly
defined, and not subject to the whims of my personal github
namespace.)
There are no actual code changes in derive_builder.
And add an imprecation in define_list_config_builder's doc comment do
do so in future for other invocations of the macro.
Add add the missing full stops.
This is an automated change made with a perl one-liner and verified
with grep -L and grep -l.
Some warnings are introduced with this change; they will be removed
in subsequent commits.
See arti#208 for older discussion on this issue.
Now the network fallbacks configuration wants to Deserialize
a Vec<FallbackDirBuilder>, rather than validated Vec<FallbackDir>.
Methods on FallbackListBuilder are as per
https://gitlab.torproject.org/tpo/core/arti/-/merge_requests/462#note_2797697
mutatis mutandi for the fact that this struct has only fallbacks in it.
This is where the FallbackList type is. We are going to want to
provide a builder too, which ought to impl Default.
This means that the default value for the type must be next to the
type. In any case, it was anomalous that it wasn't.
This commit is pure code motion.
Instead of just having a function that recalculates the latest clock
skew, instead recalculate the clock skew when it may have changed,
and notify other processes via a postage::watch.
Apparently on OSX you are not allowed to construct an Instant that is a
long time before the time when the test is running.
Also, fix the length of a year in this test.
This time, our estimator discards outliers, takes the mean of what's
left, and uses the standard deviation to try to figure out how
seriously to take our report of skew/not-skew.
These estimates are still not actually used.
The old version of this function was error-prone, and in fact had
errors: it was too easy to forget to add non-persistent fields, and
that's exactly what we forgot in a few cases
(`microdescriptor_missing`, `circ_history`, and
`suspicious_behavior_warned`).
The new version of this function consumes both of the incoming
Guards, and constructs every field explicitly so that we can't
forget to list any.
Closes#429.
Previously, we treated successfully building a circuit to a guard as
a "success", and any failure, including a directory cache failure,
as a failure. With this change, guards now have separate
success/failure and retry status for circuit usage and directory
usage.
This change is needed for guard-as-directory retry to have
reasonable behavior. Otherwise, when a guard succeeds at building a
circuit, that clears the directory-is-failing status and makes us
retry the guards to quickly.
Not all of these strictly need to be bumped to 0.2.0; many could go
to 0.1.1 instead. But since everything at the tor-rtcompat and
higher layers has had breaking API changes, it seems not so useful
to distinguish. (It seems unlikely that anybody at this stage is
depending on e.g. tor-protover but not arti-client.)
The FirstHopId type now records an enum that stores whether the hop
is a guard or a fallback. This change addresses concerns about
remembering to check the type or source of an Id before passing it
down to the FallbackState or GuardSet.
Making this change required an API change, so that dirmgr can
report success/failure status without actually knowing whether it's
using a fallback or a guard.
The code here uses a new iterator type, since I couldn't find one of
these on crates.io. I tried writing the code without it, but it was
harder to follow and test.
We do this by creating a new FallbackSet type that includes status
information, and updating the GuardMgr APIs to record success and
failure about it when appropriate. We can use this to mark
FallbackDirs retriable (or not).
With this change, FallbackDir is now stored internally as a Guard in
the GuardMgr crate. That's fine: the FallbackDir type really only
matters for configuration.
We only do this when we fail to get a regular guard (e.g., because
they're all down), and when we have been asked for a guard for a
one-hop directory.
Most of the change in this commit is plumbing to make all of the
types match up.
As before, compilation may still be broken.
We need to extend our notion of "the origin of a guard" to include
"somewhere outside the guard list"; we need the ability to return a
FallbackDir as a Guard; and we need to remember a few more pieces of
information in each pending request.
As before, this commit may break compilation; it will be restored soon.
The guard manager is responsible for handing out the first hops of
tor circuits, keeping track of their successes and failures, and
remembering their states. Given that, it makes sense to store this
information here. It is not yet used; I'll be fixing that in
upcoming commits.
Arguably, this information no longer belongs in the directory
manager: I've added a todo about moving it.
This commit will break compilation on its own in a couple of places;
subsequent commits will fix it up.
This is the logical place for it, I think: the GuardMgr's job is to
pick the first hop for a circuit depending on remembered status for
possible first hops. Making this change will let us streamline the
code that interacts with these objects.
When all guards are down, we would previously mark them all as up,
and retry aggressively. But that's far too aggressive: if there's
something wrong with our ability to connect to guards, it makes us
hammer the network over and over, ignoring all the guard retry
timeouts in practice.
Instead,
* We now allow the `pick_guard()` function to fail without
automatically retrying.
* We give different errors in the cases when all our guards are
down, and when all of the guards selected by our active usage
are down.
* Our "guards are down" error includes the time at which a guard
will next be retriable.
This is part of #407.
C tor used one schedule, and guard-spec specified another. But in
reality we should probably use a randomized schedule to retry
guards, for the reasons explained in the documentation for
RetrySchedule.
I've chosen the minima to be not too far from our previous minima
for primary and non-primary guards.
This is part of #407.
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.
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.
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.
Previously coarsetime and the traffic-timestamp feature were
enabled, since they were only required for a small corner of the
guardmgr algorithm.
