Immutable root with atomic upgrades

From Alpine Linux

What?

This article provides a basic guide to setting up a read-only-root-based Alpine Linux system with several boot environments and atomic upgrades using a modern bootloader and btrfs.

Why?

Read-only root and atomic upgrades with the ability to easily rollback or boot previous configurations is a concept that has been gaining popularity recently. Distributions providing and promoting such features, for example, are Fedora Silverblue, Opensuse MicroOS, NixOS and GNU Guix.

While Alpine Linux has its killer features, it lacks the ones mentioned above on default setup. This is a proof of concept that it's possible to implement them in a minimal way on a minimal system.

Note: Alpine Linux can also boot from RAM in diskless mode (see Installation) which supports preserving changes between reboots using lbu.

Preparation

You should have bootable Alpine media. The process to obtain it is described on the installation page.

Partitioning disks

In this guide, it's assumed that you have a fresh UEFI system without an OS and have just booted into a live Alpine system using a USB flash drive or CD. The first step is creating partition table on your HDD/SSD target device (/dev/sda here):

# apk add gptfdisk
# gdisk /dev/sda
> o ↵
> y ↵
> w ↵
> y ↵

Now we can define the partitions:

# cgdisk /dev/sda

Partition creation process consists of several steps:

  1. Start sector - you can safely use default value by pressing ↵
  2. Size
  3. Type (as hex code) - EFI is ef00, Linux filesystem is 8300, Swap is 8200.

Result table:

Part.     #     Size        Partition Type            Partition Name
----------------------------------------------------------------
1               200.0 MiB   EFI System                EFI
2               200.0 GiB   Linux filesystem          ROOT
3               32.0 GiB    Linux swap                SWAP

ROOT partition name will later be used in rEFInd configuration to identify boot volume. Next step is creating filesystems:

# mkfs.vfat -F32 /dev/sda1
# mkfs.btrfs /dev/sda2
# mkswap /dev/sda3

Now we can mount our root volume:

# mount -t btrfs /dev/sda2 /mnt

File system structure

Now we should create the file structure that would provide reliable atomic system upgrades.
Start with following directories:

# mkdir /mnt/next

Stores next current link, is necessary due to how busybox mv does atomic link replacement.

# mkdir /mnt/commons

Stores common non-snapshotting subvolumes.
We may populate it right away:

# btrfs subvolume create /mnt/commons/@var
# btrfs subvolume create /mnt/commons/@home

Next, most important directories:

# mkdir /mnt/snapshots

Stores directories containing snapshots belonging to one generation.

# mkdir /mnt/links

Stores generations of directories containing links to snapshot generations.
Let's create first generation and populate it with one OS root snapshot @:

# NEWSNAPSHOTS="$(date -u +"%Y%m%d%H%M%S")$(cat /dev/urandom | tr -dc 'a-zA-Z' | fold -w 8 | head -n 1)"
# mkdir "/mnt/snapshots/$NEWSNAPSHOTS"
# btrfs subvolume create /mnt/snapshots/$NEWSNAPSHOTS/@

Populate links:

# NEWLINKS="$(date -u +"%Y%m%d%H%M%S")$(cat /dev/urandom | tr -dc 'a-zA-Z' | fold -w 8 | head -n 1)"
# mkdir "/mnt/links/$NEWLINKS"
# ln -s "../../snapshots/$NEWSNAPSHOTS" "/mnt/links/$NEWLINKS/0"
# ln -s "../../snapshots/$NEWSNAPSHOTS" "/mnt/links/$NEWLINKS/1"
# ln -s "../../snapshots/$NEWSNAPSHOTS" "/mnt/links/$NEWLINKS/2"
# ln -s "../../snapshots/$NEWSNAPSHOTS" "/mnt/links/$NEWLINKS/3"

You can have as many links as you like, just apply changes to rEFInd config and upgrade scripts described below accordingly.
Link that will point to latest links generation:

