kata-containers/tools/osbuilder/image-builder/image_builder.sh
Jianyong Wu 2b0d5b252e image_build: align image size to 128M for arm64
There is an inconformity between qemu and kernel of memory alignment
check in memory hotplug. Both of qemu and kernel will do the start
address alignment check in memory hotplug. But it's 2M in qemu
while 128M in kernel. It leads to an issue when memory hotplug.

Currently, the kata image is a nvdimm device, which will plug into the VM as
a dimm. If another dimm is pluged, it will reside on top of that nvdimm.
So, the start address of the second dimm may not pass the alginment
check in kernel if the nvdimm size doesn't align with 128M.

There are 3 ways to address this issue I think:
1. fix the alignment size in kernel according to qemu. I think people
in linux kernel community will not accept it.
2. do alignment check in qemu and force the start address of hotplug
in alignment with 128M, which means there maybe holes between memory blocks.
3. obey the rule in user end, which means fix it in kata.

I think the second one is the best, but I can't do that for some reason.
Thus, the last one is the choice here.

Fixes: #1769
Signed-off-by: Jianyong Wu <jianyong.wu@arm.com>
2021-05-03 10:44:30 +08:00

532 lines
14 KiB
Bash
Executable File

#!/usr/bin/env bash
#
# Copyright (c) 2017-2019 Intel Corporation
#
# SPDX-License-Identifier: Apache-2.0
set -e
[ -n "${DEBUG}" ] && set -x
DOCKER_RUNTIME=${DOCKER_RUNTIME:-runc}
readonly script_name="${0##*/}"
readonly script_dir=$(dirname "$(readlink -f "$0")")
readonly lib_file="${script_dir}/../scripts/lib.sh"
readonly ext4_format="ext4"
readonly xfs_format="xfs"
# ext4: percentage of the filesystem which may only be allocated by privileged processes.
readonly reserved_blocks_percentage=3
# Where the rootfs starts in MB
readonly rootfs_start=1
# Where the rootfs ends in MB
readonly rootfs_end=-1
# DAX header size
# * NVDIMM driver reads the device namespace information from nvdimm namespace (4K offset).
# The MBR #1 + DAX metadata are saved in the first 2MB of the image.
readonly dax_header_sz=2
# DAX aligment
# * DAX huge pages [2]: 2MB alignment
# [2] - https://nvdimm.wiki.kernel.org/2mib_fs_dax
readonly dax_alignment=2
# The list of systemd units and files that are not needed in Kata Containers
readonly -a systemd_units=(
"systemd-coredump@"
"systemd-journald"
"systemd-journald-dev-log"
"systemd-journal-flush"
"systemd-random-seed"
"systemd-timesyncd"
"systemd-tmpfiles-setup"
"systemd-udevd"
"systemd-udevd-control"
"systemd-udevd-kernel"
"systemd-udev-trigger"
"systemd-update-utmp"
)
readonly -a systemd_files=(
"systemd-bless-boot-generator"
"systemd-fstab-generator"
"systemd-getty-generator"
"systemd-gpt-auto-generator"
"systemd-tmpfiles-cleanup.timer"
)
# Set a default value
AGENT_INIT=${AGENT_INIT:-no}
# Align image to 128M
readonly mem_boundary_mb=128
# shellcheck source=../scripts/lib.sh
source "${lib_file}"
usage() {
cat <<EOT
Usage: ${script_name} [options] <rootfs-dir>
This script will create a Kata Containers image file of
an adequate size based on the <rootfs-dir> directory.
Options:
-h Show this help
-o path to generate image file ENV: IMAGE
-r Free space of the root partition in MB ENV: ROOT_FREE_SPACE
Extra environment variables:
AGENT_BIN: Use it to change the expected agent binary name
AGENT_INIT: Use kata agent as init process
IMAGE_REGISTRY: Hostname for the image registry used to pull down the rootfs build image.
FS_TYPE: Filesystem type to use. Only xfs and ext4 are supported.
NSDAX_BIN: Use to specify path to pre-compiled 'nsdax' tool.
USE_DOCKER: If set will build image in a Docker Container (requries docker)
DEFAULT: not set
USE_PODMAN: If set and USE_DOCKER not set, will build image in a Podman Container (requries podman)
DEFAULT: not set
Following diagram shows how the resulting image will look like
.-----------.----------.---------------.-----------.
| 0 - 512 B | 4 - 8 Kb | 2M - 2M+512B | 3M |
|-----------+----------+---------------+-----------+
| MBR #1 | DAX | MBR #2 | Rootfs |
'-----------'----------'---------------'-----------+
| | ^ | ^
| '-data-' '--------'
| |
'--------rootfs-partition---------'
MBR: Master boot record.
