1/* SPDX-License-Identifier: GPL-2.0-or-later */
  2/*
  3 * Glue Code for the AVX/AES-NI/GFNI assembler implementation of the ARIA Cipher
  4 *
  5 * Copyright (c) 2022 Taehee Yoo <ap420073@gmail.com>
  6 */
  7
  8#include <crypto/algapi.h>
  9#include <crypto/internal/simd.h>
 10#include <crypto/aria.h>
 11#include <linux/crypto.h>
 12#include <linux/err.h>
 13#include <linux/module.h>
 14#include <linux/types.h>
 15
 16#include "ecb_cbc_helpers.h"
 17#include "aria-avx.h"
 18
 19asmlinkage void aria_aesni_avx_encrypt_16way(const void *ctx, u8 *dst,
 20					     const u8 *src);
 21EXPORT_SYMBOL_GPL(aria_aesni_avx_encrypt_16way);
 22asmlinkage void aria_aesni_avx_decrypt_16way(const void *ctx, u8 *dst,
 23					     const u8 *src);
 24EXPORT_SYMBOL_GPL(aria_aesni_avx_decrypt_16way);
 25asmlinkage void aria_aesni_avx_ctr_crypt_16way(const void *ctx, u8 *dst,
 26					       const u8 *src,
 27					       u8 *keystream, u8 *iv);
 28EXPORT_SYMBOL_GPL(aria_aesni_avx_ctr_crypt_16way);
 29#ifdef CONFIG_AS_GFNI
 30asmlinkage void aria_aesni_avx_gfni_encrypt_16way(const void *ctx, u8 *dst,
 31						  const u8 *src);
 32EXPORT_SYMBOL_GPL(aria_aesni_avx_gfni_encrypt_16way);
 33asmlinkage void aria_aesni_avx_gfni_decrypt_16way(const void *ctx, u8 *dst,
 34						  const u8 *src);
 35EXPORT_SYMBOL_GPL(aria_aesni_avx_gfni_decrypt_16way);
 36asmlinkage void aria_aesni_avx_gfni_ctr_crypt_16way(const void *ctx, u8 *dst,
 37						    const u8 *src,
 38						    u8 *keystream, u8 *iv);
 39EXPORT_SYMBOL_GPL(aria_aesni_avx_gfni_ctr_crypt_16way);
 40#endif /* CONFIG_AS_GFNI */
 41
 42static struct aria_avx_ops aria_ops;
 43
 44struct aria_avx_request_ctx {
 45	u8 keystream[ARIA_AESNI_PARALLEL_BLOCK_SIZE];
 46};
 47
 48static int ecb_do_encrypt(struct skcipher_request *req, const u32 *rkey)
 49{
 50	ECB_WALK_START(req, ARIA_BLOCK_SIZE, ARIA_AESNI_PARALLEL_BLOCKS);
 51	ECB_BLOCK(ARIA_AESNI_PARALLEL_BLOCKS, aria_ops.aria_encrypt_16way);
 52	ECB_BLOCK(1, aria_encrypt);
 53	ECB_WALK_END();
 54}
 55
 56static int ecb_do_decrypt(struct skcipher_request *req, const u32 *rkey)
 57{
 58	ECB_WALK_START(req, ARIA_BLOCK_SIZE, ARIA_AESNI_PARALLEL_BLOCKS);
 59	ECB_BLOCK(ARIA_AESNI_PARALLEL_BLOCKS, aria_ops.aria_decrypt_16way);
 60	ECB_BLOCK(1, aria_decrypt);
 61	ECB_WALK_END();
 62}
 63
 64static int aria_avx_ecb_encrypt(struct skcipher_request *req)
 65{
 66	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 67	struct aria_ctx *ctx = crypto_skcipher_ctx(tfm);
 68
 69	return ecb_do_encrypt(req, ctx->enc_key[0]);
 70}
 71
 72static int aria_avx_ecb_decrypt(struct skcipher_request *req)
 73{
 74	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 75	struct aria_ctx *ctx = crypto_skcipher_ctx(tfm);
 76
 77	return ecb_do_decrypt(req, ctx->dec_key[0]);
 78}
 79
 80static int aria_avx_set_key(struct crypto_skcipher *tfm, const u8 *key,
 81			    unsigned int keylen)
 82{
 83	return aria_set_key(&tfm->base, key, keylen);
 84}
 85
 86static int aria_avx_ctr_encrypt(struct skcipher_request *req)
 87{
 88	struct aria_avx_request_ctx *req_ctx = skcipher_request_ctx(req);
 89	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 90	struct aria_ctx *ctx = crypto_skcipher_ctx(tfm);
 91	struct skcipher_walk walk;
 92	unsigned int nbytes;
 93	int err;
 94
 95	err = skcipher_walk_virt(&walk, req, false);
 96
 97	while ((nbytes = walk.nbytes) > 0) {
 98		const u8 *src = walk.src.virt.addr;
 99		u8 *dst = walk.dst.virt.addr;
100
101		while (nbytes >= ARIA_AESNI_PARALLEL_BLOCK_SIZE) {
102			kernel_fpu_begin();
103			aria_ops.aria_ctr_crypt_16way(ctx, dst, src,
104						      &req_ctx->keystream[0],
105						      walk.iv);
106			kernel_fpu_end();
107			dst += ARIA_AESNI_PARALLEL_BLOCK_SIZE;
108			src += ARIA_AESNI_PARALLEL_BLOCK_SIZE;
109			nbytes -= ARIA_AESNI_PARALLEL_BLOCK_SIZE;
110		}
111
112		while (nbytes >= ARIA_BLOCK_SIZE) {
113			memcpy(&req_ctx->keystream[0], walk.iv, ARIA_BLOCK_SIZE);
114			crypto_inc(walk.iv, ARIA_BLOCK_SIZE);
115
116			aria_encrypt(ctx, &req_ctx->keystream[0],
117				     &req_ctx->keystream[0]);
118
119			crypto_xor_cpy(dst, src, &req_ctx->keystream[0],
120				       ARIA_BLOCK_SIZE);
121			dst += ARIA_BLOCK_SIZE;
122			src += ARIA_BLOCK_SIZE;
123			nbytes -= ARIA_BLOCK_SIZE;
124		}
125
126		if (walk.nbytes == walk.total && nbytes > 0) {
127			memcpy(&req_ctx->keystream[0], walk.iv,
128			       ARIA_BLOCK_SIZE);
129			crypto_inc(walk.iv, ARIA_BLOCK_SIZE);
130
131			aria_encrypt(ctx, &req_ctx->keystream[0],
132				     &req_ctx->keystream[0]);
133
134			crypto_xor_cpy(dst, src, &req_ctx->keystream[0],
135				       nbytes);
136			dst += nbytes;
137			src += nbytes;
138			nbytes = 0;
139		}
140		err = skcipher_walk_done(&walk, nbytes);
141	}
142
143	return err;
144}
145
146static int aria_avx_init_tfm(struct crypto_skcipher *tfm)
147{
148	crypto_skcipher_set_reqsize(tfm, sizeof(struct aria_avx_request_ctx));
149
150	return 0;
151}
152
153static struct skcipher_alg aria_algs[] = {
154	{
155		.base.cra_name		= "__ecb(aria)",
156		.base.cra_driver_name	= "__ecb-aria-avx",
157		.base.cra_priority	= 400,
158		.base.cra_flags		= CRYPTO_ALG_INTERNAL,
159		.base.cra_blocksize	= ARIA_BLOCK_SIZE,
160		.base.cra_ctxsize	= sizeof(struct aria_ctx),
161		.base.cra_module	= THIS_MODULE,
162		.min_keysize		= ARIA_MIN_KEY_SIZE,
163		.max_keysize		= ARIA_MAX_KEY_SIZE,
164		.setkey			= aria_avx_set_key,
165		.encrypt		= aria_avx_ecb_encrypt,
166		.decrypt		= aria_avx_ecb_decrypt,
167	}, {
168		.base.cra_name		= "__ctr(aria)",
169		.base.cra_driver_name	= "__ctr-aria-avx",
170		.base.cra_priority	= 400,
171		.base.cra_flags		= CRYPTO_ALG_INTERNAL,
172		.base.cra_blocksize	= 1,
173		.base.cra_ctxsize	= sizeof(struct aria_ctx),
174		.base.cra_module	= THIS_MODULE,
175		.min_keysize		= ARIA_MIN_KEY_SIZE,
176		.max_keysize		= ARIA_MAX_KEY_SIZE,
177		.ivsize			= ARIA_BLOCK_SIZE,
178		.chunksize		= ARIA_BLOCK_SIZE,
179		.walksize		= 16 * ARIA_BLOCK_SIZE,
180		.setkey			= aria_avx_set_key,
181		.encrypt		= aria_avx_ctr_encrypt,
182		.decrypt		= aria_avx_ctr_encrypt,
183		.init			= aria_avx_init_tfm,
184	}
185};
186
187static struct simd_skcipher_alg *aria_simd_algs[ARRAY_SIZE(aria_algs)];
188
189static int __init aria_avx_init(void)
190{
191	const char *feature_name;
192
193	if (!boot_cpu_has(X86_FEATURE_AVX) ||
194	    !boot_cpu_has(X86_FEATURE_AES) ||
195	    !boot_cpu_has(X86_FEATURE_OSXSAVE)) {
196		pr_info("AVX or AES-NI instructions are not detected.\n");
197		return -ENODEV;
198	}
199
200	if (!cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM,
201				&feature_name)) {
202		pr_info("CPU feature '%s' is not supported.\n", feature_name);
203		return -ENODEV;
204	}
205
206	if (boot_cpu_has(X86_FEATURE_GFNI) && IS_ENABLED(CONFIG_AS_GFNI)) {
207		aria_ops.aria_encrypt_16way = aria_aesni_avx_gfni_encrypt_16way;
208		aria_ops.aria_decrypt_16way = aria_aesni_avx_gfni_decrypt_16way;
209		aria_ops.aria_ctr_crypt_16way = aria_aesni_avx_gfni_ctr_crypt_16way;
210	} else {
211		aria_ops.aria_encrypt_16way = aria_aesni_avx_encrypt_16way;
212		aria_ops.aria_decrypt_16way = aria_aesni_avx_decrypt_16way;
213		aria_ops.aria_ctr_crypt_16way = aria_aesni_avx_ctr_crypt_16way;
214	}
215
216	return simd_register_skciphers_compat(aria_algs,
217					      ARRAY_SIZE(aria_algs),
218					      aria_simd_algs);
219}
220
221static void __exit aria_avx_exit(void)
222{
223	simd_unregister_skciphers(aria_algs, ARRAY_SIZE(aria_algs),
224				  aria_simd_algs);
225}
226
227module_init(aria_avx_init);
228module_exit(aria_avx_exit);
229
230MODULE_LICENSE("GPL");
231MODULE_AUTHOR("Taehee Yoo <ap420073@gmail.com>");
232MODULE_DESCRIPTION("ARIA Cipher Algorithm, AVX/AES-NI/GFNI optimized");
233MODULE_ALIAS_CRYPTO("aria");
234MODULE_ALIAS_CRYPTO("aria-aesni-avx");