add a new family of wrappers for EFI based images

debefivm-create is based on mmdebstrap-autopkgtest-build-qemu, which is
is co-authored with Johannes Schauer Marin Rodrigues. Also thanks to
Jochen Sprickerhof for suggesting the --rootsize option for use in
Debusine.

3 files changed, 1228 insertions, 0 deletions

diff --git a/bin/debefivm-create b/bin/debefivm-create
new file mode 100755
index 0000000..6efc204
--- /dev/null
+++ b/bin/debefivm-create
@@ -0,0 +1,498 @@
+#!/bin/sh
+# Copyright 2023 Johannes Schauer Marin Rodrigues <josch@debian.org>
+# SPDX-FileCopyrightText: 2023-2025 Helmut Grohne <helmut@subdivi.de>
+# SPDX-License-Identifier: MIT
+
+# shellcheck disable=SC2016 # Intentional quoting technique
+
+: <<'POD2MAN'
+=head1 NAME
+
+debefivm-create - Create an EFI-bootable VM image for various Debian releases and architectures
+
+=head1 SYNOPSIS
+
+B<debefivm-create> [B<-h> I<hostname>] [B<-k> F<sshkey>] [B<-o> F<output>] [B<-r> I<release>] [B<-s> <task>] [B<-z> I<size>] [B<--> I<mmdebstrap options>]
+
+=head1 DESCRIPTION
+
+Create an EFI-bootable disk image image booting into Debian.
+This essentially is a thin wrapper around B<mmdebstrap> much like B<debvm-create>.
+The resulting image contains a GPT partition table with an EFI system partition.
+A Unified Kernel Image is constructed from the installed kernel and initrd and placed into said partition.
+As a result, such an image may be booted with a variety of virtual machine emulators such as incus, libvirt or VirtualBox.
+It may also be booted using B<debefivm-run>.
+No user account is created and root can login without specifying a password.
+
+=head1 OPTIONS
+
+=over 8
+
+=item B<-a> I<architecture>, B<--architecture>=I<architecture>
+
+Set the architecture of the virtual machine.
+This option is forwarded to B<mmdebstrap> and multiple architectures may be supplied.
+The first architecture is used to assemble the Unified Kernel Image.
+It defaults to the native architecture of the installation.
+
+=item B<-h> I<hostname>, B<--hostname>=I<hostname>
+
+Set the hostname of the virtual machine.
+By default, the hostname is B<testvm>.
+
+=item B<-k> F<sshkey>, B<--sshkey>=F<sshkey>
+
+Install the given ssh public key file into the virtual machine image for the root user.
+This option also causes the ssh server to be installed.
+By default, no key or server is installed.
+To connect to the vm, pass a port number to B<debefivm-run> with the B<-s> option.
+
+=item B<-o> F<output>, B<--output>=F<output>
+
+Specify the file name of the resulting virtual machine image.
+By default, it is written to F<vm.img>.
+
+=item B<-r> I<release>, B<--release>=I<release>
+
+Use the given Debian release.
+By default, B<unstable> is being used.
+If you pass a complete F<sources.list> that includes release names to B<mmdebstrap>, you may pass an empty string here.
+
+=item B<-s> I<task>, B<--skip>=I<task>
+
+Skip a particular task or feature.
+The option may be specified multiple times or list multiple tasks to be skipped by separating them with a comma.
+By default, no tasks are skipped.
+The following tasks may be skipped.
+
+=over 4
+
+=item B<autologin>
+
+Skips adding a the customize-autologin.sh to B<mmdebstrap> that configures
+automatic root login on a serial console and also parses the C<TERM> kernel
+cmdline and passes it as C<TERM> to B<agetty>.
+
+=item B<kernel>
+
+skips installing a linux kernel image.
+This can be useful to install a kernel without a package.
+If a kernel is installed via B<mmdebstrap> option C<--include>, automtatic kernel installation is automatically skipped.
+
+=item B<packagelists>
+
+reduces the package lists inside the image.
+The B<available> database for B<dpkg> is not created.
+The package lists used by B<apt> are deleted.
+This generally produces a smaller image, but you need to run B<apt update> before installing packages and B<dpkg --set-selections> does not work.
+
+=item B<systemdnetwork>
+
+skips installing B<libnss-resolve> as well as automatic network configuration via B<systemd-networkd>.
+
+=item B<usrmerge>
+
+By default B<debefivm-create> adds a hook to enable merged-/usr without the B<usrmerge> package given a sufficiently recent Debian release.
+Without the hook, dependencies will pull the B<usrmerge> package as needed, which may result in a larger installation.
+
+=back
+
+=item B<--rootsize>=I<size>
+
+This option may be used as an alternative to B<--imagesize> to determine the size of the output image.
+It is also composed of an integer with optional unit.
+Unlike B<--imagesize> it determines the size of the root filesystem partition and the resulting image will be slightly larger than the size given here.
+
+=item B<-z> I<size>, B<--imagesize>=I<size>
+
+Specify the total image size as an integer and optional unit (example: 10K is 10*1024).
+Units are K,M,G,T (powers of 1024).
+The disk image contains an EFI system partition and an ext4 root filesystem.
+The EFI partition has a static size of about 128MB and the remaining space will be allocated to the root filesystem.
+The default is 10 GB.
+
+=item B<--> I<mmdebstrap options>
+
+All options beyond a double dash are passed to B<mmdebstrap> after the suite and target specification.
+This can be used to provide additional hooks for image customization.
+You can also request additional packages to be installed into the image using B<mmdebstrap>'s B<--include> option.
+Any positional arguments passed here will be treated as mirror specifications by B<mmdebstrap>.
+The B<--architecture> option should not be given here as B<debefivm-create> needs to know its value to construct a Unified Kernel Image.
+
+=back
+
+=head1 EXAMPLES
+
+Create an image suitable for use with B<autopkgtest-virt-qemu> B<--efi>.
