1/* 2 * Cryptographic API. 3 * 4 * Support for VIA PadLock hardware crypto engine. 5 * 6 * Copyright (c) 2004 Michal Ludvig <michal@logix.cz> 7 * 8 */ 9 10#include <crypto/algapi.h> 11#include <crypto/aes.h> 12#include <crypto/padlock.h> 13#include <linux/module.h> 14#include <linux/init.h> 15#include <linux/types.h> 16#include <linux/errno.h> 17#include <linux/interrupt.h> 18#include <linux/kernel.h> 19#include <linux/percpu.h> 20#include <linux/smp.h> 21#include <linux/slab.h> 22#include <asm/cpu_device_id.h> 23#include <asm/byteorder.h> 24#include <asm/processor.h> 25#include <asm/fpu/api.h> 26 27/* 28 * Number of data blocks actually fetched for each xcrypt insn. 29 * Processors with prefetch errata will fetch extra blocks. 30 */ 31static unsigned int ecb_fetch_blocks = 2; 32#define MAX_ECB_FETCH_BLOCKS (8) 33#define ecb_fetch_bytes (ecb_fetch_blocks * AES_BLOCK_SIZE) 34 35static unsigned int cbc_fetch_blocks = 1; 36#define MAX_CBC_FETCH_BLOCKS (4) 37#define cbc_fetch_bytes (cbc_fetch_blocks * AES_BLOCK_SIZE) 38 39/* Control word. */ 40struct cword { 41 unsigned int __attribute__ ((__packed__)) 42 rounds:4, 43 algo:3, 44 keygen:1, 45 interm:1, 46 encdec:1, 47 ksize:2; 48} __attribute__ ((__aligned__(PADLOCK_ALIGNMENT))); 49 50/* Whenever making any changes to the following 51 * structure *make sure* you keep E, d_data 52 * and cword aligned on 16 Bytes boundaries and 53 * the Hardware can access 16 * 16 bytes of E and d_data 54 * (only the first 15 * 16 bytes matter but the HW reads 55 * more). 56 */ 57struct aes_ctx { 58 u32 E[AES_MAX_KEYLENGTH_U32] 59 __attribute__ ((__aligned__(PADLOCK_ALIGNMENT))); 60 u32 d_data[AES_MAX_KEYLENGTH_U32] 61 __attribute__ ((__aligned__(PADLOCK_ALIGNMENT))); 62 struct { 63 struct cword encrypt; 64 struct cword decrypt; 65 } cword; 66 u32 *D; 67}; 68 69static DEFINE_PER_CPU(struct cword *, paes_last_cword); 70 71/* Tells whether the ACE is capable to generate 72 the extended key for a given key_len. */ 73static inline int 74aes_hw_extkey_available(uint8_t key_len) 75{ 76 /* TODO: We should check the actual CPU model/stepping 77 as it's possible that the capability will be 78 added in the next CPU revisions. */ 79 if (key_len == 16) 80 return 1; 81 return 0; 82} 83 84static inline struct aes_ctx *aes_ctx_common(void *ctx) 85{ 86 unsigned long addr = (unsigned long)ctx; 87 unsigned long align = PADLOCK_ALIGNMENT; 88 89 if (align <= crypto_tfm_ctx_alignment()) 90 align = 1; 91 return (struct aes_ctx *)ALIGN(addr, align); 92} 93 94static inline struct aes_ctx *aes_ctx(struct crypto_tfm *tfm) 95{ 96 return aes_ctx_common(crypto_tfm_ctx(tfm)); 97} 98 99static inline struct aes_ctx *blk_aes_ctx(struct crypto_blkcipher *tfm) 100{ 101 return aes_ctx_common(crypto_blkcipher_ctx(tfm)); 102} 103 104static int aes_set_key(struct crypto_tfm *tfm, const u8 *in_key, 105 unsigned int key_len) 106{ 107 struct aes_ctx *ctx = aes_ctx(tfm); 108 const __le32 *key = (const __le32 *)in_key; 109 u32 *flags = &tfm->crt_flags; 110 struct crypto_aes_ctx gen_aes; 111 int cpu; 112 113 if (key_len % 8) { 114 *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN; 115 return -EINVAL; 116 } 117 118 /* 119 * If the hardware is capable of generating the extended key 120 * itself we must supply the plain key for both encryption 121 * and decryption. 