serpent_avx_glue.c 24 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949
  1. /*
  2. * Glue Code for AVX assembler versions of Serpent Cipher
  3. *
  4. * Copyright (C) 2012 Johannes Goetzfried
  5. * <Johannes.Goetzfried@informatik.stud.uni-erlangen.de>
  6. *
  7. * Glue code based on serpent_sse2_glue.c by:
  8. * Copyright (C) 2011 Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
  9. *
  10. * This program is free software; you can redistribute it and/or modify
  11. * it under the terms of the GNU General Public License as published by
  12. * the Free Software Foundation; either version 2 of the License, or
  13. * (at your option) any later version.
  14. *
  15. * This program is distributed in the hope that it will be useful,
  16. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  18. * GNU General Public License for more details.
  19. *
  20. * You should have received a copy of the GNU General Public License
  21. * along with this program; if not, write to the Free Software
  22. * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
  23. * USA
  24. *
  25. */
  26. #include <linux/module.h>
  27. #include <linux/hardirq.h>
  28. #include <linux/types.h>
  29. #include <linux/crypto.h>
  30. #include <linux/err.h>
  31. #include <crypto/algapi.h>
  32. #include <crypto/serpent.h>
  33. #include <crypto/cryptd.h>
  34. #include <crypto/b128ops.h>
  35. #include <crypto/ctr.h>
  36. #include <crypto/lrw.h>
  37. #include <crypto/xts.h>
  38. #include <asm/i387.h>
  39. #include <asm/xcr.h>
  40. #include <asm/xsave.h>
  41. #include <asm/serpent.h>
  42. #include <crypto/scatterwalk.h>
  43. #include <linux/workqueue.h>
  44. #include <linux/spinlock.h>
  45. struct async_serpent_ctx {
  46. struct cryptd_ablkcipher *cryptd_tfm;
  47. };
  48. static inline bool serpent_fpu_begin(bool fpu_enabled, unsigned int nbytes)
  49. {
  50. if (fpu_enabled)
  51. return true;
  52. /* AVX is only used when chunk to be processed is large enough, so
  53. * do not enable FPU until it is necessary.
  54. */
  55. if (nbytes < SERPENT_BLOCK_SIZE * SERPENT_PARALLEL_BLOCKS)
  56. return false;
  57. kernel_fpu_begin();
  58. return true;
  59. }
  60. static inline void serpent_fpu_end(bool fpu_enabled)
  61. {
  62. if (fpu_enabled)
  63. kernel_fpu_end();
  64. }
  65. static int ecb_crypt(struct blkcipher_desc *desc, struct blkcipher_walk *walk,
  66. bool enc)
  67. {
  68. bool fpu_enabled = false;
  69. struct serpent_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  70. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  71. unsigned int nbytes;
  72. int err;
  73. err = blkcipher_walk_virt(desc, walk);
  74. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  75. while ((nbytes = walk->nbytes)) {
  76. u8 *wsrc = walk->src.virt.addr;
  77. u8 *wdst = walk->dst.virt.addr;
  78. fpu_enabled = serpent_fpu_begin(fpu_enabled, nbytes);
  79. /* Process multi-block batch */
  80. if (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS) {
  81. do {
  82. if (enc)
  83. serpent_enc_blk_xway(ctx, wdst, wsrc);
  84. else
  85. serpent_dec_blk_xway(ctx, wdst, wsrc);
  86. wsrc += bsize * SERPENT_PARALLEL_BLOCKS;
  87. wdst += bsize * SERPENT_PARALLEL_BLOCKS;
  88. nbytes -= bsize * SERPENT_PARALLEL_BLOCKS;
  89. } while (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS);
  90. if (nbytes < bsize)
  91. goto done;
  92. }
  93. /* Handle leftovers */
  94. do {
  95. if (enc)
  96. __serpent_encrypt(ctx, wdst, wsrc);
  97. else
  98. __serpent_decrypt(ctx, wdst, wsrc);
  99. wsrc += bsize;
  100. wdst += bsize;
  101. nbytes -= bsize;
  102. } while (nbytes >= bsize);
  103. done:
  104. err = blkcipher_walk_done(desc, walk, nbytes);
  105. }
  106. serpent_fpu_end(fpu_enabled);
  107. return err;
  108. }
  109. static int ecb_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  110. struct scatterlist *src, unsigned int nbytes)
  111. {
  112. struct blkcipher_walk walk;
  113. blkcipher_walk_init(&walk, dst, src, nbytes);
  114. return ecb_crypt(desc, &walk, true);
  115. }
  116. static int ecb_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  117. struct scatterlist *src, unsigned int nbytes)
  118. {
  119. struct blkcipher_walk walk;
  120. blkcipher_walk_init(&walk, dst, src, nbytes);
  121. return ecb_crypt(desc, &walk, false);
  122. }
  123. static unsigned int __cbc_encrypt(struct blkcipher_desc *desc,
  124. struct blkcipher_walk *walk)
  125. {
  126. struct serpent_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  127. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  128. unsigned int nbytes = walk->nbytes;
  129. u128 *src = (u128 *)walk->src.virt.addr;
  130. u128 *dst = (u128 *)walk->dst.virt.addr;
  131. u128 *iv = (u128 *)walk->iv;
  132. do {
  133. u128_xor(dst, src, iv);
  134. __serpent_encrypt(ctx, (u8 *)dst, (u8 *)dst);
  135. iv = dst;
  136. src += 1;
  137. dst += 1;
  138. nbytes -= bsize;
  139. } while (nbytes >= bsize);
  140. u128_xor((u128 *)walk->iv, (u128 *)walk->iv, iv);
  141. return nbytes;
  142. }
  143. static int cbc_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  144. struct scatterlist *src, unsigned int nbytes)
  145. {
  146. struct blkcipher_walk walk;
  147. int err;
  148. blkcipher_walk_init(&walk, dst, src, nbytes);
  149. err = blkcipher_walk_virt(desc, &walk);
  150. while ((nbytes = walk.nbytes)) {
  151. nbytes = __cbc_encrypt(desc, &walk);
  152. err = blkcipher_walk_done(desc, &walk, nbytes);
  153. }
  154. return err;
  155. }
  156. static unsigned int __cbc_decrypt(struct blkcipher_desc *desc,
  157. struct blkcipher_walk *walk)
  158. {
  159. struct serpent_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  160. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  161. unsigned int nbytes = walk->nbytes;
  162. u128 *src = (u128 *)walk->src.virt.addr;
  163. u128 *dst = (u128 *)walk->dst.virt.addr;
  164. u128 ivs[SERPENT_PARALLEL_BLOCKS - 1];
  165. u128 last_iv;
  166. int i;
  167. /* Start of the last block. */
  168. src += nbytes / bsize - 1;
  169. dst += nbytes / bsize - 1;
  170. last_iv = *src;
  171. /* Process multi-block batch */
  172. if (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS) {
  173. do {
  174. nbytes -= bsize * (SERPENT_PARALLEL_BLOCKS - 1);
  175. src -= SERPENT_PARALLEL_BLOCKS - 1;
  176. dst -= SERPENT_PARALLEL_BLOCKS - 1;
  177. for (i = 0; i < SERPENT_PARALLEL_BLOCKS - 1; i++)
  178. ivs[i] = src[i];
  179. serpent_dec_blk_xway(ctx, (u8 *)dst, (u8 *)src);
  180. for (i = 0; i < SERPENT_PARALLEL_BLOCKS - 1; i++)
  181. u128_xor(dst + (i + 1), dst + (i + 1), ivs + i);
  182. nbytes -= bsize;
  183. if (nbytes < bsize)
  184. goto done;
  185. u128_xor(dst, dst, src - 1);
  186. src -= 1;
  187. dst -= 1;
  188. } while (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS);
  189. if (nbytes < bsize)
  190. goto done;
  191. }
  192. /* Handle leftovers */
  193. for (;;) {
  194. __serpent_decrypt(ctx, (u8 *)dst, (u8 *)src);
  195. nbytes -= bsize;
  196. if (nbytes < bsize)
  197. break;
  198. u128_xor(dst, dst, src - 1);
  199. src -= 1;
  200. dst -= 1;
  201. }
  202. done:
  203. u128_xor(dst, dst, (u128 *)walk->iv);
  204. *(u128 *)walk->iv = last_iv;
  205. return nbytes;
  206. }
  207. static int cbc_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  208. struct scatterlist *src, unsigned int nbytes)
  209. {
  210. bool fpu_enabled = false;
  211. struct blkcipher_walk walk;
  212. int err;
  213. blkcipher_walk_init(&walk, dst, src, nbytes);
  214. err = blkcipher_walk_virt(desc, &walk);
  215. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  216. while ((nbytes = walk.nbytes)) {
  217. fpu_enabled = serpent_fpu_begin(fpu_enabled, nbytes);
  218. nbytes = __cbc_decrypt(desc, &walk);
  219. err = blkcipher_walk_done(desc, &walk, nbytes);
  220. }
  221. serpent_fpu_end(fpu_enabled);
  222. return err;
  223. }
  224. static inline void u128_to_be128(be128 *dst, const u128 *src)
  225. {
  226. dst->a = cpu_to_be64(src->a);
  227. dst->b = cpu_to_be64(src->b);
  228. }
  229. static inline void be128_to_u128(u128 *dst, const be128 *src)
  230. {
  231. dst->a = be64_to_cpu(src->a);
  232. dst->b = be64_to_cpu(src->b);
  233. }
  234. static inline void u128_inc(u128 *i)
  235. {
  236. i->b++;
  237. if (!i->b)
  238. i->a++;
  239. }
  240. static void ctr_crypt_final(struct blkcipher_desc *desc,
  241. struct blkcipher_walk *walk)
  242. {
  243. struct serpent_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  244. u8 *ctrblk = walk->iv;
  245. u8 keystream[SERPENT_BLOCK_SIZE];
  246. u8 *src = walk->src.virt.addr;
  247. u8 *dst = walk->dst.virt.addr;
  248. unsigned int nbytes = walk->nbytes;
  249. __serpent_encrypt(ctx, keystream, ctrblk);
  250. crypto_xor(keystream, src, nbytes);
  251. memcpy(dst, keystream, nbytes);
  252. crypto_inc(ctrblk, SERPENT_BLOCK_SIZE);
  253. }
  254. static unsigned int __ctr_crypt(struct blkcipher_desc *desc,
  255. struct blkcipher_walk *walk)
  256. {
  257. struct serpent_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  258. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  259. unsigned int nbytes = walk->nbytes;
  260. u128 *src = (u128 *)walk->src.virt.addr;
  261. u128 *dst = (u128 *)walk->dst.virt.addr;
  262. u128 ctrblk;
  263. be128 ctrblocks[SERPENT_PARALLEL_BLOCKS];
  264. int i;
  265. be128_to_u128(&ctrblk, (be128 *)walk->iv);
  266. /* Process multi-block batch */
  267. if (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS) {
  268. do {
  269. /* create ctrblks for parallel encrypt */
  270. for (i = 0; i < SERPENT_PARALLEL_BLOCKS; i++) {
  271. if (dst != src)
  272. dst[i] = src[i];
  273. u128_to_be128(&ctrblocks[i], &ctrblk);
  274. u128_inc(&ctrblk);
  275. }
  276. serpent_enc_blk_xway_xor(ctx, (u8 *)dst,
  277. (u8 *)ctrblocks);
  278. src += SERPENT_PARALLEL_BLOCKS;
  279. dst += SERPENT_PARALLEL_BLOCKS;
  280. nbytes -= bsize * SERPENT_PARALLEL_BLOCKS;
  281. } while (nbytes >= bsize * SERPENT_PARALLEL_BLOCKS);
  282. if (nbytes < bsize)
  283. goto done;
  284. }
  285. /* Handle leftovers */
  286. do {
  287. if (dst != src)
  288. *dst = *src;
  289. u128_to_be128(&ctrblocks[0], &ctrblk);
  290. u128_inc(&ctrblk);
  291. __serpent_encrypt(ctx, (u8 *)ctrblocks, (u8 *)ctrblocks);
  292. u128_xor(dst, dst, (u128 *)ctrblocks);
  293. src += 1;
  294. dst += 1;
  295. nbytes -= bsize;
  296. } while (nbytes >= bsize);
  297. done:
  298. u128_to_be128((be128 *)walk->iv, &ctrblk);
  299. return nbytes;
  300. }
  301. static int ctr_crypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  302. struct scatterlist *src, unsigned int nbytes)
  303. {
  304. bool fpu_enabled = false;
  305. struct blkcipher_walk walk;
  306. int err;
  307. blkcipher_walk_init(&walk, dst, src, nbytes);
  308. err = blkcipher_walk_virt_block(desc, &walk, SERPENT_BLOCK_SIZE);
  309. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  310. while ((nbytes = walk.nbytes) >= SERPENT_BLOCK_SIZE) {
  311. fpu_enabled = serpent_fpu_begin(fpu_enabled, nbytes);
  312. nbytes = __ctr_crypt(desc, &walk);
  313. err = blkcipher_walk_done(desc, &walk, nbytes);
  314. }
  315. serpent_fpu_end(fpu_enabled);
  316. if (walk.nbytes) {
  317. ctr_crypt_final(desc, &walk);
  318. err = blkcipher_walk_done(desc, &walk, 0);
  319. }
  320. return err;
  321. }
  322. struct crypt_priv {
  323. struct serpent_ctx *ctx;
  324. bool fpu_enabled;
  325. };
  326. static void encrypt_callback(void *priv, u8 *srcdst, unsigned int nbytes)
  327. {
  328. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  329. struct crypt_priv *ctx = priv;
  330. int i;
  331. ctx->fpu_enabled = serpent_fpu_begin(ctx->fpu_enabled, nbytes);
  332. if (nbytes == bsize * SERPENT_PARALLEL_BLOCKS) {
  333. serpent_enc_blk_xway(ctx->ctx, srcdst, srcdst);
  334. return;
  335. }
  336. for (i = 0; i < nbytes / bsize; i++, srcdst += bsize)
  337. __serpent_encrypt(ctx->ctx, srcdst, srcdst);
  338. }
  339. static void decrypt_callback(void *priv, u8 *srcdst, unsigned int nbytes)
  340. {
  341. const unsigned int bsize = SERPENT_BLOCK_SIZE;
  342. struct crypt_priv *ctx = priv;
  343. int i;
  344. ctx->fpu_enabled = serpent_fpu_begin(ctx->fpu_enabled, nbytes);
  345. if (nbytes == bsize * SERPENT_PARALLEL_BLOCKS) {
  346. serpent_dec_blk_xway(ctx->ctx, srcdst, srcdst);
  347. return;
  348. }
  349. for (i = 0; i < nbytes / bsize; i++, srcdst += bsize)
  350. __serpent_decrypt(ctx->ctx, srcdst, srcdst);
  351. }
  352. struct serpent_lrw_ctx {
  353. struct lrw_table_ctx lrw_table;
  354. struct serpent_ctx serpent_ctx;
  355. };
  356. static int lrw_serpent_setkey(struct crypto_tfm *tfm, const u8 *key,
  357. unsigned int keylen)
  358. {
  359. struct serpent_lrw_ctx *ctx = crypto_tfm_ctx(tfm);
  360. int err;
  361. err = __serpent_setkey(&ctx->serpent_ctx, key, keylen -
  362. SERPENT_BLOCK_SIZE);
  363. if (err)
  364. return err;
  365. return lrw_init_table(&ctx->lrw_table, key + keylen -
  366. SERPENT_BLOCK_SIZE);
  367. }
  368. static int lrw_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  369. struct scatterlist *src, unsigned int nbytes)
  370. {
  371. struct serpent_lrw_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  372. be128 buf[SERPENT_PARALLEL_BLOCKS];
  373. struct crypt_priv crypt_ctx = {
  374. .ctx = &ctx->serpent_ctx,
  375. .fpu_enabled = false,
  376. };
  377. struct lrw_crypt_req req = {
  378. .tbuf = buf,
  379. .tbuflen = sizeof(buf),
  380. .table_ctx = &ctx->lrw_table,
  381. .crypt_ctx = &crypt_ctx,
  382. .crypt_fn = encrypt_callback,
  383. };
  384. int ret;
  385. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  386. ret = lrw_crypt(desc, dst, src, nbytes, &req);
  387. serpent_fpu_end(crypt_ctx.fpu_enabled);
  388. return ret;
  389. }
  390. static int lrw_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  391. struct scatterlist *src, unsigned int nbytes)
  392. {
  393. struct serpent_lrw_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  394. be128 buf[SERPENT_PARALLEL_BLOCKS];
  395. struct crypt_priv crypt_ctx = {
  396. .ctx = &ctx->serpent_ctx,
  397. .fpu_enabled = false,
  398. };
  399. struct lrw_crypt_req req = {
  400. .tbuf = buf,
  401. .tbuflen = sizeof(buf),
  402. .table_ctx = &ctx->lrw_table,
  403. .crypt_ctx = &crypt_ctx,
  404. .crypt_fn = decrypt_callback,
  405. };
  406. int ret;
  407. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  408. ret = lrw_crypt(desc, dst, src, nbytes, &req);
  409. serpent_fpu_end(crypt_ctx.fpu_enabled);
  410. return ret;
  411. }
  412. static void lrw_exit_tfm(struct crypto_tfm *tfm)
  413. {
  414. struct serpent_lrw_ctx *ctx = crypto_tfm_ctx(tfm);
  415. lrw_free_table(&ctx->lrw_table);
  416. }
  417. struct serpent_xts_ctx {
  418. struct serpent_ctx tweak_ctx;
  419. struct serpent_ctx crypt_ctx;
  420. };
  421. static int xts_serpent_setkey(struct crypto_tfm *tfm, const u8 *key,
  422. unsigned int keylen)
  423. {
  424. struct serpent_xts_ctx *ctx = crypto_tfm_ctx(tfm);
  425. u32 *flags = &tfm->crt_flags;
  426. int err;
  427. /* key consists of keys of equal size concatenated, therefore
  428. * the length must be even
  429. */
  430. if (keylen % 2) {
  431. *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
  432. return -EINVAL;
  433. }
  434. /* first half of xts-key is for crypt */
  435. err = __serpent_setkey(&ctx->crypt_ctx, key, keylen / 2);
  436. if (err)
  437. return err;
  438. /* second half of xts-key is for tweak */
  439. return __serpent_setkey(&ctx->tweak_ctx, key + keylen / 2, keylen / 2);
  440. }
  441. static int xts_encrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  442. struct scatterlist *src, unsigned int nbytes)
  443. {
  444. struct serpent_xts_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  445. be128 buf[SERPENT_PARALLEL_BLOCKS];
  446. struct crypt_priv crypt_ctx = {
  447. .ctx = &ctx->crypt_ctx,
  448. .fpu_enabled = false,
  449. };
  450. struct xts_crypt_req req = {
  451. .tbuf = buf,
  452. .tbuflen = sizeof(buf),
  453. .tweak_ctx = &ctx->tweak_ctx,
  454. .tweak_fn = XTS_TWEAK_CAST(__serpent_encrypt),
  455. .crypt_ctx = &crypt_ctx,
  456. .crypt_fn = encrypt_callback,
  457. };
  458. int ret;
  459. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  460. ret = xts_crypt(desc, dst, src, nbytes, &req);
  461. serpent_fpu_end(crypt_ctx.fpu_enabled);
  462. return ret;
  463. }
  464. static int xts_decrypt(struct blkcipher_desc *desc, struct scatterlist *dst,
  465. struct scatterlist *src, unsigned int nbytes)
  466. {
  467. struct serpent_xts_ctx *ctx = crypto_blkcipher_ctx(desc->tfm);
  468. be128 buf[SERPENT_PARALLEL_BLOCKS];
  469. struct crypt_priv crypt_ctx = {
  470. .ctx = &ctx->crypt_ctx,
  471. .fpu_enabled = false,
  472. };
  473. struct xts_crypt_req req = {
  474. .tbuf = buf,
  475. .tbuflen = sizeof(buf),
  476. .tweak_ctx = &ctx->tweak_ctx,
  477. .tweak_fn = XTS_TWEAK_CAST(__serpent_encrypt),
  478. .crypt_ctx = &crypt_ctx,
  479. .crypt_fn = decrypt_callback,
  480. };
  481. int ret;
  482. desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
  483. ret = xts_crypt(desc, dst, src, nbytes, &req);
  484. serpent_fpu_end(crypt_ctx.fpu_enabled);
  485. return ret;
  486. }
  487. static int ablk_set_key(struct crypto_ablkcipher *tfm, const u8 *key,
  488. unsigned int key_len)
  489. {
  490. struct async_serpent_ctx *ctx = crypto_ablkcipher_ctx(tfm);
  491. struct crypto_ablkcipher *child = &ctx->cryptd_tfm->base;
  492. int err;
  493. crypto_ablkcipher_clear_flags(child, CRYPTO_TFM_REQ_MASK);
  494. crypto_ablkcipher_set_flags(child, crypto_ablkcipher_get_flags(tfm)
  495. & CRYPTO_TFM_REQ_MASK);
  496. err = crypto_ablkcipher_setkey(child, key, key_len);
  497. crypto_ablkcipher_set_flags(tfm, crypto_ablkcipher_get_flags(child)
  498. & CRYPTO_TFM_RES_MASK);
  499. return err;
  500. }
  501. static int __ablk_encrypt(struct ablkcipher_request *req)
  502. {
  503. struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(req);
  504. struct async_serpent_ctx *ctx = crypto_ablkcipher_ctx(tfm);
  505. struct blkcipher_desc desc;
  506. desc.tfm = cryptd_ablkcipher_child(ctx->cryptd_tfm);
  507. desc.info = req->info;
  508. desc.flags = 0;
  509. return crypto_blkcipher_crt(desc.tfm)->encrypt(
  510. &desc, req->dst, req->src, req->nbytes);
  511. }
  512. static int ablk_encrypt(struct ablkcipher_request *req)
  513. {
  514. struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(req);
  515. struct async_serpent_ctx *ctx = crypto_ablkcipher_ctx(tfm);
  516. if (!irq_fpu_usable()) {
  517. struct ablkcipher_request *cryptd_req =
  518. ablkcipher_request_ctx(req);
  519. memcpy(cryptd_req, req, sizeof(*req));
  520. ablkcipher_request_set_tfm(cryptd_req, &ctx->cryptd_tfm->base);
  521. return crypto_ablkcipher_encrypt(cryptd_req);
  522. } else {
  523. return __ablk_encrypt(req);
  524. }
  525. }
  526. static int ablk_decrypt(struct ablkcipher_request *req)
  527. {
  528. struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(req);
  529. struct async_serpent_ctx *ctx = crypto_ablkcipher_ctx(tfm);
  530. if (!irq_fpu_usable()) {
  531. struct ablkcipher_request *cryptd_req =
  532. ablkcipher_request_ctx(req);
  533. memcpy(cryptd_req, req, sizeof(*req));
  534. ablkcipher_request_set_tfm(cryptd_req, &ctx->cryptd_tfm->base);
  535. return crypto_ablkcipher_decrypt(cryptd_req);
  536. } else {
  537. struct blkcipher_desc desc;
  538. desc.tfm = cryptd_ablkcipher_child(ctx->cryptd_tfm);
  539. desc.info = req->info;
  540. desc.flags = 0;
  541. return crypto_blkcipher_crt(desc.tfm)->decrypt(
  542. &desc, req->dst, req->src, req->nbytes);
  543. }
  544. }
  545. static void ablk_exit(struct crypto_tfm *tfm)
  546. {
  547. struct async_serpent_ctx *ctx = crypto_tfm_ctx(tfm);
  548. cryptd_free_ablkcipher(ctx->cryptd_tfm);
  549. }
  550. static int ablk_init(struct crypto_tfm *tfm)
  551. {
  552. struct async_serpent_ctx *ctx = crypto_tfm_ctx(tfm);
  553. struct cryptd_ablkcipher *cryptd_tfm;
  554. char drv_name[CRYPTO_MAX_ALG_NAME];
  555. snprintf(drv_name, sizeof(drv_name), "__driver-%s",
  556. crypto_tfm_alg_driver_name(tfm));
  557. cryptd_tfm = cryptd_alloc_ablkcipher(drv_name, 0, 0);
  558. if (IS_ERR(cryptd_tfm))
  559. return PTR_ERR(cryptd_tfm);
  560. ctx->cryptd_tfm = cryptd_tfm;
  561. tfm->crt_ablkcipher.reqsize = sizeof(struct ablkcipher_request) +
  562. crypto_ablkcipher_reqsize(&cryptd_tfm->base);
  563. return 0;
  564. }
  565. static struct crypto_alg serpent_algs[10] = { {
  566. .cra_name = "__ecb-serpent-avx",
  567. .cra_driver_name = "__driver-ecb-serpent-avx",
  568. .cra_priority = 0,
  569. .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER,
  570. .cra_blocksize = SERPENT_BLOCK_SIZE,
  571. .cra_ctxsize = sizeof(struct serpent_ctx),
  572. .cra_alignmask = 0,
  573. .cra_type = &crypto_blkcipher_type,
  574. .cra_module = THIS_MODULE,
  575. .cra_list = LIST_HEAD_INIT(serpent_algs[0].cra_list),
  576. .cra_u = {
  577. .blkcipher = {
  578. .min_keysize = SERPENT_MIN_KEY_SIZE,
  579. .max_keysize = SERPENT_MAX_KEY_SIZE,
  580. .setkey = serpent_setkey,
  581. .encrypt = ecb_encrypt,
  582. .decrypt = ecb_decrypt,
  583. },
  584. },
  585. }, {
  586. .cra_name = "__cbc-serpent-avx",
  587. .cra_driver_name = "__driver-cbc-serpent-avx",
  588. .cra_priority = 0,
  589. .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER,
  590. .cra_blocksize = SERPENT_BLOCK_SIZE,
  591. .cra_ctxsize = sizeof(struct serpent_ctx),
  592. .cra_alignmask = 0,
  593. .cra_type = &crypto_blkcipher_type,
  594. .cra_module = THIS_MODULE,
  595. .cra_list = LIST_HEAD_INIT(serpent_algs[1].cra_list),
  596. .cra_u = {
  597. .blkcipher = {
  598. .min_keysize = SERPENT_MIN_KEY_SIZE,
  599. .max_keysize = SERPENT_MAX_KEY_SIZE,
  600. .setkey = serpent_setkey,
  601. .encrypt = cbc_encrypt,
  602. .decrypt = cbc_decrypt,
  603. },
  604. },
  605. }, {
  606. .cra_name = "__ctr-serpent-avx",
  607. .cra_driver_name = "__driver-ctr-serpent-avx",
  608. .cra_priority = 0,
  609. .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER,
  610. .cra_blocksize = 1,
  611. .cra_ctxsize = sizeof(struct serpent_ctx),
  612. .cra_alignmask = 0,
  613. .cra_type = &crypto_blkcipher_type,
  614. .cra_module = THIS_MODULE,
  615. .cra_list = LIST_HEAD_INIT(serpent_algs[2].cra_list),
  616. .cra_u = {
  617. .blkcipher = {
  618. .min_keysize = SERPENT_MIN_KEY_SIZE,
  619. .max_keysize = SERPENT_MAX_KEY_SIZE,
  620. .ivsize = SERPENT_BLOCK_SIZE,
  621. .setkey = serpent_setkey,
  622. .encrypt = ctr_crypt,
  623. .decrypt = ctr_crypt,
  624. },
  625. },
  626. }, {
  627. .cra_name = "__lrw-serpent-avx",
  628. .cra_driver_name = "__driver-lrw-serpent-avx",
  629. .cra_priority = 0,
  630. .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER,
  631. .cra_blocksize = SERPENT_BLOCK_SIZE,
  632. .cra_ctxsize = sizeof(struct serpent_lrw_ctx),
  633. .cra_alignmask = 0,
  634. .cra_type = &crypto_blkcipher_type,
  635. .cra_module = THIS_MODULE,
  636. .cra_list = LIST_HEAD_INIT(serpent_algs[3].cra_list),
  637. .cra_exit = lrw_exit_tfm,
  638. .cra_u = {
  639. .blkcipher = {
  640. .min_keysize = SERPENT_MIN_KEY_SIZE +
  641. SERPENT_BLOCK_SIZE,
  642. .max_keysize = SERPENT_MAX_KEY_SIZE +
  643. SERPENT_BLOCK_SIZE,
  644. .ivsize = SERPENT_BLOCK_SIZE,
  645. .setkey = lrw_serpent_setkey,
  646. .encrypt = lrw_encrypt,
  647. .decrypt = lrw_decrypt,
  648. },
  649. },
  650. }, {
  651. .cra_name = "__xts-serpent-avx",
  652. .cra_driver_name = "__driver-xts-serpent-avx",
  653. .cra_priority = 0,
  654. .cra_flags = CRYPTO_ALG_TYPE_BLKCIPHER,
  655. .cra_blocksize = SERPENT_BLOCK_SIZE,
  656. .cra_ctxsize = sizeof(struct serpent_xts_ctx),
  657. .cra_alignmask = 0,
  658. .cra_type = &crypto_blkcipher_type,
  659. .cra_module = THIS_MODULE,
  660. .cra_list = LIST_HEAD_INIT(serpent_algs[4].cra_list),
  661. .cra_u = {
  662. .blkcipher = {
  663. .min_keysize = SERPENT_MIN_KEY_SIZE * 2,
  664. .max_keysize = SERPENT_MAX_KEY_SIZE * 2,
  665. .ivsize = SERPENT_BLOCK_SIZE,
  666. .setkey = xts_serpent_setkey,
  667. .encrypt = xts_encrypt,
  668. .decrypt = xts_decrypt,
  669. },
  670. },
  671. }, {
  672. .cra_name = "ecb(serpent)",
  673. .cra_driver_name = "ecb-serpent-avx",
  674. .cra_priority = 500,
  675. .cra_flags = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
  676. .cra_blocksize = SERPENT_BLOCK_SIZE,
  677. .cra_ctxsize = sizeof(struct async_serpent_ctx),
  678. .cra_alignmask = 0,
  679. .cra_type = &crypto_ablkcipher_type,
  680. .cra_module = THIS_MODULE,
  681. .cra_list = LIST_HEAD_INIT(serpent_algs[5].cra_list),
  682. .cra_init = ablk_init,
  683. .cra_exit = ablk_exit,
  684. .cra_u = {
  685. .ablkcipher = {
  686. .min_keysize = SERPENT_MIN_KEY_SIZE,
  687. .max_keysize = SERPENT_MAX_KEY_SIZE,
  688. .setkey = ablk_set_key,
  689. .encrypt = ablk_encrypt,
  690. .decrypt = ablk_decrypt,
  691. },
  692. },
  693. }, {
  694. .cra_name = "cbc(serpent)",
  695. .cra_driver_name = "cbc-serpent-avx",
  696. .cra_priority = 500,
  697. .cra_flags = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
  698. .cra_blocksize = SERPENT_BLOCK_SIZE,
  699. .cra_ctxsize = sizeof(struct async_serpent_ctx),
  700. .cra_alignmask = 0,
  701. .cra_type = &crypto_ablkcipher_type,
  702. .cra_module = THIS_MODULE,
  703. .cra_list = LIST_HEAD_INIT(serpent_algs[6].cra_list),
  704. .cra_init = ablk_init,
  705. .cra_exit = ablk_exit,
  706. .cra_u = {
  707. .ablkcipher = {
  708. .min_keysize = SERPENT_MIN_KEY_SIZE,
  709. .max_keysize = SERPENT_MAX_KEY_SIZE,
  710. .ivsize = SERPENT_BLOCK_SIZE,
  711. .setkey = ablk_set_key,
  712. .encrypt = __ablk_encrypt,
  713. .decrypt = ablk_decrypt,
  714. },
  715. },
  716. }, {
  717. .cra_name = "ctr(serpent)",
  718. .cra_driver_name = "ctr-serpent-avx",
  719. .cra_priority = 500,
  720. .cra_flags = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
  721. .cra_blocksize = 1,
  722. .cra_ctxsize = sizeof(struct async_serpent_ctx),
  723. .cra_alignmask = 0,
  724. .cra_type = &crypto_ablkcipher_type,
  725. .cra_module = THIS_MODULE,
  726. .cra_list = LIST_HEAD_INIT(serpent_algs[7].cra_list),
  727. .cra_init = ablk_init,
  728. .cra_exit = ablk_exit,
  729. .cra_u = {
  730. .ablkcipher = {
  731. .min_keysize = SERPENT_MIN_KEY_SIZE,
  732. .max_keysize = SERPENT_MAX_KEY_SIZE,
  733. .ivsize = SERPENT_BLOCK_SIZE,
  734. .setkey = ablk_set_key,
  735. .encrypt = ablk_encrypt,
  736. .decrypt = ablk_encrypt,
  737. .geniv = "chainiv",
  738. },
  739. },
  740. }, {
  741. .cra_name = "lrw(serpent)",
  742. .cra_driver_name = "lrw-serpent-avx",
  743. .cra_priority = 500,
  744. .cra_flags = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
  745. .cra_blocksize = SERPENT_BLOCK_SIZE,
  746. .cra_ctxsize = sizeof(struct async_serpent_ctx),
  747. .cra_alignmask = 0,
  748. .cra_type = &crypto_ablkcipher_type,
  749. .cra_module = THIS_MODULE,
  750. .cra_list = LIST_HEAD_INIT(serpent_algs[8].cra_list),
  751. .cra_init = ablk_init,
  752. .cra_exit = ablk_exit,
  753. .cra_u = {
  754. .ablkcipher = {
  755. .min_keysize = SERPENT_MIN_KEY_SIZE +
  756. SERPENT_BLOCK_SIZE,
  757. .max_keysize = SERPENT_MAX_KEY_SIZE +
  758. SERPENT_BLOCK_SIZE,
  759. .ivsize = SERPENT_BLOCK_SIZE,
  760. .setkey = ablk_set_key,
  761. .encrypt = ablk_encrypt,
  762. .decrypt = ablk_decrypt,
  763. },
  764. },
  765. }, {
  766. .cra_name = "xts(serpent)",
  767. .cra_driver_name = "xts-serpent-avx",
  768. .cra_priority = 500,
  769. .cra_flags = CRYPTO_ALG_TYPE_ABLKCIPHER | CRYPTO_ALG_ASYNC,
  770. .cra_blocksize = SERPENT_BLOCK_SIZE,
  771. .cra_ctxsize = sizeof(struct async_serpent_ctx),
  772. .cra_alignmask = 0,
  773. .cra_type = &crypto_ablkcipher_type,
  774. .cra_module = THIS_MODULE,
  775. .cra_list = LIST_HEAD_INIT(serpent_algs[9].cra_list),
  776. .cra_init = ablk_init,
  777. .cra_exit = ablk_exit,
  778. .cra_u = {
  779. .ablkcipher = {
  780. .min_keysize = SERPENT_MIN_KEY_SIZE * 2,
  781. .max_keysize = SERPENT_MAX_KEY_SIZE * 2,
  782. .ivsize = SERPENT_BLOCK_SIZE,
  783. .setkey = ablk_set_key,
  784. .encrypt = ablk_encrypt,
  785. .decrypt = ablk_decrypt,
  786. },
  787. },
  788. } };
  789. static int __init serpent_init(void)
  790. {
  791. u64 xcr0;
  792. if (!cpu_has_avx || !cpu_has_osxsave) {
  793. printk(KERN_INFO "AVX instructions are not detected.\n");
  794. return -ENODEV;
  795. }
  796. xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
  797. if ((xcr0 & (XSTATE_SSE | XSTATE_YMM)) != (XSTATE_SSE | XSTATE_YMM)) {
  798. printk(KERN_INFO "AVX detected but unusable.\n");
  799. return -ENODEV;
  800. }
  801. return crypto_register_algs(serpent_algs, ARRAY_SIZE(serpent_algs));
  802. }
  803. static void __exit serpent_exit(void)
  804. {
  805. crypto_unregister_algs(serpent_algs, ARRAY_SIZE(serpent_algs));
  806. }
  807. module_init(serpent_init);
  808. module_exit(serpent_exit);
  809. MODULE_DESCRIPTION("Serpent Cipher Algorithm, AVX optimized");
  810. MODULE_LICENSE("GPL");
  811. MODULE_ALIAS("serpent");