BL1 Immutable bootloader

Author:: Raef Coles
Organization:: Arm Limited
Contact:: raef.coles@arm.com

Introduction

Some devices that use TF-M will require initial boot code that is stored in ROM. There are a variety of reasons that this might happen:

The device cannot access flash memory without a driver, so needs some setup to be done before main images on flash can be booted.
The device has no on-chip secure flash, and therefore cannot otherwise maintain a tamper-resistant root of trust.
The device has a security model that requires an immutable root of trust

Henceforth any bootloader stored in ROM will be referred to as BL1, as it would necessarily be the first stage in the boot chain.

TF-M provides a reference second-stage flash bootloader BL2, in order to allow easier integration. This bootloader implements all secure boot functionality needed to provide a secure chain of trust.

A reference ROM bootloader BL1 has now being added with the same motivation - allowing easier integration of TF-M for platforms that do not have their own BL1 and require one.

BL1 Features and Motivations

The reference ROM bootloader provides the following features:

A split between code being stored in ROM and in other non-volatile memory.
- This can allow significant cost reduction in fixing bugs compared to ROM-only bootloaders.
A secure boot mechanism that allows upgrading the next boot stage (which would usually be BL2).
- This allows for the fixing of any bugs in the BL2 image.
- Alternately, this could allow the removal of BL2 in some devices that are constrained in flash space but have ROM.
A post-quantum resistant asymmetric signature scheme for verifying the next boot stage image.
- This can allow devices to be securely updated even if attacks involving quantum computers become viable. This could extend the lifespans of devices that might be deployed in the field for many years.
A mechanism for passing boot measurements to the TF-M runtime so that they can be attested.
Tooling to create and sign images.
Fault Injection (FI) and Differential Power Analysis (DPA) mitigations.

BL1_1 and BL1_2 split bootloaders

BL1 is split into two distinct boot stages, BL1_1 which is stored in ROM and BL1_2 which is stored in other non-volatile storage. This would usually be either trusted or untrusted flash, but on platforms without flash memory can be OTP. As BL1_2 is verified against a hash stored in OTP, it is immutable after provisioning even if stored in mutable storage.

Bugs in ROM bootloaders usually cannot be fixed once a device is provisioned / in the field, as ROM code is immutable the only option is fixing the bug in newly manufactured devices.

However, it can be very expensive to change the ROM code of devices once manufacturing has begun, as it requires changes to the photolithography masks that are used to create the device. This cost varies depending on the complexity of the device and of the process node that it is being fabricated on, but can be large, both in engineering time and material/process costs.

By placing the majority of the immutable bootloader in other storage, we can mitigate the costs associated with changing ROM code, as a new BL1_2 image can be used at provisioning time with minimal changeover cost. BL1_1 contains a minimal codebase responsible mainly for the verification of the BL1_2 image.

The bootflow is as follows. For simplicity this assumes that the boot stage after BL1 is BL2, though this is not necessarily the case:

BL1_1 begins executing in place from ROM
BL1_1 copies BL1_2 into RAM
BL1_1 verifies BL1_2 against the hash stored in OTP
BL1_1 jumps to BL1_2, if the hash verification has succeeded
BL1_2 copies the primary BL2 image from flash into RAM
BL1_2 verifies the BL2 image using asymmetric cryptography
If verification fails, BL1_2 repeats 5 and 6 with the secondary BL2 image
BL1_2 jumps to BL2, if either image has successfully verified

Note

The BL1_2 image is not encrypted, so if it is placed in untrusted flash it will be possible to read the data in the image.

Some optimizations have been made specifically for the case where BL1_2 has been stored in OTP:

OTP can be very expensive in terms of chip area, though new technologies like antifuse OTP decrease this cost. Because of this, the code size of BL1_2 has been minimized. Code-sharing has been configured so that BL1_2 can call functions stored in ROM. Care should be taken that OTP is sized such that it is possible to include versions of the functions used via code-sharing, in case the ROM functions contain bugs, though less space is needed than if all code is duplicated as it is assumed that most functions will not contain bugs.

As OTP memory frequently has low performance, BL1_2 is copied into RAM before it it is executed. It also copies the next image stage into RAM before authenticating it, which allows the next stage to be stored in untrusted flash. This requires that the device have sufficient RAM to contain both the BL1_2 image and the next stage image at the same time. Note that this is done even if BL1_2 is located in XIP-capable flash, as it both allows the use of untrusted flash and simplifies the image upgrade logic.

Note

BL1_2 enables TF-M to be used on devices that contain no secure flash, though the ITS service will not be available. Other services that depend on ITS will not be available without modification.

Secure boot / Image upgrade mechanism

BL1_2 verifies the authenticity of the next stage image via asymmetric cryptography, using a public key that is provisioned into OTP.

BL1_2 implements a rollback protection counter in OTP, which is used to prevent the next stage image being downgraded to a less secure version.

BL1_2 has two image slots, which allows image upgrades to be performed. The primary slot is always booted first, and then if verification of this fails (either due to an invalid signature or due to a version lower than the rollback protection counter) the secondary slot is then booted (subject to the same checks).

