Setting up apt-cacher-ng Install apt-cacher-ng:

sudo apt install apt-cacher-ng

A pop-up will appear, if you are unsure how to answer it select no, we don t need it for this use-case. To enable apt-cacher-ng on your system, create /etc/apt/apt.conf.d/02proxy and insert:

Acquire::http::proxy "http://127.0.0.1:3142";
Acquire::https::proxy "DIRECT";

In /etc/apt-cacher-ng/acng.conf you can increase the value of ExThreshold to hold packages for a shorter or longer duration. The length depends on your specific use case and resources. A longer threshold takes more disk space, a short threshold like one day effecitvely only reduces the build time for rebuilds. If you encounter weird issues on apt update at some point the future, you can try to clean the cache from apt-cacher-ng. You can use this script:

Setting up mmdebstrap Install mmdebstrap:

sudo apt install mmdebstrap

We will create a small helper script to ease creating a chroot. Open ~/.local/bin/mmupdate and insert:

#!/bin/sh
mmdebstrap \
  --variant=buildd \
  --aptopt='Acquire::http::proxy "http://127.0.0.1:3142";' \
  --arch=amd64 \
  --components=main,contrib,non-free \
  unstable \
  ~/.cache/sbuild/unstable-amd64.tar.xz \
  http://deb.debian.org/debian

Notes:

• aptopt enables apt-cacher-ng inside the chroot.
• --arch sets the CPU architecture (see Debian Wiki).
• --components sets the archive components, if you don t want non-free pacakges you might want to remove some entries here.
• unstable sets the Debian release, you can also set for example bookworm-backports here.
• unstable-amd64.tar.xz is the output tarball containing the chroot, change accordingly to your pick of the CPU architecture and Debian release.
• http://deb.debian.org/debian is the Debian mirror, you should set this to the same one you use in your /etc.apt/sources.list.

Make mmupdate executable and run it once:

chmod +x ~/.local/bin/mmupdate
mkdir -p ~/.cache/sbuild
~/.local/bin/mmupdate

If you execute mmupdate again you can see that the downloading stage is much faster thanks to apt-cacher-ng. For me the difference is from about 115s to about 95s. Your results may vary, this depends on the speed of your internet, Debian mirror and disk. If you have used the schroot backend and sbuild-update before, you probably notice that creating a new chroot with mmdebstrap is slower. It would be a bit annoying to do this manually before we start a new Debian packaging session, so let s create a systemd service that does this for us. First create a folder for user services:

mkdir -p ~/.config/systemd/user

Create ~/.config/systemd/user/mmupdate.service and add:

[Unit]
Description=Run mmupdate
Wants=network-online.target
[Service]
Type=oneshot
ExecStart=%h/.local/bin/mmupdate

Start the service and test that it works:

systemctl --user daemon-reload
systemctl --user start mmupdate
systemctl --user status mmupdate

Create ~/.config/systemd/user/mmupdate.timer:

[Unit]
Description=Run mmupdate daily
[Timer]
OnCalendar=daily
Persistent=true
[Install]
WantedBy=timers.target

Enable the timer:

systemctl --user enable mmupdate.timer

Now every day mmupdte will be run automatically. You can adjust the period if you think daily rebuilds are a bit excessive. A neat advantage of period rebuilds is that they the base files in your apt-cacher-ng cache warm every time they run.

Setting up sbuild: Install sbuild and (optionally) autopkgtest:

sudo apt install --no-install-recommends sbuild autopkgtest

Create ~/.sbuildrc and insert:

# backend for using mmdebstrap chroots
$chroot_mode = 'unshare';
# build in tmpfs
$unshare_tmpdir_template = '/dev/shm/tmp.sbuild.XXXXXXXX';
# upgrade before starting build
$apt_update = 1;
$apt_upgrade = 1;
# build everything including source for source-only uploads
$build_arch_all = 1;
$build_arch_any = 1;
$build_source = 1;
$source_only_changes = 1;
# go to shell on failure instead of exiting
$external_commands =   "build-failed-commands" => [ [ '%SBUILD_SHELL' ] ]  ;
# always clean build dir, even on failure
$purge_build_directory = "always";
# run lintian
$run_lintian = 1;
$lintian_opts = [ '-i', '-I', '-E', '--pedantic' ];
# do not run piuparts
$run_piuparts = 0;
# run autopkgtest
$run_autopkgtest = 1;
$autopkgtest_root_args = '';
$autopkgtest_opts = [ '--apt-upgrade', '--', 'unshare', '--release', '%r', '--arch', '%a', '--prefix=/dev/shm/tmp.autopkgtest.' ];
# set uploader for correct signing
$uploader_name = 'Stephan Lachnit <stephanlachnit@debian.org>';

You should adjust uploader_name. If you don t want to run autopkgtest or lintian by default you can also disable it here. Note that for packages that need a lot of space for building, you might want to comment the unshare_tmpdir_template line to prevent a OOM build failure. You can now build your Debian packages with the sbuild command :)

Finishing touches You can add these variables to your ~/.bashrc as bonus (with adjusted name / email):

export DEBFULLNAME="<your_name>"
export DEBEMAIL="<your_email>"
export DEB_BUILD_OPTIONS="parallel=<threads>"

In particular adjust the value of parallel to ensure parallel builds. If you are new to signing / uploading your package, first install the required tools:

sudo apt install devscripts dput-ng

Create ~/.devscripts and insert:

DEBSIGN_KEYID=<your_gpg_fingerpring>
USCAN_SYMLINK=rename

You can now sign the .changes file with:

debsign ../<pkgname_version_arch>.changes

And for source-only uploads with:

debsign -S ../<pkgname_version_arch>_source.changes

If you don t introduce a new binary package, you always want to go with source-only changes. You can now upload the package to Debian with

dput ../<filename>.changes

Update Feburary 22nd Jochen Sprickerhof, who originally advised me to use the unshare backend, commented that one can also use --include=auto-apt-proxy instead of the --aptopt option in mmdebstrap to detect apt proxies automatically. He also let me know that it is possible to use autopkgtest on tmpfs (config in the blog post is updated) and added an entry on the sbuild wiki page on how to setup sbuild+unshare with ccache if you often need to build a large package. Further, using --variant=apt and --include=build-essential will produce smaller build chroots if wished. On the contrary, one can of course also use the --include option to include debhelper and lintian (or any other packages you like) to further decrease the setup time. However, staying with buildd variant is a good choice for official uploads.

Resources for further reading https://wiki.debian.org/sbuild
https://www.unix-ag.uni-kl.de/~bloch/acng/html/index.html
https://wiki.ubuntu.com/SimpleSbuild
https://wiki.archlinux.org/title/Systemd/Timers
https://manpages.debian.org/unstable/autopkgtest/autopkgtest-virt-unshare.1.en.html
Thanks for reading!

1. Creating the partition table
2. Creating and mounting the root partition
3. Bootstrapping with debootstrap
4. Chrooting into the system
5. Configure the base system
6. Define static file system information
7. Installing the kernel and bootloader
8. Finishing touches

Creating the partition partble The first step is to create the GPT partition table on the new drive. There are several tools to do this, I recommend the ArchWiki page on this topic for details. For simplicity I just went with the GParted since it has an easy GUI, but feel free to use any other tool. The layout should look like this:

Type         Partition        Suggested size
 
EFI          /dev/nvme0n1p1           512MiB
Linux swap   /dev/nvme0n1p2             1GiB
Linux fs     /dev/nvme0n1p3        remainder

Notes:

• The disk names are just an example and have to be adjusted for your system.
• Don t set disk labels, they don t appear on the new install anyway and some UEFIs might not like it on your boot partition.
• The size of the EFI partition can be smaller, in practive it s unlikely that you need more than 300 MiB. However some UEFIs might be buggy and if you ever want to install an additional kernel or something like memtest86+ you will be happy to have the extra space.
• The swap partition can be omitted, it is not strictly needed. If you need more swap for some reason you can also add more using a swap file later (see ArchWiki page). If you know you want to use suspend-to-RAM, you want to increase the size to something more than the size of your memory.
• If you used GParted, create the EFI partition as FAT32 and set the esp flag. For the root partition use ext4 or F2FS if available.

Creating and mounting the root partition To create the root partition, we need to install the f2fs-tools first:

sudo apt install f2fs-tools

Now we can create the file system with the correct flags:

mkfs.f2fs -O extra_attr,inode_checksum,sb_checksum,compression,encrypt /dev/nvme0n1p3

For details on the flags visit the ArchWiki page. Next, we need to mount the partition with the correct flags. First, create a working directory:

mkdir boostrap
cd boostrap
mkdir root
export DFS=$(pwd)/root

Then we can mount the partition:

sudo mount -o compress_algorithm=zstd:6,compress_chksum,atgc,gc_merge,lazytime /dev/nvme0n1p3 $DFS

Again, for details on the mount options visit the above mentioned ArchWiki page.

Bootstrapping with debootstrap First we need to install the debootstrap package:

sudo apt install debootstrap

Now we can do the bootstrapping:

debootstrap --arch=amd64 --components=main,contrib,non-free,non-free-firmware unstable $DFS http://deb.debian.org/debian

Notes:

• --arch sets the CPU architecture (see Debian Wiki).
• --components sets the archive components, if you don t want non-free pacakges you might want to remove some entries here.
• unstable is the Debian release, you might want to change that to testing or bookworm.
• $DFS points to the mounting point of the root partition.
• http://deb.debian.org/debian is the Debian mirror, you might want to set that to http://ftp.de.debian.org/debian or similar if you have a fast mirror in you area.

Chrooting into the system Before we can chroot into the newly created system, we need to prepare and mount virtual kernel file systems. First create the directories:

sudo mkdir -p $DFS/dev $DFS/dev/pts $DFS/proc $DFS/sys $DFS/run $DFS/sys/firmware/efi/efivars $DFS/boot/efi

Then bind-mount the directories from your system to the mount point of the new system:

sudo mount -v -B /dev $DFS/dev
sudo mount -v -B /dev/pts $DFS/dev/pts
sudo mount -v -B /proc $DFS/proc
sudo mount -v -B /sys $DFS/sys
sudo mount -v -B /run $DFS/run
sudo mount -v -B /sys/firmware/efi/efivars $DFS/sys/firmware/efi/efivars

As a last step, we need to mount the EFI partition:

sudo mount -v -B /dev/nvme0n1p1 $DFS/boot/efi

Now we can chroot into new system:

sudo chroot $DFS /bin/bash

Configure the base system The first step in the chroot is setting the locales. We need this since we might leak the locales from our base system into the chroot and if this happens we get a lot of annoying warnings.

export LC_ALL=C.UTF-8 LANG=C.UTF-8
apt install locales console-setup

Set your locales:

dpkg-reconfigure locales

Set your keyboard layout:

dpkg-reconfigure keyboard-configuration

Set your timezone:

dpkg-reconfigure tzdata

Now you have a fully functional Debian chroot! However, it is not bootable yet, so let s fix that.

Define static file system information The first step is to make sure the system mounts all partitions on startup with the correct mount flags. This is done in /etc/fstab (see ArchWiki page). Open the file and change its content to:

# file system                               mount point   type   options                                                            dump   pass
# NVME efi partition
UUID=XXXX-XXXX                              /boot/efi     vfat   umask=0077                                                         0      0
# NVME swap
UUID=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX   none          swap   sw                                                                 0      0
# NVME main partition
UUID=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX   /             f2fs   compress_algorithm=zstd:6,compress_chksum,atgc,gc_merge,lazytime   0      1

You need to fill in the UUIDs for the partitions. You can use

ls -lAph /dev/disk/by-uuid/

to match the UUIDs to the more readable disk name under /dev.

Installing the kernel and bootloader First install the systemd-boot and efibootmgr packages:

apt install systemd-boot efibootmgr

Now we can install the bootloader:

bootctl install --path=/boot/efi

You can verify the procedure worked with

efibootmgr -v

The next step is to install the kernel, you can find a fitting image with:

apt search linux-image-*

In my case:

apt install linux-image-amd64

After the installation of the kernel, apt will add an entry for systemd-boot automatically. Neat! However, since we are in a chroot the current settings are not bootable. The first reason is the boot partition, which will likely be the one from your current system. To change that, navigate to /boot/efi/loader/entries, it should contain one config file. When you open this file, it should look something like this:

title      Debian GNU/Linux bookworm/sid
version    6.1.0-3-amd64
machine-id 2967cafb6420ce7a2b99030163e2ee6a
sort-key   debian
options    root=PARTUUID=f81d4fae-7dec-11d0-a765-00a0c91e6bf6 ro systemd.machine_id=2967cafb6420ce7a2b99030163e2ee6a
linux      /2967cafb6420ce7a2b99030163e2ee6a/6.1.0-3-amd64/linux
initrd     /2967cafb6420ce7a2b99030163e2ee6a/6.1.0-3-amd64/initrd.img-6.1.0-3-amd64

The PARTUUID needs to point to the partition equivalent to /dev/nvme0n1p3 on your system. You can use

ls -lAph /dev/disk/by-partuuid/

to match the PARTUUIDs to the more readable disk name under /dev. The second problem is the ro flag in options which tell the kernel to boot in read-only mode. The default is rw, so you can just remove the ro flag. Once this is fixed, the new system should be bootable. You can change the boot order with:

efibootmgr --bootorder

However, before we reboot we might add well add a user and install some basic software.

