xref: /dragonfly/crypto/libressl/ssl/s3_cbc.c (revision c9c5aa9e)
1 /* $OpenBSD: s3_cbc.c,v 1.22 2020/06/19 21:26:40 tb Exp $ */
2 /* ====================================================================
3  * Copyright (c) 2012 The OpenSSL Project.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  *
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  *
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in
14  *    the documentation and/or other materials provided with the
15  *    distribution.
16  *
17  * 3. All advertising materials mentioning features or use of this
18  *    software must display the following acknowledgment:
19  *    "This product includes software developed by the OpenSSL Project
20  *    for use in the OpenSSL Toolkit. (http://www.openssl.org/)"
21  *
22  * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
23  *    endorse or promote products derived from this software without
24  *    prior written permission. For written permission, please contact
25  *    openssl-core@openssl.org.
26  *
27  * 5. Products derived from this software may not be called "OpenSSL"
28  *    nor may "OpenSSL" appear in their names without prior written
29  *    permission of the OpenSSL Project.
30  *
31  * 6. Redistributions of any form whatsoever must retain the following
32  *    acknowledgment:
33  *    "This product includes software developed by the OpenSSL Project
34  *    for use in the OpenSSL Toolkit (http://www.openssl.org/)"
35  *
36  * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
37  * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
38  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
39  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE OpenSSL PROJECT OR
40  * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
41  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
42  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
43  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
44  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
45  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
46  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
47  * OF THE POSSIBILITY OF SUCH DAMAGE.
48  * ====================================================================
49  *
50  * This product includes cryptographic software written by Eric Young
51  * (eay@cryptsoft.com).  This product includes software written by Tim
52  * Hudson (tjh@cryptsoft.com).
53  *
54  */
55 
56 #include "ssl_locl.h"
57 
58 #include <openssl/md5.h>
59 #include <openssl/sha.h>
60 
61 /* MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's length
62  * field. (SHA-384/512 have 128-bit length.) */
63 #define MAX_HASH_BIT_COUNT_BYTES 16
64 
65 /* MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support.
66  * Currently SHA-384/512 has a 128-byte block size and that's the largest
67  * supported by TLS.) */
68 #define MAX_HASH_BLOCK_SIZE 128
69 
70 /* Some utility functions are needed:
71  *
72  * These macros return the given value with the MSB copied to all the other
73  * bits. They use the fact that arithmetic shift shifts-in the sign bit.
74  * However, this is not ensured by the C standard so you may need to replace
75  * them with something else on odd CPUs. */
76 #define DUPLICATE_MSB_TO_ALL(x) ((unsigned int)((int)(x) >> (sizeof(int) * 8 - 1)))
77 #define DUPLICATE_MSB_TO_ALL_8(x) ((unsigned char)(DUPLICATE_MSB_TO_ALL(x)))
78 
79 /* constant_time_lt returns 0xff if a<b and 0x00 otherwise. */
80 static unsigned int
81 constant_time_lt(unsigned int a, unsigned int b)
82 {
83 	a -= b;
84 	return DUPLICATE_MSB_TO_ALL(a);
85 }
86 
87 /* constant_time_ge returns 0xff if a>=b and 0x00 otherwise. */
88 static unsigned int
89 constant_time_ge(unsigned int a, unsigned int b)
90 {
91 	a -= b;
92 	return DUPLICATE_MSB_TO_ALL(~a);
93 }
94 
95 /* constant_time_eq_8 returns 0xff if a==b and 0x00 otherwise. */
96 static unsigned char
97 constant_time_eq_8(unsigned int a, unsigned int b)
98 {
99 	unsigned int c = a ^ b;
100 	c--;
101 	return DUPLICATE_MSB_TO_ALL_8(c);
102 }
103 
104 /* tls1_cbc_remove_padding removes the CBC padding from the decrypted, TLS, CBC
105  * record in |rec| in constant time and returns 1 if the padding is valid and
106  * -1 otherwise. It also removes any explicit IV from the start of the record
107  * without leaking any timing about whether there was enough space after the
108  * padding was removed.
109  *
110  * block_size: the block size of the cipher used to encrypt the record.
111  * returns:
112  *   0: (in non-constant time) if the record is publicly invalid.
