/Users/brunogarcia/projects/bitcoin-core-dev/src/crypto/poly1305.cpp
Line | Count | Source |
1 | | // Copyright (c) 2019-present The Bitcoin Core developers |
2 | | // Distributed under the MIT software license, see the accompanying |
3 | | // file COPYING or http://www.opensource.org/licenses/mit-license.php. |
4 | | |
5 | | #include <crypto/common.h> |
6 | | #include <crypto/poly1305.h> |
7 | | |
8 | | namespace poly1305_donna { |
9 | | |
10 | | // Based on the public domain implementation by Andrew Moon |
11 | | // poly1305-donna-32.h from https://github.com/floodyberry/poly1305-donna |
12 | | |
13 | 0 | void poly1305_init(poly1305_context *st, const unsigned char key[32]) noexcept { |
14 | | /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */ |
15 | 0 | st->r[0] = (ReadLE32(&key[ 0]) ) & 0x3ffffff; |
16 | 0 | st->r[1] = (ReadLE32(&key[ 3]) >> 2) & 0x3ffff03; |
17 | 0 | st->r[2] = (ReadLE32(&key[ 6]) >> 4) & 0x3ffc0ff; |
18 | 0 | st->r[3] = (ReadLE32(&key[ 9]) >> 6) & 0x3f03fff; |
19 | 0 | st->r[4] = (ReadLE32(&key[12]) >> 8) & 0x00fffff; |
20 | | |
21 | | /* h = 0 */ |
22 | 0 | st->h[0] = 0; |
23 | 0 | st->h[1] = 0; |
24 | 0 | st->h[2] = 0; |
25 | 0 | st->h[3] = 0; |
26 | 0 | st->h[4] = 0; |
27 | | |
28 | | /* save pad for later */ |
29 | 0 | st->pad[0] = ReadLE32(&key[16]); |
30 | 0 | st->pad[1] = ReadLE32(&key[20]); |
31 | 0 | st->pad[2] = ReadLE32(&key[24]); |
32 | 0 | st->pad[3] = ReadLE32(&key[28]); |
33 | |
|
34 | 0 | st->leftover = 0; |
35 | 0 | st->final = 0; |
36 | 0 | } |
37 | | |
38 | 0 | static void poly1305_blocks(poly1305_context *st, const unsigned char *m, size_t bytes) noexcept { |
39 | 0 | const uint32_t hibit = (st->final) ? 0 : (1UL << 24); /* 1 << 128 */ |
40 | 0 | uint32_t r0,r1,r2,r3,r4; |
41 | 0 | uint32_t s1,s2,s3,s4; |
42 | 0 | uint32_t h0,h1,h2,h3,h4; |
43 | 0 | uint64_t d0,d1,d2,d3,d4; |
44 | 0 | uint32_t c; |
45 | |
|
46 | 0 | r0 = st->r[0]; |
47 | 0 | r1 = st->r[1]; |
48 | 0 | r2 = st->r[2]; |
49 | 0 | r3 = st->r[3]; |
50 | 0 | r4 = st->r[4]; |
51 | |
|
52 | 0 | s1 = r1 * 5; |
53 | 0 | s2 = r2 * 5; |
54 | 0 | s3 = r3 * 5; |
55 | 0 | s4 = r4 * 5; |
56 | |
|
57 | 0 | h0 = st->h[0]; |
58 | 0 | h1 = st->h[1]; |
59 | 0 | h2 = st->h[2]; |
60 | 0 | h3 = st->h[3]; |
61 | 0 | h4 = st->h[4]; |
62 | |
|
63 | 0 | while (bytes >= POLY1305_BLOCK_SIZE) {Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
64 | | /* h += m[i] */ |
65 | 0 | h0 += (ReadLE32(m+ 0) ) & 0x3ffffff; |
66 | 0 | h1 += (ReadLE32(m+ 3) >> 2) & 0x3ffffff; |
67 | 0 | h2 += (ReadLE32(m+ 6) >> 4) & 0x3ffffff; |
68 | 0 | h3 += (ReadLE32(m+ 9) >> 6) & 0x3ffffff; |
69 | 0 | h4 += (ReadLE32(m+12) >> 8) | hibit; |
70 | | |
71 | | /* h *= r */ |
72 | 0 | d0 = ((uint64_t)h0 * r0) + ((uint64_t)h1 * s4) + ((uint64_t)h2 * s3) + ((uint64_t)h3 * s2) + ((uint64_t)h4 * s1); |
73 | 0 | d1 = ((uint64_t)h0 * r1) + ((uint64_t)h1 * r0) + ((uint64_t)h2 * s4) + ((uint64_t)h3 * s3) + ((uint64_t)h4 * s2); |
74 | 0 | d2 = ((uint64_t)h0 * r2) + ((uint64_t)h1 * r1) + ((uint64_t)h2 * r0) + ((uint64_t)h3 * s4) + ((uint64_t)h4 * s3); |
75 | 0 | d3 = ((uint64_t)h0 * r3) + ((uint64_t)h1 * r2) + ((uint64_t)h2 * r1) + ((uint64_t)h3 * r0) + ((uint64_t)h4 * s4); |
76 | 0 | d4 = ((uint64_t)h0 * r4) + ((uint64_t)h1 * r3) + ((uint64_t)h2 * r2) + ((uint64_t)h3 * r1) + ((uint64_t)h4 * r0); |
77 | | |
78 | | /* (partial) h %= p */ |
79 | 0 | c = (uint32_t)(d0 >> 26); h0 = (uint32_t)d0 & 0x3ffffff; |
80 | 0 | d1 += c; c = (uint32_t)(d1 >> 26); h1 = (uint32_t)d1 & 0x3ffffff; |
81 | 0 | d2 += c; c = (uint32_t)(d2 >> 26); h2 = (uint32_t)d2 & 0x3ffffff; |
82 | 0 | d3 += c; c = (uint32_t)(d3 >> 26); h3 = (uint32_t)d3 & 0x3ffffff; |
83 | 0 | d4 += c; c = (uint32_t)(d4 >> 26); h4 = (uint32_t)d4 & 0x3ffffff; |
84 | 0 | h0 += c * 5; c = (h0 >> 26); h0 = h0 & 0x3ffffff; |
85 | 0 | h1 += c; |
86 | |
|
87 | 0 | m += POLY1305_BLOCK_SIZE; Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
88 | 0 | bytes -= POLY1305_BLOCK_SIZE; Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
89 | 0 | } |
90 | |
|
91 | 0 | st->h[0] = h0; |
92 | 0 | st->h[1] = h1; |
93 | 0 | st->h[2] = h2; |
94 | 0 | st->h[3] = h3; |
95 | 0 | st->h[4] = h4; |
96 | 0 | } |
97 | | |
98 | 0 | void poly1305_finish(poly1305_context *st, unsigned char mac[16]) noexcept { |
99 | 0 | uint32_t h0,h1,h2,h3,h4,c; |
100 | 0 | uint32_t g0,g1,g2,g3,g4; |
101 | 0 | uint64_t f; |
102 | 0 | uint32_t mask; |
103 | | |
104 | | /* process the remaining block */ |
105 | 0 | if (st->leftover) { |
106 | 0 | size_t i = st->leftover; |
107 | 0 | st->buffer[i++] = 1; |
108 | 0 | for (; i < POLY1305_BLOCK_SIZE; i++) {Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
109 | 0 | st->buffer[i] = 0; |
110 | 0 | } |
111 | 0 | st->final = 1; |
112 | 0 | poly1305_blocks(st, st->buffer, POLY1305_BLOCK_SIZE); Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
113 | 0 | } |
114 | | |
115 | | /* fully carry h */ |
116 | 0 | h0 = st->h[0]; |
117 | 0 | h1 = st->h[1]; |
118 | 0 | h2 = st->h[2]; |
119 | 0 | h3 = st->h[3]; |
120 | 0 | h4 = st->h[4]; |
121 | |
|
122 | 0 | c = h1 >> 26; h1 = h1 & 0x3ffffff; |
123 | 0 | h2 += c; c = h2 >> 26; h2 = h2 & 0x3ffffff; |
124 | 0 | h3 += c; c = h3 >> 26; h3 = h3 & 0x3ffffff; |
125 | 0 | h4 += c; c = h4 >> 26; h4 = h4 & 0x3ffffff; |
126 | 0 | h0 += c * 5; c = h0 >> 26; h0 = h0 & 0x3ffffff; |
127 | 0 | h1 += c; |
128 | | |
129 | | /* compute h + -p */ |
130 | 0 | g0 = h0 + 5; c = g0 >> 26; g0 &= 0x3ffffff; |
131 | 0 | g1 = h1 + c; c = g1 >> 26; g1 &= 0x3ffffff; |
132 | 0 | g2 = h2 + c; c = g2 >> 26; g2 &= 0x3ffffff; |
133 | 0 | g3 = h3 + c; c = g3 >> 26; g3 &= 0x3ffffff; |
134 | 0 | g4 = h4 + c - (1UL << 26); |
135 | | |
136 | | /* select h if h < p, or h + -p if h >= p */ |
137 | 0 | mask = (g4 >> ((sizeof(uint32_t) * 8) - 1)) - 1; |
138 | 0 | g0 &= mask; |
139 | 0 | g1 &= mask; |
140 | 0 | g2 &= mask; |
141 | 0 | g3 &= mask; |
142 | 0 | g4 &= mask; |
