?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
crypto: arm64/aes-ghash - yield NEON after every block of input
[linux/fpc-iii.git] / drivers / crypto / padlock-sha.c
blobd32c7932887697afcfc8350215805b4972dc560b
1 /*
2 * Cryptographic API.
3 *
4 * Support for VIA PadLock hardware crypto engine.
5 *
6 * Copyright (c) 2006 Michal Ludvig <michal@logix.cz>
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 */
14
15 #include <crypto/internal/hash.h>
16 #include <crypto/padlock.h>
17 #include <crypto/sha.h>
18 #include <linux/err.h>
19 #include <linux/module.h>
20 #include <linux/init.h>
21 #include <linux/errno.h>
22 #include <linux/interrupt.h>
23 #include <linux/kernel.h>
24 #include <linux/scatterlist.h>
25 #include <asm/cpu_device_id.h>
26 #include <asm/fpu/api.h>
27
28 struct padlock_sha_desc {
29 struct shash_desc fallback;
30 };
31
32 struct padlock_sha_ctx {
33 struct crypto_shash *fallback;
34 };
35
36 static int padlock_sha_init(struct shash_desc *desc)
37 {
38 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
39 struct padlock_sha_ctx *ctx = crypto_shash_ctx(desc->tfm);
40
41 dctx->fallback.tfm = ctx->fallback;
42 dctx->fallback.flags = desc->flags & CRYPTO_TFM_REQ_MAY_SLEEP;
43 return crypto_shash_init(&dctx->fallback);
44 }
45
46 static int padlock_sha_update(struct shash_desc *desc,
47 const u8 *data, unsigned int length)
48 {
49 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
50
51 dctx->fallback.flags = desc->flags & CRYPTO_TFM_REQ_MAY_SLEEP;
52 return crypto_shash_update(&dctx->fallback, data, length);
53 }
54
55 static int padlock_sha_export(struct shash_desc *desc, void *out)
56 {
57 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
58
59 return crypto_shash_export(&dctx->fallback, out);
60 }
61
62 static int padlock_sha_import(struct shash_desc *desc, const void *in)
63 {
64 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
65 struct padlock_sha_ctx *ctx = crypto_shash_ctx(desc->tfm);
66
67 dctx->fallback.tfm = ctx->fallback;
68 dctx->fallback.flags = desc->flags & CRYPTO_TFM_REQ_MAY_SLEEP;
69 return crypto_shash_import(&dctx->fallback, in);
70 }
71
72 static inline void padlock_output_block(uint32_t *src,
73 uint32_t *dst, size_t count)
74 {
75 while (count--)
76 *dst++ = swab32(*src++);
77 }
78
79 static int padlock_sha1_finup(struct shash_desc *desc, const u8 *in,
80 unsigned int count, u8 *out)
81 {
82 /* We can't store directly to *out as it may be unaligned. */
83 /* BTW Don't reduce the buffer size below 128 Bytes!
84 * PadLock microcode needs it that big. */
85 char buf[128 + PADLOCK_ALIGNMENT - STACK_ALIGN] __attribute__
86 ((aligned(STACK_ALIGN)));
87 char *result = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT);
88 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
89 struct sha1_state state;
90 unsigned int space;
91 unsigned int leftover;
92 int err;
93
94 dctx->fallback.flags = desc->flags & CRYPTO_TFM_REQ_MAY_SLEEP;
95 err = crypto_shash_export(&dctx->fallback, &state);
96 if (err)
97 goto out;
98
99 if (state.count + count > ULONG_MAX)
100 return crypto_shash_finup(&dctx->fallback, in, count, out);
101
102 leftover = ((state.count - 1) & (SHA1_BLOCK_SIZE - 1)) + 1;
103 space = SHA1_BLOCK_SIZE - leftover;
104 if (space) {
105 if (count > space) {
106 err = crypto_shash_update(&dctx->fallback, in, space) ?:
107 crypto_shash_export(&dctx->fallback, &state);
108 if (err)
109 goto out;
110 count -= space;
111 in += space;
112 } else {
113 memcpy(state.buffer + leftover, in, count);
114 in = state.buffer;
115 count += leftover;
116 state.count &= ~(SHA1_BLOCK_SIZE - 1);
117 }
118 }
119
120 memcpy(result, &state.state, SHA1_DIGEST_SIZE);
121
122 asm volatile (".byte 0xf3,0x0f,0xa6,0xc8" /* rep xsha1 */
123 : \
124 : "c"((unsigned long)state.count + count), \
125 "a"((unsigned long)state.count), \
126 "S"(in), "D"(result));
127
128 padlock_output_block((uint32_t *)result, (uint32_t *)out, 5);
129
130 out:
131 return err;
132 }
133
134 static int padlock_sha1_final(struct shash_desc *desc, u8 *out)
135 {
136 u8 buf[4];
137
138 return padlock_sha1_finup(desc, buf, 0, out);
139 }
140
141 static int padlock_sha256_finup(struct shash_desc *desc, const u8 *in,
142 unsigned int count, u8 *out)
143 {
144 /* We can't store directly to *out as it may be unaligned. */
145 /* BTW Don't reduce the buffer size below 128 Bytes!