Finishing touches Add a user:

useradd -m -G sudo -s /usr/bin/bash -c 'Full Name' username

Debian provides a TUI to install Desktop Environment. To open it, run:

tasksel

Now you can finally reboot into your new system:

reboot

Resources for further reading https://ivanb.neocities.org/blogs/y2022/debootstrap
https://www.debian.org/releases/stable/amd64/apds03.en.html
https://www.addictivetips.com/ubuntu-linux-tips/set-up-systemd-boot-on-arch-linux/
https://p5r.uk/blog/2020/using-systemd-boot-on-debian-bullseye.html
https://www.linuxfromscratch.org/lfs/view/stable/chapter07/kernfs.html
Thanks for reading!

apt-get install libvirt-daemon-system virtinst genisoimage cloud-image-utils osinfo-db-tools

mkdir -p /root/iso
cd /root/iso
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_9.0.2_amd64.iso
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_9.0.2_amd64.iso.asc
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_9.0.2_amd64.iso.sha256
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_10.0.1_amd64.iso
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_10.0.1_amd64.iso.asc
wget -q https://mirror.fsf.org/trisquel-images/trisquel-netinst_10.0.1_amd64.iso.sha256
wget -q -O- https://archive.trisquel.info/trisquel/trisquel-archive-signkey.gpg   gpg --import
sha256sum -c trisquel-netinst_9.0.2_amd64.iso.sha256
gpg --verify trisquel-netinst_9.0.2_amd64.iso.asc
sha256sum -c trisquel-netinst_10.0.1_amd64.iso.sha256
gpg --verify trisquel-netinst_10.0.1_amd64.iso.asc
wget -q https://cdbuilds.trisquel.org/aramo/trisquel-netinst_11.0-20221225_amd64.iso
echo '179566639ca8f14f0c3d5658209c59a0916d9e3bf9c026660cc07b28f2311631  trisquel-netinst_11.0-20221225_amd64.iso'   sha256sum -c

root@trana:~# cat>trisquel.preseed 
d-i debian-installer/locale select en_US
d-i keyboard-configuration/xkb-keymap select us
d-i netcfg/choose_interface select auto
d-i netcfg/disable_autoconfig boolean true
d-i netcfg/get_ipaddress string 192.168.10.201
d-i netcfg/get_netmask string 255.255.255.0
d-i netcfg/get_gateway string 192.168.10.46
d-i netcfg/get_nameservers string 192.168.10.46
d-i netcfg/get_hostname string trisquel
d-i netcfg/get_domain string sjd.se
d-i clock-setup/utc boolean true
d-i time/zone string UTC
d-i mirror/country string manual
d-i mirror/http/hostname string ftp.acc.umu.se
d-i mirror/http/directory string /mirror/trisquel/packages
d-i mirror/http/proxy string
d-i partman-auto/method string regular
d-i partman-partitioning/confirm_write_new_label boolean true
d-i partman/choose_partition select finish
d-i partman/confirm boolean true
d-i partman/confirm_nooverwrite boolean true
d-i partman-basicfilesystems/no_swap boolean false
d-i partman-auto/expert_recipe string myroot :: 1000 50 -1 ext4 \
     $primary    $bootable    method  format   \
     format    use_filesystem    filesystem  ext4   \
     mountpoint  /   \
    .
d-i partman-auto/choose_recipe select myroot
d-i passwd/root-login boolean true
d-i user-setup/allow-password-weak boolean true
d-i passwd/root-password password r00tme
d-i passwd/root-password-again password r00tme
d-i passwd/make-user boolean false
tasksel tasksel/first multiselect
d-i pkgsel/include string openssh-server
popularity-contest popularity-contest/participate boolean false
d-i grub-installer/only_debian boolean true
d-i grub-installer/with_other_os boolean true
d-i grub-installer/bootdev string default
d-i finish-install/reboot_in_progress note
d-i preseed/late_command string mkdir /target/root/.ssh ; echo ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAILzCFcHHrKzVSPDDarZPYqn89H5TPaxwcORgRg+4DagE cardno:FFFE67252015 > /target/root/.ssh/authorized_keys
^D
root@trana:~# 

root@trana:~# HOST=foo
root@trana:~# IP=192.168.10.197
root@trana:~# sed -e "s,get_ipaddress string.*,get_ipaddress string $IP," -e "s,get_hostname string.*,get_hostname string $HOST," < trisquel.preseed > vm-$HOST.preseed
root@trana:~# 

root@trana:~# cat gen-preseed-iso 
#!/bin/sh
# Copyright (C) 2018-2022 Simon Josefsson -- GPLv3+
# https://wiki.debian.org/DebianInstaller/Preseed/EditIso
# https://trisquel.info/en/wiki/customizing-trisquel-iso
set -e
set -x
ISO="$1"
PRESEED="$2"
OUTISO="$3"
LASTPWD="$PWD"
test -f "$ISO"
test -f "$PRESEED"
test ! -f "$OUTISO"
TMPDIR=$(mktemp -d)
mkdir "$TMPDIR/mnt"
mkdir "$TMPDIR/tmp"
cp "$PRESEED" "$TMPDIR"/preseed.cfg
cd "$TMPDIR"
mount "$ISO" mnt/
cp -rT mnt/ tmp/
umount mnt/
chmod +w -R tmp/
gunzip tmp/initrd.gz
echo preseed.cfg   cpio -H newc -o -A -F tmp/initrd
gzip tmp/initrd
chmod -w -R tmp/
sed -i "s/timeout 0/timeout 1/" tmp/isolinux.cfg
sed -i "s/default vesamenu.c32/default auto/" tmp/isolinux.cfg
if ! grep -q auto tmp/adtxt.cfg; then
    cat<<EOF >> tmp/adtxt.cfg
label auto
	menu label ^Automated install
	kernel linux
	append auto=true priority=critical vga=788 initrd=initrd.gz --- quiet
EOF
fi
cd tmp/
find -follow -type f   xargs md5sum  > md5sum.txt
cd ..
cd "$LASTPWD"
genisoimage -r -J -b isolinux.bin -c boot.cat \
            -no-emul-boot -boot-load-size 4 -boot-info-table \
            -o "$OUTISO" "$TMPDIR/tmp/"
rm -rf "$TMPDIR"
exit 0
^D
root@trana:~# chmod +x gen-preseed-iso 
root@trana:~# 

root@trana:~# ./gen-preseed-iso /root/iso/trisquel-netinst_10.0.1_amd64.iso vm-$HOST.preseed vm-$HOST.iso
+ ISO=/root/iso/trisquel-netinst_10.0.1_amd64.iso
+ PRESEED=vm-foo.preseed
+ OUTISO=vm-foo.iso
+ LASTPWD=/root
+ test -f /root/iso/trisquel-netinst_10.0.1_amd64.iso
+ test -f vm-foo.preseed
+ test ! -f vm-foo.iso
+ mktemp -d
+ TMPDIR=/tmp/tmp.mNEprT4Tx9
+ mkdir /tmp/tmp.mNEprT4Tx9/mnt
+ mkdir /tmp/tmp.mNEprT4Tx9/tmp
+ cp vm-foo.preseed /tmp/tmp.mNEprT4Tx9/preseed.cfg
+ cd /tmp/tmp.mNEprT4Tx9
+ mount /root/iso/trisquel-netinst_10.0.1_amd64.iso mnt/
mount: /tmp/tmp.mNEprT4Tx9/mnt: WARNING: source write-protected, mounted read-only.
+ cp -rT mnt/ tmp/
+ umount mnt/
+ chmod +w -R tmp/
+ gunzip tmp/initrd.gz
+ echo preseed.cfg
+ cpio -H newc -o -A -F tmp/initrd
5 blocks
+ gzip tmp/initrd
+ chmod -w -R tmp/
+ sed -i s/timeout 0/timeout 1/ tmp/isolinux.cfg
+ sed -i s/default vesamenu.c32/default auto/ tmp/isolinux.cfg
+ grep -q auto tmp/adtxt.cfg
+ cat
+ cd tmp/
+ find -follow -type f
+ xargs md5sum
+ cd ..
+ cd /root
+ genisoimage -r -J -b isolinux.bin -c boot.cat -no-emul-boot -boot-load-size 4 -boot-info-table -o vm-foo.iso /tmp/tmp.mNEprT4Tx9/tmp/
I: -input-charset not specified, using utf-8 (detected in locale settings)
Using GCRY_000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/gcry_sha512.mod (gcry_sha256.mod)
Using XNU_U000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/xnu_uuid.mod (xnu_uuid_test.mod)
Using PASSW000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/password_pbkdf2.mod (password.mod)
Using PART_000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/part_sunpc.mod (part_sun.mod)
Using USBSE000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/usbserial_pl2303.mod (usbserial_ftdi.mod)
Using USBSE001.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/usbserial_ftdi.mod (usbserial_usbdebug.mod)
Using VIDEO000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/videotest.mod (videotest_checksum.mod)
Using GFXTE000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/gfxterm_background.mod (gfxterm_menu.mod)
Using GCRY_001.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/gcry_sha256.mod (gcry_sha1.mod)
Using MULTI000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/multiboot2.mod (multiboot.mod)
Using USBSE002.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/usbserial_usbdebug.mod (usbserial_common.mod)
Using MDRAI000.MOD;1 for  /tmp/tmp.mNEprT4Tx9/tmp/boot/grub/x86_64-efi/mdraid09.mod (mdraid09_be.mod)
Size of boot image is 4 sectors -> No emulation
 22.89% done, estimate finish Thu Dec 29 23:36:18 2022
 45.70% done, estimate finish Thu Dec 29 23:36:18 2022
 68.56% done, estimate finish Thu Dec 29 23:36:18 2022
 91.45% done, estimate finish Thu Dec 29 23:36:18 2022
Total translation table size: 2048
Total rockridge attributes bytes: 24816
Total directory bytes: 40960
Path table size(bytes): 64
Max brk space used 46000
21885 extents written (42 MB)
+ rm -rf /tmp/tmp.mNEprT4Tx9
+ exit 0
root@trana:~#

root@trana:~# virt-install --name $HOST --disk vm-$HOST.img,size=5 --cdrom vm-$HOST.iso --osinfo linux2020 --autostart --noautoconsole --wait
Using linux2020 default --memory 4096
Starting install...
Allocating 'vm-foo.img'                                                                                                                                     0 B  00:00:00 ... 
Creating domain...                                                                                                                                          0 B  00:00:00     
Domain is still running. Installation may be in progress.
Waiting for the installation to complete.
Domain has shutdown. Continuing.
Domain creation completed.
Restarting guest.
root@trana:~# 

root@trana:~# virsh destroy foo
Domain 'foo' destroyed
root@trana:~# virsh undefine foo --remove-all-storage
Domain 'foo' has been undefined
Volume 'vda'(/root/vm-foo.img) removed.
root@trana:~# rm vm-foo.*
root@trana:~# 

CDKTF: infrastructure as code Terraform is an infrastructure-as-code tool. You can define your infrastructure by declaring resources with the HCL language. This language has some additional features like loops to declare several resources from a list, built-in functions you can call in expressions, and string templates. Terraform relies on a large set of providers to manage resources.

Managing servers Here is a short example using the Hetzner Cloud provider to spawn a virtual machine:

variable "hcloud_token"  
  sensitive = true
 
provider "hcloud"  
  token = var.hcloud_token
 
resource "hcloud_server" "web03"  
  name = "web03"
  server_type = "cpx11"
  image = "debian-11"
  datacenter = "nbg1-dc3"
 
resource "hcloud_rdns" "rdns4-web03"  
  server_id = hcloud_server.web03.id
  ip_address = hcloud_server.web03.ipv4_address
  dns_ptr = "web03.luffy.cx"
 
resource "hcloud_rdns" "rdns6-web03"  
  server_id = hcloud_server.web03.id
  ip_address = hcloud_server.web03.ipv6_address
  dns_ptr = "web03.luffy.cx"
 

HCL expressiveness is quite limited and I find a general-purpose language more convenient to describe all the resources. This is where CDK for Terraform comes in: you can manage your infrastructure using your preferred programming language, including TypeScript, Go, and Python. Here is the previous example using CDKTF and TypeScript:

import   App, TerraformStack, Fn   from "cdktf";
import   HcloudProvider   from "./.gen/providers/hcloud/provider";
import * as hcloud from "./.gen/providers/hcloud";
class MyStack extends TerraformStack  
  constructor(scope: Construct, name: string)  
    super(scope, name);
    const hcloudToken = new TerraformVariable(this, "hcloudToken",  
      type: "string",
      sensitive: true,
     );
    const hcloudProvider = new HcloudProvider(this, "hcloud",  
      token: hcloudToken.value,
     );
    const web03 = new hcloud.server.Server(this, "web03",  
      name: "web03",
      serverType: "cpx11",
      image: "debian-11",
      datacenter: "nbg1-dc3",
      provider: hcloudProvider,
     );
    new hcloud.rdns.Rdns(this, "rdns4-web03",  
      serverId: Fn.tonumber(web03.id),
      ipAddress: web03.ipv4Address,
      dnsPtr: "web03.luffy.cx",
      provider: hcloudProvider,
     );
    new hcloud.rdns.Rdns(this, "rdns6-web03",  
      serverId: Fn.tonumber(web03.id),
      ipAddress: web03.ipv6Address,
      dnsPtr: "web03.luffy.cx",
      provider: hcloudProvider,
     );
   
 
const app = new App();
new MyStack(app, "cdktf-take1");
app.synth();

Running cdktf synth generates a configuration file for Terraform, terraform plan previews the changes, and terraform apply applies them. Now that you have a general-purpose language, you can use functions.

Managing DNS records While using CDKTF for 4 web servers may seem a tad overkill, this is quite different when it comes to managing a few DNS zones. With DNSControl, which is using JavaScript as a domain-specific language, I was able to define the bernat.ch zone with this snippet of code:

D("bernat.ch", REG_NONE, DnsProvider(DNS_BIND, 0), DnsProvider(DNS_GANDI),
  DefaultTTL('2h'),
  FastMailMX('bernat.ch',  subdomains: ['vincent'] ),
  WebServers('@'),
  WebServers('vincent');

This generated 38 records. With CDKTF, I use:

new Route53Zone(this, "bernat.ch", providers.aws)
  .sign(dnsCMK)
  .registrar(providers.gandiVB)
  .www("@", servers)
  .www("vincent", servers)
  .www("media", servers)
  .fastmailMX(["vincent"]);

All the magic is in the code that I did not show you. You can check the dns.ts file in the cdktf-take1 repository to see how it works. Here is a quick explanation:

• Route53Zone() creates a new zone hosted by Route 53,
• sign() signs the zone with the provided master key,
• registrar() registers the zone to the registrar of the domain and sets up DNSSEC,
• www() creates A and AAAA records for the provided name pointing to the web servers,
• fastmailMX() creates the MX records and other support records to direct emails to Fastmail.

Here is the content of the fastmailMX() function. It generates a few records and returns the current zone for chaining:

fastmailMX(subdomains?: string[])  
  (subdomains ?? [])
    .concat(["@", "*"])
    .forEach((subdomain) =>
      this.MX(subdomain, [
        "10 in1-smtp.messagingengine.com.",
        "20 in2-smtp.messagingengine.com.",
      ])
    );
  this.TXT("@", "v=spf1 include:spf.messagingengine.com ~all");
  ["mesmtp", "fm1", "fm2", "fm3"].forEach((dk) =>
    this.CNAME( $ dk ._domainkey ,  $ dk .$ this.name .dkim.fmhosted.com. )
  );
  this.TXT("_dmarc", "v=DMARC1; p=none; sp=none");
  return this;
 

I encourage you to browse the repository if you need more information.

About Pulumi My first tentative around Terraform was to use Pulumi. You can find this attempt on GitHub. This is quite similar to what I currently do with CDKTF. The main difference is that I am using Python instead of TypeScript because I was not familiar with TypeScript at the time.² Pulumi predates CDKTF and it uses a slightly different approach. CDKTF generates a Terraform configuration (in JSON format instead of HCL), delegating planning, state management, and deployment to Terraform. It is therefore bound to the limitations of what can be expressed by Terraform, notably when you need to transform data obtained from one resource to another.³ Pulumi needs specific providers for each resource. Many Pulumi providers are thin wrappers encapsulating Terraform providers. While Pulumi provides a good user experience, I switched to CDKTF because writing providers for Pulumi is a chore. CDKTF does not require such a step. Outside the big players (AWS, Azure and Google Cloud), the existence, quality, and freshness of the Pulumi providers are inconsistent. Most providers rely on a Terraform provider and they may lag a few versions behind, miss a few resources, or have a few bugs of their own. When a provider does not exist, you can write one with the help of the pulumi-terraform-bridge library. The Pulumi project provides a boilerplate for this purpose. I had a bad experience with it when writing providers for Gandi and Vultr: the Makefile automatically installs Pulumi using a curl sh pattern and does not work with /bin/sh. There is a lack of interest for community-based contributions⁴ or even for providers for smaller players.

