xref: /dragonfly/crypto/libressl/ssl/s3_cbc.c (revision ffe53622)
1 /* $OpenBSD: s3_cbc.c,v 1.11 2015/09/11 17:17:44 jsing 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)(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
81 constant_time_lt(unsigned a, unsigned 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
89 constant_time_ge(unsigned a, unsigned 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 a, unsigned b)
98 {
99 	unsigned 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 *rec, unsigned block_size,
117     unsigned mac_size)
118 {
119 	unsigned padding_length, good, to_check, i;
120 	const unsigned 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 	if (EVP_CIPHER_flags(s->enc_read_ctx->cipher) & EVP_CIPH_FLAG_AEAD_CIPHER) {
139 		/* padding is already verified */
140 		rec->length -= padding_length + 1;
141 		return 1;
142 	}
143 
144 	good = constant_time_ge(rec->length, overhead + padding_length);
145 	/* The padding consists of a length byte at the end of the record and
146 	 * then that many bytes of padding, all with the same value as the
147 	 * length byte. Thus, with the length byte included, there are i+1
148 	 * bytes of padding.
149 	 *
150 	 * We can't check just |padding_length+1| bytes because that leaks
151 	 * decrypted information. Therefore we always have to check the maximum
152 	 * amount of padding possible. (Again, the length of the record is
153 	 * public information so we can use it.) */
154 	to_check = 255; /* maximum amount of padding. */
155 	if (to_check > rec->length - 1)
156 		to_check = rec->length - 1;
157 
158 	for (i = 0; i < to_check; i++) {
159 		unsigned char mask = constant_time_ge(padding_length, i);
160 		unsigned char b = rec->data[rec->length - 1 - i];
161 		/* The final |padding_length+1| bytes should all have the value
162 		 * |padding_length|. Therefore the XOR should be zero. */
163 		good &= ~(mask&(padding_length ^ b));
164 	}
165 
166 	/* If any of the final |padding_length+1| bytes had the wrong value,
167 	 * one or more of the lower eight bits of |good| will be cleared. We
168 	 * AND the bottom 8 bits together and duplicate the result to all the
169 	 * bits. */
170 	good &= good >> 4;
171 	good &= good >> 2;
172 	good &= good >> 1;
173 	good <<= sizeof(good)*8 - 1;
174 	good = DUPLICATE_MSB_TO_ALL(good);
175 
176 	padding_length = good & (padding_length + 1);
177 	rec->length -= padding_length;
178 	rec->type |= padding_length<<8;	/* kludge: pass padding length */
179 
180 	return (int)((good & 1) | (~good & -1));
181 }
182 
183 /* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in
184  * constant time (independent of the concrete value of rec->length, which may
185  * vary within a 256-byte window).
186  *
187  * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to
188  * this function.
189  *
190  * On entry:
191  *   rec->orig_len >= md_size
192  *   md_size <= EVP_MAX_MD_SIZE
193  *
194  * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with
195  * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into
196  * a single or pair of cache-lines, then the variable memory accesses don't
197  * actually affect the timing. CPUs with smaller cache-lines [if any] are
198  * not multi-core and are not considered vulnerable to cache-timing attacks.
199  */
200 #define CBC_MAC_ROTATE_IN_PLACE
201 
202 void
203 ssl3_cbc_copy_mac(unsigned char* out, const SSL3_RECORD *rec,
204     unsigned md_size, unsigned orig_len)
205 {
206 #if defined(CBC_MAC_ROTATE_IN_PLACE)
207 	unsigned char rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
208 	unsigned char *rotated_mac;
209 #else
210 	unsigned char rotated_mac[EVP_MAX_MD_SIZE];
211 #endif
212 
213 	/* mac_end is the index of |rec->data| just after the end of the MAC. */
214 	unsigned mac_end = rec->length;
215 	unsigned mac_start = mac_end - md_size;
216 	/* scan_start contains the number of bytes that we can ignore because
217 	 * the MAC's position can only vary by 255 bytes. */
218 	unsigned scan_start = 0;
219 	unsigned i, j;
220 	unsigned div_spoiler;
221 	unsigned rotate_offset;
222 
223 	OPENSSL_assert(orig_len >= md_size);
224 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
225 
226 #if defined(CBC_MAC_ROTATE_IN_PLACE)
227 	rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf)&63);
228 #endif
229 
230 	/* This information is public so it's safe to branch based on it. */
231 	if (orig_len > md_size + 255 + 1)
232 		scan_start = orig_len - (md_size + 255 + 1);
233 	/* div_spoiler contains a multiple of md_size that is used to cause the
234 	 * modulo operation to be constant time. Without this, the time varies
235 	 * based on the amount of padding when running on Intel chips at least.