# ln -s "./links/$NEWLINKS" /mnt/current

This setup allows us to just have static rEFInd config that points to to /current/0/@, /current/1/@, etc. while the actual underlying boot environment will change with each upgrade.
But how will fs mounting services know which snapshot generation is currently loaded?
The answer is common fstab in the btrfs root.
Get UUIDs of the partitions first:

# blkid > /mnt/fstab

Now edit fstab accordingly:

# vi /mnt/fstab

Example:

UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 / btrfs subvol=CURRENT_SNAPSHOTS_PATH/@,ro,noatime 0 0
UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 /var btrfs subvol=/commons/@var,rw,noatime 0 0
UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 /home btrfs subvol=/commons/@home,rw,noatime 0 0
# UUID=2FE6-837A /boot/efi vfat rw,noatime,discard 0 2
tmpfs /tmp tmpfs mode=1777,noatime,nosuid,nodev,size=2G 0 0
UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 swap swap rw,noatime,discard 0 0

CURRENT_SNAPSHOTS_PATH will be replaced by scripts with, for example, /snapshots/20210411212549sdBXyLxg, and the result will be piped into /etc/fstab of a created @ snapshot during new generation preparations.
Root btrfs volume structure mounted on /mnt:

|--mnt
| |--commons
| | |--@var
| | |--@home
| |--current
| |--fstab
| |--links
| | |--20210411213742qwrXAJBz
| | | |--0
| | | |--1
| | | |--2
| | | |--3
| |--next
| |--snapshots
| | |--20210411212549sdBXyLxg
| | | |--@

Base system install

With the directory structure prepared, we can begin installing a basic Alpine Linux system.
Considering that installation is done from Alpine system, we only need following parts of the process:

# apk -X https://dl-cdn.alpinelinux.org/alpine/latest-stable/main -U --allow-untrusted -p /mnt/snapshots/20210411212549sdBXyLxg/@ --initdb add alpine-base

Now we can setup basic chroot to complete the installation process:

# export SNP="/mnt/snapshots/20210411212549sdBXyLxg/@"

# mount -o bind /dev $SNP/dev
# mount -t proc none $SNP/proc
# mount -t sysfs sys $SNP/sys

# sed "s#CURRENT_SNAPSHOTS_PATH#/snapshots/20210411212549sdBXyLxg#g" /mnt/fstab > "$SNP/etc/fstab"

# cp -L /etc/resolv.conf "$SNP/etc/"
# chroot "$SNP" /bin/sh

# mount -a

# mv /etc/resolv.conf /tmp/
# ln -s /tmp/resolv.conf /etc/resolv.conf

As soon as you're in chroot, define repositories:

# echo "https://dl-cdn.alpinelinux.org/alpine/latest-stable/main" > /etc/apk/repositories

This example shows only main, but you should also add testing and community if you need any packages in those.
Now it's time for the firmware, kernel, and btrfs packages:

# apk add -U linux-firmware linux-lts btrfs-progs

You may want to change linux-firmware to a custom set of firmware packages suitable for your system, for example, linux-firmware-amd linux-firmware-amd-ucode linux-firmware-amdgpu linux-firmware-ath10k linux-firmware-qca for a typical AMD laptop.
It's also important to add the btrfs feature to mkinitfs.conf and run mkinitfs manually:

# vi /etc/mkinitfs/mkinitfs.conf
# mkinitfs

These steps prepare the kernel and generate the initramfs, which will be used later to boot from our first snapshot.
After that, you should install any package you may need on first boot.

Warning: In case your PC only has wireless connection you should also install any suitable networking software, like iwd in this example, so you will not end up severed from network on your first boot.


Note: Due to root being immutable during operation, it's recommended to install the openresolv package to support changing the network connection. In this case, /etc/resolvconf.conf should have resolv_conf=/tmp/resolv.conf, /etc/resolv.conf should be moved to /tmp/resolv.conf, and a link to the new resolv.conf location should be created: ln -sfn /tmp/resolv.conf /etc/resolv.conf.

You may also use static DNS, but this would make your network activity directly identifiable to the DNS server provider, therefore it's not recommended.

