fa70e23de7 ci: Drop libFuzzer from msan fuzz task (MarcoFalke)
Pull request description:
libFuzzer is mostly unmaintained (https://llvm.org/docs/LibFuzzer.html#status), and it isn't really needed by the CI tasks. While it provides some additional stats like rss or the max input byte size, they are not essential. Dropping libFuzzer here would also drop the "60 seconds sanity check" for empty folders, but I think this is an acceptable price to pay to silence false-positives that were hit for years.
Also, there seems to be a history of intermittent false-positive msan warnings (https://github.com/bitcoin/bitcoin/pull/33600#issuecomment-3391921802).
It is unclear what exactly is causing the false-positives, so just disable libFuzzer in this task for now, to work around them.
ACKs for top commit:
kevkevinpal:
ACK [fa70e23](fa70e23de7)
dergoegge:
ACK fa70e23de7
Tree-SHA512: c3e5958b8378ba30f51d923f97a84dec2ee60af8b9c2a4f13bc8de486a490031468371120e421384aa198ffec591db554e636935ab3c6d4de5e870238f5079f2
fabe0e07de ci: Only write docker build images to Cirrus cache (MarcoFalke)
fab64a5d6f ci: Move buildx command to python script (MarcoFalke)
fa72a2bd5c ci: Remove unused MAYBE_CPUSET (MarcoFalke)
Pull request description:
The `DOCKER_BUILD_CACHE_ARG` env var holds the options on how to use cache providers. Storing the image layers is useful for the Cirrus cache provider, because it offers 10GB per runner (https://cirrus-runners.app/setup/#speeding-up-the-cache). The cached image layers can help to avoid issues when the upstream package manager infra (apt native, apt llvm, pip, apk, git clone, ...) has outages or network issues.
However, on the GitHub Actions cache provider, a *total* cache of 10GB is offered for the whole repo. This cache must be shared with the depends cache, and the ccache, as well as the previous releases cache. So it is already full and trying to put the docker build layers into it will lead to an overflow.
Fix it by only writing to the docker cache on Cirrus.
Also, `DOCKER_BUILD_CACHE_ARG` requires a `shellcheck disable=SC2086` on the full build command. Fix that as well by using `shlex.split` from Python on just this variable.
ACKs for top commit:
m3dwards:
ACK fabe0e07de
cedwies:
reACK fabe0e0
l0rinc:
Code review ACK fabe0e07de
willcl-ark:
ACK fabe0e07de
Tree-SHA512: 4f471f080007fdd0c3bc97b0cfe0e9c0457e5029a7ccde1d784d30eb4752e5eb309cd4b122b182bce31f1b986c8a9f3e9a49da1768bedbb2b1f64f70183680ba
This has a few benefits:
* The shellcheck SC2086 warning is disabled for the whole command, but
is only needed for the DOCKER_BUILD_CACHE_ARG env var. So in Python,
only pass this one env var to shlex.split() for proper word splitting.
* Future logic improvements can be implemented in Python.
The comments are moved, which can be checked via the git options:
--color-moved=dimmed-zebra --color-moved-ws=ignore-all-space
The option is currently unused. If it is used again in the future, it
could trivially be added back.
Also, the logic is just a single undocumented python command one-liner.
So remove it for now.
444409ff2b ci: Reduce Alpine musl task to md runner size (MarcoFalke)
fa6b2e9efe ci: Turn centos config into alpine musl config (MarcoFalke)
Pull request description:
Fixes https://github.com/bitcoin/bitcoin/issues/33437
Historically, the centos task was added to add CI coverage for old packages and 32-bit depends builds, but both are now covered by different tasks.
The CentOS task aligns with Ubuntu/Debian CI tasks in terms of libc usage, but (slightly) differs in package naming and update philosophy. I am not aware of the task ever discovering a centos-related issue, so it seems fine to recycle it into an Alpine Linux task.
The main difference would be that musl libc is now used. Also, busybox is used in Alpine, so in theory the busybox install could be removed from the arm CI task in the future.
Packaging considerations: All packages should roughly be the same (gcc remains at version 14, python remains at version 3.12, etc). Also, all packages are from the Alpine main track, coming with 2 years of support. The only exception is the py3-pip package (https://pkgs.alpinelinux.org/packages?name=py3-pip&branch=v3.22&repo=&arch=riscv64) from the community track, however, I don't expect any issues arising from that.
ACKs for top commit:
janb84:
reACK 444409ff2b
willcl-ark:
ACK 444409ff2b
Tree-SHA512: fd1a1da0fd766591e44a57dbdb84f9b3b47ca92113a429bba139ee5fef54714b8fe509c321e7b3a470c29b4af7d9eab9786e1660b9effb862ecea52824f458aa
e4c04f7759 ci: add libcpp hardening flags to macOS fuzz job (fanquake)
Pull request description:
Follows up to https://github.com/bitcoin/bitcoin/pull/33425#issuecomment-3323149107.
ACKs for top commit:
maflcko:
lgtm ACK e4c04f7759. The qa-assets repo has a libc++ debug run, so this isn't required, but it seems fast enough to not hurt.