236 	 *
237 	 * The aim of right-shifting md_size is so that the compiler doesn't
238 	 * figure out that it can remove div_spoiler as that would require it
239 	 * to prove that md_size is always even, which I hope is beyond it. */
240 	div_spoiler = md_size >> 1;
241 	div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
242 	rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
243 
244 	memset(rotated_mac, 0, md_size);
245 	for (i = scan_start, j = 0; i < orig_len; i++) {
246 		unsigned char mac_started = constant_time_ge(i, mac_start);
247 		unsigned char mac_ended = constant_time_ge(i, mac_end);
248 		unsigned char b = rec->data[i];
249 		rotated_mac[j++] |= b & mac_started & ~mac_ended;
250 		j &= constant_time_lt(j, md_size);
251 	}
252 
253 	/* Now rotate the MAC */
254 #if defined(CBC_MAC_ROTATE_IN_PLACE)
255 	j = 0;
256 	for (i = 0; i < md_size; i++) {
257 		/* in case cache-line is 32 bytes, touch second line */
258 		((volatile unsigned char *)rotated_mac)[rotate_offset^32];
259 		out[j++] = rotated_mac[rotate_offset++];
260 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
261 	}
262 #else
263 	memset(out, 0, md_size);
264 	rotate_offset = md_size - rotate_offset;
265 	rotate_offset &= constant_time_lt(rotate_offset, md_size);
266 	for (i = 0; i < md_size; i++) {
267 		for (j = 0; j < md_size; j++)
268 			out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
269 		rotate_offset++;
270 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
271 	}
272 #endif
273 }
274 
275 /* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
276  * little-endian order. The value of p is advanced by four. */
277 #define u32toLE(n, p) \
278 	(*((p)++)=(unsigned char)(n), \
279 	 *((p)++)=(unsigned char)(n>>8), \
280 	 *((p)++)=(unsigned char)(n>>16), \
281 	 *((p)++)=(unsigned char)(n>>24))
282 
283 /* These functions serialize the state of a hash and thus perform the standard
284  * "final" operation without adding the padding and length that such a function
285  * typically does. */
286 static void
287 tls1_md5_final_raw(void* ctx, unsigned char *md_out)
288 {
289 	MD5_CTX *md5 = ctx;
290 	u32toLE(md5->A, md_out);
291 	u32toLE(md5->B, md_out);
292 	u32toLE(md5->C, md_out);
293 	u32toLE(md5->D, md_out);
294 }
295 
296 static void
297 tls1_sha1_final_raw(void* ctx, unsigned char *md_out)
298 {
299 	SHA_CTX *sha1 = ctx;
300 	l2n(sha1->h0, md_out);
301 	l2n(sha1->h1, md_out);
302 	l2n(sha1->h2, md_out);
303 	l2n(sha1->h3, md_out);
304 	l2n(sha1->h4, md_out);
305 }
306 #define LARGEST_DIGEST_CTX SHA_CTX
307 
308 static void
309 tls1_sha256_final_raw(void* ctx, unsigned char *md_out)
310 {
311 	SHA256_CTX *sha256 = ctx;
312 	unsigned i;
313 
314 	for (i = 0; i < 8; i++) {
315 		l2n(sha256->h[i], md_out);
316 	}
317 }
318 #undef  LARGEST_DIGEST_CTX
319 #define LARGEST_DIGEST_CTX SHA256_CTX
320 
321 static void
322 tls1_sha512_final_raw(void* ctx, unsigned char *md_out)
323 {
324 	SHA512_CTX *sha512 = ctx;
325 	unsigned i;
326 
327 	for (i = 0; i < 8; i++) {
328 		l2n8(sha512->h[i], md_out);
329 	}
330 }
331 #undef  LARGEST_DIGEST_CTX
332 #define LARGEST_DIGEST_CTX SHA512_CTX
333 
334 /* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
335  * which ssl3_cbc_digest_record supports. */
336 char
337 ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx)
338 {
339 	switch (EVP_MD_CTX_type(ctx)) {
340 	case NID_md5:
341 	case NID_sha1:
342 	case NID_sha224:
343 	case NID_sha256:
344 	case NID_sha384:
345 	case NID_sha512:
346 		return 1;
347 	default:
348 		return 0;
349 	}
350 }
351 
352 /* ssl3_cbc_digest_record computes the MAC of a decrypted, padded SSLv3/TLS
353  * record.
