On a blank example disk `/dev/sda` you can fulfill the requirements (One `EF00` partition with a file system plus one `BF00` partition) for example like so:
> `--new '1::+100M'`: Create partition number `1`. The field separator `:` separates the partition number from start sector. In this case start sector is unspecified so start sector sits at whatever the system's default is for this operation. On a blank disk on an Arch Linux live CD ISO image this will default to sector `2048`. Partition ends at whatever the beginning is `+100M` meaning plus 100 Mebibytes.
>
> `--new '2'`: Create partition number `2`. Both field number 2, the start sector, and field number 3, the end sector, are unspecified, there's no field separator `:`. Field number 2 will be the first free sector - in this case right after partition 1 - and field number 3 will be end of disk. Thus partition `2` will fill the remaining free disk space.
>
> `--typecode '1:EF00'`: Partition 1 gets partition type code `EF00`, an EFI system partition.
>
> `--typecode '2:BF00'`: Partition 2 gets partition type code `BF00`, an Solaris root partition.
The result will be something like this at which point you can start the `setup.sh` script, see [How to run this?](#how-to-run-this) below for more details.
The script will create a single ZFS zpool `zpool` on the `BF00` partition with dataset child `zpool/root` which itself has one child `zpool/root/archlinux`, that's where Arch Linux gets installed. Parallel to `zpool/root` it'll create `zpool/data` with a `zpool/data/home` child dataset that gets mounted at `/home`.
The script will use the `EF00` partition to install a ZFSBootMenu EFI executable if `efibootmgr` says that no such `ZFSBootMenu` entry exists. If ZFSBootMenu gets added to the EFI partition it'll become primary boot option.
During execution the script will call itself when it changes into its `chroot`, that's why we `export SCRIPT_URL`. Feel free to update `"${SCRIPT_URL}"` with whatever branch or revision you want to use from [quico.space/quico-os-setup/arch-zbm](https://quico.space/quico-os-setup/arch-zbm). Typically `.../branch/main/setup.sh` as shown above is what you want.
- Package manager hook: `pacman` does not have a hook to do ZFS snapshots
- See [this GitHub gist](https://gist.github.com/Soulsuke/6a7d1f09f7fef968a2f32e0ff32a5c4c#file-arch_on_zfs-txt-L238) and [zfs-snapshotter.bash](https://github.com/Soulsuke/arch-zfs-tools/blob/master/zfs-snapshotter.bash) for inspiration
- Hostname: Installation chose a pseudo-randomly generated 8-character string with `pwgen`
- Check `hostnamectl set-hostname <hostname>`
- Unprivileged user accounts: The OS was installed with `root` and unprivileged `build` users
- Passwords
- ZFS: The password for all datasets underneath `zpool` is `password`.
- Local `root` account: The local `root` account's password is `password`.
- Arch User Repository (AUR) helper: We installed [paru](https://github.com/Morganamilo/paru) as our AUR helper, we installed from GitHub via `makepkg -si`.
- In `/etc/systemd/network/50-wired.network` instead of a DHCP-based network config you can get a static one. The DHCP-based looks for reference looks like:
1. Change password in `keylocation` file, e.g. `/etc/zfs/zpool.key` or whatever other `"${zpool_name}"'.key'` file you used during setup
1. Set this key as the new encryption key:
```
zfs change-key -l zpool
```
Quoting `man 8 zfs-change-key` from `zfs-utils` version 2.1.9 for the `-l` argument: "Ensures the key is loaded before attempting to change the key." When successful the command will not output data, it'll just silently change your encryption key.
1. Rebuild initramfs:
```
mkinitcpio -P
```
Here for example with `-P` (`--allpresets`) which processes all presets contained in `/etc/mkinitcpio.d`. This step puts the changed key file into your initramfs. During setup we've adjusted `/etc/mkinitcpio.conf` so that it contains `FILES=(/etc/zfs/zpool.key)` which causes the file to be added to initramfs as-is.
## Boot flow
With your password changed in two locations (key file and initramfs) The boot process works as follows.
At boot time ZFSBootMenu will scan all pools that it can import for a `bootfs` property. If it only finds one pool with that property the dataset given as `bootfs` will be selected for boot with a 10-second countdown allowing manual interaction. With `bootfs` set ZFSBootMenu will not actively search through datasets for valid kernel and initramfs combinations, it'll instead accept `bootfs` as the default boot entry without us entering the pool decryption passphrase.