NixOS & NixOps Nix is a functional, purely-functional programming language. Nix is also the name of the package manager that is built on top of the Nix language. It allows users to declaratively install packages. nixpkgs is a repository of packages. You can install Nix on top of a regular Linux distribution. If you want more details, a good resource is the official website, and notably the learn section. There is a steep learning curve, but the reward is tremendous.

NixOS: declarative Linux distribution NixOS is a Linux distribution built on top of the Nix package manager. Here is a configuration snippet to add some packages:

environment.systemPackages = with pkgs;
  [
    bat
    htop
    liboping
    mg
    mtr
    ncdu
    tmux
  ];

It is possible to alter an existing derivation⁵ to use a different version, enable a specific feature, or apply a patch. Here is how I enable and configure Nginx to disable the stream module, add the Brotli compression module, and add the IP address anonymizer module. Moreover, instead of using OpenSSL 3, I keep using OpenSSL 1.1.⁶

services.nginx =  
  enable = true;
  package = (pkgs.nginxStable.override  
    withStream = false;
    modules = with pkgs.nginxModules; [
      brotli
      ipscrub
    ];
    openssl = pkgs.openssl_1_1;
   );

If you need to add some patches, it is also possible. Here are the patches I added in 2019 to circumvent the DoS vulnerabilities in Nginx until they were fixed in NixOS:⁷

services.nginx.package = pkgs.nginxStable.overrideAttrs (old:  
  patches = oldAttrs.patches ++ [
    # HTTP/2: reject zero length headers with PROTOCOL_ERROR.
    (pkgs.fetchpatch  
      url = https://github.com/nginx/nginx/commit/dbdd[ ].patch;
      sha256 = "a48190[ ]";
     )
    # HTTP/2: limited number of DATA frames.
    (pkgs.fetchpatch  
      url = https://github.com/nginx/nginx/commit/94c5[ ].patch;
      sha256 = "af591a[ ]";
     )
    #  HTTP/2: limited number of PRIORITY frames.
    (pkgs.fetchpatch  
      url = https://github.com/nginx/nginx/commit/39bb[ ].patch;
      sha256 = "1ad8fe[ ]";
     )
  ];
 );

If you are interested, have a look at my relatively small configuration: common.nix contains the configuration to be applied to any host (SSH, users, common software packages), web.nix contains the configuration for the web servers, isso.nix runs Isso into a systemd container.

NixOps: NixOS deployment tool On a single node, NixOS configuration is in the /etc/nixos/configuration.nix file. After modifying it, you have to run nixos-rebuild switch. Nix fetches all possible dependencies from the binary cache and builds the remaining packages. It creates a new entry in the boot loader menu and activates the new configuration. To manage several nodes, there exists several options, including NixOps, deploy-rs, Colmena, and morph. I do not know all of them, but from my point of view, the differences are not that important. It is also possible to build such a tool yourself as Nix provides the most important building blocks: nix build and nix copy. NixOps is one of the first tools available but I encourage you to explore the alternatives. NixOps configuration is written in Nix. Here is a simplified configuration to deploy znc01.luffy.cx, web01.luffy.cx, and web02.luffy.cx, with the help of the server and web functions:

let
  server = hardware: name: imports:  
    deployment.targetHost = "$ name .luffy.cx";
    networking.hostName = name;
    networking.domain = "luffy.cx";
    imports = [ (./hardware/. + "/$ hardware .nix") ] ++ imports;
   ;
  web = hardware: idx: imports:
    server hardware "web$ lib.fixedWidthNumber 2 idx " ([ ./web.nix ] ++ imports);
in  
  network.description = "Luffy infrastructure";
  network.enableRollback = true;
  defaults = import ./common.nix;
  znc01 = server "exoscale" [ ./znc.nix ];
  web01 = web "hetzner" 1 [ ./isso.nix ];
  web02 = web "hetzner" 2 [];
 

Tying everything together with Nix The Nix ecosystem is a unified solution to the various problems around software and configuration management. A very interesting feature is the declarative and reproducible developer environments. This is similar to Python virtual environments, except it is not language-specific.

Brief introduction to Nix flakes I am using flakes, a new Nix feature improving reproducibility by pinning all dependencies and making the build hermetic. While the feature is marked as experimental,⁸ it is widely used and you may see flake.nix and flake.lock at the root of some repositories. As a short example, here is the flake.nix content shipped with Snimpy, an interactive SNMP tool for Python relying on libsmi, a C library:

 
  inputs =  
    nixpkgs.url = "nixpkgs";
    flake-utils.url = "github:numtide/flake-utils";
   ;
  outputs =   self, ...  @inputs:
    inputs.flake-utils.lib.eachDefaultSystem (system:
      let
        pkgs = inputs.nixpkgs.legacyPackages."$ system ";
      in
       
        # nix build
        packages.default = pkgs.python3Packages.buildPythonPackage  
          name = "snimpy";
          src = self;
          preConfigure = ''echo "1.0.0-0-000000000000" > version.txt'';
          checkPhase = "pytest";
          checkInputs = with pkgs.python3Packages; [ pytest mock coverage ];
          propagatedBuildInputs = with pkgs.python3Packages; [ cffi pysnmp ipython ];
          buildInputs = [ pkgs.libsmi ];
         ;
        # nix run + nix shell
        apps.default =   
          type = "app";
          program = "$ self.packages."$ system ".default /bin/snimpy";
         ;
        # nix develop
        devShells.default = pkgs.mkShell  
          name = "snimpy-dev";
          buildInputs = [
            self.packages."$ system ".default.inputDerivation
            pkgs.python3Packages.ipython
          ];
         ;
       );
 

If you have Nix installed on your system:

• nix run github:vincentbernat/snimpy runs Snimpy,
• nix shell github:vincentbernat/snimpy provides a shell with Snimpy ready-to-use,
• nix build github:vincentbernat/snimpy builds the Python package, tests included, and
• nix develop . provides a shell to hack around Snimpy when run from a fresh checkout.⁹

For more information about Nix flakes, have a look at the tutorial from Tweag.

Nix and CDKTF At the root of the repository I use for CDKTF, there is a flake.nix file to set up a shell with Terraform and CDKTF installed and with the appropriate environment variables to automate my infrastructure. Terraform is already packaged in nixpkgs. However, I need to apply a patch on top of the Gandi provider. Not a problem with Nix!

terraform = pkgs.terraform.withPlugins (p: [
  p.aws
  p.hcloud
  p.vultr
  (p.gandi.overrideAttrs
    (old:  
      src = pkgs.fetchFromGitHub  
        owner = "vincentbernat";
        repo = "terraform-provider-gandi";
        rev = "feature/livedns-key";
        hash = "sha256-V16BIjo5/rloQ1xTQrdd0snoq1OPuDh3fQNW7kiv/kQ=";
       ;
     ))
]);

CDKTF is written in TypeScript. I have a package.json file with all the dependencies needed, including the ones to use TypeScript as the language to define infrastructure:

 
  "name": "cdktf-take1",
  "version": "1.0.0",
  "main": "main.js",
  "types": "main.ts",
  "private": true,
  "dependencies":  
    "@types/node": "^14.18.30",
    "cdktf": "^0.13.3",
    "cdktf-cli": "^0.13.3",
    "constructs": "^10.1.151",
    "eslint": "^8.27.0",
    "prettier": "^2.7.1",
    "ts-node": "^10.9.1",
    "typescript": "^3.9.10",
    "typescript-language-server": "^2.1.0"
   
 

I use Yarn to get a yarn.lock file that can be used directly to declare a derivation containing all the dependencies:

nodeEnv = pkgs.mkYarnModules  
  pname = "cdktf-take1-js-modules";
  version = "1.0.0";
  packageJSON = ./package.json;
  yarnLock = ./yarn.lock;
 ;

The next step is to generate the CDKTF providers from the Terraform providers and turn them into a derivation:

cdktfProviders = pkgs.stdenvNoCC.mkDerivation  
  name = "cdktf-providers";
  nativeBuildInputs = [
    pkgs.nodejs
    terraform
  ];
  src = nix-filter  
    root = ./.;
    include = [ ./cdktf.json ./tsconfig.json ];
   ;
  buildPhase = ''
    export HOME=$(mktemp -d)
    export CHECKPOINT_DISABLE=1
    export DISABLE_VERSION_CHECK=1
    export PATH=$ nodeEnv /node_modules/.bin:$PATH
    ln -nsf $ nodeEnv /node_modules node_modules
    # Build all providers we have in terraform
    for provider in $(cd $ terraform /libexec/terraform-providers; echo */*/*/*); do
      version=''$ provider##*/ 
      provider=''$ provider%/* 
      echo "Build $provider@$version"
      cdktf provider add --force-local $provider@$version   cat
    done
    echo "Compile TS   JS"
    tsc
  '';
  installPhase = ''
    mv .gen $out
    ln -nsf $ nodeEnv /node_modules $out/node_modules
  '';
 ;

Finally, we can define the development environment:

devShells.default = pkgs.mkShell  
  name = "cdktf-take1";
  buildInputs = [
    pkgs.nodejs
    pkgs.yarn
    terraform
  ];
  shellHook = ''
    # No telemetry
    export CHECKPOINT_DISABLE=1
    # No autoinstall of plugins
    export CDKTF_DISABLE_PLUGIN_CACHE_ENV=1
    # Do not check version
    export DISABLE_VERSION_CHECK=1
    # Access to node modules
    export PATH=$PWD/node_modules/.bin:$PATH
    ln -nsf $ nodeEnv /node_modules node_modules
    ln -nsf $ cdktfProviders  .gen
    # Credentials
    for p in \
      njf.nznmba.pbz/Nqzvavfgengbe \
      urgmare.pbz/ivaprag@oreang.pu \
      ihyge.pbz/ihyge@ivaprag.oreang.pu; do
        eval $(pass show $(echo $p   tr 'A-Za-z' 'N-ZA-Mn-za-m')   grep '^export')
    done
    eval $(pass show personal/cdktf/secrets   grep '^export')
    export TF_VAR_hcloudToken="$HCLOUD_TOKEN"
    export TF_VAR_vultrApiKey="$VULTR_API_KEY"
    unset VULTR_API_KEY HCLOUD_TOKEN
  '';
 ;

The derivations listed in buildInputs are available in the provided shell. The content of shellHook is sourced when starting the shell. It sets up some symbolic links to make the JavaScript environment built at an earlier step available, as well as the generated CDKTF providers. It also exports all the credentials.¹⁰ I am also using direnv with an .envrc to automatically load the development environment. This also enables the environment to be available from inside Emacs, notably when using lsp-mode to get TypeScript completions. Without direnv, nix develop . can activate the environment. I use the following commands to deploy the infrastructure:¹¹

$ cdktf synth
$ cd cdktf.out/stacks/cdktf-take1
$ terraform plan --out plan
$ terraform apply plan
$ terraform output -json > ~-automation/nixops-take1/cdktf.json

The last command generates a JSON file containing various data to complete the deployment with NixOps.

NixOps The JSON file exported by Terraform contains the list of servers with various attributes:

 
  "hardware": "hetzner",
  "ipv4Address": "5.161.44.145",
  "ipv6Address": "2a01:4ff:f0:b91::1",
  "name": "web05.luffy.cx",
  "tags": [
    "web",
    "continent:NA",
    "continent:SA"
  ]
 

In network.nix, this list is imported and transformed into an attribute set describing the servers. A simplified version looks like this:

let
  lib = inputs.nixpkgs.lib;
  shortName = name: builtins.elemAt (lib.splitString "." name) 0;
  domainName = name: lib.concatStringsSep "." (builtins.tail (lib.splitString "." name));
  server = hardware: name: imports:  
    networking =  
      hostName = shortName name;
      domain = domainName name;
     ;
    deployment.targetHost = name;
    imports = [ (./hardware/. + "/$ hardware .nix") ] ++ imports;
   ;
  cdktf-servers-json = (lib.importJSON ./cdktf.json).servers.value;
  cdktf-servers = map
    (s:
      let
        tags-maybe-import = map (t: ./. + "/$ t .nix") s.tags;
        tags-import = builtins.filter (t: builtins.pathExists t) tags-maybe-import;
      in
       
        name = shortName s.name;
        value = server s.hardware s.name tags-import;
       )
    cdktf-servers-json;
in
 
  // [ ]
  // builtins.listToAttrs cdktf-servers

For web05, this expands to:

web05 =  
  networking =  
    hostName = "web05";
    domainName = "luffy.cx";
   ;
  deployment.targetHost = "web05.luffy.cx";
  imports = [ ./hardware/hetzner.nix ./web.nix ];
 ;

As for CDKTF, at the root of the repository I use for NixOps, there is a flake.nix file to set up a shell with NixOps configured. Because NixOps do not support rollouts, I usually use the following commands to deploy on a single server:¹²

$ nix flake update
$ nixops deploy --include=web04
$ ./tests web04.luffy.cx

If the tests are OK, I deploy the remaining nodes gradually with the following command:

$ (set -e; for h in web 03..06 ; do nixops deploy --include=$h; done)

nixops deploy rolls out all servers in parallel and therefore could cause a short outage where all Nginx are down at the same time.

---

This post has been a work-in-progress for the past three years, with the content being updated and refined as I experimented with different solutions. There is still much to explore¹³ but I feel there is enough content to publish now.

---

1. It was an AMD Athlon 64 X2 5600+ with 2 GB of RAM and 2 400 GB disks with software RAID. I was paying something around 59 per month for it. While it was a good deal in 2008, by 2018 it was no longer cost-effective. It was running on Debian Wheezy with Linux-VServer for isolation, both of which were outdated in 2018.
2. I also did not use Python because Poetry support in Nix was a bit broken around the time I started hacking around CDKTF.
3. Pulumi can apply arbitrary functions with the apply() method on an output. It makes it easy to transform data that are not known during the planning stage. Terraform has functions to serve a similar purpose, but they are more limited.
4. The two mentioned pull requests are not merged yet. The second one is superseded by PR #61, submitted two months later, which enforces the use of /bin/bash. I also submitted PR #56, which was merged 4 months later and quickly reverted without an explanation.
5. You may consider packages and derivations to be synonyms in the Nix ecosystem.
6. OpenSSL 3 has outstanding performance regressions.
7. NixOS can be a bit slow to integrate patches since they need to rebuild parts of the binary cache before releasing the fixes. In this specific case, they were fast: the vulnerability and patches were released on August 13^th 2019 and available in NixOS on August 15th. As a comparison, Debian only released the fixed version on August 22^nd, which is unusually late.
8. Because flakes are experimental, many documentations do not use them and it is an additional aspect to learn.
9. It is possible to replace . with github:vincentbernat/snimpy, like in the other commands, but having Snimpy dependencies without Snimpy source code is less interesting.
10. I am using pass as a password manager. The password names are only obfuscated to avoid spam.
11. The cdktf command can wrap the terraform commands, but I prefer to use them directly as they are more flexible.
12. If the change is risky, I disable the server with CDKTF. This removes it from the web service DNS records.
13. I would like to replace NixOps with an alternative handling progressive rollouts and checks. I am also considering switching to Nomad or Kubernetes to deploy workloads.

