4 Reproducible Builds
4.1 How we define it
The Yocto Project defines reproducibility as where a given input build configuration will give the same binary output regardless of when it is built (now or in 5 years time), regardless of the path on the filesystem the build is run in, and regardless of the distro and tools on the underlying host system the build is running on.
4.2 Why it matters
The project aligns with the Reproducible Builds project, which shares information about why reproducibility matters. The primary focus of the project is the ability to detect security issues being introduced. However, from a Yocto Project perspective, it is also hugely important that our builds are deterministic. When you build a given input set of metadata, we expect you to get consistent output. This has always been a key focus but, since release 3.1 (“dunfell”), it is now true down to the binary level including timestamps.
For example, at some point in the future life of a product, you find that you need to rebuild to add a security fix. If this happens, only the components that have been modified should change at the binary level. This would lead to much easier and clearer bounds on where validation is needed.
This also gives an additional benefit to the project builds themselves, our Hash Equivalence for Shared State object reuse works much more effectively when the binary output remains the same.
We strongly advise you to make sure your project builds reproducibly before finalizing your production images. It would be too late if you only address this issue when the first updates are required.
4.3 How we implement it
There are many different aspects to build reproducibility, but some particular things we do within the build system to ensure reproducibility include:
Adding mappings to the compiler options to ensure debug filepaths are mapped to consistent target compatible paths. This is done through the DEBUG_PREFIX_MAP variable which sets the
-fdebug-prefix-mapcompiler options correctly to map to target paths.
Being explicit about recipe dependencies and their configuration (no floating configure options or host dependencies creeping in). In particular this means making sure PACKAGECONFIG coverage covers configure options which may otherwise try and auto-detect host dependencies.
Using recipe specific sysroots to isolate recipes so they only see their dependencies. These are visible as
recipe-sysroot-nativedirectories within the WORKDIR of a given recipe and are populated only with the dependencies a recipe has.
Build images from a reduced package set: only packages from recipes the image depends upon.
Filtering the tools available from the host’s
PATHto only a specific set of tools, set using the HOSTTOOLS variable.
Because of an open bug in GCC, using
DISTRO_FEATURES:append = " lto" or
-flto (Link Time Optimization) to CFLAGS makes the resulting
binary non-reproducible, in that it depends on the full absolute build path
recipe-sysroot-native, so installing the Yocto Project in a different
directory results in a different binary.
This issue is addressed by bug 14481 - Programs built with -flto are not reproducible.
4.4 Can we prove the project is reproducible?
Yes, we can prove it and we regularly test this on the Autobuilder. At the time of writing (release 3.3, “hardknott”), OpenEmbedded-Core (OE-Core) is 100% reproducible for all its recipes (i.e. world builds) apart from the Go language and Ruby documentation packages. Unfortunately, the current implementation of the Go language has fundamental reproducibility problems as it always depends upon the paths it is built in.
Only BitBake and OpenEmbedded-Core (OE-Core), which is the
layer in Poky, guarantee complete reproducibility. The moment you add
another layer, this warranty is voided, because of additional configuration
bbappend files, overridden classes, etc.
To run our automated selftest, as we use in our CI on the Autobuilder, you can run:
oe-selftest -r reproducible.ReproducibleTests.test_reproducible_builds
This defaults to including a
world build so, if other layers are added, it would
also run the tests for recipes in the additional layers. The first build will be
run using Shared State if
available, the second build explicitly disables
Shared State and builds on the
specific host the build is running on. This means we can test reproducibility
builds between different host distributions over time on the Autobuilder.
OEQA_DEBUGGING_SAVED_OUTPUT is set, any differing packages will be saved
here. The test is also able to run the
diffoscope command on the output to
generate HTML files showing the differences between the packages, to aid
debugging. On the Autobuilder, these appear under
https://autobuilder.yocto.io/pub/repro-fail/ in the form
<date> + <random ID>, e.g.
The project’s current reproducibility status can be seen at https://www.yoctoproject.org/reproducible-build-results/
You can also check the reproducibility status on supported host distributions:
4.5 Can I test my layer or recipes?
Once again, you can run a
world test using the
command provided above. This functionality is implemented
You could subclass the test and change
targets to a different target.
You may also change
sstate_targets which would allow you to “pre-cache” some
set of recipes before the test, meaning they are excluded from reproducibility
testing. As a practical example, you could set
core-image-sato, then setting
run reproducibility tests only on the targets belonging only to