# Copyright (C) 2015 Google, Inc
# SPDX-License-Identifier: GPL-2.0+
U-Boot on EFI
This document provides information about U-Boot running on top of EFI, either
as an application or just as a means of getting U-Boot onto a new platform.
In God's Name, Why?
This is useful in several situations:
- You have EFI running on a board but U-Boot does not natively support it
fully yet. You can boot into U-Boot from EFI and use that until U-Boot is
- You need to use an EFI implementation (e.g. UEFI) because your vendor
requires it in order to provide support
- You plan to use coreboot to boot into U-Boot but coreboot support does
not currently exist for your platform. In the meantime you can use U-Boot
on EFI and then move to U-Boot on coreboot when ready
- You use EFI but want to experiment with a simpler alternative like U-Boot
Only x86 is supported at present. If you are using EFI on another architecture
you may want to reconsider. However, much of the code is generic so could be
U-Boot supports running as an EFI application for 32-bit EFI only. This is
not very useful since only a serial port is provided. You can look around at
memory and type 'help' but that is about it.
More usefully, U-Boot supports building itself as a payload for either 32-bit
or 64-bit EFI. U-Boot is packaged up and loaded in its entirety by EFI. Once
started, U-Boot changes to 32-bit mode (currently) and takes over the
machine. You can use devices, boot a kernel, etc.
First choose a board that has EFI support and obtain an EFI implementation
for that board. It will be either 32-bit or 64-bit. Alternatively, you can
opt for using QEMU  and the OVMF , as detailed below.
To build U-Boot as an EFI application (32-bit EFI required), enable CONFIG_EFI
and CONFIG_EFI_APP. The efi-x86 config (efi-x86_defconfig) is set up for this.
Just build U-Boot as normal, e.g.
To build U-Boot as an EFI payload (32-bit or 64-bit EFI can be used), adjust an
existing config (like qemu-x86_defconfig) to enable CONFIG_EFI, CONFIG_EFI_STUB
and either CONFIG_EFI_STUB_32BIT or CONFIG_EFI_STUB_64BIT. All of these are
boolean Kconfig options. Then build U-Boot as normal, e.g.
You will end up with one of these files depending on what you build for:
u-boot-app.efi - U-Boot EFI application
u-boot-payload.efi - U-Boot EFI payload application
Trying it out
QEMU is an emulator and it can emulate an x86 machine. Please make sure your
QEMU version is 2.3.0 or above to test this. You can run the payload with
something like this:
cp /path/to/u-boot*.efi /tmp/efi
qemu-system-x86_64 -bios bios.bin -hda fat:/tmp/efi/
Add -nographic if you want to use the terminal for output. Once it starts
type 'fs0:u-boot-payload.efi' to run the payload or 'fs0:u-boot-app.efi' to
run the application. 'bios.bin' is the EFI 'BIOS'. Check  to obtain a
prebuilt EFI BIOS for QEMU or you can build one from source as well.
To try it on real hardware, put u-boot-app.efi on a suitable boot medium,
such as a USB stick. Then you can type something like this to start it:
(or fs0:u-boot-app.efi for the application)
This will start the payload, copy U-Boot into RAM and start U-Boot. Note
that EFI does not support booting a 64-bit application from a 32-bit
EFI (or vice versa). Also it will often fail to print an error message if
you get this wrong.
Here follow a few implementation notes for those who want to fiddle with
this and perhaps contribute patches.
The application and payload approaches sound similar but are in fact
implemented completely differently.
For the application the whole of U-Boot is built as a shared library. The
efi_main() function is in lib/efi/efi_app.c. It sets up some basic EFI
functions with efi_init(), sets up U-Boot global_data, allocates memory for
U-Boot's malloc(), etc. and enters the normal init sequence (board_init_f()
Since U-Boot limits its memory access to the allocated regions very little
special code is needed. The CONFIG_EFI_APP option controls a few things
that need to change so 'git grep CONFIG_EFI_APP' may be instructive.
The CONFIG_EFI option controls more general EFI adjustments.
The only available driver is the serial driver. This calls back into EFI
'boot services' to send and receive characters. Although it is implemented
as a serial driver the console device is not necessarilly serial. If you
boot EFI with video output then the 'serial' device will operate on your
target devices's display instead and the device's USB keyboard will also
work if connected. If you have both serial and video output, then both
consoles will be active. Even though U-Boot does the same thing normally,
These are features of EFI, not U-Boot.
Very little code is involved in implementing the EFI application feature.
U-Boot is highly portable. Most of the difficulty is in modifying the
Makefile settings to pass the right build flags. In particular there is very
little x86-specific code involved - you can find most of it in
arch/x86/cpu. Porting to ARM (which can also use EFI if you are brave
enough) should be straightforward.
Use the 'reset' command to get back to EFI.
The payload approach is a different kettle of fish. It works by building
U-Boot exactly as normal for your target board, then adding the entire
image (including device tree) into a small EFI stub application responsible
for booting it. The stub application is built as a normal EFI application
except that it has a lot of data attached to it.
The stub application is implemented in lib/efi/efi_stub.c. The efi_main()
function is called by EFI. It is responsible for copying U-Boot from its
original location into memory, disabling EFI boot services and starting
U-Boot. U-Boot then starts as normal, relocates, starts all drivers, etc.
The stub application is architecture-dependent. At present it has some
x86-specific code and a comment at the top of efi_stub.c describes this.
While the stub application does allocate some memory from EFI this is not
used by U-Boot (the payload). In fact when U-Boot starts it has all of the
memory available to it and can operate as it pleases (but see the next
The payload can pass information to U-Boot in the form of EFI tables. At
present this feature is used to pass the EFI memory map, an inordinately
large list of memory regions. You can use the 'efi mem all' command to
display this list. U-Boot uses the list to work out where to relocate
Although U-Boot can use any memory it likes, EFI marks some memory as used
by 'run-time services', code that hangs around while U-Boot is running and
is even present when Linux is running. This is common on x86 and provides
a way for Linux to call back into the firmware to control things like CPU
fan speed. U-Boot uses only 'conventional' memory, in EFI terminology. It
will relocate itself to the top of the largest block of memory it can find
U-Boot drivers typically don't use interrupts. Since EFI enables interrupts
it is possible that an interrupt will fire that U-Boot cannot handle. This
seems to cause problems. For this reason the U-Boot payload runs with
interrupts disabled at present.
While the EFI application can in principle be built as either 32- or 64-bit,
only 32-bit is currently supported. This means that the application can only
be used with 32-bit EFI.
The payload stub can be build as either 32- or 64-bits. Only a small amount
of code is built this way (see the extra- line in lib/efi/Makefile).
Everything else is built as a normal U-Boot, so is always 32-bit on x86 at
This work could be extended in a number of ways:
- Add a generic x86 EFI payload configuration. At present you need to modify
an existing one, but mostly the low-level x86 code is disabled when booting
on EFI anyway, so a generic 'EFI' board could be created with a suitable set
of drivers enabled.
- Add ARM support
- Add 64-bit application support
- Figure out how to solve the interrupt problem
- Add more drivers to the application side (e.g. video, block devices, USB,
environment access). This would mostly be an academic exercise as a strong
use case is not readily apparent, but it might be fun.
- Avoid turning off boot services in the stub. Instead allow U-Boot to make
use of boot services in case it wants to. It is unclear what it might want
Where is the code?
payload stub, application, support code. Mostly arch-neutral
helper functions for the fake DRAM init, etc. These can be used by
any board that runs as a payload.
x86 support code for running as an EFI application
x86 board code for running as an EFI application
the 'efi' command
Ben Stoltz, Simon Glass