1 The trusted boot framework on Marvell Armada 38x 2 ================================================ 3 4 Contents: 5 6 1. Overview of the trusted boot 7 2. Terminology 8 3. Boot image layout 9 4. The secured header 10 5. The secured boot flow 11 6. Usage example 12 7. Work to be done 13 8. Bibliography 14 15 1. Overview of the trusted boot 16 ------------------------------- 17 18 The Armada's trusted boot framework enables the SoC to cryptographically verify 19 a specially prepared boot image. This can be used to establish a chain of trust 20 from the boot firmware all the way to the OS. 21 22 To achieve this, the Armada SoC requires a specially prepared boot image, which 23 contains the relevant cryptographic data, as well as other information 24 pertaining to the boot process. Furthermore, a eFuse structure (a 25 one-time-writeable memory) need to be configured in the correct way. 26 27 Roughly, the secure boot process works as follows: 28 29 * Load the header block of the boot image, extract a special "root" public RSA 30 key from it, and verify its SHA-256 hash against a SHA-256 stored in a eFuse 31 field. 32 * Load an array of code signing public RSA keys from the header block, and 33 verify its RSA signature (contained in the header block as well) using the 34 "root" RSA key. 35 * Choose a code signing key, and use it to verify the header block (excluding 36 the key array). 37 * Verify the binary image's signature (contained in the header block) using the 38 code signing key. 39 * If all checks pass successfully, boot the image. 40 41 The chain of trust is thus as follows: 42 43 * The SHA-256 value in the eFuse field verifies the "root" public key. 44 * The "root" public key verifies the code signing key array. 45 * The selected code signing key verifies the header block and the binary image. 46 47 In the special case of building a boot image containing U-Boot as the binary 48 image, which employs this trusted boot framework, the following tasks need to 49 be addressed: 50 51 1. Creation of the needed cryptographic key material. 52 2. Creation of a conforming boot image containing the U-Boot image as binary 53 image. 54 3. Burning the necessary eFuse values. 55 56 (1) will be addressed later, (2) will be taken care of by U-Boot's build 57 system (some user configuration is required, though), and for (3) the necessary 58 data (essentially a series of U-Boot commands to be entered at the U-Boot 59 command prompt) will be created by the build system as well. 60 61 The documentation of the trusted boot mode is contained in part 1, chapter 62 7.2.5 in the functional specification [1], and in application note [2]. 63 64 2. Terminology 65 -------------- 66 67 CSK - Code Signing Key(s): An array of RSA key pairs, which 68 are used to sign and verify the secured header and the 69 boot loader image. 70 KAK - Key Authentication Key: A RSA key pair, which is used 71 to sign and verify the array of CSKs. 72 Header block - The first part of the boot image, which contains the 73 image's headers (also known as "headers block", "boot 74 header", and "image header") 75 eFuse - A one-time-writeable memory. 76 BootROM - The Armada's built-in boot firmware, which is 77 responsible for verifying and starting secure images. 78 Boot image - The complete image the SoC's boot firmware loads 79 (contains the header block and the binary image) 80 Main header - The header in the header block containing information 81 and data pertaining to the boot process (used for both 82 the regular and secured boot processes) 83 Binary image - The binary code payload of the boot image; in this 84 case the U-Boot's code (also known as "source image", 85 or just "image") 86 Secured header - The specialized header in the header block that 87 contains information and data pertaining to the 88 trusted boot (also known as "security header") 89 Secured boot mode - A special boot mode of the Armada SoC in which secured 90 images are verified (non-secure images won't boot); 91 the mode is activated by setting a eFuse field. 92 Trusted debug mode - A special mode for the trusted boot that allows 93 debugging of devices employing the trusted boot 94 framework in a secure manner (untested in the current 95 implementation). 96 Trusted boot framework - The ARMADA SoC's implementation of a secure verified 97 boot process. 