U-Boot FIT Signature Verification
=================================

Introduction
------------
FIT supports hashing of images so that these hashes can be checked on
loading. This protects against corruption of the image. However it does not
prevent the substitution of one image for another.

The signature feature allows the hash to be signed with a private key such
that it can be verified using a public key later. Provided that the private
key is kept secret and the public key is stored in a non-volatile place,
any image can be verified in this way.

See verified-boot.txt for more general information on verified boot.


Concepts
--------
Some familiarity with public key cryptography is assumed in this section.

The procedure for signing is as follows:

   - hash an image in the FIT
   - sign the hash with a private key to produce a signature
   - store the resulting signature in the FIT

The procedure for verification is:

   - read the FIT
   - obtain the public key
   - extract the signature from the FIT
   - hash the image from the FIT
   - verify (with the public key) that the extracted signature matches the
       hash

The signing is generally performed by mkimage, as part of making a firmware
image for the device. The verification is normally done in U-Boot on the
device.


Algorithms
----------
In principle any suitable algorithm can be used to sign and verify a hash.
At present only one class of algorithms is supported: SHA1 hashing with RSA.
This works by hashing the image to produce a 20-byte hash.

While it is acceptable to bring in large cryptographic libraries such as
openssl on the host side (e.g. mkimage), it is not desirable for U-Boot.
For the run-time verification side, it is important to keep code and data
size as small as possible.

For this reason the RSA image verification uses pre-processed public keys
which can be used with a very small amount of code - just some extraction
of data from the FDT and exponentiation mod n. Code size impact is a little
under 5KB on Tegra Seaboard, for example.

It is relatively straightforward to add new algorithms if required. If
another RSA variant is needed, then it can be added to the table in
image-sig.c. If another algorithm is needed (such as DSA) then it can be
placed alongside rsa.c, and its functions added to the table in image-sig.c
also.


Creating an RSA key and certificate
-----------------------------------
To create a new public key, size 2048 bits:

$ openssl genpkey -algorithm RSA -out keys/dev.key \
    -pkeyopt rsa_keygen_bits:2048 -pkeyopt rsa_keygen_pubexp:65537

To create a certificate for this:

$ openssl req -batch -new -x509 -key keys/dev.key -out keys/dev.crt

If you like you can look at the public key also:

$ openssl rsa -in keys/dev.key -pubout


Device Tree Bindings
--------------------
The following properties are required in the FIT's signature node(s) to
allow thes signer to operate. These should be added to the .its file.
Signature nodes sit at the same level as hash nodes and are called
signature@1, signature@2, etc.

- algo: Algorithm name (e.g. "sha1,rs2048")

- key-name-hint: Name of key to use for signing. The keys will normally be in
a single directory (parameter -k to mkimage). For a given key <name>, its
private key is stored in <name>.key and the certificate is stored in
<name>.crt.

When the image is signed, the following properties are added (mandatory):

- value: The signature data (e.g. 256 bytes for 2048-bit RSA)

When the image is signed, the following properties are optional:

- timestamp: Time when image was signed (standard Unix time_t format)

- signer-name: Name of the signer (e.g. "mkimage")

- signer-version: Version string of the signer (e.g. "2013.01")

- comment: Additional information about the signer or image

For config bindings (see Signed Configurations below), the following
additional properties are optional:

- sign-images: A list of images to sign, each being a property of the conf
node that contains then. The default is "kernel,fdt" which means that these
two images will be looked up in the config and signed if present.

For config bindings, these properties are added by the signer:

- hashed-nodes: A list of nodes which were hashed by the signer. Each is
	a string - the full path to node. A typical value might be:

	hashed-nodes = "/", "/configurations/conf@1", "/images/kernel@1",
		"/images/kernel@1/hash@1", "/images/fdt@1",
		"/images/fdt@1/hash@1";

- hashed-strings: The start and size of the string region of the FIT that
	was hashed

Example: See sign-images.its for an example image tree source file and
sign-configs.its for config signing.