354  *
355  *   ctx: the EVP_MD_CTX from which we take the hash function.
356  *     ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX.
357  *   md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written.
358  *   md_out_size: if non-NULL, the number of output bytes is written here.
359  *   header: the 13-byte, TLS record header.
360  *   data: the record data itself, less any preceeding explicit IV.
361  *   data_plus_mac_size: the secret, reported length of the data and MAC
362  *     once the padding has been removed.
363  *   data_plus_mac_plus_padding_size: the public length of the whole
364  *     record, including padding.
365  *   is_sslv3: non-zero if we are to use SSLv3. Otherwise, TLS.
366  *
367  * On entry: by virtue of having been through one of the remove_padding
368  * functions, above, we know that data_plus_mac_size is large enough to contain
369  * a padding byte and MAC. (If the padding was invalid, it might contain the
370  * padding too. ) */
371 int
372 ssl3_cbc_digest_record(const EVP_MD_CTX *ctx, unsigned char* md_out,
373     size_t* md_out_size, const unsigned char header[13],
374     const unsigned char *data, size_t data_plus_mac_size,
375     size_t data_plus_mac_plus_padding_size, const unsigned char *mac_secret,
376     unsigned mac_secret_length, char is_sslv3)
377 {
378 	union {	double align;
379 		unsigned char c[sizeof(LARGEST_DIGEST_CTX)];
380 	} md_state;
381 	void (*md_final_raw)(void *ctx, unsigned char *md_out);
382 	void (*md_transform)(void *ctx, const unsigned char *block);
383 	unsigned md_size, md_block_size = 64;
384 	unsigned sslv3_pad_length = 40, header_length, variance_blocks,
385 	len, max_mac_bytes, num_blocks,
386 	num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
387 	unsigned int bits;	/* at most 18 bits */
388 	unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
389 	/* hmac_pad is the masked HMAC key. */
390 	unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
391 	unsigned char first_block[MAX_HASH_BLOCK_SIZE];
392 	unsigned char mac_out[EVP_MAX_MD_SIZE];
393 	unsigned i, j, md_out_size_u;
394 	EVP_MD_CTX md_ctx;
395 	/* mdLengthSize is the number of bytes in the length field that terminates
396 	* the hash. */
397 	unsigned md_length_size = 8;
398 	char length_is_big_endian = 1;
399 
400 	/* This is a, hopefully redundant, check that allows us to forget about
401 	 * many possible overflows later in this function. */
402 	OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024);
403 
404 	switch (EVP_MD_CTX_type(ctx)) {
405 	case NID_md5:
406 		MD5_Init((MD5_CTX*)md_state.c);
407 		md_final_raw = tls1_md5_final_raw;
408 		md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform;
409 		md_size = 16;
410 		sslv3_pad_length = 48;
411 		length_is_big_endian = 0;
412 		break;
413 	case NID_sha1:
414 		SHA1_Init((SHA_CTX*)md_state.c);
415 		md_final_raw = tls1_sha1_final_raw;
416 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform;
417 		md_size = 20;
418 		break;
419 	case NID_sha224:
420 		SHA224_Init((SHA256_CTX*)md_state.c);
421 		md_final_raw = tls1_sha256_final_raw;
422 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
423 		md_size = 224/8;
424 		break;
425 	case NID_sha256:
426 		SHA256_Init((SHA256_CTX*)md_state.c);
427 		md_final_raw = tls1_sha256_final_raw;
428 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
429 		md_size = 32;
430 		break;
431 	case NID_sha384:
432 		SHA384_Init((SHA512_CTX*)md_state.c);
433 		md_final_raw = tls1_sha512_final_raw;
434 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
435 		md_size = 384/8;
436 		md_block_size = 128;
437 		md_length_size = 16;
438 		break;
439 	case NID_sha512:
440 		SHA512_Init((SHA512_CTX*)md_state.c);
441 		md_final_raw = tls1_sha512_final_raw;
442 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
443 		md_size = 64;
444 		md_block_size = 128;
445 		md_length_size = 16;
446 		break;
447 	default:
448 		/* ssl3_cbc_record_digest_supported should have been
449 		 * called first to check that the hash function is
450 		 * supported. */
451 		OPENSSL_assert(0);
452 		if (md_out_size)
453 			*md_out_size = 0;
454 		return 0;
455 	}
456 
457 	OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
458 	OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
459 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
460 
461 	header_length = 13;
462 	if (is_sslv3) {
463 		header_length = mac_secret_length + sslv3_pad_length +
464 		    8 /* sequence number */ +
465 		    1 /* record type */ +
466 		    2 /* record length */;
467 	}
468 
469 	/* variance_blocks is the number of blocks of the hash that we have to
470 	 * calculate in constant time because they could be altered by the
471 	 * padding value.
