dm-crypt/Specialties

Securing the unencrypted boot partition

The /boot partition and the Master Boot Record are the two areas of the disk that are not encrypted, even in an encrypted root configuration. They cannot usually be encrypted because the boot loader and BIOS (respectively) are unable to unlock a dm-crypt container in order to continue the boot process. An exception is GRUB, which gained a feature to unlock a LUKS encrypted /boot - see dm-crypt/Encrypting an entire system#Encrypted boot partition (GRUB).

This section describes steps that can be taken to make the boot process more secure.

Warning: Note that securing the /boot partition and MBR can mitigate numerous attacks that occur during the boot process, but systems configured this way may still be vulnerable to BIOS/UEFI/firmware tampering, hardware keyloggers, cold boot attacks, and many other threats that are beyond the scope of this article. For an overview of system-trust issues and how these relate to full-disk encryption, refer to [1].

Note: When using UEFI, only an ESP needs to remain unencrypted, if using Unified kernel images for example. You can then use Secure Boot to make sure the boot chain has not been tampered with, which is an easy way to achieve the same result as the below methods. See dm-crypt/Encrypting an entire system#LUKS on a partition with TPM2 and Secure Boot.

Booting from a removable device

Using a separate device to boot a system is a fairly straightforward procedure, and offers a significant security improvement against some kinds of attacks. Two vulnerable parts of a system employing an encrypted root filesystem are

• the Master Boot Record, and
• the /boot partition.

These must be stored unencrypted in order for the system to boot. In order to protect these from tampering, it is advisable to store them on a removable medium, such as a USB drive, and boot from that drive instead of the hard disk. As long as you keep the drive with you at all times, you can be certain that those components have not been tampered with, making authentication far more secure when unlocking your system.

It is assumed that you already have your system configured with a dedicated partition mounted at /boot. If you do not, please follow the steps in dm-crypt/System configuration#Kernel parameters, substituting your hard disk for a removable drive.

Note: You must make sure your system supports booting from the chosen medium, be it a USB drive, an external hard drive, an SD card, or anything else.

Prepare the removable drive (/dev/sdx).

# gdisk /dev/sdx #format if necessary. Alternatively, cgdisk, fdisk, cfdisk, gparted...
# mkfs.ext2 /dev/sdx1 #for BIOS systems
# mkfs.fat -F 32 /dev/sdx1 #for UEFI systems
# mount /dev/sdx1 /mnt

Copy your existing /boot contents to the new one.

# cp -ai /boot/* /mnt/

Mount the new partition. Do not forget to update your fstab file accordingly.

# umount /boot
# umount /mnt
# mount /dev/sdx1 /boot
# genfstab -p -U / > /etc/fstab

Update GRUB. grub-mkconfig should detect the new partition UUID automatically, but custom menu entries may need to be updated manually.

# grub-mkconfig -o /boot/grub/grub.cfg
# grub-install /dev/sdx #install to the removable device, not the hard disk. for BIOS systems
# grub-install --target=x86_64-efi --efi-directory=/boot --bootloader-id=grub #for UEFI systems

Reboot and test the new configuration. Remember to set your device boot order accordingly in your BIOS or UEFI. If the system fails to boot, you should still be able to boot from the hard drive in order to correct the problem.

chkboot

Warning: chkboot makes a /boot partition tamper-evident, not tamper-proof. By the time the chkboot script is run, you have already typed your password into a potentially compromised boot loader, kernel, or initrd. If your system fails the chkboot integrity test, no assumptions can be made about the security of your data.

Referring to an article from the ct-magazine (Issue 3/12, page 146, 01.16.2012, [2]) the following script checks files under /boot for changes of SHA-1 hash, inode, and occupied blocks on the hard drive. It also checks the Master Boot Record. The script cannot prevent certain type of attacks, but a lot are made harder. No configuration of the script itself is stored in unencrypted /boot. With a locked/powered-off encrypted system, this makes it harder for some attackers because it is not apparent that an automatic checksum comparison of the partition is done upon boot. However, an attacker who anticipates these precautions can manipulate the firmware to run their own code on top of your kernel and intercept file system access, e.g. to boot, and present the untampered files. Generally, no security measures below the level of the firmware are able to guarantee trust and tamper evidence.

The script with installation instructions is available (Author: Juergen Schmidt, ju at heisec.de; License: GPLv2). There is also package chkboot^AUR to install. The AUR package is recommended, as it has additional helpful scripts.

Juergen Schmidt's script

As /usr/local/bin/chkboot_user.sh needs to be executed right after login, you need to add it to the autostart (e.g. under KDE -> System Settings -> Startup and Shutdown -> Autostart; GNOME 3: gnome-session-properties).

With Arch Linux, changes to /boot are pretty frequent, for example by new kernels rolling-in. Therefore it may be helpful to use the scripts with every full system update. One way to do so:

#!/bin/sh
#
# Note: Insert your <user> and execute it with sudo for pacman & chkboot to work automagically
#
echo "Pacman update [1] Quickcheck before updating" &
sudo -u <user> /usr/local/bin/chkboot_user.sh
/usr/local/bin/chkboot.sh
sudo -u <user> /usr/local/bin/chkboot_user.sh
echo "Pacman update [2] Syncing repos for pacman"
pacman -Syu
/usr/local/bin/chkboot.sh
sudo -u <user> /usr/local/bin/chkboot_user.sh
echo "Pacman update [3] All done, let us roll on ..."

AUR package

After installing, enable chkboot.service.

You may want to add chkboot to the end of your mkinitcpio hooks, so that your chkboot hashes get updated every time mkinitcpio regenerates your initramfs. You can do this by adding chkboot to the end of the HOOKS array in /etc/mkinitcpio.conf.

