The application and support libraries in this directory provide a mechanism for a device's bootloader, using AVB, to store sensitive information. For a bootloader, sensitive information includes whether the device is unlocked or locked, whether it is unlockable, and what the minimum version of the OS/kernel is allowed to be booted. It may also store other sensitive flags.
The verified boot storage applet provides a mechanism to store this data in a way that enforceѕ the expected policies even if the higher level operating system is compromised or operates in an unexpected fashion.
The Verified Boot Storage Applet, VBSA, provides three purpose-built interfaces:
Each will be detailed below.
There are four supported lock types:
LOCK_CARRIER
LOCK_DEVICE
LOCK_BOOT
LOCK_OWNER
Each lock has a single byte of "lock" storage. If that byte is 0x0, then it is unlocked, or cleared. If it is non-zero, then the lock is locked. Any non-zero value is valid and may be used by the bootloader if any additional internal flagging is necessary.
In addition, a lock may have associated metadata which must be supplied during lock or unlock, or both.
See ese_boot_lock_*
in include/ese/app/boot.h for the specific interfaces.
The Carrier Lock implements a lock which can only be set when the device is not in production and can only be unlocked if provided a cryptographic signature.
This lock is available for use to implement "sim locking" or "phone locking" such that the carrier can determine if the device is allowed to boot an unsigned or unknown system image.
To provision this lock, device-specific data must be provided in an exact format. An example of this can be found in 'ese-boot-tool.cpp':collect_device_data()
. This data is then provided to the applet using ese_boot_lock_xset()
.
The following system attributes must be collected in the given order:
The data is serialized as follows:
\[length||string\]\[...\]
The length is a uint8_t
and the value is appended as a stream of uint8_t
values.
This data is then prefixed with the desired non-zero lock value before being submitted to the applet. If successful, the applet will have stored a SHA256 hash of the device data
Note, LOCK_CARRIER
can only be locked (non-zero lock value) when the applet is not in 'production' mode.
If LOCK_CARRIER
is set to a non-zero value and the applet is in production mode, then clearing the lock value requires authorization.
Authorization comes in the form of a RSA_SHA256-PKCS#1
signature over the provisioned device data SHA256 hash and a supplied montonically increasing "nonce".
The nonce value must be higher than the last seen nonce value and the signature must validate using public key internally stored in the applet (CarrierLock.java:PK_MOD
).
To perform a clear, ese_boot_lock_xset()
must be called with lock data that begins with 0x0, to clear the lock, and then contains data of the following format:
unlockToken = VERSION || NONCE || SIGNATURE SIGNATURE = RSA_Sign(SHA256(deviceData))
uint64_t
(8 bytes).uint64_t
(8 bytes).On unlock, the device data hash is cleared.
It is possible to test the key with a valid signature but a fake internal nonce and fake internal device data using ese_boot_carrier_lock_test()
. When using this interface, it expects the same unlock token as in the prior but prefixes with the fake data:
testVector = LAST_NONCE || DEVICE_DATA || unlockToken
The device lock is one of the setting used by the bootloader to determine if the boot lock can be changed. It may only be set by the operating system and is meant to protect the device from being reflashed by someone that cannot unlock or access the OS. This may also be used by an enterprise administrator to control lock policy for managed devices.
As LOCK_DEVICE
has not metadata, it can be set and retrieved using ese_boot_lock_set()
and ese_boot_lock_get()
.
The boot lock is used by the bootloader to control whether it should only boot verified system software or not. When the lock value is cleared (0x0), the bootloader will boot anything. When the lock value is non-zero, it should only boot software that is signed by a key stored in the bootloader except if LOCK_OWNER
is set. Discussion of LOCK_OWNER
will follow.
LOCK_BOOT
can only be toggled when in the bootloader/fastboot and if both LOCK_CARRIER
and LOCK_DEVICE
are cleared/unlocked.
As with LOCK_DEVICE
, LOCK_BOOT
has no metadata so it does not need the extended accessors.
The owner lock is used by the bootloader to support an alternative OS signing key provided by the device owner. LOCK_OWNER
can only be toggled if LOCK_BOOT
is cleared. LOCK_OWNER
does not require any metadata to unlock, but to lock, it requires a blob of up to 2048 bytes be provided. This is just secure storage for use by the bootloader. LOCK_OWNER
may be toggled in the bootloader or the operating system. This allows an unlocked device (LOCK_BOOT=0x0
) to install an owner key using fastboot or using software on the operating system itself.
Before LOCK_OWNER
's key should be honored by the bootloader, LOCK_BOOT
should be set (in the bootloader) to enforce use of the key and to keep malicious software in the operating system from changing it.
(Note, that the owner key should not be treated as equivalent to the bootloader's internally stored key in that the device user should have a means of knowing if an owner key is in use, but the requirement for the device to be unlocked implies both physical access the LOCK_DEVICE
being cleared.)
Verifying an operating system kernel and image is useful both for system reliability and for ensuring that the software has not been modified by a malicious party. However, if the system software is updated, malicious software may attempt to "roll" a device back to an older version in order to take advantage of mistakes in the older, verified code.
Rollback index values, or versions, may be stored securely by the bootloader to prevent these problems. The Verified Boot Storage applet provides eight 64-bit slots for storing a value. They may be read at any time, but they may only be written to when the device is in the bootloader (or fastboot).
