For testability, replace most if-based construction logic with
calls to well-tested factories living in other packages.
While there, acknowledge that a bunch of types could now be private
and make them private, modifying the code to call the public
factories allowing to construct said types instead.
Part of https://github.com/ooni/probe/issues/2121
This diff modifies netx to stop using most netxlite resolver internals
but the internal function that creates a new, unwrapped system resolver,
which will be dealt with in a subsequent pull request.
See https://github.com/ooni/probe/issues/2121
1. Use the netxlite.NewHTTPTransport factory for creating a new
HTTP2 (and HTTP1) transport;
2. Recognize the netxlite.NewOOHTTPTransport has now become
an implementation detail so make it private;
3. Recognize that netxlite.NewHTTP3Transport should call
netxlite.WrapTransport so it returns the same typechain
returned by netxlite.NewHTTPTransport (modulo, of course,
the real underlying transport), so ensure that we are
calling netxlite.WrapTransport in NewHTTP3Transport;
4. Recognize that the table based constructor inside of
netx needs a logger to create HTTPTransport instances using
either netxlite.NewHTTP{,3}Transport so pass this argument
along and ensure it's not nil using a constructor inside
model that guarantees that;
5. Cleanup netx's tests to avoid type asserting on the
typechains returned by netxlite since we already test
that inside netxlite;
6. Recognize that now we can make more legacy names inside
of netxlite private because we don't need to use them
inside tests anymore (because of previous point).
Reference issue: https://github.com/ooni/probe/issues/2121
This diff modifies netx to use netxlite to build the TLSDialer.
Building the TLSDialer entails building a TLSHandshaker.
While there, hide netxlite names we don't want to be public
and change netx tests to test for functionality.
To this end, refactor filtering to provide an easier to
use TLS server. We don't need the complexity of proxying
rather we need to provoke specific errors.
Part of https://github.com/ooni/probe/issues/2121
Tracex contained some fragile code that assembled HTTP measurements
from scattered events, which worked because we were sure we were
performing a single measurement at any given time.
This diff restructures the code to emit a transaction-start and a
transaction-done events only. We have basically removed all the other
events (which we were not using). We kept the transaction-start
though, because it may be useful to see it when reading events. In
any case, what matters here is that we're now using the transaction-done
event aline to generate the archival HTTP measurement.
Hence, the original issue has been addressed. We will possibly
do more refactoring in the future, but for now this seems sufficient.
Part of https://github.com/ooni/probe/issues/2121
Acknowledge that transports MAY be used in isolation (i.e., outside
of a Resolver) and add support for wrapping.
Ensure that every factory that creates an unwrapped type is named
accordingly to hopefully ensure there are no surprises.
Implement DNSTransport wrapping and use a technique similar to the
one used by Dialer to customize the DNSTransport while constructing
more complex data types (e.g., a specific resolver).
Ensure that the stdlib resolver's own "getaddrinfo" transport (1)
is wrapped and (2) could be extended during construction.
This work is part of my ongoing effort to bring to this repository
websteps-illustrated changes relative to netxlite.
Ref issue: https://github.com/ooni/probe/issues/2096
This diff modifies the system resolver to use a getaddrinf transport.
Obviously the transport is a fake, but its existence will allow us
to observe DNS events more naturally.
A lookup using the system resolver would be a ANY lookup that will
contain all the resolved IP addresses into the same response.
This change was also part of websteps-illustrated, albeit the way in
which I did it there was less clean than what we have here.
Ref issue: https://github.com/ooni/probe/issues/2096
Rather than passing functions to construct complex objects such
as Dialer and QUICDialer, pass interface implementations.
Ensure that a nil implementation does not cause harm.
Make Saver implement the correct interface either directly or
indirectly. We need to implement the correct interface indirectly
for TCP conns (or connected UDP sockets) because we have two
distinct use cases inside netx: observing just the connect event
and observing just the I/O events.
