Reproducible and minimal source-only tarballs

With the release of Libntlm version 1.8 the release tarball can be reproduced on several distributions. We also publish a signed minimal source-only tarball, produced by git-archive which is the same format used by Savannah, Codeberg, GitLab, GitHub and others. Reproducibility of both tarballs are tested continuously for regressions on GitLab through a CI/CD pipeline. If that wasn’t enough to excite you, the Debian packages of Libntlm are now built from the reproducible minimal source-only tarball. The resulting binaries are reproducible on several architectures.

What does that even mean? Why should you care? How you can do the same for your project? What are the open issues? Read on, dear reader…

This article describes my practical experiments with reproducible release artifacts, following up on my earlier thoughts that lead to discussion on Fosstodon and a patch by Janneke Nieuwenhuizen to make Guix tarballs reproducible that inspired me to some practical work.

Let’s look at how a maintainer release some software, and how a user can reproduce the released artifacts from the source code. Libntlm provides a shared library written in C and uses GNU Make, GNU Autoconf, GNU Automake, GNU Libtool and gnulib for build management, but these ideas should apply to most project and build system. The following illustrate the steps a maintainer would take to prepare a release:

git clone https://gitlab.com/gsasl/libntlm.git
cd libntlm
git checkout v1.8
./bootstrap
./configure
make distcheck
gpg -b libntlm-1.8.tar.gz

The generated files libntlm-1.8.tar.gz and libntlm-1.8.tar.gz.sig are published, and users download and use them. This is how the GNU project have been doing releases since the late 1980’s. That is a testament to how successful this pattern has been! These tarballs contain source code and some generated files, typically shell scripts generated by autoconf, makefile templates generated by automake, documentation in formats like Info, HTML, or PDF. Rarely do they contain binary object code, but historically that happened.

The XZUtils incident illustrate that tarballs with files that are not included in the git archive offer an opportunity to disguise malicious backdoors. I blogged earlier how to mitigate this risk by using signed minimal source-only tarballs.

The risk of hiding malware is not the only motivation to publish signed minimal source-only tarballs. With pre-generated content in tarballs, there is a risk that GNU/Linux distributions such as Trisquel, Guix, Debian/Ubuntu or Fedora ship generated files coming from the tarball into the binary *.deb or *.rpm package file. Typically the person packaging the upstream project never realized that some installed artifacts was not re-built through a typical autoconf -fi && ./configure && make install sequence, and never wrote the code to rebuild everything. This can also happen if the build rules are written but are buggy, shipping the old artifact. When a security problem is found, this can lead to time-consuming situations, as it may be that patching the relevant source code and rebuilding the package is not sufficient: the vulnerable generated object from the tarball would be shipped into the binary package instead of a rebuilt artifact. For architecture-specific binaries this rarely happens, since object code is usually not included in tarballs — although for 10+ years I shipped the binary Java JAR file in the GNU Libidn release tarball, until I stopped shipping it. For interpreted languages and especially for generated content such as HTML, PDF, shell scripts this happens more than you would like.

Publishing minimal source-only tarballs enable easier auditing of a project’s code, to avoid the need to read through all generated files looking for malicious content. I have taken care to generate the source-only minimal tarball using git-archive. This is the same format that GitLab, GitHub etc offer for the automated download links on git tags. The minimal source-only tarballs can thus serve as a way to audit GitLab and GitHub download material! Consider if/when hosting sites like GitLab or GitHub has a security incident that cause generated tarballs to include a backdoor that is not present in the git repository. If people rely on the tag download artifact without verifying the maintainer PGP signature using GnuPG, this can lead to similar backdoor scenarios that we had for XZUtils but originated with the hosting provider instead of the release manager. This is even more concerning, since this attack can be mounted for some selected IP address that you want to target and not on everyone, thereby making it harder to discover.

With all that discussion and rationale out of the way, let’s return to the release process. I have added another step here:

make srcdist
gpg -b libntlm-1.8-src.tar.gz

Now the release is ready. I publish these four files in the Libntlm’s Savannah Download area, but they can be uploaded to a GitLab/GitHub release area as well. These are the SHA256 checksums I got after building the tarballs on my Trisquel 11 aramo laptop:

91de864224913b9493c7a6cec2890e6eded3610d34c3d983132823de348ec2ca  libntlm-1.8-src.tar.gz
ce6569a47a21173ba69c990965f73eb82d9a093eb871f935ab64ee13df47fda1  libntlm-1.8.tar.gz

So how can you reproduce my artifacts? Here is how to reproduce them in a Ubuntu 22.04 container:

podman run -it --rm ubuntu:22.04
apt-get update
apt-get install -y --no-install-recommends autoconf automake libtool make git ca-certificates
git clone https://gitlab.com/gsasl/libntlm.git
cd libntlm
git checkout v1.8
./bootstrap
./configure
make dist srcdist
sha256sum libntlm-*.tar.gz

You should see the exact same SHA256 checksum values. Hooray!

This works because Trisquel 11 and Ubuntu 22.04 uses the same version of git, autoconf, automake, and libtool. These tools do not guarantee the same output content for all versions, similar to how GNU GCC does not generate the same binary output for all versions. So there is still some delicate version pairing needed.