1. Debian Installer does not support a proper BTRFS based root file system. It just allows btrfs as root but no subvolume support. Also I'm not sure about the luks support with btrfs as root.
2. I also wanted to give a try to systemd-boot as my laptop is UEFI capable and I've slowly started disliking Grub.
3. I really hate installing task-kde-desktop (Yeah you read it right, I've switched to be a KDE user for quite some time) which will pull tons of unwanted stuff and bloat. Well it's not just task-kde-desktop but any other task-desktop package does similar and I don't want to have too much of unused stuff and services running.

parted -a optimal -s /dev/sda mklabel gpt
parted -a optimal -s /dev/sda mkpart primary vfat 0% 1G
parted -a optimal -s /dev/sda set 1 esp on
mkfs.vfat -n boot_disk -F 32 /dev/sda1

cryptsetup luksFormat /dev/sda2
cryptsetup luksOpen /dev/sda2 ENC
mkfs.btrfs -L root_disk /dev/mapper/ENC

mount -o compress=zstd:1 /dev/mapper/ENC /mnt
btrfs subvol create /mnt/@
cd /mnt/@
btrfs subvol create ./home
btrfs subvol create ./opt
mkdir -p var
btrfs subvol create ./var/log
btrfs suvol set-default /mnt/@

cd /mnt/
debootstrap --include=dbus,locales,tzdata unstable @/ http://deb.debian.org/debian

umount /mnt
mount -o compress=zstd:1 /dev/mapper/ENC /mnt
mkdir -p /mnt/boot/efi
mount /dev/sda1 /mnt/boot/efi
grml-chroot /mnt /bin/bash

apt-get update
apt-get install linux-image-amd64 dracut openssh-client \
                        kde-plasma-desktop plasma-workspace-wayland \
                        plasma-nm cryptsetup btrfs-progs sudo

LABEL="root_disk" / btrfs defaults,compress=zstd:1 0 0

ENCRYPTED_ROOT UUID=xxxx none discard,x-initrd.attach

export DEBIAN_FRONTEND=noninteractive
dpkg-reconfigure locales
dpkg-reconfigure tzdata

systemd.gpt_auto=no quiet root=LABEL=root_disk

apt-get install -y systemd-boot
bootctl install --make-entry-directory=yes --entry-token=machine-id
dpkg-reconfigure linux-image-6.0.0-5-amd64

Installation Since this is a conversion (and not a new install), our procedure is slightly different than the official documentation but otherwise it's pretty much in the same spirit: we're going to use ZFS for everything, including the root filesystem. So, install the required packages, on the current system:

apt install --yes gdisk zfs-dkms zfs zfs-initramfs zfsutils-linux

We also tell DKMS that we need to rebuild the initrd when upgrading:

echo REMAKE_INITRD=yes > /etc/dkms/zfs.conf

Partitioning This is going to partition /dev/sdc with:

• 1MB MBR / BIOS legacy boot
• 512MB EFI boot
• 1GB bpool, unencrypted pool for /boot
• rest of the disk for zpool, the rest of the data

   sgdisk --zap-all /dev/sdc
   sgdisk -a1 -n1:24K:+1000K -t1:EF02 /dev/sdc
   sgdisk     -n2:1M:+512M   -t2:EF00 /dev/sdc
   sgdisk     -n3:0:+1G      -t3:BF01 /dev/sdc
   sgdisk     -n4:0:0        -t4:BF00 /dev/sdc
  

That will look something like this:

    root@curie:/home/anarcat# sgdisk -p /dev/sdc
    Disk /dev/sdc: 1953525168 sectors, 931.5 GiB
    Model: ESD-S1C         
    Sector size (logical/physical): 512/512 bytes
    Disk identifier (GUID): [REDACTED]
    Partition table holds up to 128 entries
    Main partition table begins at sector 2 and ends at sector 33
    First usable sector is 34, last usable sector is 1953525134
    Partitions will be aligned on 16-sector boundaries
    Total free space is 14 sectors (7.0 KiB)
    Number  Start (sector)    End (sector)  Size       Code  Name
       1              48            2047   1000.0 KiB  EF02  
       2            2048         1050623   512.0 MiB   EF00  
       3         1050624         3147775   1024.0 MiB  BF01  
       4         3147776      1953525134   930.0 GiB   BF00

Unfortunately, we can't be sure of the sector size here, because the USB controller is probably lying to us about it. Normally, this smartctl command should tell us the sector size as well:

root@curie:~# smartctl -i /dev/sdb -qnoserial
smartctl 7.2 2020-12-30 r5155 [x86_64-linux-5.10.0-14-amd64] (local build)
Copyright (C) 2002-20, Bruce Allen, Christian Franke, www.smartmontools.org
=== START OF INFORMATION SECTION ===
Model Family:     Western Digital Black Mobile
Device Model:     WDC WD10JPLX-00MBPT0
Firmware Version: 01.01H01
User Capacity:    1 000 204 886 016 bytes [1,00 TB]
Sector Sizes:     512 bytes logical, 4096 bytes physical
Rotation Rate:    7200 rpm
Form Factor:      2.5 inches
Device is:        In smartctl database [for details use: -P show]
ATA Version is:   ATA8-ACS T13/1699-D revision 6
SATA Version is:  SATA 3.0, 6.0 Gb/s (current: 6.0 Gb/s)
Local Time is:    Tue May 17 13:33:04 2022 EDT
SMART support is: Available - device has SMART capability.
SMART support is: Enabled

Above is the example of the builtin HDD drive. But the SSD device enclosed in that USB controller doesn't support SMART commands, so we can't trust that it really has 512 bytes sectors. This matters because we need to tweak the ashift value correctly. We're going to go ahead the SSD drive has the common 4KB settings, which means ashift=12. Note here that we are not creating a separate partition for swap. Swap on ZFS volumes (AKA "swap on ZVOL") can trigger lockups and that issue is still not fixed upstream. Ubuntu recommends using a separate partition for swap instead. But since this is "just" a workstation, we're betting that we will not suffer from this problem, after hearing a report from another Debian developer running this setup on their workstation successfully. We do not recommend this setup though. In fact, if I were to redo this partition scheme, I would probably use LUKS encryption and setup a dedicated swap partition, as I had problems with ZFS encryption as well.

Creating pools ZFS pools are somewhat like "volume groups" if you are familiar with LVM, except they obviously also do things like RAID-10. (Even though LVM can technically also do RAID, people typically use mdadm instead.) In any case, the guide suggests creating two different pools here: one, in cleartext, for boot, and a separate, encrypted one, for the rest. Technically, the boot partition is required because the Grub bootloader only supports readonly ZFS pools, from what I understand. But I'm a little out of my depth here and just following the guide.

Boot pool creation This creates the boot pool in readonly mode with features that grub supports:

    zpool create \
        -o cachefile=/etc/zfs/zpool.cache \
        -o ashift=12 -d \
        -o feature@async_destroy=enabled \
        -o feature@bookmarks=enabled \
        -o feature@embedded_data=enabled \
        -o feature@empty_bpobj=enabled \
        -o feature@enabled_txg=enabled \
        -o feature@extensible_dataset=enabled \
        -o feature@filesystem_limits=enabled \
        -o feature@hole_birth=enabled \
        -o feature@large_blocks=enabled \
        -o feature@lz4_compress=enabled \
        -o feature@spacemap_histogram=enabled \
        -o feature@zpool_checkpoint=enabled \
        -O acltype=posixacl -O canmount=off \
        -O compression=lz4 \
        -O devices=off -O normalization=formD -O relatime=on -O xattr=sa \
        -O mountpoint=/boot -R /mnt \
        bpool /dev/sdc3

I haven't investigated all those settings and just trust the upstream guide on the above.

Main pool creation This is a more typical pool creation.

    zpool create \
        -o ashift=12 \
        -O encryption=on -O keylocation=prompt -O keyformat=passphrase \
        -O acltype=posixacl -O xattr=sa -O dnodesize=auto \
        -O compression=zstd \
        -O relatime=on \
        -O canmount=off \
        -O mountpoint=/ -R /mnt \
        rpool /dev/sdc4

Breaking this down:

• -o ashift=12: mentioned above, 4k sector size
• -O encryption=on -O keylocation=prompt -O keyformat=passphrase: encryption, prompt for a password, default algorithm is aes-256-gcm, explicit in the guide, made implicit here
• -O acltype=posixacl -O xattr=sa: enable ACLs, with better performance (not enabled by default)
• -O dnodesize=auto: related to extended attributes, less compatibility with other implementations
• -O compression=zstd: enable zstd compression, can be disabled/enabled by dataset to with zfs set compression=off rpool/example
• -O relatime=on: classic atime optimisation, another that could be used on a busy server is atime=off
• -O canmount=off: do not make the pool mount automatically with mount -a?
• -O mountpoint=/ -R /mnt: mount pool on / in the future, but /mnt for now

Those settings are all available in zfsprops(8). Other flags are defined in zpool-create(8). The reasoning behind them is also explained in the upstream guide and some also in [the Debian wiki][]. Those flags were actually not used:

• -O normalization=formD: normalize file names on comparisons (not storage), implies utf8only=on, which is a bad idea (and effectively meant my first sync failed to copy some files, including this folder from a supysonic checkout). and this cannot be changed after the filesystem is created. bad, bad, bad.

[the Debian wiki]: https://wiki.debian.org/ZFS#Advanced_Topics

Side note about single-disk pools Also note that we're living dangerously here: single-disk ZFS pools are rumoured to be more dangerous than not running ZFS at all. The choice quote from this article is:

[...] any error can be detected, but cannot be corrected. This sounds like an acceptable compromise, but its actually not. The reason its not is that ZFS' metadata cannot be allowed to be corrupted. If it is it is likely the zpool will be impossible to mount (and will probably crash the system once the corruption is found). So a couple of bad sectors in the right place will mean that all data on the zpool will be lost. Not some, all. Also there's no ZFS recovery tools, so you cannot recover any data on the drives.

Compared with (say) ext4, where a single disk error can recovered, this is pretty bad. But we are ready to live with this with the idea that we'll have hourly offline snapshots that we can easily recover from. It's trade-off. Also, we're running this on a NVMe/M.2 drive which typically just blinks out of existence completely, and doesn't "bit rot" the way a HDD would. Also, the FreeBSD handbook quick start doesn't have any warnings about their first example, which is with a single disk. So I am reassured at least.

Creating mount points Next we create the actual filesystems, known as "datasets" which are the things that get mounted on mountpoint and hold the actual files.

• this creates two containers, for ROOT and BOOT

   zfs create -o canmount=off -o mountpoint=none rpool/ROOT &&
   zfs create -o canmount=off -o mountpoint=none bpool/BOOT
  

  Note that it's unclear to me why those datasets are necessary, but they seem common practice, also used in this FreeBSD example. The OpenZFS guide mentions the Solaris upgrades and Ubuntu's zsys that use that container for upgrades and rollbacks. This blog post seems to explain a bit the layout behind the installer.
• this creates the actual boot and root filesystems:

   zfs create -o canmount=noauto -o mountpoint=/ rpool/ROOT/debian &&
   zfs mount rpool/ROOT/debian &&
   zfs create -o mountpoint=/boot bpool/BOOT/debian
  

  I guess the debian name here is because we could technically have multiple operating systems with the same underlying datasets.
• then the main datasets:

   zfs create                                 rpool/home &&
   zfs create -o mountpoint=/root             rpool/home/root &&
   chmod 700 /mnt/root &&
   zfs create                                 rpool/var
  

• exclude temporary files from snapshots:

   zfs create -o com.sun:auto-snapshot=false  rpool/var/cache &&
   zfs create -o com.sun:auto-snapshot=false  rpool/var/tmp &&
   chmod 1777 /mnt/var/tmp
  

• and skip automatic snapshots in Docker:

   zfs create -o canmount=off                 rpool/var/lib &&
   zfs create -o com.sun:auto-snapshot=false  rpool/var/lib/docker
  

  Notice here a peculiarity: we must create rpool/var/lib to create rpool/var/lib/docker otherwise we get this error:

   cannot create 'rpool/var/lib/docker': parent does not exist
  

  ... and no, just creating /mnt/var/lib doesn't fix that problem. In fact, it makes things even more confusing because an existing directory shadows a mountpoint, which is the opposite of how things normally work. Also note that you will probably need to change storage driver in Docker, see the zfs-driver documentation for details but, basically, I did:

  echo '  "storage-driver": "zfs"  ' > /etc/docker/daemon.json
  

  Note that podman has the same problem (and similar solution):

  printf '[storage]\ndriver = "zfs"\n' > /etc/containers/storage.conf
  

• make a tmpfs for /run:

   mkdir /mnt/run &&
   mount -t tmpfs tmpfs /mnt/run &&
   mkdir /mnt/run/lock
  

We don't create a /srv, as that's the HDD stuff. Also mount the EFI partition:

mkfs.fat -F 32 /dev/sdc2 &&
mount /dev/sdc2 /mnt/boot/efi/

At this point, everything should be mounted in /mnt. It should look like this:

root@curie:~# LANG=C df -h -t zfs -t vfat
Filesystem            Size  Used Avail Use% Mounted on
rpool/ROOT/debian     899G  384K  899G   1% /mnt
bpool/BOOT/debian     832M  123M  709M  15% /mnt/boot
rpool/home            899G  256K  899G   1% /mnt/home
rpool/home/root       899G  256K  899G   1% /mnt/root
rpool/var             899G  384K  899G   1% /mnt/var
rpool/var/cache       899G  256K  899G   1% /mnt/var/cache
rpool/var/tmp         899G  256K  899G   1% /mnt/var/tmp
rpool/var/lib/docker  899G  256K  899G   1% /mnt/var/lib/docker
/dev/sdc2             511M  4.0K  511M   1% /mnt/boot/efi

Now that we have everything setup and mounted, let's copy all files over.