146 * PadLock microcode needs it that big. */
147 char buf[128 + PADLOCK_ALIGNMENT - STACK_ALIGN] __attribute__
148 ((aligned(STACK_ALIGN)));
149 char *result = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT);
150 struct padlock_sha_desc *dctx = shash_desc_ctx(desc);
151 struct sha256_state state;
152 unsigned int space;
153 unsigned int leftover;
154 int err;
155
156 dctx->fallback.flags = desc->flags & CRYPTO_TFM_REQ_MAY_SLEEP;
157 err = crypto_shash_export(&dctx->fallback, &state);
158 if (err)
159 goto out;
160
161 if (state.count + count > ULONG_MAX)
162 return crypto_shash_finup(&dctx->fallback, in, count, out);
163
164 leftover = ((state.count - 1) & (SHA256_BLOCK_SIZE - 1)) + 1;
165 space = SHA256_BLOCK_SIZE - leftover;
166 if (space) {
167 if (count > space) {
168 err = crypto_shash_update(&dctx->fallback, in, space) ?:
169 crypto_shash_export(&dctx->fallback, &state);
170 if (err)
171 goto out;
172 count -= space;
173 in += space;
174 } else {
175 memcpy(state.buf + leftover, in, count);
176 in = state.buf;
177 count += leftover;
178 state.count &= ~(SHA1_BLOCK_SIZE - 1);
179 }
180 }
181
182 memcpy(result, &state.state, SHA256_DIGEST_SIZE);
183
184 asm volatile (".byte 0xf3,0x0f,0xa6,0xd0" /* rep xsha256 */
185 : \
186 : "c"((unsigned long)state.count + count), \
187 "a"((unsigned long)state.count), \
188 "S"(in), "D"(result));
189
190 padlock_output_block((uint32_t *)result, (uint32_t *)out, 8);
191
192 out:
193 return err;
194 }
195
196 static int padlock_sha256_final(struct shash_desc *desc, u8 *out)
197 {
198 u8 buf[4];
199
200 return padlock_sha256_finup(desc, buf, 0, out);
201 }
202
203 static int padlock_cra_init(struct crypto_tfm *tfm)
204 {
205 struct crypto_shash *hash = __crypto_shash_cast(tfm);
206 const char *fallback_driver_name = crypto_tfm_alg_name(tfm);
207 struct padlock_sha_ctx *ctx = crypto_tfm_ctx(tfm);
208 struct crypto_shash *fallback_tfm;
209 int err = -ENOMEM;
210
211 /* Allocate a fallback and abort if it failed. */
212 fallback_tfm = crypto_alloc_shash(fallback_driver_name, 0,
213 CRYPTO_ALG_NEED_FALLBACK);
214 if (IS_ERR(fallback_tfm)) {
215 printk(KERN_WARNING PFX "Fallback driver '%s' could not be loaded!\n",
216 fallback_driver_name);
217 err = PTR_ERR(fallback_tfm);
218 goto out;
219 }
220
221 ctx->fallback = fallback_tfm;
222 hash->descsize += crypto_shash_descsize(fallback_tfm);
223 return 0;
224
225 out:
226 return err;
227 }
228
229 static void padlock_cra_exit(struct crypto_tfm *tfm)
230 {
231 struct padlock_sha_ctx *ctx = crypto_tfm_ctx(tfm);
232
233 crypto_free_shash(ctx->fallback);
234 }
235
236 static struct shash_alg sha1_alg = {
237 .digestsize = SHA1_DIGEST_SIZE,
238 .init = padlock_sha_init,
239 .update = padlock_sha_update,
240 .finup = padlock_sha1_finup,
241 .final = padlock_sha1_final,
242 .export = padlock_sha_export,
243 .import = padlock_sha_import,
244 .descsize = sizeof(struct padlock_sha_desc),
245 .statesize = sizeof(struct sha1_state),
246 .base = {
247 .cra_name = "sha1",
248 .cra_driver_name = "sha1-padlock",
249 .cra_priority = PADLOCK_CRA_PRIORITY,
250 .cra_flags = CRYPTO_ALG_TYPE_SHASH |
251 CRYPTO_ALG_NEED_FALLBACK,
252 .cra_blocksize = SHA1_BLOCK_SIZE,
253 .cra_ctxsize = sizeof(struct padlock_sha_ctx),
