Skip to content

Commit 50269b3

Browse files
sipasipa
committed
crypto: switch poly1305 to incremental implementation
This code is taken from poly1305-donna-32.h, poly1305-donna.h, poly1305-donna.c from https://github.com/floodyberry/poly1305-donna, commit e6ad6e091d30d7f4ec2d4f978be1fcfcbce72781, with the following modifications: * Coding style (braces around one-line indented if/for loops). * Rename unsigned long (long) to uint32_t and uint64_t. * Rename poly1305_block_size to POLY1305_BLOCK_SIZE. * Adding noexcept to functions. * Merging poly1305_state_internal_t and poly1305_context types. * Merging code from multiple files. * Place all imported code in the poly1305_donna namespace.
1 parent b479474 commit 50269b3

File tree

2 files changed

+233
-122
lines changed

2 files changed

+233
-122
lines changed

‎src/crypto/poly1305.cpp

+212-122
Original file line numberDiff line numberDiff line change
@@ -2,140 +2,230 @@
22
// Distributed under the MIT software license, see the accompanying
33
// file COPYING or http://www.opensource.org/licenses/mit-license.php.
44

5-
// Based on the public domain implementation by Andrew Moon
6-
// poly1305-donna-unrolled.c from https://github.com/floodyberry/poly1305-donna
7-
85
#include <crypto/common.h>
96
#include <crypto/poly1305.h>
107

118
#include <string.h>
129

13-
#define mul32x32_64(a,b) ((uint64_t)(a) * (b))
10+
namespace poly1305_donna {
1411

15-
void poly1305_auth(unsigned char out[POLY1305_TAGLEN], const unsigned char *m, size_t inlen, const unsigned char key[POLY1305_KEYLEN]) {
16-
uint32_t t0,t1,t2,t3;
17-
uint32_t h0,h1,h2,h3,h4;
12+
// Based on the public domain implementation by Andrew Moon
13+
// poly1305-donna-32.h from https://github.com/floodyberry/poly1305-donna
14+
15+
void poly1305_init(poly1305_context *st, const unsigned char key[32]) noexcept {
16+
/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
17+
st->r[0] = (ReadLE32(&key[ 0]) ) & 0x3ffffff;
18+
st->r[1] = (ReadLE32(&key[ 3]) >> 2) & 0x3ffff03;
19+
st->r[2] = (ReadLE32(&key[ 6]) >> 4) & 0x3ffc0ff;
20+
st->r[3] = (ReadLE32(&key[ 9]) >> 6) & 0x3f03fff;
21+
st->r[4] = (ReadLE32(&key[12]) >> 8) & 0x00fffff;
22+
23+
/* h = 0 */
24+
st->h[0] = 0;
25+
st->h[1] = 0;
26+
st->h[2] = 0;
27+
st->h[3] = 0;
28+
st->h[4] = 0;
29+
30+
/* save pad for later */
31+
st->pad[0] = ReadLE32(&key[16]);
32+
st->pad[1] = ReadLE32(&key[20]);
33+
st->pad[2] = ReadLE32(&key[24]);
34+
st->pad[3] = ReadLE32(&key[28]);
35+
36+
st->leftover = 0;
37+
st->final = 0;
38+
}
39+
40+
static void poly1305_blocks(poly1305_context *st, const unsigned char *m, size_t bytes) noexcept {
41+
const uint32_t hibit = (st->final) ? 0 : (1UL << 24); /* 1 << 128 */
1842
uint32_t r0,r1,r2,r3,r4;
1943
uint32_t s1,s2,s3,s4;
20-
uint32_t b, nb;
21-
size_t j;
22-
uint64_t t[5];
23-
uint64_t f0,f1,f2,f3;
24-
uint64_t g0,g1,g2,g3,g4;
25-
uint64_t c;
26-
unsigned char mp[16];
27-
28-
/* clamp key */
29-
t0 = ReadLE32(key+0);
30-
t1 = ReadLE32(key+4);
31-
t2 = ReadLE32(key+8);
32-
t3 = ReadLE32(key+12);
33-
34-
/* precompute multipliers */
35-
r0 = t0 & 0x3ffffff; t0 >>= 26; t0 |= t1 << 6;
36-
r1 = t0 & 0x3ffff03; t1 >>= 20; t1 |= t2 << 12;
37-
r2 = t1 & 0x3ffc0ff; t2 >>= 14; t2 |= t3 << 18;
38-
r3 = t2 & 0x3f03fff; t3 >>= 8;
39-
r4 = t3 & 0x00fffff;
44+
uint32_t h0,h1,h2,h3,h4;
45+
uint64_t d0,d1,d2,d3,d4;
46+
uint32_t c;
47+
48+
r0 = st->r[0];
49+
r1 = st->r[1];
50+
r2 = st->r[2];
51+
r3 = st->r[3];
52+
r4 = st->r[4];
4053

4154
s1 = r1 * 5;
4255
s2 = r2 * 5;
4356
s3 = r3 * 5;
4457
s4 = r4 * 5;
4558

46-
/* init state */
47-
h0 = 0;
48-
h1 = 0;
49-
h2 = 0;
50-
h3 = 0;
51-
h4 = 0;
52-
53-
/* full blocks */
54-
if (inlen < 16) goto poly1305_donna_atmost15bytes;
55-
poly1305_donna_16bytes:
56-
m += 16;
57-
inlen -= 16;
58-
59-
t0 = ReadLE32(m-16);
60-
t1 = ReadLE32(m-12);
61-
t2 = ReadLE32(m-8);
62-
t3 = ReadLE32(m-4);
63-
64-
h0 += t0 & 0x3ffffff;
65-
h1 += ((((uint64_t)t1 << 32) | t0) >> 26) & 0x3ffffff;
66-
h2 += ((((uint64_t)t2 << 32) | t1) >> 20) & 0x3ffffff;
67-
h3 += ((((uint64_t)t3 << 32) | t2) >> 14) & 0x3ffffff;
68-
h4 += (t3 >> 8) | (1 << 24);
69-
70-
71-
poly1305_donna_mul:
72-
t[0] = mul32x32_64(h0,r0) + mul32x32_64(h1,s4) + mul32x32_64(h2,s3) + mul32x32_64(h3,s2) + mul32x32_64(h4,s1);
73-
t[1] = mul32x32_64(h0,r1) + mul32x32_64(h1,r0) + mul32x32_64(h2,s4) + mul32x32_64(h3,s3) + mul32x32_64(h4,s2);
74-
t[2] = mul32x32_64(h0,r2) + mul32x32_64(h1,r1) + mul32x32_64(h2,r0) + mul32x32_64(h3,s4) + mul32x32_64(h4,s3);
75-
t[3] = mul32x32_64(h0,r3) + mul32x32_64(h1,r2) + mul32x32_64(h2,r1) + mul32x32_64(h3,r0) + mul32x32_64(h4,s4);
76-
t[4] = mul32x32_64(h0,r4) + mul32x32_64(h1,r3) + mul32x32_64(h2,r2) + mul32x32_64(h3,r1) + mul32x32_64(h4,r0);
77-
78-
h0 = (uint32_t)t[0] & 0x3ffffff; c = (t[0] >> 26);
79-
t[1] += c; h1 = (uint32_t)t[1] & 0x3ffffff; b = (uint32_t)(t[1] >> 26);
80-
t[2] += b; h2 = (uint32_t)t[2] & 0x3ffffff; b = (uint32_t)(t[2] >> 26);
81-
t[3] += b; h3 = (uint32_t)t[3] & 0x3ffffff; b = (uint32_t)(t[3] >> 26);
82-
t[4] += b; h4 = (uint32_t)t[4] & 0x3ffffff; b = (uint32_t)(t[4] >> 26);
83-
h0 += b * 5;
84-
85-
if (inlen >= 16) goto poly1305_donna_16bytes;
86-
87-
/* final bytes */
88-
poly1305_donna_atmost15bytes:
89-
if (!inlen) goto poly1305_donna_finish;
90-
91-
for (j = 0; j < inlen; j++) mp[j] = m[j];
92-
mp[j++] = 1;
93-
for (; j < 16; j++) mp[j] = 0;
94-
inlen = 0;
95-
96-
t0 = ReadLE32(mp+0);
97-
t1 = ReadLE32(mp+4);
98-
t2 = ReadLE32(mp+8);
99-
t3 = ReadLE32(mp+12);
100-
101-
h0 += t0 & 0x3ffffff;
102-
h1 += ((((uint64_t)t1 << 32) | t0) >> 26) & 0x3ffffff;
103-
h2 += ((((uint64_t)t2 << 32) | t1) >> 20) & 0x3ffffff;
104-
h3 += ((((uint64_t)t3 << 32) | t2) >> 14) & 0x3ffffff;
105-
h4 += (t3 >> 8);
106-
107-
goto poly1305_donna_mul;
108-
109-
poly1305_donna_finish:
110-
b = h0 >> 26; h0 = h0 & 0x3ffffff;
111-
h1 += b; b = h1 >> 26; h1 = h1 & 0x3ffffff;
112-
h2 += b; b = h2 >> 26; h2 = h2 & 0x3ffffff;
113-
h3 += b; b = h3 >> 26; h3 = h3 & 0x3ffffff;
114-
h4 += b; b = h4 >> 26; h4 = h4 & 0x3ffffff;
115-
h0 += b * 5; b = h0 >> 26; h0 = h0 & 0x3ffffff;
116-
h1 += b;
117-
118-
g0 = h0 + 5; b = g0 >> 26; g0 &= 0x3ffffff;
119-
g1 = h1 + b; b = g1 >> 26; g1 &= 0x3ffffff;
120-
g2 = h2 + b; b = g2 >> 26; g2 &= 0x3ffffff;
121-
g3 = h3 + b; b = g3 >> 26; g3 &= 0x3ffffff;
122-
g4 = h4 + b - (1 << 26);
123-
124-
b = (g4 >> 31) - 1;
125-
nb = ~b;
126-
h0 = (h0 & nb) | (g0 & b);
127-
h1 = (h1 & nb) | (g1 & b);
128-