Copying files This is a list of all the mounted filesystems

for fs in /boot/ /boot/efi/ / /home/; do
    echo "syncing $fs to /mnt$fs..." && 
    rsync -aSHAXx --info=progress2 --delete $fs /mnt$fs
done

You can check that the list is correct with:

mount -l -t ext4,btrfs,vfat   awk ' print $3 '

Note that we skip /srv as it's on a different disk. On the first run, we had:

root@curie:~# for fs in /boot/ /boot/efi/ / /home/; do
        echo "syncing $fs to /mnt$fs..." && 
        rsync -aSHAXx --info=progress2 $fs /mnt$fs
    done
syncing /boot/ to /mnt/boot/...
              0   0%    0.00kB/s    0:00:00 (xfr#0, to-chk=0/299)  
syncing /boot/efi/ to /mnt/boot/efi/...
     16,831,437 100%  184.14MB/s    0:00:00 (xfr#101, to-chk=0/110)
syncing / to /mnt/...
 28,019,293,280  94%   47.63MB/s    0:09:21 (xfr#703710, ir-chk=6748/839220)rsync: [generator] delete_file: rmdir(var/lib/docker) failed: Device or resource busy (16)
could not make way for new symlink: var/lib/docker
 34,081,267,990  98%   50.71MB/s    0:10:40 (xfr#736577, to-chk=0/867732)    
rsync error: some files/attrs were not transferred (see previous errors) (code 23) at main.c(1333) [sender=3.2.3]
syncing /home/ to /mnt/home/...
rsync: [sender] readlink_stat("/home/anarcat/.fuse") failed: Permission denied (13)
 24,456,268,098  98%   68.03MB/s    0:05:42 (xfr#159867, ir-chk=6875/172377) 
file has vanished: "/home/anarcat/.cache/mozilla/firefox/s2hwvqbu.quantum/cache2/entries/B3AB0CDA9C4454B3C1197E5A22669DF8EE849D90"
199,762,528,125  93%   74.82MB/s    0:42:26 (xfr#1437846, ir-chk=1018/1983979)rsync: [generator] recv_generator: mkdir "/mnt/home/anarcat/dist/supysonic/tests/assets/\#346" failed: Invalid or incomplete multibyte or wide character (84)
*** Skipping any contents from this failed directory ***
315,384,723,978  96%   76.82MB/s    1:05:15 (xfr#2256473, to-chk=0/2993950)    
rsync error: some files/attrs were not transferred (see previous errors) (code 23) at main.c(1333) [sender=3.2.3]

Note the failure to transfer that supysonic file? It turns out they had a weird filename in their source tree, since then removed, but still it showed how the utf8only feature might not be such a bad idea. At this point, the procedure was restarted all the way back to "Creating pools", after unmounting all ZFS filesystems (umount /mnt/run /mnt/boot/efi && umount -t zfs -a) and destroying the pool, which, surprisingly, doesn't require any confirmation (zpool destroy rpool). The second run was cleaner:

root@curie:~# for fs in /boot/ /boot/efi/ / /home/; do
        echo "syncing $fs to /mnt$fs..." && 
        rsync -aSHAXx --info=progress2 --delete $fs /mnt$fs
    done
syncing /boot/ to /mnt/boot/...
              0   0%    0.00kB/s    0:00:00 (xfr#0, to-chk=0/299)  
syncing /boot/efi/ to /mnt/boot/efi/...
              0   0%    0.00kB/s    0:00:00 (xfr#0, to-chk=0/110)  
syncing / to /mnt/...
 28,019,033,070  97%   42.03MB/s    0:10:35 (xfr#703671, ir-chk=1093/833515)rsync: [generator] delete_file: rmdir(var/lib/docker) failed: Device or resource busy (16)
could not make way for new symlink: var/lib/docker
 34,081,807,102  98%   44.84MB/s    0:12:04 (xfr#736580, to-chk=0/867723)    
rsync error: some files/attrs were not transferred (see previous errors) (code 23) at main.c(1333) [sender=3.2.3]
syncing /home/ to /mnt/home/...
rsync: [sender] readlink_stat("/home/anarcat/.fuse") failed: Permission denied (13)
IO error encountered -- skipping file deletion
 24,043,086,450  96%   62.03MB/s    0:06:09 (xfr#151819, ir-chk=15117/172571)
file has vanished: "/home/anarcat/.cache/mozilla/firefox/s2hwvqbu.quantum/cache2/entries/4C1FDBFEA976FF924D062FB990B24B897A77B84B"
315,423,626,507  96%   67.09MB/s    1:14:43 (xfr#2256845, to-chk=0/2994364)    
rsync error: some files/attrs were not transferred (see previous errors) (code 23) at main.c(1333) [sender=3.2.3]

Also note the transfer speed: we seem capped at 76MB/s, or 608Mbit/s. This is not as fast as I was expecting: the USB connection seems to be at around 5Gbps:

anarcat@curie:~$ lsusb -tv   head -4
/:  Bus 02.Port 1: Dev 1, Class=root_hub, Driver=xhci_hcd/6p, 5000M
    ID 1d6b:0003 Linux Foundation 3.0 root hub
     __ Port 1: Dev 4, If 0, Class=Mass Storage, Driver=uas, 5000M
        ID 0b05:1932 ASUSTek Computer, Inc.

So it shouldn't cap at that speed. It's possible the USB adapter is failing to give me the full speed though. It's not the M.2 SSD drive either, as that has a ~500MB/s bandwidth, acccording to its spec. At this point, we're about ready to do the final configuration. We drop to single user mode and do the rest of the procedure. That used to be shutdown now, but it seems like the systemd switch broke that, so now you can reboot into grub and pick the "recovery" option. Alternatively, you might try systemctl rescue, as I found out. I also wanted to copy the drive over to another new NVMe drive, but that failed: it looks like the USB controller I have doesn't work with older, non-NVME drives.

Boot configuration Now we need to enter the new system to rebuild the boot loader and initrd and so on. First, we bind mounts and chroot into the ZFS disk:

mount --rbind /dev  /mnt/dev &&
mount --rbind /proc /mnt/proc &&
mount --rbind /sys  /mnt/sys &&
chroot /mnt /bin/bash

Next we add an extra service that imports the bpool on boot, to make sure it survives a zpool.cache destruction:

cat > /etc/systemd/system/zfs-import-bpool.service <<EOF
[Unit]
DefaultDependencies=no
Before=zfs-import-scan.service
Before=zfs-import-cache.service
[Service]
Type=oneshot
RemainAfterExit=yes
ExecStart=/sbin/zpool import -N -o cachefile=none bpool
# Work-around to preserve zpool cache:
ExecStartPre=-/bin/mv /etc/zfs/zpool.cache /etc/zfs/preboot_zpool.cache
ExecStartPost=-/bin/mv /etc/zfs/preboot_zpool.cache /etc/zfs/zpool.cache
[Install]
WantedBy=zfs-import.target
EOF

Enable the service:

systemctl enable zfs-import-bpool.service

I had to trim down /etc/fstab and /etc/crypttab to only contain references to the legacy filesystems (/srv is still BTRFS!). If we don't already have a tmpfs defined in /etc/fstab:

ln -s /usr/share/systemd/tmp.mount /etc/systemd/system/ &&
systemctl enable tmp.mount

Rebuild boot loader with support for ZFS, but also to workaround GRUB's missing zpool-features support:

grub-probe /boot   grep -q zfs &&
update-initramfs -c -k all &&
sed -i 's,GRUB_CMDLINE_LINUX.*,GRUB_CMDLINE_LINUX="root=ZFS=rpool/ROOT/debian",' /etc/default/grub &&
update-grub

For good measure, make sure the right disk is configured here, for example you might want to tag both drives in a RAID array:

dpkg-reconfigure grub-pc

Install grub to EFI while you're there:

grub-install --target=x86_64-efi --efi-directory=/boot/efi --bootloader-id=debian --recheck --no-floppy

Filesystem mount ordering. The rationale here in the OpenZFS guide is a little strange, but I don't dare ignore that.

mkdir /etc/zfs/zfs-list.cache
touch /etc/zfs/zfs-list.cache/bpool
touch /etc/zfs/zfs-list.cache/rpool
zed -F &

Verify that zed updated the cache by making sure these are not empty:

cat /etc/zfs/zfs-list.cache/bpool
cat /etc/zfs/zfs-list.cache/rpool

Once the files have data, stop zed:

fg
Press Ctrl-C.

Fix the paths to eliminate /mnt:

sed -Ei "s /mnt/? / " /etc/zfs/zfs-list.cache/*

Snapshot initial install:

zfs snapshot bpool/BOOT/debian@install
zfs snapshot rpool/ROOT/debian@install

Exit chroot:

exit

Finalizing One last sync was done in rescue mode:

for fs in /boot/ /boot/efi/ / /home/; do
    echo "syncing $fs to /mnt$fs..." && 
    rsync -aSHAXx --info=progress2 --delete $fs /mnt$fs
done

Then we unmount all filesystems:

mount   grep -v zfs   tac   awk '/\/mnt/  print $3 '   xargs -i  umount -lf  
zpool export -a

Reboot, swap the drives, and boot in ZFS. Hurray!

Benchmarks This is a test that was ran in single-user mode using fio and the Ars Technica recommended tests, which are:

• Single 4KiB random write process:

   fio --name=randwrite4k1x --ioengine=posixaio --rw=randwrite --bs=4k --size=4g --numjobs=1 --iodepth=1 --runtime=60 --time_based --end_fsync=1
  

• 16 parallel 64KiB random write processes:

   fio --name=randwrite64k16x --ioengine=posixaio --rw=randwrite --bs=64k --size=256m --numjobs=16 --iodepth=16 --runtime=60 --time_based --end_fsync=1
  

• Single 1MiB random write process:

   fio --name=randwrite1m1x --ioengine=posixaio --rw=randwrite --bs=1m --size=16g --numjobs=1 --iodepth=1 --runtime=60 --time_based --end_fsync=1
  

Strangely, that's not exactly what the author, Jim Salter, did in his actual test bench used in the ZFS benchmarking article. The first thing is there's no read test at all, which is already pretty strange. But also it doesn't include stuff like dropping caches or repeating results. So here's my variation, which i called fio-ars-bench.sh for now. It just batches a bunch of fio tests, one by one, 60 seconds each. It should take about 12 minutes to run, as there are 3 pair of tests, read/write, with and without async. My bias, before building, running and analysing those results is that ZFS should outperform the traditional stack on writes, but possibly not on reads. It's also possible it outperforms it on both, because it's a newer drive. A new test might be possible with a new external USB drive as well, although I doubt I will find the time to do this.

Results All tests were done on WD blue SN550 drives, which claims to be able to push 2400MB/s read and 1750MB/s write. An extra drive was bought to move the LVM setup from a WDC WDS500G1B0B-00AS40 SSD, a WD blue M.2 2280 SSD that was at least 5 years old, spec'd at 560MB/s read, 530MB/s write. Benchmarks were done on the M.2 SSD drive but discarded so that the drive difference is not a factor in the test. In practice, I'm going to assume we'll never reach those numbers because we're not actually NVMe (this is an old workstation!) so the bottleneck isn't the disk itself. For our purposes, it might still give us useful results.

Rescue test, LUKS/LVM/ext4 Those tests were performed with everything shutdown, after either entering the system in rescue mode, or by reaching that target with:

systemctl rescue

The network might have been started before or after the test as well:

systemctl start systemd-networkd

So it should be fairly reliable as basically nothing else is running. Raw numbers, from the ?job-curie-lvm.log, converted to MiB/s and manually merged:

test	read I/O	read IOPS	write I/O	write IOPS
rand4k4g1x	39.27	10052	212.15	54310
rand4k4g1x--fsync=1	39.29	10057	2.73	699
rand64k256m16x	1297.00	20751	1068.57	17097
rand64k256m16x--fsync=1	1290.90	20654	353.82	5661
rand1m16g1x	315.15	315	563.77	563
rand1m16g1x--fsync=1	345.88	345	157.01	157

Peaks are at about 20k IOPS and ~1.3GiB/s read, 1GiB/s write in the 64KB blocks with 16 jobs. Slowest is the random 4k block sync write at an abysmal 3MB/s and 700 IOPS The 1MB read/write tests have lower IOPS, but that is expected.

Rescue test, ZFS This test was also performed in rescue mode. Raw numbers, from the ?job-curie-zfs.log, converted to MiB/s and manually merged:

test	read I/O	read IOPS	write I/O	write IOPS
rand4k4g1x	77.20	19763	27.13	6944
rand4k4g1x--fsync=1	76.16	19495	6.53	1673
rand64k256m16x	1882.40	30118	70.58	1129
rand64k256m16x--fsync=1	1865.13	29842	71.98	1151
rand1m16g1x	921.62	921	102.21	102
rand1m16g1x--fsync=1	908.37	908	64.30	64

Peaks are at 1.8GiB/s read, also in the 64k job like above, but much faster. The write is, as expected, much slower at 70MiB/s (compared to 1GiB/s!), but it should be noted the sync write doesn't degrade performance compared to async writes (although it's still below the LVM 300MB/s).

Conclusions Really, ZFS has trouble performing in all write conditions. The random 4k sync write test is the only place where ZFS outperforms LVM in writes, and barely (7MiB/s vs 3MiB/s). Everywhere else, writes are much slower, sometimes by an order of magnitude. And before some ZFS zealot jumps in talking about the SLOG or some other cache that could be added to improved performance, I'll remind you that those numbers are on a bare bones NVMe drive, pretty much as fast storage as you can find on this machine. Adding another NVMe drive as a cache probably will not improve write performance here. Still, those are very different results than the tests performed by Salter which shows ZFS beating traditional configurations in all categories but uncached 4k reads (not writes!). That said, those tests are very different from the tests I performed here, where I test writes on a single disk, not a RAID array, which might explain the discrepancy. Also, note that neither LVM or ZFS manage to reach the 2400MB/s read and 1750MB/s write performance specification. ZFS does manage to reach 82% of the read performance (1973MB/s) and LVM 64% of the write performance (1120MB/s). LVM hits 57% of the read performance and ZFS hits barely 6% of the write performance. Overall, I'm a bit disappointed in the ZFS write performance here, I must say. Maybe I need to tweak the record size or some other ZFS voodoo, but I'll note that I didn't have to do any such configuration on the other side to kick ZFS in the pants...

Real world experience This section document not synthetic backups, but actual real world workloads, comparing before and after I switched my workstation to ZFS.

Docker performance I had the feeling that running some git hook (which was firing a Docker container) was "slower" somehow. It seems that, at runtime, ZFS backends are significant slower than their overlayfs/ext4 equivalent:

May 16 14:42:52 curie systemd[1]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87\x2dinit-merged.mount: Succeeded.
May 16 14:42:52 curie systemd[5161]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87\x2dinit-merged.mount: Succeeded.
May 16 14:42:52 curie systemd[1]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87-merged.mount: Succeeded.
May 16 14:42:53 curie dockerd[1723]: time="2022-05-16T14:42:53.087219426-04:00" level=info msg="starting signal loop" namespace=moby path=/run/docker/containerd/daemon/io.containerd.runtime.v2.task/moby/af22586fba07014a4d10ab19da10cf280db7a43cad804d6c1e9f2682f12b5f10 pid=151170
May 16 14:42:53 curie systemd[1]: Started libcontainer container af22586fba07014a4d10ab19da10cf280db7a43cad804d6c1e9f2682f12b5f10.
May 16 14:42:54 curie systemd[1]: docker-af22586fba07014a4d10ab19da10cf280db7a43cad804d6c1e9f2682f12b5f10.scope: Succeeded.
May 16 14:42:54 curie dockerd[1723]: time="2022-05-16T14:42:54.047297800-04:00" level=info msg="shim disconnected" id=af22586fba07014a4d10ab19da10cf280db7a43cad804d6c1e9f2682f12b5f10
May 16 14:42:54 curie dockerd[998]: time="2022-05-16T14:42:54.051365015-04:00" level=info msg="ignoring event" container=af22586fba07014a4d10ab19da10cf280db7a43cad804d6c1e9f2682f12b5f10 module=libcontainerd namespace=moby topic=/tasks/delete type="*events.TaskDelete"
May 16 14:42:54 curie systemd[2444]: run-docker-netns-f5453c87c879.mount: Succeeded.
May 16 14:42:54 curie systemd[5161]: run-docker-netns-f5453c87c879.mount: Succeeded.
May 16 14:42:54 curie systemd[2444]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87-merged.mount: Succeeded.
May 16 14:42:54 curie systemd[5161]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87-merged.mount: Succeeded.
May 16 14:42:54 curie systemd[1]: run-docker-netns-f5453c87c879.mount: Succeeded.
May 16 14:42:54 curie systemd[1]: home-docker-overlay2-17e4d24228decc2d2d493efc401dbfb7ac29739da0e46775e122078d9daf3e87-merged.mount: Succeeded.