The AUR package also comes with an chkboot-desktopalert script, which will cause a graphical window to pop up with a warning if /boot changes are detected. You can make use of this script by adding it to the startup scripts of your graphical environment.

mkinitcpio-chkcryptoboot

Warning: This hook does not encrypt GRUB's core (MBR) code or EFI stub, nor does it protect against situations where an attacker is able to modify the behaviour of the boot loader to compromise the kernel and/or initramfs at run-time.

mkinitcpio-chkcryptoboot^AUR is a mkinitcpio hook that performs integrity checks during early-userspace and advises the user not to enter their root partition password if the system appears to have been compromised. Security is achieved through an encrypted boot partition, which is unlocked using GRUB's cryptodisk.mod module, and a root filesystem partition, which is encrypted with a password different from the former. This way, the initramfs and kernel are secured against offline tampering, and the root partition can remain secure even if the /boot partition password is entered on a compromised machine (provided that the chkcryptoboot hook detects the compromise, and is not itself compromised at run-time).

This hook requires grub release >=2.00 to function, and a dedicated, LUKS encrypted /boot partition with its own password in order to be secure.

Installation

Install mkinitcpio-chkcryptoboot^AUR and edit /etc/default/chkcryptoboot.conf. If you want the ability of detecting if your boot partition was bypassed, edit the CMDLINE_NAME and CMDLINE_VALUE variables, with values known only to you. You can follow the advice of using two hashes as is suggested right after the installation. Also, be sure to make the appropriate changes to the kernel command line in /etc/default/grub. Edit the HOOKS= line in /etc/mkinitcpio.conf, and insert the chkcryptoboot hook before encrypt. When finished, regenerate the initramfs.

Technical overview

mkinitcpio-chkcryptoboot^AUR consists of an install hook and a run-time hook for mkinitcpio. The install hook runs every time the initramfs is rebuilt, and hashes the GRUB EFI stub ($esp/EFI/grub_uefi/grubx64.efi) (in the case of UEFI systems) or the first 446 bytes of the disk on which GRUB is installed (in the case of BIOS systems), and stores that hash inside the initramfs located inside the encrypted /boot partition. When the system is booted, GRUB prompts for the /boot password, then the run-time hook performs the same hashing operation and compares the resulting hashes before prompting for the root partition password. If they do not match, the hook will print an error like this:

CHKCRYPTOBOOT ALERT!
CHANGES HAVE BEEN DETECTED IN YOUR BOOT LOADER EFISTUB!
YOU ARE STRONGLY ADVISED NOT TO ENTER YOUR ROOT CONTAINER PASSWORD!
Please type uppercase yes to continue:

In addition to hashing the boot loader, the hook also checks the parameters of the running kernel against those configured in /etc/default/chkcryptoboot.conf. This is checked both at run-time and after the boot process is done. This allows the hook to detect if GRUB's configuration was not bypassed at run-time and afterwards to detect if the entire /boot partition was not bypassed.

For BIOS systems the hook creates a hash of GRUB's first stage boot loader (installed to the first 446 bytes of the bootdevice) to compare at the later boot processes. The main second-stage GRUB boot loader core.img is not checked.

AIDE

Alternatively to above scripts, a hash check can be set up with AIDE which can be customized via a very flexible configuration file.

STARK

While one of these methods should serve the purpose for most users, they do not address all security problems associated with the unencrypted /boot. One approach which endeavours to provide a fully authenticated boot chain was published with POTTS as an academic thesis to implement the STARK authentication framework.

The POTTS proof-of-concept uses Arch Linux as a base distribution and implements a system boot chain with:

• POTTS - a boot menu for a one-time authentication message prompt
• TrustedGrub - a GRUB Legacy implementation which authenticates the kernel and initramfs against TPM chip PCR registers
• TRESOR - a kernel patch which implements AES but keeps the master-key not in RAM but in CPU registers during runtime.

As part of the thesis installation instructions based on Arch Linux (ISO as of 2013-01) have been published. If you want to try it, be aware these tools are not in standard repositories and the solution will be time consuming to maintain.

Using GPG, LUKS, or OpenSSL encrypted keyfiles

The following forum posts give instructions to use two factor authentication, gpg or openssl encrypted keyfiles, instead of a plaintext keyfile described earlier in this wiki article System Encryption using LUKS with GPG encrypted keys:

• GnuPG: Post regarding GPG encrypted keys This post has the generic instructions.
• OpenSSL: Post regarding OpenSSL encrypted keys This post only has the ssldec hooks.
• OpenSSL: Post regarding OpenSSL salted bf-cbc encrypted keys This post has the bfkf initcpio hooks, install, and encrypted keyfile generator scripts.
• LUKS: Post regarding LUKS encrypted keys with a lukskey initcpio hook. Or #Encrypted /boot and a detached LUKS header on USB below with a custom encrypt hook for initcpio.

Note that:

• You can follow the above instructions with only two primary partitions, one boot partition (required because of encryption) and one primary LVM partition. Within the LVM partition you can have as many partitions as you need, but most importantly it should contain at least root, swap, and home logical volume partitions. This has the added benefit of having only one keyfile for all your partitions, and having the ability to hibernate your computer (suspend to disk) where the swap partition is encrypted. If you decide to do so your hooks in /etc/mkinitcpio.conf should look like this:

  HOOKS=( ... usb usbinput (etwo or ssldec) encrypt (if using openssl) lvm2 resume ... )

  and you should add

  resume=/dev/<VolumeGroupName>/<LVNameOfSwap>

  to your kernel parameters.
• If you need to temporarily store the unencrypted keyfile somewhere, do not store them on an unencrypted disk. Even better make sure to store them to RAM such as /dev/shm.
• If you want to use a GPG encrypted keyfile, you need to use a statically compiled GnuPG version 1.4 or you could edit the hooks and use gnupg1^AUR
• It is possible that an update to OpenSSL could break the custom ssldec mentioned in the second forum post.

Remote unlocking of root (or other) partition

Imagine a system with one or more LUKS encrypted partitions (root or others) or volumes and you need to unlock these partitions/volumes during startup. In this case you need to be able to log in remotely and provide a password at the early boot phase. This can be achieved by using one or more mkinitcpio hook(s) that configure a network interface and start some kind of SSH server. Some packages listed below contribute various mkinitcpio build hooks to ease the configuration. The following tutorials add a remote unlocking method in addition to the existing local console password prompt.