Rollback storage is written to using ese_boot_rollback_index_write()
and read using ese_boot_rollback_index_read()
.
The applet supports two operational states:
On initial installation, production is false. When the applet is not in production mode, it does not enforce a number of security boundaries, such as requiring bootloader or hlos mode for lock toggling or CarrierLock verification. This allows the factory tools to run in any mode and properly configure a default lock state.
To transition to "production", a call to ese_boot_set_production(true)
is necessary.
To check the state and collect debugging information, the call ese_boot_get_state()
will return the current bootloader value, the production state, any errors codes from lock initialization, and the contents of lock storage.
After the applet is installed, a flow as follows would prepare the applet for use:
ese_boot_session_init();
ese_boot_session_open();
ese_boot_session_lock_xset(LOCK_OWNER, {0, ...});
ese_boot_session_lock_set(LOCK_BOOT, 0x1);
ese_boot_session_lock_set(LOCK_DEVICE, 0x1);
ese_boot_session_lock_set(LOCK_CARRIER, {1, deviceData});
ese_boot_session_set_production(true);
ese_boot_session_close();
Bootloader mode detection depends on hardware support to signal the applet that the application processor has been reset. Additionally, there is a mechanism for the bootloader to indicate that is loading the main OS where it flips the value. This signal provides the assurances around when storage is mutable or not during the time a device is powered on.
EseAppResult is an enum that all ese_boot_*
functions return. The enum only covers the lower 16-bits. The upper 16-bits are reserved for passing applet and secure element OS status for debugging and analysis. When the lower 16-bits are ESE_APP_RESULT_ERROR_APPLET
, then the upper bytes will be the applet code. That code can then be cross-referenced in the applet by function and code. If the lower bytes are ESE_APP_RESULT_ERROR_OS
, then the status code are the ISO7816 code from an uncaught exception or OS-level error.
ESE_APP_RESULT_ERROR_COOLDOWN
indicates that the secure element needs to stay powered on for a period of time -- either at the end of use or before the requested command can be serviced. As the behavior is implementation specific, the only effective option is to keep the secure element powered for the number of seconds specified by the response uint32_t
.
For chips that support it, like the one this applet is being tested on, the cooldown time can be requested with a special APDU to ese_transceive()
:
FFE10000
In response, a 6 byte response will contain a uint32_t
and a successful status code (90 00
). The unsigned little-endian integer indicates how many seconds the chip needs to stay powered and unused to cooldown. If this happens before the locks or rollback storage can be read, the bootloader will need to determine a safe delay or recovery path until boot can proceed securely.
There are many ways to integrate this library and the associated applet. Below are some concrete examples to guide standard approach.
# ese-boot-tool lock set carrier 1 modem-imei-string
# ese-boot-tool lock set device 1
# ese-boot-tool lock set boot 1
# ese-boot-tool lock set owner 0
# ese-boot-tool production set true
# ese-boot-tool production set false
LOCK_CARRIER
problem is being repaired, it is possible to reset the# ese-boot-tool lock reset
# ese-boot-tool lock set carrier 1 modem-imei-string
# ese-boot-tool lock set device 1
# ese-boot-tool lock set boot 1
# ese-boot-tool lock set owner 0
(To clear data from the owner lock, set owner 1 must be called with# ese-boot-tool production set true
Do not load any non-repair or non-factory OS without clearing the inBootloader signal as the applet may be transitioned out of production mode and/or the rollback state may be changed.
ese_boot_rollback_index_write()
ese_boot_rollback_index_read()
As rollback index values can only be written when inBootloader signal is set, it is critical to clear it when leaving the bootloader.
The pseudo-code and comments below should outline the basic algorithm, but it does not include integration into libavb or include use of rollback index value checking:
// Read LOCK_BOOT ese_boot_lock_get(session, kEseBootLockIdBoot, &lockBoot); if (lockBoot != 0x0) { // Boot is LOCKED. // Read the LOCK_OWNER ese_boot_lock_xget(session, kEseBootLockIdOwner, &lockOwner); if (lockOwner != 0x0) { // Owner is LOCKED // Get the lock owner value with metadata. // This is done as a second stage to avoid wasted copying when it // is not locked. uint8_t ownerData[kEseBootOwnerKeyMax + 1]; ese_boot_lock_xget(session, kEseBootLockIdOwner, ownerData sizeof(ownerData), &ownerDataUsed); // lockOwner == ownerData[0] // Parse the stored metadata into a key as per your bootloader // design. SomeBootKey key; parseDeviceOwnerKeyForBooting(ownerData + 1, ownerDataUsed, &key); // Boot using the supplied owner key // (E.g., as part of avb_validate_vbmeta_public_key()) setDeviceOwnerKeyForBooting(&key); continueBootFlow(); } else { // Boot is UNLOCKED (0x0) // Perform the boot flow. setBootIsUnverified(); continueBootFlow(); }
LOCK_BOOT
may be toggled by a fastboot command. If the conditions ofLOCK_OWNER
may be toggled and set a boot key from a fastboot commandese_boot_rollback_index_write()
on each slot with the value of 0.Note, LOCK_DEVICE
and LOCK_CARRIER
should not need to be used by fastboot.
For debugging and support, it may be desirable to connect the ese_boot_get_state()
to allow fastboot to return the current value of production, inbootloader, and the lock metadata.