Translating this:

• container setup: ~1 second
• container runtime: ~1 second
• container teardown: ~1 second
• total runtime: 2-3 seconds

Obviously, those timestamps are not quite accurate enough to make precise measurements... After I switched to ZFS:

mai 30 15:31:39 curie systemd[1]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf\x2dinit.mount: Succeeded. 
mai 30 15:31:39 curie systemd[5287]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf\x2dinit.mount: Succeeded. 
mai 30 15:31:40 curie systemd[1]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf.mount: Succeeded. 
mai 30 15:31:40 curie systemd[5287]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf.mount: Succeeded. 
mai 30 15:31:41 curie dockerd[3199]: time="2022-05-30T15:31:41.551403693-04:00" level=info msg="starting signal loop" namespace=moby path=/run/docker/containerd/daemon/io.containerd.runtime.v2.task/moby/42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142 pid=141080 
mai 30 15:31:41 curie systemd[1]: run-docker-runtime\x2drunc-moby-42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142-runc.ZVcjvl.mount: Succeeded. 
mai 30 15:31:41 curie systemd[5287]: run-docker-runtime\x2drunc-moby-42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142-runc.ZVcjvl.mount: Succeeded. 
mai 30 15:31:41 curie systemd[1]: Started libcontainer container 42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142. 
mai 30 15:31:45 curie systemd[1]: docker-42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142.scope: Succeeded. 
mai 30 15:31:45 curie dockerd[3199]: time="2022-05-30T15:31:45.883019128-04:00" level=info msg="shim disconnected" id=42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142 
mai 30 15:31:45 curie dockerd[1726]: time="2022-05-30T15:31:45.883064491-04:00" level=info msg="ignoring event" container=42a1a1ed5912a7227148e997f442e7ab2e5cc3558aa3471548223c5888c9b142 module=libcontainerd namespace=moby topic=/tasks/delete type="*events.TaskDelete" 
mai 30 15:31:45 curie systemd[1]: run-docker-netns-e45f5cf5f465.mount: Succeeded. 
mai 30 15:31:45 curie systemd[5287]: run-docker-netns-e45f5cf5f465.mount: Succeeded. 
mai 30 15:31:45 curie systemd[1]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf.mount: Succeeded. 
mai 30 15:31:45 curie systemd[5287]: var-lib-docker-zfs-graph-41ce08fb7a1d3a9c101694b82722f5621c0b4819bd1d9f070933fd1e00543cdf.mount: Succeeded.

That's double or triple the run time, from 2 seconds to 6 seconds. Most of the time is spent in run time, inside the container. Here's the breakdown:

• container setup: ~2 seconds
• container run: ~4 seconds
• container teardown: ~1 second
• total run time: about ~6-7 seconds

That's a two- to three-fold increase! Clearly something is going on here that I should tweak. It's possible that code path is less optimized in Docker. I also worry about podman, but apparently it also supports ZFS backends. Possibly it would perform better, but at this stage I wouldn't have a good comparison: maybe it would have performed better on non-ZFS as well...

Interactivity While doing the offsite backups (below), the system became somewhat "sluggish". I felt everything was slow, and I estimate it introduced ~50ms latency in any input device. Arguably, those are all USB and the external drive was connected through USB, but I suspect the ZFS drivers are not as well tuned with the scheduler as the regular filesystem drivers...

Recovery procedures For test purposes, I unmounted all systems during the procedure:

umount /mnt/boot/efi /mnt/boot/run
umount -a -t zfs
zpool export -a

And disconnected the drive, to see how I would recover this system from another Linux system in case of a total motherboard failure. To import an existing pool, plug the device, then import the pool with an alternate root, so it doesn't mount over your existing filesystems, then you mount the root filesystem and all the others:

zpool import -l -a -R /mnt &&
zfs mount rpool/ROOT/debian &&
zfs mount -a &&
mount /dev/sdc2 /mnt/boot/efi &&
mount -t tmpfs tmpfs /mnt/run &&
mkdir /mnt/run/lock

Offsite backup Part of the goal of using ZFS is to simplify and harden backups. I wanted to experiment with shorter recovery times specifically both point in time recovery objective and recovery time objective and faster incremental backups. This is, therefore, part of my backup services. This section documents how an external NVMe enclosure was setup in a pool to mirror the datasets from my workstation. The final setup should include syncoid copying datasets to the backup server regularly, but I haven't finished that configuration yet.

Partitioning The above partitioning procedure used sgdisk, but I couldn't figure out how to do this with sgdisk, so this uses sfdisk to dump the partition from the first disk to an external, identical drive:

sfdisk -d /dev/nvme0n1   sfdisk --no-reread /dev/sda --force

Pool creation This is similar to the main pool creation, except we tweaked a few bits after changing the upstream procedure:

zpool create \
        -o cachefile=/etc/zfs/zpool.cache \
        -o ashift=12 -d \
        -o feature@async_destroy=enabled \
        -o feature@bookmarks=enabled \
        -o feature@embedded_data=enabled \
        -o feature@empty_bpobj=enabled \
        -o feature@enabled_txg=enabled \
        -o feature@extensible_dataset=enabled \
        -o feature@filesystem_limits=enabled \
        -o feature@hole_birth=enabled \
        -o feature@large_blocks=enabled \
        -o feature@lz4_compress=enabled \
        -o feature@spacemap_histogram=enabled \
        -o feature@zpool_checkpoint=enabled \
        -O acltype=posixacl -O xattr=sa \
        -O compression=lz4 \
        -O devices=off \
        -O relatime=on \
        -O canmount=off \
        -O mountpoint=/boot -R /mnt \
        bpool-tubman /dev/sdb3

The change from the main boot pool are:

• no unicode normalization
• different device path (sdb used to be the M.2 device, it's now nvme0n1)
• reordered parameters

Main pool creation is:

zpool create \
        -o ashift=12 \
        -O encryption=on -O keylocation=prompt -O keyformat=passphrase \
        -O acltype=posixacl -O xattr=sa -O dnodesize=auto \
        -O compression=zstd \
        -O relatime=on \
        -O canmount=off \
        -O mountpoint=/ -R /mnt \
        rpool-tubman /dev/sdb4

First sync I used syncoid to copy all pools over to the external device. syncoid is a thing that's part of the sanoid project which is specifically designed to sync snapshots between pool, typically over SSH links but it can also operate locally. The sanoid command had a --readonly argument to simulate changes, but syncoid didn't so I tried to fix that with an upstream PR. It seems it would be better to do this by hand, but this was much easier. The full first sync was:

root@curie:/home/anarcat# ./bin/syncoid -r  bpool bpool-tubman
CRITICAL ERROR: Target bpool-tubman exists but has no snapshots matching with bpool!
                Replication to target would require destroying existing
                target. Cowardly refusing to destroy your existing target.
          NOTE: Target bpool-tubman dataset is < 64MB used - did you mistakenly run
                 zfs create bpool-tubman  on the target? ZFS initial
                replication must be to a NON EXISTENT DATASET, which will
                then be CREATED BY the initial replication process.
INFO: Sending oldest full snapshot bpool/BOOT@test (~ 42 KB) to new target filesystem:
44.2KiB 0:00:00 [4.19MiB/s] [========================================================================================================================] 103%            
INFO: Updating new target filesystem with incremental bpool/BOOT@test ... syncoid_curie_2022-05-30:12:50:39 (~ 4 KB):
2.13KiB 0:00:00 [ 114KiB/s] [===============================================================>                                                         ] 53%            
INFO: Sending oldest full snapshot bpool/BOOT/debian@install (~ 126.0 MB) to new target filesystem:
 126MiB 0:00:00 [ 308MiB/s] [=======================================================================================================================>] 100%            
INFO: Updating new target filesystem with incremental bpool/BOOT/debian@install ... syncoid_curie_2022-05-30:12:50:39 (~ 113.4 MB):
 113MiB 0:00:00 [ 315MiB/s] [=======================================================================================================================>] 100%
root@curie:/home/anarcat# ./bin/syncoid -r  rpool rpool-tubman
CRITICAL ERROR: Target rpool-tubman exists but has no snapshots matching with rpool!
                Replication to target would require destroying existing
                target. Cowardly refusing to destroy your existing target.
          NOTE: Target rpool-tubman dataset is < 64MB used - did you mistakenly run
                 zfs create rpool-tubman  on the target? ZFS initial
                replication must be to a NON EXISTENT DATASET, which will
                then be CREATED BY the initial replication process.
INFO: Sending oldest full snapshot rpool/ROOT@syncoid_curie_2022-05-30:12:50:51 (~ 69 KB) to new target filesystem:
44.2KiB 0:00:00 [2.44MiB/s] [===========================================================================>                                             ] 63%            
INFO: Sending oldest full snapshot rpool/ROOT/debian@install (~ 25.9 GB) to new target filesystem:
25.9GiB 0:03:33 [ 124MiB/s] [=======================================================================================================================>] 100%            
INFO: Updating new target filesystem with incremental rpool/ROOT/debian@install ... syncoid_curie_2022-05-30:12:50:52 (~ 3.9 GB):
3.92GiB 0:00:33 [ 119MiB/s] [======================================================================================================================>  ] 99%            
INFO: Sending oldest full snapshot rpool/home@syncoid_curie_2022-05-30:12:55:04 (~ 276.8 GB) to new target filesystem:
 277GiB 0:27:13 [ 174MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/home/root@syncoid_curie_2022-05-30:13:22:19 (~ 2.2 GB) to new target filesystem:
2.22GiB 0:00:25 [90.2MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/var@syncoid_curie_2022-05-30:13:22:47 (~ 5.6 GB) to new target filesystem:
5.56GiB 0:00:32 [ 176MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/var/cache@syncoid_curie_2022-05-30:13:23:22 (~ 627.3 MB) to new target filesystem:
 627MiB 0:00:03 [ 169MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/var/lib@syncoid_curie_2022-05-30:13:23:28 (~ 69 KB) to new target filesystem:
44.2KiB 0:00:00 [1.40MiB/s] [===========================================================================>                                             ] 63%            
INFO: Sending oldest full snapshot rpool/var/lib/docker@syncoid_curie_2022-05-30:13:23:28 (~ 442.6 MB) to new target filesystem:
 443MiB 0:00:04 [ 103MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/var/lib/docker/05c0de7fabbea60500eaa495d0d82038249f6faa63b12914737c4d71520e62c5@266253254 (~ 6.3 MB) to new target filesystem:
6.49MiB 0:00:00 [12.9MiB/s] [========================================================================================================================] 102%            
INFO: Updating new target filesystem with incremental rpool/var/lib/docker/05c0de7fabbea60500eaa495d0d82038249f6faa63b12914737c4d71520e62c5@266253254 ... syncoid_curie_2022-05-30:13:23:34 (~ 4 KB):
1.52KiB 0:00:00 [27.6KiB/s] [============================================>                                                                            ] 38%            
INFO: Sending oldest full snapshot rpool/var/lib/flatpak@syncoid_curie_2022-05-30:13:23:36 (~ 2.0 GB) to new target filesystem:
2.00GiB 0:00:17 [ 115MiB/s] [=======================================================================================================================>] 100%            
INFO: Sending oldest full snapshot rpool/var/tmp@syncoid_curie_2022-05-30:13:23:55 (~ 57.0 MB) to new target filesystem:
61.8MiB 0:00:01 [45.0MiB/s] [========================================================================================================================] 108%            
INFO: Clone is recreated on target rpool-tubman/var/lib/docker/ed71ddd563a779ba6fb37b3b1d0cc2c11eca9b594e77b4b234867ebcb162b205 based on rpool/var/lib/docker/05c0de7fabbea60500eaa495d0d82038249f6faa63b12914737c4d71520e62c5@266253254
INFO: Sending oldest full snapshot rpool/var/lib/docker/ed71ddd563a779ba6fb37b3b1d0cc2c11eca9b594e77b4b234867ebcb162b205@syncoid_curie_2022-05-30:13:23:58 (~ 218.6 MB) to new target filesystem:
 219MiB 0:00:01 [ 151MiB/s] [=======================================================================================================================>] 100%

Funny how the CRITICAL ERROR doesn't actually stop syncoid and it just carries on merrily doing when it's telling you it's "cowardly refusing to destroy your existing target"... Maybe that's because my pull request broke something though... During the transfer, the computer was very sluggish: everything feels like it has ~30-50ms latency extra:

anarcat@curie:sanoid$ LANG=C top -b  -n 1   head -20
top - 13:07:05 up 6 days,  4:01,  1 user,  load average: 16.13, 16.55, 11.83
Tasks: 606 total,   6 running, 598 sleeping,   0 stopped,   2 zombie
%Cpu(s): 18.8 us, 72.5 sy,  1.2 ni,  5.0 id,  1.2 wa,  0.0 hi,  1.2 si,  0.0 st
MiB Mem :  15898.4 total,   1387.6 free,  13170.0 used,   1340.8 buff/cache
MiB Swap:      0.0 total,      0.0 free,      0.0 used.   1319.8 avail Mem 
    PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
     70 root      20   0       0      0      0 S  83.3   0.0   6:12.67 kswapd0
4024878 root      20   0  282644  96432  10288 S  44.4   0.6   0:11.43 puppet
3896136 root      20   0   35328  16528     48 S  22.2   0.1   2:08.04 mbuffer
3896135 root      20   0   10328    776    168 R  16.7   0.0   1:22.93 zfs
3896138 root      20   0   10588    788    156 R  16.7   0.0   1:49.30 zfs
    350 root       0 -20       0      0      0 R  11.1   0.0   1:03.53 z_rd_int
    351 root       0 -20       0      0      0 S  11.1   0.0   1:04.15 z_rd_int
3896137 root      20   0    4384    352    244 R  11.1   0.0   0:44.73 pv
4034094 anarcat   30  10   20028  13960   2428 S  11.1   0.1   0:00.70 mbsync
4036539 anarcat   20   0    9604   3464   2408 R  11.1   0.0   0:00.04 top
    352 root       0 -20       0      0      0 S   5.6   0.0   1:03.64 z_rd_int
    353 root       0 -20       0      0      0 S   5.6   0.0   1:03.64 z_rd_int
    354 root       0 -20       0      0      0 S   5.6   0.0   1:04.01 z_rd_int

I wonder how much of that is due to syncoid, particularly because I often saw mbuffer and pv in there which are not strictly necessary to do those kind of operations, as far as I understand. Once that's done, export the pools to disconnect the drive:

zpool export bpool-tubman
zpool export rpool-tubman

Raw disk benchmark Copied the 512GB SSD/M.2 device to another 1024GB NVMe/M.2 device:

anarcat@curie:~$ sudo dd if=/dev/sdb of=/dev/sdc bs=4M status=progress conv=fdatasync
499944259584 octets (500 GB, 466 GiB) copi s, 1713 s, 292 MB/s
119235+1 enregistrements lus
119235+1 enregistrements  crits
500107862016 octets (500 GB, 466 GiB) copi s, 1719,93 s, 291 MB/s

... while both over USB, whoohoo 300MB/s!

