[OSFI001] Build process
[OSFI0010] There is no healthy OSF without reproducible builds
What are the reproducible builds?
Reproducible Builds and a lot of its aspects is well explained on the linked website.
Why are reproducible builds crucial OSF?
This is because knowing hashes of firmware components, we can calculate expected TPM PCRs values. By knowing good PCR values and being able to predict further values after firmware update, we are gaining the following advantages:
We can easily confirm if the firmware is valid by reading TPM PCRs and compare to reference integrity measures (RIM), RIM can be calculated by build exactly the same version of OSF.
NOTE: Please note that depending on your threat model using TPM chip with closed source firmware implementation may lead to solution which is not trustworthy.
Process of replying measurements provided in TPM Event Log is simpler because we can easily obtain software and confirm hashes used during PCR extension the process was valid and the same as produced by the software build process
- Looking for bugs by using bisection is easier since being able to build every commit of software in a reproducible manner give us confidence that there would be no change in behavior between 2 consecutive builds
- Consistent behavior for given hash can be confirmed by test results tied to given hash in that way users looking at test results may expect exactly the same behavior for firmware binaries with the same hash
- Long term maintenance should be easier if firmware build would be reproducible over a long time, because often happen some firmware land in industrial applications, e.g., robots, trains, smart city infrastructure, if the bug will be detected after a long time; it is important to have tools and be able to confirm the same binary can be generated in the future. If we know given toolchain reproduce binary bit by bit we are sure it also reproduces software/firmware behavior. In that light making small incremental improvement using a toolchain that gives reproducible results give us higher the chance that we will not introduce uncontrolled change in behavior and code the change will be reflected by the compilation process without affecting previous software behavior.
[OSFI0011] Docker containers as build systems "considered harmful."
Docker containers for a couple of last years become the default method for providing reproducible runtime environment for software as well those started to be very useful for developers to transfer exact configuration for building given software stack. We wrote about the usefulness of Docker containers in embedded software development environment on 3mdeb blog.
After a long time of using Docker containers for embedded software development and build environments we noticed problems for long time maintenance and reproducibility.
The example can be
coreboot-sdk used for building
A good summary of the problem was provided by Thrilleratplay in guix-docker repo:
Currently, the coreboot build environment,
coreboot-sdk, uses a Debian docker base image. To install additional required packages,
apt-get updatemust be run. The resulting Docker image is hosted in the Docker hub repository to be retrieved at any time in the future. However, at any time in the future, building the same docker file will generate a different image based on the latest packages used in apt-get. Over time, as packages are updated due to bug, security or feature improvements, the docker image's provenance in the docker hub repository becomes increasingly difficult, if not impossible, to audit and reproduce.
Some more detailed notes related to
coreboot-sdk issues are presented below.
- There is no meaningful versioning of SDK  - at some point, there was versioning 1.32-53, now switched to git SHA. Both seemed to have no meaning and were released at arbitrary points in time. The result is that it is hard to find which version of SDK works with the coreboot tree version.
- Validation of coreboot-sdk is not sufficient - for example, recently, python fixes were merged , somehow this change passed all QA checks, but SeaBIOS use python (not python2 or python3) and all builds using this version of SDK that compile SeaBIOS fail.
coreboot-sdkis based on moving target Debian sid. It is close to impossible to build the same Docker images at 2 different points in time.
- coreboot-sdk enforce given version of ACPI spec - this may not always be a good thing to use the most recent compiler and update code accordingly since it can easily break OSes.
Why we care?
- Whenever we deliver code or service to community or customer, we provide build environment which, for the sake of quality user support, should be stable.
- CI pipelines rely on those containers. Replacing docker image in CI pipeline whenever something change defeat the purpose of having automation and increase maintenance cost significantly and increase maintenance cost significantly
Requirements for OSF dev and build process
Open-source firmware development and build environment SHOULD have: * meaningful release process * meaningful validation process, at least basic build system and dev env capabilities should be tested * meaningful revisions with a clear explanation of what software stack can be build with what version of dev and build system, a description should include side-spec compliance, e.g., ACPI, SMBIOS, UEFI, etc. * reproducible process that works across systems and in the long run * signature, so anyone can identify where build and dev env coming from and if it is trustworthy
Ideas for solving the above issues
- Docker images are not reproducible - this is a known fact, and if we are extremely serious about stability, security and quality, we should stop using a not reproducible build environment. Of course, there is some work making Docker images reproducible, but it opens Pandora's box  of Yocto or Buildroot. Another path could be Nix or Guix, which is currently pursued in the community OSFW #guix-buildstack channel
- this is a private channel; feel free to ping anyone from 3mdeb Team to join.
- coreboot toolchain is built by its own build system, which would be hard to couple with Dockerfile or something else.
- Docker images can be signed using
docker trustcommands as described here.
- Other idea would be to use
dpkg -ifor every needed package in
Dockerfilethat create base image for reproducible toolchain. Of course that means quite extensive dependency management, but maybe simplicity is worth the effort.
- It is also possible to maintain VMs per given coreboot (or other OSF) version. That may mean long term VM image compatibility issues as well as problems with making sure images would work with various hypervisor versions. It is even more complex if multiple OS should be supported. VMs also have quite big overhead in terms of performance
I'm not tracking all activity in the community, so some claims may be plain wrong. Please let me know what I miss when I'm wrong.