xref: /freebsd/crypto/openssl/ssl/s3_cbc.c (revision 0b3105a37d7adcadcb720112fed4dc4e8040be99)
1 /* ssl/s3_cbc.c */
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 "../crypto/constant_time_locl.h"
57 #include "ssl_locl.h"
58 
59 #include <openssl/md5.h>
60 #include <openssl/sha.h>
61 
62 /*
63  * MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's
64  * length field. (SHA-384/512 have 128-bit length.)
65  */
66 #define MAX_HASH_BIT_COUNT_BYTES 16
67 
68 /*
69  * MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support.
70  * Currently SHA-384/512 has a 128-byte block size and that's the largest
71  * supported by TLS.)
72  */
73 #define MAX_HASH_BLOCK_SIZE 128
74 
75 /*-
76  * ssl3_cbc_remove_padding removes padding from the decrypted, SSLv3, CBC
77  * record in |rec| by updating |rec->length| in constant time.
78  *
79  * block_size: the block size of the cipher used to encrypt the record.
80  * returns:
81  *   0: (in non-constant time) if the record is publicly invalid.
82  *   1: if the padding was valid
83  *  -1: otherwise.
84  */
85 int ssl3_cbc_remove_padding(const SSL *s,
86                             SSL3_RECORD *rec,
87                             unsigned block_size, unsigned mac_size)
88 {
89     unsigned padding_length, good;
90     const unsigned overhead = 1 /* padding length byte */  + mac_size;
91 
92     /*
93      * These lengths are all public so we can test them in non-constant time.
94      */
95     if (overhead > rec->length)
96         return 0;
97 
98     padding_length = rec->data[rec->length - 1];
99     good = constant_time_ge(rec->length, padding_length + overhead);
100     /* SSLv3 requires that the padding is minimal. */
101     good &= constant_time_ge(block_size, padding_length + 1);
102     padding_length = good & (padding_length + 1);
103     rec->length -= padding_length;
104     rec->type |= padding_length << 8; /* kludge: pass padding length */
105     return constant_time_select_int(good, 1, -1);
106 }
107 
108 /*-
109  * tls1_cbc_remove_padding removes the CBC padding from the decrypted, TLS, CBC
110  * record in |rec| in constant time and returns 1 if the padding is valid and
111  * -1 otherwise. It also removes any explicit IV from the start of the record
112  * without leaking any timing about whether there was enough space after the
113  * padding was removed.
114  *
115  * block_size: the block size of the cipher used to encrypt the record.
116  * returns:
117  *   0: (in non-constant time) if the record is publicly invalid.
118  *   1: if the padding was valid
119  *  -1: otherwise.
120  */
121 int tls1_cbc_remove_padding(const SSL *s,
122                             SSL3_RECORD *rec,
123                             unsigned block_size, unsigned mac_size)
124 {
125     unsigned padding_length, good, to_check, i;
126     const unsigned overhead = 1 /* padding length byte */  + mac_size;
127     /* Check if version requires explicit IV */
128     if (SSL_USE_EXPLICIT_IV(s)) {
129         /*
130          * These lengths are all public so we can test them in non-constant
131          * time.
132          */
133         if (overhead + block_size > rec->length)
134             return 0;
135         /* We can now safely skip explicit IV */
136         rec->data += block_size;
137         rec->input += block_size;
138         rec->length -= block_size;
139     } else if (overhead > rec->length)
140         return 0;
141 
142     padding_length = rec->data[rec->length - 1];
143 
144     /*
145      * NB: if compression is in operation the first packet may not be of even
146      * length so the padding bug check cannot be performed. This bug
147      * workaround has been around since SSLeay so hopefully it is either
148      * fixed now or no buggy implementation supports compression [steve]
149      */
150     if ((s->options & SSL_OP_TLS_BLOCK_PADDING_BUG) && !s->expand) {
151         /* First packet is even in size, so check */
152         if ((CRYPTO_memcmp(s->s3->read_sequence, "\0\0\0\0\0\0\0\0", 8) == 0) &&
153             !(padding_length & 1)) {
154             s->s3->flags |= TLS1_FLAGS_TLS_PADDING_BUG;
155         }
156         if ((s->s3->flags & TLS1_FLAGS_TLS_PADDING_BUG) && padding_length > 0) {
157             padding_length--;
158         }
159     }
160 
161     if (EVP_CIPHER_flags(s->enc_read_ctx->cipher) & EVP_CIPH_FLAG_AEAD_CIPHER) {
162         /* padding is already verified */
163         rec->length -= padding_length + 1;
164         return 1;
165     }
166 
167     good = constant_time_ge(rec->length, overhead + padding_length);
168     /*
169      * The padding consists of a length byte at the end of the record and
170      * then that many bytes of padding, all with the same value as the length
171      * byte. Thus, with the length byte included, there are i+1 bytes of
172      * padding. We can't check just |padding_length+1| bytes because that
173      * leaks decrypted information. Therefore we always have to check the
174      * maximum amount of padding possible. (Again, the length of the record
175      * is public information so we can use it.)