Ideally, the artifacts should be possible to reproduce from the release artifacts themselves, and not only directly from git. It is possible to reproduce the full tarball in a AlmaLinux 8 container – replace almalinux:8 with rockylinux:8 if you prefer RockyLinux:

podman run -it --rm almalinux:8
dnf update -y
dnf install -y make wget gcc
wget https://download.savannah.nongnu.org/releases/libntlm/libntlm-1.8.tar.gz
tar xfa libntlm-1.8.tar.gz
cd libntlm-1.8
./configure
make dist
sha256sum libntlm-1.8.tar.gz

The source-only minimal tarball can be regenerated on Debian 11:

podman run -it --rm debian:11
apt-get update
apt-get install -y --no-install-recommends make git ca-certificates
git clone https://gitlab.com/gsasl/libntlm.git
cd libntlm
git checkout v1.8
make -f cfg.mk srcdist
sha256sum libntlm-1.8-src.tar.gz 

As the Magnus Opus or chef-d’œuvre, let’s recreate the full tarball directly from the minimal source-only tarball on Trisquel 11 – replace docker.io/kpengboy/trisquel:11.0 with ubuntu:22.04 if you prefer.

podman run -it --rm docker.io/kpengboy/trisquel:11.0
apt-get update
apt-get install -y --no-install-recommends autoconf automake libtool make wget git ca-certificates
wget https://download.savannah.nongnu.org/releases/libntlm/libntlm-1.8-src.tar.gz
tar xfa libntlm-1.8-src.tar.gz
cd libntlm-v1.8
./bootstrap
./configure
make dist
sha256sum libntlm-1.8.tar.gz

Yay! You should now have great confidence in that the release artifacts correspond to what’s in version control and also to what the maintainer intended to release. Your remaining job is to audit the source code for vulnerabilities, including the source code of the dependencies used in the build. You no longer have to worry about auditing the release artifacts.

I find it somewhat amusing that the build infrastructure for Libntlm is now in a significantly better place than the code itself. Libntlm is written in old C style with plenty of string manipulation and uses broken cryptographic algorithms such as MD4 and single-DES. Remember folks: solving supply chain security issues has no bearing on what kind of code you eventually run. A clean gun can still shoot you in the foot.

Side note on naming: GitLab exports tarballs with pathnames libntlm-v1.8/ (i.e.., PROJECT-TAG/) and I’ve adopted the same pathnames, which means my libntlm-1.8-src.tar.gz tarballs are bit-by-bit identical to GitLab’s exports and you can verify this with tools like diffoscope. GitLab name the tarball libntlm-v1.8.tar.gz (i.e., PROJECT-TAG.ARCHIVE) which I find too similar to the libntlm-1.8.tar.gz that we also publish. GitHub uses the same git archive style, but unfortunately they have logic that removes the ‘v’ in the pathname so you will get a tarball with pathname libntlm-1.8/ instead of libntlm-v1.8/ that GitLab and I use. The content of the tarball is bit-by-bit identical, but the pathname and archive differs. Codeberg (running Forgejo) uses another approach: the tarball is called libntlm-v1.8.tar.gz (after the tag) just like GitLab, but the pathname inside the archive is libntlm/, otherwise the produced archive is bit-by-bit identical including timestamps. Savannah’s CGIT interface uses archive name libntlm-1.8.tar.gz with pathname libntlm-1.8/, but otherwise file content is identical. Savannah’s GitWeb interface provides snapshot links that are named after the git commit (e.g., libntlm-a812c2ca.tar.gz with libntlm-a812c2ca/) and I cannot find any tag-based download links at all. Overall, we are so close to get SHA256 checksum to match, but fail on pathname within the archive. I’ve chosen to be compatible with GitLab regarding the content of tarballs but not on archive naming. From a simplicity point of view, it would be nice if everyone used PROJECT-TAG.ARCHIVE for the archive filename and PROJECT-TAG/ for the pathname within the archive. This aspect will probably need more discussion.

Side note on git archive output: It seems different versions of git archive produce different results for the same repository. The version of git in Debian 11, Trisquel 11 and Ubuntu 22.04 behave the same. The version of git in Debian 12, AlmaLinux/RockyLinux 8/9, Alpine, ArchLinux, macOS homebrew, and upcoming Ubuntu 24.04 behave in another way. Hopefully this will not change that often, but this would invalidate reproducibility of these tarballs in the future, forcing you to use an old git release to reproduce the source-only tarball. Alas, GitLab and most other sites appears to be using modern git so the download tarballs from them would not match my tarballs – even though the content would.

Side note on ChangeLog: ChangeLog files were traditionally manually curated files with version history for a package. In recent years, several projects moved to dynamically generate them from git history (using tools like git2cl or gitlog-to-changelog). This has consequences for reproducibility of tarballs: you need to have the entire git history available! The gitlog-to-changelog tool also output different outputs depending on the time zone of the person using it, which arguable is a simple bug that can be fixed. However this entire approach is incompatible with rebuilding the full tarball from the minimal source-only tarball. It seems Libntlm’s ChangeLog file died on the surgery table here.