+
+    debefivm-create ... -- --include=linux-image-generic,libpam-systemd,passwd,python3 --hook-dir=/usr/share/mmdebstrap/hooks/useradd --customize-hook=/usr/share/autopkgtest/setup-commands/setup-testbed
+
+=head1 SEE ALSO
+
+    debefivm-run(1) debefivm-ukify(1) debvm-create(1) mmdebstrap(1)
+
+=cut
+POD2MAN
+
+set -eu
+
+DEBARCH=$(dpkg --print-architecture)
+EFILABEL=efisys
+FAT_SIZE_KIB=$((126*1024))
+IMAGE=vm.img
+SKIP=,
+SSHKEY=
+SUITE=unstable
+ROOTLABEL=rootfs
+TOT_SIZE_KIB=$((10*1024*1024))
+VMNAME=testvm
+
+SHARE_DIR="${0%/*}/../share"
+BIN_DIR="${0%/*}/../bin/"
+
+# Disk layout:
+#  * GPT partition table in the front. Oversize to align.
+GPT_FRONT_SIZE_KIB=1024
+#  * FAT partition containing the EFI boot loader. Size configured.
+FAT_OFFSET_KIB=$GPT_FRONT_SIZE_KIB
+#  * Main ext4 root partition. Compute as remaining size.
+EXT4_OFFSET_KIB=$((FAT_OFFSET_KIB + FAT_SIZE_KIB))
+# EXT4_SIZE_KIB and TOT_SIZE_KIB depend on one another.
+#  * GPT backup table in the back. Oversize to align.
+GPT_BACK_SIZE_KIB=1024
+
+NON_EXT4_SIZE_KIB=$((EXT4_OFFSET_KIB + GPT_BACK_SIZE_KIB))
+EXT4_SIZE_KIB=$((TOT_SIZE_KIB - NON_EXT4_SIZE_KIB))
+
+nth_arg() {
+	shift "$1"
+	printf "%s" "$1"
+}
+
+die() {
+	echo "$*" >&2
+	exit 1
+}
+usage() {
+	die "usage: $0 [-a architecture] [-h hostname] [-k sskey] [-o output] [-r release] [-s task] [-z imagesize] [-- mmdebstrap options]"
+}
+usage_error() {
+	echo "error: $*" 1>&2
+	usage
+}
+
+opt_architecture() {
+	DEBARCH="$1"
+}
+opt_hostname() {
+	VMNAME=$1
+}
+opt_output() {
+	IMAGE=$1
+	if test "${IMAGE#-}" != "$IMAGE"; then
+		IMAGE="./$IMAGE"
+	fi
+}
+opt_release() {
+	SUITE="$1"
+}
+opt_rootsize() {
+	EXT4_SIZE_KIB="$(numfmt --to=none --to-unit=Ki --from=iec "$1")" ||
+		die "failed to parse size '$1'"
+	TOT_SIZE_KIB=$((EXT4_SIZE_KIB + NON_EXT4_SIZE_KIB))
+}
+opt_skip() {
+	SKIP="$SKIP$1,"
+}
+opt_sshkey() {
+	SSHKEY=$1
+}
+opt_imagesize() {
+	TOT_SIZE_KIB="$(numfmt --to=none --to-unit=Ki --from=iec "$1")" ||
+		die "failed to parse size '$1'"
+	EXT4_SIZE_KIB=$((TOT_SIZE_KIB - NON_EXT4_SIZE_KIB))
+}
+
+while getopts :a:h:k:o:r:s:z:-: OPTCHAR; do
+	case "$OPTCHAR" in
+		a)	opt_architecture "$OPTARG"	;;
+		h)	opt_hostname "$OPTARG"		;;
+		k)	opt_sshkey "$OPTARG"		;;
+		o)	opt_output "$OPTARG"		;;
+		r)	opt_release "$OPTARG"		;;
+		s)	opt_skip "$OPTARG"		;;
+		z)	opt_imagesize "$OPTARG"		;;
+		-)
+			case "$OPTARG" in
+				help)
+					usage
+				;;
+				architecture|hostname|imagesize|output|release|rootsize|skip|sshkey)
+					test "$OPTIND" -gt "$#" && usage_error "missing argument for --$OPTARG"
+					"opt_$OPTARG" "$(nth_arg "$OPTIND" "$@")"
+					OPTIND=$((OPTIND+1))
+				;;
+				architecture=*|hostname=*|imagesize=*|output=*|release=*|rootsize=*|size=*|skip=*|sshkey=*)
+					"opt_${OPTARG%%=*}" "${OPTARG#*=}"
+				;;
+				*)
+					usage_error "unrecognized option --$OPTARG"
+				;;
+			esac
+		;;
+		:)
+			usage_error "missing argument for -$OPTARG"
+		;;
+		'?')
+			usage_error "unrecognized option -$OPTARG"
+		;;
+		*)
+			die "internal error while parsing command options, please report a bug"
+		;;
+	esac
+done
+shift "$((OPTIND - 1))"
+
+if test -n "$SSHKEY" && ! test -f "$SSHKEY"; then
+	die "error: ssh keyfile '$SSHKEY' not found"
+fi
+
+check_skip() {
+	case "$SKIP" in
+		*",$1,"*)	return 0 ;;
+		*)		return 1 ;;
+	esac
+}
+
+test "$EXT4_SIZE_KIB" -lt "$((1024*64))" &&
+	die "implausibly small disk size $TOT_SIZE_KIB"
+
+# In a multiarch setting, we cannot tell in advance which architecture will be
+# supplying the Linux kernel image, but that's the architecture we need the EFI
+# stub for. Hence fail early.