176      */
177     to_check = 255;             /* maximum amount of padding. */
178     if (to_check > rec->length - 1)
179         to_check = rec->length - 1;
180 
181     for (i = 0; i < to_check; i++) {
182         unsigned char mask = constant_time_ge_8(padding_length, i);
183         unsigned char b = rec->data[rec->length - 1 - i];
184         /*
185          * The final |padding_length+1| bytes should all have the value
186          * |padding_length|. Therefore the XOR should be zero.
187          */
188         good &= ~(mask & (padding_length ^ b));
189     }
190 
191     /*
192      * If any of the final |padding_length+1| bytes had the wrong value, one
193      * or more of the lower eight bits of |good| will be cleared.
194      */
195     good = constant_time_eq(0xff, good & 0xff);
196     padding_length = good & (padding_length + 1);
197     rec->length -= padding_length;
198     rec->type |= padding_length << 8; /* kludge: pass padding length */
199 
200     return constant_time_select_int(good, 1, -1);
201 }
202 
203 /*-
204  * ssl3_cbc_copy_mac copies |md_size| bytes from the end of |rec| to |out| in
205  * constant time (independent of the concrete value of rec->length, which may
206  * vary within a 256-byte window).
207  *
208  * ssl3_cbc_remove_padding or tls1_cbc_remove_padding must be called prior to
209  * this function.
210  *
211  * On entry:
212  *   rec->orig_len >= md_size
213  *   md_size <= EVP_MAX_MD_SIZE
214  *
215  * If CBC_MAC_ROTATE_IN_PLACE is defined then the rotation is performed with
216  * variable accesses in a 64-byte-aligned buffer. Assuming that this fits into
217  * a single or pair of cache-lines, then the variable memory accesses don't
218  * actually affect the timing. CPUs with smaller cache-lines [if any] are
219  * not multi-core and are not considered vulnerable to cache-timing attacks.
220  */
221 #define CBC_MAC_ROTATE_IN_PLACE
222 
223 void ssl3_cbc_copy_mac(unsigned char *out,
224                        const SSL3_RECORD *rec,
225                        unsigned md_size, unsigned orig_len)
226 {
227 #if defined(CBC_MAC_ROTATE_IN_PLACE)
228     unsigned char rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
229     unsigned char *rotated_mac;
230 #else
231     unsigned char rotated_mac[EVP_MAX_MD_SIZE];
232 #endif
233 
234     /*
235      * mac_end is the index of |rec->data| just after the end of the MAC.
236      */
237     unsigned mac_end = rec->length;
238     unsigned mac_start = mac_end - md_size;
239     /*
240      * scan_start contains the number of bytes that we can ignore because the
241      * MAC's position can only vary by 255 bytes.
242      */
243     unsigned scan_start = 0;
244     unsigned i, j;
245     unsigned div_spoiler;
246     unsigned rotate_offset;
247 
248     OPENSSL_assert(orig_len >= md_size);
249     OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
250 
251 #if defined(CBC_MAC_ROTATE_IN_PLACE)
252     rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf) & 63);
253 #endif
254 
255     /* This information is public so it's safe to branch based on it. */
256     if (orig_len > md_size + 255 + 1)
257         scan_start = orig_len - (md_size + 255 + 1);
258     /*
259      * div_spoiler contains a multiple of md_size that is used to cause the
260      * modulo operation to be constant time. Without this, the time varies
261      * based on the amount of padding when running on Intel chips at least.
262      * The aim of right-shifting md_size is so that the compiler doesn't
263      * figure out that it can remove div_spoiler as that would require it to
264      * prove that md_size is always even, which I hope is beyond it.