So how would a distribution build these minimal source-only tarballs? I happen to help on the libntlm package in Debian. It has historically used the generated tarballs as the source code to build from. This means that code coming from gnulib is vendored in the tarball. When a security problem is discovered in gnulib code, the security team needs to patch all packages that include that vendored code and rebuild them, instead of merely patching the gnulib package and rebuild all packages that rely on that particular code. To change this, the Debian libntlm package needs to Build-Depends on Debian’s gnulib package. But there was one problem: similar to most projects that use gnulib, Libntlm depend on a particular git commit of gnulib, and Debian only ship one commit. There is no coordination about which commit to use. I have adopted gnulib in Debian, and add a git bundle to the *_all.deb binary package so that projects that rely on gnulib can pick whatever commit they need. This allow an no-network GNULIB_URL and GNULIB_REVISION approach when running Libntlm’s ./bootstrap with the Debian gnulib package installed. Otherwise libntlm would pick up whatever latest version of gnulib that Debian happened to have in the gnulib package, which is not what the Libntlm maintainer intended to be used, and can lead to all sorts of version mismatches (and consequently security problems) over time. Libntlm in Debian is developed and tested on Salsa and there is continuous integration testing of it as well, thanks to the Salsa CI team.

Side note on git bundles: unfortunately there appears to be no reproducible way to export a git repository into one or more files. So one unfortunate consequence of all this work is that the gnulib *.orig.tar.gz tarball in Debian is not reproducible any more. I have tried to get Git bundles to be reproducible but I never got it to work — see my notes in gnulib’s debian/README.source on this aspect. Of course, source tarball reproducibility has nothing to do with binary reproducibility of gnulib in Debian itself, fortunately.

One open question is how to deal with the increased build dependencies that is triggered by this approach. Some people are surprised by this but I don’t see how to get around it: if you depend on source code for tools in another package to build your package, it is a bad idea to hide that dependency. We’ve done it for a long time through vendored code in non-minimal tarballs. Libntlm isn’t the most critical project from a bootstrapping perspective, so adding git and gnulib as Build-Depends to it will probably be fine. However, consider if this pattern was used for other packages that uses gnulib such as coreutils, gzip, tar, bison etc (all are using gnulib) then they would all Build-Depends on git and gnulib. Cross-building those packages for a new architecture will therefor require git on that architecture first, which gets circular quick. The dependency on gnulib is real so I don’t see that going away, and gnulib is a Architecture:all package. However, the dependency on git is merely a consequence of how the Debian gnulib package chose to make all gnulib git commits available to projects: through a git bundle. There are other ways to do this that doesn’t require the git tool to extract the necessary files, but none that I found practical — ideas welcome!

Finally some brief notes on how this was implemented. Enabling bootstrappable source-only minimal tarballs via gnulib’s ./bootstrap is achieved by using the GNULIB_REVISION mechanism, locking down the gnulib commit used. I have always disliked git submodules because they add extra steps and has complicated interaction with CI/CD. The reason why I gave up git submodules now is because the particular commit to use is not recorded in the git archive output when git submodules is used. So the particular gnulib commit has to be mentioned explicitly in some source code that goes into the git archive tarball. Colin Watson added the GNULIB_REVISION approach to ./bootstrap back in 2018, and now it no longer made sense to continue to use a gnulib git submodule. One alternative is to use ./bootstrap with --gnulib-srcdir or --gnulib-refdir if there is some practical problem with the GNULIB_URL towards a git bundle the GNULIB_REVISION in bootstrap.conf.

The srcdist make rule is simple:

git archive --prefix=libntlm-v1.8/ -o libntlm-1.8-src.tar.gz HEAD

Making the make dist generated tarball reproducible can be more complicated, however for Libntlm it was sufficient to make sure the modification times of all files were set deterministically to the timestamp of the last commit in the git repository. Interestingly there seems to be a couple of different ways to accomplish this, Guix doesn’t support minimal source-only tarballs but rely on a .tarball-timestamp file inside the tarball. Paul Eggert explained what TZDB is using some time ago. The approach I’m using now is fairly similar to the one I suggested over a year ago. If there are problems because all files in the tarball now use the same modification time, there is a solution by Bruno Haible that could be implemented.

Side note on git tags: Some people may wonder why not verify a signed git tag instead of verifying a signed tarball of the git archive. Currently most git repositories uses SHA-1 for git commit identities, but SHA-1 is not a secure hash function. While current SHA-1 attacks can be detected and mitigated, there are fundamental doubts that a git SHA-1 commit identity uniquely refers to the same content that was intended. Verifying a git tag will never offer the same assurance, since a git tag can be moved or re-signed at any time. Verifying a git commit is better but then we need to trust SHA-1. Migrating git to SHA-256 would resolve this aspect, but most hosting sites such as GitLab and GitHub does not support this yet. There are other advantages to using signed tarballs instead of signed git commits or git tags as well, e.g., tar.gz can be a deterministically reproducible persistent stable offline storage format but .git sub-directory trees or git bundles do not offer this property.

