xref: /openbsd/lib/libssl/s3_cbc.c (revision c9675a23)
1 /* $OpenBSD: s3_cbc.c,v 1.26 2022/11/26 16:08:55 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 <openssl/md5.h>
57 #include <openssl/sha.h>
58 
59 #include "ssl_local.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
constant_time_lt(unsigned int a,unsigned int b)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
constant_time_ge(unsigned int a,unsigned int b)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
constant_time_eq_8(unsigned int a,unsigned int b)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 /* ssl3_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
ssl3_cbc_remove_padding(SSL3_RECORD_INTERNAL * rec,unsigned int eiv_len,unsigned int mac_size)116 ssl3_cbc_remove_padding(SSL3_RECORD_INTERNAL *rec, unsigned int eiv_len,
117     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 	/*
123 	 * These lengths are all public so we can test them in
124 	 * non-constant time.
125 	 */
126 	if (overhead + eiv_len > rec->length)
127 		return 0;
128 
129 	/* We can now safely skip explicit IV, if any. */
130 	rec->data += eiv_len;
131 	rec->input += eiv_len;
132 	rec->length -= eiv_len;
133 
134 	padding_length = rec->data[rec->length - 1];
135 
136 	good = constant_time_ge(rec->length, overhead + padding_length);
137 	/* The padding consists of a length byte at the end of the record and
138 	 * then that many bytes of padding, all with the same value as the
139 	 * length byte. Thus, with the length byte included, there are i+1
140 	 * bytes of padding.
141 	 *
142 	 * We can't check just |padding_length+1| bytes because that leaks
143 	 * decrypted information. Therefore we always have to check the maximum
144 	 * amount of padding possible. (Again, the length of the record is
145 	 * public information so we can use it.) */
146 	to_check = 256; /* maximum amount of padding, inc length byte. */
147 	if (to_check > rec->length)
148 		to_check = rec->length;
149 
150 	for (i = 0; i < to_check; i++) {
151 		unsigned char mask = constant_time_ge(padding_length, i);
152 		unsigned char b = rec->data[rec->length - 1 - i];
153 		/* The final |padding_length+1| bytes should all have the value
154 		 * |padding_length|. Therefore the XOR should be zero. */
155 		good &= ~(mask&(padding_length ^ b));
156 	}
157 
158 	/* If any of the final |padding_length+1| bytes had the wrong value,
159 	 * one or more of the lower eight bits of |good| will be cleared. We
160 	 * AND the bottom 8 bits together and duplicate the result to all the
161 	 * bits. */
162 	good &= good >> 4;
163 	good &= good >> 2;
164 	good &= good >> 1;
165 	good <<= sizeof(good)*8 - 1;
166 	good = DUPLICATE_MSB_TO_ALL(good);
167 
168 	padding_length = good & (padding_length + 1);
169 	rec->length -= padding_length;
170 	rec->padding_length = padding_length;
171 
172 	return (int)((good & 1) | (~good & -1));
173 }
174 
175 /* ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in
176  * constant time (independent of the concrete value of rec->length, which may
177  * vary within a 256-byte window).
178  *
179  * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to
180  * this function.
181  *
182  * On entry:
183  *   rec->orig_len >= md_size
184  *   md_size <= EVP_MAX_MD_SIZE
185  *
186  * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with
187  * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into
188  * a single or pair of cache-lines, then the variable memory accesses don't
189  * actually affect the timing. CPUs with smaller cache-lines [if any] are
190  * not multi-core and are not considered vulnerable to cache-timing attacks.
191  */
192 #define CBC_MAC_ROTATE_IN_PLACE
193 
194 void
ssl3_cbc_copy_mac(unsigned char * out,const SSL3_RECORD_INTERNAL * rec,unsigned int md_size,unsigned int orig_len)195 ssl3_cbc_copy_mac(unsigned char* out, const SSL3_RECORD_INTERNAL *rec,
196     unsigned int md_size, unsigned int orig_len)
197 {
198 #if defined(CBC_MAC_ROTATE_IN_PLACE)
199 	unsigned char rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
200 	unsigned char *rotated_mac;
201 #else
202 	unsigned char rotated_mac[EVP_MAX_MD_SIZE];
203 #endif
204 
205 	/* mac_end is the index of |rec->data| just after the end of the MAC. */
206 	unsigned int mac_end = rec->length;
207 	unsigned int mac_start = mac_end - md_size;
208 	/* scan_start contains the number of bytes that we can ignore because
209 	 * the MAC's position can only vary by 255 bytes. */
210 	unsigned int scan_start = 0;
211 	unsigned int i, j;
212 	unsigned int div_spoiler;
213 	unsigned int rotate_offset;
214 
215 	OPENSSL_assert(orig_len >= md_size);
216 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
217 
218 #if defined(CBC_MAC_ROTATE_IN_PLACE)
219 	rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf)&63);
220 #endif
221 
222 	/* This information is public so it's safe to branch based on it. */
223 	if (orig_len > md_size + 255 + 1)
224 		scan_start = orig_len - (md_size + 255 + 1);
225 	/* div_spoiler contains a multiple of md_size that is used to cause the
226 	 * modulo operation to be constant time. Without this, the time varies
227 	 * based on the amount of padding when running on Intel chips at least.