Public Key Storage
------------------
In order to verify an image that has been signed with a public key we need to
have a trusted public key. This cannot be stored in the signed image, since
it would be easy to alter. For this implementation we choose to store the
public key in U-Boot's control FDT (using CONFIG_OF_CONTROL).

Public keys should be stored as sub-nodes in a /signature node. Required
properties are:

- algo: Algorithm name (e.g. "sha1,rs2048")

Optional properties are:

- key-name-hint: Name of key used for signing. This is only a hint since it
is possible for the name to be changed. Verification can proceed by checking
all available signing keys until one matches.

- required: If present this indicates that the key must be verified for the
image / configuration to be considered valid. Only required keys are
normally verified by the FIT image booting algorithm. Valid values are
"image" to force verification of all images, and "conf" to force verfication
of the selected configuration (which then relies on hashes in the images to
verify those).

Each signing algorithm has its own additional properties.

For RSA the following are mandatory:

- rsa,num-bits: Number of key bits (e.g. 2048)
- rsa,modulus: Modulus (N) as a big-endian multi-word integer
- rsa,exponent: Public exponent (E) as a 64 bit unsigned integer
- rsa,r-squared: (2^num-bits)^2 as a big-endian multi-word integer
- rsa,n0-inverse: -1 / modulus[0] mod 2^32


Signed Configurations
---------------------
While signing images is useful, it does not provide complete protection
against several types of attack. For example, it it possible to create a
FIT with the same signed images, but with the configuration changed such
that a different one is selected (mix and match attack). It is also possible
to substitute a signed image from an older FIT version into a newer FIT
(roll-back attack).

As an example, consider this FIT:

/ {
	images {
		kernel@1 {
			data = <data for kernel1>
			signature@1 {
				algo = "sha1,rsa2048";
				value = <...kernel signature 1...>
			};
		};
		kernel@2 {
			data = <data for kernel2>
			signature@1 {
				algo = "sha1,rsa2048";
				value = <...kernel signature 2...>
			};
		};
		fdt@1 {
			data = <data for fdt1>;
			signature@1 {
				algo = "sha1,rsa2048";
				vaue = <...fdt signature 1...>
			};
		};
		fdt@2 {
			data = <data for fdt2>;
			signature@1 {
				algo = "sha1,rsa2048";
				vaue = <...fdt signature 2...>
			};
		};
	};
	configurations {
		default = "conf@1";
		conf@1 {
			kernel = "kernel@1";
			fdt = "fdt@1";
		};
		conf@1 {
			kernel = "kernel@2";
			fdt = "fdt@2";
		};
	};
};

Since both kernels are signed it is easy for an attacker to add a new
configuration 3 with kernel 1 and fdt 2:

	configurations {
		default = "conf@1";
		conf@1 {
			kernel = "kernel@1";
			fdt = "fdt@1";
		};
		conf@1 {
			kernel = "kernel@2";
			fdt = "fdt@2";
		};
		conf@3 {
			kernel = "kernel@1";
			fdt = "fdt@2";
		};
	};

With signed images, nothing protects against this. Whether it gains an
advantage for the attacker is debatable, but it is not secure.

To solved this problem, we support signed configurations. In this case it
is the configurations that are signed, not the image. Each image has its
own hash, and we include the hash in the configuration signature.