DAX: Metadata required by the NVDIMM driver to enable DAX in the guest [1][2] (struct nd_pfn_sb).
Rootfs: partition that contains the root filesystem (/usr, /bin, ect).
Kernels and hypervisors that support DAX/NVDIMM read the MBR #2, otherwise MBR #1 is read.
[1] - https://github.com/kata-containers/kata-containers/blob/main/tools/osbuilder/image-builder/nsdax.gpl.c
[2] - https://github.com/torvalds/linux/blob/master/drivers/nvdimm/pfn.h
EOT
}
# build the image using container engine
build_with_container() {
local rootfs="$1"
local image="$2"
local fs_type="$3"
local block_size="$4"
local root_free_space="$5"
local agent_bin="$6"
local agent_init="$7"
local container_engine="$8"
local nsdax_bin="$9"
local container_image_name="image-builder-osbuilder"
local shared_files=""
image_dir=$(readlink -f "$(dirname "${image}")")
image_name=$(basename "${image}")
REGISTRY_ARG=""
if [ -n "${IMAGE_REGISTRY}" ]; then
REGISTRY_ARG="--build-arg IMAGE_REGISTRY=${IMAGE_REGISTRY}"
fi
"${container_engine}" build \
${REGISTRY_ARG} \
--build-arg http_proxy="${http_proxy}" \
--build-arg https_proxy="${https_proxy}" \
-t "${container_image_name}" "${script_dir}"
readonly mke2fs_conf="/etc/mke2fs.conf"
if [ -f "${mke2fs_conf}" ]; then
shared_files+="-v ${mke2fs_conf}:${mke2fs_conf}:ro "
fi
#Make sure we use a compatible runtime to build rootfs
# In case Clear Containers Runtime is installed we dont want to hit issue:
#https://github.com/clearcontainers/runtime/issues/828
"${container_engine}" run \
--rm \
--runtime "${DOCKER_RUNTIME}" \
--privileged \
--env AGENT_BIN="${agent_bin}" \
--env AGENT_INIT="${agent_init}" \
--env FS_TYPE="${fs_type}" \
--env BLOCK_SIZE="${block_size}" \
--env ROOT_FREE_SPACE="${root_free_space}" \
--env NSDAX_BIN="${nsdax_bin}" \
--env DEBUG="${DEBUG}" \
-v /dev:/dev \
-v "${script_dir}":"/osbuilder" \
-v "${script_dir}/../scripts":"/scripts" \
-v "${rootfs}":"/rootfs" \
-v "${image_dir}":"/image" \
${shared_files} \
${container_image_name} \
bash "/osbuilder/${script_name}" -o "/image/${image_name}" /rootfs
}
check_rootfs() {
local rootfs="${1}"
[ -d "${rootfs}" ] || die "${rootfs} is not a directory"
# The kata rootfs image expect init and kata-agent to be installed
init_path="/sbin/init"
init="${rootfs}${init_path}"
if [ ! -x "${init}" ] && [ ! -L "${init}" ]; then
error "${init_path} is not installed in ${rootfs}"
return 1
fi
OK "init is installed"
candidate_systemd_paths="/usr/lib/systemd/systemd /lib/systemd/systemd"
# check agent or systemd
case "${AGENT_INIT}" in
"no")
for systemd_path in $candidate_systemd_paths; do
systemd="${rootfs}${systemd_path}"
if [ -x "${systemd}" ] || [ -L "${systemd}" ]; then
found="yes"
break
fi
done
if [ ! $found ]; then
error "None of ${candidate_systemd_paths} is installed in ${rootfs}"
return 1
fi
OK "init is systemd"
;;
"yes")
agent_path="/sbin/init"
agent="${rootfs}${agent_path}"
if [ ! -x "${agent}" ]; then
error "${agent_path} is not installed in ${rootfs}. Use AGENT_BIN env variable to change the expected agent binary name"
return 1
fi
# checksum must be different to system
for systemd_path in $candidate_systemd_paths; do
systemd="${rootfs}${systemd_path}"
if [ -f "${systemd}" ] && cmp -s "${systemd}" "${agent}"; then
error "The agent is not the init process. ${agent_path} is systemd"
return 1
fi
done
OK "Agent installed"
;;
*)
error "Invalid value for AGENT_INIT: '${AGENT_INIT}'. Use to 'yes' or 'no'"
return 1
;;
esac
return 0
}
calculate_required_disk_size() {
local rootfs="$1"
local fs_type="$2"
local block_size="$3"
readonly rootfs_size_mb=$(du -B 1MB -s "${rootfs}" | awk '{print $1}')
