1/* ====================================================================
2 * Copyright (c) 2012 The OpenSSL Project.  All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 *
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 *
11 * 2. Redistributions in binary form must reproduce the above copyright
12 *    notice, this list of conditions and the following disclaimer in
13 *    the documentation and/or other materials provided with the
14 *    distribution.
15 *
16 * 3. All advertising materials mentioning features or use of this
17 *    software must display the following acknowledgment:
18 *    "This product includes software developed by the OpenSSL Project
19 *    for use in the OpenSSL Toolkit. (http://www.openssl.org/)"
20 *
21 * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
22 *    endorse or promote products derived from this software without
23 *    prior written permission. For written permission, please contact
24 *    openssl-core@openssl.org.
25 *
26 * 5. Products derived from this software may not be called "OpenSSL"
27 *    nor may "OpenSSL" appear in their names without prior written
28 *    permission of the OpenSSL Project.
29 *
30 * 6. Redistributions of any form whatsoever must retain the following
31 *    acknowledgment:
32 *    "This product includes software developed by the OpenSSL Project
33 *    for use in the OpenSSL Toolkit (http://www.openssl.org/)"
34 *
35 * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
36 * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
37 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
38 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE OpenSSL PROJECT OR
39 * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
41 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
42 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
43 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
44 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
45 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
46 * OF THE POSSIBILITY OF SUCH DAMAGE.
47 * ====================================================================
48 *
49 * This product includes cryptographic software written by Eric Young
50 * (eay@cryptsoft.com).  This product includes software written by Tim
51 * Hudson (tjh@cryptsoft.com). */
52
53#include <assert.h>
54#include <string.h>
55
56#include <openssl/digest.h>
57#include <openssl/obj.h>
58#include <openssl/sha.h>
59
60#include "../internal.h"
61
62
63/* TODO(davidben): unsigned should be size_t. The various constant_time
64 * functions need to be switched to size_t. */
65
66/* MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's length
67 * field. (SHA-384/512 have 128-bit length.) */
68#define MAX_HASH_BIT_COUNT_BYTES 16
69
70/* MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support.
71 * Currently SHA-384/512 has a 128-byte block size and that's the largest
72 * supported by TLS.) */
73#define MAX_HASH_BLOCK_SIZE 128
74
75int EVP_tls_cbc_remove_padding(unsigned *out_len,
76                               const uint8_t *in, unsigned in_len,
77                               unsigned block_size, unsigned mac_size) {
78  unsigned padding_length, good, to_check, i;
79  const unsigned overhead = 1 /* padding length byte */ + mac_size;
80
81  /* These lengths are all public so we can test them in non-constant time. */
82  if (overhead > in_len) {
83    return 0;
84  }
85
86  padding_length = in[in_len - 1];
87
88  good = constant_time_ge(in_len, overhead + padding_length);
89  /* The padding consists of a length byte at the end of the record and
90   * then that many bytes of padding, all with the same value as the
91   * length byte. Thus, with the length byte included, there are i+1
92   * bytes of padding.
93   *
94   * We can't check just |padding_length+1| bytes because that leaks
95   * decrypted information. Therefore we always have to check the maximum
96   * amount of padding possible. (Again, the length of the record is
97   * public information so we can use it.) */
98  to_check = 256; /* maximum amount of padding, inc length byte. */
99  if (to_check > in_len) {
100    to_check = in_len;
101  }
102
103  for (i = 0; i < to_check; i++) {
104    uint8_t mask = constant_time_ge_8(padding_length, i);
105    uint8_t b = in[in_len - 1 - i];
106    /* The final |padding_length+1| bytes should all have the value
107     * |padding_length|. Therefore the XOR should be zero. */
108    good &= ~(mask & (padding_length ^ b));
109  }
110
111  /* If any of the final |padding_length+1| bytes had the wrong value,
112   * one or more of the lower eight bits of |good| will be cleared. */
113  good = constant_time_eq(0xff, good & 0xff);
114
115  /* Always treat |padding_length| as zero on error. If, assuming block size of
116   * 16, a padding of [<15 arbitrary bytes> 15] treated |padding_length| as 16
117   * and returned -1, distinguishing good MAC and bad padding from bad MAC and
118   * bad padding would give POODLE's padding oracle. */
119  padding_length = good & (padding_length + 1);
120  *out_len = in_len - padding_length;
121
122  return constant_time_select_int(good, 1, -1);
123}
124
125/* If CBC_MAC_ROTATE_IN_PLACE is defined then EVP_tls_cbc_copy_mac is performed
126 * with variable accesses in a 64-byte-aligned buffer. Assuming that this fits
127 * into a single or pair of cache-lines, then the variable memory accesses don't
128 * actually affect the timing. CPUs with smaller cache-lines [if any] are not
129 * multi-core and are not considered vulnerable to cache-timing attacks. */
130#define CBC_MAC_ROTATE_IN_PLACE
131
132void EVP_tls_cbc_copy_mac(uint8_t *out, unsigned md_size,
133                          const uint8_t *in, unsigned in_len,
134                          unsigned orig_len) {
135#if defined(CBC_MAC_ROTATE_IN_PLACE)
136  uint8_t rotated_mac_buf[64 + EVP_MAX_MD_SIZE];
137  uint8_t *rotated_mac;
138#else
139  uint8_t rotated_mac[EVP_MAX_MD_SIZE];
140#endif
141
142  /* mac_end is the index of |in| just after the end of the MAC. */
143  unsigned mac_end = in_len;
144  unsigned mac_start = mac_end - md_size;
145  /* scan_start contains the number of bytes that we can ignore because
146   * the MAC's position can only vary by 255 bytes. */
147  unsigned scan_start = 0;
148  unsigned i, j;
149  unsigned div_spoiler;
150  unsigned rotate_offset;
151
152  assert(orig_len >= in_len);
153  assert(in_len >= md_size);
154  assert(md_size <= EVP_MAX_MD_SIZE);
155
156#if defined(CBC_MAC_ROTATE_IN_PLACE)
157  rotated_mac = rotated_mac_buf + ((0 - (size_t)rotated_mac_buf) & 63);
158#endif
159
160  /* This information is public so it's safe to branch based on it. */
161  if (orig_len > md_size + 255 + 1) {
162    scan_start = orig_len - (md_size + 255 + 1);
163  }
164  /* div_spoiler contains a multiple of md_size that is used to cause the
165   * modulo operation to be constant time. Without this, the time varies
166   * based on the amount of padding when running on Intel chips at least.
167   *
168   * The aim of right-shifting md_size is so that the compiler doesn't
169   * figure out that it can remove div_spoiler as that would require it
170   * to prove that md_size is always even, which I hope is beyond it. */
171  div_spoiler = md_size >> 1;
172  div_spoiler <<= (sizeof(div_spoiler) - 1) * 8;
173  rotate_offset = (div_spoiler + mac_start - scan_start) % md_size;
174
175  memset(rotated_mac, 0, md_size);
176  for (i = scan_start, j = 0; i < orig_len; i++) {
177    uint8_t mac_started = constant_time_ge_8(i, mac_start);
178    uint8_t mac_ended = constant_time_ge_8(i, mac_end);
179    uint8_t b = in[i];
180    rotated_mac[j++] |= b & mac_started & ~mac_ended;
181    j &= constant_time_lt(j, md_size);