With this change, the construction of composed Dialers and
QUICDialers is greatly simplified and more obvious.
Part of https://github.com/ooni/probe/issues/2121
This diff modifies the construction of a dialer to allow one
to insert custom dialer wrappers into the dialers chain.
The point of the chain in which we allow custom wrappers is the
optimal one for connect, read, and write measurements.
This new design is better than the previous netx design since
we don't need to construct the whole chain manually now.
The work in this diff is part of the effort to make engine/netx
just a tiny wrapper around netxlite.
See https://github.com/ooni/probe/issues/2121.
This diff replaces engine/netx code with netxlite code in
the engine/session.go file. To this end, we needed to move
some code from engine/netx to netxlite. While there, we
did review and improve the unit tests.
A notable change in this diff is (or seems to be) that in
engine/session.go we're not filtering for bogons anymore so
that, in principle, we could believe a resolver returning
to us bogon IP addresses for OONI services. However, I did
not bother with changing bogons filtering because the
sessionresolver package is already filtering for bogons,
so it is actually okay to avoid doing that again the
session.go code. See:
https://github.com/ooni/probe-cli/blob/v3.15.0-alpha.1/internal/engine/internal/sessionresolver/resolvermaker.go#L88
There are two reference issues for this cleanup:
1. https://github.com/ooni/probe/issues/2115
2. https://github.com/ooni/probe/issues/2121
In https://github.com/ooni/probe/issues/2029#issuecomment-1140805266, we
explained why calling it "netgo" would be incorrect.
In other words, we can get the platform's `getaddrinfo` as long as
we're not cross compiling. We do cross compile `ooniprobe`, actually
it's not even possible to cross compile it.
For increased accuracy, we should stop cross compiling `miniooni`
as well, so it would also directly use `getaddrinfo`.
This diff fixes at the same time ooni/probe-cli and ooni/spec
and we'll open two pull requests in parallel.
After https://github.com/ooni/probe-cli/pull/764, the build for
CGO_ENABLED=0 has been broken for miniooni:
https://github.com/ooni/probe-cli/runs/6636995859?check_suite_focus=true
Likewise, it's not possible to run tests with CGO_ENABLED=0.
To make tests work with `CGO_ENABLED=0`, I needed to sacrifice some
unit tests run for the CGO case. It is not fully clear to me what was happening
here, but basically `getaddrinfo_cgo_test.go` was compiled with CGO
being disabled, even though the ``//go:build cgo` flag was specified.
Additionally, @hellais previously raised a valid point in the review
of https://github.com/ooni/probe-cli/pull/698:
> Another issue we should consider is that, if I understand how
> this works correctly, depending on whether or not we have built
> with CGO_ENABLED=0 on or not, we are going to be measuring
> things in a different way (using our cgo inspired getaddrinfo
> implementation or using netgo). This might present issues when
> analyzing or interpreting the data.
>
> Do we perhaps want to add some field to the output data format that
> gives us an indication of which DNS resolution code was used to
> generate the the metric?
This comment is relevant to the current commit because
https://github.com/ooni/probe-cli/pull/698 is the previous
iteration of https://github.com/ooni/probe-cli/pull/764.
So, while fixing the build and test issues, let us also distinguish
between the CGO_ENABLED=1 and CGO_ENABLED=0 cases.
Before this commit, OONI used "system" to indicate the case where
we were using net.DefaultResolver. This behavior dates back to the
Measurement Kit days. While it is true that ooni/probe-engine and
ooni/probe-cli could have been using netgo in the past when we
said "system" as the resolver, it also seems reasonable to continue
to use "system" top indicate getaddrinfo.
So, the choice here is basically to use "netgo" from now on to
indicate the cases in which we were built with CGO_ENABLED=0.
This change will need to be documented into ooni/spec along with
the introduction of the `android_dns_cache_no_data` error.