122 */ 123 ctx->D = ctx->E; 124 125 ctx->E[0] = le32_to_cpu(key[0]); 126 ctx->E[1] = le32_to_cpu(key[1]); 127 ctx->E[2] = le32_to_cpu(key[2]); 128 ctx->E[3] = le32_to_cpu(key[3]); 129 130 /* Prepare control words. */ 131 memset(&ctx->cword, 0, sizeof(ctx->cword)); 132 133 ctx->cword.decrypt.encdec = 1; 134 ctx->cword.encrypt.rounds = 10 + (key_len - 16) / 4; 135 ctx->cword.decrypt.rounds = ctx->cword.encrypt.rounds; 136 ctx->cword.encrypt.ksize = (key_len - 16) / 8; 137 ctx->cword.decrypt.ksize = ctx->cword.encrypt.ksize; 138 139 /* Don't generate extended keys if the hardware can do it. */ 140 if (aes_hw_extkey_available(key_len)) 141 goto ok; 142 143 ctx->D = ctx->d_data; 144 ctx->cword.encrypt.keygen = 1; 145 ctx->cword.decrypt.keygen = 1; 146 147 if (crypto_aes_expand_key(&gen_aes, in_key, key_len)) { 148 *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN; 149 return -EINVAL; 150 } 151 152 memcpy(ctx->E, gen_aes.key_enc, AES_MAX_KEYLENGTH); 153 memcpy(ctx->D, gen_aes.key_dec, AES_MAX_KEYLENGTH); 154 155ok: 156 for_each_online_cpu(cpu) 157 if (&ctx->cword.encrypt == per_cpu(paes_last_cword, cpu) || 158 &ctx->cword.decrypt == per_cpu(paes_last_cword, cpu)) 159 per_cpu(paes_last_cword, cpu) = NULL; 160 161 return 0; 162} 163 164/* ====== Encryption/decryption routines ====== */ 165 166/* These are the real call to PadLock. */ 167static inline void padlock_reset_key(struct cword *cword) 168{ 169 int cpu = raw_smp_processor_id(); 170 171 if (cword != per_cpu(paes_last_cword, cpu)) 172#ifndef CONFIG_X86_64 173 asm volatile ("pushfl; popfl"); 174#else 175 asm volatile ("pushfq; popfq"); 176#endif 177} 178 179static inline void padlock_store_cword(struct cword *cword) 180{ 181 per_cpu(paes_last_cword, raw_smp_processor_id()) = cword; 182} 183 184/* 185 * While the padlock instructions don't use FP/SSE registers, they 186 * generate a spurious DNA fault when cr0.ts is '1'. These instructions 187 * should be used only inside the irq_ts_save/restore() context 188 */ 189 190static inline void rep_xcrypt_ecb(const u8 *input, u8 *output, void *key, 191 struct cword *control_word, int count) 192{ 193 asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ 194 : "+S"(input), "+D"(output) 195 : "d"(control_word), "b"(key), "c"(count)); 196} 197 198static inline u8 *rep_xcrypt_cbc(const u8 *input, u8 *output, void *key, 199 u8 *iv, struct cword *control_word, int count) 200{ 201 asm volatile (".byte 0xf3,0x0f,0xa7,0xd0" /* rep xcryptcbc */ 202 : "+S" (input), "+D" (output), "+a" (iv) 203 : "d" (control_word), "b" (key), "c" (count)); 204 return iv; 205} 206 207static void ecb_crypt_copy(const u8 *in, u8 *out, u32 *key, 208 struct cword *cword, int count) 209{ 210 /* 211 * Padlock prefetches extra data so we must provide mapped input buffers. 212 * Assume there are at least 16 bytes of stack already in use. 213 */ 214 u8 buf[AES_BLOCK_SIZE * (MAX_ECB_FETCH_BLOCKS - 1) + PADLOCK_ALIGNMENT - 1]; 215 u8 *tmp = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT); 216 217 memcpy(tmp, in, count * AES_BLOCK_SIZE); 218 rep_xcrypt_ecb(tmp, out, key, cword, count); 219} 220 221static u8 *cbc_crypt_copy(const u8 *in, u8 *out, u32 *key, 222 u8 *iv, struct cword *cword, int count) 223{ 224 /* 225 * Padlock prefetches extra data so we must provide mapped input buffers. 