113  *   1: if the padding was valid
114  *  -1: otherwise. */
115 int
116 tls1_cbc_remove_padding(const SSL* s, SSL3_RECORD_INTERNAL *rec,
117     unsigned int block_size, unsigned int mac_size)
118 {
119 	unsigned int padding_length, good, to_check, i;
120 	const unsigned int overhead = 1 /* padding length byte */ + mac_size;
121 
122 	/* Check if version requires explicit IV */
123 	if (SSL_USE_EXPLICIT_IV(s)) {
124 		/* These lengths are all public so we can test them in
125 		 * non-constant time.
126 		 */
127 		if (overhead + block_size > rec->length)
128 			return 0;
129 		/* We can now safely skip explicit IV */
130 		rec->data += block_size;
131 		rec->input += block_size;
132 		rec->length -= block_size;
133 	} else if (overhead > rec->length)
134 		return 0;
135 
136 	padding_length = rec->data[rec->length - 1];
137 
138 	good = constant_time_ge(rec->length, overhead + padding_length);
139 	/* The padding consists of a length byte at the end of the record and
140 	 * then that many bytes of padding, all with the same value as the
141 	 * length byte. Thus, with the length byte included, there are i+1
142 	 * bytes of padding.
143 	 *
144 	 * We can't check just |padding_length+1| bytes because that leaks
145 	 * decrypted information. Therefore we always have to check the maximum
146 	 * amount of padding possible. (Again, the length of the record is
147 	 * public information so we can use it.) */
148 	to_check = 256; /* maximum amount of padding, inc length byte. */
149 	if (to_check > rec->length)
150 		to_check = rec->length;
151 
152 	for (i = 0; i < to_check; i++) {
153 		unsigned char mask = constant_time_ge(padding_length, i);
154 		unsigned char b = rec->data[rec->length - 1 - i];
155 		/* The final |padding_length+1| bytes should all have the value
156 		 * |padding_length|. Therefore the XOR should be zero. */
157 		good &= ~(mask&(padding_length ^ b));
158 	}
159 
160 	/* If any of the final |padding_length+1| bytes had the wrong value,
161 	 * one or more of the lower eight bits of |good| will be cleared. We
162 	 * AND the bottom 8 bits together and duplicate the result to all the
163 	 * bits. */
164 	good &= good >> 4;
165 	good &= good >> 2;
166 	good &= good >> 1;
167 	good <<= sizeof(good)*8 - 1;
168 	good = DUPLICATE_MSB_TO_ALL(good);
169 
170 	padding_length = good & (padding_length + 1);
171 	rec->length -= padding_length;
172 	rec->padding_length = padding_length;
173 
174 	return (int)((good & 1) | (~good & -1));
175 }
176 
177 /* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in
178  * constant time (independent of the concrete value of rec->length, which may
179  * vary within a 256-byte window).
180  *
181  * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to
182  * this function.
183  *
184  * On entry:
185  *   rec->orig_len >= md_size
186  *   md_size <= EVP_MAX_MD_SIZE
187  *
188  * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with
189  * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into
190  * a single or pair of cache-lines, then the variable memory accesses don't
191  * actually affect the timing. CPUs with smaller cache-lines [if any] are
192  * not multi-core and are not considered vulnerable to cache-timing attacks.
193  */
194 #define CBC_MAC_ROTATE_IN_PLACE
195 
196 void
197 ssl3_cbc_copy_mac(unsigned char* out, const SSL3_RECORD_INTERNAL *rec,
198     unsigned int md_size, unsigned int orig_len)
199 {
200 #if defined(CBC_MAC_ROTATE_IN_PLACE)
201 	unsigned char rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
202 	unsigned char *rotated_mac;
203 #else
204 	unsigned char rotated_mac[EVP_MAX_MD_SIZE];
205 #endif
206 
207 	/* mac_end is the index of |rec->data| just after the end of the MAC. */
208 	unsigned int mac_end = rec->length;
209 	unsigned int mac_start = mac_end - md_size;
210 	/* scan_start contains the number of bytes that we can ignore because
211 	 * the MAC's position can only vary by 255 bytes. */
212 	unsigned int scan_start = 0;
213 	unsigned int i, j;
214 	unsigned int div_spoiler;
215 	unsigned int rotate_offset;
216 
217 	OPENSSL_assert(orig_len >= md_size);
218 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
219 
220 #if defined(CBC_MAC_ROTATE_IN_PLACE)
221 	rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf)&63);
222 #endif
223 
224 	/* This information is public so it's safe to branch based on it. */
225 	if (orig_len > md_size + 255 + 1)
226 		scan_start = orig_len - (md_size + 255 + 1);
227 	/* div_spoiler contains a multiple of md_size that is used to cause the
228 	 * modulo operation to be constant time. Without this, the time varies
229 	 * based on the amount of padding when running on Intel chips at least.