228 	 *
229 	 * The aim of right-shifting md_size is so that the compiler doesn't
230 	 * figure out that it can remove div_spoiler as that would require it
231 	 * to prove that md_size is always even, which I hope is beyond it. */
232 	div_spoiler = md_size >> 1;
233 	div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
234 	rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
235 
236 	memset(rotated_mac, 0, md_size);
237 	for (i = scan_start, j = 0; i < orig_len; i++) {
238 		unsigned char mac_started = constant_time_ge(i, mac_start);
239 		unsigned char mac_ended = constant_time_ge(i, mac_end);
240 		unsigned char b = rec->data[i];
241 		rotated_mac[j++] |= b & mac_started & ~mac_ended;
242 		j &= constant_time_lt(j, md_size);
243 	}
244 
245 	/* Now rotate the MAC */
246 #if defined(CBC_MAC_ROTATE_IN_PLACE)
247 	j = 0;
248 	for (i = 0; i < md_size; i++) {
249 		/* in case cache-line is 32 bytes, touch second line */
250 		((volatile unsigned char *)rotated_mac)[rotate_offset^32];
251 		out[j++] = rotated_mac[rotate_offset++];
252 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
253 	}
254 #else
255 	memset(out, 0, md_size);
256 	rotate_offset = md_size - rotate_offset;
257 	rotate_offset &= constant_time_lt(rotate_offset, md_size);
258 	for (i = 0; i < md_size; i++) {
259 		for (j = 0; j < md_size; j++)
260 			out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
261 		rotate_offset++;
262 		rotate_offset &= constant_time_lt(rotate_offset, md_size);
263 	}
264 #endif
265 }
266 
267 #define l2n(l,c)	(*((c)++)=(unsigned char)(((l)>>24)&0xff), \
268 			 *((c)++)=(unsigned char)(((l)>>16)&0xff), \
269 			 *((c)++)=(unsigned char)(((l)>> 8)&0xff), \
270 			 *((c)++)=(unsigned char)(((l)    )&0xff))
271 
272 #define l2n8(l,c)	(*((c)++)=(unsigned char)(((l)>>56)&0xff), \
273 			 *((c)++)=(unsigned char)(((l)>>48)&0xff), \
274 			 *((c)++)=(unsigned char)(((l)>>40)&0xff), \
275 			 *((c)++)=(unsigned char)(((l)>>32)&0xff), \
276 			 *((c)++)=(unsigned char)(((l)>>24)&0xff), \
277 			 *((c)++)=(unsigned char)(((l)>>16)&0xff), \
278 			 *((c)++)=(unsigned char)(((l)>> 8)&0xff), \
279 			 *((c)++)=(unsigned char)(((l)    )&0xff))
280 
281 /* u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
282  * little-endian order. The value of p is advanced by four. */
283 #define u32toLE(n, p) \
284 	(*((p)++)=(unsigned char)(n), \
285 	 *((p)++)=(unsigned char)(n>>8), \
286 	 *((p)++)=(unsigned char)(n>>16), \
287 	 *((p)++)=(unsigned char)(n>>24))
288 
289 /* These functions serialize the state of a hash and thus perform the standard
290  * "final" operation without adding the padding and length that such a function
291  * typically does. */
292 static void
tls1_md5_final_raw(void * ctx,unsigned char * md_out)293 tls1_md5_final_raw(void* ctx, unsigned char *md_out)
294 {
295 	MD5_CTX *md5 = ctx;
296 	u32toLE(md5->A, md_out);
297 	u32toLE(md5->B, md_out);
298 	u32toLE(md5->C, md_out);
299 	u32toLE(md5->D, md_out);
300 }
301 
302 static void
tls1_sha1_final_raw(void * ctx,unsigned char * md_out)303 tls1_sha1_final_raw(void* ctx, unsigned char *md_out)
304 {
305 	SHA_CTX *sha1 = ctx;
306 	l2n(sha1->h0, md_out);