+test "${DEBARCH%,*}" = "$DEBARCH" ||
+	die "multiarch images are not yet supported"
+
+case "${DEBARCH%%,*}" in
+	amd64)
+		EFIARCH=x64
+		ROOTGUID=4F68BCE3-E8CD-4DB1-96E7-FBCAF984B709
+	;;
+	arm64)
+		EFIARCH=aa64
+		ROOTGUID=B921B045-1DF0-41C3-AF44-4C6F280D3FAE
+	;;
+	armhf)
+		EFIARCH=arm
+		ROOTGUID=69DAD710-2CE4-4E3C-B16C-21A1D49ABED3
+	;;
+	i386)
+		EFIARCH=ia32
+		ROOTGUID=44479540-F297-41B2-9AF7-D131D5F0458A
+	;;
+	riscv64)
+		EFIARCH=riscv64
+		ROOTGUID=72EC70A6-CF74-40E6-BD49-4BDA08E8F224
+	;;
+	*)
+		die "unsupported architecture: $DEBARCH"
+	;;
+esac
+
+test -e "$IMAGE" &&
+	die "output file $IMAGE already exists"
+
+if command -v ukify >/dev/null; then
+	UKIFY=ukify
+	UKIFY_VERBOSE=
+else
+	UKIFY="${BIN_DIR}debefivm-ukify"
+	UKIFY_VERBOSE=--verbose
+fi
+
+WORKDIR=
+
+cleanup() {
+	test -n "$WORKDIR" && rm -Rf "$WORKDIR"
+}
+cleanup_abort() {
+	cleanup
+	echo aborted >&2
+	exit 2
+}
+trap cleanup EXIT
+trap cleanup_abort HUP INT TERM QUIT
+
+WORKDIR=$(mktemp -d)
+
+truncate --size="${TOT_SIZE_KIB}K" "$IMAGE"
+
+# Stuff most of our options before the user options to allow them to override
+# ours.
+#
+# We skip replacing start-stop-daemon, because policy-rc.d nowadays works well
+# enough and then we have to not restore it before calling mkfs.ext4.
+set -- \
+	--mode=root \
+	--variant=apt \
+	--architecture="$DEBARCH" \
+	--skip=chroot/start-stop-daemon \
+	"$@"
+
+if ! check_skip kernel; then
+	set -- "--customize-hook=$SHARE_DIR/customize-kernel.sh" "$@"
+fi
+
+set -- \
+	"--customize-hook=printf 'LABEL=%s / ext4 defaults 0 0' '$ROOTLABEL' >"'"$1/etc/fstab"' \
+	"--include=?narrow(?exact-name(systemd-boot),?architecture(${DEBARCH%%,*}))" \
+	'--include=systemd-sysv' \
+	"$SUITE" \
+	/dev/null \
+	"$@"
+
+# set up a hostname
+ETC_HOSTS_TEMPLATE='127.0.0.1 localhost\n127.0.1.1 %s\n::1 ip6-localhost ip6-loopback\n'
+set -- \
+	"--customize-hook=echo '$VMNAME' >"'"$1/etc/hostname"' \
+	"--customize-hook=printf '$ETC_HOSTS_TEMPLATE' '$VMNAME' >"'"$1/etc/hosts"' \
+	"$@"
+
+# allow password-less root login
+set -- "--customize-hook=$SHARE_DIR/customize-delete-rootpw.sh" "$@"
+
+if ! check_skip systemdnetwork; then
+	# dhcp on all network interfaces, and add a dns resolver
+	set -- \
+		"--customize-hook=$SHARE_DIR/customize-networkd.sh" \
+		'--include=?not(?virtual)?exact-name(libnss-resolve)' \
+		"--customize-hook=$SHARE_DIR/customize-resolved.sh" \
+		"$@"
+fi
+
+# add ssh key for root
+if test -n "$SSHKEY"; then
+	set -- \
+		--include=openssh-server \
+		'--customize-hook=mkdir -m700 -p "$1/root/.ssh"' \
+		"--customize-hook=upload $SSHKEY /root/.ssh/authorized_keys" \
+		"$@"
+fi
+
+if ! check_skip packagelists; then
+	set -- --skip=cleanup/apt/lists "$@"
+	set -- "--customize-hook=$SHARE_DIR/customize-dpkgavailable.sh" "$@"
+fi
+
+if ! check_skip usrmerge; then
+	# Avoid the usrmerge package
+	set -- --hook-dir=/usr/share/mmdebstrap/hooks/maybe-merged-usr "$@"
+fi
+
+if ! check_skip autologin; then
+	set -- \
+		--include=login \
+		"--customize-hook=$SHARE_DIR/customize-autologin.sh" "$@"
+fi
+
+# By default, Debian mounts the EFI System Partition to /boot/efi. Keep that
+# default.
+BOOT_EFI_MOUNT_TEMPLATE='[Unit]\nDescription=EFI System Partition\n\n[Mount]\nWhat=LABEL=%s\nWhere=/boot/efi\nType=vfat\nOptions=umask=0077,rw,nodev,nosuid,noexec,nosymfollow\n\n[Install]\nWantedBy=local-fs.target\n'
+set -- \
+	'--customize-hook=mkdir "$1/boot/efi"' \
+	"--customize-hook=printf '$BOOT_EFI_MOUNT_TEMPLATE' '$EFILABEL' >"'"$1/etc/systemd/system/boot-efi.mount"' \
+	'--customize-hook=systemctl --root "$1" enable boot-efi.mount' \
+	"$@"
+
+# Pass the remaining hooks last. These collect data from the generated tree and
+# create the resulting filesystem. All user-supplied customizations should have
+# run before these.
+#
+# We download the artifacts that end up being turned into the EFI.
+EXT4_OPTIONS="offset=$((EXT4_OFFSET_KIB * 1024)),assume_storage_prezeroed=1"
+set -- "$@" \
+	"--customize-hook=download vmlinuz '$WORKDIR/kernel'" \
+	"--customize-hook=download initrd.img '$WORKDIR/initrd'" \
+	"--customize-hook=download '/usr/lib/systemd/boot/efi/linux$EFIARCH.efi.stub' '$WORKDIR/stub'" \
+	'--customize-hook=mount --bind "$1" "$1/mnt"' \
+	'--customize-hook=mount -t tmpfs tmpfs "$1/mnt/dev" -o mode=0755' \
+	'--customize-hook=rm -f "$1/usr/sbin/policy-rc.d"' \
+	'--customize-hook=/sbin/mkfs.ext4 -d "$1/mnt" -L '"'$ROOTLABEL' -E '$EXT4_OPTIONS' '$IMAGE' '${EXT4_SIZE_KIB}K'" \
+	'--customize-hook=umount --lazy "$1/mnt"'
+
+set -- mmdebstrap "$@"
+
+# We need to write the resulting $IMAGE from inside the namespace used by
+# mmdebstrap. If we were to have mmdebstrap --mode=unshare set up that
+# namespace, we could end up being unable to write it if a leading directory
+# component in $IMAGE were not not grant execute permission to any of the
+# subuids mapped. This typically is the case when using PrivateTmp=yes. With
+# that setting, there is a 0700 directory owned by the current user that
+# happens to not be mapped by mmdebstrap.