265      */
266     div_spoiler = md_size >> 1;
267     div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
268     rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
269 
270     memset(rotated_mac, 0, md_size);
271     for (i = scan_start, j = 0; i < orig_len; i++) {
272         unsigned char mac_started = constant_time_ge_8(i, mac_start);
273         unsigned char mac_ended = constant_time_ge_8(i, mac_end);
274         unsigned char b = rec->data[i];
275         rotated_mac[j++] |= b & mac_started & ~mac_ended;
276         j &= constant_time_lt(j, md_size);
277     }
278 
279     /* Now rotate the MAC */
280 #if defined(CBC_MAC_ROTATE_IN_PLACE)
281     j = 0;
282     for (i = 0; i < md_size; i++) {
283         /* in case cache-line is 32 bytes, touch second line */
284         ((volatile unsigned char *)rotated_mac)[rotate_offset ^ 32];
285         out[j++] = rotated_mac[rotate_offset++];
286         rotate_offset &= constant_time_lt(rotate_offset, md_size);
287     }
288 #else
289     memset(out, 0, md_size);
290     rotate_offset = md_size - rotate_offset;
291     rotate_offset &= constant_time_lt(rotate_offset, md_size);
292     for (i = 0; i < md_size; i++) {
293         for (j = 0; j < md_size; j++)
294             out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
295         rotate_offset++;
296         rotate_offset &= constant_time_lt(rotate_offset, md_size);
297     }
298 #endif
299 }
300 
301 /*
302  * u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
303  * little-endian order. The value of p is advanced by four.
304  */
305 #define u32toLE(n, p) \
306         (*((p)++)=(unsigned char)(n), \
307          *((p)++)=(unsigned char)(n>>8), \
308          *((p)++)=(unsigned char)(n>>16), \
309          *((p)++)=(unsigned char)(n>>24))
310 
311 /*
312  * These functions serialize the state of a hash and thus perform the
313  * standard "final" operation without adding the padding and length that such
314  * a function typically does.
315  */
316 static void tls1_md5_final_raw(void *ctx, unsigned char *md_out)
317 {
318     MD5_CTX *md5 = ctx;
319     u32toLE(md5->A, md_out);
320     u32toLE(md5->B, md_out);
321     u32toLE(md5->C, md_out);
322     u32toLE(md5->D, md_out);
323 }
324 
325 static void tls1_sha1_final_raw(void *ctx, unsigned char *md_out)
326 {
327     SHA_CTX *sha1 = ctx;
328     l2n(sha1->h0, md_out);
329     l2n(sha1->h1, md_out);
330     l2n(sha1->h2, md_out);
331     l2n(sha1->h3, md_out);
332     l2n(sha1->h4, md_out);
333 }
334 
335 #define LARGEST_DIGEST_CTX SHA_CTX
336 
337 #ifndef OPENSSL_NO_SHA256
338 static void tls1_sha256_final_raw(void *ctx, unsigned char *md_out)
339 {
340     SHA256_CTX *sha256 = ctx;
341     unsigned i;
342 
343     for (i = 0; i < 8; i++) {
344         l2n(sha256->h[i], md_out);
345     }
346 }
347 
348 # undef  LARGEST_DIGEST_CTX
349 # define LARGEST_DIGEST_CTX SHA256_CTX
350 #endif
351 
352 #ifndef OPENSSL_NO_SHA512
353 static void tls1_sha512_final_raw(void *ctx, unsigned char *md_out)
354 {
355     SHA512_CTX *sha512 = ctx;
356     unsigned i;
357 
358     for (i = 0; i < 8; i++) {
359         l2n8(sha512->h[i], md_out);
360     }
361 }
362 
363 # undef  LARGEST_DIGEST_CTX
364 # define LARGEST_DIGEST_CTX SHA512_CTX
365 #endif
366 
367 /*
368  * ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
369  * which ssl3_cbc_digest_record supports.
370  */
371 char ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx)
372 {
373 #ifdef OPENSSL_FIPS
374     if (FIPS_mode())
375         return 0;
376 #endif
377     switch (EVP_MD_CTX_type(ctx)) {
378     case NID_md5:
379     case NID_sha1:
380 #ifndef OPENSSL_NO_SHA256
381     case NID_sha224:
382     case NID_sha256:
383 #endif
384 #ifndef OPENSSL_NO_SHA512
385     case NID_sha384:
386     case NID_sha512:
387 #endif
388         return 1;
389     default:
390         return 0;
391     }
392 }
393 
394 /*-
395  * ssl3_cbc_digest_record computes the MAC of a decrypted, padded SSLv3/TLS
396  * record.
397  *
398  *   ctx: the EVP_MD_CTX from which we take the hash function.
399  *     ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX.
400  *   md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written.
401  *   md_out_size: if non-NULL, the number of output bytes is written here.
402  *   header: the 13-byte, TLS record header.
403  *   data: the record data itself, less any preceeding explicit IV.
404  *   data_plus_mac_size: the secret, reported length of the data and MAC
405  *     once the padding has been removed.
406  *   data_plus_mac_plus_padding_size: the public length of the whole
407  *     record, including padding.
408  *   is_sslv3: non-zero if we are to use SSLv3. Otherwise, TLS.
409  *
410  * On entry: by virtue of having been through one of the remove_padding
411  * functions, above, we know that data_plus_mac_size is large enough to contain
412  * a padding byte and MAC. (If the padding was invalid, it might contain the
413  * padding too. )
414  * Returns 1 on success or 0 on error
415  */
416 int ssl3_cbc_digest_record(const EVP_MD_CTX *ctx,
417                             unsigned char *md_out,
418                             size_t *md_out_size,
419                             const unsigned char header[13],
420                             const unsigned char *data,
421                             size_t data_plus_mac_size,
422                             size_t data_plus_mac_plus_padding_size,
423                             const unsigned char *mac_secret,
424                             unsigned mac_secret_length, char is_sslv3)
425 {
426     union {
427         double align;
428         unsigned char c[sizeof(LARGEST_DIGEST_CTX)];
429     } md_state;
430     void (*md_final_raw) (void *ctx, unsigned char *md_out);
431     void (*md_transform) (void *ctx, const unsigned char *block);
432     unsigned md_size, md_block_size = 64;
433     unsigned sslv3_pad_length = 40, header_length, variance_blocks,
434         len, max_mac_bytes, num_blocks,
435         num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
436     unsigned int bits;          /* at most 18 bits */
437     unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
438     /* hmac_pad is the masked HMAC key. */
439     unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
440     unsigned char first_block[MAX_HASH_BLOCK_SIZE];
441     unsigned char mac_out[EVP_MAX_MD_SIZE];
442     unsigned i, j, md_out_size_u;
443     EVP_MD_CTX md_ctx;
444     /*
445      * mdLengthSize is the number of bytes in the length field that
446      * terminates * the hash.
447      */
448     unsigned md_length_size = 8;
449     char length_is_big_endian = 1;
450 
451     /*
452      * This is a, hopefully redundant, check that allows us to forget about
453      * many possible overflows later in this function.
454      */
455     OPENSSL_assert(data_plus_mac_plus_padding_size < 1024 * 1024);
456 
457     switch (EVP_MD_CTX_type(ctx)) {
458     case NID_md5:
459         if (MD5_Init((MD5_CTX *)md_state.c) <= 0)
460             return 0;
461         md_final_raw = tls1_md5_final_raw;
462         md_transform =
463             (void (*)(void *ctx, const unsigned char *block))MD5_Transform;
464         md_size = 16;
465         sslv3_pad_length = 48;
466         length_is_big_endian = 0;
467         break;
468     case NID_sha1:
469         if (SHA1_Init((SHA_CTX *)md_state.c) <= 0)
470             return 0;
471         md_final_raw = tls1_sha1_final_raw;
472         md_transform =
473             (void (*)(void *ctx, const unsigned char *block))SHA1_Transform;
474         md_size = 20;
475         break;
476 #ifndef OPENSSL_NO_SHA256
477     case NID_sha224:
478         if (SHA224_Init((SHA256_CTX *)md_state.c) <= 0)
479             return 0;
480         md_final_raw = tls1_sha256_final_raw;
481         md_transform =
482             (void (*)(void *ctx, const unsigned char *block))SHA256_Transform;
483         md_size = 224 / 8;
484         break;
485     case NID_sha256:
486         if (SHA256_Init((SHA256_CTX *)md_state.c) <= 0)
487             return 0;
488         md_final_raw = tls1_sha256_final_raw;
489         md_transform =
490             (void (*)(void *ctx, const unsigned char *block))SHA256_Transform;
491         md_size = 32;
492         break;
493 #endif
494 #ifndef OPENSSL_NO_SHA512
495     case NID_sha384:
496         if (SHA384_Init((SHA512_CTX *)md_state.c) <= 0)
497             return 0;
498         md_final_raw = tls1_sha512_final_raw;
499         md_transform =
500             (void (*)(void *ctx, const unsigned char *block))SHA512_Transform;
501         md_size = 384 / 8;
502         md_block_size = 128;
503         md_length_size = 16;
504         break;
505     case NID_sha512:
506         if (SHA512_Init((SHA512_CTX *)md_state.c) <= 0)
507             return 0;
508         md_final_raw = tls1_sha512_final_raw;
509         md_transform =
510             (void (*)(void *ctx, const unsigned char *block))SHA512_Transform;
511         md_size = 64;
512         md_block_size = 128;
513         md_length_size = 16;
514         break;
515 #endif
516     default:
517         /*
518          * ssl3_cbc_record_digest_supported should have been called first to
519          * check that the hash function is supported.
520          */
521         OPENSSL_assert(0);
522         if (md_out_size)
523             *md_out_size = 0;
524         return 0;
525     }
526 
527     OPENSSL_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
528     OPENSSL_assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
529     OPENSSL_assert(md_size <= EVP_MAX_MD_SIZE);
530 
531     header_length = 13;
532     if (is_sslv3) {
533         header_length = mac_secret_length + sslv3_pad_length + 8 /* sequence
534                                                                   * number */  +
535             1 /* record type */  +
536             2 /* record length */ ;
537     }
538 
539     /*
540      * variance_blocks is the number of blocks of the hash that we have to
541      * calculate in constant time because they could be altered by the
542      * padding value. In SSLv3, the padding must be minimal so the end of
543      * the plaintext varies by, at most, 15+20 = 35 bytes. (We conservatively
544      * assume that the MAC size varies from 0..20 bytes.) In case the 9 bytes
545      * of hash termination (0x80 + 64-bit length) don't fit in the final
546      * block, we say that the final two blocks can vary based on the padding.
547      * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
548      * required to be minimal. Therefore we say that the final six blocks can
549      * vary based on the padding. Later in the function, if the message is
550      * short and there obviously cannot be this many blocks then
551      * variance_blocks can be reduced.
552      */
553     variance_blocks = is_sslv3 ? 2 : 6;
554     /*
555      * From now on we're dealing with the MAC, which conceptually has 13
556      * bytes of `header' before the start of the data (TLS) or 71/75 bytes
557      * (SSLv3)
558      */
559     len = data_plus_mac_plus_padding_size + header_length;
560     /*
561      * max_mac_bytes contains the maximum bytes of bytes in the MAC,
562      * including * |header|, assuming that there's no padding.
563      */
564     max_mac_bytes = len - md_size - 1;
565     /* num_blocks is the maximum number of hash blocks. */
566     num_blocks =
567         (max_mac_bytes + 1 + md_length_size + md_block_size -
568          1) / md_block_size;
569     /*
570      * In order to calculate the MAC in constant time we have to handle the
571      * final blocks specially because the padding value could cause the end
572      * to appear somewhere in the final |variance_blocks| blocks and we can't
573      * leak where. However, |num_starting_blocks| worth of data can be hashed
574      * right away because no padding value can affect whether they are
575      * plaintext.
576      */
577     num_starting_blocks = 0;
578     /*
579      * k is the starting byte offset into the conceptual header||data where
580      * we start processing.
581      */
582     k = 0;
583     /*
584      * mac_end_offset is the index just past the end of the data to be MACed.
585      */
586     mac_end_offset = data_plus_mac_size + header_length - md_size;
587     /*
588      * c is the index of the 0x80 byte in the final hash block that contains
589      * application data.
590      */
591     c = mac_end_offset % md_block_size;
592     /*
593      * index_a is the hash block number that contains the 0x80 terminating
594      * value.
595      */
596     index_a = mac_end_offset / md_block_size;
597     /*
598      * index_b is the hash block number that contains the 64-bit hash length,
599      * in bits.
600      */
601     index_b = (mac_end_offset + md_length_size) / md_block_size;
602     /*
603      * bits is the hash-length in bits. It includes the additional hash block
604      * for the masked HMAC key, or whole of |header| in the case of SSLv3.
605      */
606 
607     /*
608      * For SSLv3, if we're going to have any starting blocks then we need at
609      * least two because the header is larger than a single block.
610      */
611     if (num_blocks > variance_blocks + (is_sslv3 ? 1 : 0)) {
612         num_starting_blocks = num_blocks - variance_blocks;
613         k = md_block_size * num_starting_blocks;
614     }
615 
616     bits = 8 * mac_end_offset;
617     if (!is_sslv3) {
618         /*
619          * Compute the initial HMAC block. For SSLv3, the padding and secret
620          * bytes are included in |header| because they take more than a
621          * single block.
622          */
623         bits += 8 * md_block_size;
624         memset(hmac_pad, 0, md_block_size);
625         OPENSSL_assert(mac_secret_length <= sizeof(hmac_pad));
626         memcpy(hmac_pad, mac_secret, mac_secret_length);
627         for (i = 0; i < md_block_size; i++)
628             hmac_pad[i] ^= 0x36;
629 
630         md_transform(md_state.c, hmac_pad);
631     }
632 
633     if (length_is_big_endian) {
634         memset(length_bytes, 0, md_length_size - 4);
635         length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24);
636         length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16);
637         length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8);
638         length_bytes[md_length_size - 1] = (unsigned char)bits;
639     } else {
640         memset(length_bytes, 0, md_length_size);
641         length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24);
642         length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16);
643         length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8);
644         length_bytes[md_length_size - 8] = (unsigned char)bits;
645     }
646 
647     if (k > 0) {
648         if (is_sslv3) {
649             unsigned overhang;
650 
651             /*
652              * The SSLv3 header is larger than a single block. overhang is
653              * the number of bytes beyond a single block that the header
654              * consumes: either 7 bytes (SHA1) or 11 bytes (MD5). There are no
655              * ciphersuites in SSLv3 that are not SHA1 or MD5 based and
656              * therefore we can be confident that the header_length will be
657              * greater than |md_block_size|. However we add a sanity check just
658              * in case
659              */
660             if (header_length <= md_block_size) {
661                 /* Should never happen */
662                 return 0;
663             }
664             overhang = header_length - md_block_size;
665             md_transform(md_state.c, header);
666             memcpy(first_block, header + md_block_size, overhang);
667             memcpy(first_block + overhang, data, md_block_size - overhang);
668             md_transform(md_state.c, first_block);
669             for (i = 1; i < k / md_block_size - 1; i++)
670                 md_transform(md_state.c, data + md_block_size * i - overhang);
671         } else {
672             /* k is a multiple of md_block_size. */
673             memcpy(first_block, header, 13);
674             memcpy(first_block + 13, data, md_block_size - 13);
675             md_transform(md_state.c, first_block);
676             for (i = 1; i < k / md_block_size; i++)
677                 md_transform(md_state.c, data + md_block_size * i - 13);
678         }
679     }
680 
681     memset(mac_out, 0, sizeof(mac_out));
682 
683     /*
684      * We now process the final hash blocks. For each block, we construct it
685      * in constant time. If the |i==index_a| then we'll include the 0x80
686      * bytes and zero pad etc. For each block we selectively copy it, in
687      * constant time, to |mac_out|.
688      */
689     for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks;
690          i++) {
691         unsigned char block[MAX_HASH_BLOCK_SIZE];
692         unsigned char is_block_a = constant_time_eq_8(i, index_a);
693         unsigned char is_block_b = constant_time_eq_8(i, index_b);
694         for (j = 0; j < md_block_size; j++) {
695             unsigned char b = 0, is_past_c, is_past_cp1;
696             if (k < header_length)
697                 b = header[k];
698             else if (k < data_plus_mac_plus_padding_size + header_length)
699                 b = data[k - header_length];
700             k++;
701 
702             is_past_c = is_block_a & constant_time_ge_8(j, c);
703             is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1);
704             /*
705              * If this is the block containing the end of the application
706              * data, and we are at the offset for the 0x80 value, then
707              * overwrite b with 0x80.
708              */
709             b = constant_time_select_8(is_past_c, 0x80, b);
710             /*
711              * If this the the block containing the end of the application
712              * data and we're past the 0x80 value then just write zero.
713              */
714             b = b & ~is_past_cp1;
715             /*
716              * If this is index_b (the final block), but not index_a (the end
717              * of the data), then the 64-bit length didn't fit into index_a
718              * and we're having to add an extra block of zeros.
719              */
720             b &= ~is_block_b | is_block_a;
721 
722             /*
723              * The final bytes of one of the blocks contains the length.
724              */
725             if (j >= md_block_size - md_length_size) {
726                 /* If this is index_b, write a length byte. */
727                 b = constant_time_select_8(is_block_b,
728                                            length_bytes[j -
729                                                         (md_block_size -
730                                                          md_length_size)], b);
731             }
732             block[j] = b;
733         }
734 
735         md_transform(md_state.c, block);
736         md_final_raw(md_state.c, block);
737         /* If this is index_b, copy the hash value to |mac_out|. */
738         for (j = 0; j < md_size; j++)
739             mac_out[j] |= block[j] & is_block_b;
740     }
741 
742     EVP_MD_CTX_init(&md_ctx);
743     if (EVP_DigestInit_ex(&md_ctx, ctx->digest, NULL /* engine */ ) <= 0)
744         goto err;
745     if (is_sslv3) {
746         /* We repurpose |hmac_pad| to contain the SSLv3 pad2 block. */
747         memset(hmac_pad, 0x5c, sslv3_pad_length);
748 
749         if (EVP_DigestUpdate(&md_ctx, mac_secret, mac_secret_length) <= 0
750                 || EVP_DigestUpdate(&md_ctx, hmac_pad, sslv3_pad_length) <= 0
751                 || EVP_DigestUpdate(&md_ctx, mac_out, md_size) <= 0)
752             goto err;
753     } else {
754         /* Complete the HMAC in the standard manner. */
755         for (i = 0; i < md_block_size; i++)
756             hmac_pad[i] ^= 0x6a;
757 
758         if (EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size) <= 0
759                 || EVP_DigestUpdate(&md_ctx, mac_out, md_size) <= 0)
760             goto err;
761     }
762     EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
763     if (md_out_size)
764         *md_out_size = md_out_size_u;
765     EVP_MD_CTX_cleanup(&md_ctx);
766 
767     return 1;
768 err:
769     EVP_MD_CTX_cleanup(&md_ctx);
770     return 0;
771 }
772 
773 #ifdef OPENSSL_FIPS
774 
775 /*
776  * Due to the need to use EVP in FIPS mode we can't reimplement digests but
777  * we can ensure the number of blocks processed is equal for all cases by
778  * digesting additional data.
779  */
780 
781 void tls_fips_digest_extra(const EVP_CIPHER_CTX *cipher_ctx,
782                            EVP_MD_CTX *mac_ctx, const unsigned char *data,
783                            size_t data_len, size_t orig_len)
784 {
785     size_t block_size, digest_pad, blocks_data, blocks_orig;
786     if (EVP_CIPHER_CTX_mode(cipher_ctx) != EVP_CIPH_CBC_MODE)
787         return;
788     block_size = EVP_MD_CTX_block_size(mac_ctx);
789     /*-
790      * We are in FIPS mode if we get this far so we know we have only SHA*
791      * digests and TLS to deal with.
792      * Minimum digest padding length is 17 for SHA384/SHA512 and 9
793      * otherwise.
794      * Additional header is 13 bytes. To get the number of digest blocks
795      * processed round up the amount of data plus padding to the nearest
796      * block length. Block length is 128 for SHA384/SHA512 and 64 otherwise.
797      * So we have:
798      * blocks = (payload_len + digest_pad + 13 + block_size - 1)/block_size
799      * equivalently:
800      * blocks = (payload_len + digest_pad + 12)/block_size + 1
801      * HMAC adds a constant overhead.
802      * We're ultimately only interested in differences so this becomes
803      * blocks = (payload_len + 29)/128
804      * for SHA384/SHA512 and
805      * blocks = (payload_len + 21)/64
806      * otherwise.
807      */
808     digest_pad = block_size == 64 ? 21 : 29;
809     blocks_orig = (orig_len + digest_pad) / block_size;
810     blocks_data = (data_len + digest_pad) / block_size;
811     /*
812      * MAC enough blocks to make up the difference between the original and
813      * actual lengths plus one extra block to ensure this is never a no op.
814      * The "data" pointer should always have enough space to perform this
815      * operation as it is large enough for a maximum length TLS buffer.
816      */
817     EVP_DigestSignUpdate(mac_ctx, data,
818                          (blocks_orig - blocks_data + 1) * block_size);
819 }
820 #endif
821