307 	l2n(sha1->h1, md_out);
308 	l2n(sha1->h2, md_out);
309 	l2n(sha1->h3, md_out);
310 	l2n(sha1->h4, md_out);
311 }
312 
313 static void
tls1_sha256_final_raw(void * ctx,unsigned char * md_out)314 tls1_sha256_final_raw(void* ctx, unsigned char *md_out)
315 {
316 	SHA256_CTX *sha256 = ctx;
317 	unsigned int i;
318 
319 	for (i = 0; i < 8; i++) {
320 		l2n(sha256->h[i], md_out);
321 	}
322 }
323 
324 static void
tls1_sha512_final_raw(void * ctx,unsigned char * md_out)325 tls1_sha512_final_raw(void* ctx, unsigned char *md_out)
326 {
327 	SHA512_CTX *sha512 = ctx;
328 	unsigned int i;
329 
330 	for (i = 0; i < 8; i++) {
331 		l2n8(sha512->h[i], md_out);
332 	}
333 }
334 
335 /* Largest hash context ever used by the functions above. */
336 #define LARGEST_DIGEST_CTX SHA512_CTX
337 
338 /* Type giving the alignment needed by the above */
339 #define LARGEST_DIGEST_CTX_ALIGNMENT SHA_LONG64
340 
341 /* ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
342  * which ssl3_cbc_digest_record supports. */
343 char
ssl3_cbc_record_digest_supported(const EVP_MD_CTX * ctx)344 ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx)
345 {
346 	switch (EVP_MD_CTX_type(ctx)) {
347 	case NID_md5:
348 	case NID_sha1:
349 	case NID_sha224:
350 	case NID_sha256:
351 	case NID_sha384:
352 	case NID_sha512:
353 		return 1;
354 	default:
355 		return 0;
356 	}
357 }
358 
359 /* ssl3_cbc_digest_record computes the MAC of a decrypted, padded TLS
360  * record.
361  *
362  *   ctx: the EVP_MD_CTX from which we take the hash function.
363  *     ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX.
364  *   md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written.
365  *   md_out_size: if non-NULL, the number of output bytes is written here.
366  *   header: the 13-byte, TLS record header.
367  *   data: the record data itself, less any preceeding explicit IV.
368  *   data_plus_mac_size: the secret, reported length of the data and MAC
369  *     once the padding has been removed.
370  *   data_plus_mac_plus_padding_size: the public length of the whole
371  *     record, including padding.
372  *
373  * On entry: by virtue of having been through one of the remove_padding
374  * functions, above, we know that data_plus_mac_size is large enough to contain
375  * a padding byte and MAC. (If the padding was invalid, it might contain the
376  * padding too. )
377  */
378 int
ssl3_cbc_digest_record(const EVP_MD_CTX * ctx,unsigned char * md_out,size_t * md_out_size,const unsigned char header[13],const unsigned char * data,size_t data_plus_mac_size,size_t data_plus_mac_plus_padding_size,const unsigned char * mac_secret,unsigned int mac_secret_length)379 ssl3_cbc_digest_record(const EVP_MD_CTX *ctx, unsigned char* md_out,
380     size_t* md_out_size, const unsigned char header[13],
381     const unsigned char *data, size_t data_plus_mac_size,
382     size_t data_plus_mac_plus_padding_size, const unsigned char *mac_secret,
383     unsigned int mac_secret_length)
384 {
385 	union {
386 		/*
387 		 * Alignment here is to allow this to be cast as SHA512_CTX
388 		 * without losing alignment required by the 64-bit SHA_LONG64
389 		 * integer it contains.