+#
+# Instead, we take matters into our own hands and set up the namespaces outside
+# mmdebstrap asking it to assume root. In addition to what mmdebstrap would do
+# by itself, we also map the current user to 65536.  The precise uid does not
+# matter, but having it mapped allows CAP_DAC_OVERRIDE to be used for writing
+# the image.
+if test "$(id -u)" != 0; then
+	set -- \
+		unshare \
+		--user \
+		--mount \
+		--ipc \
+		--pid \
+		--uts \
+		--fork \
+		--kill-child=TERM \
+		--map-user=65536 \
+		--map-users=auto \
+		--map-group=65536 \
+		--map-groups=auto \
+		--propagation=private \
+		--mount-proc \
+		--setuid=0 \
+		--setgid=0 \
+		"$@"
+fi
+echo "+ $*" >&2
+"$@" || die "mmdebstrap failed"
+
+CMDLINE="root=LABEL=$ROOTLABEL rw"
+case "${DEBARCH%%,*}" in
+	amd64|i386)
+		CMDLINE="$CMDLINE console=ttyS0"
+	;;
+esac
+
+set -- "$UKIFY" \
+	build \
+	$UKIFY_VERBOSE \
+	--efi-arch "$EFIARCH" \
+	--linux "$WORKDIR/kernel" \
+	--initrd "$WORKDIR/initrd" \
+	--cmdline "$CMDLINE" \
+	--stub "$WORKDIR/stub" \
+	--output "$WORKDIR/efiimg"
+echo "+ $*" >&2
+"$@" ||
+	die "failed to generate UKI"
+
+rm -f "$WORKDIR/kernel" "$WORKDIR/initrd" "$WORKDIR/stub"
+
+truncate -s "${FAT_SIZE_KIB}K" "$WORKDIR/fat"
+/sbin/mkfs.fat -n "$EFILABEL" -F 32 --invariant "$WORKDIR/fat"
+mmd -i "$WORKDIR/fat" EFI EFI/BOOT
+mcopy -i "$WORKDIR/fat" "$WORKDIR/efiimg" "::EFI/BOOT/boot$EFIARCH.efi"
+
+rm -f "$WORKDIR/efiimg"
+
+/sbin/sfdisk "$IMAGE" <<EOF
+label: gpt
+unit: sectors
+
+start=${FAT_OFFSET_KIB}KiB, size=${FAT_SIZE_KIB}KiB, type=C12A7328-F81F-11D2-BA4B-00A0C93EC93B
+start=$((FAT_OFFSET_KIB + FAT_SIZE_KIB))KiB, type=$ROOTGUID
+EOF
+
+dd if="$WORKDIR/fat" of="$IMAGE" conv=notrunc,sparse bs=1024 "seek=$FAT_OFFSET_KIB" status=none
diff --git a/bin/debefivm-run b/bin/debefivm-run
new file mode 100755
index 0000000..c93d7d2
--- /dev/null
+++ b/bin/debefivm-run
@@ -0,0 +1,401 @@
+#!/bin/sh
+# SPDX-FileCopyrightText: 2024-2025 Helmut Grohne <helmut@subdivi.de>
+# SPDX-License-Identifier: MIT
+
+: <<'POD2MAN'
+=head1 NAME
+
+debefivm-run - Run a virtual machine from an EFI-bootable raw disk image
+
+=head1 SYNOPSIS
+
+B<-debefivm-run> [B<-a>] I<architecture>] [B<-i> F<image>] [B<-s> I<sshport>] [B<--> I<qemu options>]
+
+=head1 DESCRIPTION
+
+B<debefivm-run> is essentially a thing wrapper around B<qemu> for running a virtual machine from an UEFI bootable raw disk image.
+Such an image may be created using B<debefivm-create> or with another image creator, but its use is limited to architectures supporting EFI booting.
+
+=head1 OPTIONS
+
+=over 8
+
+=item B<-a> I<architecture>, B<--architecture>=I<architecture>
+
+Override the Debian architecture of the provided image.
+If the image uses architecture-specific type UUIDs for the root partition, the architecture can be detected.
+Otherwise, the host's architecture is assumed.
+The images created by B<debefivm-create> employ these UUIDs.
+The value is used to determine the correct emulator binary as well as suitable EFI firmware.
+
+=item B<--efi-vars>=F<variablefile>
+
+EFI variables can be changed and retained across reboots of a virtual machine if a separate variable file is supplied.
+The passed file is created from a template if absent.
+If absent, a read-only variable store will be supplied to the virtual machine.
+
+=item B<-i> F<image>, B<--image>=F<image>
+
+This option specifies the location of the virtual machine image file.
+By default F<vm.img> in the working directory is used.
+
+=item B<--netopt>=I<option>
+
+B<debefivm-run> sets up a user mode network by default.
+It therefore passes a B<-netdev> option to B<qemu>.
+Using this option, you can customize the value of that B<-netdev> option.
+For instance, you can set up additional port forwards by passing e.g. C<--netopt hostfwd=:127.0.0.1:8080-:80>.
+It can be used multiple times.
+
+=item B<--skip>=I<task>
+
+Skip a particular task or feature.
+The option may be specified multiple times or list multiple tasks to be skipped by separating them with a comma.
+By default, no tasks are skipped.
+The following tasks may be skipped.
+
+=over 4
+
+=item B<network>
+
+Do not configure a network card.
+Use this if you want to configure network on your own.
+This should also be passed in addition to passing C<-nic none> when you want to disable networking.
+
+=item B<rngdev>
+
+Do not pass a random number generator device.