226 * Assume there are at least 16 bytes of stack already in use. 227 */ 228 u8 buf[AES_BLOCK_SIZE * (MAX_CBC_FETCH_BLOCKS - 1) + PADLOCK_ALIGNMENT - 1]; 229 u8 *tmp = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT); 230 231 memcpy(tmp, in, count * AES_BLOCK_SIZE); 232 return rep_xcrypt_cbc(tmp, out, key, iv, cword, count); 233} 234 235static inline void ecb_crypt(const u8 *in, u8 *out, u32 *key, 236 struct cword *cword, int count) 237{ 238 /* Padlock in ECB mode fetches at least ecb_fetch_bytes of data. 239 * We could avoid some copying here but it's probably not worth it. 240 */ 241 if (unlikely(((unsigned long)in & ~PAGE_MASK) + ecb_fetch_bytes > PAGE_SIZE)) { 242 ecb_crypt_copy(in, out, key, cword, count); 243 return; 244 } 245 246 rep_xcrypt_ecb(in, out, key, cword, count); 247} 248 249static inline u8 *cbc_crypt(const u8 *in, u8 *out, u32 *key, 250 u8 *iv, struct cword *cword, int count) 251{ 252 /* Padlock in CBC mode fetches at least cbc_fetch_bytes of data. */ 253 if (unlikely(((unsigned long)in & ~PAGE_MASK) + cbc_fetch_bytes > PAGE_SIZE)) 254 return cbc_crypt_copy(in, out, key, iv, cword, count); 255 256 return rep_xcrypt_cbc(in, out, key, iv, cword, count); 257} 258 259static inline void padlock_xcrypt_ecb(const u8 *input, u8 *output, void *key, 260 void *control_word, u32 count) 261{ 262 u32 initial = count & (ecb_fetch_blocks - 1); 263 264 if (count < ecb_fetch_blocks) { 265 ecb_crypt(input, output, key, control_word, count); 266 return; 267 } 268 269 if (initial) 270 asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ 271 : "+S"(input), "+D"(output) 272 : "d"(control_word), "b"(key), "c"(initial)); 273 274 asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ 275 : "+S"(input), "+D"(output) 276 : "d"(control_word), "b"(key), "c"(count - initial)); 277} 278 279static inline u8 *padlock_xcrypt_cbc(const u8 *input, u8 *output, void *key, 280 u8 *iv, void *control_word, u32 count) 281{ 282 u32 initial = count & (cbc_fetch_blocks - 1); 283 284 if (count < cbc_fetch_blocks) 285 return cbc_crypt(input, output, key, iv, control_word, count); 286 287 if (initial) 288 asm volatile (".byte 0xf3,0x0f,0xa7,0xd0" /* rep xcryptcbc */ 289 : "+S" (input), "+D" (output), "+a" (iv) 290 : "d" (control_word), "b" (key), "c" (initial)); 291 292 asm volatile (".byte 0xf3,0x0f,0xa7,0xd0" /* rep xcryptcbc */ 293 : "+S" (input), "+D" (output), "+a" (iv) 294 : "d" (control_word), "b" (key), "c" (count-initial)); 295 return iv; 296} 297 298static void aes_encrypt(struct crypto_tfm *tfm, u8 *out, const u8 *in) 299{ 300 struct aes_ctx *ctx = aes_ctx(tfm); 301 int ts_state; 302 303 padlock_reset_key(&ctx->cword.encrypt); 304 ts_state = irq_ts_save(); 305 ecb_crypt(in, out, ctx->E, &ctx->cword.encrypt, 1); 306 irq_ts_restore(ts_state); 307 padlock_store_cword(&ctx->cword.encrypt); 308} 309 310static void aes_decrypt(struct crypto_tfm *tfm, u8 *out, const u8 *in) 311{ 312 struct aes_ctx *ctx = aes_ctx(tfm); 313 int ts_state; 314 315 padlock_reset_key(&ctx->cword.encrypt); 316 ts_state = irq_ts_save(); 317 ecb_crypt(in, out, ctx->D, &ctx->cword.decrypt, 1); 318 irq_ts_restore(ts_state); 319 padlock_store_cword(&ctx->cword.encrypt); 320} 321 322static struct crypto_alg aes_alg = { 323 .cra_name = "aes", 324 .cra_driver_name = "aes-padlock", 325 .cra_priority = PADLOCK_CRA_PRIORITY, 326 .cra_flags = CRYPTO_ALG_TYPE_CIPHER, 327 .cra_blocksize = AES_BLOCK_SIZE, 328 .cra_ctxsize = sizeof(struct aes_ctx), 329 .cra_alignmask = PADLOCK_ALIGNMENT - 1, 330 .cra_module = THIS_MODULE, 331 .cra_u = { 332 .cipher = { 333 .cia_min_keysize = AES_MIN_KEY_SIZE, 334 .cia_max_keysize = AES_MAX_KEY_SIZE, 335 .cia_setkey = aes_set_key, 336 .cia_encrypt = aes_encrypt, 337 .cia_decrypt = aes_decrypt, 338 } 339 } 340}; 341 342static int ecb_aes_encrypt(struct blkcipher_desc *desc, 343 struct scatterlist *dst, struct scatterlist *src, 344 unsigned int nbytes) 345{ 346 struct aes_ctx *ctx = blk_aes_ctx(desc->tfm); 347 struct blkcipher_walk walk; 348 int err; 349 int ts_state; 350 351 padlock_reset_key(&ctx->cword.encrypt); 352 353 blkcipher_walk_init(&walk, dst, src, nbytes); 354 err = blkcipher_walk_virt(desc, &walk); 355 356 ts_state = irq_ts_save(); 357 while ((nbytes = walk.nbytes)) { 358 padlock_xcrypt_ecb(walk.src.virt.addr, walk.dst.virt.addr, 359 ctx->E, &ctx->cword.encrypt, 360 nbytes / AES_BLOCK_SIZE); 361 nbytes &= AES_BLOCK_SIZE - 1; 362 err = blkcipher_walk_done(desc, &walk, nbytes); 363 } 364 irq_ts_restore(ts_state); 365 366 padlock_store_cword(&ctx->cword.encrypt); 367 368 return err; 369} 370 371static int ecb_aes_decrypt(struct blkcipher_desc *desc, 372 struct scatterlist *dst, struct scatterlist *src, 373 unsigned int nbytes) 374{ 375 struct aes_ctx *ctx = blk_aes_ctx(desc->tfm); 376 struct blkcipher_walk walk; 377 int err; 378 int ts_state; 379 380 padlock_reset_key(&ctx->cword.decrypt); 381 382 blkcipher_walk_init(&walk, dst, src, nbytes); 383 err = blkcipher_walk_virt(desc, &walk); 384 385 ts_state = irq_ts_save(); 386 while ((nbytes = walk.nbytes)) { 387 padlock_xcrypt_ecb(walk.src.virt.addr, walk.dst.virt.addr, 388 ctx->D, &ctx->cword.decrypt, 389 nbytes / AES_BLOCK_SIZE); 390 nbytes &= AES_BLOCK_SIZE - 1; 391 err = blkcipher_walk_done(desc, &walk, nbytes); 392 } 393 irq_ts_restore(ts_state); 394 395 padlock_store_cword(&ctx->cword.encrypt); 396 397 return err; 398} 399 400static struct crypto_alg ecb_aes_alg = { 401 .cra_name = "ecb(aes)", 402 .cra_driver_name = "ecb-aes-padlock", 403 .cra_priority = PADLOCK_COMPOSITE_PRIORITY, 404 .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER, 405 .cra_blocksize = AES_BLOCK_SIZE, 406 .cra_ctxsize = sizeof(struct aes_ctx), 407 .cra_alignmask = PADLOCK_ALIGNMENT - 1, 408 .cra_type = &crypto_blkcipher_type, 409 .cra_module = THIS_MODULE, 410 .cra_u = { 411 .blkcipher = { 412 .min_keysize = AES_MIN_KEY_SIZE, 413 .max_keysize = AES_MAX_KEY_SIZE, 414 .setkey = aes_set_key, 415 .encrypt = ecb_aes_encrypt, 416 .decrypt = ecb_aes_decrypt, 417 } 418 } 419}; 420 421static int cbc_aes_encrypt(struct blkcipher_desc *desc, 422 struct scatterlist *dst, struct scatterlist *src, 423 unsigned int nbytes) 424{ 425 struct aes_ctx *ctx = blk_aes_ctx(desc->tfm); 426 struct blkcipher_walk walk; 427 int err; 428 int ts_state; 429 430 padlock_reset_key(&ctx->cword.encrypt); 431 432 blkcipher_walk_init(&walk, dst, src, nbytes); 433 err = blkcipher_walk_virt(desc, &walk); 434 435 ts_state = irq_ts_save(); 436 while ((nbytes = walk.nbytes)) { 437 u8 *iv = padlock_xcrypt_cbc(walk.src.virt.addr, 438 walk.dst.virt.addr, ctx->E, 439 walk.iv, &ctx->cword.encrypt, 440 nbytes / AES_BLOCK_SIZE); 441 memcpy(walk.iv, iv, AES_BLOCK_SIZE); 442 nbytes &= AES_BLOCK_SIZE - 1; 443 err = blkcipher_walk_done(desc, &walk, nbytes); 444 } 445 irq_ts_restore(ts_state); 446 447 padlock_store_cword(&ctx->cword.decrypt); 448 449 return err; 450} 451 452static int cbc_aes_decrypt(struct blkcipher_desc *desc, 453 struct scatterlist *dst, struct scatterlist *src, 454 unsigned int nbytes) 455{ 456 struct aes_ctx *ctx = blk_aes_ctx(desc->tfm); 457 struct blkcipher_walk walk; 458 int err; 459 int ts_state; 460 461 padlock_reset_key(&ctx->cword.encrypt); 462 463 blkcipher_walk_init(&walk, dst, src, nbytes); 464 err = blkcipher_walk_virt(desc, &walk); 465 466 ts_state = irq_ts_save(); 467 while ((nbytes = walk.nbytes)) { 468 padlock_xcrypt_cbc(walk.src.virt.addr, walk.dst.virt.addr, 469 ctx->D, walk.iv, &ctx->cword.decrypt, 470 nbytes / AES_BLOCK_SIZE); 471 nbytes &= AES_BLOCK_SIZE - 1; 472 err = blkcipher_walk_done(desc, &walk, nbytes); 473 } 474 475 irq_ts_restore(ts_state); 476 477 padlock_store_cword(&ctx->cword.encrypt); 478 479 return err; 480} 481 482static struct crypto_alg cbc_aes_alg = { 483 .cra_name = "cbc(aes)", 484 .cra_driver_name = "cbc-aes-padlock", 485 .cra_priority = PADLOCK_COMPOSITE_PRIORITY, 486 .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER, 487 .cra_blocksize = AES_BLOCK_SIZE, 488 .cra_ctxsize = sizeof(struct aes_ctx), 489 .cra_alignmask = PADLOCK_ALIGNMENT - 1, 490 .cra_type = &crypto_blkcipher_type, 491 .cra_module = THIS_MODULE, 492 .cra_u = { 493 .blkcipher = { 494 .min_keysize = AES_MIN_KEY_SIZE, 495 .max_keysize = AES_MAX_KEY_SIZE, 496 .ivsize = AES_BLOCK_SIZE, 497 .setkey = aes_set_key, 498 .encrypt = cbc_aes_encrypt, 499 .decrypt = cbc_aes_decrypt, 500 } 501 } 502}; 503 504static struct x86_cpu_id padlock_cpu_id[] = { 505 X86_FEATURE_MATCH(X86_FEATURE_XCRYPT), 506 {} 507}; 508MODULE_DEVICE_TABLE(x86cpu, padlock_cpu_id); 509 510static int __init padlock_init(void) 511{ 512 int ret; 513 struct cpuinfo_x86 *c = &cpu_data(0); 514 515 if (!x86_match_cpu(padlock_cpu_id)) 516 return -ENODEV; 517 518 if (!cpu_has_xcrypt_enabled) { 519 printk(KERN_NOTICE PFX "VIA PadLock detected, but not enabled. Hmm, strange...\n"); 520 return -ENODEV; 521 } 522 523 if ((ret = crypto_register_alg(&aes_alg))) 524 goto aes_err; 525 526 if ((ret = crypto_register_alg(&ecb_aes_alg))) 527 goto ecb_aes_err; 528 529 if ((ret = crypto_register_alg(&cbc_aes_alg))) 530 goto cbc_aes_err; 531 532 printk(KERN_NOTICE PFX "Using VIA PadLock ACE for AES algorithm.\n"); 533 534 if (c->x86 == 6 && c->x86_model == 15 && c->x86_mask == 2) { 535 ecb_fetch_blocks = MAX_ECB_FETCH_BLOCKS; 536 cbc_fetch_blocks = MAX_CBC_FETCH_BLOCKS; 537 printk(KERN_NOTICE PFX "VIA Nano stepping 2 detected: enabling workaround.\n"); 538 } 539 540out: 541 return ret; 542 543cbc_aes_err: 544 crypto_unregister_alg(&ecb_aes_alg); 545ecb_aes_err: 546 crypto_unregister_alg(&aes_alg); 547aes_err: 548 printk(KERN_ERR PFX "VIA PadLock AES initialization failed.\n"); 549 goto out; 550} 551 552static void __exit padlock_fini(void) 553{ 554 crypto_unregister_alg(&cbc_aes_alg); 555 crypto_unregister_alg(&ecb_aes_alg); 556 crypto_unregister_alg(&aes_alg); 557} 558 559module_init(padlock_init); 560module_exit(padlock_fini); 561 562MODULE_DESCRIPTION("VIA PadLock AES algorithm support"); 563MODULE_LICENSE("GPL"); 564MODULE_AUTHOR("Michal Ludvig"); 565 566MODULE_ALIAS_CRYPTO("aes"); 567