230 	 *
231 	 * The aim of right-shifting md_size is so that the compiler doesn't
232 	 * figure out that it can remove div_spoiler as that would require it
233 	 * to prove that md_size is always even, which I hope is beyond it. */
234 	div_spoiler = md_size >> 1;
235 	div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
236 	rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
237 
238 	memset(rotated_mac, 0, md_size);
239 	for (i = scan_start, j = 0; i < orig_len; i++) {
240 		unsigned char mac_started = constant_time_ge(i, mac_start);
241 		unsigned char mac_ended = constant_time_ge(i, mac_end);
242 		unsigned char b = rec->data[i];
243 		rotated_mac[j++] |= b & mac_started & ~mac_ended;
244 		j &= constant_time_lt(j, md_size);
245 	}
246 
247 	/* Now rotate the MAC */
248 #if defined(CBC_MAC_ROTATE_IN_PLACE)
249 	j = 0;
250 	for (i = 0; i < md_size; i++) {
251 		/* in case cache-line is 32 bytes, touch second line */
252 		((volatile unsigned char *)rotated_mac)[rotate_offset^32];
253 		out[j++] = rotated_mac[rotate_offset++];
254 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
255 	}
256 #else
257 	memset(out, 0, md_size);
258 	rotate_offset = md_size - rotate_offset;
259 	rotate_offset &= constant_time_lt(rotate_offset, md_size);
260 	for (i = 0; i < md_size; i++) {
261 		for (j = 0; j < md_size; j++)
262 			out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
263 		rotate_offset++;
264 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
265 	}
266 #endif
267 }
268 
269 #define l2n(l,c)	(*((c)++)=(unsigned char)(((l)>>24)&0xff), \
270 			 *((c)++)=(unsigned char)(((l)>>16)&0xff), \
271 			 *((c)++)=(unsigned char)(((l)>> 8)&0xff), \
272 			 *((c)++)=(unsigned char)(((l)    )&0xff))
273 
274 #define l2n8(l,c)	(*((c)++)=(unsigned char)(((l)>>56)&0xff), \
275 			 *((c)++)=(unsigned char)(((l)>>48)&0xff), \
276 			 *((c)++)=(unsigned char)(((l)>>40)&0xff), \
277 			 *((c)++)=(unsigned char)(((l)>>32)&0xff), \
278 			 *((c)++)=(unsigned char)(((l)>>24)&0xff), \
279 			 *((c)++)=(unsigned char)(((l)>>16)&0xff), \
280 			 *((c)++)=(unsigned char)(((l)>> 8)&0xff), \
281 			 *((c)++)=(unsigned char)(((l)    )&0xff))
282 
283 /* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
284  * little-endian order. The value of p is advanced by four. */
285 #define u32toLE(n, p) \
286 	(*((p)++)=(unsigned char)(n), \
287 	 *((p)++)=(unsigned char)(n>>8), \
288 	 *((p)++)=(unsigned char)(n>>16), \
289 	 *((p)++)=(unsigned char)(n>>24))
290 
291 /* These functions serialize the state of a hash and thus perform the standard
292  * "final" operation without adding the padding and length that such a function
293  * typically does. */
294 static void
295 tls1_md5_final_raw(void* ctx, unsigned char *md_out)
296 {
297 	MD5_CTX *md5 = ctx;
298 	u32toLE(md5->A, md_out);
299 	u32toLE(md5->B, md_out);
300 	u32toLE(md5->C, md_out);
301 	u32toLE(md5->D, md_out);
302 }
303 
304 static void
305 tls1_sha1_final_raw(void* ctx, unsigned char *md_out)
306 {
307 	SHA_CTX *sha1 = ctx;
308 	l2n(sha1->h0, md_out);
309 	l2n(sha1->h1, md_out);
310 	l2n(sha1->h2, md_out);
311 	l2n(sha1->h3, md_out);
312 	l2n(sha1->h4, md_out);
313 }
314 
315 static void
316 tls1_sha256_final_raw(void* ctx, unsigned char *md_out)
317 {
318 	SHA256_CTX *sha256 = ctx;
319 	unsigned int i;
320 
321 	for (i = 0; i < 8; i++) {
322 		l2n(sha256->h[i], md_out);
323 	}
324 }
325 
326 static void
327 tls1_sha512_final_raw(void* ctx, unsigned char *md_out)
328 {
329 	SHA512_CTX *sha512 = ctx;
330 	unsigned int i;
331 
332 	for (i = 0; i < 8; i++) {
333 		l2n8(sha512->h[i], md_out);
334 	}
335 }
336 
337 /* Largest hash context ever used by the functions above. */
338 #define LARGEST_DIGEST_CTX SHA512_CTX
339 
340 /* Type giving the alignment needed by the above */
341 #define LARGEST_DIGEST_CTX_ALIGNMENT SHA_LONG64
342 