Note:

• Keep in mind to use kernel device names for the network interface (e.g. eth0) and not udev's ones (e.g. enp1s0), as those will not work.
• By default, Predictable Network Interface Names are activated and change the kernel device name during late boot. Use dmesg and look what your Network kernel module does to find the original name (e.g. eth0)
• It could be necessary to add the module for your ethernet or wireless network card to the MODULES array.

Busybox based initramfs (built with mkinitcpio)

For busybox based initramfs the packages mkinitcpio-netconf and/or mkinitcpio-ppp^AUR provide network connectivity. As SSH server you have the option of using either mkinitcpio-dropbear or mkinitcpio-tinyssh. Those hooks do not install any shell, so you also need to install the mkinitcpio-utils package. The instructions below can be used in any combination of the packages above. When there are different paths, it will be noted.

1. If you do not have an SSH key pair yet, generate one on the client system (the one which will be used to unlock the remote machine).
   Note:

   • tinyssh only supports Ed25519 and ECDSA key types without passphrase. If you chose to use mkinitcpio-tinyssh, you need to create/use one of these.
   • mkinitcpio-tinyssh currently has a bug affecting it's use of tinyssh-convert. See Github for details and a fix.
   • mkinitcpio-dropbear in version 0.0.3-5 is not compatible with the current dropbear implementation that removed dss. See Github for details and a fix.
3. Insert your SSH public key (i.e. the one you usually put onto hosts so that you can ssh in without a password, or the one you just created and which ends with .pub) into the remote machine's /etc/dropbear/root_key or /etc/tinyssh/root_key file.
   Tip: This method can later be used to add other SSH public keys as needed; In the case of simply copying the content of the remote's ~/.ssh/authorized_keys, be sure to verify that it only contains keys you intend to be using to unlock the remote machine. When adding additional keys, regenerate your initrd as well using mkinitcpio. See also OpenSSH#Protection.
4. Add all three <netconf and/or ppp> <dropbear or tinyssh> encryptssh hooks before filesystems within the "HOOKS" array in /etc/mkinitcpio.conf (the encryptssh replaces the encrypt hook). Then regenerate the initramfs.
   Note: The net hook provided by mkinitcpio-nfs-utils is not needed.
5. Configure the required cryptdevice= parameter and add the ip= kernel command parameter to your boot loader configuration with the appropriate arguments. For example, if you want the IP address to be assigned by DHCP, you can use the following parameter:

   ip=dhcp

   This is most useful if your DHCP server is configured to always assign the same IP for this host, as otherwise accessing the system via SSH is going to be difficult. In that case, you can use a static IP:

   ip=192.168.1.1:::::eth0:none

   Alternatively, you can also specify the subnet mask and gateway required by the network:

   ip=192.168.1.1::192.168.1.254:255.255.255.0::eth0:none

   Note: As of version 0.0.4 of mkinitcpio-netconf, you can nest multiple ip= parameters in order to configure multiple interfaces. You cannot mix it with ip=dhcp (ip=:::::eth0:dhcp) alone. An interface needs to be specified.

   ip=ip=192.168.1.1:::::eth0:none:ip=172.16.1.1:::::eth1:none

   If using DHCP, consider adding the netconf_timeout= kernel parameter, to prevent netconf from trying to obtain an IP forever. For a detailed description of the ip= parameter, have a look at the according mkinitcpio section. When finished, update the configuration of your boot loader.
6. Finally, restart the remote system and try to ssh to it, explicitly stating the "root" username (even if the root account is disabled on the machine, this root user is used only in the initrd for the purpose of unlocking the remote system). If you are using the mkinitcpio-dropbear package and you also have the openssh package installed, then you will probably not get any warnings before logging in, because it converts and uses the same host keys openssh uses (except Ed25519 keys, as dropbear does not support them). In the case you are using mkinitcpio-tinyssh, a script tinyssh-convert is bundled, so you can use the same keys as your openssh installation (currently only Ed25519 keys). It may be required to manually copy the host key, via tinyssh-convert, to /etc/tinyssh/sshkeydir. In either case, you should have run the ssh daemon at least once, using the provided systemd units, so the keys can be generated first. After rebooting the machine and connecting via ssh, you should be prompted for the passphrase to unlock the device, after which the system will complete its boot process and you can then ssh to it as you normally would (with the remote user of your choice).

Enabling Wi-Fi

The netconf hook is normally used with an Ethernet connection. In case you want to setup a computer with wireless only, and unlock it via Wi-Fi, you can use a predefined hook or create a custom hook to connect to a Wi-Fi network before the netconf hook is run.

Predefined hook

You can install a predefined hook based on the one in this wiki:

1. Install mkinitcpio-wifi^AUR.
2. Configure your Wi-Fi connection by creating a wpa_supplicant configuration with your network properties:

   wpa_passphrase "ESSID" "passphrase" > /etc/wpa_supplicant/initcpio.conf

3. Add the wifi hook before netconf in your /etc/mkinitcpio.conf. Your Wi-Fi-related modules should be auto-detected, if not: add them to the MODULES section.
4. Add ip=:::::wlan0:dhcp to the kernel parameters.
5. Regenerate the initramfs.
6. Update the configuration of your boot loader.

Custom hook

Below example shows a setup using a USB Wi-Fi adapter, connecting to a Wi-Fi network protected with WPA2-PSK. In case you use for example WEP or another initramfs generator, you might need to adjust accordingly.

1. Modify /etc/mkinitcpio.conf:
   • Add the needed kernel module for your specific Wi-Fi adapter.
   • Include the wpa_passphrase and wpa_supplicant binaries.
   • Add a hook wifi (or a name of your choice, this is the custom hook that will be created) before the net hook.

     MODULES=(module)
     BINARIES=(wpa_passphrase wpa_supplicant)
     HOOKS=(base udev autodetect ... wifi net ... dropbear encryptssh ...)