Now, configure the system, start with setting a password for the root:

# passwd root

Don't forget to add essential services to their respective runlevels:

rc-update add devfs sysinit
rc-update add dmesg sysinit
rc-update add mdev sysinit
rc-update add hwdrivers sysinit

rc-update add hwclock boot
rc-update add modules boot
rc-update add sysctl boot
rc-update add hostname boot
rc-update add bootmisc boot
rc-update add syslog boot

rc-update add mount-ro shutdown
rc-update add killprocs shutdown
rc-update add savecache shutdown

With the snapshot prepared and configured, we can chroot out of it and unmount everything:

# umount -a
# exit

Finish editing the snapshot by setting the ro flag and unmounting the root volume:

# btrfs property set -ts "/mnt/snapshots/20210411212549sdBXyLxg/@" ro true
# umount /mnt

Bootloader installation

Mount the EFI partition:

# mount -t vfat /dev/sda1 /mnt
# mkdir /mnt/EFI

Bootloader configuration

There are 2 options as examples of the bootloader installation: rEFInd and GRUB. Sometimes one of them will refuse to work on a system for no particular reason, in this case try the other one.

rEFInd

Check the latest version number of the refind package:

# apk info -X https://dl-cdn.alpinelinux.org/alpine/edge/testing -U refind

Download the latest version (replace 0.13.2-r3 in the example below) of the refind package:

# wget https://dl-cdn.alpinelinux.org/alpine/edge/testing/x86_64/refind-0.13.2-r3.apk

Unpack the prepared rEFInd archive and copy relevant files to /mnt/EFI/

# tar -xzf refind-0.13.2-r3.apk
# cp -r usr/share/refind /mnt/EFI/
# cd /mnt/EFI/refind

Rename config file and edit it:

# mv refind.conf-sample refind.conf
# vi refind.conf

And append following to the end of the file, remember to replace example UUIDs with your own for root (btrfs partition) and resume (swap partition). Keep in mind that if you named the btrfs volume other than ROOT during "Partitioning disks" stage, you have to change the volume field below accordingly.

menuentry "Alpine Linux" {
    icon /EFI/refind/icons/os_linux.png
    volume "ROOT"
    loader /current/0/@/boot/vmlinuz-lts
    initrd /current/0/@/boot/initramfs-lts
    options "root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/0/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash"

    submenuentry "Boot fallback 1" {
        loader /current/1/@/boot/vmlinuz-lts
        initrd /current/1/@/boot/initramfs-lts
        options "root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/1/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash"
    }

    submenuentry "Boot fallback 2" {
        loader /current/2/@/boot/vmlinuz-lts
        initrd /current/2/@/boot/initramfs-lts
        options "root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/2/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash"
    }

    submenuentry "Boot fallback 3" {
        loader /current/3/@/boot/vmlinuz-lts
        initrd /current/3/@/boot/initramfs-lts
        options "root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/3/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash"
    }
}
Note: "ROOT" is the PARTLABEL of the btrfs partition. You may also use PARTUUID instead. To get both blkid from blkid package can be used. blkid included in busybox does not provide this information.

GRUB

# apk add grub-efi

GRUB requires two configuration files this time as we will use grub-mkstandalone. The first configuration file is internal and should only point to the second file, where we store the menu:

# cd /tmp
# vi grub_internal.cfg

Set the contents to the following, but make sure to replace 2FE6-837A with your own EFI partition UUID:

insmod part_gpt
insmod fat
search --set efi --fs-uuid 2FE6-837A
configfile (${efi})/EFI/grub/grub.cfg

The second config file is the main config where we describe the entire boot menu.