472 	 *
473 	 * In SSLv3, the padding must be minimal so the end of the plaintext
474 	 * varies by, at most, 15+20 = 35 bytes. (We conservatively assume that
475 	 * the MAC size varies from 0..20 bytes.) In case the 9 bytes of hash
476 	 * termination (0x80 + 64-bit length) don't fit in the final block, we
477 	 * say that the final two blocks can vary based on the padding.
478 	 *
479 	 * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
480 	 * required to be minimal. Therefore we say that the final six blocks
481 	 * can vary based on the padding.
482 	 *
483 	 * Later in the function, if the message is short and there obviously
484 	 * cannot be this many blocks then variance_blocks can be reduced. */
485 	variance_blocks = is_sslv3 ? 2 : 6;
486 	/* From now on we're dealing with the MAC, which conceptually has 13
487 	 * bytes of `header' before the start of the data (TLS) or 71/75 bytes
488 	 * (SSLv3) */
489 	len = data_plus_mac_plus_padding_size + header_length;
490 	/* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
491 	* |header|, assuming that there's no padding. */
492 	max_mac_bytes = len - md_size - 1;
493 	/* num_blocks is the maximum number of hash blocks. */
494 	num_blocks = (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
495 	/* In order to calculate the MAC in constant time we have to handle
496 	 * the final blocks specially because the padding value could cause the
497 	 * end to appear somewhere in the final |variance_blocks| blocks and we
498 	 * can't leak where. However, |num_starting_blocks| worth of data can
499 	 * be hashed right away because no padding value can affect whether
500 	 * they are plaintext. */
501 	num_starting_blocks = 0;
502 	/* k is the starting byte offset into the conceptual header||data where
503 	 * we start processing. */
504 	k = 0;
505 	/* mac_end_offset is the index just past the end of the data to be
506 	 * MACed. */
507 	mac_end_offset = data_plus_mac_size + header_length - md_size;
508 	/* c is the index of the 0x80 byte in the final hash block that
509 	 * contains application data. */
510 	c = mac_end_offset % md_block_size;
511 	/* index_a is the hash block number that contains the 0x80 terminating
512 	 * value. */
513 	index_a = mac_end_offset / md_block_size;
514 	/* index_b is the hash block number that contains the 64-bit hash
515 	 * length, in bits. */
516 	index_b = (mac_end_offset + md_length_size) / md_block_size;
517 	/* bits is the hash-length in bits. It includes the additional hash
518 	 * block for the masked HMAC key, or whole of |header| in the case of
519 	 * SSLv3. */
520 
521 	/* For SSLv3, if we're going to have any starting blocks then we need
522 	 * at least two because the header is larger than a single block. */
523 	if (num_blocks > variance_blocks + (is_sslv3 ? 1 : 0)) {
524 		num_starting_blocks = num_blocks - variance_blocks;
525 		k = md_block_size*num_starting_blocks;
526 	}
527 
528 	bits = 8*mac_end_offset;
529 	if (!is_sslv3) {
530 		/* Compute the initial HMAC block. For SSLv3, the padding and
531 		 * secret bytes are included in |header| because they take more
532 		 * than a single block. */
533 		bits += 8*md_block_size;
534 		memset(hmac_pad, 0, md_block_size);
535 		OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad));
536 		memcpy(hmac_pad, mac_secret, mac_secret_length);
537 		for (i = 0; i < md_block_size; i++)
538 			hmac_pad[i] ^= 0x36;
539 
540 		md_transform(md_state.c, hmac_pad);
541 	}
542 
543 	if (length_is_big_endian) {
544 		memset(length_bytes, 0, md_length_size - 4);
545 		length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24);
546 		length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16);
547 		length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8);
548 		length_bytes[md_length_size - 1] = (unsigned char)bits;
549 	} else {
550 		memset(length_bytes, 0, md_length_size);
551 		length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24);
552 		length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16);
553 		length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8);
554 		length_bytes[md_length_size - 8] = (unsigned char)bits;
555 	}
556 
557 	if (k > 0) {
558 		if (is_sslv3) {
559 			/* The SSLv3 header is larger than a single block.