2. Create the wifi hook in /etc/initcpio/hooks/wifi:

   run_hook ()
   {
   	# Sleep a couple of seconds so wlan0 is setup by kernel
   	sleep 5
   
   	# Set wlan0 to up
   	ip link set wlan0 up
   
   	# Associate with Wi-Fi network
   	# 1. Save temp config file
   	wpa_passphrase "network ESSID" "pass phrase" > /tmp/wifi
   
   	# 2. Associate
   	wpa_supplicant -B -D nl80211,wext -i wlan0 -c /tmp/wifi
   
   	# Sleep a couple of seconds so that wpa_supplicant finishes connecting
   	sleep 5
   
   	# wlan0 should now be connected and ready to be assigned an ip by the net hook
   }
   
   run_cleanuphook ()
   {
   	# Kill wpa_supplicant running in the background
   	killall wpa_supplicant
   
   	# Set wlan0 link down
   	ip link set wlan0 down
   
   	# wlan0 should now be fully disconnected from the Wi-Fi network
   }

3. Create the hook installation file in /etc/initcpio/install/wifi:

   build ()
   {
   	add_runscript
   }
   help ()
   {
   cat<<HELPEOF
   	Enables Wi-Fi on boot, for dropbear ssh unlocking of disk.
   HELPEOF
   }

4. Add ip=:::::wlan0:dhcp to the kernel parameters. Remove ip=:::::eth0:dhcp so it does not conflict.
5. Optionally create an additional boot entry with kernel parameter ip=:::::eth0:dhcp.
6. Regenerate the initramfs.
7. Update the configuration of your boot loader.

Remember to setup Wi-Fi, so you are able to login once the system is fully booted. In case you are unable to connect to the Wi-Fi network, try increasing the sleep times a bit.

systemd based initramfs (built with mkinitcpio)

For systemd based initramfs the AUR package mkinitcpio-systemd-extras^AUR provides a collection of build hooks (aka install hooks) to achieve network connectivity and SSH login during early boot. Depending on the concrete setup this either gives you access to the initramfs environment via busybox' dash or just a password prompt.

A minimal setup:

/etc/mkinitcpio.conf

HOOKS=(base systemd autodetect microcode modconf kms keyboard sd-vconsole sd-network block sd-tinyssh sd-encrypt filesystems fsck)
SD_TINYSSH_COMMAND="systemd-tty-ask-password-agent --query --watch"

When building the initramfs with mkinitcpio this setup copies the already existing configuration of systemd-networkd from the main system and also tries to copy / convert existing SSH server keys from an existing TinySSH or OpenSSH installation. tinyssh needs to be installed (but not necessarily enabled) on the main system. There are additional configuration parameters in case systemd-networkd is not used by the main system. See the documentation of mkinitcpio-systemd-extras^AUR for further details.

Note:

• Predictable network interface names such as enp2s0 are not present during initramfs phase by default, so many systemd-networkd configuration files will not work during initramfs phase. See the mkinitcpio-systemd-extras wiki for workarounds.
• This section used to mention mkinitcpio-systemd-tool as another alternative to achieve remote login and LUKS unlocking during early startup. The approach is completely different as it requires only one additional hook systemd-tool and all further setup is done in special [X-SystemdTool] sections in systemd service files. Unfortunately, documentation about this approach and how to use the tool itself is very limited.

systemd based initramfs (built with dracut)

If you are using dracut instead of mkinitcpio, you might want to check out dracut-sshd as an alternative to the above options.

One-time password-less reboot

Another method that can be used to reboot a remote, headless or otherwise inaccessible system whilst not needing to be at the terminal to type the encrypted root drive password, is to use a temporary keyfile. This will need to be placed in a location that is accessible to the kernel at boot, the cryptkey boot parameter will be needed, and that particular keyfile will need to be registered as a valid key by way of the "cryptsetup luksAddKey" command.

This can be done conveniently with the help of passless-boot^AUR. The procedure described to setup that tool on the script's readme file might serve as a template for setting up a home-made solution also. Do take a look at the discussion in the Security considerations section.

Discard/TRIM support for solid state drives (SSD)

Solid state drive users should be aware that, by default, TRIM commands are not enabled by the device-mapper, i.e. block-devices are mounted without the discard option unless you override the default.

The device-mapper maintainers have made it clear that TRIM support will never be enabled by default on dm-crypt devices because of the potential security implications.[3][4] Minimal data leakage in the form of freed block information, perhaps sufficient to determine the filesystem in use, may occur on devices with TRIM enabled. An illustration and discussion of the issues arising from activating TRIM is available in the blog of a cryptsetup developer. If you are worried about such factors, keep also in mind that threats may add up: for example, if your device is still encrypted with the previous (cryptsetup <1.6.0) default cipher --cipher aes-cbc-essiv, more information leakage may occur from trimmed sector observation than with the current default.

The following cases can be distinguished:

• The device is encrypted with default dm-crypt LUKS mode:
  • By default the LUKS header is stored at the beginning of the device and using TRIM is useful to protect header modifications. If for example a compromised LUKS password is revoked, without TRIM the old header will in general still be available for reading until overwritten by another operation; if the drive is stolen in the meanwhile, the attackers could in theory find a way to locate the old header and use it to decrypt the content with the compromised password. See cryptsetup FAQ, section 5.19 What about SSDs, Flash and Hybrid Drives? and Full disk encryption on an ssd.
  • TRIM can be left disabled if the security issues stated at the top of this section are considered a worse threat than the above bullet.

See also Securely wipe disk#Flash memory.

• The device is encrypted with dm-crypt plain mode, or the LUKS header is stored separately:
  • If plausible deniability is required, TRIM should never be used because of the considerations at the top of this section, or the use of encryption will be given away.
  • If plausible deniability is not required, TRIM can be used for its performance gains, provided that the security dangers described at the top of this section are not of concern.

Warning: Before enabling TRIM on a drive, make sure the device fully supports TRIM commands, or data loss can occur. See Solid State Drives#TRIM.

Besides enabling discard support in dm-crypt, it is also required to periodically run fstrim(8) or mount the filesystem (e.g. /dev/mapper/root in this example) with the discard option in /etc/fstab. For details, please refer to the TRIM page.

