088bcae4ca
Signed-off-by: ljedrz <ljedrz@gmail.com> |
||
---|---|---|
.. | ||
patches | ||
test | ||
.gitignore | ||
package-lock.json | ||
package.json | ||
README.md | ||
rust.sh | ||
setup.sh |
Conformance testing for Rust IPFS
This directory contains the scripts used to run interface conformance testing
for Rust IPFS. It uses js-ipfsd-ctl
, ipfs-http-client
, and
interface-js-ipfs-core
. This code used to live at
https://github.com/rs-ipfs/ipfs-rust-conformance, but was integrated into
https://github.com/rs-ipfs/rust-ipfs, as working with multiple repositories got
a bit tedious.
Usage
Use setup.sh
to do npm install
and patch any of the dependencies:
$ cargo build -p ipfs-http
$ cd conformance
$ ./setup.sh
By default, there is a http
symlink to ../target/debug/ipfs-http
. You can
change this to the release binary by modifying the symlink or use a custom
binary via the environment variable IPFS_RUST_EXEC
. The default rust.sh
wrapper will record all actions taken by the tests into a single large log
file. It's not recommended to trust it to keep all log lines especially for
tests with multiple processes.
$ IPFS_RUST_EXEC="$(pwd)/rust.sh" npm test
$ cat /tmp/rust.log
Obtaining logs for tests with multiple processes
Patch the rust.sh
as follows:
-./http "$@" 2>&1 | tee -a /tmp/rust.log || retval=$?
+./http "$@" 2>&1 | tee -a /tmp/rust.log.$$ || retval=$?
Now the /tmp/rust.log
will contain only the "pointers" to other log files, for example:
>>>> new execution 24132 with args: daemon
<<<< exiting 24132 with 0
This means there is now a log file /tmp/rust.log.24132
for that invocation.
Additionally, it helps to clear out the logs often with rm -f /tmp/rust.log*
and only run selected tests using IPFS_RUST_EXEC="$(pwd)/rust.sh" npm test -- --grep 'should do foo'
.
If it's impossible to limit the number of tests to one with --grep
, you can
comment out the undesired tests in test/index.js
.
Patch management
We are currently pinned to interface-ipfs-core@0.137.0
and the fixes we have upstreamed are kept under patches/
.
To create a new patch:
- Fork https://github.com/ipfs/js-ipfs
- Clone locally
- Checkout a new branch based on the tag for the
interface-ipfs-core
version we are currently depending on (seepackage.json
) - Apply all of our patches from
patches/
withgit apply $ipfs_rust_conformance/patches/*
- Fix any new tests
- When done, cherry-pick your commits to a new branch based off the latest
main
js-ipfs
branch - Submit PR
- Squash your work, refer to your PR in the message
- Remove existing
patches/
- Recreate the
patches/
withgit format-patch -o $ipfs-rust-conformance/patches $the_tag_you_based_your_work
$variable
definitions:
$ipfs_rust_conformance
points to your checkout of this repository$the_tag_you_based_your_work
is the tag where you started working on js-ipfs
Previously this has been done by first working on a patch and then backporting it, which might be easier. Luckily the tests are low traffic. If you need help wrestling with git, don't hesitate to ask for guidance; this is simpler than its step by step explanation looks like.
Troubleshooting hangs
If the tests hang because of some deadlock on the ipfs-http
side the tests
will print the final summary from mocha
and then seem to wait forever. This
is true at least for the hangs seen so far, but you could run into a different
issues. Note that in addition to test timeouts the http client also has a
timeout, which can keep the npm test
alive for less than ten seconds after
the summary has been printed.
What has worked previously is:
- keep disabling some tests until you find the one which causes the hang
- rerun tests using the
rust.sh
wrapper which gives you logs at/tmp/rust.log
- continue debugging over at
ipfs-http
To disable other test means to comment them or to use IPFS_RUST_EXEC=... npm test -- --grep '<suspected test>'
. You should get to a single running test which
hangs. If the test does a lot, you can refactor that into a smaller one which
will only cause the hang and nothing else.
The use of rust.sh
wrapper is critical as it'll give you the logging output.
If the test has multiple running instances you might be better off separating
logs into files per invocation of rust.sh
by appending .$$
to the log files
name, which will expand to the process id of the shell running the script.
To "continue debugging" is trickier. What has worked previously is:
- attaching a
gdb
to a running process in case of livelocks - adding debugging in case everything stalls
Livelocks (I might be using a wrong term here) happen when a task (running on
either tokio or async-std) never returns from std::future::Future::poll
but
stays busy all of the time. You can spot these by seeing a core being
utilitized by ipfs-http
constantly. These are easy to track down by:
- attach
gdb -p $process_id path/to/your/ipfs-http
- find the interesting thread with
info threads
or by looking at threads' stack traces
In the "everything stalled" case, a std::future::Future::poll
has completed
with Poll::Pending
without waking up the task for a new poll. These mistakes are quite
simple to make. Good indications of such issues:
- custom
poll
methods without the [std::task::Context
] parameter: these methods will never be able to schedule wake-ups - polling some nested "pollable" thing and returning
Poll::Pending
following the nested poll returningPoll::Ready(_)
- if the inner "pollable" didn't return
Poll::Pending
, it means it had "more values to bubble up" - see this hastly written issue https://github.com/libp2p/rust-libp2p/issues/1516 and the linked commit(s)
- if the inner "pollable" didn't return
- custom [
std::future::Future
] which cannot return errors on drop like with the earlySubscriptionFuture
(see https://github.com/ipfs-rust/rust-ipfs/pull/130)
License
Same as the repository.
Trademarks
The Rust logo and wordmark are trademarks owned and protected by the Mozilla Foundation. The Rust and Cargo logos (bitmap and vector) are owned by Mozilla and distributed under the terms of the Creative Commons Attribution license (CC-BY).