Doing continous testing of all this is critical to make sure things don’t regress. Libntlm’s pipeline definition now produce the generated libntlm-*.tar.gz tarballs and a checksum as a build artifact. Then I added the 000-reproducability job which compares the checksums and fails on mismatches. You can read its delicate output in the job for the v1.8 release. Right now we insists that builds on Trisquel 11 match Ubuntu 22.04, that PureOS 10 builds match Debian 11 builds, that AlmaLinux 8 builds match RockyLinux 8 builds, and AlmaLinux 9 builds match RockyLinux 9 builds. As you can see in pipeline job output, not all platforms lead to the same tarballs, but hopefully this state can be improved over time. There is also partial reproducibility, where the full tarball is reproducible across two distributions but not the minimal tarball, or vice versa.

If this way of working plays out well, I hope to implement it in other projects too.

What do you think? Happy Hacking!

Towards reproducible minimal source code tarballs? On *-src.tar.gz

While the work to analyze the xz backdoor is in progress, several ideas have been suggested to improve the software supply chain ecosystem. Some of those ideas are good, some of the ideas are at best irrelevant and harmless, and some suggestions are plain bad. I’d like to attempt to formalize two ideas, which have been discussed before, but the context in which they can be appreciated have not been as clear as it is today.

  1. Reproducible tarballs. The idea is that published source tarballs should be possible to reproduce independently somehow, and that this should be continuously tested and verified — preferrably as part of the upstream project continuous integration system (e.g., GitHub action or GitLab pipeline). While nominally this looks easy to achieve, there are some complex matters in this, for example: what timestamps to use for files in the tarball? I’ve brought up this aspect before.
  2. Minimal source tarballs without generated vendor files. Most GNU Autoconf/Automake-based tarballs pre-generated files which are important for bootstrapping on exotic systems that does not have the required dependencies. For the bootstrapping story to succeed, this approach is important to support. However it has become clear that this practice raise significant costs and risks. Most modern GNU/Linux distributions have all the required dependencies and actually prefers to re-build everything from source code. These pre-generated extra files introduce uncertainty to that process.

My strawman proposal to improve things is to define new tarball format *-src.tar.gz with at least the following properties:

  1. The tarball should allow users to build the project, which is the entire purpose of all this. This means that at least all source code for the project has to be included.
  2. The tarballs should be signed, for example with PGP or minisign.
  3. The tarball should be possible to reproduce bit-by-bit by a third party using upstream’s version controlled sources and a pointer to which revision was used (e.g., git tag or git commit).
  4. The tarball should not require an Internet connection to download things.
    • Corollary: every external dependency either has to be explicitly documented as such (e.g., gcc and GnuTLS), or included in the tarball.
    • Observation: This means including all *.po gettext translations which are normally downloaded when building from version controlled sources.
  5. The tarball should contain everything required to build the project from source using as much externally released versioned tooling as possible. This is the “minimal” property lacking today.
    • Corollary: This means including a vendored copy of OpenSSL or libz is not acceptable: link to them as external projects.
    • Open question: How about non-released external tooling such as gnulib or autoconf archive macros? This is a bit more delicate: most distributions either just package one current version of gnulib or autoconf archive, not previous versions. While this could change, and distributions could package the gnulib git repository (up to some current version) and the autoconf archive git repository — and packages were set up to extract the version they need (gnulib’s ./bootstrap already supports this via the –gnulib-refdir parameter), this is not normally in place.
    • Suggested Corollary: The tarball should contain content from git submodule’s such as gnulib and the necessary Autoconf archive M4 macros required by the project.
  6. Similar to how the GNU project specify the ./configure interface we need a documented interface for how to bootstrap the project. I suggest to use the already well established idiom of running ./bootstrap to set up the package to later be able to be built via ./configure. Of course, some projects are not using the autotool ./configure interface and will not follow this aspect either, but like most build systems that compete with autotools have instructions on how to build the project, they should document similar interfaces for bootstrapping the source tarball to allow building.

If tarballs that achieve the above goals were available from popular upstream projects, distributions could more easily use them instead of current tarballs that include pre-generated content. The advantage would be that the build process is not tainted by “unnecessary” files. We need to develop tools for maintainers to create these tarballs, similar to make dist that generate today’s foo-1.2.3.tar.gz files.

I think one common argument against this approach will be: Why bother with all that, and just use git-archive outputs? Or avoid the entire tarball approach and move directly towards version controlled check outs and referring to upstream releases as git URL and commit tag or id. One problem with this is that SHA-1 is broken, so placing trust in a SHA-1 identifier is simply not secure. Another counter-argument is that this optimize for packagers’ benefits at the cost of upstream maintainers: most upstream maintainers do not want to store gettext *.po translations in their source code repository. A compromise between the needs of maintainers and packagers is useful, so this *-src.tar.gz tarball approach is the indirection we need to solve that. Update: In my experiment with source-only tarballs for Libntlm I actually did use git-archive output.

What do you think?

Apt archive mirrors in Git-LFS

My effort to improve transparency and confidence of public apt archives continues. I started to work on this in “Apt Archive Transparency” in which I mention the debdistget project in passing. Debdistget is responsible for mirroring index files for some public apt archives. I’ve realized that having a publicly auditable and preserved mirror of the apt repositories is central to being able to do apt transparency work, so the debdistget project has become more central to my project than I thought. Currently I track Trisquel, PureOS, Gnuinos and their upstreams Ubuntu, Debian and Devuan.