98 99 3. Boot image layout 100 -------------------- 101 102 +-- Boot image --------------------------------------------+ 103 | | 104 | +-- Header block --------------------------------------+ | 105 | | Main header | | 106 | +------------------------------------------------------+ | 107 | | Secured header | | 108 | +------------------------------------------------------+ | 109 | | BIN header(s) | | 110 | +------------------------------------------------------+ | 111 | | REG header(s) | | 112 | +------------------------------------------------------+ | 113 | | Padding | | 114 | +------------------------------------------------------+ | 115 | | 116 | +------------------------------------------------------+ | 117 | | Binary image + checksum | | 118 | +------------------------------------------------------+ | 119 +----------------------------------------------------------+ 120 121 4. The secured header 122 --------------------- 123 124 For the trusted boot framework, a additional header is added to the boot image. 125 The following data are relevant for the secure boot: 126 127 KAK: The KAK is contained in the secured header in the form 128 of a RSA-2048 public key in DER format with a length of 129 524 bytes. 130 Header block signature: The RSA signature of the header block (excluding the 131 CSK array), created using the selected CSK. 132 Binary image signature: The RSA signature of the binary image, created using 133 the selected CSK. 134 CSK array: The array of the 16 CSKs as RSA-2048 public keys in DER 135 format with a length of 8384 = 16 * 524 bytes. 136 CSK block signature: The RSA signature of the CSK array, created using the 137 KAK. 138 139 NOTE: The JTAG delay, Box ID, and Flash ID header fields do play a role in the 140 trusted boot process to enable and configure secure debugging, but they were 141 not tested in the current implementation of the trusted boot in U-Boot. 142 143 5. The secured boot flow 144 ------------------------ 145 146 The steps in the boot flow that are relevant for the trusted boot framework 147 proceed as follows: 148 149 1) Check if trusted boot is enabled, and perform regular boot if it is not. 150 2) Load the secured header, and verify its checksum. 151 3) Select the lowest valid CSK from CSK0 to CSK15. 152 4) Verify the SHA-256 hash of the KAK embedded in the secured header. 153 5) Verify the RSA signature of the CSK block from the secured header with the 154 KAK. 155 6) Verify the header block signature (which excludes the CSK block) from the 156 secured header with the selected CSK. 157 7) Load the binary image to the main memory and verify its checksum. 158 8) Verify the binary image's RSA signature from the secured header with the 159 selected CSK. 160 9) Continue the boot process as in the case of the regular boot. 161 162 NOTE: All RSA signatures are verified according to the PKCS #1 v2.1 standard 163 described in [3]. 164 165 NOTE: The Box ID and Flash ID are checked after step 6, and the trusted debug 166 mode may be entered there, but since this mode is untested in the current 167 implementation, it is not described further. 168 169 6. Usage example 170 ---------------- 171 172 ### Create key material 173 174 To employ the trusted boot framework, cryptographic key material needs to be 175 created. In the current implementation, two keys are needed to build a valid 176 secured boot image: The KAK private key and a CSK private key (both have to be 177 2048 bit RSA keys in PEM format). Note that the usage of more than one CSK is 178 currently not supported. 179 180 NOTE: Since the public key can be generated from the private key, it is 181 sufficient to store the private key for each key pair. 182 183 OpenSSL can be used to generate the needed files kwb_kak.key and kwb_csk.key 184 (the names of these files have to be configured, see the next section on 185 kwbimage.cfg settings): 186 187 openssl genrsa -out kwb_kak.key 2048 188 openssl genrsa -out kwb_csk.key 2048 189 190 The generated files have to be placed in the U-Boot root directory. 191 192 Alternatively, instead of copying the files, symlinks to the private keys can 193 be placed in the U-Boot root directory. 194 195 WARNING: Knowledge of the KAK or CSK private key would enable an attacker to 196 generate secured boot images containing arbitrary code. Hence, the private keys 197 should be carefully guarded. 