Tree-SHA512: 6c0dc90528ca867df49027eebf2d1c417a7395f9f94779076ace48e3e4b21771e7d99e8b3ed75ca56da87153418a446075429aa6b9ec5cd4b6b8cb5b0c25c1d7
b736052e39 ci: always use tag for LLVM checkout (fanquake)
Pull request description:
Rather than trying to match the apt installed clang version, which is prone to intermittent issues. i.e #33345.
ACKs for top commit:
davidgumberg:
ACK b736052e39
willcl-ark:
ACK b736052e39
Tree-SHA512: 8e3fcc8219f573cec65941576c7995f21cae3330bcdbf615f799e8c5facd1146d3239a7284e9af7b013c37170ddf7435d7df6d2966f63fe7b4a8e4937311ff36
fa96a4afea ci: Enable CI_LIMIT_STACK_SIZE=1 in i686_no_ipc task (MarcoFalke)
facfde2cdc test: Fix CLI_MAX_ARG_SIZE issues (MarcoFalke)
Pull request description:
`CLI_MAX_ARG_SIZE` has many edge case issues:
* It seems to be lower on some systems, but it is unknown how to reproduce locally: https://github.com/bitcoin/bitcoin/pull/33079#issuecomment-3139957274
* `MAX_ARG_STRLEN` is a limit per arg, but we probably want "The maximum length of [all of] the arguments": See https://www.man7.org/linux/man-pages/man3/sysconf.3.html, section `ARG_MAX - _SC_ARG_MAX`.
* It doesn't account for the additional args added by the `bitcoin` command later on: 73220fc0f9/src/bitcoin.cpp (L85-L92)
* It doesn't account for unicode encoding a string to bytes before taking its length.
The issues are mostly harmless edge cases, but it would be good to fix them. So do that here, by:
* Replacing `max()` by `sum()`, to correctly take into account all args, not just the largest one.
* Reduce `CLI_MAX_ARG_SIZE`, to account for the `bitcoin` command additional args.
Also, there is a test. The test can be called with `ulimit` to hopefully limit the max args size to the hard-coded value in the test framework. For reference:
```
$ ( ulimit -s 512 && python3 -c 'import os; print(os.sysconf("SC_ARG_MAX") )' )
131072
```
On top of this pull it should pass, ...
```
bash -c 'ulimit -s 512 && BITCOIN_CMD="bitcoin -M" ./bld-cmake/test/functional/rpc_misc.py --usecli -l DEBUG'
```
... and with the test_framework changes reverted, it should fail:
```
OSError: [Errno 7] Argument list too long: 'bitcoin'
```
Also, there is a commit to enable `CI_LIMIT_STACK_SIZE=1` in the i686 task, because it should now be possible and no longer hit the hard-to-reproduce issue mentioned above.
ACKs for top commit:
cedwies:
ACK fa96a4a
achow101:
ACK fa96a4afea
enirox001:
ACK fa96a4a — thanks for addressing the nits and clarifying the test; LGTM.
mzumsande:
Code Review ACK fa96a4afea
Tree-SHA512: d12211bd097d692d560c3615970ec0e911707d8c6cbbb145591abc548beed55f487a80b08f0a8c89d4eef4d76a9fbd6a33edc0b42b5860a93dd7b954355bc887
Install pycapnp on all (active) CI hosts which have IPC enabled and
run the functional tests.
Except for previous_releases, which uses an older version of pip
that doesn't support --break-system-packages.
ci/lint_run.sh: Only used in .cirrus.yml. Refer to test/lint/README.md on how to run locally.
ci/lint_run_all.sh: Only used in .cirrus.yml for stale re-runs of old pull request tasks.
Docker currently warns that we are missing a default value.
Set this to scratch which will error if an appropriate image tag is not
passed in to silence the warning.
Previously jobs were running on a large multi-core server where 10 jobs
as default made sense (or may even have been on the low side).
Using hosted runners with fixed (and lower) numbers of vCPUs we should
adapt compilation to match the number of cpus we have dynamically.
This is cross-platform compatible with macos and linux only.
When using hosted runners in combination with cached docker images,
there is the possibility that the host runner image is updated,
rendering the linux-headers package (stored in the cached docker image)
incompatible.
Fix this by doing a re-install of the headers package in
03_test_script.sh.
If the underlying runner kernel has not changed thie has no effect, but
prevents the job from failing if it has.
This sets the build dir at build time so that Apple SDK gets installed
in the correct/expected location for the runtime to find it.
Co-authored-by: Max Edwards <youwontforgetthis@gmail.com>
Reverts: e87429a2d0
This was added in PR #31545 with the intention that self-hosted runners
might use it to save build cache.
As we are not using hosted runners with a registry build cache, the bulk
of this commit can be reverted, simply using the value of
$DOCKER_BUILD_CACHE_ARG in the script.
link: https://github.com/bitcoin/bitcoin/pull/31545
Using buildx is required to properly load the correct driver, for use
with registry caching. Neither build, nor BUILDKIT=1 currently do this
properly.
Use of `docker buildx build` is compatible with podman.