h2 = (h2 & nb) | (g2 & b);
129-
h3 = (h3 & nb) | (g3 & b);
130-
h4 = (h4 & nb) | (g4 & b);
131-
132-
f0 = ((h0 ) | (h1 << 26)) + (uint64_t)ReadLE32(&key[16]);
133-
f1 = ((h1 >> 6) | (h2 << 20)) + (uint64_t)ReadLE32(&key[20]);
134-
f2 = ((h2 >> 12) | (h3 << 14)) + (uint64_t)ReadLE32(&key[24]);
135-
f3 = ((h3 >> 18) | (h4 << 8)) + (uint64_t)ReadLE32(&key[28]);
136-
137-
WriteLE32(&out[ 0], f0); f1 += (f0 >> 32);
138-
WriteLE32(&out[ 4], f1); f2 += (f1 >> 32);
139-
WriteLE32(&out[ 8], f2); f3 += (f2 >> 32);
140-
WriteLE32(&out[12], f3);
59+
h0 = st->h[0];
60+
h1 = st->h[1];
61+
h2 = st->h[2];
62+
h3 = st->h[3];
63+
h4 = st->h[4];
64+
65+
while (bytes >= POLY1305_BLOCK_SIZE) {
66+
/* h += m[i] */
67+
h0 += (ReadLE32(m+ 0) ) & 0x3ffffff;
68+
h1 += (ReadLE32(m+ 3) >> 2) & 0x3ffffff;
69+
h2 += (ReadLE32(m+ 6) >> 4) & 0x3ffffff;
70+
h3 += (ReadLE32(m+ 9) >> 6) & 0x3ffffff;
71+
h4 += (ReadLE32(m+12) >> 8) | hibit;
72+
73+
/* h *= r */
74+
d0 = ((uint64_t)h0 * r0) + ((uint64_t)h1 * s4) + ((uint64_t)h2 * s3) + ((uint64_t)h3 * s2) + ((uint64_t)h4 * s1);
75+
d1 = ((uint64_t)h0 * r1) + ((uint64_t)h1 * r0) + ((uint64_t)h2 * s4) + ((uint64_t)h3 * s3) + ((uint64_t)h4 * s2);
76+
d2 = ((uint64_t)h0 * r2) + ((uint64_t)h1 * r1) + ((uint64_t)h2 * r0) + ((uint64_t)h3 * s4) + ((uint64_t)h4 * s3);
77+
d3 = ((uint64_t)h0 * r3) + ((uint64_t)h1 * r2) + ((uint64_t)h2 * r1) + ((uint64_t)h3 * r0) + ((uint64_t)h4 * s4);
78+
d4 = ((uint64_t)h0 * r4) + ((uint64_t)h1 * r3) + ((uint64_t)h2 * r2) + ((uint64_t)h3 * r1) + ((uint64_t)h4 * r0);
79+
80+
/* (partial) h %= p */
81+
c = (uint32_t)(d0 >> 26); h0 = (uint32_t)d0 & 0x3ffffff;
82+
d1 += c; c = (uint32_t)(d1 >> 26); h1 = (uint32_t)d1 & 0x3ffffff;
83+
d2 += c; c = (uint32_t)(d2 >> 26); h2 = (uint32_t)d2 & 0x3ffffff;
84+
d3 += c; c = (uint32_t)(d3 >> 26); h3 = (uint32_t)d3 & 0x3ffffff;
85+
d4 += c; c = (uint32_t)(d4 >> 26); h4 = (uint32_t)d4 & 0x3ffffff;
86+
h0 += c * 5; c = (h0 >> 26); h0 = h0 & 0x3ffffff;
87+
h1 += c;
88+
89+
m += POLY1305_BLOCK_SIZE;
90+
bytes -= POLY1305_BLOCK_SIZE;
91+
}
92+
93+
st->h[0] = h0;
94+
st->h[1] = h1;
95+
st->h[2] = h2;
96+
st->h[3] = h3;
97+
st->h[4] = h4;
98+
}
99+
100+
void poly1305_finish(poly1305_context *st, unsigned char mac[16]) noexcept {
101+
uint32_t h0,h1,h2,h3,h4,c;
102+
uint32_t g0,g1,g2,g3,g4;
103+
uint64_t f;
104+
uint32_t mask;
105+
106+
/* process the remaining block */
107+
if (st->leftover) {
108+
size_t i = st->leftover;
109+
st->buffer[i++] = 1;
110+
for (; i < POLY1305_BLOCK_SIZE; i++) {
111+
st->buffer[i] = 0;
112+
}
113+
st->final = 1;
114+
poly1305_blocks(st, st->buffer, POLY1305_BLOCK_SIZE);
115+
}
116+
117+
/* fully carry h */
118+
h0 = st->h[0];
119+
h1 = st->h[1];
120+
h2 = st->h[2];
121+
h3 = st->h[3];
122+
h4 = st->h[4];
123+
124+
c = h1 >> 26; h1 = h1 & 0x3ffffff;
125+
h2 += c; c = h2 >> 26; h2 = h2 & 0x3ffffff;
126+
h3 += c; c = h3 >> 26; h3 = h3 & 0x3ffffff;
127+
h4 += c; c = h4 >> 26; h4 = h4 & 0x3ffffff;
128+