343 /* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
344  * which ssl3_cbc_digest_record supports. */
345 char
346 ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx)
347 {
348 	switch (EVP_MD_CTX_type(ctx)) {
349 	case NID_md5:
350 	case NID_sha1:
351 	case NID_sha224:
352 	case NID_sha256:
353 	case NID_sha384:
354 	case NID_sha512:
355 		return 1;
356 	default:
357 		return 0;
358 	}
359 }
360 
361 /* ssl3_cbc_digest_record computes the MAC of a decrypted, padded TLS
362  * record.
363  *
364  *   ctx: the EVP_MD_CTX from which we take the hash function.
365  *     ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX.
366  *   md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written.
367  *   md_out_size: if non-NULL, the number of output bytes is written here.
368  *   header: the 13-byte, TLS record header.
369  *   data: the record data itself, less any preceeding explicit IV.
370  *   data_plus_mac_size: the secret, reported length of the data and MAC
371  *     once the padding has been removed.
372  *   data_plus_mac_plus_padding_size: the public length of the whole
373  *     record, including padding.
374  *
375  * On entry: by virtue of having been through one of the remove_padding
376  * functions, above, we know that data_plus_mac_size is large enough to contain
377  * a padding byte and MAC. (If the padding was invalid, it might contain the
378  * padding too. )
379  */
380 int
381 ssl3_cbc_digest_record(const EVP_MD_CTX *ctx, unsigned char* md_out,
382     size_t* md_out_size, const unsigned char header[13],
383     const unsigned char *data, size_t data_plus_mac_size,
384     size_t data_plus_mac_plus_padding_size, const unsigned char *mac_secret,
385     unsigned int mac_secret_length)
386 {
387 	union {
388 		/*
389 		 * Alignment here is to allow this to be cast as SHA512_CTX
390 		 * without losing alignment required by the 64-bit SHA_LONG64
391 		 * integer it contains.
392 		 */
393 		LARGEST_DIGEST_CTX_ALIGNMENT align;
394 		unsigned char c[sizeof(LARGEST_DIGEST_CTX)];
395 	} md_state;
396 	void (*md_final_raw)(void *ctx, unsigned char *md_out);
397 	void (*md_transform)(void *ctx, const unsigned char *block);
398 	unsigned int md_size, md_block_size = 64;
399 	unsigned int header_length, variance_blocks,
400 	len, max_mac_bytes, num_blocks,
401 	num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
402 	unsigned int bits;	/* at most 18 bits */
403 	unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
404 	/* hmac_pad is the masked HMAC key. */
405 	unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
406 	unsigned char first_block[MAX_HASH_BLOCK_SIZE];
407 	unsigned char mac_out[EVP_MAX_MD_SIZE];
408 	unsigned int i, j, md_out_size_u;
409 	EVP_MD_CTX md_ctx;
410 	/* mdLengthSize is the number of bytes in the length field that terminates
411 	* the hash. */
412 	unsigned int md_length_size = 8;
413 	char length_is_big_endian = 1;
414 
415 	/* This is a, hopefully redundant, check that allows us to forget about
416 	 * many possible overflows later in this function. */
417 	OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024);
418 
419 	switch (EVP_MD_CTX_type(ctx)) {
420 	case NID_md5:
421 		MD5_Init((MD5_CTX*)md_state.c);
422 		md_final_raw = tls1_md5_final_raw;
423 		md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform;
424 		md_size = 16;
425 		length_is_big_endian = 0;
426 		break;
427 	case NID_sha1:
428 		SHA1_Init((SHA_CTX*)md_state.c);
429 		md_final_raw = tls1_sha1_final_raw;
430 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform;
431 		md_size = 20;
432 		break;
433 	case NID_sha224:
434 		SHA224_Init((SHA256_CTX*)md_state.c);
435 		md_final_raw = tls1_sha256_final_raw;
436 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
437 		md_size = 224/8;
438 		break;
439 	case NID_sha256:
440 		SHA256_Init((SHA256_CTX*)md_state.c);