Upon loading into a given dataset ZFSBootMenu will attempt to auto-load the matching decryption key. In our setup this will fail because we purposely stored the encryption key inside our `zpool/root/archlinux` dataset. ZFSBootMenu will prompt us to type in the decryption key.
Lastly ZFSBootMenu loads our OS' kernel and initramfs combination via `kexec`. For this step we don't need to enter the decryption key again. Our initramfs file contains the plain-text `/etc/zfs/zpool.key` file which allows it to seamlessly import the right dataset, load its key and mount it.
ZFS differentiates between user keys - also called wrapping keys - and the master key for any given encryption root. You never interact with the master key, you only pick your personal user key. Subsequently a user key change (in our use case we perceive this simply as a password change) has zero effect on data that's already encrypted. The operation is instant and merely reencrypts the already existing master key, the so-called _wrapped_ master key.
ZFS generates the master key exactly once when you enable encryption on a dataset - technically when it becomes an encryption root. Among other inputs it uses your user key to encrypt (to _wrap_) the master key. When you change your user key it just means that the master key stays exactly the same and only the encrypted (_wrapped_) key changes.
`man 8 zfs-change-key` from `zfs-utils` version 2.1.9 adds:
> If the user's key is compromised, `zfs change-key` does not necessarily protect existing or newly-written data from attack. Newly-written data will continue to be encrypted with the same master key as the existing data. The master key is compromised if an attacker obtains a user key and the corresponding wrapped master key. Currently, `zfs change-key` does not overwrite the previous wrapped master key on disk, so it is accessible via forensic analysis for an indeterminate length of time.
>
> In the event of a master key compromise, ideally the drives should be securely erased to remove all the old data (which is readable using the compromised master key), a new pool created, and the data copied back. This can be approximated in place by creating new datasets, copying the data (e.g. using `zfs send | zfs recv`), and then clearing the free space with `zpool trim --secure` if supported by your hardware, otherwise `zpool initialize`.
On one hand changing the ZFS encryption password is generally a good and useful thing to do. On the other hand changing your password does not currently overwrite previous wrapped master keys on disk. A sufficiently motivated party that gains access to a wrapped master key and the matching user key is able to decrypt the master key and use it to read all data encrypted with it.
By extension this means after a password change your data remains at risk until you've copied it to a new dataset and erased previously used space thereby erasing any previous wrapped master keys.
In order to generate a new master key after you've changed your user key as mentioned in `man 8 zfs-change-key` from `zfs-utils` version 2.1.9 one example workflow goes like this:
- We specifically don't `zfs send -R` (`--replicate`). While it would normally be nice to transfer all of a dataset's children at once such as all of its snapshots the `-R` argument conflicts with the `encryption` property. See [comment by Tom Caputi on GitHub openzfs/zfs issue 10507 from June 2020](https://github.com/openzfs/zfs/issues/10507#issuecomment-651162104) for details. Basically if `encryption` is set then `-R` doesn't work. We could transfer existing encryption properties with `-w`/`--raw` but we don't actually want to transfer encryption properties at all. We want them to change during transfer, see the bullet point four points down from here talking about `encryption`.
- We `zfs receive -F` destroying any target snapshots and file systems beyond the snapshot we're transferring. In this example the target `zpool/root/archlinux-frn` doesn't even exist so `-F` isn't necessary to clean anything up. It's just good practice.
- With `-v` we get verbose progress output
- Argument `-u` makes sure the dataset does not get mounted after transfer. ZFS would mount it into `/` which wouldn't be helpful since we're currently using that filesystem ourselves.
- We set encryption properties `keyformat`, `keylocation` and most importantly `encryption`. The latter will turn our transferred dataset into its own `encryptionroot` which in turn generates a new master key. The auto-generated new master key gets wrapped with our updated passphrase in `keylocation`. This basically reencrypts all data in this dataset during transfer.
The parent `zpool/root` is inheriting this property from `zpool` which will make sure that `zpool/root/archlinux-frn` essentially gets its key now from `zpool`. Both `zpool/root/archlinux-frn` and `zpool` use the same exact `keylocation` with identical content. This operation is instant.
Repeat for source dataset `zpool/root/archlinux-sxu@rekey`. You're particularly interested in parameters `DSL_CRYPTO_MASTER_KEY_1` and the initialization vector `DSL_CRYPTO_IV`. Notice that they differ between old and new dataset confirming that your new dataset has a new master key.
Next up unmap/TRIM unallocated disk areas. If your zpool runs on an entire disk and not just on a partition, and if your disk supports TRIM you're going to want to do:
The ZFS pool and dataset setup that makes this tick, explained in plain English.