560 			 * overhang is the number of bytes beyond a single
561 			 * block that the header consumes: either 7 bytes
562 			 * (SHA1) or 11 bytes (MD5). */
563 			unsigned overhang = header_length - md_block_size;
564 			md_transform(md_state.c, header);
565 			memcpy(first_block, header + md_block_size, overhang);
566 			memcpy(first_block + overhang, data, md_block_size - overhang);
567 			md_transform(md_state.c, first_block);
568 			for (i = 1; i < k/md_block_size - 1; i++)
569 				md_transform(md_state.c, data + md_block_size*i - overhang);
570 		} else {
571 			/* k is a multiple of md_block_size. */
572 			memcpy(first_block, header, 13);
573 			memcpy(first_block + 13, data, md_block_size - 13);
574 			md_transform(md_state.c, first_block);
575 			for (i = 1; i < k/md_block_size; i++)
576 				md_transform(md_state.c, data + md_block_size*i - 13);
577 		}
578 	}
579 
580 	memset(mac_out, 0, sizeof(mac_out));
581 
582 	/* We now process the final hash blocks. For each block, we construct
583 	 * it in constant time. If the |i==index_a| then we'll include the 0x80
584 	 * bytes and zero pad etc. For each block we selectively copy it, in
585 	 * constant time, to |mac_out|. */
586 	for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks; i++) {
587 		unsigned char block[MAX_HASH_BLOCK_SIZE];
588 		unsigned char is_block_a = constant_time_eq_8(i, index_a);
589 		unsigned char is_block_b = constant_time_eq_8(i, index_b);
590 		for (j = 0; j < md_block_size; j++) {
591 			unsigned char b = 0, is_past_c, is_past_cp1;
592 			if (k < header_length)
593 				b = header[k];
594 			else if (k < data_plus_mac_plus_padding_size + header_length)
595 				b = data[k - header_length];
596 			k++;
597 
598 			is_past_c = is_block_a & constant_time_ge(j, c);
599 			is_past_cp1 = is_block_a & constant_time_ge(j, c + 1);
600 			/* If this is the block containing the end of the
601 			 * application data, and we are at the offset for the
602 			 * 0x80 value, then overwrite b with 0x80. */
603 			b = (b&~is_past_c) | (0x80&is_past_c);
604 			/* If this is the block containing the end of the
605 			 * application data and we're past the 0x80 value then
606 			 * just write zero. */
607 			b = b&~is_past_cp1;
608 			/* If this is index_b (the final block), but not
609 			 * index_a (the end of the data), then the 64-bit
610 			 * length didn't fit into index_a and we're having to
611 			 * add an extra block of zeros. */
612 			b &= ~is_block_b | is_block_a;
613 
614 			/* The final bytes of one of the blocks contains the
615 			 * length. */
616 			if (j >= md_block_size - md_length_size) {
617 				/* If this is index_b, write a length byte. */
618 				b = (b&~is_block_b) | (is_block_b&length_bytes[j - (md_block_size - md_length_size)]);
619 			}
620 			block[j] = b;
621 		}
622 
623 		md_transform(md_state.c, block);
624 		md_final_raw(md_state.c, block);
625 		/* If this is index_b, copy the hash value to |mac_out|. */
626 		for (j = 0; j < md_size; j++)
627 			mac_out[j] |= block[j]&is_block_b;
628 	}
629 
630 	EVP_MD_CTX_init(&md_ctx);
631 	if (!EVP_DigestInit_ex(&md_ctx, ctx->digest, NULL /* engine */)) {
632 		EVP_MD_CTX_cleanup(&md_ctx);
633 		return 0;
634 	}
635 	if (is_sslv3) {
636 		/* We repurpose |hmac_pad| to contain the SSLv3 pad2 block. */
637 		memset(hmac_pad, 0x5c, sslv3_pad_length);
638 
639 		EVP_DigestUpdate(&md_ctx, mac_secret, mac_secret_length);
640 		EVP_DigestUpdate(&md_ctx, hmac_pad, sslv3_pad_length);
641 		EVP_DigestUpdate(&md_ctx, mac_out, md_size);
642 	} else {
643 		/* Complete the HMAC in the standard manner. */
644 		for (i = 0; i < md_block_size; i++)
645 			hmac_pad[i] ^= 0x6a;
646 
647 		EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size);
648 		EVP_DigestUpdate(&md_ctx, mac_out, md_size);
649 	}
650 	EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
651 	if (md_out_size)
652 		*md_out_size = md_out_size_u;
653 	EVP_MD_CTX_cleanup(&md_ctx);
654 
655 	return 1;
656 }
657