h0 += c * 5; c = h0 >> 26; h0 = h0 & 0x3ffffff;
129+
h1 += c;
130+
131+
/* compute h + -p */
132+
g0 = h0 + 5; c = g0 >> 26; g0 &= 0x3ffffff;
133+
g1 = h1 + c; c = g1 >> 26; g1 &= 0x3ffffff;
134+
g2 = h2 + c; c = g2 >> 26; g2 &= 0x3ffffff;
135+
g3 = h3 + c; c = g3 >> 26; g3 &= 0x3ffffff;
136+
g4 = h4 + c - (1UL << 26);
137+
138+
/* select h if h < p, or h + -p if h >= p */
139+
mask = (g4 >> ((sizeof(uint32_t) * 8) - 1)) - 1;
140+
g0 &= mask;
141+
g1 &= mask;
142+
g2 &= mask;
143+
g3 &= mask;
144+
g4 &= mask;
145+
mask = ~mask;
146+
h0 = (h0 & mask) | g0;
147+
h1 = (h1 & mask) | g1;
148+
h2 = (h2 & mask) | g2;
149+
h3 = (h3 & mask) | g3;
150+
h4 = (h4 & mask) | g4;
151+
152+
/* h = h % (2^128) */
153+
h0 = ((h0 ) | (h1 << 26)) & 0xffffffff;
154+
h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff;
155+
h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff;
156+
h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff;
157+
158+
/* mac = (h + pad) % (2^128) */
159+
f = (uint64_t)h0 + st->pad[0] ; h0 = (uint32_t)f;
160+
f = (uint64_t)h1 + st->pad[1] + (f >> 32); h1 = (uint32_t)f;
161+
f = (uint64_t)h2 + st->pad[2] + (f >> 32); h2 = (uint32_t)f;
162+
f = (uint64_t)h3 + st->pad[3] + (f >> 32); h3 = (uint32_t)f;
163+
164+
WriteLE32(mac + 0, h0);
165+
WriteLE32(mac + 4, h1);
166+
WriteLE32(mac + 8, h2);
167+
WriteLE32(mac + 12, h3);
168+
169+
/* zero out the state */
170+
st->h[0] = 0;
171+
st->h[1] = 0;
172+
st->h[2] = 0;
173+
st->h[3] = 0;
174+
st->h[4] = 0;
175+
st->r[0] = 0;
176+
st->r[1] = 0;
177+
st->r[2] = 0;
178+
st->r[3] = 0;
179+
st->r[4] = 0;
180+
st->pad[0] = 0;
181+
st->pad[1] = 0;
182+
st->pad[2] = 0;
183+
st->pad[3] = 0;
184+
}
185+
186+
void poly1305_update(poly1305_context *st, const unsigned char *m, size_t bytes) noexcept {
187+
size_t i;
188+
189+
/* handle leftover */
190+
if (st->leftover) {
191+
size_t want = (POLY1305_BLOCK_SIZE - st->leftover);
192+
if (want > bytes) {
193+
want = bytes;
194+
}
195+
for (i = 0; i < want; i++) {
196+
st->buffer[st->leftover + i] = m[i];
197+
}
198+
bytes -= want;
199+
m += want;
200+
st->leftover += want;
201+
if (st->leftover < POLY1305_BLOCK_SIZE) return;
202+
poly1305_blocks(st, st->buffer, POLY1305_BLOCK_SIZE);
203+
st->leftover = 0;
204+
}
205+
206+
/* process full blocks */
207+
if (bytes >= POLY1305_BLOCK_SIZE) {
208+
size_t want = (bytes & ~(POLY1305_BLOCK_SIZE - 1));
209+
poly1305_blocks(st, m, want);
210+
m += want;
211+
bytes -= want;
212+
}
213+
214+
/* store leftover */
215+
if (bytes) {
216+
for (i = 0; i < bytes; i++) {
217+
st->buffer[st->leftover + i] = m[i];
218+
}
219+
st->leftover += bytes;
220+
}
221+
}
222+
223+
} // namespace poly1305_donna
224+
225+
void poly1305_auth(unsigned char mac[16], const unsigned char *m, size_t bytes, const unsigned char key[32]) {
226+
using namespace poly1305_donna;
227+
poly1305_context ctx;
228+
poly1305_init(&ctx, key);
229+
poly1305_update(&ctx, m, bytes);
230+
poly1305_finish(&ctx, mac);
141231
}