Debdistget download Release/Package/Sources files and store them in a git repository published on GitLab. Due to size constraints, it uses two repositories: one for the Release/InRelease files (which are small) and one that also include the Package/Sources files (which are large). See for example the repository for Trisquel release files and the Trisquel package/sources files. Repositories for all distributions can be found in debdistutils’ archives GitLab sub-group.

The reason for splitting into two repositories was that the git repository for the combined files become large, and that some of my use-cases only needed the release files. Currently the repositories with packages (which contain a couple of months worth of data now) are 9GB for Ubuntu, 2.5GB for Trisquel/Debian/PureOS, 970MB for Devuan and 450MB for Gnuinos. The repository size is correlated to the size of the archive (for the initial import) plus the frequency and size of updates. Ubuntu’s use of Apt Phased Updates (which triggers a higher churn of Packages file modifications) appears to be the primary reason for its larger size.

Working with large Git repositories is inefficient and the GitLab CI/CD jobs generate quite some network traffic downloading the git repository over and over again. The most heavy user is the debdistdiff project that download all distribution package repositories to do diff operations on the package lists between distributions. The daily job takes around 80 minutes to run, with the majority of time is spent on downloading the archives. Yes I know I could look into runner-side caching but I dislike complexity caused by caching.

Fortunately not all use-cases requires the package files. The debdistcanary project only needs the Release/InRelease files, in order to commit signatures to the Sigstore and Sigsum transparency logs. These jobs still run fairly quickly, but watching the repository size growth worries me. Currently these repositories are at Debian 440MB, PureOS 130MB, Ubuntu/Devuan 90MB, Trisquel 12MB, Gnuinos 2MB. Here I believe the main size correlation is update frequency, and Debian is large because I track the volatile unstable.

So I hit a scalability end with my first approach. A couple of months ago I “solved” this by discarding and resetting these archival repositories. The GitLab CI/CD jobs were fast again and all was well. However this meant discarding precious historic information. A couple of days ago I was reaching the limits of practicality again, and started to explore ways to fix this. I like having data stored in git (it allows easy integration with software integrity tools such as GnuPG and Sigstore, and the git log provides a kind of temporal ordering of data), so it felt like giving up on nice properties to use a traditional database with on-disk approach. So I started to learn about Git-LFS and understanding that it was able to handle multi-GB worth of data that looked promising.

Fairly quickly I scripted up a GitLab CI/CD job that incrementally update the Release/Package/Sources files in a git repository that uses Git-LFS to store all the files. The repository size is now at Ubuntu 650kb, Debian 300kb, Trisquel 50kb, Devuan 250kb, PureOS 172kb and Gnuinos 17kb. As can be expected, jobs are quick to clone the git archives: debdistdiff pipelines went from a run-time of 80 minutes down to 10 minutes which more reasonable correlate with the archive size and CPU run-time.

The LFS storage size for those repositories are at Ubuntu 15GB, Debian 8GB, Trisquel 1.7GB, Devuan 1.1GB, PureOS/Gnuinos 420MB. This is for a couple of days worth of data. It seems native Git is better at compressing/deduplicating data than Git-LFS is: the combined size for Ubuntu is already 15GB for a couple of days data compared to 8GB for a couple of months worth of data with pure Git. This may be a sub-optimal implementation of Git-LFS in GitLab but it does worry me that this new approach will be difficult to scale too. At some level the difference is understandable, Git-LFS probably store two different Packages files — around 90MB each for Trisquel — as two 90MB files, but native Git would store it as one compressed version of the 90MB file and one relatively small patch to turn the old files into the next file. So the Git-LFS approach surprisingly scale less well for overall storage-size. Still, the original repository is much smaller, and you usually don’t have to pull all LFS files anyway. So it is net win.

Throughout this work, I kept thinking about how my approach relates to Debian’s snapshot service. Ultimately what I would want is a combination of these two services. To have a good foundation to do transparency work I would want to have a collection of all Release/Packages/Sources files ever published, and ultimately also the source code and binaries. While it makes sense to start on the latest stable releases of distributions, this effort should scale backwards in time as well. For reproducing binaries from source code, I need to be able to securely find earlier versions of binary packages used for rebuilds. So I need to import all the Release/Packages/Sources packages from snapshot into my repositories. The latency to retrieve files from that server is slow so I haven’t been able to find an efficient/parallelized way to download the files. If I’m able to finish this, I would have confidence that my new Git-LFS based approach to store these files will scale over many years to come. This remains to be seen. Perhaps the repository has to be split up per release or per architecture or similar.

Another factor is storage costs. While the git repository size for a Git-LFS based repository with files from several years may be possible to sustain, the Git-LFS storage size surely won’t be. It seems GitLab charges the same for files in repositories and in Git-LFS, and it is around $500 per 100GB per year. It may be possible to setup a separate Git-LFS backend not hosted at GitLab to serve the LFS files. Does anyone know of a suitable server implementation for this? I had a quick look at the Git-LFS implementation list and it seems the closest reasonable approach would be to setup the Gitea-clone Forgejo as a self-hosted server. Perhaps a cloud storage approach a’la S3 is the way to go? The cost to host this on GitLab will be manageable for up to ~1TB ($5000/year) but scaling it to storing say 500TB of data would mean an yearly fee of $2.5M which seems like poor value for the money.