+
+=back
+
+=item B<--transport>=I<transport>
+
+When B<debefivm-run> adds devices to B<qemu>, it has to select a transport and it most often guesses B<pci>.
+When specifying a different machine such as B<-machine microvm>, a different transport such as B<device> may be needed.
+
+=item B<-s> I<sshport>, B<--sshport>=I<sshport>
+
+If given, B<qemu> is configured to pass connections to I<127.0.0.1:sshport> to port 22 of the virtual machine.
+You can connect to your virtual machine without updating your known hosts like this:
+
+    ssh -o NoHostAuthenticationForLocalhost=yes -p $sshport root@127.0.0.1
+
+The option is a shorthand for C<--netopt hostfwd=tcp:127.0.0.1:sshport-:22>.
+
+=item B<--> I<qemu options>
+
+All options beyond a double dash are passed to B<qemu>.
+This can be used to configure additional hardware components.
+One possible use of this method is passing B<-snapshot> to avoid modifying the virtual machine image.
+
+=back
+
+=head1 SEE ALSO
+
+	debefivm-create(1) debvm-run(1) qemu(1)
+
+=cut
+POD2MAN
+
+set -u
+
+EFI_VARS=
+GRAPHICAL=
+IMAGE=vm.img
+IMAGE_ARCH=
+NATIVE_ARCH=$(dpkg --print-architecture)
+NETOPTS=
+SKIP=,
+SSHPORT=
+TRANSPORT=
+
+nth_arg() {
+	shift "$1"
+	printf "%s" "$1"
+}
+
+die() {
+	echo "error: $*" >&2
+	exit 1
+}
+usage() {
+	die "usage: $0 [-a architecture] [-g] [-i image] [-s sshport] [-- qemu options]"
+}
+usage_error() {
+	echo "error: $*" 1>&2
+	usage
+}
+
+opt_architecture() {
+	IMAGE_ARCH=$1
+}
+opt_efi_vars() {
+	EFI_VARS=$1
+}
+opt_graphical() {
+	GRAPHICAL=1
+}
+opt_help() {
+	usage
+}
+opt_image() {
+	IMAGE=$1
+}
+opt_netopt() {
+	NETOPTS="$NETOPTS,$1"
+}
+opt_skip() {
+	SKIP="$SKIP$1,"
+}
+opt_sshport() {
+	SSHPORT=$1
+}
+opt_transport() {
+	TRANSPORT=$1
+}
+
+while getopts :a:gi:s:-: OPTCHAR; do
+	case "$OPTCHAR" in
+		a)	opt_architecture "$OPTARG"	;;
+		g)	opt_graphical			;;
+		i)	opt_image "$OPTARG"		;;
+		s)	opt_sshport "$OPTARG"		;;
+		-)
+			case "$OPTARG" in
+				graphical|help)
+					"opt_$OPTARG"
+				;;
+				architecture|efi-vars|image|netopt|skip|sshport|transport)
+					test "$OPTIND" -gt "$#" && usage_error "missing argument for --$OPTARG"
+					"opt_$OPTARG" "$(nth_arg "$OPTIND" "$@")"
+					OPTIND=$((OPTIND+1))
+				;;
+				architecture=*|efi-vars=*|image=*|netopt=*|skip=*|sshport=*|transport=*)
+					"opt_${OPTARG%%=*}" "${OPTARG#*=}"
+				;;
+				*)
+					usage_error "unrecognized option --$OPTARG"
+				;;
+			esac
+		;;
+		:)
+			usage_error "missing argument for -$OPTARG"
+		;;
+		'?')
+			usage_error "unrecognized option -$OPTARG"
+		;;
+		*)
+			die "internal error while parsing command options, please report a bug"
+		;;
+	esac
+done
+shift "$((OPTIND - 1))"
+
+if test -n "$SSHPORT"; then
+	opt_netopt "hostfwd=tcp:127.0.0.1:$SSHPORT-:22"
+fi
+
+test -f "$IMAGE" || die "image '$IMAGE' not found"
+test -s "$IMAGE" || die "image '$IMAGE' is empty"
+
+check_skip() {
+	case "$SKIP" in
+		*",$1,"*)	return 0 ;;
+		*)		return 1 ;;
+	esac
+}
+
+if test -z "$IMAGE_ARCH"; then
+	PARTITIONS=$(partx --show --noheadings --output type "$IMAGE")
+
+	case "$PARTITIONS" in
+		*4f68bce3-e8cd-4db1-96e7-fbcaf984b709*)
+			IMAGE_ARCH=amd64
+		;;
+		*b921b045-1df0-41c3-af44-4c6f280d3fae*)
+			IMAGE_ARCH=arm64
+		;;
+		*72ec70a6-cf74-40e6-bd49-4bda08e8f224*)
+			IMAGE_ARCH=riscv64
+		;;
+		*69dad710-2ce4-4e3c-b16c-21a1d49abed3*)
+			IMAGE_ARCH=armhf
+		;;
+		*44479540-f297-41b2-9af7-d131d5f0458a*)
+			IMAGE_ARCH=i386
+		;;
+		*)
+			echo "cannot detect image architecture from gpt, assuming $NATIVE_ARCH" >&2
+			IMAGE_ARCH="$NATIVE_ARCH"
+		;;
+	esac
+fi
+
+# Translate IMAGE_ARCH (a Debian architecture) to a Debian CPU name.
+# This utilizes the QEMU Debian package symlink mapping that ensures that
+# calling qemu-system-${DEB_HOST_ARCH_CPU} will run the QEMU binary providing
+# the correct emulator for that CPU.