1. Create zpool with options:
1.`-R /mnt` (aka `-o cachefile=none -o altroot=/mnt`). The pool is never cached, i.e. it's considered temporary. All pool and dataset mount paths have `/mnt` prepended. From `man zpoolprops`:
> This can be used when examining an unknown pool where the mount points cannot be trusted, or in an alternate boot environment, where the typical paths are not valid. `altroot` is not a persistent property. It is valid only while the system is up.
1.`-O canmount=off`: Note the capital `-O` which makes this a file system property, not a pool property. File system cannot be mounted, and is ignored by `zfs mount -a`. This property is not inherited.
1.`-O mountpoint=none`: What it says on the tin, the pool has no mountpoint configured.
1.`-O encryption=on`: Makes this our `encryptionroot` and passes the `encryption` setting to all child datasets. Selecting `encryption=on` when creating a dataset indicates that the default encryption suite will be selected, which is currently `aes-256-gcm`.
1.`-O keylocation=file://...`: This property is only set for encrypted datasets which are encryption roots. Controls where the user's encryption key will be loaded from by default for commands such as `zfs load-key`.
1.`-O keyformat=passphrase`: Controls what format the user's encryption key will be provided as. Passphrases must be between 8 and 512 bytes long.
1. At this time the newly created zpool is not mounted anywhere. Next we create the "root" dataset, that's an arbitary term for the parent dataset of all boot environments. Boot environments in your case may be for example different operating systems all of which live on separate datasets underneath the root.
1.`-o mountpoint=none`: Same as above, the root dataset has - just like the pool - no mountpoint configured.
1.`zfs set org.zfsbootmenu:commandline=...`: Set a common kernel command line for all boot environment such as `"ro quiet"`.
1. Neither the root dataset nor the pool are mounted at this time. We now create one boot environment dataset where we want to install Arch Linux.
1.`-o mountpoint=/`: Our Arch Linux dataset will be mounted at `/`.
1.`-o canmount=noauto`: When set to `noauto`, a dataset can only be mounted and unmounted explicitly. The dataset is not mounted automatically when the dataset is created or imported, nor is it mounted by the `zfs mount -a` command or unmounted by the `zfs unmount -a` command.
1. We then `zpool set bootfs="zpool/root/archlinux" zpool`: ZFSBootMenu uses the `bootfs` property to identify suitable boot environments. If only one pool has it - as is the case here - it identifies the pool's preferred boot dataset that will be booted with a 10-second countdown allowing manual interaction in ZFSBootMenu.
1. We explicitly mount the boot environment. Since the entire pool is still subject to our initial `-R /mnt` during creation a `zfs mount zpool/root/archlinux` will mount the Arch Linux dataset not into `/` but instead into `/mnt`.
1. We also create a `data` dataset that - at least for now - we use to store only our `/home` data.
1. For `zpool/data`:
1.`-o mountpoint=/`: We use the `mountpoint` property here only for inheritance.
1.`-o canmount=off`: The `zpool/data` dataset itself cannot actually be mounted.
1. For a `zpool/data/home` child dataset:
1. We do not specify any properties. Since `canmount` cannot be inherited the parent's `canmount=off` does not apply, it instead defaults to `canmount=on`. The parent's `mountpoint=/` property on the other hand is inherited so for a `home` child dataset it conveniently equals `mountpoint=/home`.
1. In effect this `zpool/data/home` dataset is subject to `zfs mount -a` and will happily automount into `/home`.
1. We export the zpool once, we then reimport it by scanning only inside `/dev/disk/by-id`, again setting `-R /mnt` as we did during pool creation a moment ago and we do not mount any file systems.
1. We `zfs load-key <encryptionroot>` which will load the key from `keylocation` after which the `keystatus` property for `<encryptionroot>` and all child datasets will change from `unavailable` to `available`.
1. We mount our Arch Linux boot environment dataset. It automatically get prepended with `-R /mnt` since that's how we imported the pool.
1. We `zfs mount -a` which automounts `zpool/data/home` into `/home`, which again gets auto-prepended by `/mnt`.
1. We lastly mount our EFI partition into `/mnt/efi`.
1. We instruct ZFS to save its pool configuration via `zpool set cachefile=/etc/zfs/zpool.cache zpool`.
The complete ZFS structure now exists and is mounted at `/mnt` ready for any `pacstrap`, [debootstrap](https://wiki.debian.org/Debootstrap), `dnf --installroot` or other bootstrapping action.