## Checklist
- [x] I have read the [contribution guidelines](https://github.com/ooni/probe-cli/blob/master/CONTRIBUTING.md)
- [x] reference issue for this pull request: https://github.com/ooni/probe/issues/2029
- [x] if you changed anything related how experiments work and you need to reflect these changes in the ooni/spec repository, please link to the related ooni/spec pull request: https://github.com/ooni/spec/pull/242
To this end, we need to refactor the implementation to give the
DNSOverUDPChannel owenership over the net.Conn.
Once this happens, DNSOverUDPChannel.Close closes the conn.
When the conn is closed, the background goroutine will terminate
immediately because any blocking I/O operation will be immediately
unblocked and return net.ErrClosed.
See https://github.com/ooni/probe/issues/2099#issuecomment-1139066946
This diff introduces support for observing additional DNS-over-UDP
responses in some censored environments (e.g. China).
After some uncertainty around whether to use connected or unconnected
UDP sockets, I eventually settled for connected.
Here's a recap:
| | connected | unconnected |
| ----------------------- | --------- | ----------- |
| see ICMP errors | ✔️ | ❌ |
| responses from any server | ❌ | ✔️ |
Because most if not all DNS resolvers expect answers from exactly
the same servers to which they sent the query, I would say that
it's more important to have some limited ability of observing the
effect of ICMP errors (e.g., host_unreachable when we set a low
TTL and send out a query to a server).
Therefore, my choice was to modify the existing DNS-over-UDP transport.
Here's an overview of the changes:
1. introduce a new API for performing an async round trip that returns
a channel wrapper where all responses are posted. The channel will not ever
be closed, so the reader needs to use select for safely reading. If the
reader users the wrapper's Next or TryNextResponses methods, these details
do not matter because they already implement a safe reading pattern.
2. the async round trip API performs the round trip in the background
and stops processing when it sees the first error.
3. the background running code will use an overall deadline derived
from the DNSTransport.IOTimeout field to know when to stop.
4. the background running code will additionally stop running if
noone is reading the channel and there are no empty slots in the
channel's buffer.
5. the RoundTrip method has been rewritten in terms of the async API.
The design I'm using here implements the proposal for async round
trips defined at https://github.com/ooni/probe/issues/2099. I have
chosen not to make all transports async because the DNS transport
seems the only transport that needs to also work in async mode.
While there, I noticed that we were not propagating CloseIdleConnection
to the underlying dialer, which was potentially wrong, so I did it.
This diff refactors the DNSTransport model to receive in input a DNSQuery and return in output a DNSResponse.
The design of DNSQuery and DNSResponse takes into account the use case of a transport using getaddrinfo, meaning that we don't need to serialize and deserialize messages when using getaddrinfo.
The current codebase does not use a getaddrinfo transport, but I wrote one such a transport in the Websteps Winter 2021 prototype (https://github.com/bassosimone/websteps-illustrated/).
The design conversation that lead to producing this diff is https://github.com/ooni/probe/issues/2099
These two small packages could easily be merged into the model
package, since they're clearly model-like packages.
Part of https://github.com/ooni/probe/issues/2115
The objective is to make PR checks run much faster.
See https://github.com/ooni/probe/issues/2113 for context.
Regarding netxlite's tests:
Checking for every commit on master or on a release branch is
good enough and makes pull requests faster than one minute since
netxlite for windows is now 1m slower than coverage.
We're losing some coverage but coverage from integration tests
is not so good anyway, so I'm not super sad about this loss.
Rather than building for Android using ooni/go, we're now using
ooni/oocryto as the TLS dependency. Such a repository only forks
crypto/tls and some minor crypto packages and includes the
same set of patches that we have been using in ooni/go.
This new strategy should be better than the previous one in
terms of building for Android, because we can use the vanilla
go1.18.2 build. It also seems that it is easier to track and
merge from upstream with ooni/oocrypto than it is with ooni/go.
Should this assessment be wrong, we can revert back to the
previous scenario where we used ooni/go.