‎src/crypto/poly1305.h

+21
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,27 @@
1010

1111
#define POLY1305_KEYLEN 32
1212
#define POLY1305_TAGLEN 16
13+
#define POLY1305_BLOCK_SIZE 16
14+
15+
namespace poly1305_donna {
16+
17+
// Based on the public domain implementation by Andrew Moon
18+
// poly1305-donna-32.h from https://github.com/floodyberry/poly1305-donna
19+
20+
typedef struct {
21+
uint32_t r[5];
22+
uint32_t h[5];
23+
uint32_t pad[4];
24+
size_t leftover;
25+
unsigned char buffer[POLY1305_BLOCK_SIZE];
26+
unsigned char final;
27+
} poly1305_context;
28+
29+
void poly1305_init(poly1305_context *st, const unsigned char key[32]) noexcept;
30+
void poly1305_update(poly1305_context *st, const unsigned char *m, size_t bytes) noexcept;
31+
void poly1305_finish(poly1305_context *st, unsigned char mac[16]) noexcept;
32+
33+
} // namespace poly1305_donna
1334

1435
void poly1305_auth(unsigned char out[POLY1305_TAGLEN], const unsigned char *m, size_t inlen,
1536
const unsigned char key[POLY1305_KEYLEN]);

0 commit comments

Comments
 (0)