So the above example is adjusted to look like this:

/ {
	images {
		kernel@1 {
			data = <data for kernel1>
			hash@1 {
				algo = "sha1";
				value = <...kernel hash 1...>
			};
		};
		kernel@2 {
			data = <data for kernel2>
			hash@1 {
				algo = "sha1";
				value = <...kernel hash 2...>
			};
		};
		fdt@1 {
			data = <data for fdt1>;
			hash@1 {
				algo = "sha1";
				value = <...fdt hash 1...>
			};
		};
		fdt@2 {
			data = <data for fdt2>;
			hash@1 {
				algo = "sha1";
				value = <...fdt hash 2...>
			};
		};
	};
	configurations {
		default = "conf@1";
		conf@1 {
			kernel = "kernel@1";
			fdt = "fdt@1";
			signature@1 {
				algo = "sha1,rsa2048";
				value = <...conf 1 signature...>;
			};
		};
		conf@2 {
			kernel = "kernel@2";
			fdt = "fdt@2";
			signature@1 {
				algo = "sha1,rsa2048";
				value = <...conf 1 signature...>;
			};
		};
	};
};


You can see that we have added hashes for all images (since they are no
longer signed), and a signature to each configuration. In the above example,
mkimage will sign configurations/conf@1, the kernel and fdt that are
pointed to by the configuration (/images/kernel@1, /images/kernel@1/hash@1,
/images/fdt@1, /images/fdt@1/hash@1) and the root structure of the image
(so that it isn't possible to add or remove root nodes). The signature is
written into /configurations/conf@1/signature@1/value. It can easily be
verified later even if the FIT has been signed with other keys in the
meantime.


Verification
------------
FITs are verified when loaded. After the configuration is selected a list
of required images is produced. If there are 'required' public keys, then
each image must be verified against those keys. This means that every image
that might be used by the target needs to be signed with 'required' keys.

This happens automatically as part of a bootm command when FITs are used.


Enabling FIT Verification
-------------------------
In addition to the options to enable FIT itself, the following CONFIGs must
be enabled:

CONFIG_FIT_SIGNATURE - enable signing and verfication in FITs
CONFIG_RSA - enable RSA algorithm for signing

WARNING: When relying on signed FIT images with required signature check
the legacy image format is default disabled by not defining
CONFIG_IMAGE_FORMAT_LEGACY

Testing
-------
An easy way to test signing and verfication is to use the test script
provided in test/vboot/vboot_test.sh. This uses sandbox (a special version
of U-Boot which runs under Linux) to show the operation of a 'bootm'
command loading and verifying images.

A sample run is show below:

$ make O=sandbox sandbox_config
$ make O=sandbox
$ O=sandbox ./test/vboot/vboot_test.sh
Simple Verified Boot Test
=========================

Please see doc/uImage.FIT/verified-boot.txt for more information

/home/hs/ids/u-boot/sandbox/tools/mkimage -D -I dts -O dtb -p 2000
Build keys
do sha1 test
Build FIT with signed images
Test Verified Boot Run: unsigned signatures:: OK
Sign images
Test Verified Boot Run: signed images: OK
Build FIT with signed configuration
Test Verified Boot Run: unsigned config: OK
Sign images
Test Verified Boot Run: signed config: OK
check signed config on the host
Signature check OK
OK
Test Verified Boot Run: signed config: OK
Test Verified Boot Run: signed config with bad hash: OK
do sha256 test
Build FIT with signed images
Test Verified Boot Run: unsigned signatures:: OK
Sign images
Test Verified Boot Run: signed images: OK
Build FIT with signed configuration
Test Verified Boot Run: unsigned config: OK
Sign images
Test Verified Boot Run: signed config: OK
check signed config on the host
Signature check OK
OK
Test Verified Boot Run: signed config: OK
Test Verified Boot Run: signed config with bad hash: OK

Test passed


Future Work
-----------
- Roll-back protection using a TPM is done using the tpm command. This can
be scripted, but we might consider a default way of doing this, built into
bootm.


Possible Future Work
--------------------
- Add support for other RSA/SHA variants, such as rsa4096,sha512.
- Other algorithms besides RSA
- More sandbox tests for failure modes
- Passwords for keys/certificates
- Perhaps implement OAEP
- Enhance bootm to permit scripted signature verification (so that a script
can verify an image but not actually boot it)


Simon Glass
sjg@chromium.org
1-1-13