+IMAGEARCHCPU="$(dpkg-architecture --force --host-arch "$IMAGE_ARCH" --query DEB_HOST_ARCH_CPU)"
+QEMU="qemu-system-$IMAGEARCHCPU"
+CPU=
+MACHINE=
+MAX_SMP=
+
+case "$IMAGEARCHCPU" in
+	amd64)
+		QEMU=qemu-system-x86_64
+		MACHINE="type=q35"
+		BIOSCODE='/usr/share/OVMF/OVMF_CODE_4M.fd'
+		BIOSDATA='/usr/share/OVMF/OVMF_VARS_4M.fd'
+		BIOSPACKAGE=ovmf
+	;;
+	arm)
+		CPU=max
+		MACHINE="type=virt,highmem=off"
+		MAX_SMP=8
+		BIOSCODE='/usr/share/AAVMF/AAVMF32_CODE.fd'
+		BIOSDATA='/usr/share/AAVMF/AAVMF32_VARS.fd'
+		BIOSPACKAGE=qemu-efi-arm
+	;;
+	arm64)
+		CPU=max,pauth-impdef=on
+		MACHINE="type=virt,gic-version=max"
+		BIOSCODE='/usr/share/AAVMF/AAVMF_CODE.fd'
+		BIOSDATA='/usr/share/AAVMF/AAVMF_VARS.fd'
+		BIOSPACKAGE=qemu-efi-aarch64
+	;;
+	i386)
+		BIOSCODE='/usr/share/OVMF/OVMF32_CODE_4M.secboot.fd'
+		BIOSDATA='/usr/share/OVMF/OVMF32_VARS_4M.fd'
+		BIOSPACKAGE=ovfm-ia32
+	;;
+	riscv64)
+		MACHINE="type=virt"
+		BIOSCODE='/usr/share/qemu-efi-riscv64/RISCV_VIRT_CODE.fd'
+		BIOSDATA='/usr/share/qemu-efi-riscv64/RISCV_VIRT_VARS.fd'
+		BIOSPACKAGE=qemu-efi-riscv64
+	;;
+	*)
+		die "support for $IMAGE_ARCH is not implemented"
+	;;
+esac
+
+test -e "$BIOSCODE" ||
+	die "cannot file firmware file $BIOSCODE. Is $BIOSPACKAGE installed?"
+test -e "$BIOSDATA" ||
+	die "cannot file firmware file $BIOSDATA. Is $BIOSPACKAGE installed?"
+
+# Assign the default late to allow both cli and arch-specific overrides.
+: "${TRANSPORT:=pci}"
+
+comma_escape() {
+	# If a filename contains a comma, then that comma must be escaped by
+	# prefixing it with another comma or otherwise output filenames are
+	# able to inject options to qemu (and load the wrong file).
+	comma_escape_str="$1"
+	while test "${comma_escape_str%,*}" != "$comma_escape_str"; do
+		printf "%s,," "${comma_escape_str%%,*}"
+		comma_escape_str="${comma_escape_str#*,}"
+	done
+	printf "%s" "$comma_escape_str"
+}
+
+EFI_VAR_DRIVE="if=pflash,format=raw,unit=1"
+if test -n "$EFI_VARS"; then
+	if ! test -e "$EFI_VARS"; then
+		cp "$BIOSDATA" "$EFI_VARS"
+	fi
+	EFI_VAR_DRIVE="$EFI_VAR_DRIVE,read-only=off,file=$(comma_escape "$EFI_VARS")"
+else
+	EFI_VAR_DRIVE="$EFI_VAR_DRIVE,read-only=on,file=$(comma_escape "$BIOSDATA")"
+fi
+
+ENABLE_KVM=no
+if test "$NATIVE_ARCH" = "$IMAGE_ARCH"; then
+	ENABLE_KVM=yes
+fi
+if test "$ENABLE_KVM" = yes; then
+	if ! command -v "$QEMU" >/dev/null 2>&1; then
+		# Fall back to kvm in case we badly guessed qemu.
+		QEMU=kvm
+	fi
+	MACHINE="${MACHINE:+$MACHINE,}accel=kvm:tcg"
+	# While kvm will fall back gracefully, only override CPU when we expect
+	# kvm to work.
+	if test -w /dev/kvm; then
+		CPU=max
+		# kvm: "max" will become "host", intended.
+		# tcg: "max" will actually work, "host" would not.
+	fi
+fi
+
+if test -n "$MACHINE"; then
+	set -- -machine "$MACHINE" "$@"
+fi
+
+set -- \
+	-no-user-config \
+	-nodefaults \
+	-chardev stdio,id=console,mux=on,signal=off \
+	-serial chardev:console \
+	-name "debefivm-run $IMAGE" \
+	-m 1G \
+	-drive "if=pflash,format=raw,unit=0,read-only=on,file=$(comma_escape "$BIOSCODE")" \
+	-drive "$EFI_VAR_DRIVE" \
+	-drive "id=root,media=disk,format=raw,discard=unmap,file=$(comma_escape "$IMAGE"),if=none,cache=unsafe" \
+	-device "virtio-blk-$TRANSPORT,drive=root,serial=root" \
+	"$@"
+
+if test -n "$CPU"; then
+	set -- -cpu "$CPU" "$@"
+fi
+
+NPROC=$(nproc)
+if test "$NPROC" -gt 1; then
+	test -n "$MAX_SMP" && test "$NPROC" -gt "$MAX_SMP" && NPROC=$MAX_SMP
+	set -- -smp "$NPROC" "$@"
+fi
+if ! check_skip rngdev; then
+	set -- \
+		-device "virtio-rng-$TRANSPORT,rng=rng0" \
+		-object rng-random,filename=/dev/urandom,id=rng0 \
+		"$@"
+fi
+
+if test -z "$GRAPHICAL"; then
+	set -- -nographic "$@"
+else
+	case "$KERNELARCH" in
+		amd64|i386)
+			set -- -vga virtio "$@"
+		;;
+		*)
+			set -- \
+				-device "virtio-gpu-gl-$TRANSPORT" \
+				-display gtk,gl=on \
+				"$@"
+		;;
+	esac
+	set -- \
+		-device "virtio-keyboard-$TRANSPORT" \
+		-device "virtio-tablet-$TRANSPORT" \
+		"$@"
+fi
+
+DNSSEARCH=$(dnsdomainname)
+if test -z "$DNSSEARCH"; then
+	DNSSEARCH=$(sed -n 's/^\s*search\s*\.\?//p;T;q' /etc/resolv.conf)
+fi
+if test -n "$DNSSEARCH"; then
+	NETOPTS=",domainname=$DNSSEARCH$NETOPTS"
+fi
+
+if ! check_skip network; then
+	set -- \
+		-netdev "user,id=net0$NETOPTS" \
+		-device "virtio-net-$TRANSPORT,netdev=net0" \
+		"$@"
+fi
+
+echo "+ $QEMU $*" 1>&2
+exec "$QEMU" "$@"
diff --git a/bin/debefivm-ukify b/bin/debefivm-ukify
new file mode 100755
index 0000000..9267158
--- /dev/null
+++ b/bin/debefivm-ukify
@@ -0,0 +1,329 @@
+#!/bin/sh
+# SPDX-FileCopyrightText: 2023-2025 Helmut Grohne <helmut@subdivi.de>
+# SPDX-License-Identifier: MIT
+#
+
+: <<'POD2MAN'
+=head1 NAME
+
+debefivm-ukify - Create Unified Kernel Images for UEFI systems in a similar way to systemd's ukify
+
+=head1 SYNOPSIS
+
+B<debefivm-ukify> B<build> B<--linux>=F<vmlinuz> B<--output>=F<uki.efi> [options]
+
+=head1 DESCRIPTION
+
+B<debefivm-ukify> is a reimplementation of parts of B<systemd-ukify> and mainly meant for backports.