LUKS2

For a LUKS2 device, TRIM support can be enabled by using the --allow-discards --persistent options when opening it. The allow-discards flag will be written into the LUKS2 header and the option will be automatically used whenever the LUKS2 device is opened.

Note:

• Setting new persistent flags via cryptsetup --persistent replaces old flags with new ones instead of adding a new flag to the already set flags. This means if you want to enable other flags too, you have to set them all at once.
• When using OPAL encryption without dm-crypt (cryptsetup-luksFormat(8) option --hw-opal-only), discard support does not need to be explicitly enabled since there is no dm-crypt layer between the file system and the disk.

# cryptsetup --allow-discards --persistent open /dev/sdaX root

If the device is already opened, the open action will raise an error, in which case, use the cryptsetup-refresh(8) command instead:

# cryptsetup --allow-discards --persistent refresh root

You can confirm the flag is persistently set in the LUKS2 header by looking at the cryptsetup luksDump output:

# cryptsetup luksDump /dev/sdaX | grep Flags

Flags:          allow-discards

LUKS1 and plain dm-crypt

For LUKS1 and plain dm-crypt, TRIM support needs to be explicitly enabled when opening the device.

To enable TRIM support during boot, set the following kernel parameters.

If using the encrypt hook:

cryptdevice=/dev/sdaX:root:allow-discards

If using the sd-encrypt hook with systemd-based initramfs:

rd.luks.options=discard

Note: The rd.luks.options=discard kernel option does not have any effect on devices included in the initramfs image's /etc/crypttab file (/etc/crypttab.initramfs on real root). You must specify option discard in /etc/crypttab.initramfs.

For devices unlocked via /etc/crypttab use option discard, e.g.:

/etc/crypttab

luks-123abcdef-etc UUID=123abcdef-etc none discard

When manually unlocking devices on the console use --allow-discards.

For example, you can open a device with the --allow-discards option to execute a manual fstrim command:

# cryptsetup --allow-discards open /dev/sdaX root

In any case, you can verify whether the device actually was opened with discards by inspecting the dmsetup table output:

# dmsetup table

luks-123abcdef-etc: 0 1234567 crypt aes-xts-plain64 000etc000 0 8:2 4096 1 allow_discards

Disable workqueue for increased solid state drive (SSD) performance

Solid state drive users should be aware that, by default, discarding internal read and write workqueue commands are not enabled by the device-mapper, i.e. block-devices are mounted without the no_read_workqueue and no_write_workqueue option unless you override the default.

The no_read_workqueue and no_write_workqueue flags were introduced by internal Cloudflare research Speeding up Linux disk encryption made while investigating overall encryption performance. One of the conclusions is that internal dm-crypt read and write queues decrease performance for SSD drives. While queuing disk operations makes sense for spinning drives, bypassing the queue and writing data synchronously doubled the throughput and cut the SSD drives' IO await operations latency in half. The patches were upstreamed and are available since linux 5.9 and up [5].

Tip: linux-zen has dm-crypt workqueues disabled by default.

To disable workqueue for LUKS devices unlocked via crypttab use one or more of the desired no-read-workqueue or no-write-workqueue options. E.g.:

/etc/crypttab

luks-123abcdef-etc UUID=123abcdef-etc none no-read-workqueue

To disable both read and write workqueue add both flags:

/etc/crypttab

luks-123abcdef-etc UUID=123abcdef-etc none no-read-workqueue,no-write-workqueue

With LUKS2 you can set --perf-no_read_workqueue and --perf-no_write_workqueue as default flags for a device by opening it once with the option --persistent. For example:

# cryptsetup --perf-no_read_workqueue --perf-no_write_workqueue --persistent open /dev/sdaX root

When the device is already opened, the open action will raise an error. You can use the refresh option in these cases, e.g.:

# cryptsetup --perf-no_read_workqueue --perf-no_write_workqueue --persistent refresh root

You can confirm which flags are persistently set in the LUKS2 header by looking at the cryptsetup luksDump output:

# cryptsetup luksDump /dev/sdaX | grep Flags

Flags:          no-read-workqueue

In any case, you can verify whether the device actually was opened with these flags by inspecting the dmsetup table output:

# dmsetup table

luks-123abcdef-etc: 0 1234567 crypt aes-xts-plain64 000etc000 0 8:2 4096 1 no_read_workqueue

Note: Setting new persistent flags via cryptsetup --persistent replaces old flags with new ones instead of adding a new flag to the already set flags. This means if you want to enable both --perf-no_read_workqueue and --perf-no_write_workqueue (or more) you have to set them all at once.

Example for setting both no_read_workqueue and no_write_workqueue with cryptsetup:

# cryptsetup --perf-no_read_workqueue --perf-no_write_workqueue --persistent refresh root

You can confirm both flags being set by inspecting the LUKS2 cryptsetup luksDump output:

# cryptsetup luksDump /dev/sdaX | grep Flags

Flags:          no-read-workqueue no-write-workqueue

The encrypt hook and multiple disks

Note: This section only concerns unlocking in early userspace (in the initramfs phase). For unlocking devices in late userspace (after switch root), see dm-crypt/System configuration#Unlocking in late userspace.

Tip: sd-encrypt hook supports unlocking multiple devices. They can be specified on the kernel command line or in /etc/crypttab.initramfs. See dm-crypt/System configuration#Using systemd-cryptsetup-generator.

The encrypt hook only allows for a single cryptdevice= entry (archlinux/mkinitcpio/mkinitcpio#231). In system setups with multiple drives this may be limiting, because dm-crypt has no feature to exceed the physical device. For example, take "LVM on LUKS": The entire LVM exists inside a LUKS mapper. This is perfectly fine for a single-drive system, since there is only one device to decrypt. But what happens when you want to increase the size of the LVM? You cannot, at least not without modifying the encrypt hook.

The following sections briefly show alternatives to overcome the limitation. The first deals with how to expand a LUKS on LVM setup to a new disk. The second with modifying the encrypt hook to unlock multiple disks in LUKS setups without LVM.

