CoreOS systems support booting into live PXE and ISO environments. These work using a “rootfs initrd” which contains a squashfs of the actual rootfs to boot into. This rootfs, like any OSTree-based system, contains the system OSTree repo from which objects are hardlinked out to populate the root filesystem tree.
In practice, the primary use case for these environments is to simply install CoreOS to disk and reboot into the installed system. In the past, installing meant fetching the raw metal image from a remote location and writing it to disk. However, this is inefficient because the majority of the data on that image come from the same OSTree objects which are already present in the squashfs.
The osmet functionality is what now allows coreos-installer to re-use these objects to install to disk, while still matching bit-for-bit the metal image (“osmet” is a portmanteau of “OSTree” and “metal”).
At compose time (i.e. when we’re creating metal images), osmet mounts partitions from the raw metal image and uses the FIEMAP ioctl to build a table of “OSTree object checksum -> disk offsets”.
It then “packs” the raw image by going through it but skipping all the chunks which correspond to mapped OSTree objects. The resulting packed image then essentially only contains data like partition tables, the BIOS boot partition, inode metadata, etc…
This packed image is passed through an xz filter and then bundled together with the serialized OSTree object table into an “osmet” file. coreos-assembler runs the packing twice: once for (regular) 512b sector raw metal images, and once more for 4k sector images. Thus, we end up with two osmet files.
Those files are then included as part of the rootfs initrd in the live ISO and PXE environments alongside (not inside) the squashfs.
At install time (i.e. when users boot the live environment), coreos-installer detects the osmet files present and uses the appropriate one for the sector size of the target disk to recreate the metal image to write to disk. The unpacking process is the inverse of packing: it decompresses through xz, then with the deserialized lookup table, it uses the OSTree objects from the mounted squashfs to fill in the gaps which the packed object skipped over. Simultaneously, it verifies the checksum of the written image to ensure that it exactly matches the original.
All the osmet-related code is in
src/osmet/. For more information, you can also see the original PR here: