xref: /freebsd/crypto/openssl/ssl/s3_cbc.c (revision ccb59683b98360afaf5b5bb641a68fea22c68d0b)
1 /*
2  * Copyright 2012-2021 The OpenSSL Project Authors. All Rights Reserved.
3  *
4  * Licensed under the OpenSSL license (the "License").  You may not use
5  * this file except in compliance with the License.  You can obtain a copy
6  * in the file LICENSE in the source distribution or at
7  * https://www.openssl.org/source/license.html
8  */
9 
10 #include "internal/constant_time.h"
11 #include "ssl_local.h"
12 #include "internal/cryptlib.h"
13 
14 #include <openssl/md5.h>
15 #include <openssl/sha.h>
16 
17 /*
18  * MAX_HASH_BIT_COUNT_BYTES is the maximum number of bytes in the hash's
19  * length field. (SHA-384/512 have 128-bit length.)
20  */
21 #define MAX_HASH_BIT_COUNT_BYTES 16
22 
23 /*
24  * MAX_HASH_BLOCK_SIZE is the maximum hash block size that we'll support.
25  * Currently SHA-384/512 has a 128-byte block size and that's the largest
26  * supported by TLS.)
27  */
28 #define MAX_HASH_BLOCK_SIZE 128
29 
30 /*
31  * u32toLE serialises an unsigned, 32-bit number (n) as four bytes at (p) in
32  * little-endian order. The value of p is advanced by four.
33  */
34 #define u32toLE(n, p) \
35         (*((p)++)=(unsigned char)(n), \
36          *((p)++)=(unsigned char)(n>>8), \
37          *((p)++)=(unsigned char)(n>>16), \
38          *((p)++)=(unsigned char)(n>>24))
39 
40 /*
41  * These functions serialize the state of a hash and thus perform the
42  * standard "final" operation without adding the padding and length that such
43  * a function typically does.
44  */
45 static void tls1_md5_final_raw(void *ctx, unsigned char *md_out)
46 {
47     MD5_CTX *md5 = ctx;
48     u32toLE(md5->A, md_out);
49     u32toLE(md5->B, md_out);
50     u32toLE(md5->C, md_out);
51     u32toLE(md5->D, md_out);
52 }
53 
54 static void tls1_sha1_final_raw(void *ctx, unsigned char *md_out)
55 {
56     SHA_CTX *sha1 = ctx;
57     l2n(sha1->h0, md_out);
58     l2n(sha1->h1, md_out);
59     l2n(sha1->h2, md_out);
60     l2n(sha1->h3, md_out);
61     l2n(sha1->h4, md_out);
62 }
63 
64 static void tls1_sha256_final_raw(void *ctx, unsigned char *md_out)
65 {
66     SHA256_CTX *sha256 = ctx;
67     unsigned i;
68 
69     for (i = 0; i < 8; i++) {
70         l2n(sha256->h[i], md_out);
71     }
72 }
73 
74 static void tls1_sha512_final_raw(void *ctx, unsigned char *md_out)
75 {
76     SHA512_CTX *sha512 = ctx;
77     unsigned i;
78 
79     for (i = 0; i < 8; i++) {
80         l2n8(sha512->h[i], md_out);
81     }
82 }
83 
84 #undef  LARGEST_DIGEST_CTX
85 #define LARGEST_DIGEST_CTX SHA512_CTX
86 
87 /*
88  * ssl3_cbc_record_digest_supported returns 1 iff |ctx| uses a hash function
89  * which ssl3_cbc_digest_record supports.
90  */
91 char ssl3_cbc_record_digest_supported(const EVP_MD_CTX *ctx)
92 {
93     switch (EVP_MD_CTX_type(ctx)) {
94     case NID_md5:
95     case NID_sha1:
96     case NID_sha224:
97     case NID_sha256:
98     case NID_sha384:
99     case NID_sha512:
100         return 1;
101     default:
102         return 0;
103     }
104 }
105 
106 /*-
107  * ssl3_cbc_digest_record computes the MAC of a decrypted, padded SSLv3/TLS
108  * record.
109  *
110  *   ctx: the EVP_MD_CTX from which we take the hash function.
111  *     ssl3_cbc_record_digest_supported must return true for this EVP_MD_CTX.
112  *   md_out: the digest output. At most EVP_MAX_MD_SIZE bytes will be written.
113  *   md_out_size: if non-NULL, the number of output bytes is written here.
114  *   header: the 13-byte, TLS record header.
115  *   data: the record data itself, less any preceding explicit IV.
116  *   data_plus_mac_size: the secret, reported length of the data and MAC
117  *     once the padding has been removed.
118  *   data_plus_mac_plus_padding_size: the public length of the whole
119  *     record, including padding.
120  *   is_sslv3: non-zero if we are to use SSLv3. Otherwise, TLS.
121  *
122  * On entry: by virtue of having been through one of the remove_padding
123  * functions, above, we know that data_plus_mac_size is large enough to contain
124  * a padding byte and MAC. (If the padding was invalid, it might contain the
125  * padding too. )
126  * Returns 1 on success or 0 on error
127  */
128 int ssl3_cbc_digest_record(const EVP_MD_CTX *ctx,
129                            unsigned char *md_out,
130                            size_t *md_out_size,
131                            const unsigned char *header,
132                            const unsigned char *data,
133                            size_t data_plus_mac_size,
134                            size_t data_plus_mac_plus_padding_size,
135                            const unsigned char *mac_secret,
136                            size_t mac_secret_length, char is_sslv3)
137 {
138     union {
139         double align;
140         unsigned char c[sizeof(LARGEST_DIGEST_CTX)];
141     } md_state;
142     void (*md_final_raw) (void *ctx, unsigned char *md_out);
143     void (*md_transform) (void *ctx, const unsigned char *block);
144     size_t md_size, md_block_size = 64;
145     size_t sslv3_pad_length = 40, header_length, variance_blocks,
146         len, max_mac_bytes, num_blocks,
147         num_starting_blocks, k, mac_end_offset, c, index_a, index_b;
148     size_t bits;          /* at most 18 bits */
149     unsigned char length_bytes[MAX_HASH_BIT_COUNT_BYTES];
150     /* hmac_pad is the masked HMAC key. */
151     unsigned char hmac_pad[MAX_HASH_BLOCK_SIZE];
152     unsigned char first_block[MAX_HASH_BLOCK_SIZE];
153     unsigned char mac_out[EVP_MAX_MD_SIZE];
154     size_t i, j;
155     unsigned md_out_size_u;
156     EVP_MD_CTX *md_ctx = NULL;
157     /*
158      * mdLengthSize is the number of bytes in the length field that
159      * terminates * the hash.
160      */
161     size_t md_length_size = 8;
162     char length_is_big_endian = 1;
163     int ret;
164 
165     /*
166      * This is a, hopefully redundant, check that allows us to forget about
167      * many possible overflows later in this function.
168      */
169     if (!ossl_assert(data_plus_mac_plus_padding_size < 1024 * 1024))
170         return 0;
171 
172     switch (EVP_MD_CTX_type(ctx)) {
173     case NID_md5:
174         if (MD5_Init((MD5_CTX *)md_state.c) <= 0)
175             return 0;
176         md_final_raw = tls1_md5_final_raw;
177         md_transform =
178             (void (*)(void *ctx, const unsigned char *block))MD5_Transform;
179         md_size = 16;
180         sslv3_pad_length = 48;
181         length_is_big_endian = 0;
182         break;
183     case NID_sha1:
184         if (SHA1_Init((SHA_CTX *)md_state.c) <= 0)
185             return 0;
186         md_final_raw = tls1_sha1_final_raw;
187         md_transform =
188             (void (*)(void *ctx, const unsigned char *block))SHA1_Transform;
189         md_size = 20;
190         break;
191     case NID_sha224:
192         if (SHA224_Init((SHA256_CTX *)md_state.c) <= 0)
193             return 0;
194         md_final_raw = tls1_sha256_final_raw;
195         md_transform =
196             (void (*)(void *ctx, const unsigned char *block))SHA256_Transform;
197         md_size = 224 / 8;
198         break;
199     case NID_sha256:
200         if (SHA256_Init((SHA256_CTX *)md_state.c) <= 0)
201             return 0;
202         md_final_raw = tls1_sha256_final_raw;
203         md_transform =
204             (void (*)(void *ctx, const unsigned char *block))SHA256_Transform;
205         md_size = 32;
206         break;
207     case NID_sha384:
208         if (SHA384_Init((SHA512_CTX *)md_state.c) <= 0)
209             return 0;
210         md_final_raw = tls1_sha512_final_raw;
211         md_transform =
212             (void (*)(void *ctx, const unsigned char *block))SHA512_Transform;
213         md_size = 384 / 8;
214         md_block_size = 128;
215         md_length_size = 16;
216         break;
217     case NID_sha512:
218         if (SHA512_Init((SHA512_CTX *)md_state.c) <= 0)
219             return 0;
220         md_final_raw = tls1_sha512_final_raw;
221         md_transform =
222             (void (*)(void *ctx, const unsigned char *block))SHA512_Transform;
223         md_size = 64;
224         md_block_size = 128;
225         md_length_size = 16;
226         break;
227     default:
228         /*
229          * ssl3_cbc_record_digest_supported should have been called first to
230          * check that the hash function is supported.
231          */
232         if (md_out_size != NULL)
233             *md_out_size = 0;
234         return ossl_assert(0);
235     }
236 
237     if (!ossl_assert(md_length_size <= MAX_HASH_BIT_COUNT_BYTES)
238             || !ossl_assert(md_block_size <= MAX_HASH_BLOCK_SIZE)
239             || !ossl_assert(md_size <= EVP_MAX_MD_SIZE))
240         return 0;
241 
242     header_length = 13;
243     if (is_sslv3) {
244         header_length = mac_secret_length + sslv3_pad_length + 8 /* sequence
245                                                                   * number */  +
246             1 /* record type */  +
247             2 /* record length */ ;
248     }
249 
250     /*
251      * variance_blocks is the number of blocks of the hash that we have to
252      * calculate in constant time because they could be altered by the
253      * padding value. In SSLv3, the padding must be minimal so the end of
254      * the plaintext varies by, at most, 15+20 = 35 bytes. (We conservatively
255      * assume that the MAC size varies from 0..20 bytes.) In case the 9 bytes
256      * of hash termination (0x80 + 64-bit length) don't fit in the final
257      * block, we say that the final two blocks can vary based on the padding.
258      * TLSv1 has MACs up to 48 bytes long (SHA-384) and the padding is not
259      * required to be minimal. Therefore we say that the final |variance_blocks|
260      * blocks can
261      * vary based on the padding. Later in the function, if the message is
262      * short and there obviously cannot be this many blocks then
263      * variance_blocks can be reduced.
264      */
265     variance_blocks = is_sslv3 ? 2 : ( ((255 + 1 + md_size + md_block_size - 1) / md_block_size) + 1);
266     /*
267      * From now on we're dealing with the MAC, which conceptually has 13
268      * bytes of `header' before the start of the data (TLS) or 71/75 bytes
269      * (SSLv3)
270      */
271     len = data_plus_mac_plus_padding_size + header_length;
272     /*
273      * max_mac_bytes contains the maximum bytes of bytes in the MAC,
274      * including * |header|, assuming that there's no padding.
275      */
276     max_mac_bytes = len - md_size - 1;
277     /* num_blocks is the maximum number of hash blocks. */
278     num_blocks =
279         (max_mac_bytes + 1 + md_length_size + md_block_size -
280          1) / md_block_size;
281     /*
282      * In order to calculate the MAC in constant time we have to handle the
283      * final blocks specially because the padding value could cause the end
284      * to appear somewhere in the final |variance_blocks| blocks and we can't
285      * leak where. However, |num_starting_blocks| worth of data can be hashed
286      * right away because no padding value can affect whether they are
287      * plaintext.
288      */
289     num_starting_blocks = 0;
290     /*
291      * k is the starting byte offset into the conceptual header||data where
292      * we start processing.
