# Linux 7.4 fixes direct EFI boot and systemd benefits

> The kernel's EFI stub now fills in LoaderDevicePartUUID when booting without GRUB, and systemd can mount /home, /var and swap without touching fstab.
> By José Maldonado · Technology · 2026-10-28

Booting Linux directly from firmware, with no GRUB or systemd-boot in between, is a more common decision than the desktop suggests: coreboot boards, embedded devices and virtual machines that skip the bootloader. Until this week, that shortcut carried a silent cost, because the partitions systemd used to mount on its own stopped appearing.

The Linux 7.4 EFI changes, merged this Monday, October 26, into the merge window, close that gap with a firmware variable that the kernel stub now starts writing on its own. It is called LoaderDevicePartUUID and stores the identifier of the partition from which the system was launched. Vincent Mailhol authored the series, which lands in 7.4 as part of Boot Loader Interface (BLI) support, the standard with which [systemd defines what a bootloader must report](https://systemd.io/BOOT_LOADER_INTERFACE/).

The practical consequence is small in lines and large in daily friction: those who boot this way regain the automounting of /home, /var, /srv and swap described by the Discoverable Partitions Specification, an idea systemd has carried since version 211, from March 2014. Twelve years later, the missing link was not in user space but in the kernel itself. systemd does not change a single line, because it had been reading for a decade a piece of data that almost no one was writing.

## The gap left by direct boot

When EFI firmware loads the kernel image, the stub takes over the bootloader's job: it reserves memory, orders the parameters and jumps to the compiled code. What it did not do was record the partition from which it was launched.

GRUB and systemd-boot have published it for years in the BLI variable that the rest of the system queries. That information is what allows systemd-gpt-auto-generator to identify the boot disk and infer what is on each partition, without touching /etc/fstab or adding root= to the command line.

On a machine with a bootloader the flaw never shows. On a board with coreboot and edk2 as payload, or on an image that boots the kernel bare, the symptom appears within seconds: /home exists, is formatted and stays empty because no one told the system where to look for it.

Direct boot is not a lab oddity either. Disk images mounted with systemd-nspawn, or inspected from a container, are read with the same partition specification, so the data the stub now writes also helps move an installation between a physical machine and a container without reworking the layout.

## The variable systemd already knew how to read

The Linux 7.4 EFI work invents nothing: it fills a gap the specification had described for years. LoaderDevicePartUUID stores the GPT partition UUID of the medium where the bootloader lives, and it is published under the vendor GUID that systemd reserves for the Linux loader.

The [Discoverable Partitions Specification](https://uapi-group.org/specifications/specs/discoverable_partitions_specification/) maps those type UUIDs to specific mount points. These are the entries that most often appear on a disk prepared to boot without /etc/fstab and without root parameters:

Partition type

Type GUID

What systemd does when it finds it

EFI System (ESP)

c12a7328-f81f-11d2-ba4b-00a0c93ec93b

Mounts /boot or /efi on the disk from which it booted

Extended Boot

bc13c2ff-59e6-4262-a352-b275fd6f7172

Mounts /boot when it coexists with the EFI partition

amd64 root

4f68bce3-e8cd-4db1-96e7-fbcaf984b709

Mounts / and sets up dm-verity if the disk asks for it

User data

933ac7e1-2eb4-4f13-b844-0e14e2aef915

Mounts /home from the first partition with that type

Generic Linux data

0fc63daf-8483-4772-8e79-3d69d8477de4

Mounts nothing, requires a manual entry in fstab

The last row explains why the fix is useless on a disk partitioned the old way: the generic data label leads to no mount point and systemd will still require a line written by hand in fstab.

## Twelve years of specification, half the road done

The idea goes back a long way. systemd 211, released in March 2014, already knew how to discover /home and swap by partition type, and the specification later expanded the catalogue to root, /usr, /srv and /var. What never existed was firmware that filled in the data when no bootloader was involved.

It is worth noting that automounting has conditions. /var is only mounted if the partition UUID matches the first 128 bits of an HMAC-SHA256 computed over the system's machine-id, a precaution so that two installations on the same disk do not share data by accident.

## The fix ended up in the kernel because firmware said no

Mailhol came to the problem on his own. He installed coreboot with edk2 on his machine, booted the kernel from the stub and saw that several partitions stopped mounting. The trail led him to a variable that GRUB wrote and his firmware did not, and his first attempt was a change in edk2 itself, rejected as too Linux-specific.

The solution that did get through moves the work to the place where it works for any firmware. The stub first checks whether LoaderInfo is already set by an earlier stage; if it exists, it backs off without touching anything, so GRUB and systemd-boot keep priority. If it does not exist, it walks the image's device path, looks for the GPT HD() node and publishes its GUID.

The hook lives in both stub paths, the generic one and the one that only compiles on x86, so the fix works the same for a 64-bit server as for a laptop with the same firmware. The cost in size is ridiculous: the [bloat-o-meter that accompanies the patch](https://lore.kernel.org/all/20260906-efi_stub_bli-v3-1-e7dc0d6b8fcd@kernel.org/) measures 636 bytes for the function that does the work and 2.6% more libstub.

The variable is volatile, with boot-service and runtime access, so it does not survive a reboot or leave stale state in the board's NVRAM. That is the difference from a persistent variable: what is published describes the current boot, not a configuration meant to last for months.

## Who sees a change in booting and who does not

Anyone running Ubuntu with GRUB sees no change, and it is worth saying so before someone goes looking for the improvement on their machine. Nor will anyone using systemd-boot notice anything, since it already writes the variable on every boot. The change is aimed at direct boot, the scenario Mailhol found on a coreboot board and that appears whenever firmware loads the kernel with no intermediary.

There is one condition the patch does not save: the partitions must carry the specification's GUIDs. If the installer laid out the disk with the generic Linux data label, systemd will not mount anything on its own and the user will still need fstab, with a new kernel or the old one.

Checking takes one command: if LoaderDevicePartUUID is not listed in efivarfs, the filesystem that exposes firmware variables, that machine is mounting its disk blind and systemd resolves it with a line written by hand in fstab.

The rest of the merge window moves in the same direction of taking work out of booting, as in the case of [Linux 7.4 memblock](https://observatorioblockchain.com/en/technology/linux-7-4-memblock-arranque-50-ms/), where the physical memory allocator dropped from 268 seconds to 50 ms on large machines. These are improvements you do not see on the desktop but do see on a server that reboots often.

## One less /etc/fstab in the way

Linux 7.4 will not arrive as stable for several weeks, so no one will notice the change in the next update cycle. Those who will notice are those who build images with systemd-repart, those who test kernels on coreboot boards and those who deploy devices where booting is a firmware detail nobody wants to touch.

The same window also brought the improvements for laptops and handheld consoles covered in [Linux 7.4 and the handhelds](https://observatorioblockchain.com/en/technology/linux-74-soporte-portatiles-handhelds-juego/), and the pattern repeats: invisible work that is only missed when it is absent. None of these pieces sells a kernel, and all of them together are what make a machine boot without anyone opening a text editor. The underlying asymmetry remains: the standard that describes those variables is maintained by systemd, not by the forum that writes the UEFI specification, and that is why the fix ended up where it did.
