14 releases

0.2.24 Oct 11, 2024
0.2.9 Jan 20, 2023
0.2.6 Sep 8, 2021
0.2.5 Jan 6, 2021
0.2.3 Nov 17, 2020

#45 in Operating systems

Download history 128/week @ 2024-09-16 210/week @ 2024-09-23 32/week @ 2024-09-30 181/week @ 2024-10-07 25/week @ 2024-10-14

246 downloads per month

Apache-2.0

140KB
3K SLoC

bootupd: Distribution-independent updates for bootloaders

Today many Linux systems handle updates for bootloader data in an inconsistent and ad-hoc way. For example, on Fedora and Debian, a package manager update will update UEFI binaries in /boot/efi, but not the BIOS MBR data.

Transactional/"image" update systems like OSTree and dual-partition systems like the Container Linux update system are more consistent: they normally cover kernel/userspace but not anything related to bootloaders.

The reason for this is straightforward: performing bootloader updates in an "A/B" fashion requires completely separate nontrivial logic from managing the kernel and root filesystem. Today OSTree e.g. makes the choice that it does not update /boot/efi (and also doesn't update the BIOS MBR).

The goal of this project is to be a cross-distribution, OS update system agnostic tool to manage updates for things like:

  • /boot/efi
  • x86 BIOS MBR
  • Other architecture bootloaders

This project originated in this Fedora CoreOS github issue.

The scope is otherwise limited; for example, bootupd will not manage anything related to the kernel such as kernel arguments; that's for tools like grubby and ostree.

Status

bootupd supports updating GRUB and shim for UEFI firmware on x86_64 and aarch64, and GRUB for BIOS firmware on x86_64. The project is deployed in Fedora CoreOS and derivatives, and is also used by the new bootc install functionality. The bootupd CLI should be considered stable.

bootupd does not yet perform updates in a way that is safe against a power failure at the wrong moment, or against a buggy bootloader update that fails to boot the system.

Therefore, by default, bootupd updates the bootloader only when manually instructed to do so.

Relationship to other projects

dbxtool

dbxtool manages updates to the Secure Boot database - bootupd will likely need to perform any updates to the shimx64.efi binary before dbxtool.service starts. But otherwise they are independent.

fwupd

bootupd could be compared to fwupd which is a project that exists today to update hardware device firmware - things not managed by e.g. apt/zypper/yum/rpm-ostree update today.

fwupd comes as a UEFI binary today, so bootupd could take care of updating fwupd but today fwupd handles that itself. So it's likely that bootupd would only take care of GRUB and shim. See discussion in this issue.

systemd bootctl

systemd bootctl can update itself; this project would probably just proxy that if we detect systemd-boot is in use.

Other goals

One idea is that bootupd could help support redundant bootable disks. For various reasons it doesn't really work to try to use RAID1 for an entire disk; the ESP must be handled specially. bootupd could learn how to synchronize multiple EFI system partitions from a primary.

More details on rationale and integration

A notable problem today for rpm-ostree based systems is that rpm -q shim-x64 is misleading because it's not actually updated in place.

Particularly this commit makes things clear - the data from the RPM goes into /usr (part of the OSTree), so it doesn't touch /boot/efi. But that commit didn't change how the RPM database works (and more generally it would be technically complex for rpm-ostree to change how the RPM database works today).

What we ultimately want is that rpm -q shim-x64 returns "not installed" - because it's not managed by RPM or by ostree. Instead one would purely use bootupctl to manage it. However, it might still be built as an RPM, just not installed that way. The RPM version numbers would be used for the bootupd version associated with the payload, and ultimately we'd teach rpm-ostree compose tree how to separately download bootloaders and pass them to bootupctl backend.

Questions and answers

  • Why is bootupd not part of ostree?

A key advertised feature of ostree is that updates are truly transactional. There's even a a test case that validates forcibly pulling the power during OS updates. A simple way to look at this is that on an ostree-based system there is no need to have a "please don't power off your computer" screen. This in turn helps administrators to confidently enable automatic updates.

Doing that for the bootloader (i.e. bootupd's domain) is an entirely separate problem. There have been some ideas around how we could make the bootloaders use an A/B type scheme (or at least be more resilient), and perhaps in the future bootupd will use some of those.

These updates hence carry different levels of risk. In many cases actually it's OK if the bootloader lags behind; we don't need to update every time.

But out of conservatism currently today for e.g. Fedora CoreOS, bootupd is disabled by default. On the other hand, if your OS update mechanism isn't transactional, then you may want to enable bootupd by default.

  • Is bootupd a daemon?

It was never a daemon. The name was intended to be "bootloader-upDater" not "bootloader-updater-Daemon". The choice of a "d" suffix is in retrospect probably too confusing.

bootupd used to have an internally-facing bootupd.service and bootupd.socket systemd units that acted as a locking mechanism. The service would very quickly auto exit. There was nothing long-running, so it was not really a daemon.

bootupd now uses systemd-run instead to guarantee the following:

  • It provides a robust natural "locking" mechanism.
  • It ensures that critical logging metadata always consistently ends up in the systemd journal, not e.g. a transient client SSH connection.
  • It benefits from the sandboxing options available for systemd units, and while bootupd is obviously privileged we can still make use of some of this.
  • If we want a non-CLI API (whether that's DBus or Cap'n Proto or varlink or something else), we will create an independent daemon with a stable API for this specific need.

Dependencies

~13–25MB
~396K SLoC