BL1_2 contains no image upgrade logic, in order for OTA of the next stage image to be implemented, a later stage in the system must handle downloading new images and placing them in the required slot.

Post-Quantum signature verification in BL1_2

BL1_2 uses a post-quantum asymmetric signature scheme to verify the next stage. The scheme used is Leighton-Michaeli Signatures (henceforth LMS). LMS is standardised in NIST SP800-208 and IETF RFC8554.

LMS is a stateful-hash signature scheme, meaning that:

It is constructed from a cryptographic hash function, in this case SHA256.

This function can be accelerated by existing hardware accelerators, which can make LMS verification relatively fast compared to other post-quantum signature schemes that cannot be accelerated in hardware yet.

Each private key can only be used to sign a certain number of images.

BL1_2 uses the SHA256_H10 parameter set, meaning each key can sign 1024 images.

The main downside, the limited amount of possible signatures, can be mitigated by limiting the amount of image upgrades that are done. As BL2 is often currently not upgradable, it is not anticipated that this limit will be problematic. If BL1 is being used to directly boot a TF-M/NS combined image, the limit is more likely to be problematic, and care should be taken to examine the likely update amount.

LMS public keys are 32 bytes in size, and LMS signatures are 1912 bytes in size. The signature size is larger than some asymmetric schemes, though most devices should have enough space in flash to accommodate this.

The main upside of LMS, aside from the security against attacks involving quantum computers, is that it is relatively simple to implement. The software implementation that is used by BL1 is ~3KiB in size, which is considerably smaller than the corresponding RSA implementation which is at least 6.5K. This simplicity of implementation is useful to avoid bugs.

BL1 will use MbedTLS as the source for its implementation of LMS.

Note

As of the time of writing, the LMS code is still in the process of being merged into MbedTLS, so BL1 currently does not support asymmetric verification of the next boot stage. Currently, the next boot stage is hash-locked, so cannot be upgraded.

The Github pull request for LMS can be found here

BL1 boot measurements

BL1 outputs boot measurements in the same format as BL2, utilising the same shared memory area. These measurements can then be included in the attestation token, allowing the attestation of the version of the boot stage after BL1.

BL1 tooling

Image signing scripts are provided for BL1_1 and BL1_2. While the script is named create_bl2_img.py, it can be used for any next stage image.

bl1/bl1_1/scripts/create_bl1_2_img.py
bl1/bl1_2/scripts/create_bl2_img.py

These sign (and encrypt in the case of create_bl2_img.py) a given image file and append the required headers.

BL1 FI and DPA mitigations

BL1 reuses the FI countermeasures used in the TF-M runtime, which are found in lib/fih/.

BL1 implements countermeasures against DPA, which are primarily targeted towards being able to handle cryptographic material without leaking its contents. The functions with these countermeasures are found in bl1/bl1_1/shared_lib/util.c

bl_secure_memeql tests if memory regions have the same value

It does not perform early exits to prevent timing attacks.
It compares chunks in random orders to prevent DPA trace correlation analysis
It inserts random delays to prevent DPA trace correlation analysis
It performs loop integrity checks
It uses FIH constructs

bl_secure_memcpy copies memory regions

It copies chunks in random orders to prevent DPA trace correlation analysis
It inserts random delays to prevent DPA trace correlation analysis
It performs loop integrity checks
It uses FIH constructs

Using BL1 on new platforms

New platforms must define the following macros in their region_defs.h:

BL1_1_HEAP_SIZE
BL1_1_STACK_SIZE
BL1_2_HEAP_SIZE
BL1_2_STACK_SIZE
BL1_1_CODE_START
BL1_1_CODE_LIMIT
BL1_1_CODE_SIZE
BL1_2_CODE_START
BL1_2_CODE_LIMIT
BL1_2_CODE_SIZE
PROVISIONING_DATA_START
PROVISIONING_DATA_LIMIT
PROVISIONING_DATA_SIZE

The PROVISIONING_DATA_* defines are used to locate where the data to be provisioned into OTP can be found. These are required as the provisioning bundle needs to contain the entire BL1_2 image, usually >= 8KiB in size, which is too large to be placed in the static data area as is done for all other dummy provisioning data. On development platforms with reprogrammable ROM, this is often placed in unused ROM. On production platforms, this should be located in RAM and then filled with provisioning data. The format of the provisioning data that should be located in the PROVISIONING_DATA_* region can be found in bl1/bl1_1/lib/provisioning.c in the struct bl1_assembly_and_test_provisioning_data_t

If the platform is storing BL1_2 in flash, it must set BL1_2_IMAGE_FLASH_OFFSET to the flash offset of the start of BL1_2.

The platform must also implement the HAL functions defined in the following headers:

bl1/bl1_1/shared_lib/interface/trng.h
bl1/bl1_1/shared_lib/interface/crypto.h
bl1/bl1_1/shared_lib/interface/otp.h

If the platform integrates a CryptoCell-312, then it can reuse the existing implementation.

BL1 Testing

New tests have been written to test both the HAL implementation, and the integration of those functions for verifying images. These tests are stored in the tf-m-tests repository, under the test/bl1/ directory, and further subdivided into BL1_1 and BL1_2 tests.