Monitoring ZFS should be monitoring your pools regularly. Normally, the [[!debman zed]] daemon monitors all ZFS events. It is the thing that will report when a scrub failed, for example. See this configuration guide. Scrubs should be regularly scheduled to ensure consistency of the pool. This can be done in newer zfsutils-linux versions (bullseye-backports or bookworm) with one of those, depending on the desired frequency:

systemctl enable zfs-scrub-weekly@rpool.timer --now
systemctl enable zfs-scrub-monthly@rpool.timer --now

When the scrub runs, if it finds anything it will send an event which will get picked up by the zed daemon which will then send a notification, see below for an example. TODO: deploy on curie, if possible (probably not because no RAID) TODO: this should be in Puppet

Scrub warning example So what happens when problems are found? Here's an example of how I dealt with an error I received. After setting up another server (tubman) with ZFS, I eventually ended up getting a warning from the ZFS toolchain.

Date: Sun, 09 Oct 2022 00:58:08 -0400
From: root <root@anarc.at>
To: root@anarc.at
Subject: ZFS scrub_finish event for rpool on tubman
ZFS has finished a scrub:
   eid: 39536
 class: scrub_finish
  host: tubman
  time: 2022-10-09 00:58:07-0400
  pool: rpool
 state: ONLINE
status: One or more devices has experienced an unrecoverable error.  An
        attempt was made to correct the error.  Applications are unaffected.
action: Determine if the device needs to be replaced, and clear the errors
        using 'zpool clear' or replace the device with 'zpool replace'.
   see: https://openzfs.github.io/openzfs-docs/msg/ZFS-8000-9P
  scan: scrub repaired 0B in 00:33:57 with 0 errors on Sun Oct  9 00:58:07 2022
config:
        NAME        STATE     READ WRITE CKSUM
        rpool       ONLINE       0     0     0
          mirror-0  ONLINE       0     0     0
            sdb4    ONLINE       0     1     0
            sdc4    ONLINE       0     0     0
        cache
          sda3      ONLINE       0     0     0
errors: No known data errors

This, in itself, is a little worrisome. But it helpfully links to this more detailed documentation (and props up there: the link still works) which explains this is a "minor" problem (something that could be included in the report). In this case, this happened on a server setup on 2021-04-28, but the disks and server hardware are much older. The server itself (marcos v1) was built around 2011, over 10 years ago now. The hard drive in question is:

root@tubman:~# smartctl -i -qnoserial /dev/sdb
smartctl 7.2 2020-12-30 r5155 [x86_64-linux-5.10.0-15-amd64] (local build)
Copyright (C) 2002-20, Bruce Allen, Christian Franke, www.smartmontools.org
=== START OF INFORMATION SECTION ===
Model Family:     Seagate BarraCuda 3.5
Device Model:     ST4000DM004-2CV104
Firmware Version: 0001
User Capacity:    4,000,787,030,016 bytes [4.00 TB]
Sector Sizes:     512 bytes logical, 4096 bytes physical
Rotation Rate:    5425 rpm
Form Factor:      3.5 inches
Device is:        In smartctl database [for details use: -P show]
ATA Version is:   ACS-3 T13/2161-D revision 5
SATA Version is:  SATA 3.1, 6.0 Gb/s (current: 3.0 Gb/s)
Local Time is:    Tue Oct 11 11:02:32 2022 EDT
SMART support is: Available - device has SMART capability.
SMART support is: Enabled

Some more SMART stats:

root@tubman:~# smartctl -a -qnoserial /dev/sdb   grep -e  Head_Flying_Hours -e Power_On_Hours -e Total_LBA -e 'Sector Sizes'
Sector Sizes:     512 bytes logical, 4096 bytes physical
  9 Power_On_Hours          0x0032   086   086   000    Old_age   Always       -       12464 (206 202 0)
240 Head_Flying_Hours       0x0000   100   253   000    Old_age   Offline      -       10966h+55m+23.757s
241 Total_LBAs_Written      0x0000   100   253   000    Old_age   Offline      -       21107792664
242 Total_LBAs_Read         0x0000   100   253   000    Old_age   Offline      -       3201579750

That's over a year of power on, which shouldn't be so bad. It has written about 10TB of data (21107792664 LBAs * 512 byte/LBA), which is about two full writes. According to its specification, this device is supposed to support 55 TB/year of writes, so we're far below spec. Note that are still far from the "non-recoverable read error per bits" spec (1 per 10E15), as we've basically read 13E12 bits (3201579750 LBAs * 512 byte/LBA = 13E12 bits). It's likely this disk was made in 2018, so it is in its fourth year. Interestingly, /dev/sdc is also a Seagate drive, but of a different series:

root@tubman:~# smartctl -qnoserial  -i /dev/sdb
smartctl 7.2 2020-12-30 r5155 [x86_64-linux-5.10.0-15-amd64] (local build)
Copyright (C) 2002-20, Bruce Allen, Christian Franke, www.smartmontools.org
=== START OF INFORMATION SECTION ===
Model Family:     Seagate BarraCuda 3.5
Device Model:     ST4000DM004-2CV104
Firmware Version: 0001
User Capacity:    4,000,787,030,016 bytes [4.00 TB]
Sector Sizes:     512 bytes logical, 4096 bytes physical
Rotation Rate:    5425 rpm
Form Factor:      3.5 inches
Device is:        In smartctl database [for details use: -P show]
ATA Version is:   ACS-3 T13/2161-D revision 5
SATA Version is:  SATA 3.1, 6.0 Gb/s (current: 3.0 Gb/s)
Local Time is:    Tue Oct 11 11:21:35 2022 EDT
SMART support is: Available - device has SMART capability.
SMART support is: Enabled

It has seen much more reads than the other disk which is also interesting:

root@tubman:~# smartctl -a -qnoserial /dev/sdc   grep -e  Head_Flying_Hours -e Power_On_Hours -e Total_LBA -e 'Sector Sizes'
Sector Sizes:     512 bytes logical, 4096 bytes physical
  9 Power_On_Hours          0x0032   059   059   000    Old_age   Always       -       36240
240 Head_Flying_Hours       0x0000   100   253   000    Old_age   Offline      -       33994h+10m+52.118s
241 Total_LBAs_Written      0x0000   100   253   000    Old_age   Offline      -       30730174438
242 Total_LBAs_Read         0x0000   100   253   000    Old_age   Offline      -       51894566538

That's 4 years of Head_Flying_Hours, and over 4 years (4 years and 48 days) of Power_On_Hours. The copyright date on that drive's specs goes back to 2016, so it's a much older drive. SMART self-test succeeded.

Remaining issues

• TODO: move send/receive backups to offsite host, see also zfs for alternatives to syncoid/sanoid there
• TODO: setup backup cron job (or timer?)
• TODO: swap still not setup on curie, see zfs
• TODO: document this somewhere: bpool and rpool are both pools and datasets. that's pretty confusing, but also very useful because it allows for pool-wide recursive snapshots, which are used for the backup system

fio improvements I really want to improve my experience with fio. Right now, I'm just cargo-culting stuff from other folks and I don't really like it. stressant is a good example of my struggles, in the sense that it doesn't really work that well for disk tests. I would love to have just a single .fio job file that lists multiple jobs to run serially. For example, this file describes the above workload pretty well:

[global]
# cargo-culting Salter
fallocate=none
ioengine=posixaio
runtime=60
time_based=1
end_fsync=1
stonewall=1
group_reporting=1
# no need to drop caches, done by default
# invalidate=1
# Single 4KiB random read/write process
[randread-4k-4g-1x]
rw=randread
bs=4k
size=4g
numjobs=1
iodepth=1
[randwrite-4k-4g-1x]
rw=randwrite
bs=4k
size=4g
numjobs=1
iodepth=1
# 16 parallel 64KiB random read/write processes:
[randread-64k-256m-16x]
rw=randread
bs=64k
size=256m
numjobs=16
iodepth=16
[randwrite-64k-256m-16x]
rw=randwrite
bs=64k
size=256m
numjobs=16
iodepth=16
# Single 1MiB random read/write process
[randread-1m-16g-1x]
rw=randread
bs=1m
size=16g
numjobs=1
iodepth=1
[randwrite-1m-16g-1x]
rw=randwrite
bs=1m
size=16g
numjobs=1
iodepth=1

... except the jobs are actually started in parallel, even though they are stonewall'd, as far as I can tell by the reports. I sent a mail to the fio mailing list for clarification. It looks like the jobs are started in parallel, but actual (correctly) run serially. It seems like this might just be a matter of reporting the right timestamps in the end, although it does feel like starting all the processes (even if not doing any work yet) could skew the results.

Hangs during procedure During the procedure, it happened a few times where any ZFS command would completely hang. It seems that using an external USB drive to sync stuff didn't work so well: sometimes it would reconnect under a different device (from sdc to sdd, for example), and this would greatly confuse ZFS. Here, for example, is sdd reappearing out of the blue:

May 19 11:22:53 curie kernel: [  699.820301] scsi host4: uas
May 19 11:22:53 curie kernel: [  699.820544] usb 2-1: authorized to connect
May 19 11:22:53 curie kernel: [  699.922433] scsi 4:0:0:0: Direct-Access     ROG      ESD-S1C          0    PQ: 0 ANSI: 6
May 19 11:22:53 curie kernel: [  699.923235] sd 4:0:0:0: Attached scsi generic sg2 type 0
May 19 11:22:53 curie kernel: [  699.923676] sd 4:0:0:0: [sdd] 1953525168 512-byte logical blocks: (1.00 TB/932 GiB)
May 19 11:22:53 curie kernel: [  699.923788] sd 4:0:0:0: [sdd] Write Protect is off
May 19 11:22:53 curie kernel: [  699.923949] sd 4:0:0:0: [sdd] Write cache: enabled, read cache: enabled, doesn't support DPO or FUA
May 19 11:22:53 curie kernel: [  699.924149] sd 4:0:0:0: [sdd] Optimal transfer size 33553920 bytes
May 19 11:22:53 curie kernel: [  699.961602]  sdd: sdd1 sdd2 sdd3 sdd4
May 19 11:22:53 curie kernel: [  699.996083] sd 4:0:0:0: [sdd] Attached SCSI disk

Next time I run a ZFS command (say zpool list), the command completely hangs (D state) and this comes up in the logs:

May 19 11:34:21 curie kernel: [ 1387.914843] zio pool=bpool vdev=/dev/sdc3 error=5 type=2 offset=71344128 size=4096 flags=184880
May 19 11:34:21 curie kernel: [ 1387.914859] zio pool=bpool vdev=/dev/sdc3 error=5 type=2 offset=205565952 size=4096 flags=184880
May 19 11:34:21 curie kernel: [ 1387.914874] zio pool=bpool vdev=/dev/sdc3 error=5 type=2 offset=272789504 size=4096 flags=184880
May 19 11:34:21 curie kernel: [ 1387.914906] zio pool=bpool vdev=/dev/sdc3 error=5 type=1 offset=270336 size=8192 flags=b08c1
May 19 11:34:21 curie kernel: [ 1387.914932] zio pool=bpool vdev=/dev/sdc3 error=5 type=1 offset=1073225728 size=8192 flags=b08c1
May 19 11:34:21 curie kernel: [ 1387.914948] zio pool=bpool vdev=/dev/sdc3 error=5 type=1 offset=1073487872 size=8192 flags=b08c1
May 19 11:34:21 curie kernel: [ 1387.915165] zio pool=bpool vdev=/dev/sdc3 error=5 type=2 offset=272793600 size=4096 flags=184880
May 19 11:34:21 curie kernel: [ 1387.915183] zio pool=bpool vdev=/dev/sdc3 error=5 type=2 offset=339853312 size=4096 flags=184880
May 19 11:34:21 curie kernel: [ 1387.915648] WARNING: Pool 'bpool' has encountered an uncorrectable I/O failure and has been suspended.
May 19 11:34:21 curie kernel: [ 1387.915648] 
May 19 11:37:25 curie kernel: [ 1571.558614] task:txg_sync        state:D stack:    0 pid:  997 ppid:     2 flags:0x00004000
May 19 11:37:25 curie kernel: [ 1571.558623] Call Trace:
May 19 11:37:25 curie kernel: [ 1571.558640]  __schedule+0x282/0x870
May 19 11:37:25 curie kernel: [ 1571.558650]  schedule+0x46/0xb0
May 19 11:37:25 curie kernel: [ 1571.558670]  schedule_timeout+0x8b/0x140
May 19 11:37:25 curie kernel: [ 1571.558675]  ? __next_timer_interrupt+0x110/0x110
May 19 11:37:25 curie kernel: [ 1571.558678]  io_schedule_timeout+0x4c/0x80
May 19 11:37:25 curie kernel: [ 1571.558689]  __cv_timedwait_common+0x12b/0x160 [spl]
May 19 11:37:25 curie kernel: [ 1571.558694]  ? add_wait_queue_exclusive+0x70/0x70
May 19 11:37:25 curie kernel: [ 1571.558702]  __cv_timedwait_io+0x15/0x20 [spl]
May 19 11:37:25 curie kernel: [ 1571.558816]  zio_wait+0x129/0x2b0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.558929]  dsl_pool_sync+0x461/0x4f0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559032]  spa_sync+0x575/0xfa0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559138]  ? spa_txg_history_init_io+0x101/0x110 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559245]  txg_sync_thread+0x2e0/0x4a0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559354]  ? txg_fini+0x240/0x240 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559366]  thread_generic_wrapper+0x6f/0x80 [spl]
May 19 11:37:25 curie kernel: [ 1571.559376]  ? __thread_exit+0x20/0x20 [spl]
May 19 11:37:25 curie kernel: [ 1571.559379]  kthread+0x11b/0x140
May 19 11:37:25 curie kernel: [ 1571.559382]  ? __kthread_bind_mask+0x60/0x60
May 19 11:37:25 curie kernel: [ 1571.559386]  ret_from_fork+0x22/0x30
May 19 11:37:25 curie kernel: [ 1571.559401] task:zed             state:D stack:    0 pid: 1564 ppid:     1 flags:0x00000000
May 19 11:37:25 curie kernel: [ 1571.559404] Call Trace:
May 19 11:37:25 curie kernel: [ 1571.559409]  __schedule+0x282/0x870
May 19 11:37:25 curie kernel: [ 1571.559412]  ? __kmalloc_node+0x141/0x2b0
May 19 11:37:25 curie kernel: [ 1571.559417]  schedule+0x46/0xb0
May 19 11:37:25 curie kernel: [ 1571.559420]  schedule_preempt_disabled+0xa/0x10
May 19 11:37:25 curie kernel: [ 1571.559424]  __mutex_lock.constprop.0+0x133/0x460
May 19 11:37:25 curie kernel: [ 1571.559435]  ? nvlist_xalloc.part.0+0x68/0xc0 [znvpair]
May 19 11:37:25 curie kernel: [ 1571.559537]  spa_all_configs+0x41/0x120 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559644]  zfs_ioc_pool_configs+0x17/0x70 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559752]  zfsdev_ioctl_common+0x697/0x870 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559758]  ? _copy_from_user+0x28/0x60
May 19 11:37:25 curie kernel: [ 1571.559860]  zfsdev_ioctl+0x53/0xe0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.559866]  __x64_sys_ioctl+0x83/0xb0
May 19 11:37:25 curie kernel: [ 1571.559869]  do_syscall_64+0x33/0x80
May 19 11:37:25 curie kernel: [ 1571.559873]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
May 19 11:37:25 curie kernel: [ 1571.559876] RIP: 0033:0x7fcf0ef32cc7
May 19 11:37:25 curie kernel: [ 1571.559878] RSP: 002b:00007fcf0e181618 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
May 19 11:37:25 curie kernel: [ 1571.559881] RAX: ffffffffffffffda RBX: 000055b212f972a0 RCX: 00007fcf0ef32cc7
May 19 11:37:25 curie kernel: [ 1571.559883] RDX: 00007fcf0e181640 RSI: 0000000000005a04 RDI: 000000000000000b
May 19 11:37:25 curie kernel: [ 1571.559885] RBP: 00007fcf0e184c30 R08: 00007fcf08016810 R09: 00007fcf08000080
May 19 11:37:25 curie kernel: [ 1571.559886] R10: 0000000000080000 R11: 0000000000000246 R12: 000055b212f972a0
May 19 11:37:25 curie kernel: [ 1571.559888] R13: 0000000000000000 R14: 00007fcf0e181640 R15: 0000000000000000
May 19 11:37:25 curie kernel: [ 1571.559980] task:zpool           state:D stack:    0 pid:11815 ppid:  3816 flags:0x00004000
May 19 11:37:25 curie kernel: [ 1571.559983] Call Trace:
May 19 11:37:25 curie kernel: [ 1571.559988]  __schedule+0x282/0x870
May 19 11:37:25 curie kernel: [ 1571.559992]  schedule+0x46/0xb0
May 19 11:37:25 curie kernel: [ 1571.559995]  io_schedule+0x42/0x70
May 19 11:37:25 curie kernel: [ 1571.560004]  cv_wait_common+0xac/0x130 [spl]
May 19 11:37:25 curie kernel: [ 1571.560008]  ? add_wait_queue_exclusive+0x70/0x70
May 19 11:37:25 curie kernel: [ 1571.560118]  txg_wait_synced_impl+0xc9/0x110 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560223]  txg_wait_synced+0xc/0x40 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560325]  spa_export_common+0x4cd/0x590 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560430]  ? zfs_log_history+0x9c/0xf0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560537]  zfsdev_ioctl_common+0x697/0x870 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560543]  ? _copy_from_user+0x28/0x60
May 19 11:37:25 curie kernel: [ 1571.560644]  zfsdev_ioctl+0x53/0xe0 [zfs]
May 19 11:37:25 curie kernel: [ 1571.560649]  __x64_sys_ioctl+0x83/0xb0
May 19 11:37:25 curie kernel: [ 1571.560653]  do_syscall_64+0x33/0x80
May 19 11:37:25 curie kernel: [ 1571.560656]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
May 19 11:37:25 curie kernel: [ 1571.560659] RIP: 0033:0x7fdc23be2cc7
May 19 11:37:25 curie kernel: [ 1571.560661] RSP: 002b:00007ffc8c792478 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
May 19 11:37:25 curie kernel: [ 1571.560664] RAX: ffffffffffffffda RBX: 000055942ca49e20 RCX: 00007fdc23be2cc7
May 19 11:37:25 curie kernel: [ 1571.560666] RDX: 00007ffc8c792490 RSI: 0000000000005a03 RDI: 0000000000000003
May 19 11:37:25 curie kernel: [ 1571.560667] RBP: 00007ffc8c795e80 R08: 00000000ffffffff R09: 00007ffc8c792310
May 19 11:37:25 curie kernel: [ 1571.560669] R10: 000055942ca49e30 R11: 0000000000000246 R12: 00007ffc8c792490
May 19 11:37:25 curie kernel: [ 1571.560671] R13: 000055942ca49e30 R14: 000055942aed2c20 R15: 00007ffc8c795a40