198 199 ### Create/Modifiy kwbimage.cfg 200 201 The Kirkwook architecture in U-Boot employs a special board-specific 202 configuration file (kwbimage.cfg), which controls various boot image settings 203 that are interpreted by the BootROM, such as the boot medium. The support the 204 trusted boot framework, several new options were added to faciliate 205 configuration of the secured boot. 206 207 The configuration file's layout has been retained, only the following new 208 options were added: 209 210 KAK - The name of the KAK RSA private key file in the U-Boot 211 root directory, without the trailing extension of ".key". 212 CSK - The name of the (active) CSK RSA private key file in the 213 U-Boot root directory, without the trailing extension of 214 ".key". 215 BOX_ID - The BoxID to be used for trusted debugging (a integer 216 value). 217 FLASH_ID - The FlashID to be used for trusted debugging (a integer 218 value). 219 JTAG_DELAY - The JTAG delay to be used for trusted debugging (a 220 integer value). 221 CSK_INDEX - The index of the active CSK (a integer value). 222 SEC_SPECIALIZED_IMG - Flag to indicate whether to include the BoxID and FlashID 223 in the image (that is, whether to use the trusted debug 224 mode or not); no parameters. 225 SEC_BOOT_DEV - The boot device from which the trusted boot is allowed to 226 proceed, identified via a numeric ID. The tested values 227 are 0x34 = NOR flash, 0x31 = SDIO/MMC card; for 228 additional ID values, consult the documentation in [1]. 229 SEC_FUSE_DUMP - Dump the "fuse prog" commands necessary for writing the 230 correct eFuse values to a text file in the U-Boot root 231 directory. The parameter is the architecture for which to 232 dump the commands (currently only "a38x" is supported). 233 234 The parameter values may be hardcoded into the file, but it is also possible to 235 employ a dynamic approach of creating a Autoconf-like kwbimage.cfg.in, then 236 reading configuration values from Kconfig options or from the board config 237 file, and generating the actual kwbimage.cfg from this template using Makefile 238 mechanisms (see board/gdsys/a38x/Makefile as an example for this approach). 239 240 ### Set config options 241 242 To enable the generation of trusted boot images, the corresponding support 243 needs to be activated, and a index for the active CSK needs to be selected as 244 well. 245 246 Furthermore, eFuse writing support has to be activated in order to burn the 247 eFuse structure's values (this option is just needed for programming the eFuse 248 structure; production boot images may disable it). 249 250 ARM architecture 251 -> [*] Build image for trusted boot 252 (0) Index of active CSK 253 -> [*] Enable eFuse support 254 [ ] Fake eFuse access (dry run) 255 256 ### Build and test boot image 257 258 The creation of the boot image is done via the usual invocation of make (with a 259 suitably set CROSS_COMPILE environment variable, of course). The resulting boot 260 image u-boot-spl.kwb can then be tested, if so desired. The hdrparser from [5] 261 can be used for this purpose. To build the tool, invoke make in the 262 'tools/marvell/doimage_mv' directory of [5], which builds a stand-alone 263 hdrparser executable. A test can be conducted by calling hdrparser with the 264 produced boot image and the following (mandatory) parameters: 265 266 ./hdrparser -k 0 -t u-boot-spl.kwb 267 268 Here we assume that the CSK index is 0 and the boot image file resides in the 269 same directory (adapt accordingly if needed). The tool should report that all 270 checksums are valid ("GOOD"), that all signature verifications succeed 271 ("PASSED"), and, finally, that the overall test was successful 272 ("T E S T S U C C E E D E D" in the last line of output). 273 274 ### Burn eFuse structure 275 276 +----------------------------------------------------------+ 277 | WARNING: Burning the eFuse structure is a irreversible | 278 | operation! Should wrong or corrupted values be used, the | 279 | board won't boot anymore, and recovery is likely | 280 | impossible! | 281 +----------------------------------------------------------+ 282 283 After the build process has finished, and the SEC_FUSE_DUMP option was set in 284 the kwbimage.cfg was set, a text file kwb_fuses_a38x.txt should be present in 285 the U-Boot top-level directory. It contains all the necessary commands to set 286 the eFuse structure to the values needed for the used KAK digest, as well as 287 the CSK index, Flash ID and Box ID that were selected in kwbimage.cfg. 288 289 Sequentially executing the commands in this file at the U-Boot command prompt 290 will write these values to the eFuse structure. 