390 		 */
391 		LARGEST_DIGEST_CTX_ALIGNMENT align;
392 		unsigned char c[sizeof(LARGEST_DIGEST_CTX)];
393 	} md_state;
394 	void (*md_final_raw)(void *ctx, unsigned char *md_out);
395 	void (*md_transform)(void *ctx, const unsigned char *block);
396 	unsigned int md_size, md_block_size = 64;
397 	unsigned int header_length, variance_blocks,
398 	len, max_mac_bytes, num_blocks,
399 	num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
400 	unsigned int bits;	/* at most 18 bits */
401 	unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
402 	/* hmac_pad is the masked HMAC key. */
403 	unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
404 	unsigned char first_block[MAX_HASH_BLOCK_SIZE];
405 	unsigned char mac_out[EVP_MAX_MD_SIZE];
406 	unsigned int i, j, md_out_size_u;
407 	EVP_MD_CTX *md_ctx;
408 	/* mdLengthSize is the number of bytes in the length field that terminates
409 	* the hash. */
410 	unsigned int md_length_size = 8;
411 	char length_is_big_endian = 1;
412 
413 	/* This is a, hopefully redundant, check that allows us to forget about
414 	 * many possible overflows later in this function. */
415 	OPENSSL_assert(data_plus_mac_plus_padding_size < 1024*1024);
416 
417 	switch (EVP_MD_CTX_type(ctx)) {
418 	case NID_md5:
419 		MD5_Init((MD5_CTX*)md_state.c);
420 		md_final_raw = tls1_md5_final_raw;
421 		md_transform = (void(*)(void *ctx, const unsigned char *block)) MD5_Transform;
422 		md_size = 16;
423 		length_is_big_endian = 0;
424 		break;
425 	case NID_sha1:
426 		SHA1_Init((SHA_CTX*)md_state.c);
427 		md_final_raw = tls1_sha1_final_raw;
428 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA1_Transform;
429 		md_size = 20;
430 		break;
431 	case NID_sha224:
432 		SHA224_Init((SHA256_CTX*)md_state.c);
433 		md_final_raw = tls1_sha256_final_raw;
434 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
435 		md_size = 224/8;
436 		break;
437 	case NID_sha256:
438 		SHA256_Init((SHA256_CTX*)md_state.c);
439 		md_final_raw = tls1_sha256_final_raw;
440 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA256_Transform;
441 		md_size = 32;
442 		break;
443 	case NID_sha384:
444 		SHA384_Init((SHA512_CTX*)md_state.c);
445 		md_final_raw = tls1_sha512_final_raw;
446 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
447 		md_size = 384/8;
448 		md_block_size = 128;
449 		md_length_size = 16;
450 		break;
451 	case NID_sha512:
452 		SHA512_Init((SHA512_CTX*)md_state.c);
453 		md_final_raw = tls1_sha512_final_raw;
454 		md_transform = (void(*)(void *ctx, const unsigned char *block)) SHA512_Transform;
455 		md_size = 64;
456 		md_block_size = 128;
457 		md_length_size = 16;
458 		break;
459 	default:
460 		/* ssl3_cbc_record_digest_supported should have been
461 		 * called first to check that the hash function is
462 		 * supported. */
463 		OPENSSL_assert(0);
464 		if (md_out_size)
465 			*md_out_size = 0;
466 		return 0;
467 	}
468 
469 	OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
470 	OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
471 	OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
472 
473 	header_length = 13;
474 
475 	/* variance_blocks is the number of blocks of the hash that we have to
476 	 * calculate in constant time because they could be altered by the
477 	 * padding value.
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 = 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) */
488 	len = data_plus_mac_plus_padding_size + header_length;
489 	/* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
490 	* |header|, assuming that there's no padding. */
491 	max_mac_bytes = len - md_size - 1;
492 	/* num_blocks is the maximum number of hash blocks. */
493 	num_blocks = (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
494 	/* In order to calculate the MAC in constant time we have to handle
495 	 * the final blocks specially because the padding value could cause the
496 	 * end to appear somewhere in the final |variance_blocks| blocks and we
497 	 * can't leak where. However, |num_starting_blocks| worth of data can
498 	 * be hashed right away because no padding value can affect whether
499 	 * they are plaintext. */
500 	num_starting_blocks = 0;
501 	/* k is the starting byte offset into the conceptual header||data where
502 	 * we start processing. */
503 	k = 0;
504 	/* mac_end_offset is the index just past the end of the data to be
505 	 * MACed. */
506 	mac_end_offset = data_plus_mac_size + header_length - md_size;
507 	/* c is the index of the 0x80 byte in the final hash block that
508 	 * contains application data. */
509 	c = mac_end_offset % md_block_size;
510 	/* index_a is the hash block number that contains the 0x80 terminating
