This documentation’s intent is to share what we currently know about making system images build reproducibly: for example, VM and cloud images, live systems, OS installer ISO images.

General problems of reproducible system images

Usually the problems are:

  • Filesystem needs to be created at once
  • Filesystems have creation and/or modification timestamps
  • Filesystems containing UUID or a label which are not set explicit
  • Included files have timestamps
  • Included files may be generated or updated at build time in a non-reproducible manner
  • The bootloader which is integrated might have timestamps
  • The integrated initramfs images may be built non-reproducibly while building the system image
  • Timestamps which don’t depend on SOURCE_DATE_EPOCH are not reproducible.

E.g. if the filesystem has been created by mkfs.ext4, mounted, modified and saved as an image, it would not be reproducible: the allocation of the inodes is undefined.

Thankfully there are known solutions to most of these problems, read on.

ext2, ext3, ext4

Instead of using mkfs.ext, make_ext4fs can be used. make_ext4fs is creating the whole filesystem at once. make_ext4fs -T <unix_timestamp> allows to set the mtime to $SOURCE_DATE_EPOCH.

make_ext4fs should include the following commit:

  • https://git.lede-project.org/?p=project/make_ext4fs.git;a=commit;h=bd53eaafbc2a89a57b8adda38f53098a431fa8f4

ISO filesystem

When building ISO filesystems with xorriso:

  • use a recent versions of xorriso which honors $SOURCE_DATE_EPOCH for various ISO image metadata
  • pass $SOURCE_DATE_EPOCH to xorriso’s --modification-date to clamp all mtimes.

It might also be necessary to:

SquashFS metadata & compression

When compressing SquashFS images, metadata and compression can make the output unreproducible.

Using a version of mksquashfs that contains the following patches has been sufficient to make some system images reproducible:

Root filesystem content

A system image often contains a root filesystem, generated during the build, and packed in some format such as SquashFS.

Exclude unneeded files

A number of files can simply be emptied or excluded when creating the root filesystem image (some to optimize size, some because they are not needed).

E.g. Tails does this:

Beware: this can have hard to predict consequences. For example, Tails considered dropping even more stuff - such as the fontconfig cache -, but they’ve seen weird results and performance issues when doing so and finally discarded this idea.

Files metadata

The build process for a system image often creates or updates files, which generates file metadata that depends on when the build is performed.

One approach that’s been used successfully to fix this problem is to clamp mtimes of files in the root filesystem to $SOURCE_DATE_EPOCH before generating the system image.

Files generated or updated at build time

Package managers such as dpkg or RPM support postinst scripts and triggers that are run on the target system after unpacking a package, e.g. to generate or update caches and indices… potentially in a non-deterministic manner.

In order to counter this, one possible approach is to replace these scripts and do the same work later (e.g. at first boot).

Another approach is to ensure these scripts generate/update files in a reproducible manner. This approach has the advantage to fix the problem at the cause for every project that uses these programs. For example such problems were fixed in:

Finally, occasionally one may want to use strip-nondeterminism to normalize the content of such files.

gettext

GNU gettext’s POT, PO and MO files were an interesting challenge. One way to approach this problem is to: