xref: /linux/arch/arm/crypto/ghash-ce-glue.c (revision 69bfec7548f4c1595bac0e3ddfc0458a5af31f4c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Accelerated GHASH implementation with ARMv8 vmull.p64 instructions.
4  *
5  * Copyright (C) 2015 - 2018 Linaro Ltd.
6  * Copyright (C) 2023 Google LLC.
7  */
8 
9 #include <asm/hwcap.h>
10 #include <asm/neon.h>
11 #include <asm/simd.h>
12 #include <asm/unaligned.h>
13 #include <crypto/aes.h>
14 #include <crypto/gcm.h>
15 #include <crypto/b128ops.h>
16 #include <crypto/cryptd.h>
17 #include <crypto/internal/aead.h>
18 #include <crypto/internal/hash.h>
19 #include <crypto/internal/simd.h>
20 #include <crypto/internal/skcipher.h>
21 #include <crypto/gf128mul.h>
22 #include <crypto/scatterwalk.h>
23 #include <linux/cpufeature.h>
24 #include <linux/crypto.h>
25 #include <linux/jump_label.h>
26 #include <linux/module.h>
27 
28 MODULE_DESCRIPTION("GHASH hash function using ARMv8 Crypto Extensions");
29 MODULE_AUTHOR("Ard Biesheuvel <ardb@kernel.org>");
30 MODULE_LICENSE("GPL");
31 MODULE_ALIAS_CRYPTO("ghash");
32 MODULE_ALIAS_CRYPTO("gcm(aes)");
33 MODULE_ALIAS_CRYPTO("rfc4106(gcm(aes))");
34 
35 #define GHASH_BLOCK_SIZE	16
36 #define GHASH_DIGEST_SIZE	16
37 
38 #define RFC4106_NONCE_SIZE	4
39 
40 struct ghash_key {
41 	be128	k;
42 	u64	h[][2];
43 };
44 
45 struct gcm_key {
46 	u64	h[4][2];
47 	u32	rk[AES_MAX_KEYLENGTH_U32];
48 	int	rounds;
49 	u8	nonce[];	// for RFC4106 nonce
50 };
51 
52 struct ghash_desc_ctx {
53 	u64 digest[GHASH_DIGEST_SIZE/sizeof(u64)];
54 	u8 buf[GHASH_BLOCK_SIZE];
55 	u32 count;
56 };
57 
58 struct ghash_async_ctx {
59 	struct cryptd_ahash *cryptd_tfm;
60 };
61 
62 asmlinkage void pmull_ghash_update_p64(int blocks, u64 dg[], const char *src,
63 				       u64 const h[][2], const char *head);
64 
65 asmlinkage void pmull_ghash_update_p8(int blocks, u64 dg[], const char *src,
66 				      u64 const h[][2], const char *head);
67 
68 static __ro_after_init DEFINE_STATIC_KEY_FALSE(use_p64);
69 
70 static int ghash_init(struct shash_desc *desc)
71 {
72 	struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
73 
74 	*ctx = (struct ghash_desc_ctx){};
75 	return 0;
76 }
77 
78 static void ghash_do_update(int blocks, u64 dg[], const char *src,
79 			    struct ghash_key *key, const char *head)
80 {
81 	if (likely(crypto_simd_usable())) {
82 		kernel_neon_begin();
83 		if (static_branch_likely(&use_p64))
84 			pmull_ghash_update_p64(blocks, dg, src, key->h, head);
85 		else
86 			pmull_ghash_update_p8(blocks, dg, src, key->h, head);
87 		kernel_neon_end();
88 	} else {
89 		be128 dst = { cpu_to_be64(dg[1]), cpu_to_be64(dg[0]) };
90 
91 		do {
92 			const u8 *in = src;
93 
94 			if (head) {
95 				in = head;
96 				blocks++;
97 				head = NULL;
98 			} else {
99 				src += GHASH_BLOCK_SIZE;
100 			}
101 
102 			crypto_xor((u8 *)&dst, in, GHASH_BLOCK_SIZE);
103 			gf128mul_lle(&dst, &key->k);
104 		} while (--blocks);
105 
106 		dg[0] = be64_to_cpu(dst.b);
107 		dg[1] = be64_to_cpu(dst.a);
108 	}
109 }
110 
111 static int ghash_update(struct shash_desc *desc, const u8 *src,
112 			unsigned int len)
113 {
114 	struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
115 	unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
116 
117 	ctx->count += len;
118 
119 	if ((partial + len) >= GHASH_BLOCK_SIZE) {
120 		struct ghash_key *key = crypto_shash_ctx(desc->tfm);
121 		int blocks;
122 
123 		if (partial) {
124 			int p = GHASH_BLOCK_SIZE - partial;
125 
126 			memcpy(ctx->buf + partial, src, p);
127 			src += p;
128 			len -= p;
129 		}
130 
131 		blocks = len / GHASH_BLOCK_SIZE;
132 		len %= GHASH_BLOCK_SIZE;
133 
134 		ghash_do_update(blocks, ctx->digest, src, key,
135 				partial ? ctx->buf : NULL);
136 		src += blocks * GHASH_BLOCK_SIZE;
137 		partial = 0;
138 	}
139 	if (len)
140 		memcpy(ctx->buf + partial, src, len);
141 	return 0;
142 }
143 
144 static int ghash_final(struct shash_desc *desc, u8 *dst)
145 {
146 	struct ghash_desc_ctx *ctx = shash_desc_ctx(desc);
147 	unsigned int partial = ctx->count % GHASH_BLOCK_SIZE;
148 
149 	if (partial) {
150 		struct ghash_key *key = crypto_shash_ctx(desc->tfm);
151 
152 		memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
153 		ghash_do_update(1, ctx->digest, ctx->buf, key, NULL);
154 	}
155 	put_unaligned_be64(ctx->digest[1], dst);
156 	put_unaligned_be64(ctx->digest[0], dst + 8);
157 
158 	*ctx = (struct ghash_desc_ctx){};
159 	return 0;
160 }
161 
162 static void ghash_reflect(u64 h[], const be128 *k)
163 {
164 	u64 carry = be64_to_cpu(k->a) >> 63;
165 
166 	h[0] = (be64_to_cpu(k->b) << 1) | carry;
167 	h[1] = (be64_to_cpu(k->a) << 1) | (be64_to_cpu(k->b) >> 63);
168 
169 	if (carry)
170 		h[1] ^= 0xc200000000000000UL;
171 }
172 
173 static int ghash_setkey(struct crypto_shash *tfm,
174 			const u8 *inkey, unsigned int keylen)
175 {
176 	struct ghash_key *key = crypto_shash_ctx(tfm);
177 
178 	if (keylen != GHASH_BLOCK_SIZE)
179 		return -EINVAL;
180 
181 	/* needed for the fallback */
182 	memcpy(&key->k, inkey, GHASH_BLOCK_SIZE);
183 	ghash_reflect(key->h[0], &key->k);
184 
185 	if (static_branch_likely(&use_p64)) {
186 		be128 h = key->k;
187 
188 		gf128mul_lle(&h, &key->k);
189 		ghash_reflect(key->h[1], &h);
190 
191 		gf128mul_lle(&h, &key->k);
192 		ghash_reflect(key->h[2], &h);
193 
194 		gf128mul_lle(&h, &key->k);
195 		ghash_reflect(key->h[3], &h);
196 	}
197 	return 0;
198 }
199 
200 static struct shash_alg ghash_alg = {
201 	.digestsize		= GHASH_DIGEST_SIZE,
202 	.init			= ghash_init,
203 	.update			= ghash_update,
204 	.final			= ghash_final,
205 	.setkey			= ghash_setkey,
206 	.descsize		= sizeof(struct ghash_desc_ctx),
207 
208 	.base.cra_name		= "ghash",
209 	.base.cra_driver_name	= "ghash-ce-sync",
210 	.base.cra_priority	= 300 - 1,
211 	.base.cra_blocksize	= GHASH_BLOCK_SIZE,
212 	.base.cra_ctxsize	= sizeof(struct ghash_key) + sizeof(u64[2]),
213 	.base.cra_module	= THIS_MODULE,
214 };
215 
216 static int ghash_async_init(struct ahash_request *req)
217 {
218 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
219 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
220 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
221 	struct cryptd_ahash *cryptd_tfm = ctx->cryptd_tfm;
222 	struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
223 	struct crypto_shash *child = cryptd_ahash_child(cryptd_tfm);
224 
225 	desc->tfm = child;
226 	return crypto_shash_init(desc);
227 }
228 
229 static int ghash_async_update(struct ahash_request *req)
230 {
231 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
232 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
233 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
234 	struct cryptd_ahash *cryptd_tfm = ctx->cryptd_tfm;
235 
236 	if (!crypto_simd_usable() ||
237 	    (in_atomic() && cryptd_ahash_queued(cryptd_tfm))) {
238 		memcpy(cryptd_req, req, sizeof(*req));
239 		ahash_request_set_tfm(cryptd_req, &cryptd_tfm->base);
240 		return crypto_ahash_update(cryptd_req);
241 	} else {
242 		struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
243 		return shash_ahash_update(req, desc);
244 	}
245 }
246 
247 static int ghash_async_final(struct ahash_request *req)
248 {
249 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
250 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
251 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
252 	struct cryptd_ahash *cryptd_tfm = ctx->cryptd_tfm;
253 
254 	if (!crypto_simd_usable() ||
255 	    (in_atomic() && cryptd_ahash_queued(cryptd_tfm))) {
256 		memcpy(cryptd_req, req, sizeof(*req));
257 		ahash_request_set_tfm(cryptd_req, &cryptd_tfm->base);
258 		return crypto_ahash_final(cryptd_req);
259 	} else {
260 		struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
261 		return crypto_shash_final(desc, req->result);
262 	}
263 }
264 
265 static int ghash_async_digest(struct ahash_request *req)
266 {
267 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
268 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
269 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
270 	struct cryptd_ahash *cryptd_tfm = ctx->cryptd_tfm;
271 
272 	if (!crypto_simd_usable() ||
273 	    (in_atomic() && cryptd_ahash_queued(cryptd_tfm))) {
274 		memcpy(cryptd_req, req, sizeof(*req));
275 		ahash_request_set_tfm(cryptd_req, &cryptd_tfm->base);
276 		return crypto_ahash_digest(cryptd_req);
277 	} else {
278 		struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
279 		struct crypto_shash *child = cryptd_ahash_child(cryptd_tfm);
280 
281 		desc->tfm = child;
282 		return shash_ahash_digest(req, desc);
283 	}
284 }
285 
286 static int ghash_async_import(struct ahash_request *req, const void *in)
287 {
288 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
289 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
290 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
291 	struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
292 
293 	desc->tfm = cryptd_ahash_child(ctx->cryptd_tfm);
294 
295 	return crypto_shash_import(desc, in);
296 }
297 
298 static int ghash_async_export(struct ahash_request *req, void *out)
299 {
300 	struct ahash_request *cryptd_req = ahash_request_ctx(req);
301 	struct shash_desc *desc = cryptd_shash_desc(cryptd_req);
302 
303 	return crypto_shash_export(desc, out);
304 }
305 
306 static int ghash_async_setkey(struct crypto_ahash *tfm, const u8 *key,
307 			      unsigned int keylen)
308 {
309 	struct ghash_async_ctx *ctx = crypto_ahash_ctx(tfm);
310 	struct crypto_ahash *child = &ctx->cryptd_tfm->base;
311 
312 	crypto_ahash_clear_flags(child, CRYPTO_TFM_REQ_MASK);
313 	crypto_ahash_set_flags(child, crypto_ahash_get_flags(tfm)
314 			       & CRYPTO_TFM_REQ_MASK);
315 	return crypto_ahash_setkey(child, key, keylen);
316 }
317 
318 static int ghash_async_init_tfm(struct crypto_tfm *tfm)
319 {
320 	struct cryptd_ahash *cryptd_tfm;
321 	struct ghash_async_ctx *ctx = crypto_tfm_ctx(tfm);
322 
323 	cryptd_tfm = cryptd_alloc_ahash("ghash-ce-sync", 0, 0);
324 	if (IS_ERR(cryptd_tfm))
325 		return PTR_ERR(cryptd_tfm);
326 	ctx->cryptd_tfm = cryptd_tfm;
327 	crypto_ahash_set_reqsize(__crypto_ahash_cast(tfm),
328 				 sizeof(struct ahash_request) +
329 				 crypto_ahash_reqsize(&cryptd_tfm->base));
330 
331 	return 0;
332 }
333 
334 static void ghash_async_exit_tfm(struct crypto_tfm *tfm)
335 {
336 	struct ghash_async_ctx *ctx = crypto_tfm_ctx(tfm);
337 
338 	cryptd_free_ahash(ctx->cryptd_tfm);
339 }
340 
341 static struct ahash_alg ghash_async_alg = {
342 	.init			= ghash_async_init,
343 	.update			= ghash_async_update,
344 	.final			= ghash_async_final,
345 	.setkey			= ghash_async_setkey,
346 	.digest			= ghash_async_digest,
347 	.import			= ghash_async_import,
348 	.export			= ghash_async_export,
349 	.halg.digestsize	= GHASH_DIGEST_SIZE,
350 	.halg.statesize		= sizeof(struct ghash_desc_ctx),
351 	.halg.base		= {
352 		.cra_name	= "ghash",
353 		.cra_driver_name = "ghash-ce",
354 		.cra_priority	= 300,
355 		.cra_flags	= CRYPTO_ALG_ASYNC,
356 		.cra_blocksize	= GHASH_BLOCK_SIZE,
357 		.cra_ctxsize	= sizeof(struct ghash_async_ctx),
358 		.cra_module	= THIS_MODULE,
359 		.cra_init	= ghash_async_init_tfm,
360 		.cra_exit	= ghash_async_exit_tfm,
361 	},
362 };
363 
364 
365 void pmull_gcm_encrypt(int blocks, u64 dg[], const char *src,
366 		       struct gcm_key const *k, char *dst,
367 		       const char *iv, int rounds, u32 counter);
368 
369 void pmull_gcm_enc_final(int blocks, u64 dg[], char *tag,
370 			 struct gcm_key const *k, char *head,
371 			 const char *iv, int rounds, u32 counter);
372 
373 void pmull_gcm_decrypt(int bytes, u64 dg[], const char *src,
374 		       struct gcm_key const *k, char *dst,
375 		       const char *iv, int rounds, u32 counter);
376 
377 int pmull_gcm_dec_final(int bytes, u64 dg[], char *tag,
378 			struct gcm_key const *k, char *head,
379 			const char *iv, int rounds, u32 counter,
380 			const char *otag, int authsize);
381 
382 static int gcm_aes_setkey(struct crypto_aead *tfm, const u8 *inkey,
383 			  unsigned int keylen)
384 {
385 	struct gcm_key *ctx = crypto_aead_ctx(tfm);
386 	struct crypto_aes_ctx aes_ctx;
387 	be128 h, k;
388 	int ret;
389 
390 	ret = aes_expandkey(&aes_ctx, inkey, keylen);
391 	if (ret)
392 		return -EINVAL;
393 
394 	aes_encrypt(&aes_ctx, (u8 *)&k, (u8[AES_BLOCK_SIZE]){});
395 
396 	memcpy(ctx->rk, aes_ctx.key_enc, sizeof(ctx->rk));
397 	ctx->rounds = 6 + keylen / 4;
398 
399 	memzero_explicit(&aes_ctx, sizeof(aes_ctx));
400 
401 	ghash_reflect(ctx->h[0], &k);
402 
403 	h = k;
404 	gf128mul_lle(&h, &k);
405 	ghash_reflect(ctx->h[1], &h);
406 
407 	gf128mul_lle(&h, &k);
408 	ghash_reflect(ctx->h[2], &h);
409 
410 	gf128mul_lle(&h, &k);
411 	ghash_reflect(ctx->h[3], &h);
412 
413 	return 0;
414 }
415 
416 static int gcm_aes_setauthsize(struct crypto_aead *tfm, unsigned int authsize)
417 {
418 	return crypto_gcm_check_authsize(authsize);
419 }
420 
421 static void gcm_update_mac(u64 dg[], const u8 *src, int count, u8 buf[],
422 			   int *buf_count, struct gcm_key *ctx)
423 {
424 	if (*buf_count > 0) {
425 		int buf_added = min(count, GHASH_BLOCK_SIZE - *buf_count);
426 
427 		memcpy(&buf[*buf_count], src, buf_added);
428 
429 		*buf_count += buf_added;
430 		src += buf_added;
431 		count -= buf_added;
432 	}
433 
434 	if (count >= GHASH_BLOCK_SIZE || *buf_count == GHASH_BLOCK_SIZE) {
435 		int blocks = count / GHASH_BLOCK_SIZE;
436 
437 		pmull_ghash_update_p64(blocks, dg, src, ctx->h,
438 				       *buf_count ? buf : NULL);
439 
440 		src += blocks * GHASH_BLOCK_SIZE;
441 		count %= GHASH_BLOCK_SIZE;
442 		*buf_count = 0;
443 	}
444 
445 	if (count > 0) {
446 		memcpy(buf, src, count);
447 		*buf_count = count;
448 	}
449 }
450 
451 static void gcm_calculate_auth_mac(struct aead_request *req, u64 dg[], u32 len)
452 {
453 	struct crypto_aead *aead = crypto_aead_reqtfm(req);
454 	struct gcm_key *ctx = crypto_aead_ctx(aead);
455 	u8 buf[GHASH_BLOCK_SIZE];
456 	struct scatter_walk walk;
457 	int buf_count = 0;
458 
459 	scatterwalk_start(&walk, req->src);
460 
461 	do {
462 		u32 n = scatterwalk_clamp(&walk, len);
463 		u8 *p;
464 
465 		if (!n) {
466 			scatterwalk_start(&walk, sg_next(walk.sg));
467 			n = scatterwalk_clamp(&walk, len);
468 		}
469 
470 		p = scatterwalk_map(&walk);
471 		gcm_update_mac(dg, p, n, buf, &buf_count, ctx);
472 		scatterwalk_unmap(p);
473 
474 		if (unlikely(len / SZ_4K > (len - n) / SZ_4K)) {
475 			kernel_neon_end();
476 			kernel_neon_begin();
477 		}
478 
479 		len -= n;
480 		scatterwalk_advance(&walk, n);
481 		scatterwalk_done(&walk, 0, len);
482 	} while (len);
483 
484 	if (buf_count) {
485 		memset(&buf[buf_count], 0, GHASH_BLOCK_SIZE - buf_count);
486 		pmull_ghash_update_p64(1, dg, buf, ctx->h, NULL);
487 	}
488 }
489 
490 static int gcm_encrypt(struct aead_request *req, const u8 *iv, u32 assoclen)
491 {
492 	struct crypto_aead *aead = crypto_aead_reqtfm(req);
493 	struct gcm_key *ctx = crypto_aead_ctx(aead);
494 	struct skcipher_walk walk;
495 	u8 buf[AES_BLOCK_SIZE];
496 	u32 counter = 2;
497 	u64 dg[2] = {};
498 	be128 lengths;
499 	const u8 *src;
500 	u8 *tag, *dst;
501 	int tail, err;
502 
503 	if (WARN_ON_ONCE(!may_use_simd()))
504 		return -EBUSY;
505 
506 	err = skcipher_walk_aead_encrypt(&walk, req, false);
507 
508 	kernel_neon_begin();
509 
510 	if (assoclen)
511 		gcm_calculate_auth_mac(req, dg, assoclen);
512 
513 	src = walk.src.virt.addr;
514 	dst = walk.dst.virt.addr;
515 
516 	while (walk.nbytes >= AES_BLOCK_SIZE) {
517 		int nblocks = walk.nbytes / AES_BLOCK_SIZE;
518 
519 		pmull_gcm_encrypt(nblocks, dg, src, ctx, dst, iv,
520 				  ctx->rounds, counter);
521 		counter += nblocks;
522 
523 		if (walk.nbytes == walk.total) {
524 			src += nblocks * AES_BLOCK_SIZE;
525 			dst += nblocks * AES_BLOCK_SIZE;
526 			break;
527 		}
528 
529 		kernel_neon_end();
530 
531 		err = skcipher_walk_done(&walk,
532 					 walk.nbytes % AES_BLOCK_SIZE);
533 		if (err)
534 			return err;
535 
536 		src = walk.src.virt.addr;
537 		dst = walk.dst.virt.addr;
538 
539 		kernel_neon_begin();
540 	}
541 
542 
543 	lengths.a = cpu_to_be64(assoclen * 8);
544 	lengths.b = cpu_to_be64(req->cryptlen * 8);
545 
546 	tag = (u8 *)&lengths;
547 	tail = walk.nbytes % AES_BLOCK_SIZE;
548 
549 	/*
550 	 * Bounce via a buffer unless we are encrypting in place and src/dst
551 	 * are not pointing to the start of the walk buffer. In that case, we
552 	 * can do a NEON load/xor/store sequence in place as long as we move
553 	 * the plain/ciphertext and keystream to the start of the register. If
554 	 * not, do a memcpy() to the end of the buffer so we can reuse the same
555 	 * logic.
556 	 */
557 	if (unlikely(tail && (tail == walk.nbytes || src != dst)))
558 		src = memcpy(buf + sizeof(buf) - tail, src, tail);
559 
560 	pmull_gcm_enc_final(tail, dg, tag, ctx, (u8 *)src, iv,
561 			    ctx->rounds, counter);
562 	kernel_neon_end();
563 
564 	if (unlikely(tail && src != dst))
565 		memcpy(dst, src, tail);
566 
567 	if (walk.nbytes) {
568 		err = skcipher_walk_done(&walk, 0);
569 		if (err)
570 			return err;
571 	}
572 
573 	/* copy authtag to end of dst */
574 	scatterwalk_map_and_copy(tag, req->dst, req->assoclen + req->cryptlen,
575 				 crypto_aead_authsize(aead), 1);
576 
577 	return 0;
578 }
579 
580 static int gcm_decrypt(struct aead_request *req, const u8 *iv, u32 assoclen)
581 {
582 	struct crypto_aead *aead = crypto_aead_reqtfm(req);
583 	struct gcm_key *ctx = crypto_aead_ctx(aead);
584 	int authsize = crypto_aead_authsize(aead);
585 	struct skcipher_walk walk;
586 	u8 otag[AES_BLOCK_SIZE];
587 	u8 buf[AES_BLOCK_SIZE];
588 	u32 counter = 2;
589 	u64 dg[2] = {};
590 	be128 lengths;
591 	const u8 *src;
592 	u8 *tag, *dst;
593 	int tail, err, ret;
594 
595 	if (WARN_ON_ONCE(!may_use_simd()))
596 		return -EBUSY;
597 
598 	scatterwalk_map_and_copy(otag, req->src,
599 				 req->assoclen + req->cryptlen - authsize,
600 				 authsize, 0);
601 
602 	err = skcipher_walk_aead_decrypt(&walk, req, false);
603 
604 	kernel_neon_begin();
605 
606 	if (assoclen)
607 		gcm_calculate_auth_mac(req, dg, assoclen);
608 
609 	src = walk.src.virt.addr;
610 	dst = walk.dst.virt.addr;
611 
612 	while (walk.nbytes >= AES_BLOCK_SIZE) {
613 		int nblocks = walk.nbytes / AES_BLOCK_SIZE;
614 
615 		pmull_gcm_decrypt(nblocks, dg, src, ctx, dst, iv,
616 				  ctx->rounds, counter);
617 		counter += nblocks;
618 
619 		if (walk.nbytes == walk.total) {
620 			src += nblocks * AES_BLOCK_SIZE;
621 			dst += nblocks * AES_BLOCK_SIZE;
622 			break;
623 		}
624 
625 		kernel_neon_end();
626 
627 		err = skcipher_walk_done(&walk,
628 					 walk.nbytes % AES_BLOCK_SIZE);
629 		if (err)
630 			return err;
631 
632 		src = walk.src.virt.addr;
633 		dst = walk.dst.virt.addr;
634 
635 		kernel_neon_begin();
636 	}
637 
638 	lengths.a = cpu_to_be64(assoclen * 8);
639 	lengths.b = cpu_to_be64((req->cryptlen - authsize) * 8);
640 
641 	tag = (u8 *)&lengths;
642 	tail = walk.nbytes % AES_BLOCK_SIZE;
643 
644 	if (unlikely(tail && (tail == walk.nbytes || src != dst)))
645 		src = memcpy(buf + sizeof(buf) - tail, src, tail);
646 
647 	ret = pmull_gcm_dec_final(tail, dg, tag, ctx, (u8 *)src, iv,
648 				  ctx->rounds, counter, otag, authsize);
649 	kernel_neon_end();
650 
651 	if (unlikely(tail && src != dst))
652 		memcpy(dst, src, tail);
653 
654 	if (walk.nbytes) {
655 		err = skcipher_walk_done(&walk, 0);
656 		if (err)
657 			return err;
658 	}
659 
660 	return ret ? -EBADMSG : 0;
661 }
662 
663 static int gcm_aes_encrypt(struct aead_request *req)
664 {
665 	return gcm_encrypt(req, req->iv, req->assoclen);
666 }
667 
668 static int gcm_aes_decrypt(struct aead_request *req)
669 {
670 	return gcm_decrypt(req, req->iv, req->assoclen);
671 }
672 
673 static int rfc4106_setkey(struct crypto_aead *tfm, const u8 *inkey,
674 			  unsigned int keylen)
675 {
676 	struct gcm_key *ctx = crypto_aead_ctx(tfm);
677 	int err;
678 
679 	keylen -= RFC4106_NONCE_SIZE;
680 	err = gcm_aes_setkey(tfm, inkey, keylen);
681 	if (err)
682 		return err;
683 
684 	memcpy(ctx->nonce, inkey + keylen, RFC4106_NONCE_SIZE);
685 	return 0;
686 }
687 
688 static int rfc4106_setauthsize(struct crypto_aead *tfm, unsigned int authsize)
689 {
690 	return crypto_rfc4106_check_authsize(authsize);
691 }
692 
693 static int rfc4106_encrypt(struct aead_request *req)
694 {
695 	struct crypto_aead *aead = crypto_aead_reqtfm(req);
696 	struct gcm_key *ctx = crypto_aead_ctx(aead);
697 	u8 iv[GCM_AES_IV_SIZE];
698 
699 	memcpy(iv, ctx->nonce, RFC4106_NONCE_SIZE);
700 	memcpy(iv + RFC4106_NONCE_SIZE, req->iv, GCM_RFC4106_IV_SIZE);
701 
702 	return crypto_ipsec_check_assoclen(req->assoclen) ?:
703 	       gcm_encrypt(req, iv, req->assoclen - GCM_RFC4106_IV_SIZE);
704 }
705 
706 static int rfc4106_decrypt(struct aead_request *req)
707 {
708 	struct crypto_aead *aead = crypto_aead_reqtfm(req);
709 	struct gcm_key *ctx = crypto_aead_ctx(aead);
710 	u8 iv[GCM_AES_IV_SIZE];
711 
712 	memcpy(iv, ctx->nonce, RFC4106_NONCE_SIZE);
713 	memcpy(iv + RFC4106_NONCE_SIZE, req->iv, GCM_RFC4106_IV_SIZE);
714 
715 	return crypto_ipsec_check_assoclen(req->assoclen) ?:
716 	       gcm_decrypt(req, iv, req->assoclen - GCM_RFC4106_IV_SIZE);
717 }
718 
719 static struct aead_alg gcm_aes_algs[] = {{
720 	.ivsize			= GCM_AES_IV_SIZE,
721 	.chunksize		= AES_BLOCK_SIZE,
722 	.maxauthsize		= AES_BLOCK_SIZE,
723 	.setkey			= gcm_aes_setkey,
724 	.setauthsize		= gcm_aes_setauthsize,
725 	.encrypt		= gcm_aes_encrypt,
726 	.decrypt		= gcm_aes_decrypt,
727 
728 	.base.cra_name		= "gcm(aes)",
729 	.base.cra_driver_name	= "gcm-aes-ce",
730 	.base.cra_priority	= 400,
731 	.base.cra_blocksize	= 1,
732 	.base.cra_ctxsize	= sizeof(struct gcm_key),
733 	.base.cra_module	= THIS_MODULE,
734 }, {
735 	.ivsize			= GCM_RFC4106_IV_SIZE,
736 	.chunksize		= AES_BLOCK_SIZE,
737 	.maxauthsize		= AES_BLOCK_SIZE,
738 	.setkey			= rfc4106_setkey,
739 	.setauthsize		= rfc4106_setauthsize,
740 	.encrypt		= rfc4106_encrypt,
741 	.decrypt		= rfc4106_decrypt,
742 
743 	.base.cra_name		= "rfc4106(gcm(aes))",
744 	.base.cra_driver_name	= "rfc4106-gcm-aes-ce",
745 	.base.cra_priority	= 400,
746 	.base.cra_blocksize	= 1,
747 	.base.cra_ctxsize	= sizeof(struct gcm_key) + RFC4106_NONCE_SIZE,
748 	.base.cra_module	= THIS_MODULE,
749 }};
750 
751 static int __init ghash_ce_mod_init(void)
752 {
753 	int err;
754 
755 	if (!(elf_hwcap & HWCAP_NEON))
756 		return -ENODEV;
757 
758 	if (elf_hwcap2 & HWCAP2_PMULL) {
759 		err = crypto_register_aeads(gcm_aes_algs,
760 					    ARRAY_SIZE(gcm_aes_algs));
761 		if (err)
762 			return err;
763 		ghash_alg.base.cra_ctxsize += 3 * sizeof(u64[2]);
764 		static_branch_enable(&use_p64);
765 	}
766 
767 	err = crypto_register_shash(&ghash_alg);
768 	if (err)
769 		goto err_aead;
770 	err = crypto_register_ahash(&ghash_async_alg);
771 	if (err)
772 		goto err_shash;
773 
774 	return 0;
775 
776 err_shash:
777 	crypto_unregister_shash(&ghash_alg);
778 err_aead:
779 	if (elf_hwcap2 & HWCAP2_PMULL)
780 		crypto_unregister_aeads(gcm_aes_algs,
781 					ARRAY_SIZE(gcm_aes_algs));
782 	return err;
783 }
784 
785 static void __exit ghash_ce_mod_exit(void)
786 {
787 	crypto_unregister_ahash(&ghash_async_alg);
788 	crypto_unregister_shash(&ghash_alg);
789 	if (elf_hwcap2 & HWCAP2_PMULL)
790 		crypto_unregister_aeads(gcm_aes_algs,
791 					ARRAY_SIZE(gcm_aes_algs));
792 }
793 
794 module_init(ghash_ce_mod_init);
795 module_exit(ghash_ce_mod_exit);
796