182  }
183
184/* Now rotate the MAC */
185#if defined(CBC_MAC_ROTATE_IN_PLACE)
186  j = 0;
187  for (i = 0; i < md_size; i++) {
188    /* in case cache-line is 32 bytes, touch second line */
189    ((volatile uint8_t *)rotated_mac)[rotate_offset ^ 32];
190    out[j++] = rotated_mac[rotate_offset++];
191    rotate_offset &= constant_time_lt(rotate_offset, md_size);
192  }
193#else
194  memset(out, 0, md_size);
195  rotate_offset = md_size - rotate_offset;
196  rotate_offset &= constant_time_lt(rotate_offset, md_size);
197  for (i = 0; i < md_size; i++) {
198    for (j = 0; j < md_size; j++) {
199      out[j] |= rotated_mac[i] & constant_time_eq_8(j, rotate_offset);
200    }
201    rotate_offset++;
202    rotate_offset &= constant_time_lt(rotate_offset, md_size);
203  }
204#endif
205}
206
207/* u32toBE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
208 * big-endian order. The value of p is advanced by four. */
209#define u32toBE(n, p) \
210  (*((p)++)=(uint8_t)(n>>24), \
211   *((p)++)=(uint8_t)(n>>16), \
212   *((p)++)=(uint8_t)(n>>8), \
213   *((p)++)=(uint8_t)(n))
214
215/* u64toBE serialises an unsigned, 64-bit number (n) as eight bytes at (p) in
216 * big-endian order. The value of p is advanced by eight. */
217#define u64toBE(n, p) \
218  (*((p)++)=(uint8_t)(n>>56), \
219   *((p)++)=(uint8_t)(n>>48), \
220   *((p)++)=(uint8_t)(n>>40), \
221   *((p)++)=(uint8_t)(n>>32), \
222   *((p)++)=(uint8_t)(n>>24), \
223   *((p)++)=(uint8_t)(n>>16), \
224   *((p)++)=(uint8_t)(n>>8), \
225   *((p)++)=(uint8_t)(n))
226
227/* These functions serialize the state of a hash and thus perform the standard
228 * "final" operation without adding the padding and length that such a function
229 * typically does. */
230static void tls1_sha1_final_raw(void *ctx, uint8_t *md_out) {
231  SHA_CTX *sha1 = ctx;
232  u32toBE(sha1->h0, md_out);
233  u32toBE(sha1->h1, md_out);
234  u32toBE(sha1->h2, md_out);
235  u32toBE(sha1->h3, md_out);
236  u32toBE(sha1->h4, md_out);
237}
238#define LARGEST_DIGEST_CTX SHA_CTX
239
240static void tls1_sha256_final_raw(void *ctx, uint8_t *md_out) {
241  SHA256_CTX *sha256 = ctx;
242  unsigned i;
243
244  for (i = 0; i < 8; i++) {
245    u32toBE(sha256->h[i], md_out);
246  }
247}
248#undef  LARGEST_DIGEST_CTX
249#define LARGEST_DIGEST_CTX SHA256_CTX
250
251static void tls1_sha512_final_raw(void *ctx, uint8_t *md_out) {
252  SHA512_CTX *sha512 = ctx;
253  unsigned i;
254
255  for (i = 0; i < 8; i++) {
256    u64toBE(sha512->h[i], md_out);
257  }
258}
259#undef  LARGEST_DIGEST_CTX
260#define LARGEST_DIGEST_CTX SHA512_CTX
261
262int EVP_tls_cbc_record_digest_supported(const EVP_MD *md) {
263  switch (EVP_MD_type(md)) {
264    case NID_sha1:
265    case NID_sha256:
266    case NID_sha384:
267      return 1;
268
269    default:
270      return 0;
271  }
272}
273
274int EVP_tls_cbc_digest_record(const EVP_MD *md, uint8_t *md_out,
275                              size_t *md_out_size, const uint8_t header[13],
276                              const uint8_t *data, size_t data_plus_mac_size,
277                              size_t data_plus_mac_plus_padding_size,
278                              const uint8_t *mac_secret,
279                              unsigned mac_secret_length) {
280  union {
281    double align;
282    uint8_t c[sizeof(LARGEST_DIGEST_CTX)];
283  } md_state;
284  void (*md_final_raw)(void *ctx, uint8_t *md_out);
285  void (*md_transform)(void *ctx, const uint8_t *block);
286  unsigned md_size, md_block_size = 64;
287  unsigned len, max_mac_bytes, num_blocks, num_starting_blocks, k,
288           mac_end_offset, c, index_a, index_b;
289  unsigned int bits; /* at most 18 bits */
290  uint8_t length_bytes[MAX_HASH_BIT_COUNT_BYTES];
291  /* hmac_pad is the masked HMAC key. */
292  uint8_t hmac_pad[MAX_HASH_BLOCK_SIZE];
293  uint8_t first_block[MAX_HASH_BLOCK_SIZE];
294  uint8_t mac_out[EVP_MAX_MD_SIZE];
295  unsigned i, j, md_out_size_u;
296  EVP_MD_CTX md_ctx;
297  /* mdLengthSize is the number of bytes in the length field that terminates
298  * the hash. */
299  unsigned md_length_size = 8;
300
301  /* This is a, hopefully redundant, check that allows us to forget about
302   * many possible overflows later in this function. */
303  assert(data_plus_mac_plus_padding_size < 1024 * 1024);
304
305  switch (EVP_MD_type(md)) {
306    case NID_sha1:
307      SHA1_Init((SHA_CTX *)md_state.c);
308      md_final_raw = tls1_sha1_final_raw;
309      md_transform =
310          (void (*)(void *ctx, const uint8_t *block))SHA1_Transform;
311      md_size = 20;
312      break;
313
314    case NID_sha256:
315      SHA256_Init((SHA256_CTX *)md_state.c);
316      md_final_raw = tls1_sha256_final_raw;
317      md_transform =
318          (void (*)(void *ctx, const uint8_t *block))SHA256_Transform;
319      md_size = 32;
320      break;
321
322    case NID_sha384:
323      SHA384_Init((SHA512_CTX *)md_state.c);
324      md_final_raw = tls1_sha512_final_raw;
325      md_transform =
326          (void (*)(void *ctx, const uint8_t *block))SHA512_Transform;
327      md_size = 384 / 8;
328      md_block_size = 128;
329      md_length_size = 16;
330      break;
331
332    default:
333      /* EVP_tls_cbc_record_digest_supported should have been called first to
334       * check that the hash function is supported. */
335      assert(0);
336      *md_out_size = 0;
337      return 0;
338  }
339
340  assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES);
341  assert(md_block_size <= MAX_HASH_BLOCK_SIZE);
342  assert(md_size <= EVP_MAX_MD_SIZE);
343
344  static const unsigned kHeaderLength = 13;
345
346  /* kVarianceBlocks is the number of blocks of the hash that we have to
347   * calculate in constant time because they could be altered by the
348   * padding value.
349   *
350   * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
351   * required to be minimal. Therefore we say that the final six blocks
352   * can vary based on the padding. */
353  static const unsigned kVarianceBlocks = 6;
354
355  /* From now on we're dealing with the MAC, which conceptually has 13
356   * bytes of `header' before the start of the data. */
357  len = data_plus_mac_plus_padding_size + kHeaderLength;
358  /* max_mac_bytes contains the maximum bytes of bytes in the MAC, including
359  * |header|, assuming that there's no padding. */
360  max_mac_bytes = len - md_size - 1;
361  /* num_blocks is the maximum number of hash blocks. */
362  num_blocks =
363      (max_mac_bytes + 1 + md_length_size + md_block_size - 1) / md_block_size;
364  /* In order to calculate the MAC in constant time we have to handle
365   * the final blocks specially because the padding value could cause the
366   * end to appear somewhere in the final |kVarianceBlocks| blocks and we
367   * can't leak where. However, |num_starting_blocks| worth of data can
368   * be hashed right away because no padding value can affect whether
369   * they are plaintext. */
370  num_starting_blocks = 0;
371  /* k is the starting byte offset into the conceptual header||data where
372   * we start processing. */
373  k = 0;
374  /* mac_end_offset is the index just past the end of the data to be
375   * MACed. */
376  mac_end_offset = data_plus_mac_size + kHeaderLength - md_size;
377  /* c is the index of the 0x80 byte in the final hash block that
378   * contains application data. */
379  c = mac_end_offset % md_block_size;
380  /* index_a is the hash block number that contains the 0x80 terminating
381   * value. */
382  index_a = mac_end_offset / md_block_size;
383  /* index_b is the hash block number that contains the 64-bit hash
384   * length, in bits. */
385  index_b = (mac_end_offset + md_length_size) / md_block_size;
386  /* bits is the hash-length in bits. It includes the additional hash
387   * block for the masked HMAC key. */
388
389  if (num_blocks > kVarianceBlocks) {
390    num_starting_blocks = num_blocks - kVarianceBlocks;
391    k = md_block_size * num_starting_blocks;
392  }
393
394  bits = 8 * mac_end_offset;
395
396  /* Compute the initial HMAC block. */
397  bits += 8 * md_block_size;
398  memset(hmac_pad, 0, md_block_size);
399  assert(mac_secret_length <= sizeof(hmac_pad));
400  memcpy(hmac_pad, mac_secret, mac_secret_length);
401  for (i = 0; i < md_block_size; i++) {
402    hmac_pad[i] ^= 0x36;
403  }
404
405  md_transform(md_state.c, hmac_pad);
406
407  memset(length_bytes, 0, md_length_size - 4);
408  length_bytes[md_length_size - 4] = (uint8_t)(bits >> 24);
409  length_bytes[md_length_size - 3] = (uint8_t)(bits >> 16);
410  length_bytes[md_length_size - 2] = (uint8_t)(bits >> 8);
411  length_bytes[md_length_size - 1] = (uint8_t)bits;
412
413  if (k > 0) {
414    /* k is a multiple of md_block_size. */
415    memcpy(first_block, header, 13);
416    memcpy(first_block + 13, data, md_block_size - 13);
417    md_transform(md_state.c, first_block);
418    for (i = 1; i < k / md_block_size; i++) {
419      md_transform(md_state.c, data + md_block_size * i - 13);
420    }
421  }
422
423  memset(mac_out, 0, sizeof(mac_out));
424
425  /* We now process the final hash blocks. For each block, we construct
426   * it in constant time. If the |i==index_a| then we'll include the 0x80
427   * bytes and zero pad etc. For each block we selectively copy it, in
428   * constant time, to |mac_out|. */
429  for (i = num_starting_blocks; i <= num_starting_blocks + kVarianceBlocks;
430       i++) {
431    uint8_t block[MAX_HASH_BLOCK_SIZE];
432    uint8_t is_block_a = constant_time_eq_8(i, index_a);
433    uint8_t is_block_b = constant_time_eq_8(i, index_b);
434    for (j = 0; j < md_block_size; j++) {
435      uint8_t b = 0, is_past_c, is_past_cp1;
436      if (k < kHeaderLength) {
437        b = header[k];
438      } else if (k < data_plus_mac_plus_padding_size + kHeaderLength) {
439        b = data[k - kHeaderLength];
440      }
441      k++;
442
443      is_past_c = is_block_a & constant_time_ge_8(j, c);
444      is_past_cp1 = is_block_a & constant_time_ge_8(j, c + 1);
445      /* If this is the block containing the end of the
446       * application data, and we are at the offset for the
447       * 0x80 value, then overwrite b with 0x80. */
448      b = constant_time_select_8(is_past_c, 0x80, b);
449      /* If this the the block containing the end of the
450       * application data and we're past the 0x80 value then
451       * just write zero. */
452      b = b & ~is_past_cp1;
453      /* If this is index_b (the final block), but not
454       * index_a (the end of the data), then the 64-bit
455       * length didn't fit into index_a and we're having to
456       * add an extra block of zeros. */
457      b &= ~is_block_b | is_block_a;
458
459      /* The final bytes of one of the blocks contains the
460       * length. */
461      if (j >= md_block_size - md_length_size) {
462        /* If this is index_b, write a length byte. */
463        b = constant_time_select_8(
464            is_block_b, length_bytes[j - (md_block_size - md_length_size)], b);
465      }
466      block[j] = b;
467    }
468
469    md_transform(md_state.c, block);
470    md_final_raw(md_state.c, block);
471    /* If this is index_b, copy the hash value to |mac_out|. */
472    for (j = 0; j < md_size; j++) {
473      mac_out[j] |= block[j] & is_block_b;
474    }
475  }
476
477  EVP_MD_CTX_init(&md_ctx);
478  if (!EVP_DigestInit_ex(&md_ctx, md, NULL /* engine */)) {
479    EVP_MD_CTX_cleanup(&md_ctx);
480    return 0;
481  }
482
483  /* Complete the HMAC in the standard manner. */
484  for (i = 0; i < md_block_size; i++) {
485    hmac_pad[i] ^= 0x6a;
486  }
487
488  EVP_DigestUpdate(&md_ctx, hmac_pad, md_block_size);
489  EVP_DigestUpdate(&md_ctx, mac_out, md_size);
490  EVP_DigestFinal(&md_ctx, md_out, &md_out_size_u);
491  *md_out_size = md_out_size_u;
492  EVP_MD_CTX_cleanup(&md_ctx);
493
494  return 1;
495}
496