ZFS Load-Key on Boot

Bluefin
1

Question

What is the proper way of making zfs load-key on boot where the encrypted file is located in /root?

I’ve used a similar method for Debian which has worked perfectly but I want to make sure this is still the proper way of doing it in Bluefin.

Background

I am new to Bluefin, I was previously running on Debian for several years. The concept of an atomic OS is still very new to me.

I have bluefin-dx-nvidia-open:stable installed on 2 x 500GB nvme in RAID 0 with encryption (Using the installer’s automatic partitionner) but I also have 4 more drives that I’ve configured in an encrypted Raidz1.

My issue is that while the pool gets imported correctly, the encryption key does not get loaded, which prevents the zfs datasets from being mounted.

Currently, I can run these 2 commands after logging-in to resolve the issue:

sudo zfs load-key -a
sudo zfs mount -a

My goal is to automate this the right way to avoid any future updates of bluefin to override these changes.

1 Like
2

Likely the pool is being imported prior to your /var being mounted.

It then cannot fetch the key to decrypt the dataset and automount it.

You could convert to legacy mounts and specify the dependency on /var

Or…

You could modify the load-key service for your encrypted dataset to run after /var is mounted

3

Thank you for your response!

Looking at all the files in /etc/systemd/system/zfs.target.wants/, none of them performs the “zfs load-key -a” command.

So there doesn’t seem to be a built-in way at the moment to automatically load the keys needed to mount the zfs datasets.

Would the solution of creating a new file zfs-load-key.service in /etc/systemd/system/zfs.target.wants/ be the right way of doing it in Bluefin?

4

I think I would put the key in a separate partition using FAT or ext4 so it can be mounted first and more directly. Seems less risky to me.

My $0.02.

5

I did consider it but I’m not too sure how to implement that either, I guess I would have to setup an automatic mount in fstab or something and have the key in there.

My understanding is that ZFS will import pools and mount them after the primary disk is decrypted. I’ve done further digging and found similar instructions on the ArchWiki to automate the zfs load-key function (Section 6.1.1).

This leads me to believe that ZFS doesn’t attempt to load encryption keys by default, so adding this Systemd seems like the best way.

As far as where to make that change, I rechecked the bluefin documentation and found a section in the FAQ:

Follow the XDG standards for overriding core OS files. The /etc, /var, /usr/local, and /opt directories are writable, and many applications will look here for overrides of the read-only files shipped with the OS image in /usr

So I guess my conclusion is that adding the file in /etc/systemd/system/ would be the appropriate way to override this behavior without changing a part of the system that might get replaced by a future update.

I’ll give it a try later and see if it works or not.

2 Likes
6

Please do let us know what you landed on here.

I tend do agree with the systemd unit approach as well. That could be placed in $HOME so it is backed up. Less to track for reinstalls later.

7

That could be placed in $HOME so it is backed up. Less to track for reinstalls later.

Interesting! How would I go about doing that? The only way I know of would be a symlink.

8

Zed should be making a service file and mount file per dataset.

Take a look in /run/systemd.

But you would make a drop in based on the specific dataset you need to be after var.mount.

9

Place the unti file in ~/.config/systemd/user/*. Not the best reference but what I could find quickly. But it shows the resolution order early in the page.

That way the systemd unit files are in $HOME and you could place the mount points somewhere in there too. E.g., ~/mnt or ~/.local/mnt.

In my case, I would modify my ~/.local/bin/backup-home.sh script to include or exclude mount points as needed from the rsync command line.

But the unit files would be in ~/.config/systemd and would be backed up. This, of course, is to simplify the process of reinstall.

I have been working hard on reducing my manual post-install checklist. Approaches like this is doing it for me.

In short, thanks to j0rge, infy, mgiles, etc. who have been helping adjust my mental model of how to manage and use a linux dev workstation, I have reached this simple conclusion.

The OS is there to enable what I do in my home directory. Any change outside of my home directory needs to be highly scrutinized and avoided where possible.

10

Thank you @m2Giles and @klmcw for your answers on this topic.

I’m happy to report that I was able to resolve my issue by creating the unit file in /etc/systemd/system (like suggested in Blog Mounting an encrypted ZFS dataset at boot on Debian 11 Bullseye :: Alasdair Keyes :: www.akeyes.co.uk) and it’s been working fine ever since.

I did look some more into @klmcw’s suggestion but I was unable to make it work. After some more googling, I was able to figure out that this particular unit cannot be installed as a user unit because it depends on some other zfs system units.

From another great article on ArchWiki:

systemd --user runs as a separate process from the systemd --system process. User units can not reference or depend on system units or units of other users.

Big thank you to everybody that helped me with this.

2 Likes
11

Fair enough. The important point is that you tried to solve the problem by adding files in your HOME dir where it can be easily backed up.

I do NOT backup my whole desktop. Only my HOME dir because I assume I can easily reinstall.

Because you had to monkey with /etc, please add these instructions to your post-install checklist. Just in case! Stuff happens.

Glad to hear you were able to solve the issue.