Expanding LVM on multiple disks

The management of multiple disks is a basic LVM feature and a major reason for its partitioning flexibility. It can also be used with dm-crypt, but only if LVM is employed as the first mapper. In such a LUKS on LVM setup the encrypted devices are created inside the logical volumes (with a separate passphrase/key per volume). The following covers the steps to expand that setup to another disk.

Warning: Back up! While resizing filesystems may be standard, keep in mind that operations may go wrong and the following might not apply to a particular setup. Generally, extending a filesystem to free disk space is less problematic than shrinking one. This in particular applies when stacked mappers are used, as it is the case in the following example.

Adding a new drive

First, it may be desired to prepare a new disk according to dm-crypt/Drive preparation. Second, it is partitioned as a LVM, e.g. all space is allocated to /dev/sdY1 with partition type 8E00 (Linux LVM). Third, the new disk/partition is attached to the existing LVM volume group, e.g.:

# pvcreate /dev/sdY1
# vgextend MyStorage /dev/sdY1

Extending the logical volume

For the next step, the final allocation of the new diskspace, the logical volume to be extended has to be unmounted. It can be performed for the cryptdevice root partition, but in this case the procedure has to be performed from an Arch Install ISO.

In this example, it is assumed that the logical volume for /home (lv-name homevol) is going to be expanded with the fresh disk space:

# umount /home
# fsck /dev/mapper/home
# cryptsetup close /dev/mapper/home
# lvextend -l +100%FREE MyStorage/homevol

Now the logical volume is extended and the LUKS container comes next:

# cryptsetup open /dev/MyStorage/homevol home
# umount /home      # as a safety, in case it was automatically remounted
# cryptsetup --verbose resize home

Finally, the filesystem itself is resized:

# e2fsck -f /dev/mapper/home
# resize2fs /dev/mapper/home

Done! If it went to plan, /home can be remounted and now includes the span to the new disk:

# mount /dev/mapper/home /home

Note that the cryptsetup resize action does not affect encryption keys, and these have not changed.

Modifying the encrypt hook for multiple partitions

Note that sd-encrypt supports multiple partitions out of the box. If several (or all) partitions opened this way share the same passphrase, sd-encrypt will try it for each and not ask for it multiple times. This may be an easier alternative to the following.

Root filesystem spanning multiple partitions

It is possible to modify the encrypt hook to allow multiple hard drive decrypt root (/) at boot. One way:

# cp /usr/lib/initcpio/install/encrypt /etc/initcpio/install/encrypt2
# cp /usr/lib/initcpio/hooks/encrypt  /etc/initcpio/hooks/encrypt2
# sed -i "s/cryptdevice/cryptdevice2/" /etc/initcpio/hooks/encrypt2
# sed -i "s/cryptkey/cryptkey2/" /etc/initcpio/hooks/encrypt2

Add cryptdevice2= to your boot options (and cryptkey2= if needed), and add the encrypt2 hook to your mkinitcpio.conf before rebuilding it. See dm-crypt/System configuration.

Multiple non-root partitions

Maybe you have a requirement for using the encrypt hook on a non-root partition. Arch does not support this out of the box, however, you can easily change the cryptdev and cryptname values in /lib/initcpio/hooks/encrypt (the first one to your /dev/sd* partition, the second to the name you want to attribute). That should be enough.

The big advantage is you can have everything automated, while setting up /etc/crypttab with an external key file (i.e. the keyfile is not on any internal hard drive partition) can be a pain - you need to make sure the USB/FireWire/... device gets mounted before the encrypted partition, which means you have to change the order of /etc/fstab (at least).

Of course, if the cryptsetup package gets upgraded, you will have to change this script again. Unlike /etc/crypttab, only one partition is supported, but with some further hacking one should be able to have multiple partitions unlocked.

If you want to do this on a software RAID partition, there is one more thing you need to do. Just setting the /dev/mdX device in /lib/initcpio/hooks/encrypt is not enough; the encrypt hook will fail to find the key for some reason, and not prompt for a passphrase either. It looks like the RAID devices are not brought up until after the encrypt hook is run. You can solve this by putting the RAID array in /boot/grub/menu.lst, like

kernel /boot/vmlinuz-linux md=1,/dev/hda5,/dev/hdb5

If you set up your root partition as a RAID, you will notice the similarities with that setup. GRUB can handle multiple array definitions just fine:

kernel /boot/vmlinuz-linux root=/dev/md0 ro md=0,/dev/sda1,/dev/sdb1 md=1,/dev/sda5,/dev/sdb5,/dev/sdc5

Encrypted system using a detached LUKS header

This example follows the same setup as in dm-crypt/Encrypting an entire system#Plain dm-crypt, which should be read first before following this guide.

By using a detached header the encrypted blockdevice itself only carries encrypted data, which gives deniable encryption as long as the existence of a header is unknown to the attackers. It is similar to using plain dm-crypt, but with the LUKS advantages such as multiple passphrases for the masterkey and key derivation. Further, using a detached header offers a form of two factor authentication with an easier setup than using GPG or OpenSSL encrypted keyfiles, while still having a built-in password prompt for multiple retries. See Data-at-rest encryption#Cryptographic metadata for more information.

See dm-crypt/Device encryption#Encryption options for LUKS mode for encryption options before performing the first step to setup the encrypted system partition and creating a header file to use with cryptsetup:

# dd if=/dev/zero of=header.img bs=16M count=1
# cryptsetup luksFormat --offset 32768 --header header.img /dev/sdX

Tip: The --offset option allows specifying the start of encrypted data on a device. By reserving a space at the beginning of device you have the option of later reattaching the LUKS header. The value is specified in 512-byte sectors, see cryptsetup-luksFormat(8) for more details.

Open the container:

# cryptsetup open --header header.img /dev/sdX enc

Now follow the LVM on LUKS setup to your requirements. The same applies for preparing the boot partition on the removable device (because if not, there is no point in having a separate header file for unlocking the encrypted disk). Next move the header.img onto it:

# mv header.img /mnt/boot

Follow the installation procedure up to the mkinitcpio step (you should now be arch-chrooted inside the encrypted system).