readonly image="$(mktemp)"
readonly mount_dir="$(mktemp -d)"
readonly max_tries=20
readonly increment=10
for i in $(seq 1 $max_tries); do
local img_size="$((rootfs_size_mb + (i * increment)))"
create_disk "${image}" "${img_size}" "${fs_type}" "${rootfs_start}" > /dev/null 2>&1
if ! device="$(setup_loop_device "${image}")"; then
continue
fi
if ! format_loop "${device}" "${block_size}" "${fs_type}" > /dev/null 2>&1 ; then
die "Could not format loop device: ${device}"
fi
mount "${device}p1" "${mount_dir}"
avail="$(df -BM --output=avail "${mount_dir}" | tail -n1 | sed 's/[M ]//g')"
umount "${mount_dir}"
losetup -d "${device}"
if [ "${avail}" -gt "${rootfs_size_mb}" ]; then
rmdir "${mount_dir}"
rm -f "${image}"
echo "${img_size}"
return
fi
done
rmdir "${mount_dir}"
rm -f "${image}"
error "Could not calculate the required disk size"
}
# Calculate image size based on the rootfs and free space
calculate_img_size() {
local rootfs="$1"
local root_free_space_mb="$2"
local fs_type="$3"
local block_size="$4"
# rootfs start + DAX header size + rootfs end
local reserved_size_mb=$((rootfs_start + dax_header_sz + rootfs_end))
disk_size="$(calculate_required_disk_size "${rootfs}" "${fs_type}" "${block_size}")"
img_size="$((disk_size + reserved_size_mb))"
if [ -n "${root_free_space_mb}" ]; then
img_size="$((img_size + root_free_space_mb))"
fi
remaining="$((img_size % mem_boundary_mb))"
if [ "${remaining}" != "0" ]; then
img_size=$((img_size + mem_boundary_mb - remaining))
fi
echo "${img_size}"
}
setup_loop_device() {
local image="$1"
# Get the loop device bound to the image file (requires /dev mounted in the
# image build system and root privileges)
device=$(losetup -P -f --show "${image}")
#Refresh partition table
partprobe -s "${device}" > /dev/null
# Poll for the block device p1
for _ in $(seq 1 5); do
if [ -b "${device}p1" ]; then
echo "${device}"
return 0
fi
sleep 1
done
error "File ${device}p1 is not a block device"
return 1
}
format_loop() {
local device="$1"
local block_size="$2"
local fs_type="$3"
case "${fs_type}" in
"${ext4_format}")
mkfs.ext4 -q -F -b "${block_size}" "${device}p1"
info "Set filesystem reserved blocks percentage to ${reserved_blocks_percentage}%"
tune2fs -m "${reserved_blocks_percentage}" "${device}p1"
;;
"${xfs_format}")
# DAX and reflink cannot be used together!
# Explicitly disable reflink, if it fails then reflink
# is not supported and '-m reflink=0' is not needed.
if mkfs.xfs -m reflink=0 -q -f -b size="${block_size}" "${device}p1" 2>&1 | grep -q "unknown option"; then
mkfs.xfs -q -f -b size="${block_size}" "${device}p1"
fi
;;
*)
error "Unsupported fs type: ${fs_type}"
return 1
;;
esac
}
create_disk() {
local image="$1"
local img_size="$2"
local fs_type="$3"
local part_start="$4"
info "Creating raw disk with size ${img_size}M"
qemu-img create -q -f raw "${image}" "${img_size}M"
OK "Image file created"
# Kata runtime expect an image with just one partition
# The partition is the rootfs content
info "Creating partitions"
parted -s -a optimal "${image}" -- \
mklabel msdos \
mkpart primary "${fs_type}" "${part_start}"M "${rootfs_end}"M
OK "Partitions created"
}
create_rootfs_image() {
local rootfs="$1"
local image="$2"
local img_size="$3"
local fs_type="$4"
local block_size="$5"
create_disk "${image}" "${img_size}" "${fs_type}" "${rootfs_start}"
if ! device="$(setup_loop_device "${image}")"; then
die "Could not setup loop device"
fi
if ! format_loop "${device}" "${block_size}" "${fs_type}"; then
die "Could not format loop device: ${device}"
fi
info "Mounting root partition"