See https://github.com/ooni/probe/issues/2106 for extra context.
Previously, the DNS decoder did not check whether it was parsing
a DNS query or a DNS response, which was wrong.
As a side note, it seems I am using "reply" in the codebase instead
of "response". The latter seems correct DNS terminology.
This diff has been extracted from 9249d14f80
See https://github.com/ooni/probe/issues/2096.
This diff has been extracted and adapted from 8848c8c516
The reason to prefer composition over embedding is that we want the
build to break if we add new methods to interfaces we define. If the build
does not break, we may forget about wrapping methods we should
actually be wrapping. I noticed this issue inside netxlite when I was working
on websteps-illustrated and I added support for NS and PTR queries.
See https://github.com/ooni/probe/issues/2096
While there, perform comprehensive netxlite code review
and apply minor changes and improve the docs.
This diff has been extracted from c2f7ccab0e
See https://github.com/ooni/probe/issues/2096
While there, export DecodeReply to decode a raw reply without
interpreting the Rcode or parsing the results, which seems a
nice extra feature to have to more flexibly parse DNS replies
in other parts of the codebase.
This error occurred for example when querying kazemjalali.com
in websteps measurements run from Iran.
This error is relatively uncommon, but it still makes sense to
create a specific mapping rule for it.
Originally: 4269e82fbd
See https://github.com/ooni/probe/issues/2096
I've seen some measurements returning some IP addresses for HTTPSSvc
queries but not returning any ALPN value.
For example:
```
% d4
decoding DNS round trip 0:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 57768
;; flags: rd; QUERY: 1, ANSWER: 0, AUTHORITY: 0, ADDITIONAL: 0
;; QUESTION SECTION:
;psiphon.ca. IN HTTPS
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 57768
;; flags: qr rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 0
;; QUESTION SECTION:
;psiphon.ca. IN HTTPS
;; ANSWER SECTION:
psiphon.ca. 121 IN A 31.13.85.53
```
Now, the response is clearly bogus. At the time of this writing that
IP address belongs to Facebook. This measurement has been collected in
China, so it's expected for the GFW to behave like this.
Yet, I don't feel like it's accurate to report this measurement as a
"no answer" response. Rather, this response is a valid one containing
a clearly invalid IP address and should be flagged as such.
Originally: 57a023bcf4
See https://github.com/ooni/probe/issues/2096
This diff fixes the short-circuit-IP-addr resolver to
correctly handle IP addrs during LookupHTTPS.
The original diff was: 2b51d144bf
See https://github.com/ooni/probe/issues/2096
While there, add unit tests for IPv6.
* quic-go upgrade: replaced Session/EarlySession with Connection/EarlyConnection
* quic-go upgrade: added context to RoundTripper.Dial
* quic-go upgrade: made corresponding changes to tutorial
* quic-go upgrade: changed sess variable instances to qconn
* quic-go upgrade: made corresponding changes to tutorial
* cleanup: remove unnecessary comments
Those comments made sense in terms of illustrating the changes
but they're going to be less useful once we merge.
* fix(go.mod): apparently we needed `go1.18.1 mod tidy`
VSCode just warned me about this. It seems fine to apply this
change as part of the pull request at hand.
* cleanup(netxlite): http3dialer can be removed
We used to use http3dialer to glue a QUIC dialer, which had a
context as its first argument, to the Dial function used by the
HTTP3 transport, which did not have a context as its first
argument.
Now that HTTP3 transport has a Dial function taking a context as
its first argument, we don't need http3dialer
anymore, since we can use the QUIC dialer directly.
Cc: @DecFox
* Revert "cleanup(netxlite): http3dialer can be removed"
This reverts commit c62244c620cee5fadcc2ca89d8228c8db0b96add
to investigate the build failure mentioned at
https://github.com/ooni/probe-cli/pull/715#issuecomment-1119450484
* chore(netx): show that test was already broken
We didn't see the breakage before because we were not using
the created transport, but the issue of using a nil dialer was
already present before, we just didn't see it.