I realized that ultimately I would want a git repository locally with the entire content of all apt archives, including their binary and source packages, ever published. The storage requirements for a service like snapshot (~300TB of data?) is today not prohibitly expensive: 20TB disks are $500 a piece, so a storage enclosure with 36 disks would be around $18.000 for 720TB and using RAID1 means 360TB which is a good start. While I have heard about ~TB-sized Git-LFS repositories, would Git-LFS scale to 1PB? Perhaps the size of a git repository with multi-millions number of Git-LFS pointer files will become unmanageable? To get started on this approach, I decided to import a mirror of Debian’s bookworm for amd64 into a Git-LFS repository. That is around 175GB so reasonable cheap to host even on GitLab ($1000/year for 200GB). Having this repository publicly available will make it possible to write software that uses this approach (e.g., porting debdistreproduce), to find out if this is useful and if it could scale. Distributing the apt repository via Git-LFS would also enable other interesting ideas to protecting the data. Consider configuring apt to use a local file:// URL to this git repository, and verifying the git checkout using some method similar to Guix’s approach to trusting git content or Sigstore’s gitsign.

A naive push of the 175GB archive in a single git commit ran into pack size limitations:

remote: fatal: pack exceeds maximum allowed size (4.88 GiB)

however breaking up the commit into smaller commits for parts of the archive made it possible to push the entire archive. Here are the commands to create this repository:

git init
git lfs install
git lfs track 'dists/**' 'pool/**'
git add .gitattributes
git commit -m"Add Git-LFS track attributes." .gitattributes
time debmirror --method=rsync --host ftp.se.debian.org --root :debian --arch=amd64 --source --dist=bookworm,bookworm-updates --section=main --verbose --diff=none --keyring /usr/share/keyrings/debian-archive-keyring.gpg --ignore .git .
git add dists project
git commit -m"Add." -a
git remote add origin git@gitlab.com:debdistutils/archives/debian/mirror.git
git push --set-upstream origin --all
for d in pool//; do
echo $d;
time git add $d;
git commit -m"Add $d." -a
git push
done

The resulting repository size is around 27MB with Git LFS object storage around 174GB. I think this approach would scale to handle all architectures for one release, but working with a single git repository for all releases for all architectures may lead to a too large git repository (>1GB). So maybe one repository per release? These repositories could also be split up on a subset of pool/ files, or there could be one repository per release per architecture or sources.

Finally, I have concerns about using SHA1 for identifying objects. It seems both Git and Debian’s snapshot service is currently using SHA1. For Git there is SHA-256 transition and it seems GitLab is working on support for SHA256-based repositories. For serious long-term deployment of these concepts, it would be nice to go for SHA256 identifiers directly. Git-LFS already uses SHA256 but Git internally uses SHA1 as does the Debian snapshot service.

What do you think? Happy Hacking!

Classic McEliece goes to IETF and OpenSSH

My earlier work on Streamlined NTRU Prime has been progressing along. The IETF document on sntrup761 in SSH has passed several process points. GnuPG’s libgcrypt has added support for sntrup761. The libssh support for sntrup761 is working, but the merge request is stuck mostly due to lack of time to debug why the regression test suite sporadically errors out in non-sntrup761 related parts with the patch.

The foundation for lattice-based post-quantum algorithms has some uncertainty around it, and I have felt that there is more to the post-quantum story than adding sntrup761 to implementations. Classic McEliece has been mentioned to me a couple of times, and I took some time to learn it and did a cut’n’paste job of the proposed ISO standard and published draft-josefsson-mceliece in the IETF to make the algorithm easily available to the IETF community. A high-quality implementation of Classic McEliece has been published as libmceliece and I’ve been supporting the work of Jan Mojžíš to package libmceliece for Debian, alas it has been stuck in the ftp-master NEW queue for manual review for over two months. The pre-dependencies librandombytes and libcpucycles are available in Debian already.

All that text writing and packaging work set the scene to write some code. When I added support for sntrup761 in libssh, I became familiar with the OpenSSH code base, so it was natural to return to OpenSSH to experiment with a new SSH KEX for Classic McEliece. DJB suggested to pick mceliece6688128 and combine it with the existing X25519+sntrup761 or with plain X25519. While a three-algorithm hybrid between X25519, sntrup761 and mceliece6688128 would be a simple drop-in for those that don’t want to lose the benefits offered by sntrup761, I decided to start the journey on a pure combination of X25519 with mceliece6688128. The key combiner in sntrup761x25519 is a simple SHA512 call and the only good I can say about that is that it is simple to describe and implement, and doesn’t raise too many questions since it is already deployed.

After procrastinating coding for months, once I sat down to work it only took a couple of hours until I had a successful Classic McEliece SSH connection. I suppose my brain had sorted everything in background before I started. To reproduce it, please try the following in a Debian testing environment (I use podman to get a clean environment).