254 .cra_module = THIS_MODULE,
255 .cra_init = padlock_cra_init,
256 .cra_exit = padlock_cra_exit,
257 }
258 };
259
260 static struct shash_alg sha256_alg = {
261 .digestsize = SHA256_DIGEST_SIZE,
262 .init = padlock_sha_init,
263 .update = padlock_sha_update,
264 .finup = padlock_sha256_finup,
265 .final = padlock_sha256_final,
266 .export = padlock_sha_export,
267 .import = padlock_sha_import,
268 .descsize = sizeof(struct padlock_sha_desc),
269 .statesize = sizeof(struct sha256_state),
270 .base = {
271 .cra_name = "sha256",
272 .cra_driver_name = "sha256-padlock",
273 .cra_priority = PADLOCK_CRA_PRIORITY,
274 .cra_flags = CRYPTO_ALG_TYPE_SHASH |
275 CRYPTO_ALG_NEED_FALLBACK,
276 .cra_blocksize = SHA256_BLOCK_SIZE,
277 .cra_ctxsize = sizeof(struct padlock_sha_ctx),
278 .cra_module = THIS_MODULE,
279 .cra_init = padlock_cra_init,
280 .cra_exit = padlock_cra_exit,
281 }
282 };
283
284 /* Add two shash_alg instance for hardware-implemented *
285 * multiple-parts hash supported by VIA Nano Processor.*/
286 static int padlock_sha1_init_nano(struct shash_desc *desc)
287 {
288 struct sha1_state *sctx = shash_desc_ctx(desc);
289
290 *sctx = (struct sha1_state){
291 .state = { SHA1_H0, SHA1_H1, SHA1_H2, SHA1_H3, SHA1_H4 },
292 };
293
294 return 0;
295 }
296
297 static int padlock_sha1_update_nano(struct shash_desc *desc,
298 const u8 *data, unsigned int len)
299 {
300 struct sha1_state *sctx = shash_desc_ctx(desc);
301 unsigned int partial, done;
302 const u8 *src;
303 /*The PHE require the out buffer must 128 bytes and 16-bytes aligned*/
304 u8 buf[128 + PADLOCK_ALIGNMENT - STACK_ALIGN] __attribute__
305 ((aligned(STACK_ALIGN)));
306 u8 *dst = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT);
307
308 partial = sctx->count & 0x3f;
309 sctx->count += len;
310 done = 0;
311 src = data;
312 memcpy(dst, (u8 *)(sctx->state), SHA1_DIGEST_SIZE);
313
314 if ((partial + len) >= SHA1_BLOCK_SIZE) {
315
316 /* Append the bytes in state's buffer to a block to handle */
317 if (partial) {
318 done = -partial;
319 memcpy(sctx->buffer + partial, data,
320 done + SHA1_BLOCK_SIZE);
321 src = sctx->buffer;
322 asm volatile (".byte 0xf3,0x0f,0xa6,0xc8"
323 : "+S"(src), "+D"(dst) \
324 : "a"((long)-1), "c"((unsigned long)1));
325 done += SHA1_BLOCK_SIZE;
326 src = data + done;
327 }
328
329 /* Process the left bytes from the input data */
330 if (len - done >= SHA1_BLOCK_SIZE) {
331 asm volatile (".byte 0xf3,0x0f,0xa6,0xc8"
332 : "+S"(src), "+D"(dst)
333 : "a"((long)-1),
334 "c"((unsigned long)((len - done) / SHA1_BLOCK_SIZE)));
335 done += ((len - done) - (len - done) % SHA1_BLOCK_SIZE);
336 src = data + done;
337 }
338 partial = 0;
339 }
340 memcpy((u8 *)(sctx->state), dst, SHA1_DIGEST_SIZE);
341 memcpy(sctx->buffer + partial, src, len - done);
342
343 return 0;
344 }
345
346 static int padlock_sha1_final_nano(struct shash_desc *desc, u8 *out)
347 {
348 struct sha1_state *state = (struct sha1_state *)shash_desc_ctx(desc);
349 unsigned int partial, padlen;
350 __be64 bits;
351 static const u8 padding[64] = { 0x80, };
352
353 bits = cpu_to_be64(state->count << 3);
354
355 /* Pad out to 56 mod 64 */
356 partial = state->count & 0x3f;
357 padlen = (partial < 56) ? (56 - partial) : ((64+56) - partial);