Now we understand why removing the http3transport in
c62244c620cee5fadcc2ca89d8228c8db0b96add did cause the
breakage mentioned at
https://github.com/ooni/probe-cli/pull/715#issuecomment-1119450484
* fix(netx): convert broken integration test to working unit test
There's no point in using the network here. Add a fake dialer that
breaks and ensure we're getting the expected error.
We've now improved upon the original test because the original test was
not doing anything while now we're testing whether we get back a QUIC
dialer that _can be used_.
After this commit, I can then readd the cleanup commit
c62244c620cee5fadcc2ca89d8228c8db0b96add and it won't be
broken anymore (at least, this is what I expected to happen).
* Revert "Revert "cleanup(netxlite): http3dialer can be removed""
This reverts commit 0e254bfc6ba3bfd65365ce3d8de2c8ec51b925ff
because now we should have fixed the broken test.
Co-authored-by: decfox <decfox>
Co-authored-by: Simone Basso <bassosimone@gmail.com>
Here's the squash of the following patches that enable support
for go1.18 and update our dependencies.
This diff WILL need to be backported to the release/3.14 branch.
* chore: use go1.17.8
See https://github.com/ooni/probe/issues/2067
* chore: upgrade to probe-assets@v0.8.0
See https://github.com/ooni/probe/issues/2067.
* chore: update dependencies and enable go1.18
As mentioned in 7a0d17ea91,
the tree won't build with `go1.18` unless we say it does.
So, not only here we need to update dependencies but also we
need to explicitly say `go1.18` in the `go.mod`.
This work is part of https://github.com/ooni/probe/issues/2067.
* chore(coverage.yml): run with go1.18
This change will give us a bare minimum confidence that we're
going to build our tree using version 1.18 of golang.
See https://github.com/ooni/probe/issues/2067.
* chore: update user agent used for measuring
See https://github.com/ooni/probe/issues/2067
* chore: run `go generate ./...`
See https://github.com/ooni/probe/issues/2067
* fix(dialer_test.go): make test work with go1.17 and go1.18
1. the original test wanted the dial to fail, so ensure we're not
passing any domain name to exercise dialing not resolving;
2. match the end of the error rather than the whole error string.
Tested locally with both go1.17 and go1.18.
See https://github.com/ooni/probe-cli/pull/708#issuecomment-1096447186
This diff includes some final changes to be ready for blessing
a cli release. These changes are:
1. run `go generate ./...` to update the bundled CA
2. update the header we use for measuring
3. ensure `mk` uses the latest version of several tools
Reference issue: https://github.com/ooni/probe/issues/1845
This diff introduces a new package called `./internal/archival`. This package collects data from `./internal/model` network interfaces (e.g., `Dialer`, `QUICDialer`, `HTTPTransport`), saves such data into an internal tabular data format suitable for on-line processing and analysis, and allows exporting data into the OONI data format.
The code for collecting and the internal tabular data formats are adapted from `measurex`. The code for formatting and exporting OONI data-format-compliant structures is adapted from `netx/archival`.
My original objective was to _also_ (1) fully replace `netx/archival` with this package and (2) adapt `measurex` to use this package rather than its own code. Both operations seem easily feasible because: (a) this code is `measurex` code without extensions that are `measurex` related, which will need to be added back as part of the process; (b) the API provided by this code allows for trivially converting from using `netx/archival` to using this code.
Yet, both changes should not be taken lightly. After implementing them, there's need to spend some time doing QA and ensuring all nettests work as intended. However, I am planning a release in the next two weeks, and this QA task is likely going to defer the release. For this reason, I have chosen to commit the work done so far into the tree and defer the second part of this refactoring for a later moment in time. (This explains why the title mentions "1/N").