# podman run -it --rm debian:testing-slim
apt update
apt dist-upgrade -y
apt install -y wget python3 librandombytes-dev libcpucycles-dev gcc make git autoconf libz-dev libssl-dev

cd ~
wget -q -O- https://lib.mceliece.org/libmceliece-20230612.tar.gz | tar xfz -
cd libmceliece-20230612/
./configure
make install
ldconfig
cd ..

git clone https://gitlab.com/jas/openssh-portable
cd openssh-portable
git checkout jas/mceliece
autoreconf
./configure # verify 'libmceliece support: yes'
make # CC="cc -DDEBUG_KEX=1 -DDEBUG_KEXDH=1 -DDEBUG_KEXECDH=1"

You should now have a working SSH client and server that supports Classic McEliece! Verify support by running ./ssh -Q kex and it should mention mceliece6688128x25519-sha512@openssh.com.

To have it print plenty of debug outputs, you may remove the # character on the final line, but don’t use such a build in production.

You can test it as follows:

./ssh-keygen -A # writes to /usr/local/etc/ssh_host_...

# setup public-key based login by running the following:
./ssh-keygen -t rsa -f ~/.ssh/id_rsa -P ""
cat ~/.ssh/id_rsa.pub > ~/.ssh/authorized_keys

adduser --system sshd
mkdir /var/empty

while true; do $PWD/sshd -p 2222 -f /dev/null; done &

./ssh -v -p 2222 localhost -oKexAlgorithms=mceliece6688128x25519-sha512@openssh.com date

On the client you should see output like this:

OpenSSH_9.5p1, OpenSSL 3.0.11 19 Sep 2023
...
debug1: SSH2_MSG_KEXINIT sent
debug1: SSH2_MSG_KEXINIT received
debug1: kex: algorithm: mceliece6688128x25519-sha512@openssh.com
debug1: kex: host key algorithm: ssh-ed25519
debug1: kex: server->client cipher: chacha20-poly1305@openssh.com MAC: <implicit> compression: none
debug1: kex: client->server cipher: chacha20-poly1305@openssh.com MAC: <implicit> compression: none
debug1: expecting SSH2_MSG_KEX_ECDH_REPLY
debug1: SSH2_MSG_KEX_ECDH_REPLY received
debug1: Server host key: ssh-ed25519 SHA256:YognhWY7+399J+/V8eAQWmM3UFDLT0dkmoj3pIJ0zXs
...

debug1: Host '[localhost]:2222' is known and matches the ED25519 host key.
debug1: Found key in /root/.ssh/known_hosts:1
debug1: rekey out after 134217728 blocks
debug1: SSH2_MSG_NEWKEYS sent
debug1: expecting SSH2_MSG_NEWKEYS
debug1: SSH2_MSG_NEWKEYS received
debug1: rekey in after 134217728 blocks
...
debug1: Sending command: date
debug1: pledge: fork
debug1: permanently_set_uid: 0/0
Environment:
  USER=root
  LOGNAME=root
  HOME=/root
  PATH=/usr/bin:/bin:/usr/sbin:/sbin:/usr/local/bin
  MAIL=/var/mail/root
  SHELL=/bin/bash
  SSH_CLIENT=::1 46894 2222
  SSH_CONNECTION=::1 46894 ::1 2222
debug1: client_input_channel_req: channel 0 rtype exit-status reply 0
debug1: client_input_channel_req: channel 0 rtype eow@openssh.com reply 0
Sat Dec  9 22:22:40 UTC 2023
debug1: channel 0: free: client-session, nchannels 1
Transferred: sent 1048044, received 3500 bytes, in 0.0 seconds
Bytes per second: sent 23388935.4, received 78108.6
debug1: Exit status 0

Notice the kex: algorithm: mceliece6688128x25519-sha512@openssh.com output.

How about network bandwidth usage? Below is a comparison of a complete SSH client connection such as the one above that log in and print date and logs out. Plain X25519 is around 7kb, X25519 with sntrup761 is around 9kb, and mceliece6688128 with X25519 is around 1MB. Yes, Classic McEliece has large keys, but for many environments, 1MB of data for the session establishment will barely be noticeable.

./ssh -v -p 2222 localhost -oKexAlgorithms=curve25519-sha256 date 2>&1 | grep ^Transferred
Transferred: sent 3028, received 3612 bytes, in 0.0 seconds

./ssh -v -p 2222 localhost -oKexAlgorithms=sntrup761x25519-sha512@openssh.com date 2>&1 | grep ^Transferred
Transferred: sent 4212, received 4596 bytes, in 0.0 seconds

./ssh -v -p 2222 localhost -oKexAlgorithms=mceliece6688128x25519-sha512@openssh.com date 2>&1 | grep ^Transferred
Transferred: sent 1048044, received 3764 bytes, in 0.0 seconds

So how about session establishment time?

date; i=0; while test $i -le 100; do ./ssh -v -p 2222 localhost -oKexAlgorithms=curve25519-sha256 date > /dev/null 2>&1; i=`expr $i + 1`; done; date
Sat Dec  9 22:39:19 UTC 2023
Sat Dec  9 22:39:25 UTC 2023
# 6 seconds

date; i=0; while test $i -le 100; do ./ssh -v -p 2222 localhost -oKexAlgorithms=sntrup761x25519-sha512@openssh.com date > /dev/null 2>&1; i=`expr $i + 1`; done; date
Sat Dec  9 22:39:29 UTC 2023
Sat Dec  9 22:39:38 UTC 2023
# 9 seconds

date; i=0; while test $i -le 100; do ./ssh -v -p 2222 localhost -oKexAlgorithms=mceliece6688128x25519-sha512@openssh.com date > /dev/null 2>&1; i=`expr $i + 1`; done; date
Sat Dec  9 22:39:55 UTC 2023
Sat Dec  9 22:40:07 UTC 2023
# 12 seconds

I never noticed adding sntrup761, so I’m pretty sure I wouldn’t notice this increase either. This is all running on my laptop that runs Trisquel so take it with a grain of salt but at least the magnitude is clear.