293      */
294     k = 0;
295     /*
296      * mac_end_offset is the index just past the end of the data to be MACed.
297      */
298     mac_end_offset = data_plus_mac_size + header_length - md_size;
299     /*
300      * c is the index of the 0x80 byte in the final hash block that contains
301      * application data.
302      */
303     c = mac_end_offset % md_block_size;
304     /*
305      * index_a is the hash block number that contains the 0x80 terminating
306      * value.
307      */
308     index_a = mac_end_offset / md_block_size;
309     /*
310      * index_b is the hash block number that contains the 64-bit hash length,
311      * in bits.
312      */
313     index_b = (mac_end_offset + md_length_size) / md_block_size;
314     /*
315      * bits is the hash-length in bits. It includes the additional hash block
316      * for the masked HMAC key, or whole of |header| in the case of SSLv3.
317      */
318 
319     /*
320      * For SSLv3, if we're going to have any starting blocks then we need at
321      * least two because the header is larger than a single block.
322      */
323     if (num_blocks > variance_blocks + (is_sslv3 ? 1 : 0)) {
324         num_starting_blocks = num_blocks - variance_blocks;
325         k = md_block_size * num_starting_blocks;
326     }
327 
328     bits = 8 * mac_end_offset;
329     if (!is_sslv3) {
330         /*
331          * Compute the initial HMAC block. For SSLv3, the padding and secret
332          * bytes are included in |header| because they take more than a
333          * single block.
334          */
335         bits += 8 * md_block_size;
336         memset(hmac_pad, 0, md_block_size);
337         if (!ossl_assert(mac_secret_length <= sizeof(hmac_pad)))
338             return 0;
339         memcpy(hmac_pad, mac_secret, mac_secret_length);
340         for (i = 0; i < md_block_size; i++)
341             hmac_pad[i] ^= 0x36;
342 
343         md_transform(md_state.c, hmac_pad);
344     }
345 
346     if (length_is_big_endian) {
347         memset(length_bytes, 0, md_length_size - 4);
348         length_bytes[md_length_size - 4] = (unsigned char)(bits >> 24);
349         length_bytes[md_length_size - 3] = (unsigned char)(bits >> 16);
350         length_bytes[md_length_size - 2] = (unsigned char)(bits >> 8);
351         length_bytes[md_length_size - 1] = (unsigned char)bits;
352     } else {
353         memset(length_bytes, 0, md_length_size);
354         length_bytes[md_length_size - 5] = (unsigned char)(bits >> 24);
355         length_bytes[md_length_size - 6] = (unsigned char)(bits >> 16);
356         length_bytes[md_length_size - 7] = (unsigned char)(bits >> 8);
357         length_bytes[md_length_size - 8] = (unsigned char)bits;
358     }
359 
360     if (k > 0) {
361         if (is_sslv3) {
362             size_t overhang;
363 
364             /*
365              * The SSLv3 header is larger than a single block. overhang is
366              * the number of bytes beyond a single block that the header
367              * consumes: either 7 bytes (SHA1) or 11 bytes (MD5). There are no
368              * ciphersuites in SSLv3 that are not SHA1 or MD5 based and
369              * therefore we can be confident that the header_length will be
370              * greater than |md_block_size|. However we add a sanity check just
371              * in case
372              */
373             if (header_length <= md_block_size) {
374                 /* Should never happen */
375                 return 0;
376             }
377             overhang = header_length - md_block_size;
378             md_transform(md_state.c, header);
379             memcpy(first_block, header + md_block_size, overhang);
380             memcpy(first_block + overhang, data, md_block_size - overhang);
381             md_transform(md_state.c, first_block);
382             for (i = 1; i < k / md_block_size - 1; i++)
383                 md_transform(md_state.c, data + md_block_size * i - overhang);
384         } else {
385             /* k is a multiple of md_block_size. */
386             memcpy(first_block, header, 13);
387             memcpy(first_block + 13, data, md_block_size - 13);
388             md_transform(md_state.c, first_block);
389             for (i = 1; i < k / md_block_size; i++)
390                 md_transform(md_state.c, data + md_block_size * i - 13);
391         }
392     }
393 
394     memset(mac_out, 0, sizeof(mac_out));
395 
396     /*
397      * We now process the final hash blocks. For each block, we construct it
398      * in constant time. If the |i==index_a| then we'll include the 0x80
399      * bytes and zero pad etc. For each block we selectively copy it, in
400      * constant time, to |mac_out|.