On a more high-level perspective, it would also be beneficial, I guess, to explain _why_ I am doing these changes. There are two intertwined reasons. The first reason is that `netx/archival` has shortcomings deriving from its original https://github.com/ooni/netx legacy. The most relevant shortcoming is that it saves all kind of data into the same tabular structure named `Event`. This design choice is unfortunate because it does not allow one to apply data-type specific logic when processing the results. In turn, this choice results in complex processing code. Therefore, I believe that replacing the code with event-specific data structures is clearly an improvement in terms of code maintainability and would quite likely lead us to more confidently change and evolve the codebase.
The second reason why I would like to move forward these changes is to unify the codepaths used for measuring. At this point in time, we basically have two codepaths: `./internal/engine/netx` and `./internal/measurex`. They both have pros and cons and I don't think we want to rewrite whole experiments using `netx`. Rather, what we probably want is to gradually merge these two codepaths such that `netx` is a set of abstractions on top of `measurex` (which is more low-level and has a more-easily-testable design). Because saving events and generating an archival data format out of them consists of at least 50% of the complexity of both `netx` and `measurex`, it seems reasonable to unify this archival-related part of the two codebases as the first step.
At the highest level of abstraction, these changes are part of the train of changes which will eventually lead us to bless `websteps` as a first class citizen in OONI land. Because `websteps` requires different underlying primitives, I chose to develop these primitives from scratch rather than wrestling with `netx`, which used another model. The model used by `websteps` is that we perform each operation in isolation and immediately we save the results, while `netx` creates whole data structures and collects all the events happening via tracing. We believe the model used by `websteps` to be better because it does not require your code to figure out everything that happened after the measurement, which is a source of subtle bugs in the current implementation. So, when I started implementing websteps I extracted the bits of `netx` that could also be beneficial to `websteps` into a separate library, thus `netxlite` was born.
The reference issue describing merging the archival of `netx` and `measurex` is https://github.com/ooni/probe/issues/1957. As of this writing the issue still references the original plan, which I could not complete by the end of this Sprint, so I am going to adapt the text of the issue to only refer to what was done in here next. Of course, I also need follow-up issues.
* chore(netxlite): add currently failing test case
This diff introduces a test cases that will fail because of the reason
explained in https://github.com/ooni/probe/issues/1965.
* chore(netxlite/iox_test.go): add failing unit tests
These tests directly show how the Go implementation of ReadAll
and Copy has the issue of checking for io.EOF equality.
* fix(netxlite): make {ReadAll,Copy}Context robust to wrapped io.EOF
The fix is simple: we just need to check for `errors.Is(err, io.EOF)`
after either io.ReadAll or io.Copy has returned. When this condition is
true, we need to convert the error back to `nil` as it ought to be.
While there, observe that the unit tests I committed in the previous
commit are wrongly asserting that the error must be wrapped. This
assertion is not correct, because in both cases we have just ensured
that the returned error is `nil` (i.e., success).
See https://github.com/ooni/probe/issues/1965.
* cleanup: remove previous workaround for wrapped io.EOF
These workarounds were partial, meaning that they would cover some
cases in which the issue occurred but not all of them.
Handling the problem in `netxlite.{ReadAll,Copy}Context` is the
right thing to do _as long as_ we always use these functions instead
of `io.{ReadAll,Copy}`.
This is why it's now important to ensure we clearly mention that
inside of the `CONTRIBUTING.md` guide and to also ensure that we're
not using these functions in the code base.
* fix(urlgetter): repair tests who assumed to see EOF error
Now that we have established that we should normalize EOF when
reading bodies like the stdlib does and now that it's clear why
our behavior diverged from the stdlib, we also need to repair
all the tests that assumed this incorrect behavior.
* fix(all): don't use io{,util}.{Copy,ReadAll}
* feat: add checks to ensure we don't use io.{Copy,ReadAll}
* doc(netxlite): document we know how to deal w/ wrapped io.EOF
* fix(nocopyreadall.bash): add exception for i/n/iox.go