+Please prefer B<systemd>'s implementation.
+It only supports the B<build> subcommand, does not implement any functionality related to secure boot and lacks configuration file support.
+Other than that, it can be used to assemble a linux kernel image, an initrd and an EFI stub image from B<systemd> into an EFI-bootable Unified Kernel Image.
+
+=head1 OPTIONS
+
+=over 8
+
+=item B<--deb-arch>=I<debian_architecture>
+
+Set the architecture of the supplied and produced artifacts using Debian's architecture nomenclature.
+See B<--efi-arch> for another supplying it using the EFI nomenclature instead.
+This option is not available on the B<systemd> implementation.
+
+=item B<--devicetree>=F<devicetree_blob>
+
+Add a device tree file as a dtb section to the generated UKI.
+There can be at most one device tree.
+
+=item B<--cmdline>=I<kernel_cmdline>
+
+Supply extra arguments to be appended to the Linux kernel command line during boot.
+Can be given given at most once.
+If the value starts with an B<@>, it is treated as a filename whose contents will be added.
+
+=item B<--efi-arch>=I<efi_architecture>
+
+Set the architecture of the supplied and produced artifacts using EFI's architecture nomenclature.
+See B<--deb-arch> for another supplying it using the Debian nomenclature instead.
+
+=item B<--initrd>=F<initrd>
+
+Supply an initrd image to be embedded.
+There can be at most one initrd.
+
+=item B<--linux>=F<vmlinuz>
+
+Supply a Linux kernel image to be embedded.
+This option must be given exactly once.
+
+=item B<--os-release>=I<release_information>
+
+Set the os-release description of in the UKI.
+Can be given given at most once.
+If the value starts with an B<@>, it is treated as a filename whose contents will be added.
+
+=item B<--output>=<uki.efi>
+
+Set the name of the resulting UKI image.
+This option must be given exactly once and the target file must not exist already.
+
+=item B<--section>=<name:value>
+
+Add an arbitrary ELF section to the resulting UKI.
+If the value starts with an B<@>, it is treated as a filename whose contents will be added.
+Most image components actually are such ELF sections.
+For instance B<--os-release>, adds a section named B<osrel>.
+
+=item B<--stub>=F<efi_stub>
+
+Supply the location of B<systemd>'s EFI stub image.
+Unlike the B<systemd> implementation, this option accepts a directory containing stubs named by architecture.
+By default, stubs will be looked up in B</usr/lib/systemd/boot/efi>.
+
+=item B<--verbose>
+
+Print diagnostic information during image conversion.
+This option is not available on the B<systemd> implementation.
+
+=back
+
+=head1 SEE ALSO
+
+    ukify(1)
+
+=cut
+POD2MAN
+
+set -eu
+
+die() {
+	echo "$*" >&2
+	exit 1
+}
+usage() {
+	die "usage: $0 [build] <--linux=|--initrd=|--stub=> [--cmdline=|--deb-arch=|--devicetree=|--efi-arch=|--os-release=|--output=|--section=|--verbose]"
+}
+usage_error() {
+	echo "error: $*" >&2
+	usage
+}
+
+DEBARCH=
+BOOTSTUB=/usr/lib/systemd/boot/efi
+OUTPUT=
+VERBOSE=
+
+TDIR=
+cleanup() {
+	if test -n "$TDIR"; then
+		rm -Rf "$TDIR"
+	fi
+}
+cleanup_abort() {
+	cleanup
+	echo aborted >&2
+	exit 2
+}
+trap cleanup EXIT
+trap cleanup_abort HUP INT TERM QUIT
+
+# The TDIR will contain one file per section. It can be a symbolic link or
+# regular file. The section name is the filename with a leading dot.