401      */
402     for (i = num_starting_blocks; i <= num_starting_blocks + variance_blocks;
403          i++) {
404         unsigned char block[MAX_HASH_BLOCK_SIZE];
405         unsigned char is_block_a = constant_time_eq_8_s(i, index_a);
406         unsigned char is_block_b = constant_time_eq_8_s(i, index_b);
407         for (j = 0; j < md_block_size; j++) {
408             unsigned char b = 0, is_past_c, is_past_cp1;
409             if (k < header_length)
410                 b = header[k];
411             else if (k < data_plus_mac_plus_padding_size + header_length)
412                 b = data[k - header_length];
413             k++;
414 
415             is_past_c = is_block_a & constant_time_ge_8_s(j, c);
416             is_past_cp1 = is_block_a & constant_time_ge_8_s(j, c + 1);
417             /*
418              * If this is the block containing the end of the application
419              * data, and we are at the offset for the 0x80 value, then
420              * overwrite b with 0x80.
421              */
422             b = constant_time_select_8(is_past_c, 0x80, b);
423             /*
424              * If this block contains the end of the application data
425              * and we're past the 0x80 value then just write zero.
426              */
427             b = b & ~is_past_cp1;
428             /*
429              * If this is index_b (the final block), but not index_a (the end
430              * of the data), then the 64-bit length didn't fit into index_a
431              * and we're having to add an extra block of zeros.
432              */
433             b &= ~is_block_b | is_block_a;
434 
435             /*
436              * The final bytes of one of the blocks contains the length.
437              */
438             if (j >= md_block_size - md_length_size) {
439                 /* If this is index_b, write a length byte. */
440                 b = constant_time_select_8(is_block_b,
441                                            length_bytes[j -
442                                                         (md_block_size -
443                                                          md_length_size)], b);
444             }
445             block[j] = b;
446         }
447 
448         md_transform(md_state.c, block);
449         md_final_raw(md_state.c, block);
450         /* If this is index_b, copy the hash value to |mac_out|. */
451         for (j = 0; j < md_size; j++)
452             mac_out[j] |= block[j] & is_block_b;
453     }
454 
455     md_ctx = EVP_MD_CTX_new();
456     if (md_ctx == NULL)
457         goto err;
458     if (EVP_DigestInit_ex(md_ctx, EVP_MD_CTX_md(ctx), NULL /* engine */ ) <= 0)
459         goto err;
460     if (is_sslv3) {
461         /* We repurpose |hmac_pad| to contain the SSLv3 pad2 block. */
462         memset(hmac_pad, 0x5c, sslv3_pad_length);
463 
464         if (EVP_DigestUpdate(md_ctx, mac_secret, mac_secret_length) <= 0
465             || EVP_DigestUpdate(md_ctx, hmac_pad, sslv3_pad_length) <= 0
466             || EVP_DigestUpdate(md_ctx, mac_out, md_size) <= 0)
467             goto err;
468     } else {
469         /* Complete the HMAC in the standard manner. */
470         for (i = 0; i < md_block_size; i++)
471             hmac_pad[i] ^= 0x6a;
472 
473         if (EVP_DigestUpdate(md_ctx, hmac_pad, md_block_size) <= 0
474             || EVP_DigestUpdate(md_ctx, mac_out, md_size) <= 0)
475             goto err;
476     }
477     /* TODO(size_t): Convert me */
478     ret = EVP_DigestFinal(md_ctx, md_out, &md_out_size_u);
479     if (ret && md_out_size)
480         *md_out_size = md_out_size_u;
481     EVP_MD_CTX_free(md_ctx);
482 
483     return 1;
484  err:
485     EVP_MD_CTX_free(md_ctx);
486     return 0;
487 }
488