Boot chain and integrity guarantees¶
The main disk GPT partition layout is as follows (Fig. 4):
EFI system partition (ESP) which includes:
The EFI bootloader and its configuration;
An EFI binary for each version currently installed (usually two, but only one at installation time). Each EFI binary bundles a Linux kernel, its command line and an initramfs.
LVM partition which holds all LVM Volume Groups and Logical Volumes for the system. The currently available ones are:
DM-Verity block device and metadata appended at the end
Uncompressed squashfs filesystem
DM-Crypt + DM-Integrity block device
DM-Crypt with a random key generated at boot time
Boot chain order¶
As shown in Fig. 5, the boot steps (following hardware initialization) are as follows:
The UEFI Firmware executes the initial bootloader from the EFI system partition.
Following the Boot Loader Specification, the bootloader (gummiboot/systemd-boot) executes the EFI binary in the
EFI/Linuxfolder with the highest version number. The executed EFI binary is a bundle created by dracut which wraps a Linux kernel, its command line and an initramfs.
The initramfs opens the
core-<version>DM-Verity block device and mounts the SquashFS filesystem in
/sysroot. Then, using the TPM, it unseals the key required to open the DM-Crypt/DM-Integrity
core_statepartition and mounts the resulting ext4 partition in
/sysroot/mnt/state. Finally, it performs a pivot_root to switch to the final root.
Boot chain and rootfs integrity¶
The integrity of the boot chain and root filesystem is guaranteed by the following mechanisms.
All EFI binaries installed in the EFI System Partition (ESP) are signed in order to protect their integrity via Secure Boot. Their confidentiality however is not assured. Currently, two EFI binaries are necessary:
The systemd-boot EFI bootloader. Its only purpose is to execute the latest CLIP OS EFI binary. This bootloader is minimal and only acts as a switch between the available CLIP OS versions. This enables recovery in case of update failure or temporary rollback due to bugs in a new version.
The main CLIP OS EFI binary. It contains the initial kernel image, the initramfs and the kernel command line. Signing this binary guarantees the integrity of its three components. Note that the kernel command line also includes the DM-Verity root hash for the CLIP OS Core partition. This binary is chosen by the bootloader and the Secure Boot signature verification is performed by the UEFI Firmware.
No default keys are supported (e.g., Microsoft keys) thus you will need to generate you own Secure Boot signing keys and use them during the build process. Those keys must then be enrolled in hardware.
Once the kernel is booted with the initial initramfs, it will look for the
core-<version> LVM Logical Volume which includes a DM-Verity block device.
root hash used to verify the integrity of this partition is included in
the kernel command line (thus protected in integrity by Secure Boot). This
assures the integrity of the content of the DM-Verity block device, which
includes the read-only uncompressed SquashFS root filesystem. Support for
forward error correction (FEC) is also enabled thus increasing resistance to
disk read errors or failures.
We want to achieve full-disk encryption in a way that is both secure and unnoticeable to the end user. The TPM enables us to do that by sealing the encryption key and providing it at boot time if and only if the machine is in a known-good state (via PCRs extension). In other words, we want to be able to decrypt the disk only once we prove we have booted trusted code.
However, we would like to avoid re-sealing the encryption key every single time we upgrade the kernel, the initramfs or the command line.
We already have Secure Boot, ensuring the UEFI firmware boots our signed EFI binary containing the kernel, the initramfs and the command line. In other words, we already assert the boot of trusted code.
recording whether Secure Boot is turned on;
recording the key database;
recording which keys were used.
By sealing the disk-encryption key according to the PCR 7 value, we ensure that the disk is decrypted if and only if our Secure Boot policy has not been tampered with, which in turn means that we booted a trusted EFI binary, i.e. trusted kernel, initramfs and command line.
Naturally, we can in addition use most if not all PCRs 0 to 6 to measure firmware integrity. In the future, we may also use an additional PCR to measure the LUKS header(s), similarly to what TrustedGRUB2 does with PCR 12.
We rely on TPM 2.0, for various reasons including:
TPM 1.2 only supports deprecated cryptographic algorithms and TrouSerS is hard to deal with and not satisfying (partly due to tcsd);
TPM 2.0 is already replacing TPM 1.2 in new machines;
TPM 2.0 offers several new interesting functionalities, such as multiple hierarchies.
One issue with TPM 2.0 is that utilities and libraries to deal with it are still under heavy development. We chose to use the tpm2-tools, which rely on the tpm2-tss implementation of the TCG’s TPM2 Software Stack (TSS2).
Initramfs and LUKS¶
We use a Bash script located in our initramfs.
TPM-sealed LUKS keyfiles are located in the EFI System Partition.
The kernel’s Resource Manager (RM) is used to ease objects management. Basically, the RM presents each new call to a
tpm2_*tool with an empty TPM (i.e. it cleans transient objects when the file handle to
We use the Owner Hierarchy (OH) and leave the Endorsement Hierarchy (EH) for remote attestation.
One could make the primary object persistent (with
tpm2_evictcontrol) to avoid re-derivating it from the seed each time and save some time at the expense of additional complexity.
We do not directly make the loaded keyfile object persistent as we would not have enough space in the TPM for all keyfiles we are going to want to use.
We would like to use different PCR lists for a given machine’s first boots following/during its provisioning, as we may for instance change its BIOS configuration.
We would like to use keyctl in order to directly store the decrypted LUKS key in a kernel keyslot so that cryptsetup can use it without it being passed through userspace. Note that, currently, keyfiles are in memory and may be swapped to disk, but that is tolerable as we use an encrypted swap device. Another solution could be to use ramfs instead of tmpfs.
Setup for testing¶
Under QEMU with OVMF¶
As a short term solution, we use hard-coded dummy keys to sign the EFI binaries with sbsigntools and ship an OVMF VARS template file in which these keys have been manually enrolled.
The optimal way of setting up Secure Boot for a virtual machine will be to develop and use an EFI binary to automatically enroll PK, KEK and db keys, similarly to what Fedora is performing. Such keys would be generated at build time for a given deployment using the relevant PKI.
The standard CLIP OS build process includes build steps to enable Secure Boot testing under QEMU with OVMF.
QEMU (and libvirt) support two TPM backends:
TPM passthrough device: requires the end user to have a hardware TPM on its host machine, which in addition cannot be used simultaneously by anything else. There are also some problems due to the way the TPM is initialized by the host, and thus some commands used by the guest cannot work as expected, and so on.
TPM emulator: provides TPM functionality for each VM using a TPM emulator installed on the host. swtpm is currently the only supported emulator.
The second option is more adapted to our needs but requires people to install swtpm, which requires libtpms.
The EFI firmware (OVMF) needs to be built with TPM support. We provide and use
our own derived
sys-firmware/edk2-ovmf to enable Secure Boot and TPM
With real hardware¶
This is not supported at the moment. The main difference is that Secure Boot keys will need to be enrolled through the manufacturer’s UEFI firmware. However, we expect the EFI binary mentioned above to make this step as automated as possible.