441 		md_final_raw = tls1_sha256_final_raw;
442 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
443 		md_size = 32;
444 		break;
445 	case NID_sha384:
446 		SHA384_Init((SHA512_CTX*)md_state.c);
447 		md_final_raw = tls1_sha512_final_raw;
448 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
449 		md_size = 384/8;
450 		md_block_size = 128;
451 		md_length_size = 16;
452 		break;
453 	case NID_sha512:
454 		SHA512_Init((SHA512_CTX*)md_state.c);
455 		md_final_raw = tls1_sha512_final_raw;
456 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
457 		md_size = 64;
458 		md_block_size = 128;
459 		md_length_size = 16;
460 		break;
461 	default:
462 		/* ssl3_cbc_record_digest_supported should have been
463 		 * called first to check that the hash function is
464 		 * supported. */
465 		OPENSSL_assert(0);
466 		if (md_out_size)
467 			*md_out_size = 0;
468 		return 0;
469 	}
470 
471 	OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
472 	OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
473 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
474 
475 	header_length = 13;
476 
477 	/* variance_blocks is the number of blocks of the hash that we have to
478 	 * calculate in constant time because they could be altered by the
479 	 * padding value.
480 	 *
481 	 * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
482 	 * required to be minimal. Therefore we say that the final six blocks
483 	 * can vary based on the padding.
484 	 *
485 	 * Later in the function, if the message is short and there obviously
486 	 * cannot be this many blocks then variance_blocks can be reduced. */
487 	variance_blocks = 6;
488 	/* From now on we're dealing with the MAC, which conceptually has 13
489 	 * bytes of `header' before the start of the data (TLS) */
490 	len = data_plus_mac_plus_padding_size + header_length;
491 	/* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
492 	* |header|, assuming that there's no padding. */
493 	max_mac_bytes = len - md_size - 1;
494 	/* num_blocks is the maximum number of hash blocks. */
495 	num_blocks = (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
496 	/* In order to calculate the MAC in constant time we have to handle
497 	 * the final blocks specially because the padding value could cause the
498 	 * end to appear somewhere in the final |variance_blocks| blocks and we
499 	 * can't leak where. However, |num_starting_blocks| worth of data can
500 	 * be hashed right away because no padding value can affect whether
501 	 * they are plaintext. */
502 	num_starting_blocks = 0;
503 	/* k is the starting byte offset into the conceptual header||data where
504 	 * we start processing. */
505 	k = 0;
506 	/* mac_end_offset is the index just past the end of the data to be
507 	 * MACed. */
508 	mac_end_offset = data_plus_mac_size + header_length - md_size;
509 	/* c is the index of the 0x80 byte in the final hash block that
510 	 * contains application data. */
511 	c = mac_end_offset % md_block_size;
512 	/* index_a is the hash block number that contains the 0x80 terminating
513 	 * value. */
514 	index_a = mac_end_offset / md_block_size;
515 	/* index_b is the hash block number that contains the 64-bit hash
516 	 * length, in bits. */
517 	index_b = (mac_end_offset + md_length_size) / md_block_size;
518 	/* bits is the hash-length in bits. It includes the additional hash
519 	 * block for the masked HMAC key. */
520 
521 	if (num_blocks > variance_blocks) {
522 		num_starting_blocks = num_blocks - variance_blocks;
523 		k = md_block_size*num_starting_blocks;
524 	}
525 
526 	bits = 8*mac_end_offset;
527 	/* Compute the initial HMAC block. */
528 	bits += 8*md_block_size;
529 	memset(hmac_pad, 0, md_block_size);
530 	OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad));
531 	memcpy(hmac_pad, mac_secret, mac_secret_length);
532 	for (i = 0; i < md_block_size; i++)