But in 1.0 and beyond we'll be adding a bunch of other features (eg,
netflow padding, DoS prevention) that will need coarsetime all over
the place.
And since we're going to be doing coarsetime all over the place, the
previous justification for making traffic-timestamping optional (the
tiny performance hit) is no longer relevant.
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.
This needs two kinds. We have decided to treat a non-shutdown
SpawnError as "unexplained" rather than as an InternalError.
There are many crates whose
From<futures::task::SpawnError> for Error
erroneously treat it as an internal error. We will fix them in a moment.
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
Previously we stored only one guard sample, in a state file called
"default_guards". That's not future-proof, since we want to have
multiple samples in the future. (`guard-spec.txt` specifies
separate samples for highly restrictive filters, and for bridge
usage.)
This patch changes our behavior so that we can store multiple
samples in a new "guards" file.
I had thought about automatically migrating from the previous file
format and location, but I don't think that's necessary given our
current (lack of) stability guarantees.
Closes#176.
This test should only fail very rarely (around 1/2.4e8) when guards
are chosen from a list of 20 with uniform probability. But that
wasn't what we were doing on the mock test network: we were choosing
from a list of 10 viable guards, with nonuniform probability.
As a fix, we change the test network probabilities so that the
guards _are_ chosen with a uniform probability for this test, and we
use a modified version of the test network where there are indeed 20
Guard-flagged relays with the required DirCache=2 protocol.
Closes#276.
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.
If we don't know a current microdescriptor for a guard, we can't use it
for multihop circuits, since we don't know its onion keys.
This is part of a fix for #178.
`tor-rtcompat`'s `TlsConnector` trait previously included a method to
create a TLS-over-TCP connection, which implied creating a TCP stream
inside that method. This commit changes that, and makes the function
wrap a TCP stream, as returned from the runtime's `TcpProvider` trait
implementation, instead.
This means you can actually override `TcpProvider` and have it apply to
*all* connections Arti makes, which is useful for issues like arti#235
and other cases where you want to have a custom TCP stream
implementation.
This required updating the mock TCP/TLS types in `tor-rtmock` slightly;
due to the change in API, we now store whether a `LocalStream` should
actually be a TLS stream inside the stream itself, and check this
property on reads/writes in order to detect misuse. The fake TLS wrapper
checks this property and removes it in order to "wrap" the stream,
making reads and writes work again.
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.
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.
Rust nightly claims that Vec might get its own retain_mut method,
which would potentially conflict with the extension method we've
grabbed from the retain_mut crate. To solve this, we're calling the
method explicitly.
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.
We define "coming back online" as happening when a guard attempt
succeeds, if that attempt that was launched when we seemed to be
offline.
We define "seeming to be offline" as having all of our primary
guards marked unreachable, and having received no incoming network
traffic in a while.
Closes#216.
This is based on @eta's patches for !118 and !119: Since we already
have an unbounded channel, we don't need to use an elaborate mess of
one-shot senders. We can just use the unbounded_send() method,
which also lets us enqueue a message without having to await.
Closes#219.
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 Futureproof<T> type lets you serialize and deserialize types whose
representations might change (most useful for enums that might grow
additional variants). It uses #[serde(untagged)] to accomplish this.
This gets used in order to make the `disabled` field of `Guard` more
robust against future guard disablement reasons being added.
A test was also added to verify correct behaviour of the new type.
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.
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.
We now track, for every guard: the total number of successful
circuits we've built through it, along with the total number of
"indeterminate" circuits.
Recall that a circuit's status is "indeterminate" if it has failed
for a reason that _might_ be the guard's fault, or might not be the
guard's fault. For example, if extending to the second hop of the
circuit fails, we have no way to know whether the guard deliberately
refused to connect there, or whether the second hop is just offline.
But we don't want to forgive all indeterminate circuit failures: if
we did, then a malicious guard could simply reject any second hops
that it didn't like, thereby filtering the client into a chosen
set of circuits.
As a stopgap solution, this patch now makes guards become
permanently disabled if the fraction of their circuit failures
becomes too high.
See also general-purpose path bias selection (arti#65), and Mike's
idea for changing the guard reachability definition (torspec#67).
This patch doesn't do either of those.
Closes#185.
The `get_relay` function was confusing, since it would return None if
the relay was present, but wasn't actually a guard. We only used it
in one place, and in that one place we used it wrong, leading to a
panic bug.
Fixes#193.
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.
The advantage here is that we no longer have to use a futures-aware
Mutex, or a blocking send operation, and therefore can simplify a
bunch of the GuardMgr APIs to no longer be async. That'll avoid
having to propagate the asyncness up the stack.
The disadvantage is that unbounded channels are just that: nothing
in the channel prevents us from overfilling it. Fortunately, the
process that consumes from the channel shouldn't block much, and
the channel only gets filled when we're planning a circuit path.
Also, refactor our message handling to be more like the tor_proto
reactors. The previous code had a bug where, once the stream of
events was exhausted, we wouldn't actually get any more
notifications.
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.
Nick Mathewson
Ian Jackson
Orhun Parmaksız
Ian Jackson
eta
Samanta Navarro
trinity-1686a
Yuan Lyu
Daniel Eades
Neel Chauhan
dagon
Dimitris Apostolou