Here's another example, where you see the USB controller bleeping out and back into existence:

mai 19 11:38:39 curie kernel: usb 2-1: USB disconnect, device number 2
mai 19 11:38:39 curie kernel: sd 4:0:0:0: [sdd] Synchronizing SCSI cache
mai 19 11:38:39 curie kernel: sd 4:0:0:0: [sdd] Synchronize Cache(10) failed: Result: hostbyte=DID_ERROR driverbyte=DRIVER_OK
mai 19 11:39:25 curie kernel: INFO: task zed:1564 blocked for more than 241 seconds.
mai 19 11:39:25 curie kernel:       Tainted: P          IOE     5.10.0-14-amd64 #1 Debian 5.10.113-1
mai 19 11:39:25 curie kernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
mai 19 11:39:25 curie kernel: task:zed             state:D stack:    0 pid: 1564 ppid:     1 flags:0x00000000
mai 19 11:39:25 curie kernel: Call Trace:
mai 19 11:39:25 curie kernel:  __schedule+0x282/0x870
mai 19 11:39:25 curie kernel:  ? __kmalloc_node+0x141/0x2b0
mai 19 11:39:25 curie kernel:  schedule+0x46/0xb0
mai 19 11:39:25 curie kernel:  schedule_preempt_disabled+0xa/0x10
mai 19 11:39:25 curie kernel:  __mutex_lock.constprop.0+0x133/0x460
mai 19 11:39:25 curie kernel:  ? nvlist_xalloc.part.0+0x68/0xc0 [znvpair]
mai 19 11:39:25 curie kernel:  spa_all_configs+0x41/0x120 [zfs]
mai 19 11:39:25 curie kernel:  zfs_ioc_pool_configs+0x17/0x70 [zfs]
mai 19 11:39:25 curie kernel:  zfsdev_ioctl_common+0x697/0x870 [zfs]
mai 19 11:39:25 curie kernel:  ? _copy_from_user+0x28/0x60
mai 19 11:39:25 curie kernel:  zfsdev_ioctl+0x53/0xe0 [zfs]
mai 19 11:39:25 curie kernel:  __x64_sys_ioctl+0x83/0xb0
mai 19 11:39:25 curie kernel:  do_syscall_64+0x33/0x80
mai 19 11:39:25 curie kernel:  entry_SYSCALL_64_after_hwframe+0x44/0xa9
mai 19 11:39:25 curie kernel: RIP: 0033:0x7fcf0ef32cc7
mai 19 11:39:25 curie kernel: RSP: 002b:00007fcf0e181618 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
mai 19 11:39:25 curie kernel: RAX: ffffffffffffffda RBX: 000055b212f972a0 RCX: 00007fcf0ef32cc7
mai 19 11:39:25 curie kernel: RDX: 00007fcf0e181640 RSI: 0000000000005a04 RDI: 000000000000000b
mai 19 11:39:25 curie kernel: RBP: 00007fcf0e184c30 R08: 00007fcf08016810 R09: 00007fcf08000080
mai 19 11:39:25 curie kernel: R10: 0000000000080000 R11: 0000000000000246 R12: 000055b212f972a0
mai 19 11:39:25 curie kernel: R13: 0000000000000000 R14: 00007fcf0e181640 R15: 0000000000000000
mai 19 11:39:25 curie kernel: INFO: task zpool:11815 blocked for more than 241 seconds.
mai 19 11:39:25 curie kernel:       Tainted: P          IOE     5.10.0-14-amd64 #1 Debian 5.10.113-1
mai 19 11:39:25 curie kernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
mai 19 11:39:25 curie kernel: task:zpool           state:D stack:    0 pid:11815 ppid:  2621 flags:0x00004004
mai 19 11:39:25 curie kernel: Call Trace:
mai 19 11:39:25 curie kernel:  __schedule+0x282/0x870
mai 19 11:39:25 curie kernel:  schedule+0x46/0xb0
mai 19 11:39:25 curie kernel:  io_schedule+0x42/0x70
mai 19 11:39:25 curie kernel:  cv_wait_common+0xac/0x130 [spl]
mai 19 11:39:25 curie kernel:  ? add_wait_queue_exclusive+0x70/0x70
mai 19 11:39:25 curie kernel:  txg_wait_synced_impl+0xc9/0x110 [zfs]
mai 19 11:39:25 curie kernel:  txg_wait_synced+0xc/0x40 [zfs]
mai 19 11:39:25 curie kernel:  spa_export_common+0x4cd/0x590 [zfs]
mai 19 11:39:25 curie kernel:  ? zfs_log_history+0x9c/0xf0 [zfs]
mai 19 11:39:25 curie kernel:  zfsdev_ioctl_common+0x697/0x870 [zfs]
mai 19 11:39:25 curie kernel:  ? _copy_from_user+0x28/0x60
mai 19 11:39:25 curie kernel:  zfsdev_ioctl+0x53/0xe0 [zfs]
mai 19 11:39:25 curie kernel:  __x64_sys_ioctl+0x83/0xb0
mai 19 11:39:25 curie kernel:  do_syscall_64+0x33/0x80
mai 19 11:39:25 curie kernel:  entry_SYSCALL_64_after_hwframe+0x44/0xa9
mai 19 11:39:25 curie kernel: RIP: 0033:0x7fdc23be2cc7
mai 19 11:39:25 curie kernel: RSP: 002b:00007ffc8c792478 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
mai 19 11:39:25 curie kernel: RAX: ffffffffffffffda RBX: 000055942ca49e20 RCX: 00007fdc23be2cc7
mai 19 11:39:25 curie kernel: RDX: 00007ffc8c792490 RSI: 0000000000005a03 RDI: 0000000000000003
mai 19 11:39:25 curie kernel: RBP: 00007ffc8c795e80 R08: 00000000ffffffff R09: 00007ffc8c792310
mai 19 11:39:25 curie kernel: R10: 000055942ca49e30 R11: 0000000000000246 R12: 00007ffc8c792490
mai 19 11:39:25 curie kernel: R13: 000055942ca49e30 R14: 000055942aed2c20 R15: 00007ffc8c795a40

I understand those are rather extreme conditions: I would fully expect the pool to stop working if the underlying drives disappear. What doesn't seem acceptable is that a command would completely hang like this.

References See the zfs documentation for more information about ZFS, and tubman for another installation and migration procedure.

Note: V3D driver is not present in Raspberry Pi models 0-3.

Raspberry Pi 4 64-bit (arm64)

Diff short summary:

1. instead of getting the .config file from bcm2711_defconfig, get it by running make ARCH=arm64 defconfig
2. compile and install the kernel image and modules as usual, but just copy the dtb file arch/arm64/boot/dts/broadcom/bcm2711-rpi-4-b.dtb to the /boot of your target machine (no /overlays directory to copy)
3. change /boot/config.txt:
   • comment/remove the dt_overlay=vc4-kms-v3d entry
   • add a device_tree=bcm2711-rpi-4-b.dtb entry

Raspberry Pi 4 32-bits (arm)

Diff short summary:

1. get the .config file by running make ARCH=arm multi_v7_defconfig
2. using make ARCH=arm menuconfig or a similar tool, enable CONFIG_ARM_LPAE=y
3. compile and install the kernel image and modules as usual, but just copy the dtb file arch/arm/boot/dts/bcm2711-rpi-4-b.dtb to the /boot of your target machine (no /overlays directory to copy)
4. change /boot/config.txt:
   • comment/remove the dt_overlay=vc4-kms-v3d entry
   • add a device_tree=bcm2711-rpi-4-b.dtb entry

Step-by-step for remote deployment:

Set variables Raspberry Pi 4 64-bit (arm64)

cd <path-to-upstream-linux-directory>
KERNEL= make kernelrelease 
ARCH="arm64"
CROSS_COMPILE="aarch64-linux-gnu-"
DEFCONFIG=defconfig
IMAGE=Image
DTB_PATH="broadcom/bcm2711-rpi-4-b.dtb"
RPI4=<ip>
TMP= mktemp -d 

Raspberry Pi 4 32-bits (arm)

cd <path-to-upstream-linux-directory>
KERNEL= make kernelrelease 
ARCH="arm"
CROSS_COMPILE="arm-linux-gnueabihf-"
DEFCONFIG=multi_v7_defconfig
IMAGE=zImage
DTB_PATH="bcm2711-rpi-4-b.dtb"
RPI4=<ip>
TMP= mktemp -d 

Get default .config file

make ARCH=$ARCH CROSS_COMPILE=$CROSS_COMPILE $DEFCONFIG

Raspberry Pi 4 32-bit (arm) Additional step for 32-bit system. Enable CONFIG_ARM_LPAE=y using make ARCH=arm menuconfig

Cross-compile the mainline kernel

make ARCH=$ARCH CROSS_COMPILE=$CROSS_COMPILE $IMAGE modules dtbs

Install modules to send

make ARCH=$ARCH CROSS_COMPILE=$CROSS_COMPILE INSTALL_MOD_PATH=$TMP modules_install

Copy kernel image, modules and the dtb to your remote system

ssh $RPI4 mkdir -p /tmp/new_modules /tmp/new_kernel
rsync -av $TMP/ $RPI4:/tmp/new_modules/
scp arch/$ARCH/boot/$IMAGE $RPI4:/tmp/new_kernel/$KERNEL
scp arch/$ARCH/boot/dts/$DTB_PATH $RPI4:/tmp/new_kernel
ssh $RPI4 sudo rsync -av /tmp/new_modules/lib/modules/ /lib/modules/
ssh $RPI4 sudo rsync -av /tmp/new_kernel/ /boot/
rm -rf $TMP

Set config.txt of you RPi 4 In your Raspberry Pi 4, open the config file /boot/config.txt

• comment/remove the dt_overlay=vc4-kms-v3d entry
• add a device_tree=bcm2711-rpi-4-b.dtb entry
• add a kernel=<image-name> entry

Why not Kworkflow? You can safely use the steps above, but if you are hacking the kernel and need to repeat this compiling and installing steps repeatedly, why don t try the Kworkflow? Kworkflow is a set of scripts to synthesize all steps to have a custom kernel compiled and installed in local and remote machines and it supports kernel building and deployment to Raspberry Pi machines for Raspbian 32 bit and 64 bit. After learning the kernel compilation and installation step by step, you can simply use kw bd command and have a custom kernel installed in your Raspberry Pi 4.

Note: the userspace 3D acceleration (glx/mesa) is working as expected on arm64, but the driver is not loaded yet for arm. Besides that, a bunch of pte invalid errors may appear when using 3D acceleration, it s a known issue that are still under investigation.

Changes in version 0.3.7 (2022-10-23)

• Update mkdocs for material docs generator (Dirk in #102)
• Use inherits() instead comparing to class() (Trevor Davis in #104)
• Set default arguments in nanoduration() (Trevor Davis in #105)
• Add as.nanoduration.difftime() support (Trevor Davis in #106)
• Add +/- methods for nanotime and difftime objects (Trevor Davis in #110 closing #108, #107)

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. Please report excessive re-aggregation in third-party for-profit settings.