# vi grub.cfg

Set to contain, but replace UUIDs:

set timeout=3
menuentry "Alpine Linux Current" {
	search --set root --fs-uuid b9ff5e7b-e128-4e64-861a-2fdd794a9828
	linux /current/0/@/boot/vmlinuz-edge root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/0/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash
	initrd /current/0/@/boot/initramfs-edge
}
menuentry "Alpine Linux Snapshot 1" {
	search --set root --fs-uuid b9ff5e7b-e128-4e64-861a-2fdd794a9828
	linux /current/1/@/boot/vmlinuz-edge root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/1/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash
	initrd /current/1/@/boot/initramfs-edge
}
menuentry "Alpine Linux Snapshot 2" {
	search --set root --fs-uuid b9ff5e7b-e128-4e64-861a-2fdd794a9828
	linux /current/2/@/boot/vmlinuz-edge root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/2/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash
	initrd /current/2/@/boot/initramfs-edge
}
menuentry "Alpine Linux Snapshot 3" {
	search --set root --fs-uuid b9ff5e7b-e128-4e64-861a-2fdd794a9828
	linux /current/3/@/boot/vmlinuz-edge root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/3/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 quiet splash
	initrd /current/3/@/boot/initramfs-edge
}

Generate the grubx64.efi binary:

# grub-mkstandalone -O x86_64-efi -o grubx64.efi "boot/grub/grub.cfg=/tmp/grub_internal.cfg"
# mkdir /mnt/EFI/grub
# mv grubx64.efi /mnt/EFI/grub/
# mv grub.cfg /mnt/EFI/grub/

Adding EFI boot entry

To add the chosen bootloader to UEFI, efibootmgr is a suitable tool. The following example is for rEFInd, but could be easily adjusted for GRUB:

# apk add efibootmgr
# efibootmgr --create --disk /dev/sda --part 1 --loader /EFI/refind/refind_x64.efi --label "rEFInd" --verbose

/dev/sda is our disk device and 1 is the number of the FAT32 partition containing the bootloader data.

Updating or altering the system

Warning: Without the following step or an alternative you will have no easy way to mutate the installed system.


Warning: These examples are implemented using execline and require the execline package in the system.


Note: These could surely be implemented in POSIX shell, however, execline provides a number of runtime advantages and the resulting script is much more readable.
# touch /usr/sbin/sysmut
# chmod +x /usr/sbin/sysmut
# vi /usr/sbin/sysmut

Example script to mutate the the system:

#!/bin/execlineb -W
unshare --mount
importas -D 0 source 1
define mnt /media/root
if { mkdir -p ${mnt} }
if { mount -t btrfs -o rw,noatime UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 ${mnt} }
foreground {
	backtick -E dt {
		date -u +%Y%m%d%H%M%S
	}
	backtick -E rnd {
		pipeline { cat /dev/urandom }
		pipeline { tr -dc a-zA-Z }
		pipeline { fold -w 8 }
		head -n 1
	}
	define newsnap ${dt}${rnd}
	if { mkdir -p ${mnt}/snapshots/${newsnap} }
	if { btrfs subvolume snapshot ${mnt}/current/${source}/@ ${mnt}/snapshots/${newsnap}/@ }
	if {
		redirfd -w 1 ${mnt}/snapshots/${newsnap}/@/etc/fstab
			sed s#CURRENT_SNAPSHOTS_PATH#/snapshots/${newsnap}#g ${mnt}/fstab
	}
	if { mount -t proc none ${mnt}/snapshots/${newsnap}/@/proc }
	if { mount -t sysfs sys ${mnt}/snapshots/${newsnap}/@/sys }
	if { mount -o bind,ro /dev ${mnt}/snapshots/${newsnap}/@/dev }
	foreground {
		foreground { mount -o bind,ro /etc/resolv.conf ${mnt}/snapshots/${newsnap}/@/etc/resolv.conf }
		foreground {
			chroot ${mnt}/snapshots/${newsnap}/@
			foreground { mount -a }
			foreground { sh }
			importas apply ?
			foreground { umount -a }
			exit ${apply}
		}
		importas apply ?
		foreground { redirfd -w 2 /dev/null umount ${mnt}/snapshots/${newsnap}/@/etc/resolv.conf }
		ifelse { exit ${apply} } {
			if { btrfs property set -ts ${mnt}/snapshots/${newsnap}/@ ro true }
			define newlink ${dt}${rnd}
			if { mkdir -p ${mnt}/links/${newlink} }
			if { ln -s ../../snapshots/${newsnap} ${mnt}/links/${newlink}/0 }
			if { cp -P ${mnt}/current/0 ${mnt}/links/${newlink}/1 }
			if { cp -P ${mnt}/current/1 ${mnt}/links/${newlink}/2 }
			if { cp -P ${mnt}/current/2 ${mnt}/links/${newlink}/3 }
			if { mkdir -p ${mnt}/next }
			if { ln -sfn ./links/${newlink} ${mnt}/next/current }
			if { mv ${mnt}/next/current ${mnt}/ }
			echo "Changes applied"
		}
		echo "Changes discarded"
	}
	foreground { redirfd -w 2 /dev/null umount ${mnt}/snapshots/${newsnap}/@/proc }
	foreground { redirfd -w 2 /dev/null umount ${mnt}/snapshots/${newsnap}/@/sys }
	redirfd -w 2 /dev/null umount ${mnt}/snapshots/${newsnap}/@/dev
}
umount ${mnt}