143 | 0 | mask = ~mask; |
144 | 0 | h0 = (h0 & mask) | g0; |
145 | 0 | h1 = (h1 & mask) | g1; |
146 | 0 | h2 = (h2 & mask) | g2; |
147 | 0 | h3 = (h3 & mask) | g3; |
148 | 0 | h4 = (h4 & mask) | g4; |
149 | | |
150 | | /* h = h % (2^128) */ |
151 | 0 | h0 = ((h0 ) | (h1 << 26)) & 0xffffffff; |
152 | 0 | h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff; |
153 | 0 | h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff; |
154 | 0 | h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff; |
155 | | |
156 | | /* mac = (h + pad) % (2^128) */ |
157 | 0 | f = (uint64_t)h0 + st->pad[0] ; h0 = (uint32_t)f; |
158 | 0 | f = (uint64_t)h1 + st->pad[1] + (f >> 32); h1 = (uint32_t)f; |
159 | 0 | f = (uint64_t)h2 + st->pad[2] + (f >> 32); h2 = (uint32_t)f; |
160 | 0 | f = (uint64_t)h3 + st->pad[3] + (f >> 32); h3 = (uint32_t)f; |
161 | |
|
162 | 0 | WriteLE32(mac + 0, h0); |
163 | 0 | WriteLE32(mac + 4, h1); |
164 | 0 | WriteLE32(mac + 8, h2); |
165 | 0 | WriteLE32(mac + 12, h3); |
166 | | |
167 | | /* zero out the state */ |
168 | 0 | st->h[0] = 0; |
169 | 0 | st->h[1] = 0; |
170 | 0 | st->h[2] = 0; |
171 | 0 | st->h[3] = 0; |
172 | 0 | st->h[4] = 0; |
173 | 0 | st->r[0] = 0; |
174 | 0 | st->r[1] = 0; |
175 | 0 | st->r[2] = 0; |
176 | 0 | st->r[3] = 0; |
177 | 0 | st->r[4] = 0; |
178 | 0 | st->pad[0] = 0; |
179 | 0 | st->pad[1] = 0; |
180 | 0 | st->pad[2] = 0; |
181 | 0 | st->pad[3] = 0; |
182 | 0 | } |
183 | | |
184 | 0 | void poly1305_update(poly1305_context *st, const unsigned char *m, size_t bytes) noexcept { |
185 | 0 | size_t i; |
186 | | |
187 | | /* handle leftover */ |
188 | 0 | if (st->leftover) { |
189 | 0 | size_t want = (POLY1305_BLOCK_SIZE - st->leftover); Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
190 | 0 | if (want > bytes) { |
191 | 0 | want = bytes; |
192 | 0 | } |
193 | 0 | for (i = 0; i < want; i++) { |
194 | 0 | st->buffer[st->leftover + i] = m[i]; |
195 | 0 | } |
196 | 0 | bytes -= want; |
197 | 0 | m += want; |
198 | 0 | st->leftover += want; |
199 | 0 | if (st->leftover < POLY1305_BLOCK_SIZE) return; Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
200 | 0 | poly1305_blocks(st, st->buffer, POLY1305_BLOCK_SIZE); Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
201 | 0 | st->leftover = 0; |
202 | 0 | } |
203 | | |
204 | | /* process full blocks */ |
205 | 0 | if (bytes >= POLY1305_BLOCK_SIZE) {Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
206 | 0 | size_t want = (bytes & ~(POLY1305_BLOCK_SIZE - 1)); Line | Count | Source | 15 | 0 | #define POLY1305_BLOCK_SIZE 16 |
|
207 | 0 | poly1305_blocks(st, m, want); |
208 | 0 | m += want; |
209 | 0 | bytes -= want; |
210 | 0 | } |
211 | | |
212 | | /* store leftover */ |
213 | 0 | if (bytes) { |
214 | 0 | for (i = 0; i < bytes; i++) { |
215 | 0 | st->buffer[st->leftover + i] = m[i]; |
216 | 0 | } |
217 | 0 | st->leftover += bytes; |
218 | 0 | } |
219 | 0 | } |
220 | | |
221 | | } // namespace poly1305_donna |