readonly mount_dir=$(mktemp -p ${TMPDIR:-/tmp} -d osbuilder-mount-dir.XXXX)
mount "${device}p1" "${mount_dir}"
OK "root partition mounted"
info "Copying content from rootfs to root partition"
cp -a "${rootfs}"/* "${mount_dir}"
sync
OK "rootfs copied"
info "Removing unneeded systemd services and sockets"
for u in "${systemd_units[@]}"; do
find "${mount_dir}" -type f \( \
-name "${u}.service" -o \
-name "${u}.socket" \) \
-exec rm -f {} \;
done
info "Removing unneeded systemd files"
for u in "${systemd_files[@]}"; do
find "${mount_dir}" -type f -name "${u}" -exec rm -f {} \;
done
info "Creating empty machine-id to allow systemd to bind-mount it"
touch "${mount_dir}/etc/machine-id"
info "Unmounting root partition"
umount "${mount_dir}"
OK "Root partition unmounted"
if [ "${fs_type}" = "${ext4_format}" ]; then
fsck.ext4 -D -y "${device}p1"
fi
losetup -d "${device}"
rmdir "${mount_dir}"
}
set_dax_header() {
local image="$1"
local img_size="$2"
local fs_type="$3"
local nsdax_bin="$4"
# rootfs start + DAX header size
local rootfs_offset=$((rootfs_start + dax_header_sz))
local header_image="${image}.header"
local dax_image="${image}.dax"
rm -f "${dax_image}" "${header_image}"
create_disk "${header_image}" "${img_size}" "${fs_type}" "${rootfs_offset}"
dax_header_bytes=$((dax_header_sz * 1024 * 1024))
dax_alignment_bytes=$((dax_alignment * 1024 * 1024))
info "Set DAX metadata"
# Set metadata header
# Issue: https://github.com/kata-containers/osbuilder/issues/240
if [ -z "${nsdax_bin}" ] ; then
nsdax_bin="${script_dir}/nsdax"
gcc -O2 "${script_dir}/nsdax.gpl.c" -o "${nsdax_bin}"
trap "rm ${nsdax_bin}" EXIT
fi
"${nsdax_bin}" "${header_image}" "${dax_header_bytes}" "${dax_alignment_bytes}"
sync
touch "${dax_image}"
# Copy MBR #1 + DAX metadata
dd if="${header_image}" of="${dax_image}" bs="${dax_header_sz}M" count=1
# Copy MBR #2 + Rootfs
dd if="${image}" of="${dax_image}" oflag=append conv=notrunc
# final image
mv "${dax_image}" "${image}"
sync
rm -f "${dax_image}" "${header_image}"
}
main() {
[ "$(id -u)" -eq 0 ] || die "$0: must be run as root"
# variables that can be overwritten by environment variables
local agent_bin="${AGENT_BIN:-kata-agent}"
local agent_init="${AGENT_INIT:-no}"
local fs_type="${FS_TYPE:-${ext4_format}}"
local image="${IMAGE:-kata-containers.img}"
local block_size="${BLOCK_SIZE:-4096}"
local root_free_space="${ROOT_FREE_SPACE:-}"
local nsdax_bin="${NSDAX_BIN:-}"
while getopts "ho:r:f:" opt
do
case "$opt" in
h) usage; return 0;;
o) image="${OPTARG}" ;;
r) root_free_space="${OPTARG}" ;;
f) fs_type="${OPTARG}" ;;
*) break ;;
esac
done
shift $(( OPTIND - 1 ))
rootfs="$(readlink -f "$1")"
if [ -z "${rootfs}" ]; then
usage
exit 0
fi
local container_engine
if [ -n "${USE_DOCKER}" ]; then
container_engine="docker"
elif [ -n "${USE_PODMAN}" ]; then
container_engine="podman"
fi
if [ -n "$container_engine" ]; then
build_with_container "${rootfs}" \
"${image}" "${fs_type}" "${block_size}" \
"${root_free_space}" "${agent_bin}" \
"${agent_init}" "${container_engine}" \
"${nsdax_bin}"
exit $?
fi
if ! check_rootfs "${rootfs}" ; then
die "Invalid rootfs"
fi
img_size=$(calculate_img_size "${rootfs}" "${root_free_space}" "${fs_type}" "${block_size}")
# the first 2M are for the first MBR + NVDIMM metadata and were already
# consider in calculate_img_size
rootfs_img_size=$((img_size - dax_header_sz))
create_rootfs_image "${rootfs}" "${image}" "${rootfs_img_size}" \
"${fs_type}" "${block_size}"
# insert at the beginning of the image the MBR + DAX header
set_dax_header "${image}" "${img_size}" "${fs_type}" "${nsdax_bin}"
}
main "$@"