+TDIR=$(mktemp --directory --tmpdir mini-ukify.XXXXXXXXXX)
+
+add_section_content() {
+	test -e "$TDIR/$1" && die "section '$1' already defined"
+	printf "%s" "$2" > "$TDIR/$1"
+}
+add_section_file() {
+	test -e "$TDIR/$1" && die "section '$1' already defined"
+	test -f "$2" || die "file '$2' does not exist"
+	if test "${2#/}" = "$2"; then
+		ln --symbolic --relative "$2" "$TDIR/$1"
+	else
+		ln --symbolic "$2" "$TDIR/$1"
+	fi
+}
+add_section_either() {
+	if test "${2#@}" = "$2"; then
+		add_section_content "$1" "$2"
+	else
+		add_section_file "$1" "${2#@}"
+	fi
+}
+
+opt_cmdline() {
+	add_section_either cmdline "$1"
+}
+opt_deb_arch() {
+	case "$1" in
+		amd64)
+			EFIARCH=x64
+			GNU_TYPE=x86_64-linux-gnu
+		;;
+		arm64)
+			EFIARCH=aa64
+			GNU_TYPE=aarch64-linux-gnu
+		;;
+		armhf)
+			EFIARCH=arm
+			GNU_TYPE=arm-linux-gnueabihf
+		;;
+		i386)
+			EFIARCH=ia32
+			GNU_TYPE=i686-linux-gnu
+		;;
+		riscv64)
+			EFIARCH=riscv64
+			GNU_TYPE=riscv64-linux-gnu
+		;;
+		*)
+			die "unsupported Debian architecture: $1"
+		;;
+	esac
+	DEBARCH="$1"
+}
+opt_devicetree() {
+	add_section_file dtb "$1"
+}
+opt_efi_arch() {
+	case "$1" in
+		aa64) opt_deb_arch arm64 ;;
+		arm) opt_deb_arch armhf ;;
+		ia32) opt_deb_arch i386 ;;
+		x64) opt_deb_arch amd64 ;;
+		riscv64) opt_deb_arch "$1" ;;
+		*)
+			die "unsupported EFI architecture: $1"
+		;;
+	esac
+}
+opt_initrd() {
+	add_section_file initrd "$1"
+}
+opt_linux() {
+	add_section_file linux "$1"
+	test -z "$OUTPUT" && OUTPUT="$1.unsigned.efi"
+}
+opt_os_release() {
+	add_section_either osrel "$1"
+}
+opt_output() {
+	OUTPUT="$1"
+}
+opt_section() {
+	case "$1" in
+		*[!a-z]*:*)
+			die "invalid section name '${1%%:*}'"
+		;;
+		*:*)
+		;;
+		*)
+			die "missing section name separated by colon in '$1'"
+		;;
+	esac
+	add_section_either "${1%%:*}" "${1#*:}"
+}
+opt_stub() {
+	BOOTSTUB="$1"
+}
+opt_verbose() {
+	VERBOSE=yes
+}
+positional=1
+positional_1() {
+	test "$1" = "build" || usage_error "command not understood: '$1'"
+}
+positional_2() {
+	usage_error "too many positional arguments"
+}
+
+while test "$#" -gt 0; do
+	case "$1" in
+		--cmdline=*|--deb-arch=*|--devicetree=*|--efi-arch=*|--initrd=*|--linux=*|--os-release=*|--output=*|--section=*|--stub=*)
+			optname="${1%%=*}"
+			optname="${optname#--}"
+			test "${optname#*-}" = "$optname" ||
+				optname="${optname%%-*}_${optname#*-}"
+			"opt_$optname" "${1#*=}"
+		;;
+		--cmdline|--deb-arch|--devicetree|--efi-arch|--initrd|--linux|--os-release|--output|--section|--stub)
+			test "$#" -ge 2 || usage_error "missing argument for $1"
+			optname="${1#--}"
+			test "${optname#*-}" = "$optname" ||
+				optname="${optname%%-*}_${optname#*-}"
+			"opt_$optname" "$2"
+			shift
+		;;
+		--verbose)
+			"opt_${1#--}"
+		;;
+		--*)
+			usage_error "unrecognized option $1"
+		;;
+		*)
+			"positional_$positional" "$1"
+			positional=$((positional + 1))
+		;;
+	esac
+	shift
+done
+
+test "$positional" = 2 || usage_error "missing subcommand, only 'build' is supported"
+test -z "$DEBARCH" && opt_deb_arch "$(dpkg --print-architecture)"
+test -e "$TDIR/linux" || usage_error "missing --linux argument"
+test -e "$BOOTSTUB" || die "stub image or directory '$BOOTSTUB' does not exist"
+test -e "$OUTPUT" && die "output '$OUTPUT' already exists"
+
+test -d "$BOOTSTUB" && BOOTSTUB="$BOOTSTUB/linux$EFIARCH.efi.stub"
+test -f "$BOOTSTUB" || die "efi boot stub $BOOTSTUB not found"
+
+GNU_PREFIX="$GNU_TYPE-"
+test "$(dpkg-query --showformat='${db:Status-Status}' --show binutils-multiarch)" = installed &&
+	GNU_PREFIX=
+
+# Compute the next multiple of $2 greater than or equal to $1.
+align_size() {
+	echo "$((($1) + ($2) - 1 - (($1) + ($2) - 1) % ($2)))"
+}
+alignment=$("${GNU_PREFIX}objdump" --private-headers "$BOOTSTUB" | sed 's/^SectionAlignment\s\+\([0-9]\)/0x/;t;d')
+test -z "$alignment" &&
+	die "failed to discover the alignment of the efi stub"
+test -n "$VERBOSE" &&
+	echo "determined efi vma alignment as $alignment"
+
+# Discover the last section in terms of vma + size. We'll append sections
+# beyond.
+lastoffset=0
+# shellcheck disable=SC2034  # unused variables serve documentation
+lastoffset="$("${GNU_PREFIX}objdump" --section-headers "$BOOTSTUB" \
+	| while read -r idx name size vma lma fileoff algn behind; do
+		test -z "$behind" -a "${algn#"2**"}" != "$algn" || continue
+		offset=$((0x$vma + 0x$size))
+		test "$offset" -gt "$lastoffset" || continue
+		lastoffset="$offset"
+		echo "$lastoffset"
+	done | tail -n1)"
+lastoffset=$(align_size "$lastoffset" "$alignment")
+test -n "$VERBOSE" &&
+	echo "determined minimum efi vma offset as $lastoffset"
+
+# Compute the objdump invocation that constructs the UKI. Successively add
+# sections and compute non-overlapping VMAs.
+set -- \
+	"${GNU_PREFIX}objcopy" \
+	--enable-deterministic-archives
+for sectionfile in "$TDIR/"*; do
+	section="${sectionfile##*/}"
+	size=$(stat -Lc%s "$sectionfile")
+	size=$(align_size "$size" "$alignment")
+	set -- "$@" \
+		--add-section ".$section=$sectionfile" \
+		--change-section-vma ".$section=$lastoffset"
+	lastoffset=$((lastoffset + size))
+done
+set -- "$@" "$BOOTSTUB" "$OUTPUT"
+if test -n "$VERBOSE"; then
+	ls -lA "$TDIR"
+	printf "%s\n" "$*"
+fi
+"$@" || die "failed to construct UKI"