Tip: You will notice that since the system partition only has "random" data, it does not have a partition table and by that an UUID or a LABEL. But you can still have a consistent mapping using the Persistent block device naming#by-id and by-path. E.g. using disk id from /dev/disk/by-id/.

There are two options for initramfs to support a detached LUKS header.

Using systemd hook

Set the following kernel parameters:

rd.luks.name=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX=enc rd.luks.options=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX=header=/header.img:UUID=ZZZZZZZZ-ZZZZ-ZZZZ-ZZZZ-ZZZZZZZZZZZZ rd.luks.data=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX=/dev/disk/by-id/your-disk_id

• Replace XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX with the LUKS super block UUID. It can be acquired with cryptsetup luksDump header.img or blkid -s UUID -o value header.img.
• Replace UUID=ZZZZZZZZ-ZZZZ-ZZZZ-ZZZZ-ZZZZZZZZZZZZ with the block device of volume in which the header file is located.

Alternatively, instead of using the rd.luks kernel parameters, the options can be specified in a /etc/crypttab.initramfs file:

/etc/crypttab.initramfs

enc	/dev/disk/by-id/your-disk_id	none	header=/header.img:UUID=ZZZZZZZZ-ZZZZ-ZZZZ-ZZZZ-ZZZZZZZZZZZZ

Next, modify /etc/mkinitcpio.conf to use systemd and to include the file system module for the volume in which the header is located. For example, if it is a FAT volume:

/etc/mkinitcpio.conf

...
MODULES=(vfat)
...
HOOKS=(base systemd autodetect microcode modconf kms keyboard sd-vconsole block sd-encrypt lvm2 filesystems fsck)
...

Regenerate the initramfs and you are done.

Note:

• When using /etc/crypttab.initramfs, no cryptsetup parameters need to be passed to the kernel command line, since /etc/crypttab.initramfs will be added as /etc/crypttab in the initramfs.
• Refrain from using the rd.luks kernel parameters together with /etc/crypttab.initramfs as it can cause conflicts. See the warning in dm-crypt/System configuration#Using systemd-cryptsetup-generator for details.

Modifying encrypt hook

This method shows how to modify the encrypt hook in order to use a detached LUKS header. Now the encrypt hook has to be modified to let cryptsetup use the separate header (archlinux/mkinitcpio/mkinitcpio#234; base source and idea for these changes published on the BBS). Make a copy so it is not overwritten on a mkinitcpio update:

# cp /usr/lib/initcpio/hooks/encrypt /etc/initcpio/hooks/encrypt2
# cp /usr/lib/initcpio/install/encrypt /etc/initcpio/install/encrypt2

/etc/initcpio/hooks/encrypt2 (around line 52)

warn_deprecated() {
    echo "The syntax 'root=${root}' where '${root}' is an encrypted volume is deprecated"
    echo "Use 'cryptdevice=${root}:root root=/dev/mapper/root' instead."
}

local headerFlag=false
for cryptopt in ${cryptoptions//,/ }; do
    case ${cryptopt} in
        allow-discards)
            cryptargs="${cryptargs} --allow-discards"
            ;;
        header)
            cryptargs="${cryptargs} --header /boot/header.img"
            headerFlag=true
            ;;
        *)
            echo "Encryption option '${cryptopt}' not known, ignoring." >&2
            ;;
    esac
done

if resolved=$(resolve_device "${cryptdev}" ${rootdelay}); then
    if $headerFlag || cryptsetup isLuks ${resolved} >/dev/null 2>&1; then
        [ ${DEPRECATED_CRYPT} -eq 1 ] && warn_deprecated
        dopassphrase=1

Now edit the mkinitcpio.conf to add the encrypt2 and lvm2 hooks, the header.img to FILES and the loop to MODULES, apart from other configuration the system requires:

/etc/mkinitcpio.conf

...
MODULES=(loop)
...
FILES=(/boot/header.img)
...
HOOKS=(base udev autodetect microcode modconf kms keyboard keymap consolefont block encrypt2 lvm2 filesystems fsck)
...

This is required so the LUKS header is available on boot allowing the decryption of the system, exempting us from a more complicated setup to mount another separate USB device in order to access the header. After this set up the initramfs is created.

Next the boot loader is configured to specify the cryptdevice= also passing the new header option for this setup:

cryptdevice=/dev/disk/by-id/your-disk_id:enc:header

To finish, following dm-crypt/Encrypting an entire system#Post-installation is particularly useful with a /boot partition on an USB storage medium.

Encrypted /boot and a detached LUKS header on USB

Rather than embedding the header.img and keyfile into the initramfs image, this setup will make your system depend entirely on the usb key rather than just the image to boot, and on the encrypted keyfile inside of the encrypted boot partition. Since the header and keyfile are not included in the initramfs image and the custom encrypt hook is specifically for the usb's by-id, you will literally need the usb key to boot.

For the usb drive, since you are encrypting the drive and the keyfile inside, it is preferred to cascade the ciphers as to not use the same one twice. Whether a meet-in-the-middle attack would actually be feasible is debatable. You can do twofish-serpent or serpent-twofish.

Preparing the disk devices

sdb will be assumed to be the USB drive, sda will be assumed to be the main hard drive.

Prepare the devices according to dm-crypt/Drive preparation.

Preparing the USB key

Use gdisk to partition the disk according to the layout shown here, with the exception that it should only include the first two partitions. So as follows:

# gdisk /dev/sdb

Number  Start (sector)    End (sector)  Size       Code  Name
   1            2048         1050623   512.0 MiB   EF00  EFI System
   2         1050624         1460223   200.0 MiB   8300  Linux filesystem

Before running cryptsetup, look at the Encryption options for LUKS mode and Ciphers and modes of operation first to select your desired settings.

Prepare the boot partition but do not mount any partition yet and Format the EFI system partition.