Future work items include:

  • Use a better hybrid mode combiner than SHA512? See for example KEM Combiners.
  • Write IETF document describing the Classic McEliece SSH protocol
  • Submit my patch to the OpenSSH community for discussion and feedback, please review meanwhile!
  • Implement a mceliece6688128sntrup761x25519 multi-hybrid mode?
  • Write a shell script a’la sntrup761.sh to import a stripped-down mceliece6688128 implementation to avoid having OpenSSH depend on libmceliece?
  • My kexmceliece6688128x25519.c is really just kexsntrup761x25519.c with the three calls to sntrup761 replaced with mceliece6688128. This should be parametrized to avoid cut’n’paste of code, if the OpenSSH community is interested.
  • Consider if the behaviour of librandombytes related to closing of file descriptors is relevant to OpenSSH.

Happy post-quantum SSH’ing!

Update: Changing the mceliece6688128_keypair call to mceliece6688128f_keypair (i.e., using the fully compatible f-variant) results in McEliece being just as fast as sntrup761 on my machine.

Update 2023-12-26: An initial IETF document draft-josefsson-ssh-mceliece-00 published.

Streamlined NTRU Prime sntrup761 goes to IETF

The OpenSSH project added support for a hybrid Streamlined NTRU Prime post-quantum key encapsulation method sntrup761 to strengthen their X25519-based default in their version 8.5 released on 2021-03-03. While there has been a lot of talk about post-quantum crypto generally, my impression has been that there has been a slowdown in implementing and deploying them in the past two years. Why is that? Regardless of the answer, we can try to collaboratively change things, and one effort that appears strangely missing are IETF documents for these algorithms.

Building on some earlier work that added X25519/X448 to SSH, writing a similar document was relatively straight-forward once I had spent a day reading OpenSSH and TinySSH source code to understand how it worked. While I am not perfectly happy with how the final key is derived from the sntrup761/X25519 secrets – it is a SHA512 call on the concatenated secrets – I think the construct deserves to be better documented, to pave the road for increased confidence or better designs. Also, reusing the RFC5656§4 structs makes for a worse specification (one unnecessary normative reference), but probably a simpler implementation. I have published draft-josefsson-ntruprime-ssh-00 here. Credit here goes to Jan Mojžíš of TinySSH that designed the earlier sntrup4591761x25519-sha512@tinyssh.org in 2018, Markus Friedl who added it to OpenSSH in 2019, and Damien Miller that changed it to sntrup761 in 2020. Does anyone have more to add to the history of this work?

Once I had sharpened my xml2rfc skills, preparing a document describing the hybrid construct between the sntrup761 key-encapsulation mechanism and the X25519 key agreement method in a non-SSH fashion was easy. I do not know if this work is useful, but it may serve as a reference for further study. I published draft-josefsson-ntruprime-hybrid-00 here.

Finally, how about a IETF document on the base Streamlined NTRU Prime? Explaining all the details, and especially the math behind it would be a significant effort. I started doing that, but realized it is a subjective call when to stop explaining things. If we can’t assume that the reader knows about lattice math, is a document like this the best place to teach it? I settled for the most minimal approach instead, merely giving an introduction to the algorithm, included SageMath and C reference implementations together with test vectors. The IETF audience rarely understands math, so I think it is better to focus on the bits on the wire and the algorithm interfaces. Everything here was created by the Streamlined NTRU Prime team, I merely modified it a bit hoping I didn’t break too much. I have now published draft-josefsson-ntruprime-streamlined-00 here.

I maintain the IETF documents on my ietf-ntruprime GitLab page, feel free to open merge requests or raise issues to help improve them.

To have confidence in the code was working properly, I ended up preparing a branch with sntrup761 for the GNU-project Nettle and have submitted it upstream for review. I had the misfortune of having to understand and implement NIST’s DRBG-CTR to compute the sntrup761 known-answer tests, and what a mess it is. Why does a deterministic random generator support re-seeding? Why does it support non-full entropy derivation? What’s with the key size vs block size confusion? What’s with the optional parameters? What’s with having multiple algorithm descriptions? Luckily I was able to extract a minimal but working implementation that is easy to read. I can’t locate DRBG-CTR test vectors, anyone? Does anyone have sntrup761 test vectors that doesn’t use DRBG-CTR? One final reflection on publishing known-answer tests for an algorithm that uses random data: are the test vectors stable over different ways to implement the algorithm? Just consider of some optimization moved one randomness-extraction call before another, then wouldn’t the output be different? Are there other ways to verify correctness of implementations?

As always, happy hacking!