291 292 If the SEC_FUSE_DUMP option was not set, the commands needed to burn the fuses 293 have to be crafted by hand. The needed fuse lines can be looked up in [1]; a 294 rough overview of the process is: 295 296 * Burn the KAK public key hash. The hash itself can be found in the file 297 pub_kak_hash.txt in the U-Boot top-level directory; be careful to account for 298 the endianness! 299 * Burn the CSK selection, BoxID, and FlashID 300 * Enable trusted boot by burning the corresponding fuse (WARNING: this must be 301 the last fuse line written!) 302 * Lock the unused fuse lines 303 304 The command to employ is the "fuse prog" command previously enabled by setting 305 the corresponding configuration option. 306 307 For the trusted boot, the fuse prog command has a special syntax, since the 308 ARMADA SoC demands that whole fuse lines (64 bit values) have to be written as 309 a whole. The fuse prog command itself allows lists of 32 bit words to be 310 written at a time, but this is translated to a series of single 32 bit write 311 operations to the fuse line, where the individual 32 bit words are identified 312 by a "word" counter that is increased for each write. 313 314 To work around this restriction, we interpret each line to have three "words" 315 (0-2): The first and second words are the values to be written to the fuse 316 line, and the third is a lock flag, which is supposed to lock the fuse line 317 when set to 1. Writes to the first and second words are memoized between 318 function calls, and the fuse line is only really written and locked (on writing 319 the third word) if both words were previously set, so that "incomplete" writes 320 are prevented. An exception to this is a single write to the third word (index 321 2) without previously writing neither the first nor the second word, which 322 locks the fuse line without setting any value; this is needed to lock the 323 unused fuse lines. 324 325 As an example, to write the value 0011223344556677 to fuse line 10, we would 326 use the following command: 327 328 fuse prog -y 10 0 00112233 44556677 1 329 330 Here 10 is the fuse line number, 0 is the index of the first word to be 331 written, 00112233 and 44556677 are the values to be written to the fuse line 332 (first and second word) and the trailing 1 is the value for the third word 333 responsible for locking the line. 334 335 A "lock-only" command would look like this: 336 337 fuse prog -y 11 2 1 338 339 Here 11 is the fuse number, 2 is the index of the first word to be written 340 (notice that we only write to word 2 here; the third word for fuse line 341 locking), and the 1 is the value for the word we are writing to. 342 343 WARNING: According to application note [4], the VHV pin of the SoC must be 344 connected to a 1.8V source during eFuse programming, but *must* be disconnected 345 for normal operation. The AN [4] describes a software-controlled circuit (based 346 on a N-channel or P-channel FET and a free GPIO pin of the SoC) to achieve 347 this, but a jumper-based circuit should suffice as well. Regardless of the 348 chosen circuit, the issue needs to be addressed accordingly! 349 350 7. Work to be done 351 ------------------ 352 353 * Add the ability to populate more than one CSK 354 * Test secure debug 355 * Test on Armada XP 356 357 8. Bibliography 358 --------------- 359 360 [1] ARMADA(R) 38x Family High-Performance Single/Dual CPU System on Chip 361 Functional Specification; MV-S109094-00, Rev. C; August 2, 2015, 362 Preliminary 363 [2] AN-383: ARMADA(R) 38x Families Secure Boot Mode Support; MV-S302501-00 364 Rev. A; March 11, 2015, Preliminary 365 [3] Public-Key Cryptography Standards (PKCS) #1: RSA Cryptography 366 Specifications Version 2.1; February 2003; 367 https://www.ietf.org/rfc/rfc3447.txt 368 [4] AN-389: ARMADA(R) VHV Power; MV-S302545-00 Rev. B; January 28, 2016, 369 Released 370 [5] Marvell Armada 38x U-Boot support; November 25, 2015; 371 https://github.com/MarvellEmbeddedProcessors/u-boot-marvell 372 373 2017-01-05, Mario Six <mario.six (a] gdsys.cc> 374