# mount /dev/mapper/cryptboot /mnt
# dd if=/dev/urandom of=/mnt/key.img bs=filesize count=1 iflag=fullblock
# cryptsetup luksFormat /mnt/key.img
# cryptsetup open /mnt/key.img lukskey

filesize is in bytes but can be followed by a suffix such as M. Having too small of a file will get you a nasty Requested offset is beyond real size of device /dev/loop0 error. As a rough reference, creating a 4M file will encrypt it successfully. You should make the file larger than the space needed since the encrypted loop device will be a little smaller than the file's size.

With a big file, you can use --keyfile-offset=offset and --keyfile-size=size to navigate to the correct position (see Gentoo:Custom Initramfs#Encrypted keyfile).

Now you should have lukskey opened in a loop device (underneath /dev/loop1), mapped as /dev/mapper/lukskey.

The main drive

# truncate -s 16M /mnt/header.img
# cryptsetup --key-file=/dev/mapper/lukskey --keyfile-offset=offset --keyfile-size=size luksFormat /dev/sda --offset 32768 --header /mnt/header.img

Pick an offset and size in bytes (8192 KiB is the maximum keyfile size for cryptsetup).

# cryptsetup open --header /mnt/header.img --key-file=/dev/mapper/lukskey --keyfile-offset=offset --keyfile-size=size /dev/sda enc
# cryptsetup close lukskey
# umount /mnt

Follow Preparing the logical volumes to set up LVM on LUKS.

See Partitioning#Discrete partitions for recommendations on the size of your partitions.

Once your root partition is mounted, mount your encrypted boot partition as /mnt/boot and your EFI system partition as /mnt/efi.

Installation procedure and custom encrypt hook

Follow the installation guide up to the mkinitcpio step but do not do it yet, and skip the partitioning, formatting, and mounting steps as they have already been done.

In order to get the encrypted setup to work, you need to build your own hook, which is thankfully easy to do and here is the code you need. You will have to follow Persistent block device naming#by-id and by-path to figure out your own by-id values for the usb and main hard drive (they are linked -> to sda or sdb).

You should be using the by-id instead of just sda or sdb because sdX can change and this ensures it is the correct device.

You can name customencrypthook anything you want, and custom build hooks can be placed in the hooks and install folders of /etc/initcpio. Keep a backup of both files (cp them over to the /home partition or your user's /home directory after you make one). /usr/bin/ash is not a typo.

/etc/initcpio/hooks/customencrypthook

#!/usr/bin/ash

run_hook() {
    modprobe -a -q dm-crypt >/dev/null 2>&1
    modprobe loop
    [ "${quiet}" = "y" ] && CSQUIET=">/dev/null"

    while [ ! -L '/dev/disk/by-id/usbdrive-part2' ]; do
     echo 'Waiting for USB'
     sleep 1
    done

    cryptsetup open /dev/disk/by-id/usbdrive-part2 cryptboot
    mount --mkdir /dev/mapper/cryptboot /mnt
    cryptsetup open /mnt/key.img lukskey
    cryptsetup --header /mnt/header.img --key-file=/dev/mapper/lukskey --keyfile-offset=''offset'' --keyfile-size=''size'' open /dev/disk/by-id/harddrive enc
    cryptsetup close lukskey
    umount /mnt
}

usbdrive is your USB drive by-id, and harddrive is your main hard drive by-id.

Tip: You could also close cryptboot using cryptsetup close, but having it open makes it easier to mount for system updates using Pacman and regenerating the initramfs with mkinitcpio. The /boot partition must be mounted for updates that affect the kernel or Initramfs, and the initramfs will be automatically regenerated after these updates.

# cp /usr/lib/initcpio/install/encrypt /etc/initcpio/install/customencrypthook

Now edit the copied file and remove the help() section as it is not necessary.

/etc/mkinitcpio.conf (edit this only do not replace it, these are just excerpts of the necessary parts)

MODULES=(loop)
...
HOOKS=(base udev autodetect microcode modconf kms keyboard keymap consolefont block customencrypthook lvm2 filesystems fsck)

The files=() and binaries=() arrays are empty, and you should not have to replace HOOKS=(...) array entirely just edit in customencrypthook lvm2 after block and before filesystems, and make sure systemd and encrypt are removed.

Boot loader

Finish the Installation Guide from the mkinitcpio step. To boot you would need either GRUB or efibootmgr. Note you can use GRUB to support the encrypted disks by Configuring the boot loader but editing the GRUB_CMDLINE_LINUX is not necessary for this set up.

Or use direct UEFI Secure Boot by generating keys with cryptboot^AUR then signing the initramfs and kernel and creating a bootable .efi file for your EFI system partition with sbupdate-git^AUR. Before using cryptboot or sbupdate note this excerpt from Secure Boot#Using your own keys:

Tip: Note that cryptboot^AUR requires the encrypted boot partition to be specified in /etc/crypttab before it runs, and if you are using it in combination with sbupdate-git^AUR, sbupdate expects the /boot/efikeys/db.* files created by cryptboot to be capitalized like DB.* unless otherwise configured in /etc/default/sbupdate. Users who do not use systemd to handle encryption may not have anything in their /etc/crypttab file and would need to create an entry.

# efibootmgr --create --disk /dev/device --part partition_number --label "Arch Linux Signed" --loader "EFI\Arch\linux-signed.efi" --unicode

See efibootmgr(8) for an explanation of the options.

Make sure the boot order puts Arch Linux Signed first. If not change it with efibootmgr --bootorder XXXX,YYYY,ZZZZ --unicode.

Changing the LUKS keyfile

# cryptsetup --header /boot/header.img --key-file=/dev/mapper/lukskey --keyfile-offset=offset --keyfile-size=size luksChangeKey /dev/mapper/enc /dev/mapper/lukskey2 --new-keyfile-size=newsize --new-keyfile-offset=newoffset

Afterwards, cryptsetup close lukskey and shred or dd the old keyfile with random data before deleting it, then make sure that the new keyfile is renamed to the same name of the old one: key.img or other name.