The bitcoin-node binary is built on all platforms which have
multiprocess enabled, but for functional tests it's only used in
CentOS native (depends) job. The next commit will also add a
non-depends job.
This causes IPC binaries (bitcoin-node, bitcoin-gui) to be included
in releases.
The effect on CI is that this causes more depends builds to build IPC
binaries, but still the only build running functional tests with them
is the i686_multiprocess one.
Except for Windows.
f49840dd90 doc: Fix typo in files.md (Ryan Ofsky)
f5cf0b1ccc bitcoin wrapper: improve help output (Ryan Ofsky)
c810b168b8 doc: Add description of installed files to files.md (Ryan Ofsky)
94ffd01a02 doc: Add release notes describing libexec/ binaries (Ryan Ofsky)
cd97905ebc cmake: Move internal binaries from bin/ to libexec/ (Ryan Ofsky)
Pull request description:
This change moves binaries that are not typically invoked directly by users from the `bin/` directory to the `libexec/` directory in CMake installs and binary releases. The goal of the PR is to introduce a distinction between internal and external binaries so starting with #31802, we can use IPC to implement features in new binaries without adding those binaries to the CLI. The change also helps reduce clutter in `bin/`, making it easier for users to identify useful tools to run. Summary of changes:
- For **source builds** (i.e. developer builds) — There are no changes.
- For **source installs** (i.e. `cmake --install` result) — `test_bitcoin`, `test_bitcoin-qt`, and `bench_bitcoin` are installed in `${CMAKE_PREFIX_PATH}/libexec` instead of `${CMAKE_PREFIX_PATH}/bin`, so they are no longer on the system `PATH`. However, they can still be invoked from the `libexec/` directory, or from the CLI as `bitcoin test`, `bitcoin test-gui`, and `bitcoin bench`, respectively.
- For **binary releases** — Since `test_bitcoin` is the only test binary enabled in releases, the only change is moving `test_bitcoin` from `bin/` to `libexec/`.
<details><summary>Details</summary>
<p>
The table below shows the install location of each binary after this change, and the availability of each binary.
| Binary | Location | Availability | Change |
|----------------------|--------------|----------------------|-------------------------------|
| `bitcoin` | `bin/` | 📦 Binary release (since #31375) | Unchanged |
| `bitcoin-cli` | `bin/` | 📦 Binary release | Unchanged |
| `bitcoind` | `bin/` | 📦 Binary release | Unchanged |
| `bitcoin-qt` | `bin/` | 📦 Binary release | Unchanged |
| `bitcoin-tx` | `bin/` | 📦 Binary release | Unchanged |
| `bitcoin-util` | `bin/` | 📦 Binary release | Unchanged |
| `bitcoin-wallet` | `bin/` | 📦 Binary release | Unchanged |
| `bench_bitcoin` | `libexec/` | 🛠 Source build only | Moved from `bin/` |
| `bitcoin-chainstate` | `libexec/` | 🛠 Source build only | Newly installed (was built) |
| `bitcoin-gui` | `libexec/` | 🛠 Source build only (until #31802) | Moved from `bin/` |
| `bitcoin-node` | `libexec/` | 🛠 Source build only (until #31802) | Moved from `bin/` |
| `test_bitcoin` | `libexec/` | 📦 Binary release | Moved from `bin/` |
| `test_bitcoin-qt` | `libexec/` | 🛠 Source build only | Moved from `bin/` |
</p>
</details>
---
This PR is part of the [process separation project](https://github.com/bitcoin/bitcoin/issues/28722).
ACKs for top commit:
l0rinc:
re-ACK f49840dd90
Sjors:
re-ACK f49840dd90
achow101:
ACK f49840dd90
janb84:
re ACK f49840dd90
BrandonOdiwuor:
Tested ACK f49840dd90
hodlinator:
re-ACK f49840dd90
willcl-ark:
utACK f49840dd90
Tree-SHA512: 858a2e1a53db11ee3c5c759bfdeea566f242b9ce5e8a898fa435222e41662b8184577c0dc2c4c058294b4de41d8cb3ba3e5d24c748c280efa4a3f84e3ec4344d
3333d3f75f ci: Only pass documented env vars (MarcoFalke)
Pull request description:
The CI currently inherits almost all env vars from the host. This was problematic in the past and causing non-determinism, e.g. the fix in commit fa12558d21. It is still problematic today, see e.g. https://github.com/bitcoin/bitcoin/pull/31349#issuecomment-2586828644, or https://github.com/bitcoin/bitcoin/issues/32935
This fixes https://github.com/bitcoin/bitcoin/issues/32935 by only passing env vars documented in `./ci/test/00_setup_env.sh`.
Implementation-wise, instead of cramming the python code into the `python -c ""` statement, just start a fresh py file, which is easier to handle.
ACKs for top commit:
willcl-ark:
ACK 3333d3f75f
Tree-SHA512: f922e481a844128d7fbf773563278a3992c178ead60a3050eceb9ded2aad979afc815a5cbdb9f68494493c5d8d942cdd1111c21e32a5746d19505b87745cb84a