533 		hmac_pad[i] ^= 0x36;
534 
535 	md_transform(md_state.c, hmac_pad);
536 
537 	if (length_is_big_endian) {
538 		memset(length_bytes, 0, md_length_size - 4);
539 		length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24);
540 		length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16);
541 		length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8);
542 		length_bytes[md_length_size - 1] = (unsigned char)bits;
543 	} else {
544 		memset(length_bytes, 0, md_length_size);
545 		length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24);
546 		length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16);
547 		length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8);
548 		length_bytes[md_length_size - 8] = (unsigned char)bits;
549 	}
550 
551 	if (k > 0) {
552 		/* k is a multiple of md_block_size. */
553 		memcpy(first_block, header, 13);
554 		memcpy(first_block + 13, data, md_block_size - 13);
555 		md_transform(md_state.c, first_block);
556 		for (i = 1; i < k/md_block_size; i++)
557 			md_transform(md_state.c, data + md_block_size*i - 13);
558 	}
559 
560 	memset(mac_out, 0, sizeof(mac_out));
561 
562 	/* We now process the final hash blocks. For each block, we construct
563 	 * it in constant time. If the |i==index_a| then we'll include the 0x80
564 	 * bytes and zero pad etc. For each block we selectively copy it, in
565 	 * constant time, to |mac_out|. */
566 	for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks; i++) {
567 		unsigned char block[MAX_HASH_BLOCK_SIZE];
568 		unsigned char is_block_a = constant_time_eq_8(i, index_a);
569 		unsigned char is_block_b = constant_time_eq_8(i, index_b);
570 		for (j = 0; j < md_block_size; j++) {
571 			unsigned char b = 0, is_past_c, is_past_cp1;
572 			if (k < header_length)
573 				b = header[k];
574 			else if (k < data_plus_mac_plus_padding_size + header_length)
575 				b = data[k - header_length];
576 			k++;
577 
578 			is_past_c = is_block_a & constant_time_ge(j, c);
579 			is_past_cp1 = is_block_a & constant_time_ge(j, c + 1);
580 			/* If this is the block containing the end of the
581 			 * application data, and we are at the offset for the
582 			 * 0x80 value, then overwrite b with 0x80. */
583 			b = (b&~is_past_c) | (0x80&is_past_c);
584 			/* If this is the block containing the end of the
585 			 * application data and we're past the 0x80 value then
586 			 * just write zero. */
587 			b = b&~is_past_cp1;
588 			/* If this is index_b (the final block), but not
589 			 * index_a (the end of the data), then the 64-bit
590 			 * length didn't fit into index_a and we're having to
591 			 * add an extra block of zeros. */
592 			b &= ~is_block_b | is_block_a;
593 
594 			/* The final bytes of one of the blocks contains the
595 			 * length. */
596 			if (j >= md_block_size - md_length_size) {
597 				/* If this is index_b, write a length byte. */
598 				b = (b&~is_block_b) | (is_block_b&length_bytes[j - (md_block_size - md_length_size)]);
599 			}
600 			block[j] = b;
601 		}
602 
603 		md_transform(md_state.c, block);
604 		md_final_raw(md_state.c, block);
605 		/* If this is index_b, copy the hash value to |mac_out|. */
606 		for (j = 0; j < md_size; j++)
607 			mac_out[j] |= block[j]&is_block_b;
608 	}
609 
610 	EVP_MD_CTX_init(&md_ctx);
611 	if (!EVP_DigestInit_ex(&md_ctx, ctx->digest, NULL /* engine */)) {
612 		EVP_MD_CTX_cleanup(&md_ctx);
613 		return 0;
614 	}
615 
616 	/* Complete the HMAC in the standard manner. */
617 	for (i = 0; i < md_block_size; i++)
618 		hmac_pad[i] ^= 0x6a;
619 
620 	EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size);
621 	EVP_DigestUpdate(&md_ctx, mac_out, md_size);
622 
623 	EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
624 	if (md_out_size)
625 		*md_out_size = md_out_size_u;
626 	EVP_MD_CTX_cleanup(&md_ctx);
627 
628 	return 1;
629 }
630