511 	 * value. */
512 	index_a = mac_end_offset / md_block_size;
513 	/* index_b is the hash block number that contains the 64-bit hash
514 	 * length, in bits. */
515 	index_b = (mac_end_offset + md_length_size) / md_block_size;
516 	/* bits is the hash-length in bits. It includes the additional hash
517 	 * block for the masked HMAC key. */
518 
519 	if (num_blocks > variance_blocks) {
520 		num_starting_blocks = num_blocks - variance_blocks;
521 		k = md_block_size*num_starting_blocks;
522 	}
523 
524 	bits = 8*mac_end_offset;
525 	/* Compute the initial HMAC block. */
526 	bits += 8*md_block_size;
527 	memset(hmac_pad, 0, md_block_size);
528 	OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad));
529 	memcpy(hmac_pad, mac_secret, mac_secret_length);
530 	for (i = 0; i < md_block_size; i++)
531 		hmac_pad[i] ^= 0x36;
532 
533 	md_transform(md_state.c, hmac_pad);
534 
535 	if (length_is_big_endian) {
536 		memset(length_bytes, 0, md_length_size - 4);
537 		length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24);
538 		length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16);
539 		length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8);
540 		length_bytes[md_length_size - 1] = (unsigned char)bits;
541 	} else {
542 		memset(length_bytes, 0, md_length_size);
543 		length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24);
544 		length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16);
545 		length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8);
546 		length_bytes[md_length_size - 8] = (unsigned char)bits;
547 	}
548 
549 	if (k > 0) {
550 		/* k is a multiple of md_block_size. */
551 		memcpy(first_block, header, 13);
552 		memcpy(first_block + 13, data, md_block_size - 13);
553 		md_transform(md_state.c, first_block);
554 		for (i = 1; i < k/md_block_size; i++)
555 			md_transform(md_state.c, data + md_block_size*i - 13);
556 	}
557 
558 	memset(mac_out, 0, sizeof(mac_out));
559 
560 	/* We now process the final hash blocks. For each block, we construct
561 	 * it in constant time. If the |i==index_a| then we'll include the 0x80
562 	 * bytes and zero pad etc. For each block we selectively copy it, in
563 	 * constant time, to |mac_out|. */
564 	for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks; i++) {
565 		unsigned char block[MAX_HASH_BLOCK_SIZE];
566 		unsigned char is_block_a = constant_time_eq_8(i, index_a);
567 		unsigned char is_block_b = constant_time_eq_8(i, index_b);
568 		for (j = 0; j < md_block_size; j++) {
569 			unsigned char b = 0, is_past_c, is_past_cp1;
570 			if (k < header_length)
571 				b = header[k];
572 			else if (k < data_plus_mac_plus_padding_size + header_length)
573 				b = data[k - header_length];
574 			k++;
575 
576 			is_past_c = is_block_a & constant_time_ge(j, c);
577 			is_past_cp1 = is_block_a & constant_time_ge(j, c + 1);
578 			/* If this is the block containing the end of the
579 			 * application data, and we are at the offset for the
580 			 * 0x80 value, then overwrite b with 0x80. */
581 			b = (b&~is_past_c) | (0x80&is_past_c);
582 			/* If this is the block containing the end of the
583 			 * application data and we're past the 0x80 value then
584 			 * just write zero. */
585 			b = b&~is_past_cp1;
586 			/* If this is index_b (the final block), but not
587 			 * index_a (the end of the data), then the 64-bit
588 			 * length didn't fit into index_a and we're having to
589 			 * add an extra block of zeros. */
590 			b &= ~is_block_b | is_block_a;
591 
592 			/* The final bytes of one of the blocks contains the
593 			 * length. */
594 			if (j >= md_block_size - md_length_size) {
595 				/* If this is index_b, write a length byte. */
596 				b = (b&~is_block_b) | (is_block_b&length_bytes[j - (md_block_size - md_length_size)]);
597 			}
598 			block[j] = b;
599 		}
600 
601 		md_transform(md_state.c, block);
602 		md_final_raw(md_state.c, block);
603 		/* If this is index_b, copy the hash value to |mac_out|. */
604 		for (j = 0; j < md_size; j++)
605 			mac_out[j] |= block[j]&is_block_b;
606 	}
607 
608 	if ((md_ctx = EVP_MD_CTX_new()) == NULL)
609 		return 0;
610 	if (!EVP_DigestInit_ex(md_ctx, EVP_MD_CTX_md(ctx), NULL /* engine */)) {
611 		EVP_MD_CTX_free(md_ctx);
612 		return 0;
613 	}
614 
615 	/* Complete the HMAC in the standard manner. */
616 	for (i = 0; i < md_block_size; i++)
617 		hmac_pad[i] ^= 0x6a;
618 
619 	EVP_DigestUpdate(md_ctx, hmac_pad, md_block_size);
620 	EVP_DigestUpdate(md_ctx, mac_out, md_size);
621 
622 	EVP_DigestFinal(md_ctx, md_out, &md_out_size_u);
623 	if (md_out_size)
624 		*md_out_size = md_out_size_u;
625 	EVP_MD_CTX_free(md_ctx);
626 
627 	return 1;
628 }
629