358 padlock_sha1_update_nano(desc, padding, padlen);
359
360 /* Append length field bytes */
361 padlock_sha1_update_nano(desc, (const u8 *)&bits, sizeof(bits));
362
363 /* Swap to output */
364 padlock_output_block((uint32_t *)(state->state), (uint32_t *)out, 5);
365
366 return 0;
367 }
368
369 static int padlock_sha256_init_nano(struct shash_desc *desc)
370 {
371 struct sha256_state *sctx = shash_desc_ctx(desc);
372
373 *sctx = (struct sha256_state){
374 .state = { SHA256_H0, SHA256_H1, SHA256_H2, SHA256_H3, \
375 SHA256_H4, SHA256_H5, SHA256_H6, SHA256_H7},
376 };
377
378 return 0;
379 }
380
381 static int padlock_sha256_update_nano(struct shash_desc *desc, const u8 *data,
382 unsigned int len)
383 {
384 struct sha256_state *sctx = shash_desc_ctx(desc);
385 unsigned int partial, done;
386 const u8 *src;
387 /*The PHE require the out buffer must 128 bytes and 16-bytes aligned*/
388 u8 buf[128 + PADLOCK_ALIGNMENT - STACK_ALIGN] __attribute__
389 ((aligned(STACK_ALIGN)));
390 u8 *dst = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT);
391
392 partial = sctx->count & 0x3f;
393 sctx->count += len;
394 done = 0;
395 src = data;
396 memcpy(dst, (u8 *)(sctx->state), SHA256_DIGEST_SIZE);
397
398 if ((partial + len) >= SHA256_BLOCK_SIZE) {
399
400 /* Append the bytes in state's buffer to a block to handle */
401 if (partial) {
402 done = -partial;
403 memcpy(sctx->buf + partial, data,
404 done + SHA256_BLOCK_SIZE);
405 src = sctx->buf;
406 asm volatile (".byte 0xf3,0x0f,0xa6,0xd0"
407 : "+S"(src), "+D"(dst)
408 : "a"((long)-1), "c"((unsigned long)1));
409 done += SHA256_BLOCK_SIZE;
410 src = data + done;
411 }
412
413 /* Process the left bytes from input data*/
414 if (len - done >= SHA256_BLOCK_SIZE) {
415 asm volatile (".byte 0xf3,0x0f,0xa6,0xd0"
416 : "+S"(src), "+D"(dst)
417 : "a"((long)-1),
418 "c"((unsigned long)((len - done) / 64)));
419 done += ((len - done) - (len - done) % 64);
420 src = data + done;
421 }
422 partial = 0;
423 }
424 memcpy((u8 *)(sctx->state), dst, SHA256_DIGEST_SIZE);
425 memcpy(sctx->buf + partial, src, len - done);
426
427 return 0;
428 }
429
430 static int padlock_sha256_final_nano(struct shash_desc *desc, u8 *out)
431 {
432 struct sha256_state *state =
433 (struct sha256_state *)shash_desc_ctx(desc);
434 unsigned int partial, padlen;
435 __be64 bits;
436 static const u8 padding[64] = { 0x80, };
437
438 bits = cpu_to_be64(state->count << 3);
439
440 /* Pad out to 56 mod 64 */
441 partial = state->count & 0x3f;
442 padlen = (partial < 56) ? (56 - partial) : ((64+56) - partial);
443 padlock_sha256_update_nano(desc, padding, padlen);
444
445 /* Append length field bytes */
446 padlock_sha256_update_nano(desc, (const u8 *)&bits, sizeof(bits));
447
448 /* Swap to output */
449 padlock_output_block((uint32_t *)(state->state), (uint32_t *)out, 8);
450
451 return 0;
452 }
453
454 static int padlock_sha_export_nano(struct shash_desc *desc,
455 void *out)
456 {
457 int statesize = crypto_shash_statesize(desc->tfm);
458 void *sctx = shash_desc_ctx(desc);
459
460 memcpy(out, sctx, statesize);
461 return 0;
462 }
463
464 static int padlock_sha_import_nano(struct shash_desc *desc,
465 const void *in)
466 {
467 int statesize = crypto_shash_statesize(desc->tfm);
468 void *sctx = shash_desc_ctx(desc);
469
470 memcpy(sctx, in, statesize);
471 return 0;
472 }
473
474 static struct shash_alg sha1_alg_nano = {
475 .digestsize = SHA1_DIGEST_SIZE,
476 .init = padlock_sha1_init_nano,
477 .update = padlock_sha1_update_nano,
478 .final = padlock_sha1_final_nano,
479 .export = padlock_sha_export_nano,
480 .import = padlock_sha_import_nano,
481 .descsize = sizeof(struct sha1_state),
482 .statesize = sizeof(struct sha1_state),
483 .base = {
484 .cra_name = "sha1",
485 .cra_driver_name = "sha1-padlock-nano",
486 .cra_priority = PADLOCK_CRA_PRIORITY,
487 .cra_flags = CRYPTO_ALG_TYPE_SHASH,
488 .cra_blocksize = SHA1_BLOCK_SIZE,
489 .cra_module = THIS_MODULE,
490 }
491 };
492
493 static struct shash_alg sha256_alg_nano = {
494 .digestsize = SHA256_DIGEST_SIZE,
495 .init = padlock_sha256_init_nano,
496 .update = padlock_sha256_update_nano,
497 .final = padlock_sha256_final_nano,
498 .export = padlock_sha_export_nano,
499 .import = padlock_sha_import_nano,
500 .descsize = sizeof(struct sha256_state),
501 .statesize = sizeof(struct sha256_state),
502 .base = {
503 .cra_name = "sha256",
504 .cra_driver_name = "sha256-padlock-nano",
505 .cra_priority = PADLOCK_CRA_PRIORITY,
506 .cra_flags = CRYPTO_ALG_TYPE_SHASH,
507 .cra_blocksize = SHA256_BLOCK_SIZE,
508 .cra_module = THIS_MODULE,
509 }
510 };
511
512 static const struct x86_cpu_id padlock_sha_ids[] = {
513 X86_FEATURE_MATCH(X86_FEATURE_PHE),
514 {}
515 };
516 MODULE_DEVICE_TABLE(x86cpu, padlock_sha_ids);
517
518 static int __init padlock_init(void)
519 {
520 int rc = -ENODEV;
521 struct cpuinfo_x86 *c = &cpu_data(0);
522 struct shash_alg *sha1;
523 struct shash_alg *sha256;
524
525 if (!x86_match_cpu(padlock_sha_ids) || !boot_cpu_has(X86_FEATURE_PHE_EN))
526 return -ENODEV;
527
528 /* Register the newly added algorithm module if on *
529 * VIA Nano processor, or else just do as before */
530 if (c->x86_model < 0x0f) {
531 sha1 = &sha1_alg;
532 sha256 = &sha256_alg;
533 } else {
534 sha1 = &sha1_alg_nano;
535 sha256 = &sha256_alg_nano;
536 }
537
538 rc = crypto_register_shash(sha1);
539 if (rc)
540 goto out;
541
542 rc = crypto_register_shash(sha256);
543 if (rc)
544 goto out_unreg1;
545
546 printk(KERN_NOTICE PFX "Using VIA PadLock ACE for SHA1/SHA256 algorithms.\n");
547
548 return 0;
549
550 out_unreg1:
551 crypto_unregister_shash(sha1);
552
553 out:
554 printk(KERN_ERR PFX "VIA PadLock SHA1/SHA256 initialization failed.\n");
555 return rc;
556 }
557
558 static void __exit padlock_fini(void)
559 {
560 struct cpuinfo_x86 *c = &cpu_data(0);
561
562 if (c->x86_model >= 0x0f) {
563 crypto_unregister_shash(&sha1_alg_nano);
564 crypto_unregister_shash(&sha256_alg_nano);
565 } else {
566 crypto_unregister_shash(&sha1_alg);
567 crypto_unregister_shash(&sha256_alg);
568 }
569 }
570
571 module_init(padlock_init);
572 module_exit(padlock_fini);
573
574 MODULE_DESCRIPTION("VIA PadLock SHA1/SHA256 algorithms support.");
575 MODULE_LICENSE("GPL");
576 MODULE_AUTHOR("Michal Ludvig");
577
578 MODULE_ALIAS_CRYPTO("sha1-all");
579 MODULE_ALIAS_CRYPTO("sha256-all");
580 MODULE_ALIAS_CRYPTO("sha1-padlock");
581 MODULE_ALIAS_CRYPTO("sha256-padlock");
Linux with added FPC-III support