It will get you into the root shell chrooted into the new snapshot, where you can apply any change you like. The origin of the new snapshot is defined by the first and only argument, in form of number. If no argument provided the 0 (current latest) is taken as origin.
If chroot shell exits with an error, there will be no switch to the new snapshots. This means you can manually discard changes while in the chroot by:

# exit 1

Deleting unused snapshots

Unused snapshots can be garbage-collected by:

# touch /usr/sbin/syscln
# chmod +x /usr/sbin/syscln
# vi /usr/sbin/syscln
#!/bin/execlineb -W
unshare --mount
define mnt /media/root
if { mkdir -p ${mnt} }
if { mount -t btrfs -o rw,noatime,compress=zstd:3 UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 ${mnt} }
foreground {
	foreground {
		pipeline {
			foreground {
				pipeline {
					find -H ${mnt}/snapshots/ -maxdepth 1 -mindepth 1 -print0
				}
				xargs -0 -r realpath
			}
			pipeline {
				find -H ${mnt}/current/ -maxdepth 1 -mindepth 1 -print0
			}
			xargs -0 -r realpath
		}
		pipeline { tr \\n \\0 }
		pipeline { sort -z }
		pipeline { uniq -u -z }
		pipeline { xargs -0 -r -n 1 -I [] find -H [] -maxdepth 1 -mindepth 1 -print0 }
		xargs -0 -r btrfs subvolume delete
	}
	foreground { find -H ${mnt}/snapshots/ -maxdepth 1 -mindepth 1 -empty -type d -delete }
	foreground {
		pipeline {
			foreground {
				pipeline {
					find -H ${mnt}/links/ -maxdepth 1 -mindepth 1 -print0
				}
				xargs -0 -r realpath
			}
			realpath ${mnt}/current
		}
		pipeline { tr \\n \\0 }
		pipeline { sort -z }
		pipeline { uniq -u -z }
		pipeline { xargs -0 -r -n 1 -I [] find -H [] -maxdepth 1 -mindepth 1 -print0 }
		xargs -0 -r -n 1 unlink
	}
	find -H ${mnt}/links/ -maxdepth 1 -mindepth 1 -empty -type d -delete
}
umount ${mnt}

Allowing temporary runtime alterations

You can use overlayfs with tmpfs built into Alpine's init script to allow changes in the rootfs which will be automatically reverted upon reboot.
To make use of this, add overlaytmpfs to the kernel boot options in refind.conf, e.g.:

...
    initrd /current/0/@/boot/initramfs-lts
    options "root=UUID=b9ff5e7b-e128-4e64-861a-2fdd794a9828 rootfstype=btrfs rootflags=subvol=/current/0/@,ro,noatime resume=UUID=f0239163-9d46-47c1-67a4-3ee1d63d0676 overlaytmpfs quiet splash"

    submenuentry "Boot fallback 1" {
...